text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Generic second-order macroscopic traffic node model for general
multi-input multi-output road junctions via a dynamic system
approach
arXiv:1707.09346v1 [cs.SY] 28 Jul 2017
Matthew A. Wright and Roberto Horowitz
July 31, 2017
Abstract
This paper addresses an open problem in traffic modeling: the second-order macroscopic node problem. A
second-order macroscopic traffic model, in contrast to a first-order model, allows for variation of driving behavior
across subpopulations of vehicles in the flow. The second-order models are thus more descriptive (e.g., they have
been used to model variable mixtures of behaviorally-different traffic, like car/truck traffic, autonomous/humandriven traffic, etc.), but are much more complex. The second-order node problem is a particularly complex
problem, as it requires the resolution of discontinuities in traffic density and mixture characteristics, and solving
of throughflows for arbitrary numbers of input and output roads to a node (in other words, this is an arbitrarydimensional Riemann problem with two conserved quantities). We propose a solution to this problem by making
use of a recently-introduced dynamic system characterization of the first-order node model problem, which gives
insight and intuition as to the continuous-time dynamics implicit in first-order node models. We use this intuition
to extend the dynamic system node model to the second-order setting. We also extend the well-known “Generic
Class of Node Model” constraints to the second order and present a simple solution algorithm to the second-order
node problem. This node model has immediate applications in allowing modeling of behaviorally-complex traffic
flows of contemporary interest (like partially-autonomous-vehicle flows) in arbitrary road networks.
1
Introduction
The macroscopic approximation of vehicle traffic has proven a valuable tool for the study of traffic’s nonlinear
dynamics, and the design of methods for mitigating and controlling undesirable outcomes like congestion. This
macroscopic theory describes the dynamics of vehicles along roads with partial differential equations (PDEs) inspired
by fluid flow. The most basic macroscopic formulation is the so-called “kinematic wave” or “Lighthill-WhithamRichards” (LWR) due to [12, 14], which describes traffic with a one-dimensional conservation equation,
∂ρ ∂(ρv)
+
=0
∂t
∂z
(1)
where ρ(z, t) is the density of vehicles, t is time, z is the lineal direction along the road, and v(ρ) is the flow speed.
The total flow, ρv, is often expressed in terms of a flux function, f (ρ) = ρv (the flux function on a long straight road
is often called the fundamental diagram).
The LWR-type formulation (1) is a simple nonlinear model and cannot capture many characteristics of real
traffic flows. For example, a flux function f (·) of ρ only does not admit the oft-empirically-observed phenomenon
of accelerating and decelerating flows tracing a hysteresis loop in the (ρ, v) plane [20]. One extension of the LWR
model that can express a richer variety of dynamics is the so-called Aw-Rascle-Zhang (ARZ) [1, 20] family of models.
These models fit into the so-called “generic second order” 1 or “extended ARZ” class of traffic models [11], which can
be written as
∂ρ ∂(ρv)
+
=0
∂t
∂z
∂w
∂w
+v
=0
∂t
∂z
where v = V (ρ, w)
(2)
1 As seen in (2), the “second-order” model actually consists of two first-order partial differential equations (that is, they only contain
first derivatives). In a case of overloaded mathematical terminology, the name “second order” here comes from a system-theoretic view,
where a second-order system is one that has two state variables: in this case, ρ and w (or, equivalently, ρ and v).
1
where w(x, t) is a property or invariant that is conserved along trajectories [11].
The property w in (2) can be described as a characteristic of vehicles that determines their density-velocity
relationship. Members of the generic second order model (GSOM) family are differentiated by the choice of w and
its relationship on the ρ-v behavior. Examples of chosen w’s include the difference between vehicles’ speed and an
equilibrium speed [1], driver spacing [20], or the flow’s portion of autonomous vehicles [17]. An intuitive way of
describing the effect of the property w in (2) is that it parameterizes a family of flow models, f (ρ, w) = ρV (ρ, w),
with different flow models for different values of w [11, 5].
For application of macroscopic traffic simulation, road networks are often modeled as directed graphs. Edges
that represent individual roads are called links, and junctions where links meet are called nodes. Typically, the flow
model f (·) on links is called the “link model,” and the flow model at nodes is called the “node model.” Development
of accurate link and node models have been areas of much research activity in transportation engineering for many
years.
This paper focuses on node models for first- and second-order macroscopic models. The node model resolves the
discontinuities in ρ and/or w between links and determines a Neumann boundary For nodes with merges, diverges,
or both, this Riemann problem becomes multidimensional. Through this, the node model determines how the state
of an individual link affects and is affected by its connected links, their own connected links, and so on through the
network. As a result, it has recently been recognized that the specific node model used can have a very large role in
describing the network-scale congestion dynamics that emerge in complex and large networks (for more on this, see
the discussions in, e.g., the introduction sections of [16] and [8]).
In [18], we introduced a novel characterization of node models as dynamic systems. Traditional studies of node
models (see, e.g., [16, 6, 2, 15, 8, 19]) usually present the node model as an optimization problem (where the node
flows are found by solving this problem) or in algorithmic form (where an explicit set of steps are performed to
compute the flows across the node). In contrast, the dynamic system characterization describes the flows across
the node as themselves evolving over some period of time (in application, this means that the dynamic system
characterization presents time-varying dynamics that are said to occur during the simulation timesteps of the link
PDEs). The dynamic system characterization can be thought of as making explicit the time-varying behavior of
the flows at nodes of many algorithmic node models: it was shown in [18] that the dynamic system characterization
produces the same solutions as the algorithm introduced in [19], which also reduces to the one introduced in [16] as
a special case.
The dynamic system characterization has proven useful in imparting an intuition as to what physical processes
over time are implicit in these algorithmic node models (see the discussions referring to [18] in [19] for some examples).
In this paper, we develop a dynamic system characterization of a second-order node model, and use it to solve the
general node problem for second-order models.
This paper has several main contributions. The first is an extension of the dynamic system characterization of firstorder node models as introduced in [18] to a simple, closed-form solution algorithm. This represents the completion of
an argument began in Section 4.1 of that reference. The second contribution is the extension of the dynamic system
characterization to the generic second-order models. As we will see, the dynamic system characterization lends itself
to an intuitive incorporation of the second PDE in (2) that is not obvious in the traditional, optimization-problem
presentation of node models. The third contribution, and the principal contribution of this paper, parallels the first
by using the second-order dynamic system node model to derive an intuitive, closed-form algorithm for computing
node flows for second-order flow models for general, multi-input multi-output nodes. To the best of our knowledge,
this represents the first proposed generic (applicable to multi-input multi-output nodes) node model for second-order
traffic flow modeling2 .
The remainder of this paper is organized as follows. Section 2 reviews the first-order node flow problem, the
first-order dynamic system characterization introduced in [18], and presents the aforementioned closed-form solution
algorithm (contribution one in the above paragraph). Section 3 reviews the link discretization of the GSOM (2) as
presented in, e.g., [11, 5], which produces the inputs to our second-order node model, and the standard one-input
one-output second-order flow problem and its solution. Section 4 presents the extension of the second-order flow
problem to the multi-input multi-output case, the dynamic system characterization to the GSOM family (2) and
the solution algorithm for the general node problem (contributions two and three). Finally, Section 5 concludes and
notes some open problems.
2 A note on naming: as we will see in Section 2.1, we build off the so-called “generic class of first-order node models” to develop our
second-order node model. Given that the relevant second-order model used (2) is itself called the “generic second order model,” it might
be accurate to describe this paper’s results as the “genericization of the generic class of node models to the generic second-order model,”
but this description likely loses in comprehensibility what it might gain in accuracy.
2
2
First-order node model
In this Section, we review the general first-order node problem and a particular node model (and its solution algorithm). This node model will be extended to the second-order node problem in Section 3.
The traffic node problem is defined on a junction of M input links, indexed by i, and N output links, indexed by j.
We further define C classes (sometimes called “commodities”) of vehicle, indexed by c. The first-order node problem
c
takes as inputs the incoming links’ per-class demands Sic , split ratios βi,j
(which define the portion of vehicles of
class c in link i that wish to exit to link j), and outgoing links’ supplies Rj , and gives as outputs the set of flows from
c
c
c
i to j for class c, fi,j
. We denote as a shorthand the per-class directed demand Si,j
, βi,j
Sic . Nodes are generally
infinitesimally small and have no storage, so all the flow that enters the node must exit the node.
The rest of this Section is organized as follows. Section 2.1 defines our first-order node problem as an optimization
problem defined by explicit requirements, following the example set by [16]. Section 2.2 reviews the dynamic system
of [18] whose executions produce solutions to the node problem. Finally, Section 2.3 uses the dynamic system
formulation as a base to develop a node model solution algorithm. This algorithm represents the completion of an
argument began in [18].
2.1
“Generic Class of Node Model” requirements
The node problem’s history begins with the original formulation of macroscopic discretized first-order traffic flow
models [4]. There have been many developments in the node model theory since, but we reflect only some more
recent results.
We can divide the node model literature into pre- and post-[16] epochs. They drew from the literature several
earlier-proposed node model requirements to develop a set of conditions for first-order nodel models they call the
“generic class of first-order node models” (GCNM). These set of conditions give an excellent starting point for our
discussion of the mathematical technicalities of node models, and have been used as a starting point by many
subsequent papers, such as [6, 2, 15, 8, 19]. In the following list, we present the variant of first-order GCNM
requirements used in [19], which includes a modification of the first-in-first-out (FIFO) requirement (item 6 below)
to [19]’s “partial FIFO” requirement.
1. Applicability to general numbers of input links M and output links N . In the case of multi-class flow, this also
extends to general numbers of classes c.
P
c
.
2. Maximization of the total flow through the node. Mathematically, this may be expressed as max i,j,c fi,j
According to [16], this means that “each flow should be actively restricted by one of the constraints, otherwise
it would increase until it hits some constraint.” When a node model is formulated as a constrained optimization
problem, its solution will automatically satisfy this requirement. However, what this requirement really means
is that constraints should be stated correctly and not be overly simplified and, thus, overly restrictive for the
sake of convenient problem formulation. See the literature review in [16] for examples of node models that
inadvertently do not maximize node throughput by oversimplifying their requirements.
c
3. Non-negativity of all input-output flows. Mathematically, fi,j
≥ 0 for all i, j, c.
4. P
Flow conservation:
Total flow entering the node must be equal to total flow exiting the node. Mathematically,
P c
c
f
for
all c.
f
=
j i,j
i i,j
P c
P c
≤ Rj .
≤ Sic and i fi,j
5. Satisfaction of demand and supply constraints. Mathematically, j fi,j
6. Satisfaction of the (partial) first-in-first-out (FIFO) constraint: if a single destination j ′ for a given i is not
able to accept all demand from i to j ′ , then all other flows from i are constrained by the queue of j ′ -destined
c
is partially defined
vehicles that builds up in i. The degree to which this queue restricts the other flows fi,j
i
by the restriction intervals ηj ′ ,j = [y, z] ⊆ [0, 1]. This interval means that a queue in the i, j ′ movement
will block the portion of i, j-serving lanes in i with leftmost extent y and rightmost extent z (e.g., if i, j is a
through movement that uses two lanes and i, j ′ is a right-turn movement that uses the right of those two lanes,
then ηji ′ ,j = [1/2, 1].3 The traditional, full FIFO behavior, where any queue in i blocks all of i’s lanes, can be
recovered by setting all ηji ′ ,j = [0, 1] 4 .
3 Continuing this example, we will have η i
= [0, 1] since the only lane in i that serves movement i, j ′ (the right lane) will be blocked
j,j ′
by a queue for the through movement, which will queue on both lanes).
4 To help keep the meaning of η i
clear, we find it helpful to read it as “the restriction interval of j ′ onto j for i”
j ′ ,j
3
Another item that defines the partial first-in-first-out behavior is the amount of time that a restriction interval
ηji ′ ,j is active. That is, if j ′ is a low-supply link with relatively high demands and j ′′ is a high-supply link with
relatively low demands, it should be the case that ηji ′ ,j is active on a greater portion of the directed demand
c
Si,j
than ηj ′′ ,j . 5 We will see how this effect of time is captured in the dynamic system formulation of Section
2.2.
Finally, we require that we consider the cumulative effect of restriction intervals. Suppose that a movement i, j
has an active restriction from a queue for movement i, j ′ . Then, say that another downstream link j ′′ exhausts
its supply, and vehicles begin queueing for the movement i, j ′′ . Then, the new restriction on i, j (after this
second queue forms) is ηji ′ ,j ∪ ηji ′′ ,j .
This requirement is stated mathematically as
c
c
fi,j
≤ Si,j
[
fi,j ′
− A
ηji ′ ,j ×
Si,j , Si,j
S
i,j ′
′
(3)
j 6=j
c
where A(·) denotes the area of a two-dimensional object, × denotes a Cartesian product, fi,j
,
P c
Si,j , c Si,j .
P
c
fi,j , and
The formulation in (3) is complex in order to state it as an optimization constraint, and not as a consequence of
the time-varying queue formation intuition outlined in the third paragraph of this item. A major contribution
of the dynamic system approach to node modeling is the explicit encoding of this more intuitive description.
See [19, Sections 3.2-3.3] or [18, Section 3.2] for a much more in-depth discussion of this requirement.
7. Satisfaction of the invariance principle. If the flow from some input link i is restricted by the available output
supply, this input link enters a congested regime. This creates a queue in this input link and causes its demand
Si to jump to capacity Fi in an infinitesimal time, and therefore, a node model should yield solutions that are
invariant to replacing Si with Fi when flow from input link i is supply-constrained [9].
8. Supply restrictions on a flow from any given input link arePimposed on class P
components of this flow proporc
c
c
c
tionally to their per-class demands. Mathematically, fi,j
/( c fi,j
) = βi,j
Sic /( c βi,j
Sic ).
This assumes that the classes are mixed isotropically. This means that all vehicles attempting to take movement
i, j will be queued in roughly random order, and not, for example, having all vehicles of commodity c = 1 queued
in front of all vehicles of c = 2, in which case the c = 2 vehicles would be disproportionally affected by spillback.
We feel this is a reasonable assumption for situations where the demand at the node is dependent mainly on
the vehicles near the end of the link (e.g., in a small cell at the end).
In addition to the above numbered requirements, two other elements are needed to define a node model. The first
is a rule for the portioning of output link supplies Rj among the input links. Following [7], in [16] it was proposed to
allocate supply for incoming flows proportionally to input link capacities, which we will denote Fi . In this paper, we
allocate supply proportionally to the links’ “priorities” pi (in the spirit of [4, 13, 6, 19]). In the dynamic system view,
priorities represent the relative rate at which vehicles exit each link i to claim downstream space (one reasonable
formulation might be to follow the capacity-proportional example, pi = Fi , if, as in [16], it assumed that vehicles
exit a link at rate Fi ).
The second necessary element is a redistribution of “leftover supply.” Following the initial partitioning of supplies
Rj , if one or more of the supply-receiving input links does not fill its allocated supply, some rule must redistribute
the difference to other input links who may still fill it. This second element is meant to model the selfish behavior of
drivers to take any space available, and ties in closely with requirement 2 above. [16] referred to these two elements
collectively as a “supply constraint interaction rule” (SCIR). For some discussion of choices of SCIRs in recent papers,
see [19, Section 2.1].
In this paper, we consider a SCIR of the form [19]
X
X
fi,j <
Sic , Wi 6= ∅
(4a)
∀ i s.t.
j
c
5 This amount-of-time requirement is encoded in the right part of the Cartesian product in (3). This amount-of-time requirement
appears as a component of a Cartesian product in the two-dimensional rectangles that appear in [19, Section 3.2]. However, it is much
more intuitive to understand this as an explicit temporal property as it appears in the dynamic system characterization, so we will not
discuss [19, Section 3.2]’s derivation here.
4
pi,j
Rj ,
∀ i s.t. Wi 6= ∅, fi,j ≥ PM
i′ =1 pi′ ,j
X
c
c
where Wi = {j ∗ :
βi,j
∗ Si > 0,
∀j ∈ Wi
(4b)
c
pi′ ,j ∗ fi,j ∗ ≥ pi,j ∗ fi′ ,j ∗ ∀i′ 6= i}
where
P c
Si,j
(5)
pij = Pc c pi .
c Si
is the “oriented priority,” which distributes input
priority proportionally to the actual vehicles using that
P links’
c
(as before).
priority to claim downstream supply, and fi,j , c fi,j
The set Wi denotes all output links that restrict the flow from
P i. cThe cconditions for membership in Wi can be
read as “there is some nonzero demand for the movement i, j ∗ ( c βi,j
∗ Si > 0), and i claims at least its priorityproportional allocation of j ∗ ’s supply (pi′ ,j ∗ fi,j ∗ ≥ pi,j ∗ fi′ ,j ∗ ∀i′ 6= i).” Note that if a link j ∗ is in Wi ∩ Wi′ , then
pi′ ,j ∗ fi,j ∗ = pi,j ∗ fi′ ,j ∗ by construction.
Constraint (4a) says that if a link i is not able to fill its demand, then there is at least one output link in Wi that
restricts i, and that i’s movements claim at least as much as their oriented-priority-proportional allocation of supply.
Constraint (4b) captures the reallocation of “leftover” supply, which states that a link i that cannot fulfill all of its
demand to the links in Wi will continue to send vehicles after links i′ : j ∈
/ Wi′ have fulfilled their demands to the
j ∈ Wi .
This concludes the setup of our generic node model problem. A solution will have flows that are constrained by
at least one of the constraints outlined above. An algorithm to solve this problem and proof of optimality is given
in [19].
2.1.1
Other first-order node model requirements
Note that the list of first-order node requirements presented in Section 2.1 (which is the particular node problem of
interest for the remainder of this paper) is not an exhaustive list of all “reasonable” node model requirements. Since
the statement of the GCNM requirements in [16], several authors have proposed extensions or modifications (as we
have in the “partial FIFO” relaxation). Beyond what we have covered here, one of the most discussed are nodal
supply constraints. These supply constraints, as their name suggests, describe supply limitations at the node rather
than in one of the output links. They are meant to describe restrictions on traffic that occur due to interference
between flows in the junction (rather than vehicles being blocked in the input link), or the exhaustion of some “shared
resource” such as green light time at a signalized intersection. Each movement through the node may or may not
consume an amount of a node supply proportional to its throughflow.
The node supply constraints in the GCNM framework were originally proposed by in [16]. In [2] it was noted that
these node supplies may lead to non-unique solutions. Very recently, [8] revisited the node supply constraints (mostly
in the context of distribution of green time) to address [2]’s critique of non-uniqueness of solutions and proposed a
generalization of the flow-maximization objective that still enforces that drivers will take any available space they
can.
We do not explicitly include the node supply constraints in the dynamic system node models and resulting solution
algorithms in this paper. The path towards their inclusion in the first- and second-order cases is straightforward
but notationally cumbersome and somewhat beyond this paper’s scope of fusing the GCNM and second-order link
models.
2.2
Review of first-order node dynamic system
This Section reviews the node dynamic system characterization of node models presented in [18]. This dynamic
system is a hybrid system, which means that it contains both continuous and discrete states (also called discrete
modes). Here, the continuous states evolve in time according to differential equations, the differential equations
themselves change between discrete states, and the discrete state transitions are activated when conditions on the
continuous states are satisfied.
• Let there be N · M · C time-varying continuous states xci,j (t), each representing the number of vehicles of class
c that have taken movement i, j through the node. The continuous state space is denoted X.
• Let J be the set of all output links j. Let there be 2M discrete states qν , ν ∈ 2J (recall 2J refers to the power
set of J ), the index ν representing the set of downstream
links that have become congested. A downstream
P P
link j is said to “become congested” at time t if i c xcij (t) = Rj . The discrete state space is denoted Q.
5
• Init ⊆ Q × X defines the set of permissible initial states of the system at t = 0.
• Dom: Q → X denotes the domain of a discrete state, which is the space of permissible continuous states while
the discrete state is active.
• Φ : Q × X → Q × X is a reset relation, which defines the transitions between discrete states and the conditions
for those transitions.
• The hybrid system execution begins at time t = 0.
P
c
when i, j has a
• Each link is given a “time limit” Ti , ( c Sic )/pi . This is necessary to ensure that xci,j < Si,j
partial FIFO constraint active (which appears in the dynamic system as a flow rate attenuation).
Our hybrid system (Q, X, Init, ẋci,j , Dom, Φ) is
Q = {qν }, ν ∈ 2J
X =R
Init =
ẋci,j
=
(6a)
M·N ·C
(6b)
Q × {xci,j (t =
c
Si,j
P
p
ij
Sc
c i,j
·
∀i, j, c}
0) = 0
[
ηji ′ ,j
1−
′
j ∈ν,
c
∃ c: xci,j′ <Si,j
′
c
if xcij (t) < Sij
and t < Ti
0
otherwise
XX
x:
xci,j = Rj ∀j ∈ ν and
c
i
Dom(qν ) =
XX
/ν
xci,j ≤ Rj ∀j ∈
(6c)
(6d)
(6e)
c
i
Φ(qν , x) = (qν ′ , x) if
XX
xci,j ∗ = Rj ∗
(6f)
c
′
i
where ν = ν ∪ j ∗ .
When ẋcij = 0 for all i, j, c, the execution is complete and fijc = xcij .
It was shown in [18] that the hybrid system (6) produces the same solutions as [19]’s algorithm. In the following
Section, we show how to quickly compute executions of the hybrid system, which, since it is based on the continuoustime dynamics of (6), presents a more intuitive algorithm than the one in [19].
2.3
Execution of the first-order node dynamic system as a simple algorithm
Evaluating continuous-time or hybrid systems typically involves forward integration of the differential equation(s)
with fixed or varying step sizes. However, in the case of (6), evaluation can be performed in a much simpler manner.
This is due to the particular dynamics of the system - since the continuous-time dynamics and the condition for
discrete mode switching are very simple, the time that the next mode switch will occur can be found in closed form.
Equations (6e) and (6f) say that a mode switch where link j enters ν will occur when
XX
xcij = Rj .
(7)
i
c
Say we are currently at time t0 . Combining (7) with (6d), we can find the time that the mode switch occurs, which
we denote tj .
Z tj X X
XX
ẋcij dt.
(8)
Rj =
xcij (t0 ) +
i
t0
c
6
i
c
Solving the integral in (8),
Z
tj
t0
=
Z
XX
i
tj
t0
ẋci,j dt
c
XX
i
= (tj − t0 )
c
X
i
Then, plugging (9) into (8),
c
[
Si,j
i
pi,j P c 1 −
ηj ′ ,j dt
c Si,j
′
j ∈ν,
c
∃ c: xci,j′ <Si,j
′
[
i
ηj ′ ,j .
pij 1 −
(9)
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
P P
Rj − i c xci,j (t0 )
.
tj = t0 +
S i
P
1
−
η
p
j ′ ,j
i i,j
(10)
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
This value can be computed for each output link j. Then, the j with the smallest tj will be the first link to fill and
join ν. We had used j ∗ for this output link, so let tj ∗ , min tj . However, one of the input links may have its time
limit Ti expire. This would also change the dynamics, as it stops sending vehicles at that time.
Therefore, evaluation of the system trajectory beginning from t0 can be done by (i) evaluating (10) for each
output link, (ii) identifying t∗j , and (iii) checking whether any of the time limits Ti occur before tj ∗ . This is an
event-triggered simulation: it is only necessary to determine when the next event will occur. The equations for ẋcij
over [t0 , min({Ti }, tj ∗ )] can then be evaluated in closed form under qν .
Note that the ẋci,j ’s for an i may change to zero from nonzero without a change in the discrete state qν , if the
c
conditional of xci,j (t) < Si,j
in (6d) is broken. This can be understood as the i running out of vehicles that it is able
to send. This may happen if pi > Si for that i, and some (partial) FIFO constraint becomes active on i. In the
following algorithm, we introduce a new set, µ, that was not present in the dynamic system definition and contains
the i’s that either exhaust their supply or have their time limits expire (i.e., those i’s whose ẋci,j become zero without
j necessarily entering ν).
These steps are summarized in Algorithm 1. This algorithm represents the completion of an argument began in
[18].
3
3.1
Review of second-order flow modeling
Introduction
The formulation of the GSOM seen in (2) has been called the “advective form” [5]. In this form, the property w is
advected with the vehicles at speed v. That is, it is constant along trajectories. This form makes the statement that
the property w is a property of vehicles and is easy to understand conceptually.
However, to apply a discretization, it is useful to consider the total property ρw, and rewrite (2) in “conservative
form” [11, 5],
∂ρ ∂(ρv)
+
=0
∂t
∂z
∂(ρw) ∂(ρwv)
+
=0
∂t
∂z
where v = V (ρ, w).
(11)
We will review the relevant finite-volume discretization using the Godunov scheme of (11) [11] in the next Section.
For a deeper analysis on the physical properties of (11), see, e.g., [11].
We make one note on constraints imposed on the form of v(ρ, w) in (11). It has been stated [11, (19)] that, to
apply the Godunov discretization to (11), one is restricted to choices of V (ρ, w) for which there is a unique ρ for
every (v, w) and a unique w for every (v, ρ). That is, V (ρ, w) must be invertible in both its arguments.
7
Algorithm 1 First-order node model solution algorithm
c
Input: Sic , βi,j
, Rj , pi , ηji ′ ,j
c
Output: fi,j
c
c
Ti , Si,j
, fi,j
, pi,j , µ, ẋci,j , ν ← (Algorithm 2) {Setup and initialization}
t ← 0 {Begin main loop}
while ∃ i, j, c : ẋci,j 6= 0 do
for all i, j, c : i ∈
/ µ do
c
S i
Si,j
c
ẋi,j ← pi,j P S c · 1 −
ηj ′ ,j {Compute continuous-time dynamics}
c
i,j
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
end for
dt ← (Algorithm 3)
for all i, j, c do
c
c
fi,j
← fi,j
+ ẋci,j · dt {Integrate forward in time}
end for
c
c
for all i : ∀j, c, Si,j
− fi,j
= 0 do
µ ← µ ∪ i {Account for all emptied input links}
ẋci,j ← 0 ∀ j, c
end for
for all j : t + dt = tj do
ν ← ν ∪ j {Account for all filled output links}
end for
t ← t + dt
end while
c
return fi,j
Algorithm 2 Setup and initialization
for all i P
do
Ti ← ( c Sic )/pi
for all j do
for all c do
c
c
Si,j
← βi,j
Sic
c
fi,j ← 0
end for P
P
c
)/( c Sic )
pi,j ← pi · ( c Si,j
end for
end for
µ←∅
for all i : ∀c, Sic = 0 do
µ ←µ∪i
ẋci,j ← 0 ∀j, c
end for
ν←∅
for all j : Rj = 0 do
ν ←ν∪j
end for
c
c
return Ti , Si,j
, fi,j
, pi,j , µ, ẋci,j , ν
8
Algorithm 3 Computing the time to integrate forward, first-order case
for all j ∈
/ ν do P P
c
)/ẋci,j {Compute the filling time for every output link}
tj ← t + (Rj − i c fi,j
end for
for all i, j : i ∈
/ µ and j ∈
/ ν do
c
c
ti,j ← (Si,j
− fi,j
)/ẋci,j for any c {by construction (Section 2.1, item 8), all c’s fulfill their demands at the same
time.}
end for
dt ← min({ti,j }i∈µ,j
/
∈ν
/ , {tj }j ∈ν
/ , {Ti }i∈µ
/ )
return dt
3.2
Godunov discretization of the GSOM
The Godunov discretization of the first-order (LWR) model (1), first introduced as the Cell Transmission Model [3]
is well-known. The Godunov scheme discretizes a conservation law into small finite-volume cells. Each cell has a
constant value of the conserved quantity, and inter-cell fluxes are computed by solving Riemann problems at each
boundary. The Godunov scheme is a first-order method, so it is useful for simulating solutions to PDEs with no
second- or higher-order derivatives like the LWR formulation. In the CTM, the Riemann problem is stated in the
form of the demand and supply functions.
Since (11) is also a conservation law with no second- or higher-order derivatives, the Godunov scheme is applicable
as well [11]. However, due to the second PDE for ρw, an intermediate state arises in the Riemann problem and its
solution [1, 20, 10, 11]. This intermediate state has not always had a clear physical meaning, and this lack of
clarity likely inhibited the extension of the Godunov discretization to the multi-input multi-output node case. In
our following outline of the discretized one-input one-output flow problem, we make use of a physical interpretation
of the intermediate state due to [5].
A final note: in the first-order node model, we were able to ignore the first-order demand and supply functions
that generated the supplies Rj and per-class demands Sic . That is, we were agnostic to the method by which they
were computed (and to the input and output link densities), as they did not change during evaluation of the node
problem. As we will see shortly, this is not the case for the second-order flow problem (due to the intermediate state
and its interactions with the downstream link). Therefore, our explanation below makes use of the second-order
demand and supply functions S(ρ, w) and R(ρ, w), respectively.
3.2.1
Preliminaries
In this paper, we say that each vehicle class c has its own property value wc . The net (averaged over vehicle classes)
property of a link ℓ, denoted wℓ , is
P c c
w ρℓ
wℓ = c
(12)
ρℓ
P
where ρℓ , c ρcℓ is the total density of link ℓ.
In the second-order model, the fundamental diagram of a link is a function of both net density and net property
as defined above. This carries over to the demand and supply functions in the Godunov discretization [11, 5]. That
means that the supply and demand are defined at the link level with the net quantities ρℓ and wℓ . For an input link
i,
(
ρi vi
if ρi ≤ ρc (wi )
Si = S(ρi , wi ) =
(13)
F (wi ) if ρi > ρc (wi )
where ρc (wi ) is the critical density for property value wi and F (wi ) is the capacity for property value wi .
The demand from (13) is split among the classes and movements proportional to their densities and split ratios,
ρci
ρi
c
= βi,j
Sic .
Sic = Si
c
Si,j
The oriented priorities pi,j are computed according to (5), as before.
9
3.2.2
Computing supply
Solving for an output link’s supply is a much more complicated problem. We will begin our discussion with a review
of the one-input-one-output case [5, Sections 3.3, 3.4].
The supply R of the output link in this one-to-one case, where i is the input link and j is the output link, is
(
F (wi ) if ρM ≤ ρc (wi )
Rj = R(ρM , wi ) =
(14)
ρM vM if ρM > ρc (wi ).
We see that the supply of the downstream link is actually a function of the upstream link’s vehicles’ property, and
the density and speed of some “middle” state, M . The middle state is given by [5, (16)]
wM = wi
(
V (0, wi ) if V (0, wi ) < vj
vM =
vj
otherwise
ρM s.t. vM = V (ρM , wM )
(15a)
(15b)
(15c)
where vj = V (ρj , wj ) is the velocity of the downstream link’s vehicles and V (·) is the velocity function as given by
the fundamental diagram.
In [5], the intuition behind the meaning of the middle state is given as follows: the middle state vehicles are
actually those that are leaving the upstream link i and entering the downstream link j. As they leave i and enter
j, they clearly carry their own property (15a), but their velocity is upper-bounded by the velocity at which that the
downstream vehicles exit link j and free up the space that the i-to-j vehicles enter (15b). The middle density, ρM
(and therefore the downstream supply R), is then determined by both the upstream vehicles’ characteristics (i.e.,
wi ) and the downstream link’s flow characteristics (through vj ). In other words, the number of vehicles that can
fit into whatever space is freed up in the downstream link is a function of the drivers’ willingness to pack together
(defined by wi ). Since the meaning of supply Rj is “the number of vehicles that j can accept,” this means that Rj
is dependent on wi (15c).
Note that (15b) is also the equation by which congestion spills back from j to i: if j is highly congested, then vj
will be low. This then makes ρM large in (15c), which in turn leads to a small Rj in (14).
Now that we have reviewed the 1-to-1 case, we can consider how to generalize this to a multi-input-multi-output
node when we determine supply for several links.
4
4.1
Second-order node model
Multi-input-multi-output case
We saw that the reasoning behind the dependence of Rj on wi was that the spacing tendencies of i’s vehicles determine
the number of vehicles that can fit in j. Therefore, in generalizing to a multi-input-multi-output node, it makes sense
to define a link j’s “middle state” as being dependent on the vehicles actually entering link j. That is, if wj − , the w
just upstream of j, is the “middle state” of link j, then we say
P P
c
i
c wi ẋi,j
P
.
(16a)
wj − = P
c
i
c ẋi,j
The “j-upstream middle state” velocity and density, vj − and ρj − , are then
(
V (0, wj − ) if V (0, wj − ) < vj
vj − =
vj
otherwise
ρj − s.t. vj − = V (ρj − , wj − )
(16b)
(16c)
and the supply Rj is
Rj = R(ρj − , wj − ) =
(
F (wj − ) if ρj − ≤ ρc (wj − )
ρj − vj − if ρj − > ρc (wj − ).
(17)
Note that in (16a), we defined wj − as a function of ẋcij . Recall from the first-order node model that the ẋcij ’s
can change as (i) upstream links i exhaust their demand or (ii) downstream links j run out of supply. These two
10
events correspond to discrete state changes in our hybrid system. This, of course, carries over to the second-order
node model. This means that the j − quantities, and thus the supply Rj , change as ẋcij ’s change. Therefore, at each
discrete state transition, we need to determine the new supply for each output link j for the new mixture of vehicles
that will be entering j in the next discrete state.
We will explain how this is done through the following example. Suppose that at time t0 , we compute some
wj − , vj − , ρj − , and Rj with (16)-(17). Then, at time t1 , one of the ẋci,j for that j changes. At that point, we
recompute ρj and wj ,
ρcj (t1 ) = ρcj (t0 ) +
ρj (t1 ) =
wj (t1 ) =
X
c
P
1 X c
x (t1 )
Lj i ij
ρcj (t1 )
(18)
wc ρcj (t1 )
ρj (t1 )
c
where Lj is the length of j. Then, we recompute all the “middle state” variables and Rj using (16)-(17). Critically,
note that in this recomputation, the new vj at t1 is vj (t1 ) = V (ρj (t1 ), wj (t1 )). This means that vj − (t1 ) will also be
different than vj − (t0 ). This will carry through to create a Rj (t1 ) that is different from Rj (t0 ), and takes into account
both the vehicles that have moved into j between t0 and t1 , and the difference in properties wj − (t0 ) and wj − (t1 ).
Note that if wj − (t1 ) leads to significantly tighter packing (i.e., smaller inter-vehicle spacing) than wj − (t0 ), it is
conceivable that we will have Rj (t1 ) > Rj (t0 ) (especially if ρj (t0 ) is not that much smaller than ρj (t1 )).
Of course, the description above assumes isotropic mixing of all vehicle classes in the link j (recall we stated this
assumption for input links i in item (8) of the first-order GCNM requirements in Section 2.1).
Unlike supply, demand does not need to be recomputed since we assume the mixture of vehicles demanding each
movement remains the same (due to our isotropic mixture assumption)
In summary, we state the second-order generalization of the GCNM requirements as the same as the first-order
requirements P
stated in Section
P c2.1,c with the addition of a constraint enforcing the conservation of property via the
c
w , and the modification of the supply constraint such that the supply is computed
wc = j fi,j
second PDE: i fi,j
from the second-order fundamental diagram using the property of its incoming flows. This second point, where the
supply constraint is also dependent on the flow solution, only worsens the nonconvexity of the node problem. Indeed,
we are drifting away from a setting where the optimization-problem makes the most sense, and it may be more
helpful to understanding to consider the physical dynamics encoded by the solution methods.
In any case, we now have all the ingredients necessary to extend our first-order hybrid system node model to the
second-order formulation.
4.2
Dynamic system definition
We state the second-order node dynamic system as an extension to the first-order one presented in 2.2. Most of the
symbols remain the same. However, we make a few changes:
• Let µ ∈ 2I (where I is the set of all input links i) be the set of all exhausted input links (this set was introduced
in the first-order algorithm in Section 2.3). This is necessary to state the recalculations of supply according to
the steps in Section 4.1 when a link exhausts its demand and the net property of a j − changes.
• Paralleling j ∗ , let i∗ denote an exhausted input link. An input link is said to be exhausted at time t if
c
Si,j
− xci,j (t) = 0 ∀j, c.PNote that the formula for the time of demand exhaustion remains the same as in the
first-order case, Ti = ( c Sic )/pi .
• To accommodate recomputing of supply using (18), we will add more continuous states: the N ·M ·C quantities
x̄ci,j , which will denote the flow of movement i, j for class c for movement c since the last time that supplies
have been recalculated, and the M · C per-class densities of the output links, ρcj . This is necessary because,
following (16)-(18), the new supplies Rj for the new wj − will also take into account the vehicles xci,j , that have
already made the movement, so when determining when a link j is filled with its new supply we will need a
“fresh” counter of vehicles that have entered it.
• We assume we have the initial ρcj (0) for all j, c.
11
Our hybrid system (Q, X, Init, ẋci,j , x̄˙ ci,j , ρ̇cj , Dom, Φ) is
Q = {qµ,ν }, µ ∈ 2I , ν ∈ 2J
(19a)
X = RM·N ·C × RM·N ·C × RM·C
c
∀i, j, c;
xi,j (t = 0) = 0
∀i, j, c;
Init = Q × x̄ci,j (t = 0) = 0
c
ρj (t = 0) = ρcj (0)
∀j, c
c
Si,j
P
p
ij
c
c Si,j
[
if i ∈
/µ
i
c
ηj ′ ,j
· 1−
ẋi,j =
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
0
otherwise
x̄˙ ci,j = ẋci,j
X
ẋci,j
ρ̇cj =
(19b)
(19c)
(19d)
(19e)
(19f)
i
c
x : t ≥ Ti ∨ ∀j, c : xci,j = Si,j
c
t ≤ Ti , ∃j, c : xci,j ≤ Si,j
XX
q
x̄ci,j = Rj µ,ν
Dom(qµ,ν ) =
c
i
XX
q
x̄ci,j ≤ Rj µ,ν
i
c
′
(qµ,ν ′ , x )
Φ(qµ,ν , x) = (qµ′ ,ν , x′ )
∀i ∈ µ,
∀i ∈
/ µ,
∀j ∈ ν,
∀j ∈
/ ν
P P c
qµ,ν
if
i
c x̄i,j ∗ = Rj ∗
where ν ′ = ν ∪ j ∗
if t = Ti
c
∨ ∀j, c xci,j = Si,j
′
∗
where µ = µ ∪ i .
(19g)
(19h)
′c
c
c
c
where x′ = ({x′c
i,j }, {x̄i,j }, {ρj }) = ({xi,j }, {0}, ρj )
P c c
q
c w ρj
and Rj µ,ν from (16), (17), with wj = P
c
c ρj
When ẋcij = 0 for all i, j, c, the execution is complete and fijc = xcij .
Unsurprisingly, the second-order dynamic system is more complicated than the first-order one. The reader will
note that the discrete dynamics, as discussed before, are triggered by links j ∗ filling and links i∗ emptying. The filling
of a j ∗ and its entering into ν remains the same as the first-order system. The emptying or time-expiry of input
links, rather than being encoded in the continuous dynamics as was done in the first-order system’s (6d), is now in
the discrete dynamics in (19g),(19h). While it was possible to reduce the number of discrete states in the first-order
system by including i-emptying in the continuous dynamics, in the second order system, any change in the continuous
dynamics changes the output links’ wj − , so all continuous dynamics changes must trigger a recomputation of Rj ,
which, in (19), we do when µ or ν change.
Thankfully, although the second-order system seems much more complex than the first-order system, the secondorder solution algorithm is not that much more complicated than the solution method of the first-order system. We
will see why in the next Section.
4.3
Solution algorithm
Note that, just as in the first-order system, the second order system has constant continuous dynamics in each
discrete state. This means that, just as in the first-order case, we can easily compute the time that the next discrete
state transition occurs. Like in Section 2.3,
Pthis is the smallest of the tj ’s and Ti ’s. As we said, the input link “time
limits” remain the same as before, Ti = ( c Sic )/pi . The time that an output link runs out of supply and is filled
under the discrete state qµ,ν , if t0 is the time that the discrete state switched to qµ,ν and j’s supply was recomputed,
12
is similar to (10),
q
Rj µ,ν
tj = t0 + P P P
i
j
q
c
c ẋi,j
= t0 +
R µ,ν
j
S i
P
1
−
η
p
i,j
j ′ ,j
i∈µ
/
(20)
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
q
but differs in two key ways. First, the term for supply is the recomputed Rj µ,ν from (16), (17) (this also accounts
for why the numerator in (20) does not have a subtracted quantity as in (10), as that subtraction of already-filled
supply is accounted for in the recomputed supply. Second is that the denominator is summed over i ∈
/ µ rather than
all i, as the set µ is not in the definition of the first-order dynamic system as stated in Section 2.2.
We now state solution algorithm for the second-order dynamic system. It follows the same logic as the first-order
case: identifying the next Ti or tj to occur, finding the constant continuous-time dynamics that the system will
evolve under until that time, integrating forward in time, a new step of recomputing supply, and repeating.
Algorithm 4 Second-order node model solution algorithm
c
, pi , ηji ′ ,j , Rj {Rj is only the initial Rj and will be re-computed}
Input: Sic , ρcj , wc , βi,j
c
Output: fi,j
c
c
Ti , Si,j
, fi,j
, pi,j , µ, ẋci,j , ν, wi , x̄ci,j , ← (Algorithm 5)
t ← 0 {Begin main loop}
while ∃ i, j, c : ẋci,j 6= 0 do
for all i, j, c : i ∈
/ µ do
c
S i
Si,j
c
ηj ′ ,j {Compute continuous-time dynamics}
ẋi,j ← pi,j P S c · 1 −
c
i,j
j ′ ∈ν,
c
∃ c: xci,j′ <Si,j
′
end for
dt ← (Algorithm 6)
for all i, j, c do
c
c
fi,j
← fi,j
+ ẋci,j · dt
c
c
x̄i,j ← ẋi,j · dt
end for
ρj , wj ← (Algorithm 8)
c
c
for all i : ∀j, c, Si,j
− fi,j
= 0 do
µ ← µ ∪ i {Account for all emptied input links}
ẋci,j ← 0 ∀ j, c
end for
for all j : t + dt = tj do
ν ← ν ∪ j {Account for all filled output links}
end for
t ← t + dt
end while
c
return fi,j
Algorithm 5 Setup and initialization, second-order case
c
c
Ti , Si,j
, fi,j
, pi,j , µ, ẋci,j , ν ← (Algorithm 2)
for all iPdo c c
w ρ
wi = cρi i
end for
for all i, j, c do
x̄ci,j ← 0
end for
c
c
return Ti , Si,j
, fi,j
, pi,j , µ, ẋci,j , ν, wi , x̄ci,j
13
Algorithm 6 Computing the time to integrate forward, second-order case
for all j do
Rj ← (Algorithm 7)
end for
for all j ∈
/ ν do
tj ← t + Rj /ẋci,j {Compute the filling time for every output link}
end for
for all i, j : i ∈
/ µ and j ∈
/ ν do
c
c
ti,j ← (Si,j
− fi,j
)/ẋci,j for any c
end for
dt ← min({ti,j }i∈µ,j
/
∈ν
/ , {tj }j ∈ν
/ , {Ti }i∈µ
/ )
return dt
Algorithm 7 Second-order computation of supply
wj − =
P P
c
i P
c wi ẋi,j
P
c
i
c ẋi,j
if V (0, wj − ) < vj then
vj − = V (0, wj − )
else
vj − = vj
end if
ρj − s.t. vj − = V (ρj − , wj − )
if ρj − ≤ ρc (wj − ) then
Rj = R(ρj − , wj − )
else
Rj = F (wj − )
end if
return Rj
Algorithm 8 Recomputing the downstream links’ density and property
P
ρcj ← ρcj + L1j i xcij
P c
ρj ← Pc ρj
w c ρc
wj ← cρj j
return ρj , wj
14
4.4
Second-order extension of the GCNM requirements?
In solving the second-order node problem, the fact that the supply must be continually recalculated can be interpreted
as indicating that the use of the supply and demand quantities is not as natural as in the first-order case. We see
that demand S(ρ, w) and supply R(ρ, w) alone (i.e., not including the ρ and w) is not enough to solve the node
problem in the second-order case: the link ρ and w quantities are required. This is not unnatural: the node problem
is, after all, a Riemann problem to resolve discontinuities in ρ and w. In the first-order case, the node problem is
often stated in terms of supply and demand instead of the actual conserved quantity ρ because i) they have a more
intuitive physical meaning, and ii) since link densities are not needed beyond their use in S(ρ) and R(ρ) for the
first-order problem, beginning with S(ρ) and R(ρ) simplifies the problem by removing one step. However, we have
seen that using the beginning-with-supply-and-demand framework in the second-order case does not simplify the
problem along the lines of ii, as we still need to make use of ρ and w. Therefore, in the future it may make more
sense to state the second-order node problem as taking inputs of ρ and w for all links, rather than its inputs being
S(·) and R(·). That would remove the unintuitive nature of needing to “recompute supply.”
5
Conclusion
This paper presented a generalization of the widely-used “Generic Class of Node Model” macroscopic traffic junction
models to the so-called “General Second Order Model” flow model. This paper’s results allow the extension of
macroscopic modeling of variable-behavior flows (based on different mixtures of driving behavior) to complex general
networks. Many of these flows and networks had been only able to be modeled by microscopic models that consider
the behavioral variability on a per-car level, but macroscopic models that can capture the aggregate features as a
more granular model can greatly increase the scale of problems that we are able to study. As stated before, the
second-order flow models have been used to represent flows of great contemporary interest, such as mixtures of
human-driven and autonomous vehicles [17]. Researchers and practitioners will need to use every tool available to
understand and predict the system-level changes that will arise from the traffic demand changing not just in size,
but in characteristics.
Some immediate avenues for future refinement of second-order macroscopic models presented themselves during
this paper. As mentioned in Section 2.1.1, we do not address node supply constraints in this paper’s node models.
However, the immediate application of a general, multi-input-multi-output second-order node model, macroscopic
simulation of mixed-human-driven-and-autonomous traffic on complex networks, is of particular concern in scheduling
problems involving green light timing. Future work, then, should incorporate the node supply constraints into the
general second-order node problem so that they may be used in signal optimization and the still-developing potential
that connected and automated vehicles bring to traffic control.
References
[1] A. Aw and M. Rascle. Resurrection of ‘second order’ models of traffc flow. SIAM Journal on Applied Mathematics, 63(1):916–938, 2000.
[2] R. Corthout, G. Flötteröd, F. Viti, and C. M. Tampére. Non-unique flows in macroscopic first-order intersection
models. Transportation Research Part B: Methodological, 46(3):343–359, Mar. 2012.
[3] C. Daganzo. The cell transmission model: A dynamic representation of highway traffic consistent with the
hydrodynamic theory. Transportation Research Part B: Methodological, 28(4):269–287, 1994.
[4] C. Daganzo. The cell transmission model, Part II: Network traffic. Transportation Research Part B: Methodological, 29(2):79–93, 1995.
[5] S. Fan, Y. Sun, B. Piccoli, B. Seibold, and D. B. Work. A Collapsed Generalized Aw-Rascle-Zhang Model and
Its Model Accuracy. arXiv preprint arXiv:1702.03624, 2017.
[6] G. Flötteröd and J. Rohde. Operational macroscopic modeling of complex urban road intersections. Transportation Research Part B: Methodological, 45(6):903–922, July 2011.
[7] G. Gentile, L. Meschini, and N. Papola. Spillback congestion in dynamic traffic assignment: a macroscopic flow
model with time-varying bottlenecks. Transportation Research Part B: Methodological, 41(10):1114–1138, 2007.
15
[8] S. E. Jabari. Node modeling for congested urban road networks. Transportation Research Part B: Methodological,
91:229–249, Sept. 2016.
[9] J. Lebacque and M. Khoshyaran. First-order macroscopic traffic flow models: intersection modeling, network
modeling. In The 16th International Symposium on Transportation and Traffic Theory (ISTTT), pages 365–386,
2005.
[10] J.-P. Lebacque, S. Mammar, and H. Haj-Salem. The Aw–Rascle and Zhang’s model: Vacuum problems, existence
and regularity of the solutions of the Riemann problem. Transportation Research Part B: Methodological,
41(7):710–721, Aug. 2007.
[11] J.-P. Lebacque, S. Mammar, and H. H. Salem. Generic second order traffic flow modelling. In Transportation
and Traffic Theory 2007. Papers Selected for Presentation at ISTTT17, pages 755–776, 2007.
[12] M. Lighthill and G. Whitham. On kinematic waves I: Flow movement in long rivers. II: A theory of traffic flow
on long crowded roads. Proc. Royal Society of London, Part A, 229(1178):281–345, 1955.
[13] D. Ni and J. D. Leonard. A simplified kinematic wave model at a merge bottleneck. Applied Mathematical
Modelling, 29(11):1054–1072, Nov. 2005.
[14] P. Richards. Shock waves on the highway. Operations Research, 4(1):42–51, 1956.
[15] E.-S. Smits, M. C. Bliemer, A. J. Pel, and B. van Arem. A family of macroscopic node models. Transportation
Research Part B: Methodological, 74:20–39, Apr. 2015.
[16] C. M. J. Tampère, R. Corthout, D. Cattrysse, and L. H. Immers. A generic class of first order node models for
dynamic macroscopic simulation of traffic flows. Transportation Research, Part B, 45(1):289–309, 2011.
[17] R. Wang, Y. Li, and D. B. Work. Comparing traffic state estimators for mixed human and automated traffic
flows. Transportation Research Part C: Emerging Technologies, 78:95–110, May 2017.
[18] M. Wright, R. Horowitz, and A. A. Kurzhanskiy. A dynamic system characterization of road network node
models. In Proceedings of the 10th IFAC Symposium on Nonlinear Control Systems, volume 49, pages 1054–
1059, August 2016.
[19] M. A. Wright, G. Gomes, R. Horowitz, and A. A. Kurzhanskiy. On node and route choice models for highdimensional road networks. Submitted to Transportation Research Part B, 2016.
[20] H. M. Zhang. A non-equilibrium traffic model devoid of gas-like behavior. Transportation Research Part B:
Methodological, 36(3):275–290, 2002.
16
| 3cs.SY
|
From the Skin-Depth Equation to the Inverse RFEC
Sensor Model
arXiv:1609.04810v1 [cs.CE] 15 Sep 2016
Raphael Falque, Teresa Vidal-Calleja, Gamini Dissanayake, and Jaime Valls Miro
University of Technology Sydney, Australia
Emails: [email protected], [email protected],
[email protected], [email protected]
Abstract—In this paper, we tackle the direct and inverse
problems for the Remote-Field Eddy-Current (RFEC) technology.
The direct problem is the sensor model, where given the geometry
the measurements are obtained. Conversely, the inverse problem
is where the geometry needs to be estimated given the field
measurements. These problems are particularly important in
the field of Non-Destructive Testing (NDT) because they allow
assessing the quality of the structure monitored. We solve the
direct problem in a parametric fashion using Least Absolute
Shrinkage and Selection Operation (LASSO). The proposed
inverse model uses the parameters from the direct model to
recover the thickness using least squares producing the optimal
solution given the direct model. This study is restricted to the
2D axisymmetric scenario. Both, direct and inverse models, are
validated using a Finite Element Analysis (FEA) environment
with realistic pipe profiles.
Keywords: Remote Field Eddy Current (RFEC), direct
problem, inverse problem, Non Destructive Evaluation (NDE)
I. I NTRODUCTION
The Remote Field Eddy Current (RFEC) technology allows
in-line inspection of ferromagnetic pipelines. Tools based on
this technology are usually composed of an exciter coil and
one or several receivers. The exciter coil, driven by a lowfrequency alternative current, generates an electromagnetic
field that flows outside the pipe near the exciter coil and flows
back inward the pipe at a remote area as shown in Fig. 1(a).
The receivers are located in the remote part and record the
magnetic field. As shown in the figure the magnetic field
passes twice the pipe’s wall; this phenomenon is commonly
referred as the double through wall in the literature [1].
When the magnetic field flows through a ferromagnetic
medium (i.e. the pipe), the amplitude of the magnetic field
is attenuated, and the phase is delayed. Due to the double
through wall penetration, the magnetic field recorded by the
receiver has been modified by different areas of the pipe: when
it flows outward the pipe near the exciter coil and when it
flows backwards the pipe in the remote area. Hence inferring
the geometry of the pipe from the signal information is a
challenging task since a single measurement is correlated with
different areas of the geometry.
Inferring the pipe’s geometry from the tool signal corresponds to solving the inverse problem of the RFEC. This
problem has been studied in the literature for the 2D axisymmetrical case of a perfect pipe with a single crack. The
problem is then formulated as recovering the shape (size and
Fig. 1: Representation of the RFEC phenomenon. From the
global phenomenon (a), we propose a parametric direct model
that consider independently the flow of the magnetic field in
the air (b), and the local attenuation due to the magnetic field
flowing through the pipe thickness (c).
width) of the single defect [2–4]. These approaches solve the
problem using data-driven techniques and bypass the problem
of recovering the full pipe’s geometry. These solutions fit the
case of steel material, where most pipe bursts are due to cracks;
in the case of cast-iron pipes, the material is more sensitive to
corrosion. Hence the geometry of the pipe has a more organic
shape rather than a single isolated crack. Therefore, for castiron pipes recovering the full pipe’s geometry is critical.
Some other approaches from the literature consist of modifying the tool design. The use of several receivers located
at different axial locations from the exciter coil allows using
redundancy of the information provided by passing through
the same location to recover the full pipe’s geometry [5], [6].
However, this approach leads to longer tools and require more
electrical power to operate the multiple sensors and/or exciter
coils. Due to the in-line nature of the RFEC tools, the mobility
and the battery consumption have to be optimised.
Particularly for this work, to allow simple hardware design,
we consider the case of an elementary RFEC tool composed
of a single exciter coil and a single receiver. The aim of
this paper is to obtain an inverse sensor model of the RFEC
phenomenon which, given a set of continuous magnetic field
measurements, allows to recover the full pipe’s geometry for
a 2D axisymmetric scenario.
The remainder of the paper is organised as follows. In
Sec. II, we give conceptual ideas about the behaviour of the
magnetic field. We then propose a direct model solved using
Least Absolute Shrinkage and Selection Operation (LASSO).
From the direct model, we derive an inverse model formulated
in a least-square form. The dataset generated with Finite
Element Analysis (FEA) and the experimental results are given
in Sec. III. We finally discuss the performance and limitations
of the proposed model in Sec. IV.
II. M ODELLING OF THE RFEC PHENOMENON
The direct problem of the RFEC phenomenon consists
of mapping the pipe’s geometry to the sensor measurement
through a sensor model. Conversely, the inverse (or indirect)
problem consists of finding the model that maps the sensor
measurements into the pipe’s geometry. The main goal is to
solve the inverse problem, however, solving the direct problem
provides qualitative and quantitative information on the form
of the inverse model. Before to consider the direct and inverse
problem, we discuss some high-level insight of the RFEC
technology. Particular attention is dedicated to understanding
how the geometry near to the exciter coil impacts the sensor
measurements. Qualitative descriptions of the overall RFEC
phenomenon have been broadly studied and in depth descriptions are available in the literature [1], [7], [8].
A. Background information
As shown in [1], it is possible to consider a defect in the
pipe’s geometry as an anomalous source model. The defect
is then replaced by an independent source of magnetic field
superposed to the pipe (see Fig. 12 in [1]). Knowing that
the magnetic field gets attenuated while travelling through
a ferromagnetic medium, the idea is to replace the lack of
attenuation from the defect by a source of magnetic field
superposed to a perfect pipe. Following the same idea, one
could consider the pipe’s thickness as an attenuation of the
signal.
Let us consider a pipe with an organic geometry (i.e. a
corroded pipe), defined by a piecewise constant profile (as
shown in Fig. 1). Each piece is then considered as a local
source of attenuation. We then dissociate the global RFEC
phenomenon shown in Fig. 1(a) in two part: (i) the attenuation
due to the magnetic field flowing in the air, and (ii) the
local attenuation due to the magnetic field flowing through
the pipe (the former one is shown in Fig. 1(b) and the latter
in Fig. 1(c)).
The global attenuation of the magnetic field propagating in
the air (i), mostly due to field radiating from the coil, is a
constant term for a given excitation and global geometry. The
definition of this value is however complex since it involves
many parameters (e.g. dimensions and excitation of the coil,
diameter of the pipe, distance between the exciter and the
receiver).
The local interaction of the electromagnetic wave with the
pipe (ii) can be described as a plane wave propagating through
a homogeneous, isotropic, and conductive medium (i.e. the
pipe). This phenomenon can be described by deriving the skin
depth equation from the Maxwell equations and can be written
as follow;
s
s
ωµσ
ωµσ
t+ωt
−j
t
−
2
2 e
(1)
B(t) = B0 e
{z
},
|
{z
}|
phase contribution
amplitude
with B the magnetic field, B0 the initial value of the magnetic
field, ω the frequency, µ the magnetic permeability of the
medium, σ the electrical conductivity, and t the distance
travelled by the wave. The amplitude and the phase-lag are
usually the measurements recorded by the RFEC tools since
they have a log-linear or linear relationship with the thickness
of the conductive medium:
r
ωµσ
t
φlocal =
2
r
(2)
ωµσ
t
ln(B)local = ln(B0 ) −
2
In this paper, we model the direct and inverse problem
uniquely with the amplitude. However, a similar study could
be done with the phase-lag.
B. Direct problem
We now consider the direct problem, which consists of
finding a function h such that h : t −→ y, where y is
the sensor measurements and t a set of thickness values that
describe the pipe’s geometry around the RFEC tool.
Let us first consider the case of a single measurement. Using
the wave superposition principle, we can then add (i) and (ii)
as follow;
k
X
y = y0 −
wi ti + ,
(3)
i=1
with y0 the constant term described in (i), ti the ith thicknesses
of the pipe piece (the pipe’s geometry is approximated as a
piecewise constant profile a shown in Fig.
p 1(a)), and wi the
unknown parameters that embeds both − ωµσ
2 from (ii) and
a location weight. Since this approach is an approximation of
the actual phenomenon, we consider the noise contribution
that contains both the actual sensor noise and an unmodelled
non-linearity. Given enough independent measurements, the
optimal values for the weights can be found using a least
square formulation.
Let us now consider a set of m measurements where each
measurement is associated to k local average thicknesses that
are regularly spaced over the length of the tool. This can be
seen as moving the tool within the pipe simultaneously to
gathering pipe thickness information in a sliding window. The
sliding window approximates the geometry as a piecewiseconstant profile as describe in Fig. 1(a). We then formulate
Eq. (3) in a matrix form to combine the m set of measurement
and thickness values together,
y = T w + ,
(4)
The constant term y0 from Eq. (3) is unknown (it depends
on the excitation, the number of turn in the coil, the electromagnetic properties of the air and the distance between the
exciter and the sensor). It is, however, possible to estimate
y0 from the measurements, therefore, we include it into the
vector of the model parameters w which is defined as,
y0
w1
w = ... ,
(5)
wk−1
wk
T is the matrix that contains the local average thickness
information;
1 t11 t12 . . . t1k
1 t21 t22 . . . t2k
T = .
(6)
..
..
.. ,
..
..
.
.
.
.
1 tm1 tm2 . . . tmk
and y the vector with all the sensor measurements.
y1
y2
y = . .
..
(7)
ym
In order to select parameters ŵ that reflects the attenuation
of the magnetic field through its path, there is a need for an
optimisation method that sets the weights of the non-relevant
thicknesses to zero. It can be obtained by learning the model
parameters with LASSO [9]. Using this parameter selection
also allows avoiding over-fitting the irrelevant parameter that
would be performed by a closed form solution. More formally,
LASSO corresponds to the least square formulation with L1 regularisation as
min k T ŵ − y k2 +α k ŵ k1 ,
(8)
with α the regularisation parameter, which is learned with
an iterative process. Finally, the direct problem is solved by
estimating ŷ as
ŷ = |{z}
T ŵ ,
(9)
h(T )
with the proposed model h(T ).
C. Inverse problem
After estimating the parameters of the direct model (3), we
now consider the inverse problem. More formally, we want to
find the inverse function h−1 such that h−1 : y −→ t. Due
to the double-through wall phenomenon, h cannot be simply
inverted as the geometry under the exciter coil and the receiver
are convoluted in the measurements. Instead, having the direct
problem expressed as a linear model, allows formulating the
inverse problem in a closed form solution, which can be
obtained with least squares.
We consider here solving the inverse problem for a long
pipe section as one system (i.e. recovering the thickness of the
full pipe at the same time). To solve the optimisation problem
through least squares, the degree of freedom (which is equal
to the number of equations minus the number of parameters)
of the system has to be positive or null. As a rule of thumb,
to avoid over-fitting, the degree of freedom should be superior
to ten.
Let us consider the inspection of a long pipe section using
a RFEC tool. During the inspection, a set of m discrete
measurements are collected at regular intervals along the pipe.
We approximate the pipeline geometry as a piecewise-constant
profile with n steps of average thickness t. n is chosen to be
ten times smaller than m.
We then re-write Eq. (3) so it can be formulated as a global
optimisation problem, where all the sensor measurements y
are related to all the piecewise thicknesses ti as
y = W t̂ + y0 ,
(10)
with y and y0 defined in Eq. (3) and (7). t̂ is the set of the
all the thickness estimates for each value of the piecewiseconstant pipeline profile, defined as
t1
t2
t̂ = . .
(11)
..
tn
W is an m × n matrix that contains the relationship
between thickness values and sensor measurements and is
defined by the parameters learned from the direct model. In
practice, each line of W contains the weights w for the local
thickness values and is set to 0 for the others thickness values.
Since there are multiple measurements between the ith and
(i + 1)th values, spatial weights ai and bi are used to define
the influence of the piece proximity as:
a1 w1
..
.
aj w1
W =
0
.
..
0
b1 w1 + a¯1 w2
..
.
b1 wk
..
.
0
..
.
0
..
.
bj w1 + aj w2
aj+1 w1
..
.
...
..
.
...
...
..
.
bj wk
bj+1 wk−1 + aj+1 wk−1
..
.
0
bj+1 wk
..
.
0
0
..
.
...
..
.
...
...
..
.
0
...
0
0
0
...
0
..
.
0
,
0
...
am wk
(12)
Fig. 2: 3D representation of the 2D axisymmetric simulation.
(1) the air box, present all around the pipe, (2) the exciter coil
as a rectangular cross multi-turn copper coil, (3) the receiver,
simplified as a point measurement, and (4) the cast-iron pipe.
with ai and bi defined as follow,
bi
ai + bi
ai
b¯i ,
ai + bi
a¯i ,
(13)
(14)
where ai is the distance from the point measurement to the
centre of the ith step, and bi the distance from the point
measurement to the centre of the (i+1)th step. We then obtain
the thickness estimates t by solving the linear least squares in
closed form,
t̂ = (W t W )−1 W t (y − y0 ) .
|
{z
}
which is simplified to a point measurement (e.g. it could
simulate a hall effect sensor), and (4) the cast-iron pipe that has
its geometry defined from pipe segments extracted from the
decommissioned pipeline. A schematic of the global system
is shown in Fig. 2 with the thickness gains corresponding to
Bell and Spigot (B&S) joints that link pipe segments together.
All medium are approximated as homogeneous and
isotropic. The air and copper material properties are defined
using built-in materials from the COMSOL library. To get
a realistic 2D axisymmetric modelisation of the pipe, the
pipe’s magnetic properties are obtained by analysing a pipe
sample with a Superconducting QUantum Interference Device
(SQUID). We then have both the geometry and the material
properties that come from a real pipeline. The material properties used in the model are displayed in Tab. I. The conductivity
of the air is set to a non-zero value to avoid computational
singularities. The stability of the simulation has been validated
for different meshing sizes, air box sizes and other parameters.
The meshing size is defined according to the wavelength λ
of the magnetic field in each material (i.e. at least five times
smaller than the wavelength), with λ defined as:
λ= r
(15)
h−1 (y)
III. R ESULTS
FEA simulations with a 2D-axisymmetric geometry have
been used to validate the proposed methods in a controlled
environment. We look here at the performance of both the
direct and inverse model applied to a long pipe section with a
known geometry. Note that although the validation has been
done for a 2D-axisymmetric scenario, the proposed models
can be adapted for any RFEC axisymmetric tool.
A. FEA environment
This section describes how the data for validation was
obtained. In the context of our particular research project that
motivates this paper, we used data from a cast-iron pipeline,
which has been decommissioned and is currently dedicated for
research purposes. This particular pipeline was laid more than
hundred years ago, and some parts of the pipe are significantly
corroded. Some pipes section have been exhumed, grid-blasted
and analysed. The material properties of the cast-iron have
been measured, and the corrosion’s profile have been captured
with a laser scanner, using the process described in [10]. We
generated a 60 m long 1-D profile based on the geometry
of exhumed pipe segments. Once incorporated into a FEA
simulation environment, this realistic profile has provided
sufficient data for validation.
The FEA used here is done using COMSOL Multiphysics in
a 2D-axisymmetric scenario. The FEA geometry is composed
of four different components: (1) the air box defining the limits
of the FEA scenario, (2) the exciter coil, which is modelled
as a rectangular cross multi-turn copper coil, (3) the receiver
2π
.
ω.µ.σ
2
(16)
Using Eq.(16) with the magnetic properties of each material
we can define the minimum size for the meshing at each part
of the scenario. The minimum size of each element in the
meshing is given in Tab. I.
The pipeline’s inspection has been simulated using a parameter sweep for the position of the RFEC tool within the pipe
for the 60m length. The amplitude and the phase-lag of the
electromagnetic phase have been recorded for each position of
the parameter sweep.
B. Application of the direct model
We now consider the direct problem applied to a dataset
generated from the FEA environment described previously.
The aim here is to learn the parameters defined in Eq. (5).
As shown in Fig. 2, note that to make it more realistic the
simulated thickness profile contains B&S joints. The thickness
of the B&S joints are much larger than the other parts of the
pipe, hence, due to the linear nature of proposed model, these
data that relate to the B&S are expected to perform poorly. We
solve the direct model for three datasets: (a) the first dataset
include the complete set of data, (b) the data with a B&S joint
located near the receiver have been removed in the second
dataset, and (c) the data near both the exciter and receiver
have been removed in the third dataset.
TABLE I: Properties of each material
material
Air
Cast-Iron
Copper (coil)
µr
1
4.96
1
r
1
1
1
ρ [S/m]
10
1.12e7
5.99e7
l [m]
0.1
0.012
0.0115
(a)
(b)
(c)
Fig. 3: Due to the presence of the B&S joints inducing a sort of non-linearity in the data, the model described in Eq. (8) is
not longer valid. Therefore, we remove data where the B&S joint has an impact on the exciter coil (a), and where it has an
impact on the receivers (b). The model learned from the filtered data is shown in (c).
We compared the estimated ŷ and the actual sensor measurement y in Fig. 3 with each sub-figure dedicated to each
dataset. In Fig. 3(a), we set the colour information to reflect
the impact of the B&S joints located near the receiver (the
yellow points are more influenced by the B&S). In Fig. 3(a),
the blue points represent the estimation with the B&S located
on top of the receiver. The third dataset shown in Fig. 3(c)
shows a better regression.
Since the simulations are done in a controlled environment,
the locations of the B&S joints are known, thus removing these
particular data is a trivial task. In the case of an unknown
environment, one could classify the construction features of
the pipeline, which can be done using a Support Vector
Machine (SVM) classifier such as in [11]. An alternative
would consist of automating the data selection with methods
such as Peirce’s [12] or Chauvenet’s [13] criterion.
The parameter α from Eq. (8) is chosen using ten-fold crossvalidation. The estimated parameters w and measurements
Fig. 4: Evolution of the Mean Square Error (MSE) versus
the value of the α parameter (using cross-validation). The α
indicated in blue corresponds to the sparsest solution within
one standard error of the MSE (it is the chosen one).
of the goodness of fitting (Mean Square Error MSE and the
coefficient of determination R2 ) are available in Tab. II. As
expected, the constant is a positive term, and the attenuation
coefficients are negative terms. Moreover, we can see that the
geometry near to the receiver and near to the exciter coil have
a more important role which is reflected by higher weights.
C. Application of the inverse problem
After solving the direct problem, all the parameters required
for the inverse problem are known. We consider recovering the
60 metres of pipe thickness as a global problem and the full
geometry is recovered from the set of all the measurement
using the formulation established in Eq. (15).
The inverse problem relies on the parameters learnt for
the direct problem. In the case where all the parameters
from the RFEC tool and the magnetic properties of the
pipe specimen are known, it is possible to obtain direct and
inverse models through FEA simulation. Otherwise, multiple
thickness measurements have to be collected from the studied
pipe. These thickness measurements at specific locations are
needed to learn the parameters. In practice, collecting such
measurements is a feasible task considering the few parameters
that are present in the proposed model.
A zoom-in of the pipe profile reconstructed is shown in
Fig. 5. The estimation is shown in blue, and the ground
truth is shown in orange. The spikes in the ground-truth
correspond to the B&S joints. As predicted, the proposed
inverse model cannot recover these thicknesses due to the nonlinear behaviour of the magnetic field in these regions.
The estimation error for the thickness of the 60 metre-long
pipe is of 0.1mm for the MSE and 10.3mm for RMSE (with
the average thickness of the pipe being around 30mm). If
we remove the areas with the B&S joints, the RMSE falls
to 2.6mm.
TABLE II: Output of the Least Absolute Shrinkage and Selection Operation. The localised increase of thickness (e.g. B&S joints)
lead to the spread the weights. This is visible by comparing the lines of the table.
coef
dataset (a)
dataset (b)
dataset (c)
cst
y0
w1
w2
Exciter
w3
w4
-17.6
-26.0
-20.9
-5.7
-16.7
-22.4
-19.9
-3.3
-16.0
-48.3
-103.0
0
w19
w20
w21
w22
-5.7
-3.2
-0.2
-2.0
-2.7
-7.8
-10.3
-5.5
-1.9
-1.1
-1.6
0
-6.2
-41.5
0
0
0
0
0
-2.9
-11.2
IV. D ISCUSSION
In this paper, we tackle the direct and inverse problems for
a 2D-axisymmetric RFEC tool composed of a single exciter
coil and a single receiver. We have shown, using FEA, that
both direct and inverse model are accurate for recovering pipe
sections with organic geometry (which is often the case for
corroded cast-iron pipes). The FEA model used to generate the
dataset is based on a realistic geometry and material properties
obtained from old cast-iron pipes.
The proposed direct model is solved using LASSO. The
L1 -regularisation allows selecting automatically the important
thickness areas for the model while reducing the number of
parameters. This result into a simplistic model, with most
important thicknesses located next to the exciter coil and the
receiver. The inverse problem relies on the parameters from
the direct problem and is solved using least squares. For
training the proposed inverse model, thickness measurements
have to be collected from the pipe. In practice, collecting such
measurements is a feasible task considering the few parameters
of the proposed model.
The main limitation of the proposed method lies in the form
of the proposed model. The linear model allows solving the
inverse problem in a closed-form. The model gives accurate
results apart for the B&S joints. For these extremely thick
thicknesses, the magnetic field would flow through the path of
least resistance which cannot be captured by a linear model.
Furthermore, to the outstanding thicknesses, the magnetic
properties are considered constant for the full pipeline. In
practice, pipes can have a variation of magnetic properties;
this case is not studied here.
In future work, we are planning to apply this method for
a tool with a sensor array (3D case). It can be shown that
Thickness (mm)
50
estimation
ground-truth
40
30
20
10
0
36
38
40
42
44
Receiver
w23
w24
w5
46
48
50
52
distance (m)
Fig. 5: Thickness estimated in closed-form. The estimation is
shown in blue, and the ground-truth in orange.
w25
w26
w27
-19.6
-19.1
-8.6
-9.1
-24.5
-63.1
-2.5
0
-46.3
-36.9
-29.0
-3.9
Goodness of Fit
MSE
R2
0.296
0.5754
0.174
0.6391
0.016
0.8779
the attenuation from the exciter behaves as a circumferential
offset [14]. Therefore, it is possible to deconvolute the signal
in a similar fashion.
ACKNOWLEDGMENT
This publication is an outcome from the Critical Pipes
Project funded by Sydney Water Corporation, Water Research
Foundation of the USA, Melbourne Water, Water Corporation
(WA), UK Water Industry Research Ltd, South Australia Water
Corporation, South East Water, Hunter Water Corporation,
City West Water, Monash University, University of Technology
Sydney and University of Newcastle. The research partners are
Monash University (lead), University of Technology Sydney
and University of Newcastle.
R EFERENCES
[1] D. L. Atherton, “Remote field eddy current inspection,” IEEE Transactions on Magnetics, vol. 31, no. 6, pp. 4142–4147, 1995.
[2] M.-È. Davoust, L. Brusquet, and G. Fleury, “Robust estimation of flaw
dimensions using remote field eddy current inspection,” Measurement
Science and Technology, vol. 17, no. 11, pp. 3006–3014, nov 2006.
[3] M.-È. Davoust, L. Brusquet, and G. Fleury, “Robust Estimation of Hidden Corrosion Parameters Using an Eddy Current Technique,” Journal
of Nondestructive Evaluation, vol. 29, no. 3, pp. 155–167, jun 2010.
[4] A. Tao, W. Zhang, Z. Wang, and Q. Luo, “Design on forward modeling
of RFEC inspection for cracks,” Proceedings - 2014 International Conference on Information Science, Electronics and Electrical Engineering,
ISEEE 2014, vol. 1, pp. 579–584, 2014.
[5] E. Cardelli, N. Esposito, and M. Raugi, “Electromagnetic Analysis of
RFEC Differential Probes,” IEEE Transactions on Magnetics, vol. 29,
no. 2, pp. 1849–1852, 1993.
[6] A. Skarlatos, G. Pichenot, D. Lesselier, M. Lambert, and B. Duchêne,
“Electromagnetic Modeling of a Damaged Ferromagnetic Metal Tube
by a Volume Integral Equation Formulation,” IEEE Transactions on
Magnetics, vol. 44, no. 5, pp. 623–632, 2008.
[7] W. Lord, Y. S. Sun, S. Udpa, and S. Nath, “A Finite Element Study
of the Remote Field Eddy Current Phenomen,” IEEE Transactions on
Magnetics, vol. 24, no. 1, pp. 435–438, 1988.
[8] Y. S. Sun, D. Cooley, H. Han, S. Udpa, and W. Lord, “Efforts
towards gaining a better understanding of the remote field eddy current
phenomenon and expanding its applications,” IEEE Transactions on
Magnetics, vol. 32, no. 3, pp. 1589–1592, may 1996.
[9] R. Tibshirani, “Regression Selection and Shrinkage via the Lasso,”
Journal of the Royal Statistical Society, vol. 58, no. 1, pp. 267–288,
1996.
[10] B. Skinner, T. Vidal-calleja, J. Valls Miro, F. D. Bruijn, and R. Falque,
“3D Point Cloud Upsampling for Accurate Reconstruction of Dense
2.5D Thickness Maps 3D Point Cloud Acquisition,” in Australasian
Conference on Robotics and Automation (ACRA), 2014.
[11] T. Vidal-calleja, J. V. Miro, and F. Mart, “Automatic Detection and
Verification of Pipeline Construction Elements with Multi-modal data,”
in IROS, 2014.
[12] B. Peirce, “Criterion for the rejection of doubtful observations,” The
Astronomical Journal, vol. 2, no. 45, pp. 161–163, 1852.
[13] C. William, A Manual of Spherical and Practical Astronomy. Philadelphia, J. B. Lippincott {&} co.; London, Trubner {&} co., 1863.
[14] R. Falque, T. Vidal-calleja, J. Valls Miro, D. C. Lingnau, and D. E.
Russell, “Background Segmentation to Enhance Remote Field Eddy Current Signals,” in Australasian Conference on Robotics and Automation
(ACRA), 2014, pp. 1–9.
| 3cs.SY
|
arXiv:1509.03122v1 [math.AC] 10 Sep 2015
REDUCTION OF LOCAL UNIFORMIZATION TO THE CASE OF
RANK ONE VALUATIONS FOR RINGS WITH ZERO DIVISORS
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
Abstract. This is a continuation of a previous paper by the same authors.
In the former paper, it was proved that in order to obtain local uniformization
for valuations centered on local domains, it is enough to prove it for rank
one valuations. In this paper, we extend this result to the case of valuations
centered on rings which are not necessarily integral domains and may even
contain nilpotents.
1. Introduction
For an algebraic variety X over a field k, the problem of resolution of singularities is whether there exists a proper birational morphism X ′ −→ X such that X ′
is regular. The problem of local uniformization can be seen as the local version
of resolution of singularities for an algebraic variety. For a valuation ν of k(X)
having a center on X, the local uniformization problem asks whether there exists
a proper birational morphism X ′ −→ X such that the center of ν on X ′ is regular.
This problem was introduced by Zariski in the 1940’s as an important step to prove
resolution of singularities. Zariski’s approach consists in proving first that every
valuation having a center on the given algebraic variety admits local uniformization. Then one has to glue these local solutions to obtain a global resolution of all
singularities.
Zariski succeeded in proving local uniformization for valuations centered on algebraic varieties over a field of characteristic zero (see [11]). He used this to prove
resolution of singularities for algebraic surfaces and threefolds over a field of characteristic zero (see [12]). Abhyankar proved (see [1]) that local uniformization can
be obtained for valuations centered on algebraic surfaces in any characteristic and
used this fact to prove resolution of singularities for surfaces (see [2] and [4]). He
also proved local uniformization and resolution of singularities for threefolds over
fields of characteristic other than 2, 3 and 5 (see [3]). Very recently, Cossart and
Piltant proved resolution of singularities (and, in particular, local uniformization)
for threefolds over any field of positive characteristic, as well as in the arithmetic
case (see [5] and [6]). They proved it using the approach of Zariski. However, the
2010 Mathematics Subject Classification. Primary 14B05; Secondary 14E15, 13H05.
Key words and phrases. Local uniformization, resolution of singularities, reduced varieties.
During the realization of this project the first author was supported by a grant from the
program “Ciência sem Fronteiras” from the Brazilian government.
1
2
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
problem of local uniformization remains open for valuations centered on algebraic
varieties of dimension greater than three over fields of positive characteristic.
Since local uniformization is a local problem, we can work with local rings instead
of algebraic varieties. A valuation ν centered on a local integral domain R is said
to admit local unifomization if there exists a local local ring R(1) dominated by Oν
and dominating R such that R(1) is regular. Let N be the category of all noetherian
local domains and M ⊆ N be a subcategory of N which is closed under taking
homomorphic images and localizing any finitely generated birational extension at
a prime ideal. We want to know for which subcategories M with these properties,
all valuations centered on objects of M admit local uniformization. In Section
7.8 of [7], Grothendieck proved that any category of schemes, closed under passing
to subschemes and finite radical extensions, in which resolution of singularities
holds, is a subcategory of quasi-excellent schemes (it is known that the category
of quasi-excellent schemes is closed under all the operations mentioned above). He
conjectured (see Remark 7.9.6 of [7]) that resolution of singularities holds in this
most general possible context: that of quasi-excellent schemes. Translated into our
local situation, this conjecture says that the subcategory of N which optimizes local
uniformization is the category of all quasi-excellent local rings. This subcategory
has the properties above. For a discussion on quasi-excellent and excellent local
rings see Section 7.8 of [7]. However, this conjecture is widely open.
In most of the successful cases, including those mentioned above, local uniformization was first proved for rank one valuations. Then the general case was
reduce to this a priori weaker one. In [9], we prove that this reduction works under
very general assumptions. Namely, we consider a subcategory M of the category
of all noetherian local integral domains, closed under taking homomorphic images
and localizing any finitely generated birational extension at a prime ideal. The
main result of [9] is that if every rank one valuation centered on an object of M
admits local uniformization, then all the valuations centered on objects of M admit local uniformization. The main goal of this paper is to extend this result to
rings which are not necessarily integral domains and, in particular, may contain
nilpotent elements. The importance of non-integral and non-reduced schemes in
modern algebraic geometry is well known. Even if one were only interested in reduced schemes to start with, one is led to consider non-reduced ones as they are
produced by natural constructions, for example, in deformation theory. Therefore,
it appears desirable to study the problem of local uniformization for such schemes
and, in particular, to extend our earlier results on reducing the problem to the rank
one case to this more general context.
If R is not reduced we cannot expect, in general, to make R(1) be regular by
blowings up. The natural extension to this case is to require R(1) red to be regular
n+1
n
and I(1)
/I(1)
to be an R(1) red -free module for every n ∈ N (here I(1) denotes
the nilradical of R(1) ). For more precise definitions see Section 2. Let N be the
category of all noetherian local rings and M ⊆ N be a subcategory of N which
LOCAL UNIFORMIZATION FOR NON-DOMAINS
3
is closed under taking homomorphic images and localizing any finitely generated
birational extension at a prime ideal. Our main result is the following:
Theorem 1.1. Assume that for every noetherian local ring R in Ob(M), every rank
one valuation centered on R admits local uniformization. Then all the valuations
centered on objects of M admit local uniformization.
The proof of Theorem 1.1 consists of three main steps. The first step is to
prove that for every local ring R and every valuation ν centered on R, there exists
a local blowing up (see Definition 2.2) R −→ R(1) such that R(1) has only one
associated prime ideal. Then we consider a decomposition ν = ν1 ◦ ν2 of ν such
that rk(ν1 ) < rk(ν) and rk(ν2 ) < rk(ν). Using induction, we can assume that both
ν1 and ν2 admit local uniformization. The second main step consists in using this
to prove that there exists a local blowing up R(1) −→ R(2) such that R(2) red is
regular. The third and final step is to prove that there exists a further local blowing
n+1
n
up R(2) −→ R(3) such that R(3) red is regular and I(3)
/I(3)
is an R(3) red -free
module for every n ∈ N (here I(3) denotes the nilradical of R(3) ).
This paper is divided as follows. In Section 2 we present the basic definitions
and results that will be used in the sequel. Sections 3, 4 and 5 are dedicated to
prove the results related to the first, second and third steps, respectively. In the
last section we present a proof of our main theorem.
2. Preliminaries
Let R be a noetherian commutative ring with unity and Γ an ordered abelian
group. Set Γ∞ := Γ ∪ {∞} and extend the addition and order from Γ to Γ∞ as
usual.
Definition 2.1. A valuation ν on R is a mapping ν : R −→ Γ∞ with the following
properties:
(V1)
(V2)
(V3)
(V4)
ν(ab) = ν(a) + ν(b) for every a, b ∈ R;
ν(a + b) ≥ min{ν(a), ν(b)} for every a, b ∈ R;
ν(1) = 0 and ν(0) = ∞;
The support of ν, which is defined by supp(ν) := {a ∈ R | ν(a) = ∞}, is
a minimal prime ideal of R.
Take a multiplicative system S of R such that supp(ν) ⊆ R \ S. Then the
extension (which we call again ν) of ν to RS given by ν(a/s) := ν(a) − ν(s) is again
a valuation. Indeed, the three first axioms are easily checked. The minimality of
supp(ν) as a prime ideal of RS follows from the fact that the prime ideals of RS are
in a bijective correspondence to the prime ideals of R contained in R \ S. From now
on, we will freely make such extensions of ν to RS without mentioning it explicitly.
A valuation ν on R is said to have a center if ν(a) ≥ 0 for every a ∈ R. In this
case, the center of ν on R is defined by Cν (R) := {a ∈ R | ν(a) > 0}. Moreover, if
R is a local ring with unique maximal ideal m (in which case we say “the local ring
4
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
(R, m)”), then a valuation ν on R is said to be centered at R if ν(a) ≥ 0 for every
a ∈ R and ν(a) > 0 for every a ∈ m. We observe that if ν is a valuation having a
center on R, then ν is centered on RCν (R) . The value group of ν, denoted by νR,
is defined as the subgroup of Γ generated by {ν(a) | a ∈ R}. The rank of ν is the
number of proper convex subgroups of νR.
For an element b ∈ R \ supp(ν) we consider the canonical map Φ : R −→ Rb
given by Φ(a) = a/1. Let
J(b) := ker Φ =
∞
[
annR (bi ).
i=1
We have a natural embedding R/J(b) ⊆ Rb . Take a1 , . . . , ar ∈ R such that
ν(ai ) ≥ ν(b)
for each i, 1 ≤ i ≤ r.
Consider the subring R′ := R/J(b)[a1 /b, . . . , ar /b] of Rb . Then the restriction of ν
to R′ has a center Cν (R′ ) in R′ . We set R(1) := RC′ ν (R′ ) .
Definition 2.2. The canonical map R −→ R(1) will be called the local blowing
up of R with respect to ν along the ideal (b, a1 , · · · , ar ). For a valuation µ having
a center on R we will say that R −→ R(1) is µ-compatible if b ∈
/ Cµ (R) and
ai ∈ Cµ (R) for every i, 1 ≤ i ≤ r.
Lemma 2.3. The composition of finitely many local blowings up is again a local
blowing up. Moreover, if each of these local blowings up is µ-compatible, then their
composition is again µ-compatible.
Proof. It is enough to prove that for two local blowings up π : R −→ R(1) and
π ′ : R(1) −→ R(2) with respect to ν, there exists a local blowing up R −→ R(3)
with respect to ν such that R(3) ≃ R(2) . We write
R(1) = RC′ ν (R′ ) for R′ = R/J(b)[a1 /b, . . . , ar /b]
for some a1 , . . . , ar , b ∈ R and
R(2) = RCν (R′(1) ) for R′(1) = R(1) /J(β)[α1 /β, . . . , αs /β]
′(1)
for some α1 , . . . , αs , β ∈ R(1) . Then there exist ar+1 , . . . , ar+s , b′ ∈ R such that
αi /β = π(ar+i )/π(b′ ) for each i, 1 ≤ i ≤ s. Consider the local blowing up
R −→ R(3)
given by
R(3) = RC′′ ν (R′′ ) for R′′ = R/J(bb′ )[a1 b′ /bb′ , . . . , ar b′ /bb′ , ar+1 b/bb′ . . . , ar+s b/bb′].
It is straightforward to prove that R(2) ≃ R(3) .
In view of Lemma 2.3, we will freely use the fact that the composition of finitely
many local blowings up is itself a local blowing up without mentioning it explicitly.
For simplicity of notation, we denote the nilradical of R by I, i.e.,
I = Nil(R) := {a ∈ R | al = 0 for some l ∈ N}.
LOCAL UNIFORMIZATION FOR NON-DOMAINS
5
Definition 2.4. We say that Spec(R) is normally flat along Spec(Rred ) if
I n /I n+1 is an Rred -free module for every n ∈ N.
Since R is noetherian, there exists N ∈ N such that I n = (0) for every n > N .
Hence, the condition in Definition 2.4 is equivalent to the freeness of the finitely
many modules I/I 2 , . . . , I N /I N +1 = I N .
Definition 2.5. For a local ring R, a valuation ν centered on R is said to admit local uniformization if there exists a local blowing up R −→ R(1) with
respect to ν such that R(1) red is regular and Spec R(1) is normally flat along
Spec R(1) red .
Let ν = ν1 ◦ ν2 be a fixed decomposition of ν. For simplicity of notation, we
set p := Cν1 (R) and for a local blowing up R −→ R(1) we set p(1) := Cν1 R(1) .
We need to guarantee that the main structure of Rp and R/p are preserved under
ν1 -compatible local blowings up. More precisely, we have to prove the following:
Proposition 2.6. Let π : R −→ R(1) be a ν1 -compatible local blowing up. Then the
(1)
canonical maps Rp −→ Rp(1) and R/p −→ R(1) /p(1) induced by π are isomorphisms.
In order to prove Proposition 2.6 we need the following basic lemma.
Lemma 2.7. Let S be a multiplicative system of R contained in R \ Cν (R). Then
the canonical map Φ : RCν (R) −→ (RS )Cν (RS ) given by Φ(a/b) = (a/1)/(b/1) is an
isomorphism.
Proof. For an element (a/b)/(c/d) ∈ (RS )Cν (RS ) we have
ν(b) = ν(c) = ν(d) = 0.
Consequently, ν(bc) = 0 and ad/bc ∈ RCν (R) . Then
(a/b)/(c/d) = (ad/1)/(bc/1) = Φ(ad/bc).
Suppose that Φ(a/b) = 0. This means that there exists c/d ∈ RS \ Cν (RS ) such
that ac/d = 0 in RS . Thus, there exists s ∈ S such that sac = 0. Moreover, since
c/d ∈
/ Cν (RS ) we also have that c ∈
/ Cν (R). This and the fact that s ∈ S ⊆ R\Cν (R)
imply that sc ∈
/ Cν (R). Hence, a/b = 0 in RCν (R) which is what we wanted to
prove.
Proof of Proposition 2.6. Applying Lemma 2.7 to R (with S = {1, b, b2, . . .}) and
R′ (with S ′ = R′ \ Cν (R′ )) and the valuation ν1 , we obtain that the canonical maps
(1)
Rp −→ (Rb )Cν (Rb ) and RC′ ν (R′ ) −→ Rp(1) , respectively, are isomorphisms. Hence,
1
1
in order to prove the first assertion, it is enough to show that the canonical map
(Rb )Cν (Rb ) ←− RC′ ν (R′ ) is an isomorphism.
1
1
Since R′ ⊆ Rb and Cν1 (R′ ) = R′ ∩Cν1 (Rb ) we have that RC′ ν (R′ ) −→ (Rb )Cν (Rb )
1
1
is injective. On the other hand, any element (a/bn )/(c/bm ) in (Rb )Cν (Rb ) can be
1
written as (abm /1)/(cbn /1) which is the image of abm /cbn . Hence the map
RC′ ν
1 (R
′)
−→ (Rb )Cν
1 (Rb )
6
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
is surjective and consequently it is an isomorphism.
Set R0 = R/J(b) and consider the induced map R0 −→ R(1) . Since the canonical
map R −→ R0 is surjective, in order to prove the surjectivity of R −→ R(1) /p(1) ,
it is enough to show that R0 −→ R(1) /p(1) is surjective. For an element α ∈ R(1)
we write α = p/q where p = P (a1 /b, . . . , ar /b) and q = Q(a1 /b, . . . , ar /b) for some
P (X1 , . . . , Xr ), Q(X1 , . . . , Xr ) ∈ R0 [X1 , . . . , Xr ].
Set p0 = P (0, . . . , 0) and q0 = Q(0, . . . , 0). Then
p1 := p − p0 =
r
X
ai /b · Pi (a1 /b, · · · , ar /b)
r
X
ai /b · Qi (a1 /b, · · · , ar /b)
i=1
and
q1 := q − q0 =
i=1
for some Pi , Qi ∈ R0 [X1 , . . . , Xr ], 1 ≤ i ≤ r. Since ν1 (ai /b) > 0 we obtain that
ν1 (p1 ) > 0 and ν1 (q1 ) > 0. This implies that
(1)
ν1 (q0 ) =
0,
(2)
ν1 (q0 q) =
0
and
ν1 (q0 p1 − p0 q1 ) > 0.
(3)
Therefore,
p/q − p0 /q0 = (q0 p1 − p0 q1 )/q0 q ∈ p(1) .
It remains to prove that p0 /q0 ∈ R0 . Since ν1 (q1 ) > 0, also ν(q1 ) > 0. Hence,
ν(q0 ) = ν(q − q1 ) = 0 and consequently q0 is a unit in R0 . Therefore, p0 /q0 ∈ R0 .
To finish our proof it is enough to show that the kernel of R −→ R(1) /p(1) is p.
This follows immediately from the definition of p and p(1) as the centers of ν1 on
R and R(1) , respectively.
Lemmas 2.8 and 2.9 below are generalizations of Lemma 2.18 and Corollary 2.20
of [9], respectively. The proofs presented there can be adapted to our more general
case. We present sketches of the proofs for the convenience of the reader.
e(1) with respect to ν1 , there exists
Lemma 2.8. For each local blowing up Rp −→ R
e(1) ≃ R(1)
.
a local blowing up R −→ R(1) with respect to ν such that R
p(1)
e(1) given by
Proof. We consider the local blowing up Rp −→ R
e (1) = R
e′
R
C
e′ )
ν 1 (R
e ′ = Rp /J(β)[α1 /β, . . . , αr /β].
for R
Choose a1 , . . . , ar , b ∈ R such that for each i, 1 ≤ i ≤ r we have Φ(ai )/Φ(b) = αi /β
where Φ : R −→ Rp is the canonical map. If ν(ai ) < ν(b) for some i, 1 ≤ i ≤ r,
LOCAL UNIFORMIZATION FOR NON-DOMAINS
7
then we have ν1 (αi ) = ν1 (β). Choose i so as to minimize the value ν(ai ), in other
words, so that ν(ai ) ≤ ν(aj ) for all j ∈ {1, . . . , r}. Set
e′′ := Rp /J(αi ) α1 , . . . , αi−1 , β , αi+1 , . . . , αr .
R
αi
αi αi αi
αi
e′′
Then R(1) ≃ R
C
e′′ ) .
ν 1 (R
Hence, after a suitable permutation of the set {a1 , . . . , ar , b},
we may assume that ν(ai ) ≥ ν(b) for every i, 1 ≤ i ≤ r. Consider the local blowing
up
R(1) = RC′ ν (R′ ) for R′ = R/J(b)[a1 /b, . . . , ar /b]
(1)
e(1) .
with respect to ν. It is straightforward to prove that Rp(1) ≃ R
(1)
Lemma 2.9. For each local blowing up R/p −→ R
exists a local blowing up R −→ R
(1)
and Rp ≃ Rp(1) .
(1)
with respect to ν2 , there
(1)
with respect to ν such that R(1) /p(1) ≃ R
Proof. For an element a ∈ R we denote its image under the canonical map
R −→ R/p
by a. Then
R
(1)
′
′
= RCν2 (R′ ) with R = (R/p) /J(b)[a1 /b, . . . , ar /b]
for some a1 , . . . , ar , b ∈ R \ p. Since ν2 (ai ) ≥ ν2 (b) we have ν(ai ) ≥ ν(b) for every
i, 1 ≤ i ≤ r. Then we can consider the local blowing up
R(1) = RC′ ν (R′ ) with R′ = R/J(b)[a1 /b, . . . , ar /b]
with respect to ν. It is again straightforward to prove that R(1) /p(1) ≃ R
(1)
Rp ≃ Rp(1) .
(1)
and
3. Associated prime ideals of R
Let R be a local ring and ν a valuation centered on R. The main result of this
section is the following:
Proposition 3.1. There exists a local blowing up R −→ R(1) with respect to ν such
that Nil R(1) is the only associated prime of R(1) .
In order to prove Proposition 3.1, we need the following result.
Lemma 3.2. Let R′ = R/J(b)[a1 /b, . . . , ar /b] ⊆ Rb for some b, a1 , . . . , ar ∈ R with
ν(b) ≤ ν(ai ) for every i, 1 ≤ i ≤ r. Then for every c′ ∈ R′ the ideal annR′ (c′ ) can
be written as annR′ (c/1) for some c ∈ R. Moreover, if annR′ (c′ ) is prime, then
annR bN c is a prime ideal of R for some N ∈ N.
8
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
Proof. Choose c ∈ R such that c′ = c/bl for some l ∈ N. Fix a′ ∈ R′ and write
a′ = a/bm for some m ∈ N and a ∈ R. Then we have
a′ ∈ annR′ (c′ ) ⇐⇒ acbn = 0 for some n ∈ N ⇐⇒ a′ ∈ annR′ (c/1).
Now assume that annR′ (c′ ) is prime and set R0 := R/J(b). Then
annR0 (c/1) = annR′ (c′ ) ∩ R0
is also prime. Moreover,
π −1 (annR0 (c/1)) =
(4)
∞
[
annR (bn c),
n=1
where π : R −→ R/J(b) is the canonical epimorphism. Indeed,
a ∈ π −1 (annR0 (c/1)) ⇐⇒ ac/1 = 0 in Rb
⇐⇒ bn ac = 0 in R for some n ∈ N
∞
[
⇐⇒ a ∈
annR (bn c).
n=1
Since R is noetherian and
we have that
(5)
annR (bc) ⊆ annR b2 c ⊆ · · · ⊆ annR (bn c) ⊆ · · ·
∞
[
annR (bn c) for some N ∈ N.
annR bN c =
n=1
By (4) and (5) we conclude that annR bN c is a prime ideal of R.
Corollary 3.3. For a local blowing up R −→ R(1) , if Nil(R) is the only associated
prime ideal of R, then Nil R(1) is the only associated prime ideal of R(1) .
Proof. Let R(1) = RC′ ν (R′ ) for some R′ as in Lemma 3.2. Theorem 6.2 of [8] gives
us that Ass R(1) = Ass(R′ ) ∩ Spec R(1) . This and Lemma 3.2 guarantee that
|Ass R(1) | ≤ |Ass (R) | = 1. Consequently, R(1) has only one associated prime
ideal, say q. The primary decomposition theorem now gives us that q = Nil R(1) ,
which is what we wanted to prove.
We will use Corollary 3.3 throughout this paper without always mentioning it
explicitly.
Proof of Proposition 3.1. Since supp(ν) is a minimal prime ideal, there exists at
most one associated prime ideal of R contained in (hence equal to) supp(ν). We
will prove that if |Ass(R)| > 1, then there exists a local blowing up R −→ R(1)
such that |Ass R(1) | < |Ass(R)|.
Take an associated prime ideal q of R such that q 6⊆ supp(ν). Write
q = (b, a1 , . . . , ar )
LOCAL UNIFORMIZATION FOR NON-DOMAINS
9
with ν(b) ≤ ν(ai ) for every i, 1 ≤ i ≤ r. Blowing up R with respect to ν along q
gives us a local ring
R(1) = RC′ ν (R′ ) where R′ = R/J(b)[a1 /b, . . . , ar /b].
Observe that this is indeed a local blowing up because ν(b) ≤ ν(ai ) for every i and
q 6⊆ supp(ν) implies that b ∈
/ supp(ν). Since
Ass R(1) = Ass(R′ ) ∩ Spec R(1) (see Theorem 6.2 of [8]),
it remains to show that |Ass(R′ )| < |Ass(R)|.
By Lemma 3.2, we obtain that R′ has at most |Ass(R)| many associated prime
ideals. Moreover, for the chosen associated prime ideal q = annR (c) of R and
for every r ∈ N, the ideal annR′ (c/br ) is not prime in R′ . Indeed, since q =
(b, a1 , . . . , ar ) = annR (c) we have bc = 0 in R. This means that
c/1 = 0/1
in R and consequently annR′ (c/1) = R′ (which is not prime). Therefore
′
|Ass(R′ )| < |Ass(R)|.
Remark 3.4. If I is the only associated prime ideal of R, then for every b ∈
/ I
we have J(b) = (0). In this particular case, we can eliminate the ideal J(b) in the
definition of a local blowing up. We will use this throughout this paper without
mentioning it explicitly.
4. Making Rred regular
Let R be a local ring and ν a valuation centered on R. Assume that ν = ν1 ◦ ν2
and denote by p the center of ν1 on R. As usual, we denote by I the nilradical
of R and for a local blowing up R −→ R(1) we denote the nilradical of R(1) by
I(1) . Assume that I is the only associated prime ideal of R. The main goal of this
section is to prove the following proposition.
Proposition 4.1. Assume that (Rp )red and R/p are regular. Then there exists a
ν1 -compatible local blowing up R −→ R(1) such that R(1) red is regular. Moreover,
for every local blowing up R(1) −→ R(2) along an ideal (b, a1 , . . . , ar ) with b ∈
/ p(1)
(2)
and a1 , . . . , ar ∈ I(1) we have that R
is regular.
red
In order to prove Proposition 4.1 we will need a few lemmas.
Lemma 4.2. Assume that (Rp )red is regular. Then there exists
a ν1-compatible
2
(1)
(1)
(1)
(1)
local blowing up R −→ R such that the R /p -module p / p(1) + I(1) is
(1)
(1)
are elements
of p(1) whose images in
free. Moreover, if y1 , · · · , yr
2
2
p(1) / p(1) + I(1) form a basis of p(1) / p(1) + I(1) then their images in
(1)
(1)
.
form a regular system of parameters of Rp(1)
Rp(1)
red
red
10
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
Lemma 4.3. Let π : R −→ R(1) be a local blowing up along an ideal (b, a1 , . . . , ar )
with b ∈
/ p and a1 , . . . , ar ∈ I. If p/ p2 + I is a free R/p-module, then
2
(1)
(1)
+ I(1)
p /
p
is a free R(1) /p(1) -module.
Lemma 4.4. Take y1 , . . . , yr ∈ p and x1 , . . . , xt ∈ m \ p whose images form a
regular system of parameters of (Rp )red and R/p, respectively. If p/(p2 + I) is an
R/p-free module with basis y1 + (p2 + I), . . . , yr + (p2 + I), then Rred is regular.
Proof of Proposition 4.1, assuming Lemmas 4.2, 4.3 and 4.4. We apply Lemma 4.2
(1)
(1)
(1)
to obtain a ν1 -compatible local blowing up
y1 , . . . , yr ∈ R(1) such
R −→ R and
2
that their images in p(1) / p(1) + I(1) form an R(1) red - basis and their images
(1)
form a regular system of parameters. Moreover, by Proposition 2.6,
in Rp(1)
red
R(1) /p(1) is regular. Also, by Lemma 4.3 and Proposition 2.6, for every local blowing up R(1) −→ R(2) along an ideal (b, a1 , . . . , ar ) with b ∈
/ p(1) and a1 , . . . , ar ∈ I(1)
the hypotheses of Lemma 4.4 are satisfied for R(2) . Hence, we obtain that R(1) red
and R(2) red are regular.
We now proceed with the proofs of Lemmas 4.2, 4.3 and 4.4.
Lemma 4.5. Take generators y1 , . . . , yr , yr+1 , . . . , yr+s of p and b ∈
/ p. Let
π : R −→ R(1)
be the local blowing up along the ideal (b, y1 , . . . , yr ). Set
(1)
yi
(1)
= π(yi )/b for 1 ≤ i ≤ r and yr+k = π(yr+k ) for 1 ≤ k ≤ s.
(1)
(1)
Then p(1) is generated by y1 , . . . , yr+s .
(1)
Proof. Obviously yi ∈ p(1) for every i, 1 ≤ i ≤ r + s. Take an element p/q ∈ p(1) .
This implies that p = p(y1 /b, . . . , yr /b) for some p(X1 , . . . , Xr ) ∈ R[X1 , . . . , Xr ]
(see Remark 3.4). If we set p0 = p(0, . . . , 0), then
yr
y1
p = p0 + p1 + . . . + pr , for some p1 , . . . , pr ∈ R′ .
b
b
This implies that p0 ∈ p. Hence, there exist a1 , . . . , ar+s ∈ R such that p0 =
a1 y1 + . . . + ar+s yr+s . Thus
s
r
p X π(bai ) + pi (1) X π(ar+k ) (1)
(1)
(1)
=
yi +
yr+k ∈ y1 , . . . , yr+s R(1) .
q
q
q
i=1
k=1
This concludes our proof.
Proof of Lemma 4.2. Since (Rp )red is regular there are elements y1 , . . . , yr ∈ p such
that their images in (Rp )red form a regular system of parameters. The first step is
to reduce to the case when y1 , . . . , yr generate p.
LOCAL UNIFORMIZATION FOR NON-DOMAINS
11
Assume that y1 , . . . , yr do not generate p. Choose yr+1 , . . . , yr+s ∈ p such that
y1 , . . . , yr , yr+1 , . . . , yr+s generate p. For each k, 1 ≤ k ≤ s, we can find bk ∈ R \ p,
b1k , . . . , brk ∈ R and hk ∈ (y1 , . . . , yr )2 such that
bk yr+k + b1k y1 + . . . + brk yr + hk ∈ I.
Consider the local blowing up π : R −→ R(1) along (b1 , y1 , . . . , yr ). It follows that
(1)
(1)
(1)
π(b1 ) yr+1 + π(b11 )y1 + . . . + π(br1 )yr(1) + h1
∈ I(1)
and
(1)
(1)
(1)
π(bk )yr+k + π(b1 b1k )y1 + . . . + π(b1 brk )yr(1) + hk ∈ I(1) for 2 ≤ k ≤ s,
(1)
(1)
(1)
where yi = π(yi )/b1 for 1 ≤ i ≤ r and yr+k = π(yr+k ) and some hk ∈
2
(1)
(1)
y1 , . . . , yr
for 1 ≤ i ≤ s. Since I(1) is prime and π(b1 ) ∈
/ I(1) we obtain
that
(1)
(1)
(1)
yr+1 + π(b11 )y1 + . . . + π(br1 )yr(1) + h1 ∈ I(1) .
Consequently,
(1)
(1)
(1)
(1)
(1)
(1)
y1 , . . . , yr(1) , yr+1 , . . . , yr+s R(1) = y1 , . . . , yr(1) , yr+2 , . . . , yr+s R(1) .
We proceed inductively to obtain a ν1 -compatible local blowing up R −→ R(s) such
that
(s)
(s)
(s)
(s)
y1 , . . . , yr(s) , yr+1 , . . . , yr+s R(s) = y1 , . . . , yr(s) R(s) .
(s)
(s) (s)
(s)
By Lemma 4.5, we have p(s) = y1 , . . . , yr , yr+1 , . . . , yr+s R(s) and by Lemma
(s)
(s)
(s)
form a regular system of parameters.
2.6 the images of y1 , . . . , yr in Rp(s)
(s)
(s)
red
This means that y1 , . . . , yr generate p(s) . Thus we have reduced the problem to
the case when (y1 , . . . , yr ) generate p and will make this assumption from now on.
Now, the only non-trivial fact that remains to be checked is that the images of
y1 , . . . , yr in p/p2 + I are R/p-linearly independent. Take a1 , . . . , ar ∈ R such that
a1 y1 + . . . + ar yr ∈ p2 + I.
Since the images of y1 , . . . , yr in (Rp )red form a regular system of parameters, their
images in pRp /(p2 + I)Rp form an Rp /pRp -basis of pRp /(p2 + I)Rp . This implies
that a1 /1, . . . , ar /1 ∈ pRp and consequently a1 , . . . , ar ∈ p.
This completes the proof of the Lemma.
Proof of Lemma 4.3. Take y1 , . . . , ys ∈ p such that their images form an R/p-basis
of p/ p2 + I . We claim that
the images of π(y1 ), . . . , π(ys ) form an R(1) /p(1)
2
basis of p(1) / p(1) + I(1) . Take an element α ∈ p(1) . Then α = p/q where
p, q ∈ R′ := R[a1 /b, . . . , ar /b] with ν1 (p) > 0 and ν(q) = 0. Set p0 = p(0, . . . , 0)
and write
a1
ar
p = p0 + p1 + . . . + pr for some p1 , . . . , pr ∈ R′ .
b
b
12
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
This implies that p0 ∈ p. By our assumption, there exist c1 , . . . , cs ∈ p, g ∈ p2 and
h ∈ I such that
p0 = c1 y1 + . . . + cs yr + g + h.
Consequently,
α=
π(cs )
π(g) π(h) a1 p1
ar p r
π(c1 )
π(y1 ) + . . . +
π(ys ) +
+
+
+ ...+
.
q
q
q
q
b q
b q
Since a1 , . . . , ar , h ∈ I, we have that
ar p r
π(h) a1 p1
+
+ ...+
∈ I(1) .
q
b q
b q
2
This and the fact
π(g)/q
∈ p(1) imply that the images of π(y1 ), . . . , π(yr )
that
2
generate p(1) / p(1) + I(1) .
Now assume that there exists αi = (ai /bli )/(ci /bmi ) ∈ R(1) , 1 ≤ i ≤ r, such that
2
α1 π(y1 ) + . . . + αr π(yr ) ∈ p(1) + I(1) .
Then there exists n ∈ N such that
a1 bn y1 + . . . + ar bn yr ∈ p2 + I.
This implies that ai bn ∈ p for every i, 1 ≤ i ≤ r. Since b ∈
/ p, this implies that
(1)
a1 , . . . , ar ∈ p. Therefore, α1 , . . . , αr ∈ p , which concludes our proof.
Proof of Lemma 4.4. Set p′ = {a + I ∈ Rred | a ∈ p}. Since the images of the yi ’s
in p/(p2 + I) form a basis of p/(p2 + I), we conclude that (y1 , . . . , yr ) + p2 + I = p.
Applying Nakayama’s Lemma (corollary of Theorem 2.2 of [8]) we conclude that
(y1 , . . . , yr ) + I = p and consequently y1 + I, . . . , yr + I generate p′ .
Since the images of y1 , . . . , yr , x1 , . . . , xt in Rred generate m′ = {a + I ∈ Rred |
a ∈ m} we conclude that r + t ≥ dim Rred . Also, since r = dim (Rp )red = ht (p′ )
and t = dim (R/p) = ht(m/p) = ht (m′ /p′ ) we have
dim(Rred ) = ht (m′ ) ≥ ht (p′ ) + ht (m′ /p′ ) = r + t ≥ dim(Rred ).
Therefore, r + t = dim(Rred ) and hence Rred is regular.
5. Making I n /I n+1 free
Let R be a local ring and ν a valuation centered on R. Assume that
ν = ν1 ◦ ν2
and denote by p the center of ν1 on R. As usual, we set I = Nil(R) and Ip :=
(k)
(k)
Nil(Rp ). Also,
for
a
local
blowing
up
R
−→
R
we
set
I
=
Nil
R
and
(k)
(k)
Ip(k) := Nil Rp(k) . Assume that I is the only associated prime ideal of R. The
main goal of this section is to prove the following proposition.
LOCAL UNIFORMIZATION FOR NON-DOMAINS
13
Proposition 5.1. Assume that Ipn /Ipn+1 is an (Rp )red -free module for every n ∈ N.
Then there exists a local blowing up R −→ R(1) with respect to ν along an ideal
n+1
n
(b, a1 , . . . , ar ) with b ∈
/ p and a1 , . . . , ar ∈ I such that the R(1) red -module I(1)
/I(1)
is free for every n ∈ N.
In order to prove Proposition 5.1, we will need some preliminary results.
Lemma 5.2. Take elements y1 , . . . , yr+s ∈ I n such that their images in I n /I n+1
generate I n /I n+1 as an Rred -module. Consider the local blowing up π : R −→ R(1)
along the ideal (b, y1 , . . . , yr ) for some b ∈ R \ I. Set
(1)
yi
(1)
= π(yi )/b for 1 ≤ i ≤ r and yr+k = π(yr+k ) for 1 ≤ k ≤ s.
(1)
(1)
n+1
n
Then the images of y1 , . . . , yr+s in I(1)
/I(1)
form a set of generators of this
module.
n
Proof. Take an element p/q ∈ I(1)
. As in proof of the Lemma 4.5, we can write
p = p0 +
yr
y1
p1 + . . . + pr , for some p1 , . . . , pr ∈ R′
b
b
with p0 ∈ I n . This means that there exists a1 , . . . , ar+s ∈ R such that p0 − a1 y1 −
. . . − ar+s yr+s = y0 ∈ I n+1 . Consequently,
r+s
r
X
π(ai ) (1)
π(y0 )
p X π(bai ) + pi (1)
n+1
−
yi −
yi =
∈ I(1)
.
q i=1
q
q
q
i=r+1
This concludes our proof.
Lemma 5.3. Under the same assumptions as in the previous lemma, if the images of y1 , . . . , yr in I n /I n+1 are Rred -linearly independent, then the images of
(1)
(1)
n+1
n
a1 , . . . , ar in I(1)
/I(1)
are R(1) red -linearly independent.
Proof. Take elements α1 , . . . , αr ∈ R(1) such that
(6)
n+1
α1 y (1) + . . . + αr y (1) ∈ I(1)
.
We have to show that α1 , . . . , αr ∈ I(1) . For each i, 1 ≤ i ≤ r, we write αi =
(ai /bri )/(ci /bsi ) for some ai , ci ∈ R and ri , si ∈ N. Then equation (6) implies that
there exists l ∈ N and c ∈ R \ p such that
a1 bl cy1 + . . . + ar bl cyr ∈ I n+1 .
Since y1 + I n+1 , . . . , yr + I n+1 are Rred -linearly independent, this implies that
ai b l c ∈ I
for every i, 1 ≤ i ≤ r.
Since I is prime (this is a consequence of the fact that it is the only associated
prime ideal of R) and b, c ∈ R \ I, we obtain that a1 , . . . , ar ∈ I. Consequently,
α1 , . . . , αr ∈ I(1) , which concludes our proof.
14
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
Proof of Proposition 5.1. By assumption, we have that Ipn /Ipn+1 is (Rp )red -free for
every n ∈ N. Hence, by Proposition 2.6
for every ν1 -compatible local blowing up
(1)
n+1
(1)
n
-free for every n ∈ N. Therefore, it
R −→ R we have that Ip(1) /Ip(1) is Rp(1)
red
is enough to show that for a fixed n ∈ N, there exists a local blowing up R −→ R(1)
n+1
n
along an ideal (b, a1 , . . . , ar ) with b ∈
/ p and a1 , . . . , ar ∈ I such that I(1)
/I(1)
is
R(1) red -free.
Take elements y1 /b1 , . . . , yr /br ∈ Ipn , y1 , . . . , yr ∈ R and b1 , . . . , br ∈ R \ p such
that
y1 /b1 + Ipn+1 , . . . , yr /br + Ipn+1
form a basis of Ipn /Ipn+1 . We observe first that since I is prime and yi /bi ∈ Ipn , we
have yi ∈ I n for each i, 1 ≤ i ≤ r. We claim that if
y1 + I n+1 , . . . , yr + I n+1
generate I n /I n+1 as an Rred -module, then this module is free. Indeed, if there
exists ai + I ∈ Rred such that a1 y1 + . . . + ar yr ∈ I n+1 , then
a1 b1 /1 · y1 /b1 + . . . + ar br /1 · yr /br = (a1 y1 + . . . + ar yr )/1 ∈ Ipn+1 .
This implies that for each i, 1 ≤ i ≤ r, ai bi /1 ∈ Ip and consequently ai bi ci ∈ I
for some ci ∈ R \ p . Since I is prime and b1 c1 , . . . , br cr ∈ R \ I, we conclude that
a1 , . . . , ar ∈ I, which is what we wanted to prove.
If y1 + I n+1 , . . . , yr + I n+1 do not generate I n /I n+1 (as an Rred -module), then
we take yr+1 , . . . , yr+s ∈ I n such that y1 + I n+1 , . . . , yr+s + I n+1 generate I n /I n+1 .
For each k, 1 ≤ k ≤ s, since yr+k ∈ I n there exist bk ∈ R \ p, such that
bk yr+k − b1k y1 − . . . − brk yr ∈ I n+1 ,
(7)
for some b1k , . . . , brk ∈ R. Consider now the local blowing up along the ideal
(b1 , y1 , . . . , yr ). Set
(1)
yi
:= π(yi )/b1 ∈ R(1)
for each i, 1 ≤ i ≤ r
and
(1)
yr+k := π(yr+k ) ∈ R(1)
for each k, 1 ≤ k ≤ s.
From equation (7) we obtain that
(1)
(1)
n+1
yr+1 − π(b11 )y1 − . . . − π(br1 )yr(1) ∈ I(1)
,
and
(1)
(1)
n+1
π(bk )yr+k − π(b1 b1k )y1 − . . . − π(br brk )yr(1) ∈ I(1)
,
(1)
n+1
is generated in the R(1)
for every k, 2 ≤ k ≤ s. Consequently, yr+1 + I(1)
(1)
(1)
n+1
n+1
n+1
n
. Moreover,
, . . . , yr + I(1)
module I(1)
/I(1)
by y1 + I(1)
(1)
n+1
n
obtain that I(1) /I(1) is generated as an Rred -module by the
(1)
(1)
(1)
y1 , . . . , yr(1) , yr+2 , . . . , yr+s .
red
-
using Lemma 5.2, we
images of
LOCAL UNIFORMIZATION FOR NON-DOMAINS
15
(1)
(1)
n+1
n
Also, by Lemma 5.3, the images of y1 , . . . , yr in I(1)
/I(1)
are R(1) red -linearly
independent.
We proceed inductively to obtain a local blowing up R −→ R(s) such that the
(s)
(s)
n+1
n
/I(s)
is generated by the images of y1 , . . . , yr and the imR(s) red -module I(s)
(s)
(s)
n+1
n
/I(s)
are R(s) red -linearly independent.
ages of y1 , . . . , yr in I(s)
6. Proof of the main Theorem
In this section we present the proof of our main theorem.
Proof of Theorem 1.1. We will prove the assertion by induction on the rank. Since
all rank one valuations admit local uniformization by assumption, we fix n ∈ N and
will prove that if all valuations of rank smaller than n admit local uniformization,
then also valuations of rank n admit local uniformization.
Let ν be a valuation centered in the local ring R ∈ Ob(M) such that rk(ν) = n.
By Lemma 3.1, there exists a local blowing up R −→ R(1) with respect to ν such
that Nil R(1) is the only associated prime ideal of R(1) . Hence, replacing R by
R(1) , we may assume that the only associated prime ideal of R is Nil(R).
Decompose ν as ν = ν1 ◦ ν2 for valuations ν1 and ν2 with rank smaller than
n. By assumption, we know that ν1 and ν2 admit local uniformization. Since ν1
admits local uniformization, by use of Lemma 2.8, there exists a local
blowing
up
(1)
(1)
n+1
(1)
n
R −→ R with respect to ν such that Rp(1) is regular and Ip(1) /Ip(1) is Rp(1)
red
free for every n ∈ N. Replacing R by R(1) we may assume that (Rp )red is regular
and Ipn /Ipn+1 is (Rp )red -free for every n ∈ N.
Since ν2 admits local uniformization, we can use Lemma 2.9 to obtain
that
there
(1)
(1)
and
exists a local blowing up R −→ R with respect to ν such that Rp(1)
red
(1)
n+1
-free for every n ∈ N. Replacing
R(1) /p(1) are regular and Ipn(1) /Ip(1) is Rp(1)
red
R by R(1) we can assume that (Rp )red and R/p are regular and that Ipn /Ipn+1 is
(Rp )red -free for every n ∈ N.
Since (Rp )red and R/p are regular, we apply Proposition 4.1 to obtain a ν1
compatible local blowing up R −→ R(1) such that R(1) red is regular. Using
(1)
-module for every n ∈ N.
Proposition 2.6, we have that Ipn(1) /Ipn+1
Rp(1)
(1) is a free
red
By Proposition 5.1, there exists a p(1) -compatible local blowing up R(1) −→ R(2)
n+1
n
such that I(2)
/I(2)
is an R(2) red -free module for every n ∈ N. Moreover, since this
local blowing up is along an ideal (b, a1 , . . . , ar ) with b ∈
/ p(1) and a1 , . . . , ar ∈ I(1) ,
(2)
we conclude using Proposition 4.1 that R
is regular. This concludes our
red
proof.
References
[1] S. Abhyankar, Local uniformization on algebraic surfaces over ground fields of characteristic
p 6= 0, Ann. of Math. 63 (1956), 491–526.
[2] S. Abhyankar, On the valuations centered in a local domain, Amer. J. Math. 78 (1956),
321–348.
16
JOSNEI NOVACOSKI AND MARK SPIVAKOVSKY
[3] S. Abhyankar, Resolution of singularities of embedded algebraic surfaces, Pure and Applied
Mathematics 24, Academic Press, New York and London 1966.
[4] S. Abhyankar, Simultaneous resolution for algebraic surfaces, Amer. J. Math. 78 (1956),
761–790.
[5] V. Cossart and O. Piltant, Resolution of singularities of threefolds in positive characteristic.
I. Reduction to local uniformization on Artin-Schreier and purely inseparable coverings, J.
Algebra 320 no. 3 (2008), 1051–1082.
[6] V. Cossart and O. Piltant, Resolution of singularities of threefolds in positive characteristic.
II, J. Algebra 321 no. 7 (2009), 1836–1976.
[7] A. Grothendieck, Éléments de géométrie algébrique. IV. Étude locale des schémas et des
morphismes de schémas. II, Inst. Hautes Études Sci. Publ. Math. 24 (1965), 5–231.
[8] H. Matsumura, Commutative ring theory, Cambridge University Press, 1986.
[9] J. Novacoski and M. Spivakovsky, Reduction of local uniformization to the rank one case,
Proceedings of the Second International Conference on Valuation Theory, EMS Series of
Congress Reports (2014), 404–431.
[10] O. Zariski and P. Samuel, Commutative algebra, Vol. II, New York/Heidelberg/Berlin (1960).
[11] O. Zariski, Local uniformization theorem on algebraic varieties, Ann. of Math. 41 (1940),
852–896.
[12] O. Zariski, Reduction of singularities of algebraic three dimensional varieties, Ann. of Math.
45 (1944), 472 – 542.
Josnei Novacoski
CAPES Foundation
Ministry of Education of Brazil
Brası́lia/DF 70040-020
Brazil
E-mail address: [email protected]
Mark Spivakovsky
Institut de Mathématiques de Toulouse and CNRS
Université Paul Sabatier
118 route de Narbonne
F-31062 Toulouse cedex 9
France
E-mail address: [email protected]
| 0math.AC
|
arXiv:1502.02478v1 [cs.NE] 9 Feb 2015
Efficient batchwise dropout training using
submatrices
Ben Graham
[email protected]
Jeremy Reizenstein
[email protected]
Leigh Robinson
[email protected]
February 10, 2015
Abstract
Dropout is a popular technique for regularizing artificial neural networks. Dropout
networks are generally trained by minibatch gradient descent with a dropout mask
turning off some of the units—a different pattern of dropout is applied to every
sample in the minibatch. We explore a very simple alternative to the dropout mask.
Instead of masking dropped out units by setting them to zero, we perform matrix
multiplication using a submatrix of the weight matrix—unneeded hidden units are
never calculated. Performing dropout batchwise, so that one pattern of dropout is
used for each sample in a minibatch, we can substantially reduce training times.
Batchwise dropout can be used with fully-connected and convolutional neural networks.
1
Independent versus batchwise dropout
Dropout is a technique to regularize artificial neural networks—it prevents overfitting
[8]. A fully connected network with two hidden layers of 80 units each can learn to
classify the MNIST training set perfectly in about 20 training epochs—unfortunately
the test error is quite high, about 2%. Increasing the number of hidden units by a factor
of 10 and using dropout results in a lower test error, about 1.1%. The dropout network
takes longer to train in two senses: each training epoch takes several times longer,
and the number of training epochs needed increases too. We consider a technique for
speeding up training with dropout—it can substantially reduce the time needed per
epoch.
Consider a very simple `-layer fully connected neural network with dropout. To
train it with a minibatch of b samples, the forward pass is described by the equations:
xk+1 = [xk · dk ] × Wk
1
k = 0, . . . , ` − 1.
Here xk is a b × nk matrix of input/hidden/output units, dk is a b × nk dropout-mask
matrix of independent Bernoulli(1 − pk ) random variables, pk denotes the probability
of dropping out units in level k, and Wk is an nk × nk+1 matrix of weights connecting
level k with level k + 1. We are using · for (Hadamard) element-wise multiplication
and × for matrix multiplication. We have forgotten to include non-linear functions
(e.g. the rectifier function for the hidden units, and softmax for the output units) but for
the introduction we will keep the network as simple as possible.
The network can be trained using the backpropagation algorithm to calculate the
gradients of a cost function (e.g. negative log-likelihood) with respect to the Wk :
∂cost
∂cost
= [xk · dk ]T ×
∂Wk
∂xk+1
∂cost
∂cost
T
=
× Wk · dk .
∂xk
∂xk+1
With dropout training, we are trying to minimize the cost function averaged over an
ensemble of closely related networks. However, networks typically contain thousands
of hidden units, so the size of the ensemble is much larger than the number of training
samples that can possibly be ‘seen’ during training. This suggests that the independence of the rows of the dropout mask matrices dk might not be terribly important; the
success of dropout simply cannot depend on exploring a large fraction of the available
dropout masks. Some machine learning libraries such as Pylearn2 allow dropout to
be applied batchwise instead of independently1 . This is done by replacing dk with a
1×nk row matrix of independent Bernoulli(1−pk ) random variables, and then copying
it vertically b times to get the right shape.
To be practical, it is important that each training minibatch can be processed quickly.
A crude way of estimating the processing time is to count the number of floating point
multiplication operations needed (naively) to evaluate the × matrix multiplications
specified above:
`−1
X
k=0
b × nk × nk+1 + nk × b × nk+1 + b × nk+1 × nk .
|
{z
} |
{z
} |
{z
}
forwards
∂cost/∂W
backwards
However, when we take into account the effect of the dropout mask, we see that many
of these multiplications are unnecessary. The (i, j)-th element of the Wk weight matrix
effectively ‘drops-out’ of the calculations if unit i is dropped in level k, or if unit j is
dropped in level k + 1. Applying 50% dropout in levels k and k + 1 renders 75% of
the multiplications unnecessary.
If we apply dropout independently, then the parts of Wk that disappear are different for each sample. This makes it effectively impossible to take advantage of the
redundancy—it is slower to check if a multiplication is necessary than to just do the
multiplication. However, if we apply dropout batchwise, then it becomes easy to take
advantage of the redundancy. We can literally drop-out redundant parts of the calculations.
1 Pylearn2:
see function apply dropout in mlp.py
2
●
4000N
●
2000N
●
●
1000N
●
●
●
●
●
●
●
●
1000
3000
70
●
●
●
4000N
●
2000N
●
1000N
●
●
●
●
●
●
●
●
●
500N
●
●
100
300
●
0.1
500N
●
●
●
50
●
●
60
8000N
●
●
% Time Saving
●
●
●
●
●
40
●
●
●
8000N
30
30.0
10.0
3.0
1.0
●
●
●
●
0.3
Epoch training time, seconds
No dropout
Batchwise
●
100
300
Minibatch size
1000
3000
Minibatch size
Figure 1: Left: MNIST training time for three layer networks (log scales) on an
NVIDIA GeForce GTX 780 graphics card. Right: Percentage reduction in training
times moving from no dropout to batchwise dropout. The time saving for the 500N
network with minibatches of size 100 increases from 33% to 42% if you instead compare batchwise dropout with independent dropout.
The binary 1 × nk batchwise dropout matrices dk naturally define submatrices of
the weight and hidden-unit matrices. Let xdropout
:= xk [ : , dk ] denote the submatrix
k
of xk consisting of the level-k hidden units that survive dropout. Let Wkdropout :=
Wk [dk , dk+1 ] denote the submatrix of Wk consisting of weights that connect active
units in level k to active units in level k + 1. The network can then be trained using the
equations:
xdropout
= xdropout
× Wkdropout
k+1
k
∂cost
∂cost
= (xdropout
)T ×
k
dropout
∂Wk
∂xdropout
k+1
∂cost
∂cost
dropout T
=
× (Wk
)
∂xdropout
∂xdropout
k
k+1
The redundant multiplications have been eliminated. There is an additional benefit
in terms of memory needed to store the hidden units: xdropout
needs less space than
k
xk . In Section 2 we look at the performance improvement that can be achieved using
CUDA/CUBLAS code running on a GPU. Roughly speaking, processing a minibatch
with 50% batchwise dropout takes as long as training a 50% smaller network on the
same data. This explains the nearly overlapping pairs of lines in Figure 1.
We should emphasize that batchwise dropout only improves performance during
training; during testing the full Wk matrix is used as normal, scaled by a factor of
1 − pk . However, machine learning research is often constrained by long training times
and high costs of equipment. In Section 3 we show that all other things being equal,
batchwise dropout is similar to independent dropout, but faster. Moreover, with the
3
increase in speed, all other things do not have to be equal. With the same resources,
batchwise dropout can be used to
• increase the number of training epochs,
• increase the number of hidden units,
• increase the number of validation runs used to optimize “hyper-parameters”, or
• to train a number of independent copies of the network to form a committee.
These possibilities will often be useful as ways of improving generalization/reducing
test error.
In Section 4 we look at batchwise dropout for convolutional networks. Dropout
for convolutional networks is more complicated as weights are shared across spatial
locations. A minibatch passing up through a convolutional network might be represented at an intermediate hidden layer by an array of size 100 × 32 × 12 × 12: 100
samples, the output of 32 convolutional filters, at each of 12 × 12 spatial locations. It
is conventional to use a dropout mask with shape 100 × 32 × 12 × 12; we will call this
independent dropout. In contrast, if we want to apply batchwise dropout efficiently by
adapting the submatrix trick, then we will effectively be using a dropout mask with
shape 1 × 32 × 1 × 1. This looks like a significant change: we are modifying the
ensemble over which the average cost is optimized. During training, the error rates are
higher. However, testing the networks gives very similar error rates.
1.1
Fast dropout
We might have called batchwise dropout fast dropout but that name is already taken
[11]. Fast dropout is very different approach to solving the problem of training large
neural network quickly without overfitting. We discuss some of the differences of the
two techniques in the appendix.
2
Implementation
In theory, for n × n matrices, addition is an O(n2 ) operation, and multiplication is
O(n2.37... ) by the Coppersmith–Winograd algorithm. This suggests that the bulk of
our processing time should be spent doing matrix multiplication, and that a performance improvement of about 60% should be possible compared to networks using independent dropout, or no dropout at all. In practice, SGEMM functions use Strassen’s
algorithm or naive matrix multiplication, so performance improvement of up to 75%
should be possible.
We implemented batchwise dropout for fully-connected and convolutional neural
networks using CUDA/CUBLAS2 . We found that using the highly optimized cublasSgemm function to do the bulk of the work, with CUDA kernels used to form the submatrices Wkdropout and to update the Wk using ∂cost/∂Wkdropout , worked well. Better
2 Software
available at http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic-research/graham/
4
performance may well be obtained by writing a SGEMM-like matrix multiplication
function that understands submatrices.
For large networks and minibatches, we found that batchwise dropout was substantially faster, see Figure 1. The approximate overlap of some of the lines on the left
indicates that 50% batchwise dropout reduces the training time in a similar manner to
halving the number of hidden units.
The graph on the right show the time saving obtained by using submatrices to implement dropout. Note that for consistency with the left hand side, the graph compares
batchwise dropout with dropout-free networks, not with networks using independent
dropout. The need to implement dropout masks for independent dropout means that
Figure 1 slightly undersells the performance benefits of batchwise dropout as an alternative to independent dropout.
For smaller networks, the performance improvement is lower—bandwidth issues
result in the GPU being under utilized. If you were implementing batchwise dropout
for CPUs, you would expect to see greater performance gains for smaller networks as
CPUs have a lower processing-power to bandwidth ratio.
2.1
Efficiency tweaks
If you have n = 2000 hidden units and you drop out p = 50% of them, then the
number of dropped units is approximately np = 1000, but with some small variation as
you
standard deviation is
p are really dealing with a Binomial(n, p) random variable—its
np(1 − p) = 22.4. The sizes of the submatrices Wkdropout and xdropout
are therefore
k
slightly random. In the interests of efficiency and simplicity, it is convenient to remove
this randomness. An alternative to dropping each unit independently with probability p
is to dropa subset of exactly np of the hidden units, uniformly at random from the set
n
of all np
such subsets. It is still the case that each unit is dropped out with probability
p. However, within a hidden layer we no longer have strict independence regarding
which units are dropped out. The probability of dropping out the first two hidden units
changes very slightly, from
p2 = 0.25
to
np np − 1
·
= 0.24987....
n n−1
Also, we used a modified form of NAG-momentum minibatch gradient descent [9].
After each minibatch, we only updated the elements of Wkdropout , not all the element
of Wk . With vk and vkdropout denoting the momentum matrix/submatrix corresponding
to Wk and Wkdropout , our update was
vkdropout ← µvkdropout − ε(1 − µ)∂cost/∂Wkdropout
Wkdropout ← Wkdropout + vkdropout .
The momentum still functions as an autoregressive process, smoothing out the gradients, we are just reducing the rate of decay µ by a factor of (1 − pk )(1 − pk+1 ).
5
3.5
3.0
1.0
1.5
2.0
2.5
Test
Train
0.5
0.0
% Errors after 1000 epochs
●
●
1
●
●
4
●
●
16
●
●
64
●
●
256
●
●
●
●
●
●
1024 4096
Number of dropout patterns used
Figure 2: Dropout networks trained using a restricted the number of dropout patterns
(each × is from an independent experiment). The blue line marks the test error for a
network with half as many hidden units trained without dropout.
3
Results for fully-connected networks
The fact that batchwise dropout takes less time per training epoch would count for
nothing if a much larger number of epochs was needed to train the network, or if a
large number of validation runs were needed to optimize the training process. We
have carried out a number of simple experiment to compare independent and batchwise
dropout. In many cases we could have produced better results by increasing the training
time, annealing the learning rate, using validation to adjust the learning process, etc.
We choose not to do this as the primary motivation for batchwise dropout is efficiency,
and excessive use of fine-tuning is not efficient.
For datasets, we used:
• The MNIST3 set of 28 × 28 pixel handwritten digits.
• The CIFAR-10 dataset of 32x32 pixel color pictures ([4]).
• An artificial dataset designed to be easy to overfit.
Following [8], for MNIST and CIFAR-10 we trained networks with 20% dropout in
the input layer, and 50% dropout in the hidden layers. For the artificial dataset we
increased the input-layer dropout to 50% as this reduced the test error. In some cases,
we have used relatively small networks so that we would have time to train a number
of independent copies of the networks. This was useful in order to see if the apparent
differences between batchwise and independent dropout are significant or just noise.
3 http://yann.lecun.com/exdb/mnist/
6
3.1
MNIST
Our first experiment explores the effect of dramatically restricting the number of dropout
patterns seen during training. Consider a network with three hidden layers of size 1000,
trained for 1000 epochs using minibatches of size 100. The number of distinct dropout
patterns, 23784 , is so large that we can assume that we will never generate the same
dropout mask twice. During independent dropout training we will see 60 million different dropout patterns, during batchwise dropout training we will see 100 times fewer
dropout patterns.
For both types of dropout, we trained 12 independent networks for 1000 epochs,
with batches of size 100. For batchwise dropout we got a mean test error of 1.04%
[range (0.92%,1.1%), s.d. 0.057%] and for independent dropout we got a mean test
errors of 1.03% [range (0.98%,1.08%), s.d. 0.033%]. The difference in the mean test
errors is not statistically significant.
To explore further the reduction in the number of dropout patterns seen, we changed
our code for (pseudo)randomly generating batchwise dropout patterns to restrict the
number of distinct dropout patterns used. We modified it to have period n minibatches,
with n = 1, 2, 4, 8, . . . ; see Figure 2. For n = 1 this corresponds to only ever using one dropout mask, so that 50% of the network’s 3000 hidden weights are never
actually trained (and 20% of the 784 input features are ignored). During training this
corresponds to training a dropout-free network with half as many hidden units—the test
error for such a network is marked by a blue line in Figure 2. The error during testing
is higher than the blue line because the untrained weights add noise to the network.
If n is less than thirteen, is it likely that some of the networks 3000 hidden units
are dropped out every time and so receive no training. If n is in the range thirteen to
fifty, then it is likely that every hidden unit receives some training, but some pairs of
hidden units in adjacent layers will not get the chance to interact during training, so
the corresponding connection weight is untrained. As the number of dropout masks
increases into the hundreds, we see that it is quickly a case of diminishing returns.
3.2
Artificial dataset
To test the effect of changing network size, we created an artificial dataset. It has 100
classes, each containing 1000 training samples and 100 test samples. Each class is defined using an independent random walk of length 1000 in the discrete cube {0, 1}1000 .
For each class we generated the random walk, and then used it to produce the training
and test samples by randomly picking points along the length of walk (giving binary sequences of length 1000) and then randomly flipping 40% of the bits. We trained three
layer networks with n ∈ {250, 500, 1000, 2000} hidden units per layer with minibatches of size 100. See Figure 3.
Looking at the training error against training epochs, independent dropout seems to
learn slightly faster. However, looking at the test errors over time, there does not seem
to be much difference between the two forms of dropout. Note that the x-axis is the
number of training epochs, not the training time. The batchwise dropout networks are
learning much faster in terms of real time.
7
10
50
Independent
Batchwise
6
250N
4
30
500N
% Test Error
40
35
250N
25
% Train Error
8
45
Independent
Batchwise
500N
2000N
0
100
1000N
2000N
2
20
1000N
200
0
Epoch
100
200
Epoch
Figure 3: Artificial dataset. 100 classes each corresponding to noisy observations of a
one dimensional manifold in {0, 1}1000 .
3.3
CIFAR-10 fully-connected
Learning CIFAR-10 using a fully connected network is rather difficult. We trained
three layer networks with n ∈ {125, 250, 500, 1000, 2000} hidden units per layer with
minibatches of size 1000. We augmented the training data with horizontal flips. See
Figure 4.
4
Convolutional networks
Dropout for convolutional networks is more complicated as weights are shared across
spatial locations. Suppose layer k has spatial size sk × sk with nk features per spatial
location, and if the k-th operation is a convolution with f × f filters. For a minibatch
of size b, the convolution involves arrays with sizes:
layer k : b × nk × sk × sk
weights Wk : nk+1 × nk × f × f
Dropout is normally applied using dropout masks with the same size as the layers. We
will call this independent dropout—independent decisions are mode at every spatial
location. In contrast, we define batchwise dropout to mean using a dropout mask with
shape 1 × nk × 1 × 1. Each minibatch, each convolutional filter is either on or off—
across all spatial locations.
These two forms of regularization seem to be doing quite different things. Consider a filter that detects the color red, and a picture with a red truck in it. If dropout is
applied independently, then by the law of averages the message “red” will be transmitted with very high probability, but with some loss of spatial information. In contrast,
8
60
Independent
Batchwise
50
50
Independent
Batchwise
30
45
250N
40
20
500N
% Test Error
40
125N
250N
1000N
500N
2000N
1000N
2000N
0
35
10
% Train Error
125N
0
500
1000
0
500
Epoch
1000
Epoch
Figure 4: Results for CIFAR-10 using fully-connected networks of different sizes.
with batchwise dropout there is a 50% chance we delete the entire filter output. Experimentally, the only substantial difference we could detect was that batchwise dropout
resulted in larger errors during training.
To implement batchwise dropout efficiently, notice that the 1 × nk × 1 × 1 dropout
masks corresponds to forming subarrays Wkdropout of the weight arrays Wk with size
(1 − pk+1 )nk+1 × (1 − pk )nk × f × f.
The forward-pass is then simply a regular convolutional operation using Wkdropout ; that
makes it possible, for example, to take advantage of the highly optimized cudnnConvolutionForward
function from the NVIDIA cuDNN package.
4.1
MNIST
For MNIST, we trained a LeNet-5 type CNN with two layers of 5 × 5 filters, two
layers of 2 × 2 max-pooling, and a fully connected layer [6]. There are three places for
applying 50% dropout:
50%
50%
50%
32C5 − MP2 − 64C5 − MP2 − 512N − 10N.
The test errors for the two dropout methods are similar, see Figure 5.
4.2
CIFAR-10 with varying dropout intensity
For a first experiment with CIFAR-10 we used a small convolutional network with
small filters. The network is a scaled down version of the network from [1]; there are
four places to apply dropout:
p
p
p
p
128C3 − MP2 − 256C2 − MP2 − 384C2 − MP2 − 512N − 10N.
9
0.7
0.6
0.4
0.5
% Test errror
0.8
Independent
Batchwise
50
100
150
200
Epochs
Figure 5: MNIST test errors, training repeated three times for both dropout methods.
The input layer is 24 × 24. We trained the network for 1000 epochs using randomly
chosen subsets of the training images, and reflected each image horizontally with probability one half. For testing we used the centers of the images.
In Figure 6 we show the effect of varying the dropout probability p. The training
errors are increasing with p, and the training errors are higher for batchwise dropout.
The test-error curves both seem to have local minima around p = 0.2. The batchwise
test error curve seems to be shifted slightly to the left of the independent one, suggesting that for any given value of p, batchwise dropout is a slightly stronger form of
regularization.
4.3
CIFAR-10 with many convolutional layers
We trained a deep convolutional network on CIFAR-10 without data augmentation.
Using the notation of [2], our network has the form
√
3
(64nC2 − F M P 2)12 − 832C2 − 896C1 − output,
i.e. it consists of 12 2 × 2 convolutions with 64n filters in the n-th layer, 12 layers
max-pooling, followed by two fully connected layers; the network has 12.6 million parameters. We used an increasing amount of dropout per layer, rising linearly from 0%
dropout after the third layer to 50% dropout after the 14th. Even though the amount
of dropout used in the middle layers is small, batchwise dropout took less than half as
long per epoch as independent dropout; this is because applying small amounts of independent dropout in large hidden-layers creates a bandwidth performance-bottleneck.
As the network’s max-pooling operation is stochastic, the test errors can be reduced
by repetition. Batchwise dropout resulted in a average test error of 7.70% (down to
5.78% with 12-fold testing). Independent dropout resulted in an average test error of
7.63% (reduced to 5.67% with 12-fold testing).
10
●
●
●
15
Independent
Testing
●
●
●
●
●
10
●
●
●
●
●
●
Batchwise
Training ●
5
% error
Batchwise
Testing
●
●
●
●
●
●
●
Independent
Training
●
0
●
●
●
0.0
0.1
0.2
0.3
0.4
p
Figure 6: CIFAR-10 results using a convolutional network with dropout probability
p ∈ (0, 0.4). Batchwise dropout produces a slightly lower minimum test error.
5
Conclusions and future work
We have implemented an efficient form of batchwise dropout. All other things being
equal, it seems to learn at roughly the same speed as independent dropout, but each
epoch is faster. Given a fixed computational budget, it will often allow you to train
better networks.
There are other potential uses for batchwise dropout that we have not explored yet:
• Restricted Boltzmann Machines can be trained by contrastive divergence [3] with
dropout [8]. Batchwise dropout could be used to increase the speed of training.
• When a fully connected network sits on top of a convolutional network, training the top and bottom of the network can be separated over different computational nodes [5]. The fully connected top-parts of the network typically contains
95% of the parameters—keeping the nodes synchronized is difficult due to the
large size of the matrices. With batchwise dropout, nodes could communicate
∂cost/∂Wkdropout instead of ∂cost/∂Wk and so reducing the bandwidth needed.
• Using independent dropout with recurrent neural networks can be too disruptive
to allow effective learning; one solution is to only apply dropout to some parts
of the network [12]. Batchwise dropout may provide a less damaging form of
dropout, as each unit will either be on or off for the whole time period.
• Dropout is normally only used during training. It is generally more accurate
use the whole network for testing purposes; this is equivalent to averaging over
the ensemble of dropout patterns. However, in a “real-time” setting, such as
11
analyzing successive frames from a video camera, it may be more efficient to
use dropout during testing, and then to average the output of the network over
time.
• Nested dropout [7] is a variant of regular dropout that extends some of the properties of PCA to deep networks. Batchwise nested dropout is particularly easy
to implement as the submatrices are regular enough to qualify as matrices in the
context of the SGEMM function (using the LDA argument).
• DropConnect is an alternative form of regularization to dropout [10]. Instead of
dropping hidden units, individual elements of the weight matrix are dropped out.
Using a modification similar to the one in Section 2.1, there are opportunities for
speeding up DropConnect training by approximately a factor of two.
References
[1] D. Ciresan, U. Meier, and J. Schmidhuber. Multi-column deep neural networks
for image classification. In Computer Vision and Pattern Recognition (CVPR),
2012 IEEE Conference on, pages 3642–3649, 2012.
[2] Ben Graham. Fractional max-pooling, 2014. http://arxiv.org/abs/1412.6071.
[3] Hinton and Salakhutdinov. Reducing the Dimensionality of Data with Neural
Networks. SCIENCE: Science, 313, 2006.
[4] Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images. Technical report, 2009.
[5] Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks,
2014. http://arxiv.org/abs/1404.5997.
[6] Y. L. Le Cun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of IEEE, 86(11):2278–2324, November 1998.
[7] Oren Rippel, Michael A. Gelbart, and Ryan P. Adams. Learning ordered representations with nested dropout, 2014. http://arxiv.org/abs/1402.0915.
[8] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A Simple Way to Prevent Neural Networks from
Overfitting. Journal of Machine Learning Research, 15:1929–1958, 2014.
[9] Ilya Sutskever, James Martens, George E. Dahl, and Geoffrey E. Hinton. On the
importance of initialization and momentum in deep learning. In ICML, volume 28
of JMLR Proceedings, pages 1139–1147. JMLR.org, 2013.
[10] Li Wan, Matthew Zeiler, Sixin Zhang, Yann Lecun, and Rob Fergus. Regularization of Neural Networks using DropConnect, 2013. JMLR W&CP 28 (3) :
1058–1066, 2013.
12
[11] Sida Wang and Christopher Manning. Fast dropout training. JMLR W & CP,
28(2):118–126, 2013.
[12] Wojciech Zaremba, Ilya Sutskever, and Oriol Vinyals. Recurrent neural network
regularization, 2014. http://arxiv.org/abs/1409.2329.
13
A
Fast dropout
We might have called batchwise dropout fast dropout but that name is already taken
[11]. Fast dropout is an alternative form of regularization that uses a probabilistic
modeling technique to imitate the effect of dropout; each hidden unit is replaced with
a Gaussian probability distribution. The fast relates to reducing the number of training
epochs needed compared to regular dropout (with reference to results in a preprint4
of [8]). Training a network 784-800-800-10 on the MNIST dataset with 20% input
dropout and 50% hidden-layer dropout, fast dropout converges to a test error of 1.29%
after 100 epochs of L-BFGS. This appears to be substantially better than the test error
obtained in the preprint after 100 epochs of regular dropout training.
However, this is a dangerous comparison to make. The authors of [8] used a
learning-rate scheme designed to produce optimal accuracy eventually, not after just
one hundred epochs. We tried using batchwise dropout with minibatches of size 100
and an annealed learning rate of 0.01e−0.01×epoch . We trained a network with two
hidden layers of 800 rectified linear units each. Training for 100 epochs resulted in a
test error of 1.22% (s.d. 0.03%). After 200 epochs the test error has reduced further
to 1.12% (s.d. 0.04%). Moreover, per epoch, batchwise-dropout is faster than regular
dropout while fast-dropout is slower. Assuming we can make comparisons across different programs5 , the 200 epochs of batchwise dropout training take less time than the
100 epoch of fast dropout training.
4 http://arxiv.org/abs/1207.0580
5 Using our software to implement the network, each batchwise dropout training epoch take 0.67 times
as long as independent dropout. In [11] a figures of 1.5 is given for the ratio between fast- and independentdropout when using minibatch SGD; when using L-BFGS to train fast-dropout networks the training time per
epoch will presumably be even more than 1.5 times longer, as L-BFGS use line-searches requiring additional
forward passes through the neural network.
14
| 9cs.NE
|
Matching while Learning
Ramesh Johari∗
Vijay Kamble†
Yash Kanoria‡
June 20, 2017
arXiv:1603.04549v2 [cs.LG] 18 Jun 2017
Abstract
We consider the problem faced by a service platform that needs to match supply with
demand, but also to learn attributes of new arrivals in order to match them better in the
future. We introduce a benchmark model with heterogeneous workers and jobs that arrive
over time. Job types are known to the platform, but worker types are unknown and must be
learned by observing match outcomes. Workers depart after performing a certain number
of jobs. The payoff from a match depends on the pair of types and the goal is to maximize
the steady-state rate of accumulation of payoff.
Our main contribution is a complete characterization of the structure of the optimal policy in the limit that each worker performs many jobs. The platform faces a trade-off for each
worker between myopically maximizing payoffs (exploitation) and learning the type of the
worker (exploration). This creates a multitude of multi-armed bandit problems, one for each
worker, coupled together by the constraint on availability of jobs of different types (capacity
constraints). We find that the platform should estimate a shadow price for each job type,
and use the payoffs adjusted by these prices, first, to determine its learning goals and then,
for each worker, (i) to balance learning with payoffs during the “exploration phase”, and (ii)
to myopically match after it has achieved its learning goals during the “exploitation phase."
Keywords: matching, learning, two-sided platform, multi-armed bandit, capacity constraints.
1
Introduction
This paper considers a central operational challenge faced by platforms that serve as matchmakers between supply and demand. Such platforms face a fundamental exploration-exploitation
trade-off: on the one hand, efficient operation involves making matches that generate the most
value (“exploitation”); on the other hand, the platform must continuously learn about newly
arriving participants, so that they can be efficiently matched (“exploration”). In this paper, we
develop a structurally simple and nearly optimal approach to resolving this trade-off.
In the model we consider, there are two groups of participants: workers and jobs. The
terminology is inspired by online labor markets (e.g., Upwork for remote work, Handy for housecleaning, Thumbtack and Taskrabbit for local tasks, etc.); however, our model can be viewed
as a stylized abstraction of many other matching platforms as well. Time is discrete, and new
workers and jobs arrive at the beginning of every time period. Workers depart after performing
∗
Stanford University ([email protected])
Stanford University ([email protected])
‡
Columbia Business School ([email protected])
†
1
a specified number of jobs. Each time a worker and job are matched, a (random) payoff is
generated and observed by the platform, where the payoff distribution depends on the worker
type and the job type.
As our emphasis is on the interaction between matching and learning, our model has several
features that focus our analysis in this paper. First, we assume that the platform centrally
controls matching: at the beginning of each time period, the platform matches each worker in
the system to an available job. Second, strategic considerations are not modeled; this remains an
interesting direction for future work. Finally, we focus on the goal of maximizing the steady-state
rate of payoff generation.1
We now describe the learning challenge faced by the platform. In most platforms, more is
known about one side of the platform than the other; accordingly, we assume job types are
known, while the type of a new worker is unknown. The platform learns about workers’ types
through the payoffs obtained when they are matched to jobs. However, because the supply of
jobs is limited, using jobs to learn can reduce immediate payoffs, as well as deplete the supply
of jobs available to the rest of the marketplace. Thus the presence of capacity constraints forces
us to carefully design both exploration and exploitation in the matching algorithm in order to
optimize the rate of payoff generation.
Our main contribution in this paper is the development of a matching policy that is nearly
payoff optimal. Our algorithm is divided into two phases in each worker’s lifetime: exploration
(identification of the worker type) and exploitation (optimal matching given the worker’s identified type). We refer to our policy as DEEM: Decentralized Explore-then-Exploit for Matching.
To develop intuition for our solution, consider a simple example with two types of jobs (“easy”
and “hard”) and two types of workers (“expert” and “novice”). Experts can do both types of tasks
well; but novices can only do easy tasks well. Suppose that there is a limited supply of easy jobs:
more than the mass of novices available, but less than the total mass of novices and experts. In
particular, to maximize payoff the platform must learn enough to match some experts to hard
jobs.
DEEM has several key features, each of which can be understood in the context of this
example. First, DEEM has a natural decentralization property: it determines the choice of
job type for a worker based only on that worker’s history. This decentralization is arguably
essential in large-scale online platforms, where matching is typically carried out on an individual
basis, rather than centrally. In order to accomplish this decentralization, it is essential for the
algorithm to account for the externality to the rest of the market when a worker is matched to a
given job. For example, if easy jobs are relatively scarce, then matching a worker to such a job
makes it unavailable to the rest of the market. Our approach is to “price” this externality: we
find shadow prices for the capacity constraints, and adjust per-match payoffs downward using
these prices.
Second, our algorithm design specifies learning goals that ensure an efficient balance between
exploration and exploitation. In particular, in our example, we note that there are two kinds of
errors possible while exploring: misclassifying a novice as an expert, and vice versa. Occasionally
mislabeling experts as novices is not catastrophic: some experts need to do easy jobs anyway,
and so the algorithm can account for such errors in the exploitation phase. Thus, relatively less
effort can be invested in minimizing this error type. However, mistakenly labeling novices as
experts can be catastrophic: in this case, novices will be matched to hard jobs in the exploitation
1
This is a reasonable proxy for the goal of a platform that, say, takes a fraction of the total surplus generated
through matches. More generally, we believe that this is a benchmark problem whose solution informs algorithmic
design for settings with other related objectives, such as revenue maximization.
2
phase, causing substantial loss of payoff; thus the probability of such errors must be kept very
small. A major contribution of our work is to precisely identify the correct learning goals in
the exploration phase, and to then design DEEM to meet these learning goals while maximizing
payoff generation.
Third, DEEM involves a carefully constructed exploitation phase to ensure that capacity
constraints are met while maximizing payoffs. A naive approach during the exploitation phase
would match a worker to any job type that yields the maximum externality-adjusted payoff
corresponding to his type label. It turns out that such an approach leads to significant violations
of capacity constraints, and hence poor performance. The reason is that in a generic capacitated
problem instance, one or more worker types are indifferent between multiple job types, and
suitable tie-breaking is necessary to achieve good performance. In our theoretical development,
we achieve this by modifying the solution to the static optimization problem with known worker
types, whereas our practical implementation of DEEM achieves appropriate tie-breaking via
simple but dynamically updated shadow prices.
Our main result (Theorem 5.1) shows that DEEM achieves essentially optimal regret as the
number of jobs N performed by each worker during their lifetime grows, where regret is the loss
in payoff accumulation rate relative to the maximum achievable with known worker types. In
our setting, a lower bound on the regret is (C log N/N )(1 + o(1)) for some C ∈ [0, ∞) that is a
function of system parameters. DEEM achieves this level of regret to leading order when C > 0,
while it achieves a regret of O(log log N/N ) if C = 0.
Situations where C > 0 are those in which there is an inherent tension between the goals
of learning and payoff maximization. To develop intuition, consider an expanded version of the
example above, where each worker can be either an expert or novice programmer, as well as an
expert or novice graphic designer. Suppose the supply of jobs is such that if worker types were
known, only expert graphic designers who are also novice programmers would be matched to
graphic design jobs.2 But if we are learning worker types, then expert graphic designers must
be matched to approximately O(log N ) programming jobs to learn whether they are novice
or expert programmers – and in turn, whether they should be matched to graphic design or
programming jobs, respectively. Thus O(log N/N ) average regret per period is incurred relative
to the optimal solution with known types. DEEM precisely minimizes the regret incurred while
these distinctions are made, thus achieving the lower bound on the regret.
Our theory is complemented by a practical heuristic that we call DEEM+ , which optimizes
performance for small values of N , and an implementation and simulation that demonstrates a
natural way of translating our work into practice. In particular, our simulations reveal substantial benefit from jointly managing capacity constraints and learning, as we do in DEEM and
DEEM+ .
The remainder of the paper is organized as follows. After discussing related work in Section 2,
we present our model and outline the optimization problem of interest to the platform in Section
3. In Section 4, we discuss the three key ideas above in the design of DEEM, and present its
formal definition. In Section 5, we present our main theorem, and discuss the optimal regret
scaling. In Section 6 we present a sketch of the proof of the main result. In Section 7, discuss
practical implementation of DEEM and present the heuristic DEEM+ . In Section 8, we use
simulations to compare the performance of DEEM and DEEM+ with well-known multi-armed
bandit algorithms. We conclude in Section 9. All proofs are in the appendices.
2
This would be the case, e.g., if programming jobs are both in high demand and more valuable (conditional
on successful completion) than graphic design jobs.
3
2
Related literature
A foundational model for investigating the exploration-exploitation tradeoff is the stochastic
multi-armed bandit (MAB) problem [9, 18, 23]. The goal is to find an adaptive expected-regretminimizing policy for choosing among arms with unknown payoff distributions (where regret
is measured against the expected payoff of the best arm) [4, 10, 28]. The closest work in this
literature to our paper is by Agrawal et al. [1]; in their model, they assume the joint vector
of arm distributions can only take on one of finitely many values. This introduces correlation
across different arms; depending on certain identifiability conditions, the optimal regret is either
Θ(1/N ) or Θ(log N/N ). In our model, the analog is that job types are arms, and for each worker
we solve a MAB problem to identify the true type of a worker, from among a finite set of possible
worker types.
Our work is also related to recent literature on MAB problems with capacity constraints; we
refer to these broadly as bandits with knapsacks. The formulation is the same as the classical
MAB problem, with the modification that every pull of an arm depletes a vector of resources
which are limited in supply [14]. The formulation subsumes several related problems in revenue
management under demand uncertainty [11, 16, 17, 34, 37], and budgeted dynamic procurement
[13, 36]; there have been a variety of extensions [2, 15], with recently a significant generalization
of the problem to a contextual bandit setting, with concave rewards and convex constraints
[3, 5]. There is considerable difference between our model and bandits with knapsacks. Bandits
with knapsacks consider a single MAB problem over a fixed time horizon. Our setting on the
other hand can be seen as a system with an ongoing arriving stream of MAB problems, one
per worker; these MAB problems are coupled together by the capacity constraints on arriving
jobs. Indeed, as noted in the introduction, a significant structural point for us is to solve these
problems in a decentralized manner, to ease their implementation in large-scale online platforms.
We conclude by discussing some other directions of work that are related to this paper.
There are a number of recent pieces of work that consider efficient matching in dynamic twosided matching markets [6, 7, 12, 20, 21, 24, 25, 27]; a related class of dynamic resource allocation
problems, online bipartite matching, is also well studied in the computer science community (see
[32] for a survey). Similar to the current paper, Fershtman and Pavan [22] also study matching
with learning, mediated by a central platform. Relative to our model, their work does not
have constraints on the number of matches per agent, while it does consider agent incentives.
Finally, a recent work [31] studies a pure learning problem in a setting similar to ours with
capacity constraints on each type of server/expert; while there are some similarities in the style
of analysis, that paper focuses exclusively on learning the exact type, rather than balancing
exploration and exploitation as we do in this paper.
3
The model and the optimization problem
In this section we first describe our model. In particular, we describe the primitives of our
platform (“workers” and “jobs”), and givea formal specification of the matching process we study.
We conclude by precisely defining the optimization problem of interest that we solve in this paper.
3.1
Preliminaries
Workers and jobs. For convenience we adopt the terminology of workers and jobs to describe
the two sides of the market. We assume a fixed set of job types J , and a fixed set of worker
4
types I.
A key point is that the model we consider is a continuum model, and so the evolution of the
system will be described by masses of workers and jobs.3 In particular, at each time step, a mass
ρ̂(i) > 0 of workers of type i and a mass µ(j) > 0 of jobs of type j arrive. In what follows, we
model the scenario where type uncertainty exists only for workers; i.e., the platform will know
the types of arriving jobs exactly, but will need to learn the types of arriving workers. We also
assume for now that the arrival rates of jobs and workers are known to the platform; later in
Section 7, we discuss how the platform might account for the possibility that these parameters
are unknown.
Matching and the payoff matrix. If a mass of workers of type i is matched to a mass
of jobs of type j, we assume that a fraction A(i, j) of this mass of matches generates a reward
of 1 (per unit mass), while a fraction 1 − A(i, j) generates a reward of zero (per unit mass).
This formal specification is meant to capture a large-system model in a setting where matches
between type i workers and type j jobs generate a Bernoulli(A(i, j)) payoff. We do not concern
ourselves with the division of payoffs between workers and employers in this paper; instead we
assume that the platform’s goal is to maximize the total rate of payoff generation.4 We call the
matrix A the payoff matrix; throughout, we assume that no two rows of A are identical.5
A key assumption in our work is that the platform knows the matrix A. In particular, we
are considering a platform that has enough aggregate information to understand compatibility
between different worker and job types; however, for any given worker newly arriving to the
platform, the platform does not know the worker’s type. Thus, from the perspective of the
platform, there will be uncertainty in payoffs in each period because although the platform
knows that a given mass of workers of type i exist in the platform, the identity of the workers
of type i is not known.
We define an “empty” job type κ, such that all worker types matched to κ generate zero
reward, i.e., A(i, κ) = 0 for all i. We view κ as representing the possibility that a worker
goes unmatched, and thus assume that an unbounded capacity of job type κ is available, i.e.,
µ(κ) = ∞.
Worker lifetimes. We imagine that each arriving worker lives in the system for N time
steps,6 and has the opportunity to be matched to a job in each time step (so each job takes one
unit of time to complete). We assume the platform knows N .
Note that we have ρ(i) = ρ̂(i)N as the total mass of workers of type i in the system at each
time step. For our theoretical analysis, we later consider a scaling regime where N → ∞, and
ρ̂(i) → 0, while ρ(i) remains fixed. In this regime, worker lifetimes grow to infinity, and arrival
rates scale down, but the total mass of workers of each type available in each time period remains
fixed.
Generalized imbalance. Throughout our technical development, we make a mild structural
assumption on the problem instance, defined by the tuple (ρ, µ, A). This is captured by the
following definition. We say that arrival rates ρ = (ρ(i))i∈I and µ = (µ(j))j∈J satisfy the
generalized imbalance condition if there is no pair of nonempty subsets of worker types and job
3
Formally, this can be seen as a continuum scaling of a discrete system; see, e.g., [19, 29, 30]
This would be the case, e.g., in a platform where the operator takes a fixed percentage of the total payoff
generated from a match.
5
This mild requirement simply ensures that it is possible, in principle, to distinguish between each pair of
worker types.
6
Our analysis and results generalize to random worker lifetimes that are i.i.d. across workers of different types,
with mean N and any distribution such that the lifetime exceeds N/polylog(n) with high probability.
4
5
types (I 0 , J 0 ), such that the total worker arrival rate of I 0 exactly matches the total job arrival
rate of J 0 . Formally,
X
X
ρ(i) 6=
µ(j) ∀I 0 ⊆ I, J 0 ⊆ J , I 0 6= φ .
i∈I 0
j∈J 0
The generalized imbalance condition holds generically.7 Note that this condition does not depend
on the matrix A.
Worker history. To define the state of the system and the resulting matching dynamics,
we need the notion of a worker history. A worker history is a tuple Hk = ((j1 , x1 ), . . . , (jk , xk )),
where jm is the job type this worker was matched to at her m-th time step in the system, for
1 ≤ m ≤ k; and xm ∈ {0, 1} is the corresponding reward obtained. Note that since workers live
for N jobs, the histories will have k = 0, . . . , N − 1. We let φ denote the empty history (for
k = 0).
3.2
System dynamics
Our goal is to model the following process. The operator observes, at any point in time, the
distribution of histories of workers in the platform, and also knows the job arrival rate µ. The
matching policy of the platform amounts to determining what mass of workers of each type of
history will be matched to which type of jobs. Ultimately, for this process to generate high
payoffs over time, the platform must choose jobs to learn worker types in order to optimize
payoffs.
With this intuition in mind, we now give a formal specification of our system dynamics.
System profile. A system profile ν is a joint measure over worker histories and worker types;
i.e., ν(Hk , i) is the mass of workers in the system with history Hk and type i. The evolution of
the system is a discrete-time dynamical system ν0 , ν1 , ν2 , . . ., where each νt is a system profile.8
Matching policy. To describe the dynamics we assume that the platform uses a matching
policy to match the entire mass of workers to jobs in each time step (we think of unmatched
workers as being matched to the empty job type κ). We assume that any mass of jobs left
unmatched in a given period disappears at the end of that period (our results do not depend on
this assumption).
Suppose that the system starts at time t = 0 with no workers in the system before this time.9
A matching policy π0 , π1 , . . . for the system specifies, at each time t, given a system profile νt ,
the mass of workers with each history that is matched to jobs of each type. In particular, let
πt (Hk , j|νt ) denote the fraction of workers
with history Hk matched to jobs of type j at time
P
t, given a system profile νt . (Thus j πt (Hk , j|νt ) = 1 for all t, Hk , and νt .) Note that the
matching policy acts on each worker’s history, not on the true type of each worker: this is because
the platform is assumed to not know worker types, except as learned through the history itself.
Dynamics. These features completely determine the evolution of the system profile {νt }.
Observe that νt (Hk , i)πt (Hk , j|νt ) is the total mass of workers of type i with history Hk who are
|I|+|J |
7
The set (ρ̂, µ) for which the condition holds is open and dense in R++
, where R++ are the strictly positive
real numbers.
8
The platform cannot directly observePthe system profile,
but can infer it. The platform observes the mass
of workers with each possible history
ν(H
,
i)
.
It
can then infer ν(Hk , i)’s individually by using
k
i∈I
Hk
knowledge of arrival rates ρ̂(i)’s, and the A matrix (which allows it to calculate the likelihood of seeing the
sequence of outcomes in Hk under the worker type i), together with Bayes’ rule.
9
In what follows we ultimately consider a steady-state analysis of the dynamical system, and initial conditions
will be irrelevant as long as the initial mass of workers is bounded.
6
matched to jobs of type j at time t, given policy πt and system profile νt . For all i, j, and t, we
have
νt+1 (φ, i) = ρ̂(i);
(1)
νt+1 ((Hk , (j, 1)), i) = νt (Hk , i)πt (Hk , j|νt )A(i, j),
k = 0, . . . , N − 2;
νt+1 ((Hk , (j, 0)), i) = νt (Hk , i)πt (Hk , j|νt )(1 − A(i, j)), k = 0, . . . , N − 2.
(2)
(3)
Decentralization through worker-history-only (WHO) policies. Note that, in general, policies may be time-varying, and may have complex dependence on the system profile νt .
We consider a much simpler class of policies that we call worker-history-only (WHO) policies.
These are policies where there exists a π such that
πt (Hk , j|νt ) = π(Hk , j).
In other words, in a WHO policy, the fraction of workers with history Hk who are matched to
jobs of type j does not depend on either time or on the full system profile. Thus WHO policies
are decentralized.
An obvious concern at this point is that a policy cannot allocate more jobs of type j than
there are. We formalize this capacity constraint in (8) below: in particular, a WHO policy does
not exceed the capacity of any job type in any period if and only if it satisfies (8).
Let ΠN denote the class of WHO policies, for a given N . In Section D.1 in Appendix D, we
establish that it suffices to restrict attention to policies in ΠN that satisfy (8).
Remark 1. For any feasible policy, there exists a WHO policy satisfying capacity constraints
that achieves a payoff accumulation rate arbitrarily close to that of the former policy. In particular, WHO policies satisfying capacity constraints suffice to achieve the highest possible payoff
accumulation rate.
Steady state of a WHO policy π. First, suppose that there are no capacity constraints,
and consider the system dynamics (1)–(3), assuming the system initially starts empty. The
dynamics (1)–(3) yields a unique steady state that can be inductively computed for k = 0, 1, . . .:
νπ (φ, i) = ρ̂(i);
(4)
νπ ((Hk , (j, 1)), i) = νπ (Hk , i)π(Hk , j)A(i, j), k = 0, . . . , N − 2;
(5)
νπ ((Hk , (j, 0)), i) = νπ (Hk , i)π(Hk , j)(1 − A(i, j)), k = 0, . . . , N − 2.
(6)
We refer to the measure νπ as the steady state induced by the policy π.
Routing matrix of a WHO policy π. If the system is in steady state, then at any time
period, π induces a steady-state fraction
P xπ (i, j) of the mass
P of workers of type i that are assigned
ν (H,i)π(H,j)
ν (H,i)π(H,j)
to type j jobs. We have xπ (i, j) = HPπ νπ (H,i)
= H π ρ(i)
. We call {xπ (i, j)}I×J
H
the routing matrix achieved by the policy π. This is a (row) stochastic matrix; i.e., each row
sums to 1. Observe that the mass of demand for jobs of type j from workers of type i in any
time
P period is ρ(i)xπ (i, j), and the total mass of demand for jobs of type j in any time period
is i∈I ρ(i)xπ (i, j).
Let X N = xπ : π ∈ ΠN ⊆ [0, 1]|I|×|J | be the set of routing matrices achievable (when
each worker does N jobs) by WHO policies. (Again, we note that capacity constraints are
ignored in the definition of X N .) In Appendix D, we show that X N is a convex polytope (see
Proposition D.4).
7
3.3
The optimization problem
Our paper focuses on maximization of the steady-state rate of payoff accumulation, subject to
the capacity constraints. This leads to the following optimization problem:
X
X
maximize W N (π) ,
ρ(i)
xπ (i, j)A(i, j)
(7)
i∈I
subject to
X
j∈J
ρ(i)xπ (i, j) ≤ µ(j)
∀j ∈ J ;
(8)
i∈I
xπ ∈ X N .
(9)
The objective is the steady-state rate of payoff accumulation per time period, expressed in terms
of the routing matrix induced by a (WHO) policy π. The constraint is the capacity constraint:
the system will be stable if and only if the total “demand” for jobs of type j is not greater than
the arrival rate of jobs of type j.
Since X N is a convex polytope, this is a linear program, albeit a complex one. The complexity
of this problem is hidden in the complexity of the set X N , which includes all possible routing
matrices that can be obtained using WHO policies. The remainder of our paper is devoted to
solving this problem and characterizing its value, by considering an asymptotic regime where
N → ∞.
3.4
The benchmark: Known worker types
We evaluate our performance relative to a natural benchmark: the maximal rate of payoff accumulation possible if worker types are perfectly known upon arrival. In this case, any stochastic
matrix is feasible as a routing matrix. Let D denote the set of all stochastic matrices:
X
|I|×|J |
D= x∈R
: x(i, j) ≥ 0;
x(i, j) = 1 .
(10)
j∈J
Note that any routing matrix in D is implementable by a simple policy under known worker
types: given a desired routing matrix x, route a fraction x(i, j) of workers of type i to jobs of
type j.
Thus, with known worker types, the maximal rate of payoff accumulation is given by the
solution to the following optimization problem:
X
X
maximize
ρ(i)
x(i, j)A(i, j)
(11)
i∈I
subject to
X
j∈J
ρ(i)x(i, j) ≤ µ(j)
∀j ∈ J ;
(12)
i∈I
x ∈ D.
(13)
We let V ∗ denote the maximal value of the preceding optimization problem, and let x∗ denote
the solution. This linear program is a special case of the “static planning problem” that arises
frequently in the operations literature (see, e.g. [8]). The problem can also be viewed as a
version of the assignment problem due to Shapley and Shubik [35], in which the resources are
divisible.
8
3.5
Regret
We evaluate the performance of a given policy in terms of its regret relative to V ∗ . In particular,
given N and a WHO policy π satisfying (8), we define the regret of π as V ∗ − W N (π).
We focus on the asymptotic regime where N → ∞, and try to find policies that have “small”
regret in this regime. This asymptotic regime allows us to identify structural aspects of policies
that perform well. In Appendix D (see Proposition D.3), we show that it is relatively “easy” to
design policies that achieve a vanishing regret (and even regret that is within a constant factor
of the smallest possible). The idea is straightforward: informally, when N is large, policies that
“explore” for a vanishing fraction of worker lifetimes will be able to learn the worker’s true type
sufficiently well to yield a rate of payoff accumulation such that regret converges to zero in the
limit.
For this reason, our analysis focuses on a more refined notion of asymptotic optimality. In
particular, we focus on developing policies that achieve a nearly optimal rate at which the regret
V ∗ − W N (πN ) approaches zero. This is formalized in Theorem 5.1 below.
3.6
A note on terminology
Note that, intuitively, WHO policies have the feature that decisions are taken on the basis of
the history of a given worker, not on the basis of the system profile as a whole. In the sequel, we
will typically refer to π(Hk , j) as “the probability that a worker of history Hk is matched to a job
of type j.” We use this terminology to make the presentation more intuitive, since the intention
is that our algorithms be implemented at the level of each individual worker’s history. However,
to formalize all our arguments, we emphasize that our proofs translate π(Hk , j) as the fraction
of workers of history Hk matched to a job type j; this correspondence applies throughout the
technical development.
4
Decentralized Explore-then-Exploit for Matching (DEEM): A
payoff-maximizing policy
In this section we present the design of a sequence of policies
∗ that achieves a nearly optimal rate of convergence of reπN
∗ ). We refer to our policy design as DEEM:
gret V ∗ − W N (πN
Decentralized Explore-then-Exploit for Matching. Our main
result, stated in the next section, is Theorem 5.1: there we
exactly quantify the regret performance of DEEM (an upper
bound on its regret), and characterize it as nearly optimal
(a lower bound on the regret of any feasible WHO policy).
To begin to understand the challenges involved, consider
the example in Figure 1. In this example, there are two types
of workers: “novice” and “expert,” with a mass of ρ = 0.5 of
each present in steady state. There are two types of jobs:
“easy” and “hard,” each arriving at rate 0.6.
Jobs
0.6
𝜇
0.6
Workers
Easy
Hard
0.5 Expert
0.9
0.8
0.9
0.1
𝜌
0.5 Novice
Figure 1: An example
We make several observations regarding this example that inform our subsequent work.
1. The benchmark. In this example, the optimal solution to the benchmark problem (11)–
9
(13) with known types routes all novices to easy jobs, a mass 0.1 of experts to easy jobs, and a
mass 0.4 of experts to hard jobs. Of course, our problem is that we do not know worker types
on arrival.
2. Capacity constraints affect an optimal WHO policy’s need to learn. If easy and hard jobs
are in infinite supply, then the WHO policy π that matches all workers to easy jobs is optimal.
However, with the finite supply of available easy jobs, some workers must do hard jobs. But
which workers?
Clearly, for payoff optimality, an optimal policy should aim to match experts to hard jobs.
But this is only possible if it first learns that a worker is an expert. Because of the structure of
A, the type of a worker can only be learnt by matching it to hard jobs; those who perform well
on these jobs are experts, and those who fail are novices.
3. Minimizing regret requires learning up front. Assigning workers of unknown type to hard
jobs necessarily incurs regret relative to the benchmark. Indeed, novices unknowingly matched
to hard jobs lead to a regret of 0.8 per unit mass of such workers in each period. Minimizing this
regret therefore requires that the algorithm not only learn worker types, but also do so relatively
early in their lifetime, so that workers identified as experts can be assigned many hard jobs.
In our work, this leads to a structure where we separate our policy into exploration and
exploitation phases: the policy first tries to learn a worker’s type, and then “exploits” by assigning
this worker to jobs while assuming that the learned type is correct. The exploration phase will
be of length O(log N ), which is short relative to the worker’s lifetime.
4. Some mistakes in the exploration phase are worse than others. There are two kinds of
mistakes that the policy can make while learning: it can mistakenly identify novices as experts,
and it can mistakenly identify experts and novices. These mistakes differ in their impact on
regret.
Suppose that at the end of the exploration phase, the algorithm misclassifies a novice as
an expert. This has a dire impact on regret: the novice is then assigned to hard jobs in the
exploitation phase, and as noted above, this incurs a regret of 0.8 per unit mass (of workers
misclassified this way) per unit time. Thus we must work hard in the exploration phase to avoid
such errors.
On the other hand, suppose that at the end of the exploration phase, the algorithm misclassifies an expert as a novice. This mistake is far less consequential: workers misclassified in this
way will be assigned to easy jobs. But a mass 0.1 of experts must be assigned to easy jobs even
in the benchmark solution with known types. Therefore, as long as this misclassified mass is not
too large, we can adjust for it in the exploitation phase.
This discussion highlights the need to precisely identify the learning goals of the algorithm:
to minimize regret, how strongly does each worker type need to be distinguished from others?
A major contribution of our work is to demonstrate an optimal construction of learning goals
for regret minimization. As noted above, the capacity constraints fundamentally influence the
learning goals of the algorithm.
In the remainder of the section, we describe key ideas behind the construction of our policy,
highlighted by the issues raised in the preceding example. We formally describe DEEM in
Section 4.4. We state our main theorem in Section 5.
10
4.1
Key idea 1: Use shadow prices as an “externality adjustment” to payoffs
We begin by first noticing an immediate difficulty that arises in using WHO policies in the
presence of capacity constraints. WHO policies are decentralized, i.e., they act only on the
history of the worker; as such, they cannot use aggregate state information about the system,
that conveys whether capacity constraints are being met or not. In order to solve (7)–(9),
therefore, we need to find a way to “adjust” for capacity constraints despite the fact that our
policy acts only at the level of worker histories.
Our key insight is to use shadow prices for the capacity constraints to adjust payoffs; we then
measure regret with respect to these adjusted payoffs. Recall that (7)–(9) is a linear program.
Let pN be the optimal shadow prices (dual variables) for the capacity constraints (8). Then by
standard duality results, it follows that the policy that is optimal for (7)–(9) is also optimal for
the following unconstrained optimization problem:
X
X
maximizexπ ∈X N
ρ(i)
xπ (i, j)(A(i, j) − pN (j)).
(14)
i∈I
j∈J
Thus one may attempt to account for capacity constraints using shadow prices10 pN (j).
The challenge here is that the set X N is quite complex, and thus characterizing the optimal
shadow prices of (7)–(9) is not a reasonable path forward. Instead, we use the optimal shadow
prices in the benchmark linear program with known types (11)–(13) to adjust payoffs; we then
measure regret with respect to these adjusted payoffs (the practical heuristic we implement uses
a different, instance-independent approach to estimate shadow prices; see Section 7).
We let p∗ denote the vector of optimal shadow prices for the capacity constraint (12) in the
problem with known types (11)–(13). Using the generalized imbalance condition, we show that
these prices are uniquely determined; see Proposition D.2 in Appendix D.
Although p∗ (j) 6= pN (j), for large N , the platform should be able to learn the type of a
worker type early in her lifetime, leading to small |p∗ (j) − pN (j)|. This motivates an analog of
(14):
X
X
maximizexπ ∈X N
ρ(i)
xπ (i, j)(A(i, j) − p∗ (j)).
(15)
i∈I
j∈J
We develop a near-optimal algorithm for problem (15), such that (1) constraints on job
capacities are not violated, and (2) complementary slackness conditions are satisfied, i.e., if
p∗ (j) > 0, then the job type j is fully utilized. We then show this leads to the upper bound in
the main result.
4.2
Key idea 2: Meet required learning goals while minimizing regret
As noted in our discussion of the example in Figure 1, we must carefully define the learning
goals of the algorithm: which worker types need to be distinguished from which others, and
with what level of confidence? A key contribution of our work is to formalize the learning goals
of our algorithm. In this section we define the learning goals of the algorithm, and outline the
exploration phase that meets these goals.
Let the set of optimal job types for worker type i be defined by J (i) = arg maxj∈J A(i, j) −
(A standard duality argument demonstrates that in any optimal solution of the benchmark
(11)–(13), a worker type i is assigned only to jobs in J (i).)
p∗ (j).
10
Further effort is needed to ensure the policy does not violate capacity constraints, and that complementary
slackness holds.
11
Recall that in the example in Figure 1, it is far more important not to misclassify a novice as
an expert than to misclassify an expert as a novice. We formalize this distinction through the
following definition.
Definition 1. We say that a type i needs to be strongly distinguished from a type i0 if J (i) \
J (i0 ) 6= ∅. For each worker type i, let Str(i) be the set of all types i0 from which i needs to be
strongly distinguished, i.e., Str(i) = {i0 : J (i) \ J (i0 ) 6= ∅}.
In words, this means that i needs to be strongly distinguished from i0 if it has at least one
optimal job type that is not optimal for i0 , whereas it needs to be only weakly distinguished
from i0 if all optimal job types for i are also optimal for i0 . This definition is most easily
understood through the example in Figure 1 and our subsequent discussion. In particular, note
that for that example, the benchmark shadow prices are p∗ (easy) = 0.1 and p∗ (hard) = 0, and
thus J (novice) = {easy}, while J (expert) = {easy, hard}. Thus experts need to be strongly
distinguished from novices, since hard jobs are optimal for experts but not for novices; on the
other hand, novices need to be only weakly distinguished from experts, since easy jobs are
optimal for experts as well.
In the exploration phase of our algorithm, our goal is to classify a worker’s type as quickly as
possible: the preceding definition is what we use to formalize the learning goals in this phase.
In particular, consider making an error where the true type is i0 , but we misclassify it as i. If
i0 is not in Str(i), any probability of an error of o(1) for such misclassification error is tolerable
as N grows large (as in the example in Figure 1). We choose Θ(1/ log N ) as the target error
probability for this kind of error. On the other hand, for any i0 ∈ Str(i), the optimal target error
probability is much smaller. In particular, the optimal target error probability can be shown
to be approximately 1/N : if we choose a larger target, we will incur a relatively large expected
regret during exploitation due to misclassification, if we choose a smaller target, the exploration
phase is unnecessarily long, and we thus incur a relatively large regret in the exploration phase.
With the learning goals defined, the exploration phase of DEEM operates in one of two
subphases: either “guessing” or “confirmation,” as follows. At every job allocation opportunity,
we check whether the posterior probability of the maximum a posteriori (MAP) estimate of the
worker type is sufficiently high. If this probability is low, we say the policy is in the “guessing”
subphase of the exploration phase, and a job type is chosen at random for the next match. On
the other hand, if it is high (in particular, greater than log N times the posterior probability
of any other worker type), then we say that the policy is in the “confirmation” subphase of the
exploration phase: in this regime, the policy works to confirm the MAP estimate. Specifically,
in the confirmation subphase, the policy focuses only on strongly distinguishing the MAP from
all other types in Str(i); the trade-off is that this must be done with minimum regret. We frame
this as an optimization problem (see (16) below): essentially, the goal is to find a distribution
over job types that minimizes the expected regret until the confirmation goals are met. In the
confirmation subphase, the policy allocates the worker to jobs according to this distribution,
until the type is confirmed.
We conclude by briefly explaining the role of the guessing phase in minimizing regret. Informally, guessing is necessary so that confirmation minimizes regret for the correct worker type
with high probability. In particular, suppose that there are two worker types i and i0 that have
the same optimal job types, i.e., J (i) = J (i0 ), and with A(i, j) = A(i0 , j) for all j ∈ J (i).
In this case, payoff maximization does not require distinguishing i from i0 . Nevertheless, it is
possible that the confirmation policies for i and i0 differ, without necessarily distinguishing i
from i0 . In this case, i first needs to be distinguished from i0 with probability of error o(1) to
achieve optimal regret to leading order. Concretely, if there is no guessing phase and the MAP is
12
i early in the worker’s lifetime, the policy will never discover its mistake and ultimately confirm
using the wrong policy, incurring an additional (leading order) regret of Θ(log N/N ).
4.3
Key idea 3: Optimally allocate in the exploitation phase while meeting
capacity constraints
When the algorithm completes the exploration phase, it enters the exploitation phase; in this
phase, the algorithm aims to match a worker to jobs that maximize the rate of payoff generation,
given the confirmed type label. A naive approach would match a worker labeled type i to any job
type in J (i), since these are the optimal job types for worker type i after externality adjustment.
This approach turns out to fail spectacularly and generically leads to Ω(1) regret (this occurs
for any set of fixed shadow prices). To see why, we need the following fact.
Fact 1. Under generalized imbalance, as long as there is at least one capacity constraint that
is binding in some optimal solution x∗ to the benchmark problem (11)–(13) with known types,
there is at least one worker i such that x∗ (i, ·) is supported on multiple job types.
This fact implies that appropriate tie-breaking between multiple optimal job types is necessary
during exploitation for one or more worker types in order to achieve vanishing regret.
In order to implement appropriate tie-breaking, suppose that we assign jobs during the exploitation phase using the routing matrix x∗ that solves the benchmark problem (11)–(13); in
this case, each worker with confirmed type i is matched to job type j with probability x∗ (i, j).
However, this naive approach needs further modification to overcome two issues. First, some
capacity is being used in the exploration phase and the effective routing matrix during the exploration phase does not match x∗ . Second, the exploration phase can end with an incorrectly
classified worker type.
Our policy in the exploitation phase chooses a routing matrix y ∗ that resembles x∗ , but
addresses the two concerns raised in the preceding paragraph. Crucially, the chosen y ∗ should
ensure that only job types in J (i) are assigned with positive probability, and satisfy the complementary slackness conditions. We show (in Proposition A.5, using Fact 1) that such a y ∗ indeed
exists for an N large enough under the generalized imbalance condition, and we show how to
compute it. Note that as y ∗ is a fixed routing matrix, it can be implemented in a decentralized
manner.
We comment here that y ∗ is largely a theoretical device used to obtain the provable regret
optimality of our policy. In our implementation of DEEM (see Section 7), we propose a far
simpler solution: we use dynamically updated shadow prices to automatically achieve appropriate
tie-breaking. The shadow prices respond in a “tâtonnement” manner based on the currently
available supply of different job types: the price of job type j rises when the available supply
falls. In particular, fluctuations in these shadow prices naturally lead to the necessary tiebreaking for efficient exploitation.
4.4
Formal definition of DEEM
∗ , based on the discussion above.
In this section we provide a formal definition of the policy πN
First, for each i define the maximal externality-adjusted utility U (i) = maxj∈J A(i, j)−p∗ (j).
13
Then choose α(i) such that:
αj U (i) − [A(i, j) − p∗ (j)]
P
,
mini0 ∈Str(i) j∈J αj KL(i, i0 |j)
P
α(i) ∈ A(i) = arg min
α∈∆(J )
j∈J
(16)
where KL(i, i0 |j) is the Kullback–Leibler divergence11 between the distributions Bernoulli(A(i, j))
and Bernoulli(A(i0 , j)), and ∆(J ) is the set of distributions over J . The idea is that sampling
job types from α(i) allows the policy to distinguish i simultaneously from all i0 ∈ Str(i), while
incurring the smallest possible externality-adjusted regret. In Appendix B, we show that (16)
can be written as a small linear program. If the optimization problem in (16) has multiple
solutions, we pick the one that has the largest denominator (and hence the largest numerator
as well), thus maximizing learning rate subject to optimality; i.e., we choose
X
α(i) ∈ arg max min
αj KL(i, i0 |j).
(17)
0
α∈A(i) i ∈Str(i) j∈J
We discuss details in the Appendix B.
For m = 1, · · · , N , let the job type chosen at opportunity m be jm and the outcome be
Xm . For any i ∈ I and j ∈ J , let l(X, i, j) = A(i, j)1{X=1} + (1 − A(i, j))1{X=0} . Define
Q
λ0 (i) = 1, and for k ≥ 1, let λk (i) = km=1 l(Xm , i, jm ) denote the likelihood of the observed
history until the k-th job under worker type i. Let MAPk = arg maxi∈I ρ̂(i)λk (i) be the MAP
estimate based on the history, and define Λk (i, i0 ) = ρ̂(i)λk (i)/ρ̂(i)λk (i0 ), i.e., the ratio of the
posterior probabilities of type i and i0 . For convenience, we refer to Λ0 (i, i0 ) as the prior odds of
i relative to i0 , and Λk (i, i0 ) as the posterior odds of i relative to i0 after k jobs.
DEEM is defined as follows.
1. Phase 1: Exploration. Suppose that i = MAPk .
(a) Guessing subphase. If mini0 6=i Λk (i, i0 ) < log N , choose the next job type uniformly at
random in J .
(b) Confirmation subphase to strongly distinguish i from types in Str(i). If we
have
mini0 6=i Λk (i, i0 ) ≥ log N but mini0 ∈Str(i) Λk (i, i0 ) < N , draw the next job type i.i.d. from the
distribution α(i).
(c) Exit condition for the exploration phase. If mini0 6=i Λk (i, i0 ) ≥ log N and
mini0 ∈Str(i) Λk (i, i0 ) ≥ N , then the worker is labeled as being of type i and the policy moves to
the exploitation phase. (The worker is never returned to the exploration phase.)
2. Phase 2: Exploitation. For every job opportunity, for a worker confirmed to be of
type i, choose a job in J (i) with probability y ∗ (i, j), where y ∗ is a routing matrix (specified in
Proposition A.5 in Appendix A) such that system capacity constraints are not violated in steady
state.
5
Main result
Our main result is the following theorem. In particular, we prove a lower bound on the regret of
∗ constructed in the preceding section (essentially)
any policy, and show the sequence of policies πN
achieves this lower bound.
The KL divergence between a Bernoulli(q) and a Bernoulli(q 0 ) distribution is defined as q log
1−q
q) log 1−q
0.
11
14
q
q0
+ (1 −
Theorem 5.1. Fix (ρ, µ, A) such that: (a) no two rows of A are identical; and (b) the generalized
imbalance condition holds. Then there is a constant C = C(ρ, µ, A) ∈ [0, ∞) such that
1. (Lower bound) For any N and any WHO policy π that is feasible for (7)–(9),
V ∗ − W N (xπ ) ≥
C log N
1 + o(1) and
N
(18)
∗ is feasible for (7)–(9) for each N , with:
2. (Upper bound) The sequence of policies πN
∗
V ∗ − W N (πN
)≤
log log N
C log N
1 + o(1) + O
.
N
N
(19)
The constant C that appears in the theorem depends on the primitives of the problem, i.e.,
(ρ, µ, A); it is defined as follows:
P
∗
X
j∈J αj U (i) − [A(i, j) − p (j)]
P
C(i) = min
;
C
=
ρ(i)C(i) .
(20)
mini0 ∈Str(i) j∈J αj KL(i, i0 |j)
α∈∆(J )
i∈I
(Note that C(i) captures the regret per unit mass of service opportunities from workers of type
i.) Informally, instances in which there is a conflict between exploration (i.e., learning worker
type) and exploitation (i.e., maximizing short-term payoffs) have larger values of C. The case
C = 0 corresponds to instances where the goals of learning and regret minimization are aligned;
i.e., learning does not require regret of Ω(log N/N ). In this case, our result establishes that our
chosen policies are nearly asymptotically optimal, to within O(log log N/N ). On the other hand,
the instances with C > 0 are those instances with a non-trivial tension between learning and
∗ )
short-term payoffs. For these instances, our result establishes that our chosen policies (πN
N ≥1
achieve asymptotically optimal regret upto leading order in N .
The constant C is best understood in terms of the definition
of α in the exploration phase (cf. (16)). Note
Jobs
𝜇
0.6
0.6
that for a fixed α, for workers of true type i, the smallest
Easy
Workers
Hard
0
value of the log posterior odds, mini0 ∈Str(i)
P log Λn (i, i ), 0increases at an expected rate of mini0 ∈Str(i) j∈J αj KL(i, i |j)
0.5 Expert
0.6
0.8
during confirmation. Thus, when N is large, the time
𝜌
taken to confirm i against worker
types in Str(i) is apP
proximately log N/(mini0 ∈Str(i) j∈J αj KL(i, i0 |j)). Hence,
0.6
0.1
0.5 Novice
the externality-adjusted regret incurred until confirmation is complete, per P
unit mass of workers of
type i, is approximately
j∈J αj U (i) − [A(i, j) −
P
Figure 2: An example where ∗
p (j)] log N/(mini0 ∈Str(i) j∈J αj KL(i, i0 |j)). Optimizing
Ω(log N/N ) regret is unavoidable.
over α results in an expected regret of nearly C(i) log N
that must be incurred until the strong distinguishing goals
are met for a unit mass of workers of type i. This translates to an expected regret of nearly
ρ̂(i)C(i) log N = ρ(i)C(i) log N/N owing to workers of type i per time unit. This reasoning
forms the basis of our lower bound, formalized in Proposition A.1 in Appendix A.
Now, a regret of Θ(log N/N ) is unavoidable when C(i) > 0 for some i. To develop some
intuition for this case, consider the same example as before, but with a modified payoff matrix
shown in Figure 2. It can be shown that in this case, a regret of Ω(log N/N ) is unavoidable
in the event that the true type of the worker is novice [1]. The problem is the following: to
distinguish novices from experts, the policy must allocate workers to hard jobs. But hard jobs
15
are strictly suboptimal for novices, and so if the true type of the worker is novice, some regret
is unavoidable.
In particular, to develop intuition for the magnitude of this regret, imagine a policy that
assigns workers hard jobs for the first k steps (leading to Θ(k) absolute regret per unit mass
of workers), and based on the realized payoffs, estimates the worker type (with confidence
1 − exp(−Θ(k))). If the worker is estimated to be a novice, the policy can choose to assign only
easy jobs to the worker. However, this means that there will be no further learning about the
worker type, and the expected contribution to absolute regret is about N times the probability
that the worker is truly an expert, i.e., N exp(−Θ(k)), per unit mass of workers. Combining,
we see the total absolute regret is at least Θ(k) + N exp(−Θ(k)) ≥ Θ(log N ) over the lifetime
of a unit mass of workers. (Here k = Θ(log N ) is needed to achieve Θ(log N ) absolute regret.)
We then divide by N to obtain the regret per unit mass of service opportunities by workers.
This discussion motivates the following definition.
Definition 2. Consider a worker type i. Suppose that there exists another type i0 such that
A(i, j) = A(i0 , j) for all j ∈ J (i) and J (i) ∩ J (i0 ) = φ. Then we say that the ordered pair (i, i0 )
is a difficult type pair.
A similar definition also appears in [1]; the modification here is that the sets J (i) are defined
with respect to externality-adjusted payoffs to account for capacity constraints.
The constant C(i) > 0 if and only if there is some other i0 such that (i, i0 ) is a difficult type
pair. In general, if: (1) none of the job types in J (i) allow us to distinguish between i and i0 ;
and (2) all the jobs in J (i) are strictly suboptimal for i0 , then any policy that achieves small
regret must distinguish between i and i0 and must assign the worker to jobs outside J (i) to
make this distinction. This leads to a regret of Ω(log N ) per unit mass of workers of type i (over
the lifetime of the workers).
On the other hand, if there is no difficult type pair, then there is no conflict between learning
and regret minimization. Here, one can show that C(i) = 0 for each i, and this value is attained
by some distribution α(i) that is supported on J (i). To see this note that if α is fully supported
on J (i) (i.e., αj > 0 for all j ∈ J (i)), then the numerator is 0; however if there is no type
i0 such that (i, i0 ) is a difficult type pair, then the denominator is strictly positive, and thus
C(i) = 0. In this case, C = 0 and our main result says that our algorithm achieves a regret of
O(log log N/N ) asymptotically.12 This regret basically results from the uniform sampling of the
job types during the guessing phase, which accounts for O(log log N/N ) fraction of the lifetime
of the worker.
6
Proof sketch
The proof of Theorem 5.1 can be found in Appendix A. Here we present a sketch. The critical
ingredient in the proof is the following relaxed optimization problem in which there are no
capacity constraints, but capacity violations are charged with non-negative prices p∗ from the
optimization problem (11) with known worker types.
hX
i
X
X
X
WpN∗ = max
ρ(i)
x(i, j)A(i, j) −
p∗ (j)
ρ(i)x(i, j) − µ(j) .
(21)
x∈X N
i∈I
j∈J
j∈J
12
i∈I
In fact, our proof demonstrates that this regret can be brought down to any O(fN /N ) such that fN = o(1)
by choosing a different threshold in the guessing phase.
16
Lower bound on regret. If C > 0 (i.e., if there is at least one difficult pair of worker
types; cf. Section 5), there is an upper bound on the performance of any policy in this problem,
expressed relative to V ∗ . This result follows directly from [1]:
WpN∗ ≤ V ∗ −
C log N
(1 + o(1)),
N
where C ≥ 0 is precisely the constant appearing in (20). By a standard duality argument,
we know that W N ≤ WpN∗ , and hence this bound holds for W N as well (see Proposition A.1),
yielding the lower bound on regret on our original problem (7).
∗ is feasible for problem
Upper bound on regret. There are two key steps in proving that πN
∗ ) ≥ V ∗ − C(log N/N )(1 + o(1)).
(7)–(9), and W N (πN
∗ , with an arbitrary exploitation-phase routing matrix
1. First, we show that our policy πN
supported on J (i) for each i ∈ I, achieves near optimal performance for the single multi-armed
bandit problem (21). Formally, if (with some abuse of notation) we let WpN∗ (π) denote the value
attained by a policy π in problem (21), i.e.,
WpN∗ (π) =
X
i∈I
ρ(i)
X
xπ (i, j)A(i, j) −
j∈J
X
j∈J
hX
i
ρ(i)xπ (i, j) − µ(j) ,
p∗ (j)
i∈I
then
∗
V ∗ − WpN∗ (πN
) = C(log N/N )(1 + o(1)) + O(log log N/N ).
∗ ) ≥ W N − o(log N/N ), i.e., π ∗ is
This is shown in13 Proposition A.2. Thus we have WpN∗ (πN
p∗
N
near-optimal in problem (21).
2. In the next part of the proof, we show that we can design a routing matrix y ∗ (that
∗ , such that the following conditions
depends on N ) for the exploitation phase of the policy πN
are satisfied:
P
∗ (i, j) − µ(j) = 0 for all j such that p∗ (j) > 0,
(a) (Complementary slackness) i∈I ρ(i)xπN
and,
P
∗ (i, j) − µ(j) ≤ 0 for all other j ∈ J .
(b) (Feasibility) i∈I ρ(i)xπN
∗ with this choice of y ∗ in the exploitation
This is shown in Proposition A.5. We deduce that πN
phase is feasible for problem (7)–(9) and the complementarity slackness property implies that
∗ ) = W N (π ∗ ), yielding our upper bound on regret.
W N (πN
p∗
N
∗ , the correct label of the worker
Construction of y ∗ . At the end of the exploration phase of πN
is learned with a confidence of at least (1 − o(1)). This fact, coupled with the generalized
imbalance condition (leading to flexibility in modifying x∗ ; cf. Fact 1), is sufficient to ensure an
appropriate and feasible choice of y ∗ = x∗ + o(1) will correct the deviations from x∗ in terms
of capacity utilizations of job types with p∗ (j) > 0 arising because of the (short) exploration
phase, and because of the infrequent cases in which exploitation is based on an incorrect worker
label coming out of the exploration phase.
13
[1] proves this result for a similar policy.
17
7
Practical considerations and a heuristic
Our theoretical analysis of DEEM focused on an asymptotic regime where N → ∞. In this
section, we focus on a number of practical considerations that arise when considering implementation of a policy like DEEM. First, we discuss a practical approach to managing capacity
constraints via dynamic queue-based shadow prices. Second, we discuss two modifications to the
algorithm that improve performance when N is finite, and suggest a modified heuristic that we
call DEEM+ . In the next section, we simulate both DEEM and DEEM+ , and evaluate their
performance.
7.1
Dynamic queue-based shadow prices
A key step in making DEEM practical is to use dynamic shadow prices based on supply-demand
imbalances in the market. In our mathematical model, we had assumed that the masses of new
workers and jobs arrive instantaneously at the beginning of each period, after which they are
instantaneously matched, and further, each job either gets matched immediately on arrival or
disappears at the end of the period. However, in real platforms, the arrivals, departures, and
matchings of workers and jobs occur sequentially in continuous time. In these settings, it is
common for platforms to maintain a “queue” of jobs of each type that grows when new jobs
arrive (in continuous time) and shrinks when existing jobs are matched. In this scenario, the
queue length at any time can be leveraged to compute an instantaneous shadow price for the
job type that can be utilized for externality adjustment of the payoffs.
A reasonable approach is to set the shadow price on each job type via a decreasing function
of the corresponding queue length. One natural way to do this is as follows: assume that in
practice, the arriving jobs accumulate in queues for the different types, each with a finite capacity
B (if the capacity is exceeded, jobs are lost). If the queue length of job type j at any instant
is q(j), we set the price of j at that instant to pq (j) = (B − q(j))/B (thus the price lies in
[0, 1]). Note that pq (j) changes every time a job is assigned to a worker, or a new job arrives,
because the queue length changes. We implement a discrete-time analog of this approach in our
simulated marketplace in the next section.
Computing the prices in this fashion obviates the need to explicitly compute y ∗ in the exploitation phase of our policy; instead the exploitation phase can be implemented by allocating
optimally for each worker given the current queue-based prices (still a fully decentralized solution). The natural fluctuation in these prices ensures appropriate tie-breaking in allocation
(cf. Fact 1).
These queue-based prices can be incorporated in the implementation of DEEM in the following way.
1. Modifying the exploration and exploitation phases. While computing α(i) for each
i ∈ I in the confirmation phase of DEEM, replace p∗ (j) by the instantaneous queue-based shadow
prices pq (j) in Eq. (16). Similarly, in the exploitation phase, instead of explicitly computing the
routing matrix y ∗ , use these prices to decide assignments in the following manner. Define the
sets J ∗ (i) as:
J ∗ (i) = arg max A(i, j) − pq (j).
j∈J
If an assignment has to be made in the exploitation phase for some worker who has already been
labeled as being of type i, then a job type j ∗ ∈ J ∗ (i) is chosen (note that typically, J ∗ (i) will
18
be a singleton).14
2. Price-smoothing for learning goals. The platform can also determine the strong
distinction requirements (see Definition 1), i.e., the set Str(i) for each i based on the sets J ∗ (i)
induced by the instantaneous prices defined above. But this approach suffers from the drawback
that random fluctuations of the shadow prices around their mean values can result in changes in
the sets J ∗ (i), and hence in the learning goals, which could be detrimental to the performance
of our policy. On the other hand these fluctuations are essential for appropriate tie-breaking
across multiple optimal job types in the exploitation phase.
Thus we propose the following modification: we utilize an average of recent prices within a
fixed recent window of time, and modify the definition of J ∗ (i) to incorporate a small tolerance
so that the set Str(i) remains unaffected by the fluctuations in the prices. To be precise, for a
window size W , let p̄q (i) be the unweighted average of the queue length based prices seen over
the past W epochs of changes in the price (again, note that pq (i) changes every time a job is
assigned to a worker, and also when new jobs arrive). Next, for a tolerance ε > 0, we define
Jε∗ (i) = {j ∈ J : A(i, j) − p̄q (j) ≥ max
[A(i, j 0 ) − p̄q (j 0 )] − ε}.
0
j
Then the set Str(i) for each i (see Definition 1), is defined based on Jε∗ (i).
7.2
Improving performance in the finite N regime
We propose two changes that improve performance in the finite N regime. First, recall that
if a worker type i has an optimal job type j that is not optimal for some worker type i0 , then
DEEM tries to achieve a probability 1/N of misclassifying a worker of type i0 as type i. For
small N , however, we can do better: the desired probability of error should explicitly depend
on how much regret type i0 incurs by performing job j – if this regret is very small, then it isn’t
worth trying to make this distinction with high precision.
In particular, for each i0 ∈ Str(i), define:
R(i, i0 ) ,
max
j∈J (i), j ∈J
/ (i0 )
U (i0 ) − [A(i0 , j) − p∗ (j)].
R(i, i0 ) is the highest regret incurred if type i0 is matched to a suboptimal job that is optimal
for type i. A reasonable approach is to aim for a probability of misclassification of i0 as i of
1/N R(i, i0 ) instead of 1/N , thus accounting for the fact that if R(i, i0 ) is small, then we can
tolerate a higher probability of error.
The second change we propose is to explicitly incorporate the posterior into the exploration
phase. Recall that in DEEM, we guess and then confirm in the exploration phase; and guessing
is not optimized, but rather involves exploration uniformly at random. When N is finite, we can
gain by instead leveraging the posterior at each round to appropriately allocate confirmation
effort across the different types, until the learning goals are met for some type i. In principle, this
approach can subsume the guessing and confirmation phases into a uniformly defined exploration
phase; the challenge is to precisely describe how the posterior is used to guide the exploration
phase.
14
In practice, we can continue to benefit from learning during the exploitation phase: instead of optimizing
the price-adjusted payoff for the confirmed worker label, we can optimize for the current MAP estimate, thus
accounting for the possibility that we may have confirmed incorrectly. Clearly, doing so can only improve
performance.
19
7.3
DEEM+ : A practical heuristic for finite N
In this subsection, we incorporate the two suggestions of the preceding subsection into a formal
heuristic we refer to as DEEM+ .
It will be convenient to define R(i, i0 ) = 1 for all i0 ∈
/ Str(i), and define Λ̂k (i, i0 ) as follows:
1. If i0 ∈ Str(i), then Λ̂k (i, i0 ) , Λk (i, i0 )/R(i, i0 ).
2. If i0 ∈
/ Str(i), then Λ̂k (i, i0 ) , Λk (i, i0 )/R(i, i0 ) if Λk (i, i0 )/R(i, i0 ) < log N , and Λ̂k (i, i0 ) , N
otherwise.
Next, after k matches, for each type i define:
Lk (i) , {i0 6= i : Λ̂k (i, i0 ) < N }.
Lk (i) is the set of types that i remains to be distinguished from after k opportunities. In case
the true worker type is i, effort should ideally be directed towards these distinctions in order to
speed up confirmation. Next, define the posterior probability of the worker being of type i after
opportunity k as:
ρ̂(i)λk (i)
gk (i) , P
.
0
0
i0 ∈I ρ̂(i )λk (i )
Then DEEM+ is defined as follows:
1. Phase 1: Exploration. DEEM+ is in this phase as long as there is no i such that
mini0 6=i Λ̂k (i, i0 ) ≥ N . After k allocations, choose a job from a distribution αk that satisfies:
P
P
∗ (j)]
g
(i)
α
U
(i)
−
[A(i,
j)
−
p
j
k
i
j∈J
P
αk ∈ A = arg min
, (22)
0
0
α∈∆(J ) mini∈I; i0 ∈Lk (i) (1/gk (i))
j∈J αj KL(i, i |j)/ (1 + log R(i, i )/ log N )
αk can be computed as a solution to a linear program as shown in Appendix B.
15
Exit from exploration. If at some opportunity k, there is a worker type i such that
0
mini0 ∈i
/ Λ̂k (i, i ) ≥ N , then label the worker as being of type i and enter the exploitation phase.
2. Phase 2: Exploitation. Exploitation is the same as defined in DEEM.
DEEM+ shares the same explore-then-exploit structure as DEEM, but with changes to the
exploration phase to optimize learning at finite N ; as we will see in our simulations, these
optimizations allow DEEM+ to substantially outperform DEEM at small N .
Although an exact analysis is beyond the scope of this work, we conjecture that DEEM+ inherits
the asymptotic performance bounds that hold for DEEM. Informally, consider the periods in
DEEM+ where Λ̂k (i, i0 ) < log N for the MAP estimate i, for some type i0 6= i; these periods are analogous to the guessing phase in DEEM. Similar to DEEM, one can argue that this
phase accounts for at most O(log log N/N ) fraction of the worker lifetime. The stages where
Λ̂k (i, i0 ) > log N for all i0 6= i but Λ̂k (i, i0 ) < N for some i0 are similarly analogous to the confirmation subphase of DEEM. Here we can informally argue that on the event that i is the true
type, the posterior distribution quickly concentrates sufficiently on i, and the policy defined by
15
The adjustment in the denominator precisely accounts for differences in regret incurred in making each
distinction.
20
(22) asymptotically achieves the same regret until confirmation (in the leading order term) as
the stationary randomized policy α(i) defined in DEEM. To see this, observe that as gk (i) → 1
for some i, the objective function in (22) converges to the objective function for computing α(i)
in (16), modulo the (1 + log R(i, i0 )/ log N ) factors that simply capture the fact that the learning
goals have been adjusted for small N .
Just like DEEM, DEEM+ can be analogously implemented using queue-based shadow prices
instead of p∗ (j) (with the sets Str(i) computed using smoothed prices).
8
Simulations
In this section, we simulate DEEM and DEEM+ in a market environment with queue-based
shadow prices. We compare performance of these policies against a greedy policy, as well as
benchmark MAB approaches. We consider instances with 3 types of workers and 3 types of
jobs. We assume that N = 30, and ρ(i) = 1 for each i (so that ρ̂(i) = 1/30). We generated
350 instances, where for each instance, independently: (1) µ(j) is sampled from a uniform
distribution on [0.5, 1.5]; and (2) each entry of the expected payoff matrix is sampled from a
uniform distribution on [0, 1].
Given an instance (ρ̂, N, µ, A), our simulated marketplace is described as follows.
Arrival process. Time is discrete, t = 1, 2, · · · , T, where we assume T = 150. At the
beginning of each time period t, Mt (i) number of workers of type i and Lt (j) jobs of type j
arrive, such that M1 (i), M2 (i), · · · and L1 (i), L2 (i), · · · are i.i.d. sequences. For a scaling constant
τ = 900, we assumed that Mt (i) = τ ρ̂(i) = 900 × (1/30) = 30 for all t (recall that ρ̂(i) = 1/30
for all i in all our instances), i.e., Mt (i) is deterministic. We generated Lt (j) from a binomial
distribution with mean τ µ(j) = 900µ(j).16 Each worker stays in the system for N periods and
then leaves. Each job requires one period to perform.
Queues. We assume that the arriving jobs accumulate in queues for the different types,
each with a finite buffer of capacity B, where we choose B = 10, 000. If the buffer capacity is
exceeded for some job type then the remaining jobs are lost.
Matching process. In the beginning of each period, once all the new workers and jobs
have arrived, the platform sequentially considers each worker in the platform,17 and generates
an assignment based on the history of the worker and the chosen policy. If a job of the required
type is unavailable, then the worker remains unmatched. For each worker-job match, a random
payoff is realized, drawn from the distribution specified by A, and the assignment-payoff tuple
is added to the history of the worker.
Queue-based prices. The platform maintains queue-based prices for the jobs in the following way: if the queue length of job type j at any instant is q(j), the price of j at that instant is
set to be pq (j) = (B − q(j))/B. The prices thus change when either (1) new jobs arrive at the
beginning of each period, or (2) a job gets matched to a worker.
A remark on the choice of instances. In all our test instances, all the entries of the
expected payoff matrix A are distinct. We conjecture that this would typically be the case in
many settings in practice – exact indistinguishability of different worker types using a particular
16
A binomial distribution has two parameters (n, p), where n is the number of trials, and p is the probability of
success at each trial; we chose n = 2700 and p = µ(j)/3 for generating each Lt (j). Note that since µ(j) ∈ [0.5, 1.5],
we have p < 1.
17
This consists of the new workers and all the workers who have arrived in the past N − 1 periods, or from the
beginning of time if t < N .
21
1.0
0.8
0.6
Policy
PA-UCB
PA-TS
DEEM
PA-Greedy
DEEM+
PA-UCB
DEEM
PA-TS
PA-Greedy
DEEM+
0.4
0.2
0.0
0.60
0.65
0.70
0.75 0.80 0.85
Performance ratio
0.90
0.95
1.00
Avg. Perf. Ratio (± std. error)
0.797 (± 0.007)
0.904 (± 0.005)
0.898 (± 0.005)
0.931 (± 0.003)
0.943 (± 0.003)
Table 1: Average performance ratios of
different policies across 350 instances,
along with standard errors.
Figure 3: The empirical CDF of the performance ratios of the different policies.
job type wouldn’t be commonly encountered. For such instances, as we discussed in section 5,
there is no conflict between learning and regret minimization in the confirmation subphase of
DEEM, and it incurs a regret of O(log log N/N ), where the leading order term results entirely
from the regret incurred due to uniform sampling of the job types while guessing. In fact, in
this case we can show that a “greedy” policy that maximizes the price-adjusted payoff for the
MAP estimate throughout the exploration phase and enters exploitation after strong distinction
requirements are met for some worker type, incurs a regret of O(1/N ).18 It would thus appear
that this price-adjusted greedy policy, which is attractive for its simplicity, would be a reasonable
solution in such cases. However our simulations show that when N is small, DEEM+ can lead to
significant gains over the price-adjusted greedy approach. We discuss this result further below.
8.1
Results
We implemented five policies: DEEM, DEEM+ , and price-adjusted versions of UCB [10],
Thompson sampling (TS) [4] and Greedy, and compared their performance. All algorithms
measure payoffs adjusted by the queue-based shadow prices; in this way, they all effectively
account for capacity constraints. We have already described this implementation for the DEEM
variants earlier. Price-adjusted Greedy (PA-Greedy) simply chooses the job type that maximizes
the instantaneous shadow price adjusted payoff for the MAP estimate throughout the lifetime of
the worker.19 UCB and TS are well known algorithms for the standard stochastic multi-armed
bandit problem and the details of their implementation in the presence of shadow prices, which
we will denote PA-UCB and PA-TS, can be found in Appendix C.
Figure 3 shows the cumulative distribution function – over the 350 instances – of the ratio
of the payoff generation rate attained by a policy and the optimal payoff generation rate if the
worker types are known, for the five candidate policies. The average of these ratios over the
sample space for each policy is given in Table 2. As one can observe, DEEM+ significantly outperforms PA-UCB, PA-Greedy, DEEM and PA-TS. PA-Greedy, PA-TS, DEEM and DEEM+ perform
considerably better than UCB on average, presumably benefiting from the knowledge of the ex18
Informally, since every job can make every possible distinction between worker types, the probability that
the true worker type is not identified as the MAP estimate at opportunity t decays as exp(−ηt), where η is an
instance dependent constant. Thus the total expected regret over the lifetime of a worker is bounded as O(1) in
expectation.
19
With queue-based shadow prices, tie-breaking during exploitation is rendered unnecessary. Moreover, we
allow the algorithm to continue to benefit from learning during exploitation by optimizing for the current MAP
estimate rather than for the confirmed type. Thus the distinction between exploration and exploitation disappears.
22
pected payoff matrix A. In contrast to PA-Greedy, both DEEM and PA-TS actively experiment
in order to learn quickly: DEEM experiments during its guessing phase where it uniformly samples job types, and PA-TS experiments due to sampling from the posterior (see Appendix C for
details), especially in the early stages when the posterior is not sufficiently concentrated. While
this experimentation is desirable, neither of them efficiently trade off between payoff maximization and learning, resulting in a degraded performance in comparison to PA-Greedy. On the
other hand PA-Greedy suffers from excessive exploitation, resulting in performance which, although better than DEEM and PA-TS, is still significantly worse than DEEM+ . We now focus
on this latter difference.
DEEM+ vs. PA-Greedy. We had discussed in our remark earlier that in instances without
exactly indistinguishable type pairs, PA-Greedy is expected to perform reasonably well. However
in our simulations we see that on an average across all instances, DEEM+ results in about 15%
reduction in regret as compared to PA-Greedy.
In order to gain intuition for this gain, observe that although exactly indistinguishable type
pairs are rarely encountered, it could frequently be the case that two type pairs i and i0 have
their expected payoffs A(i, j) and A(i0 , j) under some job type j close enough that practically
it would take too long to distinguish them with a reasonably small probability of error. This
results in approximately difficult type pairs – two types i and i0 that have different optimal job
types where none of the optimal job types for i is able to distinguish between i and i0 reasonably
quickly. In these situations, if at any point the MAP estimate under the greedy policy is i when
the true type is i0 , then exploiting for i may not allow the policy to recover from its bad estimate
within a reasonable number of jobs, thus incurring high regret. There is high probability of
encountering such a situation in the early stages of the algorithm when the confidence in the
MAP estimate is not sufficiently high.
This is where DEEM+ ’s approach of appropriately allocating confirmation efforts depending
on the posterior results in significant gains in performance over the greedy approach. In particular, there are situations where DEEM+ will appropriately prioritize learning and will actively
explore instead of simply choosing the optimal job type for the MAP estimate. For example,
suppose there are only two types i and i0 , where (i, i0 ) is close to being a difficult pair, i.e., the
learning rate offered by the optimal job type for i towards the (i, i0 ) distinction is close to 0. In
this case, even if the current MAP estimate is i with high confidence (although Λ̂k (i, i0 ) < N ,
so that we are still in the exploration phase), instead of choosing the optimal job type for i,
DEEM+ may choose some other job type that will quickly distinguish i from i0 .
Thus we expect DEEM+ to outperform PA-Greedy more significantly in situations where
there are approximately difficult type pairs. In order to verify that this is indeed the case, first,
we formally define a simple notion of approximate indistinguishability and difficulty. We say
that the type pair (i, i0 ) is ∆-indistinguishable using a job type j if KL(i, i0 |j) < ∆; otherwise
we say that it is ∆-distinguishable. We say that type pair (i, i0 ) is ∆-difficult if KL(i, i0 |j) < ∆
for all j ∈ J (i) and J (i) ∩ J (i0 ) = φ.
For ∆ ∈ {0.1, 0.4}, we picked those instances out of the 350 in which there is at least one pair
(i, i0 ) such that (1) (i, i0 ) is ∆-difficult, and (2) there is some j such that KL(i, i0 |j) > 0.5; i.e.,
instances with at least one ∆-difficult type pair, such that there exists a job type under which this
pair is 0.5-distinguishable. We will call such instances ∆-DBD (∆-difficult-but-distinguishable)
instances. These are precisely the instances where measured exploration in cases where the MAP
estimate is a worker type that forms a ∆-difficult type pair with some other type can lead to
significant gains in performance.20 Note that as ∆ increases, the set of instances that satisfy
20
Note that if KL(i, i0 |j) > 0.5, the job j can distinguish i from i0 with a misclassification error of 1/30 for
23
Sample
A
B
Number of instances
78
82
Avg. regret reduction
21%
12%
Table 2: DEEM+ ’s percentage reduction in regret relative to Greedy on average in sample A,
consisting of 0.1-DBD instances, and in sample B consisting of 0.4-DBD instances that aren’t
0.1-DBD.
these conditions grows progressively larger (a ∆-DBD instance is also a ∆0 -DBD instance for
∆0 > ∆).
We next considered two sets of samples: sample A is the set of 0.1-DBD instances, and
sample B is the set of 0.4-DBD instances that are not 0.1-DBD. Based on the discussion above,
we should expect a substantial reduction of regret in the 0.1-DBD instances, relative to those
0.4-DBD instances that are not 0.1-DBD. Indeed, consistent with our intuition, a one tailed two
sample t-test showed that the mean percentage reduction of regret in sample A is larger than
that in sample B with a p-value of 0.009. The sample average percentage reduction of regret in
the two samples is given in Table 2.
9
Conclusion
This work suggests a novel and practical algorithm for learning while matching, applicable across
a range of online matching platforms. Several directions of generalization remain open for future
work. First, while we consider a finite-type model, a richer model of types would admit a wider
range of applications; e.g., workers and jobs may be characterized by features in a vector-valued
space, with compatibility determined by the inner product between feature vectors. Second,
while our model includes only one-sided uncertainty, in general a market will include two-sided
uncertainty (i.e., both supply and demand will exhibit type uncertainty). We expect that a
similar approach using externality prices to first set learning objectives, and then achieve them
while incurring minimum regret, should be applicable even in these more general settings.
Third, recall that we assumed the expected surplus from a match between a worker type
and a job type (i.e., the A matrix) is known to the platform. This reflects a first order concern
of many platforms, where aggregate knowledge is available, but learning individual user types
quickly is challenging. Nevertheless, it may also be of interest to study how A can be efficiently
learned by the platform. This direction may be related to issues addressed by the literature on
(a single) multi-armed bandit under capacity constraints [2].
We conclude by noting that our model ignores strategic behavior by participants. A simple
extension might be to presume that workers are less likely to return after several bad experiences;
this would dramatically alter the model, forcing the policy to become more conservative. The
modeling and analysis of these and other strategic behaviors remain important challenges.
References
[1] Rajeev Agrawal, Demosthenis Teneketzis, and Venkatachalam Anantharam. Asymptotically
efficient adaptive allocation schemes for controlled iid processes: finite parameter space.
strong distinction, in at most about 7 jobs on average, which is reasonably quick.
24
Automatic Control, IEEE Transactions on, 34(3):258–267, 1989.
[2] Shipra Agrawal and Nikhil R Devanur. Bandits with concave rewards and convex knapsacks.
In Proceedings of the fifteenth ACM conference on Economics and computation, pages 989–
1006. ACM, 2014.
[3] Shipra Agrawal and Nikhil R Devanur. Linear contextual bandits with global constraints
and objective. arXiv preprint arXiv:1507.06738, 2015.
[4] Shipra Agrawal and Navin Goyal. Analysis of thompson sampling for the multi-armed
bandit problem. arXiv preprint arXiv:1111.1797, 2011.
[5] Shipra Agrawal, Nikhil R Devanur, and Lihong Li. Contextual bandits with global constraints and objective. arXiv preprint arXiv:1506.03374, 2015.
[6] Mohammad Akbarpour, Shengwu Li, and Shayan Oveis Gharan. Dynamic matching market
design. Available at SSRN 2394319, 2014.
[7] Ross Anderson, Itai Ashlagi, David Gamarnik, and Yash Kanoria. A dynamic model of
barter exchange. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 1925–1933. SIAM, 2015.
[8] Baris Ata and Sunil Kumar. Heavy traffic analysis of open processing networks with complete resource pooling: asymptotic optimality of discrete review policies. The Annals of
Applied Probability, 15(1A):331–391, 2005.
[9] J.-Y. Audibert and R. Munos. Introduction to bandits: Algorithms and theory. In ICML,
2011.
[10] Peter Auer, Nicolo Cesa-Bianchi, and Paul Fischer. Finite-time analysis of the multiarmed
bandit problem. Machine learning, 47(2-3):235–256, 2002.
[11] Moshe Babaioff, Shaddin Dughmi, Robert Kleinberg, and Aleksandrs Slivkins. Dynamic
pricing with limited supply. ACM Transactions on Economics and Computation, 3(1):4,
2015.
[12] Mariagiovanna Baccara, SangMok Lee, and Leeat Yariv. Optimal dynamic matching. Available at SSRN 2641670, 2015.
[13] Ashwinkumar Badanidiyuru, Robert Kleinberg, and Yaron Singer. Learning on a budget:
posted price mechanisms for online procurement. In Proceedings of the 13th ACM Conference on Electronic Commerce, pages 128–145. ACM, 2012.
[14] Ashwinkumar Badanidiyuru, Robert Kleinberg, and Aleksandrs Slivkins. Bandits with
knapsacks. In Foundations of Computer Science (FOCS), 2013 IEEE 54th Annual Symposium on, pages 207–216. IEEE, 2013.
[15] Ashwinkumar Badanidiyuru, John Langford, and Aleksandrs Slivkins. Resourceful contextual bandits. In Proceedings of The 27th Conference on Learning Theory, pages 1109–1134,
2014.
[16] Omar Besbes and Assaf Zeevi. Dynamic pricing without knowing the demand function:
Risk bounds and near-optimal algorithms. Operations Research, 57(6):1407–1420, 2009.
25
[17] Omar Besbes and Assaf Zeevi. Blind network revenue management. Operations research,
60(6):1537–1550, 2012.
[18] Sébastien Bubeck and Nicolo Cesa-Bianchi. Regret analysis of stochastic and nonstochastic
multi-armed bandit problems. Machine Learning, 5(1):1–122, 2012.
[19] Jim G Dai. On positive harris recurrence of multiclass queueing networks: a unified approach via fluid limit models. The Annals of Applied Probability, pages 49–77, 1995.
[20] Ettore Damiano and Ricky Lam. Stability in dynamic matching markets. Games and
Economic Behavior, 52(1):34–53, 2005.
[21] Sanmay Das and Emir Kamenica. Two-sided bandits and the dating market. In Proceedings
of the 19th international joint conference on Artificial intelligence, pages 947–952. Morgan
Kaufmann Publishers Inc., 2005.
[22] Daniel Fershtman and Alessandro Pavan. Dynamic matching: experimentation and cross
subsidization. Technical report, Citeseer, 2015.
[23] John Gittins, Kevin Glazebrook, and Richard Weber. Multi-armed bandit allocation indices.
John Wiley & Sons, 2011.
[24] Ming Hu and Yun Zhou. Dynamic matching in a two-sided market. Available at SSRN,
2015.
[25] Sangram V Kadam and Maciej H Kotowski. Multi-period matching. Technical report,
Harvard University, John F. Kennedy School of Government, 2015.
[26] Emilie Kaufmann, Nathaniel Korda, and Rémi Munos. Thompson sampling: An asymptotically optimal finite-time analysis. In Algorithmic Learning Theory, pages 199–213. Springer,
2012.
[27] Morimitsu Kurino. Credibility, efficiency, and stability: A theory of dynamic matching
markets. 2005.
[28] Tze Leung Lai and Herbert Robbins. Asymptotically efficient adaptive allocation rules.
Advances in applied mathematics, 6(1):4–22, 1985.
[29] Constantinos Maglaras and Assaf Zeevi. Pricing and capacity sizing for systems with shared
resources: Approximate solutions and scaling relations. Management Science, 49(8):1018–
1038, 2003.
[30] Constantinos Maglaras and Assaf Zeevi. Pricing and design of differentiated services: Approximate analysis and structural insights. Operations Research, 53(2):242–262, 2005.
[31] Laurent Massoulie and Kuang Xu. On the capacity of information processing systems, 2016.
Unpublished.
[32] Aranyak Mehta. Online matching and ad allocation. Theoretical Computer Science, 8(4):
265–368, 2012.
[33] Daniel Russo and Benjamin Van Roy. Learning to optimize via posterior sampling. Mathematics of Operations Research, 39(4):1221–1243, 2014.
[34] Denis Sauré and Assaf Zeevi. Optimal dynamic assortment planning with demand learning.
Manufacturing & Service Operations Management, 15(3):387–404, 2013.
26
[35] Lloyd S Shapley and Martin Shubik. The assignment game i: The core. International
Journal of game theory, 1(1):111–130, 1971.
[36] Adish Singla and Andreas Krause. Truthful incentives in crowdsourcing tasks using regret
minimization mechanisms. In Proceedings of the 22nd international conference on World
Wide Web, pages 1167–1178. International World Wide Web Conferences Steering Committee, 2013.
[37] Zizhuo Wang, Shiming Deng, and Yinyu Ye. Close the gaps: A learning-while-doing algorithm for single-product revenue management problems. Operations Research, 62(2):
318–331, 2014.
Appendices
A
Proof of Theorem 5.1
For the rest of this section, let C be the quantity defined in (20). We present it again for the
convenience of the reader:
P
∗
X
j∈J αj U (i) − [A(i, j) − p (j)]
P
C(i) = min
;
C=
ρ(i)C(i) .
0
mini0 ∈Str(i) j∈J αj KL(i, i |j)
α∈∆(J )
i∈I
Recall problem 7. We will first show the following lower bound on the difference between V ∗
and W N , which follows directly from Agrawal et al. [1].
Proposition A.1.
lim sup
N →∞
N
V ∗ − W N ≥ C.
log N
Proof. Consider the following relaxed problem:
X
X
X
X
WpN∗ = max
ρ(i)
x(i, j)A(i, j) −
p∗ (j)[
ρ(i)x(i, j) − µ(j)].
x∈X N
i∈I
j∈J
j∈J
(23)
i∈I
By a standard duality argument, we know that WpN∗ ≥ W N . The optimal policy in this problem
is a solution to
X
maximizeπ∈ΠN
ρ(i)xπ (i, j)(A(i, j) − p∗ (j)) .
(24)
i∈I, j∈J
Then from Theorem 3.1 in Agrawal et al. [1], we know that
lim sup
N →∞
N
V ∗ − WpN∗ ≥ C.
log N
The result then follows from the fact that W N ≤ WpN∗ .
Let WpN∗ (π ∗ ) be the value attained by DEEM in optimization problem (21) (same as (23)),
assuming that the routing matrix y in the exploitation phase is such that y(i, .) is supported on
27
J (i). We will prove an upper bound on the difference between V ∗ and WpN∗ (π ∗ ). Note that the
difference in these values of the two problems is the same as the difference in
X
X
ρ(i)
xπ∗ (i, j)(A(i, j) − p∗ (j)),
i∈I
and
P
i∈I
j∈J
ρ(i)U (i). Following is the result.
Proposition A.2. Consider the sequence of policies (π ∗ (N ))N ≥1 such that the routing matrix
y used in the exploitation phase satisfies y(i, .) ∈ ∆(J (i)). Then,
lim sup
N →∞
N
V ∗ − WpN∗ (π ∗ ) ≤ C.
log N
Further, suppose that there are no difficult type pairs. Then,
lim sup
N →∞
N
V ∗ − WpN∗ (π ∗ ) ≤ K
log log N
where K = K(ρ, µ, A) ∈ [0, ∞) is some constant.
In order to prove this Proposition, we need the following result that follows from Theorem
4.1 in [1].
Lemma A.3. Let X1 , X2 , · · · be i.i.d. random variables where Xi is the outcome of choosing a
job type j ∈ J according to a distribution α ∈ ∆(J ). Suppose i ∈ I and B ⊆ I \ {i} are such
that
X
αj KL(i, i0 |j) > 0
j∈J
for each
i0
∈ B. Let
ΛB
k (i)
1.
lim sup
N →∞
= mini0 ∈B Λk (i, i0 ). Then,
Ei [inf{k ≥ 0|ΛB
1
k (i) ≥ f (N )}]
P
≤
,
log f (N )
mini0 ∈B j∈J αj KL(i, i0 |j)
2.
Pi0 (Λk (i, i0 ) ≥ a for some k ≤ N ) ≤
ρ̂(i)/ρ̂(i0 )
.
a
for any a > 0.
Next, we also need the following result.
Lemma A.4. Let X1j , X2j , · · · be i.i.d. random variables for each j = 1, · · · , K, such that
|Xij | ≤ M and E(Xij ) = mj > 0. Let a, b and k j be such that a < k j < b for each j. Let
P
Snj = k j + ni=1 Xij and let B ⊆ {1, · · · , K}. Let E be the event:
0
{Snj < a for some j 0 before Snj > b for all j ∈ B}.
Let T = inf{n : Snj < a for some j}. Then
E[T 1E ] ≤ G
for some 0 < G < ∞ that does not depend on a, b, or k j for any j.
28
0
Proof. Define k j − a , z j . If we define Ej = {Snj < a for some n} and Tj = inf{n : Snj < a},
PK
P
then we have E ⊆ ∪j Ej , and thus we have E[T 1E ] ≤ K
j=1 E[Tj 1Ej ]. Now we
j=1 E[T 1Ej ] ≤
have
E[Tj 1Ej ] =
≤
≤
=
≤
∞
X
n=1
∞
X
n=1
∞
X
n=1
∞
X
n=1
∞
X
n=1
nP(T = n)
n
X
nP(
Xij ≤ −zj )
i=1
n exp(
−(nmj + zj )2
)
4nM 2
n exp(−
zj2
nm2j
mj zj
−
−
)
4M 2
2M 2
4nM 2
n exp(−
nm2j
) = G(mj , M ) < ∞,
4M 2
where the second inequality results from the Hoeffding bound. Taking G =
proves the result.
PK
j
j=1 G(m , M )
Proof of Proposition A.2. Let X denote the type of the worker. Let R(i) denote the expected
total regret over the lifetime of a worker on the event {X = i}, defined as
X
R(i) = N max[A(i, j) − p∗ (j)] − N
xπ∗ (i, j)[A(i, j) − p∗ (j))].
j∈J
j∈J
Here N xπ∗ (i, j) is the expected total number of times a job of type j is allotted to a worker of
type i under the policy π ∗ (N ). We will refer to the above quantity as just regret. For the rest
of the proof, all the expectations are on the event {X = i}. The proof will utilize the fact that
the log of the ratio of the posteriors, log(Λk (i, i0 )), for any i and i0 is a random walk, such that
if αk is the probability distribution over job types chosen at opportunity k, then
log(Λk+1 (i, i0 )) − log(Λk (i, i0 )) = log(
p(Xk , i, jk )
),
p(Xk , i0 , jk )
p(Xk ,i,jk )
where the random variables {log( p(X
)}k are independent random variables with a finite
0
k ,i ,jk )
P
support (since Xk and jk take finite values), and with mean j αjk KL(i, i0 |j). (Note here that if
P k
0
0
0
j αj KL(i, i |j) = 0 then since KL(i, i |j) ≥ 0, it must be that KL(i, i |j) = 0 for all j such that
p(Xk ,i,jk )
αjk > 0, and in this case we must have A(i, j) = A(i, j 0 ) for all such j. Thus log( p(X
) = 0,
0
k ,i ,jk )
i.e., the if the drift of the random walk is 0 at some k then the random walk has stopped.) Recall
that log(Λ0 (i, i0 )) = log(ρ̂(i)/ρ̂(i0 )).
Our goal is to compute an upper bound on R(i). To do so we first compute the expected regret
incurred till the end of the exploration phase in our algorithm. Denote this by Re (i). Below,
we will find an upper bound on this regret assuming that the worker performs an unbounded
number of jobs. Clearly the same bound holds on the expected regret until the end of exploration
phase if the worker leaves after N jobs.
Our strategy is as follows: we will decompose the regret till the end of exploration into the
regret incurred till the first time one of the following two events occurs:
29
1. Event A: mini0 6=i log Λk (i, i0 ) ≥ log log N (or mini0 6=i Λk (i, i0 ) ≥ log N ) and
2. Event B: mini0 6=i log Λk (i, i0 ) ≤ − log log N (or mini0 6=i Λk (i, i0 ) ≤
1
log N ),
followed by the residual regret, which will depend on which event occurred first. Note that one
of these two events will occur with probability 1.
We will compute two different upper bounds, depending on two different regimes of initial
posterior distributions of the different types (note that the posterior probabilities of the different
types i under the observed history is a sufficient statistic at any opportunity under our policy).
First, suppose that R̄(i) is highest expected regret incurred over all possible starting posteriors
that a) do not satisfy the conditions of both A and B and b) such that mini0 6=i log Λ0 (i, i0 ) ≥
mini0 6=i log(ρ̂(i)/ρ̂(i0 )). Let L1 be the set of starting posteriors that satisfy these conditions.
Next, suppose that R̃(i) is the highest expected regret incurred, where the supremum is taken
over all possible starting posteriors that a) do not satisfy the conditions of both A and B and
b) such that mini0 6=i log Λn (i, i0 ) < log(ρ̂(i)/ρ̂(i0 )). Let L2 be the set of posteriors that satisfy
these conditions. Clearly, Re (i) ≤ R̄(i).
Let G(i) denote the maximum expected regret incurred by the algorithm till one of A or B
occurs, where the maximum is taken over all possible starting posteriors that do not satisfy the
conditions of both A and B, i.e., L1 ∪ L2 . For convenience, we denote A < B as the event that
A occurs before B and vice versa (similarly for any two events). Thus we have
R̄(i) ≤ G(i) + sup P(A < B|l1 )E(Residual regret|A, l1 )
l1 ∈L1
+ sup P(B < A|l1 )E(Residual regret|B, l1 )
l1 ∈L1
and
R̃(i) ≤ G(i) + sup P(A < B|l2 )E(Residual regret|A, l2 )
l2 ∈L2
+ sup P(B < A|l2 )E(Residual regret|B, l2 ).
l2 ∈L2
First, let us find a bound on G(i). This is easy, because, G(i) ≤ E(inf{k > 0 : mini6=i0 Λk (i, i0 ) ≥
log N }}) = O(log log N ) from Lemma A.3 (since if neither condition A, nor B is satisfied, then
the policy in the guessing phase, and thus all job types are utilized with positive probability, and
hence the condition in the Lemma of the requirement of a positive learning rate for each distinction is satisfied). Also, from the second statement in Lemma A.3, since the posteriors in L1 are
such that mini0 6=i Λn (i, i0 ) ≥ mini0 6=i ρ̂(i)/ρ̂(i0 ), we have that P(B < A|l1 ) ≤ P(B ever occurs) ≤
O(1)/ log N . Finally we have supl2 ∈L2 P(B < A|l2 ) = w < 1. We thus have
2
R̄(i) ≤ O(log log N ) + sup E(Residual regret|A, l1 ) +
l1 ∈L1
1
sup E(Residual regret|B, l1 ) and
log N l1 ∈L1
(25)
R̃(i) ≤ O(log log N ) + sup E(Residual regret|A, l2 ) + w sup E(Residual regret|B, l2 ).
l2 ∈L2
(26)
l2 ∈L2
Next, consider suplr ∈Lr E(Residual regret|A, lr ). This depends on which of the following two
events happens next:
1. Event A0 : mini0 6=i log Λk (i, i0 ) < log log N (or mini0 6=i Λk (i, i0 ) < log N ),
30
2. Event A00 : i gets confirmed, i.e., mini0 ∈Str(i) log Λk (i, i0 ) > log N (or mini0 ∈Str(i) Λk (i, i0 ) >
N ).
Again conditional on A, one of the two events will occur with probability 1. We have
sup E(Residual regret|A, lr ) = sup [E(Residual regret|A, A0 < A00 , lr )P(A0 < A00 |A, lr )
lr ∈Lr
lr ∈Lr
+ E(Residual regret|A, A0 > A00 , lr )P(A0 > A00 |A, lr )].
Now from Lemma A.4 it follows that
E(Residual regret|A, A0 < A00 , lr )P(A0 < A00 |A, lr )
= E(Residual regret I{A0 <A00 } |A, lr ) ≤ M + R̄(i)P(A0 < A00 |A, lr )
for some constant M that does not depend on lr or N . To see this, note that A0 < A00 is the event
that, starting from some values between log log N and log N , some random walk Λk (i, i0 ) for i0 6= i
crosses the lower threshold log log N before all the random walks Λk (i, i0 ) for each i0 ∈ Str(i)
cross the upper threshold log N . Now between these two thresholds, the job distribution αk
equals α(i) for all k. Hence the drift for any of the random walks Λk (i, i0 ) for each i0 ∈ Str(i)
is strictly positive and finite. Further, as we argued earlier, if the drift for any of these random
walks is 0, then that random walk has stopped, and such random walks can be ignored. Thus
the conditions of Lemma A.4 are satisfied, and hence E((Time till A0 ) I{A0 <A00 } |A, lr ) = G < ∞.
Since the regret per unit time is bounded, the deduction follows. Moving on, we have
E(Residual regret|A, A00 < A0 , lr )
≤ E(inf{k > 0 : min Λk (i, i0 ) ≥ N })
i0 ∈Str(i)
X
αj U (i) − [A(i, j) − p∗ (j)] .
j∈J
Thus we have
sup E(Residual regret|A, lr ) ≤ O(1) + sup P(A0 < A00 |A, lr )R̄(i)
lr ∈Lr
(27)
lr ∈Lr
+ (1 − P(A0 < A00 |A, lr ))E(inf{k > 0 : min Λk (i, i0 ) ≥ N })
i0 ∈Str(i)
X
αj U (i) − [A(i, j) − p∗ (j)] .
j∈J
(28)
Thus we have
sup E(Residual regret|A, lr ) ≤ O(1) + qk R̄(i)
(29)
lr ∈Lr
+ (1 − qk )E(inf{k > 0 : min Λk (i, i0 ) ≥ N })
i0 ∈Str(i)
X
αj U (i) − [A(i, j) − p∗ (j)] ,
(30)
j∈J
for some qk where qk < 1 since suplr ∈Lr P(A0 < A00 |A, lr ) < 1.
Next, consider suplr ∈Lr E(Residual regret|B, lr ). This depends on which of the following two
events occurs next:
1. Event B 0 : mini0 6=i log Λk (i, i0 ) ≥ − log log N (or mini0 6=i Λk (i, i0 ) ≥
1
log N ),
2. Event B 00 : Some i0 6= i gets confirmed, i.e., mini00 ∈Str(i0 ) log Λk (i0 , i00 ) > log N
(or mini00 ∈Str(i0 ) Λk (i0 , i00 ) > N ).
31
Again, conditional on B, one of the two events will occur with probability 1. Let K(i) be the
maximum expected regret incurred till either B or B 0 occurs given that B has occurred and
the starting likelihoods were in Lr . Note that if B 00 < B 0 then the exploration phase ends
and hence there is no residual regret (Although note that if i0 is such that i ∈ Str(i0 ), then
P(B 00 < B 0 |B, lr ) ≤ O(1)/N from the second statement in Lemma A.3.). Then we have
sup E(Residual regret|B, lr ) ≤ K(i) + sup P(B 0 < B 00 |B, lr )R̃(i).
lr ∈Lr
lr ∈Lr
Now we can show that if there is a type i0 such that i ∈ I \ Str(i0 ), then K(i) ≤ O(log N ),
where as if there is no such type, then K(i) = O(1). We first show this. Let T (i) be the
maximum expected time taken till either B or B 0 occurs given that B has occurred and the
starting likelihoods were in Lr . Clearly K(i) ≤ T (i) since the price adjusted payoffs lie in [0, 1].
Now, let T1 be the time spent after B has occurred, before B or B 0 occurs, while either a)
algorithm is in the guessing phase or b) the algorithm is in the confirmation phase for some
guessed type i0 such that α(i0 )j > 0 for some j such that KL(i, i0 |j) > 0. Under this case, we
will say that the algorithm is in state 1, and let 1k be the event that the algorithm is in state
2 at time k. Next let T2 be the time spent after B has occurred, before B or B 0 occurs, while
the algorithm is in the confirmation phase for some guessed type i0 such that α(i0 )j = 0 for all
j such that KL(i, i0 |j) > 0 (clearly this can happen only for i0 such that i ∈ I \ Str(i0 ); thus
if such an i0 doesn’t exist, then T2 = 0). Under this case, we will say that the algorithm is in
state 2, and let 2k be the event that the algorithm is in state 2 at time k. Now we clearly have
T (i) ≤ supr ∈Lr E(T1 |B, lr ) + supr ∈Lr E(T2 |B, lr ).
Let Γk (i) = mini0 6=i log Λk (i, i0 ). Then observe that E[(Γk+1 (i)−Γk (i))I1k |B, lr ] > m for some
m > 0 that depends only on the primitives of the problem and E[(Γk+1 (i) − Γk (i))I2k |B, lr ] = 0,
i.e, when the algorithm is in state 1, the drift in Γk (i) is strictly positive where as when the
algorithm is in state 2, then Γk (i) does not change.
Now consider E(T1 |B, lr ). Let k ∗ be the opportunity that B occurred for the first time. Then
clearly Γk∗ (i) = − log log N − ε, where ε ≥ 0 is bounded by some constant M depending only
on the problem instance. Thus P(T1 > t|B, lr ) ≤ P(Γk0 (i) − Γk∗ (i) ≤ ε|B, lr ) for some k 0 > k ∗
such that the algorithm has been in state 1 t times at opportunity k 0 . Thus our observation
above implies that P(T1 > t|B, lr ) ≤ exp(−ct) for some c > 0, by a standard application of a
concentration inequality. Thus E(T1 |B, lr ) = O(1).
Next consider E(T2 |B, lr ). Consider the successive returns of the algorithm to state 2. Conditional on the algorithm having entered state 2, the expected time spent in that state is bounded
by the expected time till the guessed type i0 is confirmed, which is O(log N ) from Lemma A.3,
and the conditional probability that i0 gets confirmed is some q > 0. Thus the total expected
number of returns to state 2 is bounded by 1/q. Thus E(T2 |B, lr ) = O(log N ) as well. Thus
K(i) ≤ O(log N ) and we have
sup E(Residual regret|B, lr ) ≤ O(log N ) + R̃(i).
lr ∈Lr
32
And thus we finally have
1
O(log(N ) + R̃(i)
log N
X
+ (1 − q1 )E(inf{k > 0 : min Λk (i, i0 ) ≥ N })
αj U (i) − [A(i, j) − p∗ (j)] ;
R̄(i) ≤ O(log log N ) + q1 R̄(i) +
i0 ∈Str(i)
(31)
j∈J
R̃(i) ≤ O(log log N ) + q2 R̄(i) + wO(log N ) + wR̃(i)
X
+ (1 − q2 )E(inf{k > 0 : min Λk (i, i0 ) ≥ N })
αj U (i) − [A(i, j) − p∗ (j)] .
i0 ∈Str(i)
(32)
j∈J
Combining the above two equations, we deduce that
1 − q1
O(log log N )
1 − q1 − q2 / log N
X
+ E[inf{k > 0 : min Λn (i, i0 ) ≥ N }]
αj U (i) − [A(i, j) − p∗ (j)]
Re (i) ≤ R̄(i) ≤
i0 ∈Str(i)
j∈J
= O(log log N )
+ (1 + o(1))E[inf{k > 0 : min Λn (i, i0 ) ≥ N }]
i0 ∈Str(i)
X
αj U (i) − [A(i, j) − p∗ (j)] . (33)
j∈J
Now, we observed earlier that P(i0 gets confirmed | {X = i}) ≤ 1/N if i0 ∈ Str(i). Thus the
regret in the exploitation phase is in the worst case of order O(N ) with probability 1/N and 0
otherwise. Thus the total expected regret in the exploitation phase is O(1). Thus
X
R(i) ≤ O(log log N )+(1+o(1))E[inf{k > 0 : min Λk (i, i0 ) ≥ N }]
αj U (i)−[A(i, j)−p∗ (j)] .
i0 ∈Str(i)
j∈J
Thus Lemma A.3 implies the result (note that if there are no difficult type pairs, then
[A(i, j) − p∗ (j)] = 0).
P
j∈J
αj U (i)−
Next, we prove that for a large enough N , one can choose a routing matrix y ∗ in the exploitation phase of DEEM that will ensure that matches optimize price-adjusted payoffs, and
such that the capacity and complementary slackness conditions are satisfied.
Proposition A.5. Suppose that the generalized imbalance condition is satisfied. Consider any
optimal routing matrix x∗ that is an optimal solution to problem (11). Then in the policy π ∗ (N )
for the N -job problem, for any N large enough, one can choose a routing matrix y ∗ such that
y ∗ (i, .) ∈ ∆(J (i)) and that satisfies:
1.
P
2.
P
i∈I
ρ(i)xπ∗ (i, j) = µ(j) for any j such that
i∈I
ρ(i)xπ∗ (i, j) < µ(j) for any other j.
P
i∈I
ρ(i)x∗ (i, j) = µ(j), and
We remark that the y ∗ we construct satisfies ky ∗ − x∗ k = o(1). In order to prove this
proposition, we will need the following Lemma.
Lemma A.6. Suppose that the generalized imbalance condition
P is satisfied. Consider any feasible
routing matrix [x(i, j)]I×J . Consider any job j such that i∈I ρ(i)x(i, j) = µ(j). Then there is
a path on the complete bipartite graph between worker types I and job types J with the following
properties:
33
• One end point is job j.
• The other end point is a job type whose capacity is under-utilized (it is permitted to be κ).
• For every job type on the path in between, they are operating at capacity/all jobs are being
served. (All worker types are fully utilized by definition, since we formally consider an
unassigned worker as being assigned to job type κ.)
• For every undirected edge on the path, there is a positive rate of jobs routed on that edge
in x.
Proof. Consider a bi-partite graph with jobs representing nodes on one side and workers on the
other. There is an edge between a job j 0 and a worker i if x(i, j) > 0. Consider the connected
component of job type j in this graph. Suppose it includes no job type that is underutilized.
Then the arrival rate of jobs from the set of workers in the connected component exactly matches
the total effective service rate of the sellers in connected component. But this is a contradiction
since generalized imbalance holds. Hence there exists an underutilized job type j 0 that can be
reached from j. Take any path from j to j 0 . Traverse it starting from j and terminate it the
first time it hits any underutilized job type.
Proof of Proposition A.5. Recall that for a given routing matrix [y(i, j)]I×J , xπ∗ (i, j), is the
resulting fraction of jobs of type j directed to worker type i. In the course of this proof, we will
suppress the subscript π ∗ . Clearly, there exist εi0 (i, j) for each i ∈ I, i0 ∈ I ∪ {0}, j ∈ J such
that we have
X
x(i, j) = ε0 (i, j) + (1 − εi (i, j))y(i, j) +
εi0 (i, j)y(i0 , j) .
(34)
i0 ∈I\{i}
The ε’s depend on the guessing and confirmation phases but not on y. (In particular, ε0 arises
from the overall routing contribution of the guessing and confirmation phases, and εi ’s arise
from the small likelihood that a worker who is confirmed as type i is actually some other type.)
A key fact that we will use is that all ε’s are uniformly bounded by o(1).
P
P
Let Jx∗ = {s : i∈I ρ(i)x∗ (i, j) = µ(j)} and Jπ∗ = {j : i∈I ρ(i)xπ∗ (i, j) = µ(j)}. Now we
want to find a y such that y(i, ·) ∈ ∆(J (i)) for all i ∈ I (call (i, j) a “permissible edge” in the
bipartite graph between workers and jobs if j ∈ J (i)), and such that:
• For each j ∈ Jx∗ we also have j ∈ Jπ∗ , i.e., Jx∗ ⊆ Jπ∗ .
• ky − x∗ k = o(1).
Note that the two P
bullets together willPimply the proposition, since kx−x∗ k = o(1) from Eq. (34),
and this leads to i∈I ρ(i)x(i, j) = i∈I ρ(i)x∗ (i, j) + o(1) < µ(j) for all j ∈ J \Jx∗ , for large
enough N .
The requirement in the first bullet can be written as a set of linear equations using Eq. (34).
Here we write y (and later also x∗ ) as a column vector with |I||J | elements:
By + ε̂ = (µ(j))j∈Jx∗ .
Here we have kε̂k = o(1) and matrix B can be written as B = B0 + Bε , where B0 has 1’s in
columns corresponding to dimensions (·, s) and 0’s everywhere else, and kBε k = o(1). Expressing
y as y = x + z, we are left with the following equation for z,
Bz = −(Bε x∗ + ε̂)
34
(35)
using the fact that B0 x∗ = (µ(j))j∈Jx∗ by definitions of B0 and Jx∗ . We will look for a solution
to this underdetermined set of equations with a specific structure: we want z to be a linear
combination of flows along |Jx∗ | paths coming from Lemma A.6, one path λj for each j ∈ Jx∗ .
Each λj can be written as a column vector with +1’s on the odd edges (including the edge
incident on j) and −1’s on the even edges. Let Λ = [λj ]j∈Jx∗ be the path matrix. Then z with
the desired structure can be expressed as Λη, where η is the vector of flows along each of the
paths. Now note that Bz = (B0 + Bε )Λη = (I + Bε Λ)η. Here we deduced B0 Λ = I from the
fact that λj is a path which has j as one end point, and a worker or else a job not in Jx∗ as the
other end point. Our system of equations reduces to
(I + Bε Λ)η = −(Bε x∗ + ε̂) ,
Since kBε k = o(1), the coefficient matrix is extremely well behaved being o(1) different from
the identity, and we deduce that this system of equations has a unique solution η ∗ that satisfies
kη ∗ k = o(1). This yields us z ∗ = Λη ∗ that is also of size o(1), and supported on permissible
edges since each of the paths is supported on permissible edges (Lemma A.6). Thus, we finally
obtain y ∗ = x∗ + z ∗ possessing all the desired properties. Notice that the (permissible) edges on
which y ∗ differs from x∗ had strictly positive values in x∗ by Lemma A.6, and hence this is also
the case in y ∗ for large enough N .
Finally, we show that with the choice of y ∗ constructed in Proposition A.5 in the exploitation
phase, the sequence of policies (π ∗ (N )) asymptotically achieve the required upper bound on
regret.
Proposition A.7. Suppose that the generalized imbalance condition is satisfied. Consider the
sequence of policies (π ∗ (N ))N ≥1 , with the routing matrix y ∗ proposed in Proposition A.5. Let
W N (π ∗ ) be the value attained by this policy in optimization problem (7). Then
lim sup
N →∞
N
V ∗ − W N (π ∗ ) ≤ C.
log N
Further, suppose that there are no difficult type pairs. Then,
lim sup
N →∞
N
V ∗ − W N (π ∗ ) ≤ K
log log N
where K = K(ρ, µ, A) ∈ (0, ∞) is some constant.
Proof. From Proposition A.5 it follows that the policy π ∗ is feasible in problem 7, and further
X
X
X
X
WpN∗ (π ∗ ) =
ρ(i)
xπ∗ (i, j)A(i, j) −
p∗ (j)[
ρ(i)xπ∗ (i, j) − µ(j)].
(36)
=
i∈I
j∈J
X
X
i∈I
ρ(i)
j∈J
xπ∗ (i, j)A(i, j),
i∈I
(37)
j∈J
P
∗
where the second equality follows from the fact that if p∗ (j) > 0, then i∈I ρ(i)x
P (i, j)−µ(j) = 0
by complementary slackness, and hence from Proposition A.5 we obtain that i∈I ρ(i)xπ∗ (i, j)−
µ(j) = 0 as well for these j. Thus we have a policy π ∗ that is feasible, and that gives a rate
WpN∗ (π ∗ ) of accumulation of payoff in problem (7). Thus the result follows from Proposition A.2.
35
B
Computation of the policy in the confirmation subphase
Denoting mini0 ∈Str(i)
lem is the same as:
P
j∈J
αj KL(i, i0 |j) as h (where h is non-negative), the optimization prob-
X αj
U (i) − [A(i, j) − p∗ (j)]
h
j∈J
X αj
s.t.
KL(i, i0 |j) ≥ 1 for all i0 ∈ Str(i),
h
min
j∈J
X αj
αj
1 1
= ,
≥ 0 and
≥ 0 for all j.
h
h h
h
j∈J
Now redefine
αj
h
, ᾱj and
1
h
, h̄ to obtain the linear program:
min
X
ᾱj U (i) − [A(i, j) − p∗ (j)]
j∈J
s.t.
X
ᾱj KL(i, i0 |j) ≥ 1 for all i0 ∈ Str(i),
j∈J
ᾱj ≥ 0 for all j,
P
P
where at any optimal solution ᾱ∗ , we have h̄∗ = j∈J ᾱj∗ , and thus α(i)j = ᾱj∗ / j∈J ᾱj∗ . Note
that a feasible solution exists to this linear program as long as KL(i, i0 |j) > 0 for some j for
each i0 . When there are multiple solutions, we choosePthe solution with the largest learning rate;
i.e., we choose a solution with the smallest h̄∗ , i.e., j∈J ᾱ∗ . One way to accomplish this is to
P
P
modify the objective to minimize j∈J ᾱj U (i) − [A(i, j) − p∗ (j)] + τ j∈J ᾱj for some small
τ > 0.
For small
can simply evaluate all the finite extreme points of the constrained
P problems, we
0 |j) ≥ 1 for all i0 ∈ Str(i); ᾱ ≥ 0 for all j}; i.e., all the extreme points
ᾱ
KL(i,
i
set {ᾱ :
j
j∈J j
such that ᾱj < ∞ for all j. This is sufficient because we can show that there always exists a
finite solution to the linear program. To see this, note that
ᾱj = ᾱjm ,
X
i0 ∈Str(i)
1
0
I
KL(i, i0 |j) {KL(i,i |j)>0}
is feasible and finite. Further, for any solution such that αj∗ > ᾱjm , αj∗ can be reduced to ᾱjm
without loss in objective while maintaining feasibility.
For the practical heuristic DEEM+ , αk can be computed as a solution to the following
optimization problem.
X
X αj
min
gk (i)
U (i) − [A(i, j) − p∗ (j)]
h
i
j∈J
X αj
s.t.
KL(i, i0 |j) ≥ gk (i) 1 + log R(i, i0 )/ log N for all i ∈ I, i0 ∈ Lk (i),
h
j∈J
X αj
αj
1 1
= ,
≥ 0 and
≥ 0 for all j.
h
h h
h
j∈J
We can again redefine
αj
h
, ᾱj and
1
h
, h̄ to obtain a linear program.
36
C
Practical implementation of other policies
1. PA-UCB: The upper confidence bound (UCB) algorithm is a popular multi-armed bandit algorithm [9, 10, 18], that embodies the well known approach of “optimism in the face of
uncertainty" to solving these problems. In its classical implementation, one keeps track of highprobability confidence intervals for the expected payoffs of each arm, and at each step chooses
an arm that has the highest upper confidence bound, i.e., the highest upper boundary of the
confidence interval. To be precise, if rj (t) is the average reward seen for some arm j, that has
been pulled nj times until time t, then the upper confidence bound for the mean reward for this
arm is given by:
q
uj (t) = rj (t) +
2 log t/nj .
The algorithm chooses arm j = arg maxj µj (t). In our context, the arms are the job types, and
if k jobs have already been allotted to a worker, and r̄j (k) is the average payoff obtained from
past assignments of job j, and nj is the number of these assignments, we will define
q
uj (k) = rj (k) + 2 log k/nj − pq (j),
where pq (j) is the current queue length based price for job j. The PA-UCB algorithm then
chooses job type j = arg maxj∈J µj to be assigned to the worker next. Note that this algorithm
does not require the knowledge of the instance primitives (ρ̂, N, µ, A).
2. Thompson sampling (PA-TS): Thompson sampling is another popular multi-armed
bandit algorithm [9, 18] employing a Bayesian approach to the problem of arm selection. The
description of the algorithm is simple: starting with a prior, at every step select an arm with a
probability equal to the posterior probability of that arm being optimal. These posterior probabilities are updated based on observations made at each step. One can incorporate information
about correlation between the rewards of the different arms in computing these posteriors, which
makes it a versatile algorithm that exploits the reward structure in multiple settings. It is known
to give asymptotically tight regret guarantees in many multi-armed bandit problems of interest
[4, 26, 33].
In our simulations, a price-adjusted version of TS is implemented as follows. The prior
probability of the worker being of type i is P ρ̂(i)
With this as the prior, depending on
0 .
i0 ρ̂(i )
each worker’s history, a posterior distribution of the type of the worker is computed using the
knowledge of the expected payoff matrix A. Then a worker type is sampled from this distribution.
Suppose this type is i, then a job type j ∗ = arg maxj A(i, j) − pq (j) is assigned to the worker.
In contrast to the PA-UCB algorithm, Thompson sampling does utilize the knowledge of the
expected payoff matrix A as well as the arrival rates ρ̂ (the latter to construct a starting prior,
and the former for the posterior updates).
D
D.1
Other proofs
Sufficiency of worker-history-only policies
We show that there is a worker-history-only (WHO) policy that achieves a rate of payoff accumulation that is arbitrarily close to the maximum possible. We will think of N as being fixed
throughout this section.
37
Suppose the system starts at time t = 1 with no workers already present before the period.21
Arrivals thereafter occur as described in Section 3. Consider any (arbitrary, time varying)
policy π and let xπ,t (i, j) denote the derived quantity representing the fraction of workers of
type i who are assigned jobs on type j in period t under π. Then the largest possible rate of
payoff accumulation under policy π over long horizons is
V (π) = limsupT →∞ VT (π)
where VT (π) =
1
T
T X
X
ρ(i)
t=1 i∈I
X
(38)
xπ,t (i, j)A(i, j) .
(39)
j∈J
Note that we have ignored the effect of less than ρ(i) workers of type i being present for the first
N periods, but this does not change the limiting value V . Also, note that randomization in π
cannot increase the achievable value of V , since one can always do as well by picking the most
favorable sample path.
Claim D.1. Fix any policy π and any ε > 0. Then there is a worker-history-only (WHO) policy
that achieves a steady state rate of payoff accumulation exceeding V (π) − ε.
Proof. We suppress dependence on π. By definition of V , we know that there exists an increasing
sequence of times T1 , T2 , . . . such that VTi > V − ε/2 for all i = 1, 2, . . .. We will construct a
suitable WHO policy by using a sufficiently large time in this sequence. Let νt (Hk ) be the
measure of workers in the system with history Hk just before the start of time t, and abusing
notation, let νt (Hk )j be the measure of such workers who are assigned to job type j at time t.
Since the policy cannot assign more jobs than have arrived in any period, we have
N X
X
νt (Hk )j ≤ µ(j)
for all t ≥ 1 .
(40)
k=1 Hk
Fix T , which we think of as a large member of the sequence above. The average measure of
workers with history Hk who are present is
T
1X
ν̄(Hk ) =
νt (Hk )
T
for all Hk and k = 1, 2, . . . , N .
(41)
t=1
The average measure of such workers who are assigned job j is similarly defined and denoted by
ν̄(Hk , j). We immediately have that
N X
X
ν̄(Hk )j ≤ µ(j) ,
(42)
k=1 Hk
by averaging Eq. (40) over times until T . Now, consider a worker with history Hk assigned a job
of type j. Using the known A matrix and arrival rates ρ, we can infer the posterior distribution
of the worker type based on Hk , and hence, the likelihood of the job of type j being successfully
completed. Let p(Hk , j) denote the probability of success. Then the distribution of Hk+1 for
the worker is simply given by
Hk+1 = Hk , j, Bernoulli(p(Hk , j)) .
21
The analysis would be very similar and produce the same results if the starting state is an arbitrary one with
a bounded mass of workers already present.
38
Barring the edge effect at time T caused by workers whose history was Hk at time T , this
allows us to uniquely determine ν̄(Hk+1 ) based on ν̄(Hk , j)’s. In particular, for any δ1 > 0, if
T ≥ maxi∈I ρ(i)/(N δ1 ) we have that
δ1
ν̄(Hk , (j, 1)) ≈ ν̄(Hk )j p(Hk , j)
δ1
ν̄(Hk , (j, 0)) ≈ ν̄(Hk )j 1 − p(Hk , j) .
(43)
δ
Here, a ≈ b represents the bound |a − b| ≤ δ. Note that we have
VT =
N X
X
ν̄(Hk )j p(Hk , j) .
(44)
k=1 Hk
We are now ready to define our WHO policy π. For every Hk such that ν̄(Hk ) ≥ δ2 , this
policy will attempt to assign a fraction ν̄(Hk )j /ν̄(Hk ) of workers with history Hk to jobs of type
j. Ignore capacity constraints for the present. (We will find that the capacity constraints will
be almost satisfied.) Leave the choice of δ2 for later; we will choose δ2 small and then choose
0 < δ1 < δ2 /N so as to achieve the desired value of δ3 below. Workers with rare histories, i.e.,
histories such that ν̄(Hk ) < δ2 , will not be assigned jobs under π. Note that the definition of
rare histories refers to frequency of occurrence under π. Then, this uniquely specifies π as well
as the steady state mix of workers before any time t. In particular, the steady state mass ν(Hk )
under π of workers with history Hk that are not rare is bounded as
(1 − δ1 /δ2 )k−1 ν̄(Hk ) ≤ ν(Hk ) ≤ (1 + δ1 /δ2 )k−1 ν̄(Hk )
(45)
using Eq. (43), and the fact that all subhistories of Hk are also not rare. It follows that
δ3
ν(Hk ) ≈ ν̄(Hk ) where δ3 = max(exp(N δ1 /δ2 ) − 1, δ2 ) ,
(46)
for all histories (including rare histories), using k ≤ N and ν(Hk ) ≤ 1. Violation of the j-capacity
constraint under π is given by
N X
N X
X
X
ν(Hk )j − µ(j) ≤
ν̄(Hk )j − µ(j) + 2N |J |N −1 δ3 = 2N |J |N −1 δ3
k=1 Hk
+
k=1 Hk
+
using Eq. (46) and Eq. (42), and the fact that there are k≤N (2|J |)N −1 ≤ 2N |J |N −1 possible
histories. It follows that the sum of capacity constraint violations across j ∈ J is bounded by
(2|J |)N δ3 . Pick an arbitrary set of workers to go unmatched to get rid of any capacity violations
(this can be done while remaining within the class of WHO policies). In worst case, this will
cause payoff loss of 1 for each period remaining in the worker’s lifetime. Thus, the loss caused
by the need to remedy capacity violations is bounded by δ4 = N (2|J |)N δ3 per period.
P
Ignoring capacity violations, the steady state rate of accumulation of payoff under π is
N X
X
k=1 Hk
δ5
ν(Hk )j p(Hk , j) ≈
N X
X
ν̄(Hk )j p(Hk , j) = VT (π)
k=1 Hk
where δ5 = 2N |J |N −1 δ3 < δ4 .
39
(47)
again using Eq. (46) and the fact that there are
P
k≤N (2|J |)
N −1
≤ 2N |J |N −1 possible histories.
Let V (π) denote the true steady state rate of accumulation of payoff under π when capacity
constraints are considered. Combining the above, we deduce that V (π) ≥ VT (π) − 2δ4 . The time
T will be chosen as a member of the sequence defined at the beginning of the proof, ensuring
VT (π) ≥ V (π) − ε/2; hence it will suffice to show V (π) ≥ VT (π) − ε/2. Hence, it suffices to have
δ4 = ε/4, which can achieved using δ3 = δ2 = ε/(4N (2|J |)N ) and δ1 = δ3 log(1 + δ3 )/N and T
a member of the sequence satisfying T ≥ maxi∈I ρ(i)/(N δ1 ). This yields the required bound of
V (π) ≥ V (π) − ε.
D.2
Uniqueness of prices under generalized imbalance
Proposition D.2. Under the generalized imbalance condition, the job shadow prices p∗ are
uniquely determined.
Proof of Proposition D.2. The dual to problem 11 can be written as
X
X
minimize
µ(j)p(j) +
ρ(i)v(i)
j∈J
i∈I
subject to
p(j) + v(i) ≥ A(i, j) ∀i ∈ I, j ∈ J ,
p(j) ≥ 0 ∀j ∈ J ,
v(i) ≥ 0 ∀i ∈ I .
The dual variables are (P, V ) where “job prices” P = (p(j))j∈J and “worker values” V =
(v(i))i∈I . We will prove the result by contradiction. Suppose there are multiple dual optima.
Let D be the set of dual optima. Let J 0 be the set of jobs such that the prices of those jobs
take multiple values in D. Formally,
J 0 = {j ∈ J : p(j) takes multiple values in D} .
(48)
Similarly, let I 0 be the set of workers such that the prices of those workers take multiple values
in D. Formally,
I 0 = {i ∈ I : v(i) takes multiple values in D} .
(49)
For each j ∈ J 0 , we immediately deduce that there exists a dual optimum with p(j) > 0, and
hence the capacity constraint of job type j is tight in all primal optima.PSimilarly, we deduce
that for each i ∈ I 0 , worker type i is assigned a job in all periods, i.e., j∈J x(i, j) = 1. By
assumption, we have
X
X
ρ(i) 6=
µ(j) .
i∈I 0
j∈J 0
Suppose the left hand side is larger than the right (the complementary case can be dealt with
similarly). Take any primal optimum x∗ . The jobs in J 0 do not have enough capacity to serve
all workers in I 0 , hence there must be some worker i ∈ I 0 and a job s ∈
/ J 0 such that x∗ (i, s) > 0.
0
Since s ∈
/ J , we must have that p(j) has a unique optimum value in D. Call this value p∗ (j).
Let the largest and smallest values of v(i) in D be v max (i) and v min (i). By complementary
slackness, we know that
v max (i) + p∗ (j) = A(i, j) = v min (i) + p∗ (j)
⇒ v max (i) = v min (i) .
40
But since i ∈ I 0 we must have v max (i) > v min (i). Thus we have obtained a contradiction.
The proof of the next proposition shows that a very simple “learn then exploit" strategy
achieves a regret of O(log N/N ). This follows from the fact that, under an identifiability condition, the sequence of sets (X N ) converges to the set D in an appropriately defined distance.
Proposition D.3. Suppose that no two rows in A are identical. Then supx∈D inf y∈X N kx−yk =
O logNN
Proof. It is clear that xN ⊆ D. We will find an inner approximation X̃ N to X N such that
X̃ N ⊆ X N , and X̃ N converges to D in an appropriate sense as N goes to infinity. To define this
approximation, suppose that in the learning problem corresponding to a fixed N , one starts off
with a exploration phase of a fixed length O(log N ), where each job j is presented to the worker
Os number of times (where Os = O(log N ), fixed a priori), so that after this phase, the type of
the worker becomes known with a probability of error at most O(1/N ). This will then allow us
to relate the problem to the problem in which the user type is known.
Suppose after this phase, the probability that a worker of type i is correctly identified is p(i)
and the probability that she is mis-identified as some other type i0 is p(i, i0 ). Note that since
no two rows in A are identical, p(i, i0 ) = O(1/N ) for all i 6= i0 . Let d(i, j) denote the expected
number of times a worker that has been identified as being of type i (correctly or incorrectly)
is directed towards job j after the exploration phase, i.e., from job Os + 1 till the N th job. Let
¯ j) = d(i, j)/N . Then we can see that, one can attain all x in the following set:
d(i,
X
N
¯ j) ≥ 0;
X̃
=
x ∈ R|I|×|J | : d(i,
x(ib, s) = 1;
(50)
j∈J
X
Os
0 ¯ 0
¯
x(i, j) =
+ p(i)d(i, j) +
p(i, i )d(i , s)
N
0
(51)
i 6=i
¯ j) ≤ 1, and since p(i, i0 ) ≤ O(1/N ), we can express the above set as:
Now since d(i,
X
¯ j) ≥ 0;
¯ j) + O log N
X̃ N =
x ∈ R|I|×|J | : d(i,
x(i, j) = 1; x(i, j) = d(i,
(52)
N
j∈J
This in turn is the same as:
log N X
N
|I|×|J |
X̃
=
x∈R
: x(i, j) ≥ O
;
x(i, j) = 1
N
(53)
j∈J
Note that by construction, X̃ N ⊆ xN . But we can now see that X̃ N converges to D in the sense
that
log N
sup inf kx − yk = O
,
N
x∈D y∈X̃ N
and hence,
log N
sup inf kx − yk = O
N
x∈D y∈X N
as well.
Proposition D.4. The set X N is a convex polytope.
41
Proof. For the purpose of this proof, let
X
N
= {N x : x ∈ X N }.
N
We will show that X is a polytope, from which the result will follow. We will prove this using
N
an induction argument. We will represent each point in X as a |I| × |J | matrix (x(i, j))|I|×|J | .
Let worker types in I be labeled as i1 , . . . , i|I| and let job types in J be labeled as j1 , . . . , j|J | .
Now clearly, X
0
= {(0)|I|×|J | } which is a convex polytope. We will show that if X
N
is
(N +1)
a convex polytope, then X
is one as well, and hence the result will follow. To do so, we
decompose the assignment problem with (N +1) jobs, into the first job and the remaining N jobs.
A policy in the (N + 1)- jobs problem is a choice of a randomization over the jobs in J for the
first job, and depending on whether a reward was obtained or not with the chosen job, a choice
N
of a point in X to be achieved for the remaining N jobs. Each such policy gives a point in the
(N +1)
N
X
. Suppose that η1 ∈ ∆(J ) is the randomization chosen for job 1, and let R(j, 1) ∈ X
N
and R(j, 0) ∈ X be the points chosen to be achieved from job 2 onwards depending on the job
j that was chosen, and whether a reward was obtained or not, i.e.. R(., .) is a mapping from
N
J × {0, 1} to the set X . Then this policy achieves the following point in the (N + 1)- jobs
problem:
η1 (j1 ) η1 (j2 ) . . . η1 (j|J | )
..
..
.. + X η (j) Diag[A(., j)]R(j, 1) + Diag[Ā(., j)]R(j, 0) ,
.
.
.
1
.
.
.
j∈J
η1 (j1 ) η1 (j2 ) . . . η1 (j|J | )
where
A(i1 , j)
0
...
0
A(i
,
j)
...
2
Diag[A(., j)] = .
.
..
..
..
.
0
0
...
and
1 − A(i1 , j)
0
...
0
1 − A(i2 , j) . . .
Diag[Ā(., j)] =
..
..
..
.
.
.
0
0
...
0
0
0
A(i|I| , j)
0
0
.
0
1 − A(i|I| , j)
And thus we have
(N +1)
X =
η1 (j1 ) η1 (j2 ) . . .
..
..
..
.
.
.
η1 (j1 ) η1 (j2 ) . . .
η1 (j|J | )
.. + X η (j) Diag[A(., j)]R(j, 1) + Diag[Ā(., j)]R(j, 0)
1
.
η1 (j|J | )
N
.
: η1 ∈ ∆(J ), R(., .) ∈ X
j∈J
Let 1s be the |I| × |J | matrix with ones along column corresponding to job type j and all
other entries 0. Then the set
N
,
J (s) = 1s + Diag[A(., j)]R(j, 1) + Diag[Ā(., j)]R(j, 0) : R(s, .) ∈ X
42
is a convex polytope, being a linear combination of two convex polytopes, followed by an affine
(N +1)
shift. It is easy to see that X
is just a convex combination of the polytopes J (s) for j ∈ J ,
(N +1)
is a convex polytope as well.
and hence X
43
| 8cs.DS
|
USING MATCHING TO DETECT INFEASIBILITY OF SOME INTEGER PROGRAMS
arXiv:1703.01532v1 [cs.DS] 4 Mar 2017
S.J. GISMONDI∗, E.R. SWART†
Abstract. A novel matching based heuristic algorithm designed to detect specially formulated infeasible {0, 1} IPs is
presented. The algorithm’s input is a set of nested doubly stochastic subsystems and a set E of instance defining variables set
at zero level. The algorithm deduces additional variables at zero level until either a constraint is violated (the IP is infeasible),
or no more variables can be deduced zero (the IP is undecided). All feasible IPs, and all infeasible IPs not detected infeasible
are undecided. We successfully apply the algorithm to a small set of specially formulated infeasible {0, 1} IP instances of the
Hamilton cycle decision problem. We show how to model both the graph and subgraph isomorphism decision problems for input
to the algorithm. Increased levels of nested doubly stochastic subsystems can be implemented dynamically. The algorithm is
designed for parallel processing, and for inclusion of techniques in addition to matching.
Key words. integer program, matching, permutations, decision problem
MSC Subject classifications. 05C45 68R10 90C05 90C10
1. Introduction. We present a novel matching based heuristic algorithm deigned to detect specially
formulated infeasible {0, 1} IPs. It either detects an infeasible IP or exits undecided. It does not solve an IP.
We call it the triple overlay matching based closure algorithm (the algorithm). Input to the algorithm is an
IP whose constraints are a set of nested doubly stochastic boolean subsystems [12] together with a set E of
instance defining variables set at zero level. The IP’s solution set is a subset of the set of n! nxn permutation
matrices P , written as n2 xn2 block permutation matrices Q each with block structure P . The algorithm is a
polynomial time search that deduces additional variables at zero level via matching until either a constraint
is violated in which case the IP is infeasible, or we can go no further in which case the IP is undecided. If
the IP is decided infeasible, a set of variables deduced to be at zero level can be used to test and display
a set of violated constraints. If the IP is undecided, additional variables deduced zero can be added to E,
and nothing more can be concluded. While some infeasible IPs may fail to be detected infeasible (not yet
found), feasible IPs can only fall in the undecided category.
In section 2 we present the generic IP required as input to the algorithm, and we view the set of all
solutions of the IP as an n2 xn2 block permutation matrix Q whose components are {0, 1} variables. Each
nxn block (u, i) is nxn permutation matrix P where block (u, i) contains pu,i in position (u, i). An instance
is modelled by setting certain variables of Q to zero level. In sections 3, 4 and 5, we present the algorithm,
an application / matching model of the Hamilton cycle decision problem (HCP), empirical results and two
conjectures. In section 6, we present generalizations of the algorithm, matching models for both the graph
and subgraph isomorphism decision problems, and other uses. We also propose more development. Its
success, effectiveness and practicality can then be evaluated in comparison to other algorithms. We invite
researchers to collaborate with us. Contact the corresponding author for FORTRAN code.
The ideas presented in this paper originated from a polyhedral model of cycles not in graphs [9]. At that
time we thought about how to recognize the Birkhoff polytope as an image of a solution set of a compact
formulation for non-Hamiltonian graphs. We’ve accomplished part of that goal in this paper. That is, the
convex hull of all excluded permutations for infeasible IPs is the Birkhoff polytope, and its easy to build
a compact formulation from E. In this paper, over 2,100 non-Hamiltonian graphs ranging from 10 - 104
vertices are correctly decided as infeasible IPs. None failed that are not reported. Although counterexamples
surely exist, we believe there is an insightful theory to be discovered that explains these early successes.
2. About Specially Constructed {0, 1} IPs and Terminology. Imagine a {0, 1} integer program
modelled such that P is a solution if and only if the integer program is feasible e.g. matching. Also imagine
an arbitrary set of instance defining constraints of the form pu,i + pv,j ≤ 1. It’s not obvious how to apply
matching to help in its solution. Now imagine that we create a compact formulation whose solution set is
isomorphic (i.e. equal under an orthogonal projection), where we convert each linear constraint into all of its
instantiated discrete states via creation of a set of discrete {0, 1} variables. Then it becomes easy to exploit
matching. Hence the algorithm.
∗ University
of Guelph, Canada, Email: [email protected] (Corresponding Author)
British Columbia, Canada, Email: [email protected]
Ted and I dedicate this paper to the late Pal Fischer (16/11/2016) - friend, colleague and mentor.
† Kelowna,
Code the IP above so that each of the instance defining constraints is a set of two distinct components
of P {pu,i , pv,j }, interchangeably playing the role of a {0, 1} variable for which {pu,i , pv,j } = 0 if and only if
pu,i = pv,j = 0, or pu,i = 1 and pv,j = 0, or pu,i = 0 and pv,j = 1. Create an instance of the IP by creating
an instance of exclusion set E whose elements are the set of these {pu,i , pv,j }. If there exists P satisfying
{pu,i , pv,j } = 0 for all {pu,i , pv,j } ∈ E, then P is a solution of the IP. Otherwise P satisfies {pu,i , pv,j } = 1
for at least one {pu,i , pv,j } ∈ E and P is excluded from the solution set of the IP. We view elements of E as
coding precisely the set of permutation matrices excluded from the solution set of the IP. That is, E excludes
the union of sets of (n − 2)! P , each set satisfying {pu,i , pv,j }=1, for each {pu,i , pv,j } ∈ E. An example of
the modelling technique needed to create E is presented in section 4, originally presented in [9]. We exclude
these permutation matrices by setting {pu,i , pv,j } = 0 for each {pu,i , pv,j } ∈ E.
The complement of exclusion set E, with respect to all {pu,i , pv,j } is called available set V . The IP is
feasible if and only if there exists P whose set of n(n−1)
distinct pairs of components {pu,i , pv,j } that satisfy
2
pu,i = pv,j = 1 and define P are in V . P is said to be covered by V if there exists a subset of n(n−1)
2
{pu,i , pv,j } ∈ V such that pu,i = pv,j = 1 defines P and each {pu,i , pv,j } participates in P ’s cover.
Definition 2.1. Clos(E) (closed exclusion set E) is the set of all {pu,i , pv,j } not participating in any
cover of any P .
Note that {pu,i , pv,j } ∈ E is code for the set of (n − 2)! permutation matrices for which pu,i = pv,j = 1.
Clearly if clos(E) is such that all n! permutation matrices are accounted, then there is no P covered by V
i.e. V is empty.
Definition 2.2. Open(V ) (open available set V ) is the complement of clos(E) w.r.t. all {pu,i , pv,j },
the set of all {pu,i , pv,j } participating in a cover of at least one P .
Theorem 2.1.
P The IP is infeasible if and only if open(V ) = ∅.
System 1: P i pu,i = 1, u = 1, 2, ..., n
u pu,i = 1, i = 1, 2, ..., n
F or
Pall u, i = 1, 2, ..., n
Pj6=i {pu,i , pv,j } = pu,i , v = 1, 2, ..., n, v 6= u.
v6=u {pu,i , pv,j } = pu,i , j = 1, 2, ..., n, j 6= i.
{pu,i , pv,j } ∈ E ⇒ Assign {pu,i , pv,j } = 0.
pu,i , {pu,i , pv,j } ∈ {0, 1}
Visualize system 1 in the form of n2 xn2 permutation matrix Q, n2 blocks of P . Block (u, i) contains
pu,i in position (u, i), the remaining entries in row u and column i being zero. The rest of the entries in
block (u, i) have the form {pu,i , pv,j }, v 6= u, j 6= i. It’s assumed variables in Q have been initialized by E.
Henceforth, we present the algorithm in terms of matrix Q. See Figure 2.1, an example of the general form
of matrix Q for n = 4. For E = ∅, the set of n! nxn permutation matrices each written as a Q matrix i.e. in
n2 xn2 block form, is the set of integer extrema of the solution set of system 1§ . See Figure 2.2, an example
of an integer solution to system 1 in matrix Q form, for n = 4.
p
11
0
0
0
0
p
21
0
0
0
0
p31
0
0
0
0
p41
0
p12
0
0
0
0
p13
0
0
0
0
p14
0
0
0
{p11 p22 } {p11 p23 } {p11 p24 } {p12 p21 } 0 {p12 p23 } {p12 p24 } {p13 p21 } {p13 p22 } 0 {p13 p24 } {p14 p21 } {p14 p22 } {p14 p23 } 0
{p11 p32 } {p11 p33 } {p11 p34 } {p12 p31 } 0 {p12 p33 } {p12 p34 } {p13 p31 } {p13 p32 } 0 {p13 p34 } {p14 p31 } {p14 p32 } {p14 p33 } 0
{p11 p42 } {p11 p43 } {p11 p44 } {p12 p41 } 0 {p12 p43 } {p12 p44 } {p13 p41 } {p13 p42 } 0 {p13 p44 } {p14 p41 } {p14 p42 } {p14 p43 } 0
{p21 p12 } {p21 p13 } {p21 p14 } {p22 p11 } 0 {p22 p13 } {p22 p14 } {p23 p11 } {p23 p12 } 0 {p23 p14 } {p24 p11 } {p24 p12 } {p24 p13 } 0
0
p22
0
0
0
0
p23
0
0
0
0
p24
0
0
0
{p21 p32 } {p21 p33 } {p21 p34 } {p22 p31 } 0 {p22 p33 } {p22 p34 } {p23 p31 } {p23 p32 } 0 {p23 p34 } {p24 p31 } {p24 p32 } {p24 p33 } 0
{p21 p42 } {p21 p43 } {p21 p44 } {p22 p41 } 0 {p42 p33 } {p22 p44 } {p23 p41 } {p23 p42 } 0 {p23 p44 } {p24 p41 } {p24 p42 } {p24 p43 } 0
{p31 p12 } {p31 p13 } {p31 p14 } {p32 p11 } 0 {p32 p13 } {p32 p14 } {p33 p11 } {p33 p12 } 0 {p33 p14 } {p34 p11 } {p34 p12 } {p34 p13 } 0
{p31 p22 } {p31 p23 } {p31 p24 } {p32 p21 } 0 {p32 p23 } {p32 p24 } {p33 p21 } {p33 p22 } 0 {p33 p24 } {p34 p21 } {p34 p22 } {p34 p23 } 0
0
0
0
0
p32
0
0
0
0
p33
0
0
0
0
p34
{p31 p42 } {p31 p43 } {p31 p44 } {p32 p41 } 0 {p32 p43 } {p32 p44 } {p33 p41 } {p33 p42 } 0 {p33 p44 } {p34 p41 } {p34 p42 } {p34 p43 } 0
{p41 p12 } {p41 p13 } {p41 p14 } {p42 p11 } 0 {p42 p13 } {p42 p14 } {p43 p11 } {p43 p12 } 0 {p43 p14 } {p44 p11 } {p44 p12 } {p44 p13 } 0
{p41 p22 } {p41 p23 } {p41 p24 } {p42 p21 } 0 {p42 p23 } {p42 p24 } {p43 p21 } {p43 p22 } 0 {p43 p24 } {p44 p21 } {p44 p22 } {p44 p23 } 0
{p41 p32 } {p41 p33 } {p41 p34 } {p42 p31 } 0 {p42 p33 } {p42 p34 } {p43 p31 } {p43 p32 } 0 {p43 p34 } {p44 p31 } {p44 p32 } {p44 p33 } 0
0
0
0
0
p42
0
0
0
0
p43
0
0
0
0
p44
Fig. 2.1. General Form of Matrix Q, n = 4.
§ An
integer solution of system 1 exists if and only if it’s an nxn permutation matrix in n2 xn2 block form [7].
2
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
Fig. 2.2. An Integer Solution to System 1 in Matrix Q Form, n = 4.
3. About Triple Overlay Matching Based Closure. We first present an overview of the algorithm,
followed by the formal algorithm. Let E be given. Encode Q and create V .
3.1. Overview of the Triple Overlay Matching Based Closure Algorithm. Rather that search
for the existence of P covered by V , we attempt to shrink V so that {pu,i , pv,j } ∈ V if and only if {pu,i , pv,j }
participates in a cover of at least one P . The algorithm deduces which {pu,i , pv,j } ∈ V do not participate in
any cover of any P , removes them from V , and adds them to E. Its success depends upon whether or not
it’s true that for infeasible IPs, when we initialize Q via E, it’s sufficient to deduce open(V ) = ∅¶ . While
it’s impossible for a feasible IP to yield open(V ) = ∅, infeasible IPs cause the algorithm to either deduce
infeasibility or exit undecided. We say undecided because although we deduce some of these {pu,i , pv,j } ∈ V
that do not participate in any cover of any P , it’s not known if we deduce all of these {pu,i , pv,j }. Brief
details about how the algorithm deduces variables at zero level in every solution of the IP now follow.
The algorithm systematically tests a set of necessary conditions assuming a feasible IP each time a qu,i,v,j
is set at unit level. That is, if pu,i = pv,j =1, blocks (u, i) and (v, j) are assumed to cover a match, a necessary
condition for the existence of a n2 xn2 block permutation matrix solution of the IP. But rather than test for
a match covered by these two blocks, we exhaust all choices of a third variable common to these blocks, set
at unit level, and test for the existence of a match covered by all three blocks. After exhausting all possible
choices of a third‡‡ variable, if no match exists, the given qu,i,v,j variable is deduced zero. Otherwise we
conclude nothing. In both cases we continue on to the next variable not yet deduced zero. Eventually no
more variables can be deduced zero, none of the constraints appear violated and the IP is undecided; or
enough of the variables are deduced zero such that a constraint is violated and the IP is infeasible.
3.2. The Triple Overlay Matching Based Closure Algorithm. Interchangeably associate matrix
Q with a {0, 1} matrix that has entries at zero level where matrix Q has entries at zero level, and unit entries
where matrix Q has pu,i or {pu,i , pv,j } entries. We’ll now reference {0, 1} variables qu,i,u,i and qu,i,v,j . A unit
entry in the uth row and ith column of non-empty block (u, i) represents variable pu,i . The remaining unit
entries in the v th row and j th column of block (u, i) with u 6= v and i 6= j can be regarded as representing
pv,j variables (which is what they really do represent in the case of a {0, 1} solution) and, they can also be
regarded as representing {pu,i , pv,j } variables. We think of this associated matrix in terms of patterns in Q
that cover n2 xn2 block permutation matrices, and then we’ll exploit matching.
Definition 3.1. Match(*) is a logical function. Input * is an nxn {0, 1} matrix. Row labels are viewed
as vertices 1 through n (set A), and column labels are viewed as vertices n + 1 through 2n (set B). Match(*)
¶ In earlier work [12], we create an equivalence class, the set of all possible V ’s none of which cover any P , whose class
representative is ∅.
‡‡ Hence the term triple overlay. Every variable not deduced zero participates in a match in an overlay of three blocks of
Q. There exists quadrupal, quintuple overlay through to exhaustion where the algorithm tests factorial numbers of n (n − 2 is
sufficient) overlays for a match.
3
returns TRUE if there exists a match between A and B. Otherwise Match(*) returns FALSE.
Definition 3.2. Overlay(*1 ,*2 ) is a binary AND function applied component-wise to two nxn {0, 1}
matrices. Its output is a {0, 1} matrix. We loosely use the terms double and triple overlay in place of
Overlay(*1 ,*2 ) and Overlay(Overlay(*1 ,*2 ),*3 ) etc.
Definition 3.3. Check RowsColumns Q is a routine that returns TRUE if a row or column in matrix Q is all 0, in which case the algorithm terminates and the graph is deduced infeasible. Otherwise
Check RowsColumns Q returns FALSE. In our FORTRAN implementation of the algorithm, before testing
for termination, we also implement Boolean closure within and between blocks in Q. This efficiently deduces
some of the non-zero components of Q to be at zero level and we note significant speed increases.
Note that Boolean closure in Check RowsColumns Q can be replaced by LP. Temporarily set a nonzero component of matrix Q to unit level and check for infeasibility subject to doubly stochastic constraints
of matrix Q. Infeasibility implies that the component can be set to zero level.
Whenever the algorithm exits undecided, then for every non-zero qu,i,v,j , there exists a match in a triple
overlay of blocks (u, i), (v, j) and at least one (w, k) block. The IP is then not deduced infeasible and we call
the corresponding matrix Q the non-empty triple overlay closure of the IP. Otherwise the algorithm exits
and the IP is deduced infeasible i.e. open(V ) is deduced to be empty.
Input: {open(V ) ← V , Q}
Output: {open(V ), decision}
if Check RowsColumns Q EXIT {open(V ), infeasible};
CONTINUE TRIPLE CLOSURE;
oldQ ← Q;
for u, i = 1, 2, ..., n; and qu,i,u,i 6= 0 do
if ˜Match(Q(u, i)) then
qu,i,u,i ← 0;
for v, j = 1, 2, ..., n; and u 6= v and i 6= j and qu,i,v,j 6= 0 do
qu,i,v,j ← 0; qv,j,u,i ← 0; open(V ) ← open(V )\{pu,i , pv,j }\{pv,j , pu,i };
end
if Check RowsColumns Q EXIT {open(V ), infeasible};
⇒ NEXT i ;
end
for v, j = 1, 2, ..., n; and u 6= v and i 6= j and Q(u, i)v,j 6= 0 do
if ˜Match(Overlay(Q(u, i), Q(v, j))) then
qu,i,v,j ← 0; qv,j,u,i ← 0; open(V ) ← open(V )\{pu,i , pv,j }\{pv,j , pu,i };
⇒ NEXT j ;
end
DoubleOverlay ← Overlay(Q(u, i), Q(v, j));
TRIPLE CLOSURE;
for w, k = 1, 2, ..., n; and u 6= w 6= v, i 6= k 6= j and DoubleOverlayw,k 6= 0 do
if ˜Match(Overlay(DoubleOverlay, Q(w, k))) then
DoubleOverlay w,k ← 0;
⇒ TRIPLE CLOSURE;
end
end
if ˜Match(DoubleOverlay) then
qu,i,v,j ← 0; qv,j,u,i ← 0; open(V ) ← open(V )\{pu,i , pv,j }\{pv,j , pu,i };
end
end
end
if oldQ 6= Q ⇒ CONTINUE TRIPLE CLOSURE;
EXIT {open(V ), undecided};
Algorithm 1: The Triple Overlay Matching Based Closure Algorithm
4
4. Application to the HCP. Let G be an n + 1 vertex graph also referenced by its adjacency matrix.
We model the HCP for simple, connected 3-regular graphs, as do others [5, 10, 3], called the 3HCP.
4.1. Background Information and Classification of 3-Regular Graphs. The 3HCP is a well
known decision problem and is NP-complete [6]. G is 3-colourable (edge) if G is Hamiltonian, and since
3-regular graphs are either 3 or 4-colourable, it follows that if G is 4-colourable then G is non-Hamiltonian.
These graphs were initially studied by Peter Tait in the 1880s (named Snarks by Martin Gardner in 1976).
Tait conjectured (1884) that every 3-connected, planar, 3-regular graph has a Hamilton cycle, later disproved
by Tutte in 1946 via construction of a 46 vertex counterexample. This was a significant conjecture, and had
it been true, it implied the famous 4-colour theorem. These ideas are summarized in the figure below.
All Simple, Connected, 3-Regular Graphs
Hamiltonian Graphs
3-Colourable
non-Hamiltonian Graphs
non-Snarks
3-Colourable
Tutte’s Counterexample
Snarks
4-Colourable
Fig. 4.1. Classification of Simple, Connected, 3-Regular Graphs
4.2. The Matching Model of the HCP. Regard paths of length n + 1 that start and stop at the
same vertex and pass through every vertex, as directed graphs on n + 1 vertices. For undirected graphs,
every cycle is accompanied by a counter-directed companion cycle. No matter that G is Hamiltonian or nonHamiltonian, assign vertex n + 1 as the origin and terminal vertex for all cycles, and assign each directed
Hamilton cycle to be in correspondence with each nxn permutation matrix P where pu,i = 1 if and only if
the ith arc in a cycle enters vertex u. We encode each cycle as a permutation of vertex labels. For example,
the path sequence {6,2,4,3,5,1,6} is code for the first arc enters vertex 2, the second arc enters vertex 4
and so on. Since pn+1,n+1 = 1 for all cycles by definition, it’s sufficient to code cycles as nxn permutation
matrices. Note that an arc is directed, and an edge is undirected i.e. the pair of arcs (u, i) & (i, u) is the
edge (u, i). Unless otherwise stated, all graphs are simple, connected and 3-regular.
We next encode graph instance G by examining G’s adjacency matrix, adding to E all pairs of components of P , {pu,i , pv,j } that encode paths of length j − i (j > i) from vertex u to vertex v in cycles not in G.
This encodes precisely the set of cycles not in G i.e. every cycle not in G uses at least one arc not in G.
See the algorithm (How to Initialize Exclusion Set E) below and recall that G is connected. For arc
(u, v) not in G we can assign {pu,i , pv,i+1 } ∈ E. But we also compute additional {pu,i , pv,i+m } whenever it’s
possible to account for no paths of length m in G, from vertex u to vertex v. We do this by implementing
Dijkstras algorithm with equally weighted arcs to find minimal length paths between all pairs of vertices,
coded to return m = n + 1 if no path exists. We account for all paths of length one not in G (arcs not in
G), and, all paths of length two not in G by temporarily deleting the arc between adjacent vertices.
Begin as follows. If u is adjacent to v then temporarily delete arc (u,v) and apply Dijkstras algorithm to
discover a minimal path of length m > 1, a simple ‘speed-up’. No paths of length k can exist, k = 1, ..., m − 1
and {pu,i , pv,i+k } are discovered that 1) for k = 1 and u not adjacent to v correspond with arcs in cycles not
in G, and 2) for k > 1 correspond with paths of length k in cycles not in G. Accounting for all arcs not in G
is sufficient to model precisely all cycles not in G, and we account for paths in cycles not in G to bolster E.
Two special cases arise. Case 1. Last arc in cycle: Recall that every n + 1th arc in a cycle enters vertex
n + 1 by definition. Therefore observe arcs (u,n + 1) not in G, temporarily deleted or otherwise, noting
how corresponding sets of cycles not in G can be encoded by permutation matrices for which the nth arc
in a cycle enters vertex u i.e. pu,n = 1. This is the case for k=1, and u not adjacent to v when Dijkstras
algorithm returns m = 2. If Dijkstras algorithm returns m = 3, then again for k=1 and if u is not adjacent
to v set pu,n = 1, and for k=2, no paths of length two exist and these sets of cycles not in G can be encoded
by permutation matrices for which the n − 1th arc in a cycle enters vertex u i.e. pu,n−1 = 1. Continuing in
5
this way, encode all possible n + 1 − k th arcs in cycles not in G, in paths of length k not in G, to enter vertex
u i.e. pu,n+1−k = 1, k = 1, 2, ..., m − 1. Case 2. First arc in cycle: Recall every first arc in every cycle exits
vertex n + 1. Observe and code all arcs (n + 1,v) in cycles not in G, in paths of length k not in G by coding
all possible k th arcs to enter vertex v i.e. pv,k =1, k = 1, 2, ..., m − 1.
For the general case, an exclusion set can be constructed by noting that a cycle not in G uses at least
one arc not in G e.g. (u, v). The complete set of permutation matrices corresponding to these cycles not
in G are characterized by {{pu,l , pv,l+1 }, l = 1, 2, ..., n − 1}, added to E. By indexing l, arc (u, v) can play
the role of (n-1) sequence positions in disjoint sets of cycles not in G. Considering all O(n2 ) arcs not in G,
each playing the role of all O(n) possible sequence positions, it’s possible to construct the set of permutation
matrices corresponding to the set of cycles not in G accounted by the union of O(n3 ) {pu,l , pv,l+1 }, added to
E. We generalize this idea via Dijkstras algorithm and account for some sets of paths of length k not in G.
Recall that G is strongly connected. But if an arc is temporarily deleted, it’s possible for no path to
exist between a given pair of vertices. This useful information indicates that an arc is essential under the
assumption of the existence of a Hamilton cycle that uses this arc. In case 1, this implies that a particular
pu,n is necessary, and by integrality must be at unit level in every assignment of variables, assuming the graph
is Hamiltonian (until deduced otherwise, if ever). Thus all other P ’s in the same row (and column) can be
set at zero level. This is accounted for when we initialize E. Recall that m = n + 1 in the case that Dijkstras
algorithm returns no minimal path. The k loop appends the necessary set of {pv,j , pu,n+1−k } to E effectively
setting variables in blocks (u,1,) through (u,n − 1) at zero level. When implemented in the algorithm, pu,n
must attain unit level via double stochastity, and this implies that the other P ’s in the same column are
deduced to be at zero level. Similarily for Case 2. In the general case, it’s also possible for no path to exist
between a given pair of vertices (u,v) (when an arc is temporarily deleted). Under the assumption of the
existence of a Hamilton cycle, this arc is essential and can play the role of sequence position 2 through n-1
and so in each case, all complementary row (and column) {pu,i , pv,j } are assigned to E. When implemented,
a single {pu,i , pv,j } variable remains in each row and therefore is equated with that block’s pu,i variable via
‘scaled’ double stochastity within the block i.e. rows and columns in the block sum to pu,i . Complementary
{pu,i , pv,j } variables in the corresponding column are therefore set 0 in each block. Thus essential arcs also
contribute to new information by adding their complementary row / column {pu,i , pv,j } to E.
Finally, encode E into matrix Q i.e. assign qu,i,v,j = 0 for each {pu,i , pv,j } ∈ E, and then create V .
6
Input: {Arc Adjacency matrix for G}
Output: {E}
E ← ∅;
Case 1: for u = 1, 2, ..., n do
Arc ← G(u, n + 1); G(u, n + 1) ← 0; m ← DijkstrasAlgorithm(G,u,n + 1);
for k = Arc + 1, Arc + 2, ..., m − 1 do
E ← E ∪ {pv,j , pu,n+1−k }, v = 1, 2, ..., n, v 6= u; j = 1, 2, ..., n, j 6= n + 1 − k;
end
G(u, n + 1) ← Arc;
end
Case 2: for v = 1, 2, ..., n do
Arc ← G(n + 1, v); G(n + 1, v) ← 0; m ← DijkstrasAlgorithm(G,n + 1,v);
for k = Arc + 1, Arc + 2, ..., m − 1 do
E ← E ∪ {pv,k , pu,i }, u = 1, 2, ..., n, u 6= v; i = 1, 2, ..., n, i 6= k;
end
G(n + 1, v) ← Arc;
end
General Case. for u = 1, 2, ..., n do
for v = 1, 2, ..., n; v 6= u do
Arc ← G(u, v); G(u, v) ← 0; m ← DijkstrasAlgorithm(G,u,v);
for k = Arc + 1, Arc + 2, ..., m − 1 do
E ← E ∪ {pu,l , pv,l+k }, l = 1, 2, ..., n − k;
end
G(u, v) ← Arc;
end
end
EXIT {E} ;
Algorithm 2: How to Initialize Exclusion Set E
5. Empirical Results and Two Conjectures. Table 5.1 below lists some details of 25 applications
(all 3-regular graphs) of the algorithm. Table 5.2 below lists some details of 20 applications (mostly 3regular graphs) of an earlier version of the matching based closure algorithm called the WCA∗∗ [12] (a
subset from over 2,100 applications). For both algorithms, all of the graphs are decided non-Hamiltonian
and no application of either algorithm to any other graphs failed that are not reported.
5.1. Empirical Results. In both tables, heading # pu,i (|V |) is the count of non-zero pu,i variables
and the size of initial available set V (the number of non-zero qu,i,v,j components in Q) after initializing E,
before implementing the algorithm. Note pu,i = qu,i,u,i . We only count qu,i,v,j i < j (distinct qu,i,v,j ).
In Table 5.2, heading |open(V )| ≤ refers to an upper bound on |open(V )| for 11 selected graphs, each
modified to include the cycle 1 − 2 − ... − n − (n + 1) − 1, simply to observe open(V ). Two of these graphs
are also hypohamiltonian. The count in parentheses is an upper bound on |open(V )| after removing a vertex
and re-running the WCA.
5.2. Two Conjectures.
2
2
distinct {pu,i , pv,j } ∈ E exists
Conjecture 5.1. Polynomial sized proof of membership of all n (n−1)
2
for all simple, connected, 3-regular, non-Hamiltonian graphs.
Conjecture 5.2. Triple overlay matching based closure deduces open(V ) = ∅ for all simple, connected,
3-regular, non-Hamiltonian graphs.
∗∗ The
WCA is a breadth-first closure, exhausting the middle v, j loop before returning to label CONTINUE TRIPLE
CLOSURE. It’s followed by triple closure, also applied breadth-first, exhausting the interior w, k loop before returning to label
TRIPLE CLOSURE. Many more applications of Boolean closure across all of Q at many more intermediate steps are also
implemented, unlike triple overlay matching based closure as we have presented (although these checks can also be included).
Block overlays are also restricted to be of the form Q(u, i) and Q(v, j), i < j. In this way we can solve problems in the 50-100
vertex range. The WCA is designed to be parallelized, and the FORTRAN code is written for distributed computing.
7
Table 5.1
Applications of the Triple Overlay Matching Based Closure Algorithm
Name of Graph
# Vertices (All 3-Regular Graphs)
# pu,i (|V |)
10
12, 20, 28
12
18, 26
18
22
24
26
26
26
28
30
57 (858)
87 (2,199), 271 (26,380), 567 (126,128)
87 (2,257)
223 (16,630), 495 (88,968)
219 (16,262)
345 (43,719)
419 (65,711)
≈ 500 (≈ 1, 000, 000)
497 (89,188)
505 (96,175), 509 (95,543)
583 (129,018)
691 (187,398)
Petersen Snark
3 Flower Snarks
Tietzs Snark
2 Blanusa Snarks
House of Graphs[1]
A Loupekine Snark
A Goldberg Snark
10 House of Graphs[1]
Jan Goedgebeur Snark
2 Celmins-Swart Snarks
House of Graphs[1]
A Double Star Snark
Table 5.2
Applications of A Matching Based Closure Algorithm (WCA) [12]
Name of Graph
Petersen Snark
Herschel Graph
A Kleetope
1
Matteo[4]††
1
Coxeter
House of Graphs #3337 Snark[1]
Zamfirescu Snark
1
Barnette-Bosk-Lederberg††
1
A Hypohamiltonian
1
Tutte††
A Grinberg Graph
1
Georges††
Szekeres Snark
Watkins Snark
1
Ellingham-Horton
Thomassen
Meredith
A Flower Snark
1
Horton††
A Goldberg Snark
# Vertices (Edges)
# pu,i (|V |)
|open(V )| ≤
10 (15)
11 (18)
14 (36)
20 (30)
28 (42)
34 (51)
36 (54)
38 (57)
45 (70)
46 (69)
46 (69)
50 (75)
50 (75)
50 (75)
54 (81)
60 (99)
70 (140)
76 (114)
96 (144)
104 (156)
57 (858)
No 2-Factor
147 (8,166)
275 (26,148)
597 (136,599)
897 (308,234)
983 (363,987)
1077 (440,318)
1,656 (1,109,738)
1,649 (1,060,064)
1,737 (1,204,722)
2037 (1,701,428?)
2045 (1,718,336)
2051(1,708,987)
2,315 (2,135,948)
3,105 (4,071,600)
4,221 (7,526,996)
4,851 (9,720,420)
8,205 (29,057,118)
9,339 (37,802,124)
792
1,980
5,809
27,093
135,453(1,241)2
335,697
7,749
96,834
296,668 (29,724)2
382,400
- Not run yet - Not run yet - Not run yet - Not run yet 1,045,041
- Not run yet - Not run yet - Not run yet - Not run yet - Not run yet -
1
Simple, connected, 3-regular, and 3-colourable.
2
Hypohamiltonian. Confirmed existence of non-empty open(V ) after removing a vertex and re-running the WCA.
†† Historical Note: Ignoring the planarity condition on Tait’s conjecture, the Matteo graph [4] is the smallest non-planar
counterexample, while the Barnette-Bosk-Lederberg graph [13] is the smallest planar, 3-colourable, 3-connected counterexample to Tait’s conjecture (Tutte’s graph is a larger counterexample). We also note that the Georges graph is the smallest
counterexample to Tutte’s conjecture, and Horton’s graph was the first counterexample to Tutte’s conjecture.
8
6. Discussion.
6.1. About Practical Generalizations of the Algorithm. The algorithm can be designed to invoke
arbitrary levels of overlay i.e. adaptive strategies that change the level of overlay if more depth is desired
/ needed to deduce variables at zero level. But in order to make use of increased overlay, it’s necessary to
add more variables to retain information about tests for matching. For example, if we create a quadrupal
overlay version of the algorithm, we then introduce {0, 1} {pu,i , pv,j , pw,k } variables and redefine system 1
and matrix Q in terms of triply nested Birkhoff polyhedra. See the discussion in [7] for a description of these
polyhedra as feasible regions of LP formulations (relaxed IPs). There exists a sequence of feasible regions in
correspondence with increasing levels of nested Birkhoff polyhedra whose end feasible region is the convex
hull of the set of integer extrema of system 1. See [2] for a discussion of facet-inducing inequalities.
The term closure has so far been reserved for deducing variables added to E by invoking the algorithm.
But other polynomial time techniques can be used to deduce variables at zero level. For example, prior to
matching, we could implement LP and maximize each variable in system 1, and if its maximum is less than
unit level, the variable can be set zero. In our implementation, we use Boolean closure. See [12] for more
details. We also note there exist entire conferences devoted to Matching Under Preferences [11]. Perhaps
many more innovative heuristics exist and can be included in the algorithm.
The algorithm is designed for parallel processing. Each qu,i,v,j variable not yet deduced zero can be
tested independent of the others by making a copy of matrix Q and implementing the algorithm. If an
independent process deduces a qu,i,v,j variable at zero level, simply update the corresponding qu,i,v,j variable
in each Q across all processes.
For some applications, there exist model specific dependencies between variables i.e. undirected HCP
implies {pu,i , pv,j } = 0 if and only if {pu,n+1−i , pv,n+1−j } = 0. In this way we account for companion cycles.
6.2. About Study of the Algorithm. Exclusion set E is the focus of study. We propose to classify
different E by the pattern that remains in matrix Q after exit from the algorithm (up to isomorphism) i.e.
Q covers the set of all possible solutions to the IP. It would be useful to know what kinds of E cause the
algorithm to generate Q as a minimal cover, since it then follows that the algorithm would decide feasibility
of the IP. Even if there exist classes of E for which infeasible IPs provably exit the algorithm infeasible, no
matter that Q is or is not a minimal cover, it still follows that the algorithm decides feasibility of the IP.
We plan to investigate counterexamples via the matching model for HCP. Graph C7-21 (not 3-regular)
fails an earlier version of the algorithm [12]. We will convert [3] C7-21 and study it as instance of 3HCP.
6.3. Two More Matching Model Applications for Input to the Algorithm. We now present two
more matching models as applications for the algorithm‡‡ . Q’s components no longer have the interpretation
as sequenced arcs in a cycle. Instead, let Q be an m2 xm2 block permutation matrix, whose blocks are mxm
permutation matrices P . We note from [7] that F is a subgraph of G if and only if there exists permutation
matrix P such that P T GP covers F (and we add) if and only if Q·g covers f, where f and g are column vectors
of adjacency matrices F and G formatted as {F (1, 1), F (1, 2), ..., F (1, m), F (2, 1), F (2, 2), ..., F (m, m)} and
{G(1, 1), G(1, 2), ..., G(1, m), G(2, 1), G(2, 2), ..., G(m, m)}.
We now model both the graph and subgraph isomorphism decision problems as matching models, the
single difference being that in the case of graph isomorphism, more information appears to be added to
E. First note that Q · g covers f means Q · g is required to place ones in the same positions as those of
f. So for each of these equations, a subset of row components sum to one implying that the complement
row components must therefore all be set at zero level. Add them to E. This completes the subgraph
isomorphism matching model and only part of the graph isomorphism model. For graph isomorphism, cover
means equality. The remaining equations to be satisfied are those for which Q · g is required to place zeroes
in the same positions as those of f. So for each of these equations, a subset of row components sum to zero
implying that these row components must therefore all be set at zero level. Add them to E. This completes
the graph isomorphism matching model.
6.4. Other Applications of the Algorithm. We originally intended for the algorithm to decide
feasibility of a matching model. When it decides infeasibility, the algorithm has served its purpose. Otherwise
it’s not known if the model is feasible or infeasible. We note that open(V ) is a refined cover of possible
solutions to the IP and we believe that this is useful. We propose that the algorithm can be developed as
‡‡ See
[8] for more information about these modelling techniques.
9
part of other search based algorithms, either to provide refined information prior to a search, or incorporated
and updated alongside a search based algorithm to provide more information during a search.
There is one last thought about an academic use for the algorithm. Suppose we are given a correctly
guessed infeasible IP, and the algorithm exits undecided. We can attribute the failure to E as lacking the
necessary / right kind of {pu,i , pv,j } that could induce closure. We could then theoretically augment E
with additional {pu,i , pv,j } until we deduce infeasibility, and discover extra information needed to generate
open(V ) = ∅. So for application, when the algorithm gets stuck and open(V ) 6= ∅, simply augment E with
additional {pu,i , pv,j } ∈ open(V ), and test if open(V ) becomes empty. While it might be difficult to guess
minimal sized sets of additional {pu,i , pv,j }, if they can be guessed, we will then have articulated what critical
information is needed to solve the problem. Of course it’s not known if these additional {pu,i , pv,j } can be
efficiently computed or validated as members in E. See conjecture 5.1.
7. Acknowledgements and Dedication. Thank you to: Adrian Lee for preparing and running some
of the examples presented in Tables 5.1 and 5.2; Nicholas Swart for testing and implementing 2000 31-vertex
non-Hamiltonian graphs in 2013; Catherine Bell for suggestions and contributions early on in this project.
We dedicate this paper to the late Pal Fischer (16/11/2016). For Ted, Pal was a colleague, co-author
and friend. For myself (Gismondi), Pal taught me analysis, an understanding of convex polyhedra, and later
became a colleague. Ted and I both already miss him very much.
REFERENCES
[1] G. Brinkmann, K. Coolsaet, J. Goedgebeur, and H. Melot, House of Graphs: a database of interesting graphs,
Discrete Applied Mathematics. Available at http://hog.grinvin.org, 161 (2013), pp. 311–314.
[2] M. Demers and S.J. Gismondi, Enumerating facets of Q0/1 , Util. Math., 77 (2008), pp. 125–134.
[3] V. Ejov, M. Haythorpe, and S. Rossomakhine, A Linear-size Conversion of HCP to 3HCP, Australasian Journal of
Combinatorics, 62 (2015), pp. 45–58.
[4] Mathematics Stack Exchange, graph6 string:
Sspp?wc g ?w?o???i???ag?bo?g. Retrieved May 7, 2016,
http://math.stackexchange.com/questions/367671/smallest-nonhamiltonian-3-connected-graph-with-chromatic-index-3,
(2016).
[5] J.A. Filar, M. Haythorpe, and S. Rossomakhine, A new heuristic for detecting non-hamiltonicity in cubic graphs,
Computers & Operations Research, 64 (2015), pp. 283–292.
[6] M.R. Gary, D.S. Johnson, and R.E. Tarjan, The planar hamilton circuit problem is np-complete, SIAM J. Comput.,
5 (1976), pp. 704–714.
[7] S.J. Gismondi, Subgraph isomorphism and the Hamilton tour decision problem using a linearized form of P GP t , Util.
Math., 76 (2008), pp. 229–248.
[8]
, Modelling decision problems via birkhoff polyhedra, Journal of Algorithms and Computation, 44 (2013), pp. 61–81.
[9] S.J. Gismondi and E.R. Swart, A model of the coNP-complete non-Hamilton tour decision problem, Math. Prog. Ser.
A, 100 (2004), pp. 471–483.
[10] M. Haythorpe, FHCP challenge set. Retrieved July 16, 2016, http://fhcp.edu.au/fhcpcs, (2015).
[11] Microsoft Research Lab. New England, Match-up 2017, Cambridge, MA, USA, https://www.microsoft.com/enus/research/event/match-up-2017/, (2017).
[12] E.R. Swart, S.J. Gismondi, N.R. Swart, C.E. Bell, and A. Lee, Deciding graph non-hamiltonicity via a closure
algorithm, Journal of Algorithms and Computation, 48 (2016), pp. 1–35.
[13] Wolfram
Math
World,
Barnette-bosk-lederberg
graph.
Retrieved
May
7,
2016,
http://mathworld.wolfram.com/Barnette-Bosak-LederbergGraph.html, (2016).
10
| 8cs.DS
|
arXiv:1706.05744v2 [cs.LG] 4 Nov 2017
Learning Hierarchical Information Flow
with Recurrent Neural Modules
Danijar Hafner ∗
Google Brain
[email protected]
Alex Irpan
Google Brain
[email protected]
James Davidson
Google Brain
[email protected]
Nicolas Heess
Google DeepMind
[email protected]
Abstract
We propose ThalNet, a deep learning model inspired by neocortical communication
via the thalamus. Our model consists of recurrent neural modules that send features
through a routing center, endowing the modules with the flexibility to share features
over multiple time steps. We show that our model learns to route information
hierarchically, processing input data by a chain of modules. We observe common
architectures, such as feed forward neural networks and skip connections, emerging
as special cases of our architecture, while novel connectivity patterns are learned
for the text8 compression task. Our model outperforms standard recurrent neural
networks on several sequential benchmarks.
1
Introduction
Deep learning models make use of modular building blocks such as fully connected layers, convolutional layers, and recurrent layers. Researchers often combine them in strictly layered or task-specific
ways. Instead of prescribing this connectivity a priori, our method learns how to route information as
part of learning to solve the task. We achieve this using recurrent modules that communicate via a
routing center that is inspired by the thalamus.
Warren McCulloch and Walter Pitts invented the perceptron in 1943 as the first mathematical model
of neural information processing [24], laying the groundwork for modern research on artificial neural
networks. Since then, researchers have continued looking for inspiration from neuroscience to identify
new deep learning architectures [11, 13, 18, 34].
While some of these efforts have been directed at learning biologically plausible mechanisms in an
attempt to explain brain behavior, our interest is to achieve a flexible learning model. In the neocortex,
communication between areas can be broadly classified into two pathways: Direct communication
and communication via the thalamus [31]. In our model, we borrow this latter notion of a centralized
routing system to connect specializing neural modules.
In our experiments, the presented model learns to form connection patterns that process input
hierarchically, including skip connections as known from ResNet [12], Highway networks [32], and
DenseNet [15] and feedback connections, which are known to both play an important role in the
neocortex and improve deep learning [7, 22]. The learned connectivity structure is adapted to the
task, allowing the model to trade-off computational width and depth. In this paper, we study these
properties with the goal of building an understanding of the interactions between recurrent neural
modules.
∗
Work done during an internship with Google Brain.
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
x1
x2
x3
f1
f1
f1
f2
f2
x1
f1
f4
Φ
f3
Φ1
y4
y3
(a) Module f 1 receives the task input,
f 2 can be used for side computation, f 3
is trained on an auxiliary task, and f 4
produces the output for the main task.
f2
Φ2
f2
f3
f3
f3
y1
y2
y3
(b) Computation of 3 modules unrolled in time. One possible path of
hierarchical information flow is highlighted in green. We show that
our model learns hierarchical information flow, skip connections
and feedback connections in Section 4.
Figure 1: Several modules share their learned features via a routing center. Dashed lines are used for
dynamic reading only. We define both static and dynamic reading mechanisms in Section 2.2.
Section 2 defines our computational model. We point out two critical design axes, which we explore
experimentally in the supplementary material. In Section 3 we compare the performance of our model
on three sequential tasks, and show that it consistently outperforms multi-layer recurrent networks.
In Section 4, we apply the best performing design to a language modeling task, where we observe
that the model automatically learns hierarchical connectivity patterns.
2
Thalamus Gated Recurrent Modules
We find inspiration for our work in the neurological structure of the neocortex. Areas of the neocortex
communicate via two principal pathways: The cortico-cortico-pathway comprises direct connections
between nuclei, and the cortico-thalamo-cortico comprises connections relayed via the thalamus.
Inspired by this second pathway, we develop a sequential deep learning model in which modules
communicate via a routing center. We name the proposed model ThalNet.
2.1
Model Definition
Our system comprises a tuple of computation modules F = (f 1 , · · · , f I ) that route their respective
features into a shared center vector Φ. An example instance of our ThalNet model is shown in
Figure 1a. At every time step t, each module f i reads from the center vector via a context input cit
and an optional task input xit . The features φit = f i (cit , xit ) that each module produces are directed
into the center Φ.2 Output modules additionally produce task output from their feature vector as a
function oi (φi ) = y i .
All modules send their features to the routing center, where they are merged to a single feature vector
Φt = m(φ1t , · · · , φIt ). In our experiments, we simply implement m as the concatenation of all φi .
At the next time step, the center vector Φt is then read selectively by each module using a reading
mechanism to obtain the context input cit+1 = ri (Φt , φit ).3 This reading mechanism allows modules
to read individual features, allowing for complex and selective reuse of information between modules.
The initial center vector Φ0 is the zero vector.
2
In practice, we experiment with both feed forward and recurrent implementations of the modules f i . For
simplicity, we omit the hidden state used in recurrent modules in our notation.
3
The reading mechanism is conditioned on both Φt and φit separately as the merging does not preserve φit in
the general case.
2
y
φ1
φ2
φ3
φ4
Φ
×
x
f2
=
c
Figure 2: The ThalNet model from the perspective of a single module. In this example, the module
receives input xi and produces features to the center Φ and output y i . Its context input ci is determined
as a linear mapping of the center features from the previous time step. In practice, we apply weight
normalization to encourage interpretable weight matrices (analyzed in Section 4).
In summary, ThalNet is governed by the following equations:
Module features:
φit = f i (cit , xit )
(1)
Module output:
yti
=
(2)
Center features:
Φt
i
ct+1
=
Read context input:
=
i
o (φit )
m(φ1t , · · · , φIt )
ri (Φt , φit )
(3)
(4)
The choice of input and output modules depends on the task at hand. In a simple scenario (e.g., single
task), there is exactly one input module receiving task input, some number of side modules, and
exactly one output module producing predictions. The output modules get trained using appropriate
loss functions, with their gradients flowing backwards through the fully differentiable routing center
into all modules.
Modules can operate in parallel as reads target the center vector from the previous time step. An
unrolling of the multi-step process can be seen in Figure 1b. This figure illustrates the ability to
arbitrarily route between modules between time steps This suggest a sequential nature of our model,
even though application to static input is possible by allowing observing the input for multiple time
steps.
We hypothesize that modules will use the center to route information through a chain of modules
before producing the final output (see Section 4). For tasks that require producing an output at every
time step, we repeat input frames to allow the model to process through multiple modules first, before
producing an output. This is because communication between modules always spans a time step.4
2.2
Reading Mechanisms
We now discuss implementations of the reading mechanism ri (Φ, φi ) and modules f i (ci , xi ), as
defined in Section 2.1. We draw a distinction between static and dynamic reading mechanisms for
ThalNet. For static reading, ri (Φ) is conditioned on independent parameters. For dynamic reading,
ri (Φ, φi ) is conditioned on the current corresponding module state, allowing the model to adapt its
connectivity within a single sequence. We investigate the following reading mechanisms:
• Linear Mapping. In its simplest form, static reading consists of a fully connected layer
r(Φ, ·) = W Φ with weights W ∈ R|c|×|Φ| as illustrated in Figure 2. This approach
performs reasonably well, but can exhibit unstable learning dynamics and learns noisy
weight matrices that are hard to interpret. Regularizing weights using L1 or L2 penalties
does not help here since it can cause side modules to not get read from anymore.
• Weight Normalization. We found linear mappings with weight normalization [28] paW
rameterization to be effective. For this, the context input is computed as r(Φ, ·) = β |W
|Φ
with scaling factor β ∈ R, weights W ∈ R|c|×|Φ| , and the Euclidean matrix norm |W |.
4
Please refer to Graves [8] for a study of a similar approach.
3
Normalization results in interpretable weights since increasing one weight pushes other, less
important, weights closer to zero, as demonstrated in Section 4.
• Fast Softmax. To achieve dynamic routing, we condition the reading weight matrix on
the current module features φi . This can be seen as a form of fast weights, providing
a biologically plausible method for attention [2, 29]. We then apply softmax normalization to the computed weights so that each element of the context is computed as a
weighted average over center elements, rather than just a weighted sum. Specifically,
P|Φ|
r(Φ, φ)(j) = e(W φ+b)(j) / k=1 e(W φ+b)(jk) Φ with weights W ∈ R|φ|×|Φ|×|c| , and biases b ∈ R|Φ|×|c| . While this allows for a different connectivity pattern at each time step, it
introduces |φi + 1| × |Φ| × |ci | learned parameters per module.
• Fast Gaussian. As a compact parameterization for dynamic routing, we consider choosing each context element as a Gaussian weighted average of Φ, with only mean and variance vectors learned conditioned on φi . The context input is computed as r(Φ, φ)(j) =
f (1, 2, · · · , |Φ|)|(W φ + b)(j) , (U φ + d)(j) Φ with weights W, U ∈ R|c|×|φ| , biases
b, d ∈ R|c| , and the Gaussian density function f (x|µ, σ 2 ). The density is evaluated for each
index in Φ based on its distance from the mean. This reading mechanism only requires
|φi + 1| × 2 × |ci | parameters per module and thus makes dynamic reading more practical.
Reading mechanisms could also select between modules on a high level, instead of individual feature
elements. We do not explore this direction since it seems less biologically plausible. Moreover,
we demonstrate that such knowledge about feature boundaries is not necessary, and hierarchical
information flow emerges when using fine-grained routing (see Figure 4). Theoretically, this also
allows our model to perform a wider class of computations.
3
Performance Comparison
We investigate the properties and performance of our model on several benchmark tasks. First, we
compare reading mechanisms and module designs on a simple sequential task, to obtain a good
configuration for the later experiments. Please refer to the supplementary material for the precise
experiment description and results. We find that the weight normalized reading mechanism provides
best performance and stability during training. We will use ThalNet models with four modules of
configuration for all experiments in this section. To explore the performance of ThalNet, we now
conduct experiments on three sequential tasks of increasing difficulty:
• Sequential Permuted MNIST. We use images from the MNIST [21] data set, the pixels
of every image by a fixed random permutation, and show them to the model as a sequence
of rows. The model outputs its prediction of the handwritten digit at the last time step, so
that it must integrate and remember observed information from previous rows. This delayed
prediction combined with the permutation of pixels makes the task harder than the static
image classification task, with a multi-layer recurrent neural network achieving ~65 % test
error. We use the standard split of 60,000 training images and 10,000 testing images.
• Sequential CIFAR-10. In a similar spirit, we use the CIFAR-10 [19] data set and feed
images to the model row by row. We flatten the color channels of every row so that the
model observes a vector of 96 elements at every time step. The classification is given after
observing the last row of the image. This task is more difficult than the MNIST task, as
the image show more complex and often ambiguous objects. The data set contains 50,000
training images and 10,000 testing images.
• Text8 Language Modeling. This text corpus consisting of the first 108 bytes of the
English Wikipedia is commonly used as a language modeling benchmark for sequential
models. At every time step, the model observes one byte, usually corresponding to 1
character, encoded as a one-hot vector of length 256. The task it to predict the distribution
of the next character in the sequence. Performance is measured in bits per character (BPC)
PN
computed as − N1 i=1 log2 p(xi ). Following Cooijmans et al. [4], we train on the first
90% and evaluate performance on the following 5% of the corpus.
For the two image classification tasks, we compare variations of our model to a stacked Gated
Recurrent Unit (GRU) [3] network of 4 layers as baseline. The variations we compare are different
4
Sequential CIFAR-10 Testing
Sequential Permuted MNIST Testing
0.6
20
40
60
Epochs
80
0.45
ThalNet GRU-FF
ThalNet FF-GRU
ThalNet GRU
GRU Baseline
ThalNet FF-GRU-FF
ThalNet FF
0.40
0.35
100
Sequential Permuted MNIST Training
Bits per character (BPC)
ThalNet FF-GRU-FF
ThalNet FF-GRU
ThalNet FF
ThalNet GRU-FF
ThalNet GRU
GRU Baseline
20
40
60
Epochs
80
100
Sequential CIFAR-10 Training
1.0
0.8
0.7
ThalNet FF-GRU
ThalNet FF-GRU-FF
ThalNet GRU-FF
ThalNet FF
ThalNet GRU
GRU Baseline
0.6
0.5
20
40
60
Epochs
80
100
Accuracy (%)
0.9
0.55
0.50
GRU Baseline
ThalNet GRU
ThalNet FF-GRU
ThalNet GRU-FF
ThalNet FF-GRU-FF
ThalNet FF
0.45
0.40
20
40
60
Epochs
80
1.46
GRU (1 step)
GRU (2 steps)
ThalNet (2 steps)
1.44
1.42
1.40
10
20
30
Epochs
40
50
Text8 Language Modeling Training
0.60
100
Bits per character (BPC)
0.7
0.5
Accuracy (%)
Text8 Language Modeling Evaluation
0.50
0.8
Accuracy (%)
Accuracy (%)
0.9
1.45
1.40
1.35
1.30
1.25
1.20
ThalNet (2 steps)
GRU (1 step)
GRU (2 steps)
10
20
30
Epochs
40
50
Figure 3: Performance on the permuted sequential MNIST, sequential CIFAR, and text8 language
modeling tasks. The stacked GRU baseline reaches higher training accuracy on CIFAR, but fails to
generalize well. On both tasks, ThalNet clearly outperforms the baseline in testing accuracy. On
CIFAR, we see how recurrency within the modules speeds up training. The same pattern is shows for
the text8 experiment, where ThalNet using 12M parameters matches the performance of the baseline
with 14M parameters. The step number 1 or 2 refers to repeated inputs as discussed in Section 2. We
had to smooth the graphs using a running average since the models were evaluated on testing batches
on a rolling basis.
choices of feed-forward layers and GRU layers for implementing the modules f i (ci , xi ): We test with
two fully connected layers (FF), a GRU layer (GRU), fully connected followed by GRU (FF-GRU),
GRU followed by fully connected (GRU-FF), and a GRU sandwiched between fully connected layers
(FF-GRU-FF).5 For all models, we pick the largest layer sizes such that the number of parameters does
not exceed 50,000. Training is performed for 100 epochs on batches of size 50 using RMSProp [33]
with a learning rate of 10−3 .
For language modeling, we simulate ThalNet for 2 steps per token, as described in Section 2 to allow
the output module to read information about the current input before making its prediction. Note
that on this task, our model uses only half of its capacity directly, since its side modules can only
integrate longer-term dependencies from previous time steps. We run the baseline once without extra
steps and once with 2 steps per token, allowing it to apply its full capacity once and twice on each
token, respectively. This makes the comparison a bit difficult, but only by favouring the baseline.
This suggests that architectural modifications, such as explicit skip-connections between modules,
could further improve performance.
The Text8 task requires larger models. We train ThalNet with 4 modules of a size 400 feed forward
layer and a size 600 GRU layer each, totaling in 12 million model parameters. We compare to
a standard baseline in language modeling, a single GRU with 2000 units, totaling in 14 million
parameters. We train on batches of 100 sequences, each containing 200 bytes, using the Adam
optimizer [17] with a default learning rate of 10−3 . We scale down gradients exceeding a norm of 1.
Results for 50 epochs of training are shown in Figure 3. The training took about 8 days for ThalNet
with 2 steps per token, 6 days for the baseline with 2 steps per token, and 3 days for the baseline
without extra steps.
Figure 3 shows the training and testing and training curves for the three tasks described in this section.
ThalNet outperforms standard GRU networks in all three tasks. Interestingly, ThalNet experiences a
5
Note that the modules require some amount of local structure to allow them to specialize. Implementing the
modules as a single fully connected layer recovers a standard recurrent neural network with one large layer.
5
much smaller gap between training and testing performance than our baseline – a trend we observed
across all experimental results.
On the Text8 task, ThalNet scores 1.39 BPC using 12M parameters, while our GRU baseline scores
1.41 BPC using 14M parameters (lower is better). Our model thus slightly improves on the baseline
while using fewer parameters. This result places ThalNet in between the baseline and regularization
methods designed for language modeling, which can also be applied to our model. The baseline
performance is consistent with published results of LSTMs with similar number of parameters [20].
We hypothesize the information bottleneck at the reading mechanism acting as an implicit regularizer
that encourages generalization. Compared to using one large RNN that has a lot of freedom of
modeling the input-output mapping, ThalNet imposes local structure to how the input-output mapping
can be implemented. In particular, it encourages the model to decompose into several modules that
have stronger intra-connectivity than extra-connectivity. Thus, to some extend every module needs to
learn a self-contained computation.
4
Hierarchical Connectivity Patterns
Using its routing center, our model is able to learn its structure as part of learning to solve the
task. In this section, we explore the emergent connectivity patterns. We show that our model learns
to route features in hierarchical ways as hypothesized, including skip connections and feedback
connections. For this purpose, we choose the text8 corpus, a medium-scale language modeling
benchmark consisting of the first 108 bytes of Wikipedia, preprocessed for the Hutter Prize [23]. The
model observes one one-hot encoded byte per time step, and is trained to predict its future input at
the next time step.
We use comparably small models to be able to run experiments quickly, comparing ThalNet models
of 4 FF-GRU-FF modules with layer sizes 50, 100, 50 and 50, 200, 50. Both experiments use
weight normalized reading. Our focus here is on exploring learned connectivity patterns. We show
competitive results on the task using larger models in Section 3.
We simulate two sub time steps to allow for the output module to receive information of the current
input frame as discussed in Section 2. Models are trained for 50 epochs on batches of size 10
containing sequences of length 50 using RMSProp with a learning rate of 10−3 . In general, we
observe different random seeds converging to similar connectivity patterns with recurring elements.
4.1
Trained Reading Weights
Figure 4 shows trained reading weights for various reading mechanisms, along with their connectivity
graphs that were manually deduced.6 Each image represents a reading weight matrix for the modules
1 to 4 (top to bottom). Each pixel row shows the weight factors that get multiplied with Φ to produce
a single element of the context vector of that module. The weight matrices thus has dimensions of
|Φ| × |ci |. White pixels represent large magnitudes, suggesting focus on features at those positions.
The weight matrices of weight normalized reading clearly resemble the boundaries of the four
concatenated module features φ1 , · · · , φ4 in the center vector Φ, even though the model has no notion
of the origin and ordering of elements in the center vector.
A similar structure emerges with fast softmax reading. These weight matrices are sparser than
the weights from weight normalization. Over the course of a sequence, we observe some weights
staying constant while others change their magnitudes at each time step. This suggests that optimal
connectivity might include both static and dynamic elements. However, this reading mechanism
leads to less stable training. This problem could potentially alleviated by normalizing the fast weight
matrix.
With fast Gaussian reading, we see that the distributions occasionally tighten on specific features
in the first and last modules, the modules that receive input and emit output. The other modules
learn large variance parameters, effectively spanning all center features. This could potentially be
addressed by reading using mixtures of Gaussians for each context element instead. We generally
find that weight normalized and fast softmax reading select features with in a more targeted way.
6
Developing formal measurements for this deduction process seems beneficial in the future.
6
skip connection
skip connection
feedback
connection
x
1
2
3
4
y
x
skip connection
(a) Weight Normalization
1
2
3
4
y
feedback connection
(b) Fast Softmax
(c) Fast Gaussian
Figure 4: Reading weights learned by different reading mechanisms with 4 modules on the text8
language modeling task, alongside manually deducted connectivity graphs. We plot the weight
matrices that produce the context inputs to the four modules, top to bottom. The top images show
focus of the input modules, followed by side modules, and output modules at the bottom. Each pixel
row gets multiplied with the center vector Φ to produce one scalar element of the context input ci .
We visualize the magnitude of weights between the 5 % to the 95 % percentile. We do not include the
connectivity graph for Fast Gaussian reading as its reading weights are not clearly structured.
4.2
Commonly Learned Structures
The top row in Figure 4 shows manually deducted connectivity graphs between modules. Arrows
represent the main direction of information flow in the model. For example, the two incoming arrows
to module 4 in Figure 4a indicate that module 4 mainly attends to features produced by modules 1
and 3. We infer the connections from the larger weight magnitudes in the first and third quarters of
the reading weights for module 4 (bottom row).
A typical pattern that emerges during the experiments can be seen in the connectivity graphs of both
weight normalized and fast softmax reading (Figures 4a and 4b). Namely, the output module reads
features directly from the input module. This direction connection is established early on during
training, likely because this is the most direct gradient path from output to input. Later on, the side
modules develop useful features to support the input and output modules.
In another pattern, one module reads from all other modules and combines their information. In
Figure 4b, module 2 takes this role, reading from modules 1, 3, 4, and distributing these features via
the input module. In additional experiments with more than four modules, we observed this pattern to
emerge predominantly. This connection pattern provides a more efficient way of information sharing
than cross-connecting all modules.
Both connectivity graphs in Figure 4 include hierarchical computation paths through the modules.
They include learn skip connections, which are known to improve gradient flow from popular models
such as ResNet [12], Highway networks [32], and DenseNet [15]. Furthermore, the connectivity
graphs contain backward connections, creating feedback loops over two or more modules. Feedback
connections are known to play a critical role in the neocortex, which inspired our work [7].
5
Related Work
We describe a recurrent mixture of experts model, that learns to dynamically pass information between
the modules. Related approaches can be found in various recurrent and multi-task methods as outlined
in this section.
7
Modular Neural Networks. ThalNet consists of several recurrent modules that interact and exploit
each other. Modularity is a common property of existing neural models. [5] learn a matrix of
tasks and robot bodies to improve both multitask and transfer learning. [1] learn modules modules
specific to objects present in the scene, which are selected by an object classifier. These approaches
specify modules corresponding to a specific task or variable manually. In contrast, our model
automatically discovers and exploits the inherent modularity of the task and does not require a
one-to-one correspondence of modules to task variables.
The Column Bundle model [25] consists of a central column and several mini-columns around it.
While not applied to temporal data, we observe a structural similarity between our modules and the
mini-columns, in the case where weights are shared among layers of the mini-columns, which the
authors mention as a possibility.
Learned Computation Paths. We learn the connectivity between modules alongside the task.
There are various methods in the multi-task context that also connectivity between modules. Fernando
et al. [6] learn paths through multiple layers of experts using an evolutionary approach. Rusu et al. [27]
learn adapter connections to connect to fixed previously trained experts and exploit their information.
These approaches focus on feed-forward architectures. The recurrency in our approach allows for
complex and flexible computational paths. Moreover, we learn interpretable weight matrices that can
be examined directly without performing costly sensitivity analysis.
The Neural Programmer Interpreted presented by Reed and De Freitas [26] is related to our dynamic
gating mechanisms. In their work, a network recursively calls itself in a parameterized way to
perform tree-shaped computations. In comparison, our model allows for parallel computation
between modules and for unrestricted connectivity patterns between modules.
Memory Augmented RNNs. The center vector in our model can be interpreted as an external
memory, with multiple recurrent controllers operating on it. Preceding work proposes recurrent
neural networks operating on external memory structures. The Neural Turing Machine proposed by
Graves et al. [9], and follow-up work [10], investigate differentiable ways to address a memory for
reading and writing. In the ThalNet model, we use multiple recurrent controllers accessing the center
vector. Moreover, our center vector is recomputed at each time step, and thus should not be confused
with a persistent memory as is typical for model with external memory.
6
Conclusion
We presented ThalNet, a recurrent modular framework that learns to pass information between neural
modules in a hierarchical way. Experiments on sequential and permuted variants of MNIST and
CIFAR-10 are a promising sign of the viability of this approach. In these experiments, ThalNet
learns novel connectivity patterns that include hierarchical paths, skip connections, and feedback
connections.
In our current implementation, we assume the center features to be a vector. Introducing a matrix
shape for the center features would open up ways to integrate convolutional modules and similaritybased attention mechanisms for reading from the center. While matrix shaped features are easily
interpretable for visual input, it is less clear how this structure will be leveraged for other modalities.
A further direction of future work is to apply our paradigm to tasks with multiple modalities for
inputs and outputs. It seems natural to either have a separate input module for each modality, or to
have multiple output modules that can all share information through the center. We believe this could
be used to hint specialization into specific patterns and create more controllable connectivity patterns
between modules. Similarly, we an interesting direction is to explore the proposed model can be
leveraged to learn and remember a sequence of tasks.
We believe modular computation in neural networks will become more important as researchers
approach more complex tasks and employ deep learning to rich, multi-modal domains. Our work
provides a step in the direction of automatically organizing neural modules that leverage each other
in order to solve a wide range of tasks in a complex world.
8
References
[1] J. Andreas, M. Rohrbach, T. Darrell, and D. Klein. Neural module networks. In IEEE Conference
on Computer Vision and Pattern Recognition, pages 39–48, 2016.
[2] J. Ba, G. E. Hinton, V. Mnih, J. Z. Leibo, and C. Ionescu. Using fast weights to attend to the
recent past. In Advances in Neural Information Processing Systems, pages 4331–4339, 2016.
[3] K. Cho, B. van Merriënboer, D. Bahdanau, and Y. Bengio. On the properties of neural machine
translation: Encoder–decoder approaches. Syntax, Semantics and Structure in Statistical
Translation, page 103, 2014.
[4] T. Cooijmans, N. Ballas, C. Laurent, Ç. Gülçehre, and A. Courville. Recurrent batch normalization. arXiv preprint arXiv:1603.09025, 2016.
[5] C. Devin, A. Gupta, T. Darrell, P. Abbeel, and S. Levine. Learning modular neural network
policies for multi-task and multi-robot transfer. arXiv preprint arXiv:1609.07088, 2016.
[6] C. Fernando, D. Banarse, C. Blundell, Y. Zwols, D. Ha, A. A. Rusu, A. Pritzel, and D. Wierstra. Pathnet: Evolution channels gradient descent in super neural networks. arXiv preprint
arXiv:1701.08734, 2017.
[7] C. D. Gilbert and M. Sigman. Brain states: top-down influences in sensory processing. Neuron,
54(5):677–696, 2007.
[8] A. Graves. Adaptive computation time for recurrent neural networks.
arXiv:1603.08983, 2016.
arXiv preprint
[9] A. Graves, G. Wayne, and I. Danihelka. Neural turing machines. arXiv preprint arXiv:1410.5401,
2014.
[10] A. Graves, G. Wayne, M. Reynolds, T. Harley, I. Danihelka, A. Grabska-Barwińska, S. G.
Colmenarejo, E. Grefenstette, T. Ramalho, J. Agapiou, et al. Hybrid computing using a neural
network with dynamic external memory. Nature, 538(7626):471–476, 2016.
[11] J. Hawkins and D. George. Hierarchical temporal memory: Concepts, theory and terminology.
Technical report, Numenta, 2006.
[12] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE
Conference on Computer Vision and Pattern Recognition, pages 770–778, 2016.
[13] G. Hinton, A. Krizhevsky, and S. Wang. Transforming auto-encoders. Artificial Neural Networks
and Machine Learning (ICANN), pages 44–51, 2011.
[14] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
[15] G. Huang, Z. Liu, K. Q. Weinberger, and L. van der Maaten. Densely connected convolutional
networks. arXiv preprint arXiv:1608.06993, 2016.
[16] R. A. Jacobs, M. I. Jordan, and A. G. Barto. Task decomposition through competition in a
modular connectionist architecture: The what and where vision tasks. Cognitive Science, 15(2):
219–250, 1991.
[17] D. Kingma and J. Ba. Adam: A method for stochastic optimization. In International Conference
on Learning Representations, 2015.
[18] J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan,
J. Quan, T. Ramalho, A. Grabska-Barwinska, et al. Overcoming catastrophic forgetting in
neural networks. Proceedings of the National Academy of Sciences, page 201611835, 2017.
[19] A. Krizhevsky. Learning multiple layers of features from tiny images, 2009.
[20] D. Krueger, T. Maharaj, J. Kramár, M. Pezeshki, N. Ballas, N. R. Ke, A. Goyal, Y. Bengio,
H. Larochelle, A. Courville, et al. Zoneout: Regularizing rnns by randomly preserving hidden
activations. arXiv preprint arXiv:1606.01305, 2016.
9
[21] Y. LeCun and C. Cortes. The MNIST database of handwritten digits, 1998.
[22] T. P. Lillicrap, D. Cownden, D. B. Tweed, and C. J. Akerman. Random synaptic feedback
weights support error backpropagation for deep learning. Nature Communications, 7, 2016.
[23] M. Mahoney. About the test data. http://mattmahoney.net/dc/textdata, 2011.
[24] W. S. McCulloch and W. Pitts. A logical calculus of the ideas immanent in nervous activity.
The bulletin of mathematical biophysics, 5(4):115–133, 1943.
[25] T. Pham, T. Tran, and S. Venkatesh. One size fits many: Column bundle for multi-x learning.
arXiv preprint arXiv:1702.07021, 2017.
[26] S. Reed and N. De Freitas. Neural programmer-interpreters. In International Conference on
Learning Representations, 2015.
[27] A. A. Rusu, N. C. Rabinowitz, G. Desjardins, H. Soyer, J. Kirkpatrick, K. Kavukcuoglu,
R. Pascanu, and R. Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671,
2016.
[28] T. Salimans and D. P. Kingma. Weight normalization: A simple reparameterization to accelerate
training of deep neural networks. In Advances in Neural Information Processing Systems, pages
901–901, 2016.
[29] J. Schmidhuber. Learning to control fast-weight memories: An alternative to dynamic recurrent
networks. Neural Computation, 4(1):131–139, 1992.
[30] N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint
arXiv:1701.06538, 2017.
[31] S. M. Sherman. Thalamus plays a central role in ongoing cortical functioning. Nature neuroscience, 16(4):533–541, 2016.
[32] R. K. Srivastava, K. Greff, and J. Schmidhuber.
arXiv:1505.00387, 2015.
Highway networks.
arXiv preprint
[33] T. Tieleman and G. Hinton. Lecture 6.5-rmsprop: Divide the gradient by a running average of
its recent magnitude. COURSERA: Neural networks for machine learning, 4(2), 2012.
[34] F. Zenke, B. Poole, and S. Ganguli. Improved multitask learning through synaptic intelligence.
arXiv preprint arXiv:1703.04200, 2017.
10
Supplementary Material for Learning Hierarchical
Information Flow with Recurrent Neural Modules
A
Module Designs and Reading Mechanisms
1.0
Sequential MNIST Testing
1.0
0.9
0.8
ThalNet FF-GRU
ThalNet FF-GRU-FF
ThalNet GRU-FF
GRU Baseline
ThalNet GRU
ThalNet FF
0.7
0.6
0.5
20
40
60
Epochs
80
Accuracy (%)
Accuracy (%)
0.9
Sequential MNIST Testing
0.8
0.7
0.6
0.5
100
(a) Module designs
ThalNet Weight Norm
ThalNet Linear
GRU Baseline
ThalNet Fast Softmax
20
40
60
Epochs
80
100
(b) Reading mechanisms
Figure 5: Test performance on the Sequential MNIST task grouped by module design (left) and
reading mechanism (right). Plots show the top, median, and bottom accuracy over the other design
choices. Recurrent modules train faster than pure fully connected modules and weight normalized
reading is both stable and performs best. FF-GRU-FF modules perform similarly to FF-GRU while
limiting the size of the center.
We use a sequential variant of MNIST [21] to compare the reading mechanisms described in Section 2.2, along with implementations of the module function. In Sequential MNIST, the model
observes handwritten digits of 28 × 28 pixels from top to bottom, one row per time step. The
prediction is given at the last time step, so that the model has to integrate and remember observed
information over the sequence. This makes the task more challenging than in the static setting with a
multi-layer recurrent network achieving ~7 % error on this task.
To implement the modules f i (ci , xi ) we test various combinations of fully connected and recurrent
layers of Gated Recurrent Units (GRU) [3]. Modules require some amount of local structure to
allow them to specialize.7 We test with two fully connected layers (FF), a GRU layer (GRU), fully
connected followed by GRU (FF-GRU), GRU followed by fully connected (GRU-FF), and a GRU
sandwiched between fully connected layers (FF-GRU-FF). In addition, we compare performance to
a stacked GRU baseline with 4 layers. For all models, we pick the largest layer sizes such that the
number of parameters does not exceed 50,000.
We train for 100 epochs on batches of size 50 using RMSProp [33] with a learning rate of 10−3 .
Figure 5 shows the test accuracy of module designs and reading mechanisms. ThalNet outperforms
the stacked GRU baseline in most configurations. We assume that the structure imposed by our model
acts as a regularizer. We perform a further performance comparison in Section 3.
Results for module designs are shown in Figure 5a in the appendix. We observe a benefit of recurrent
modules as they exhibit faster and more stable training than fully connected modules. This could
be explained by the fact that pure fully connected modules have to learn to use the routing center to
store information over time, which is a long feedback loop. Having a fully connected layer before the
recurrent layer also significantly improves performance. A fully connected layer after the GRU let us
produce compact feature vectors φi that scale better to large modules, although we find FF-GRU to
be beneficial in later experiments (Section 3).
7
Implementing the modules as a single fully connected layer recovers a standard recurrent neural network
with one large layer.
11
Results for the reading mechanisms area shown in Figure 5b. The reading mechanism only has
a small impact on the model performance. We find weight normalized reading to yield more
stable performance than linear or fast softmax reading. For all further experiments, we use weight
normalized reading due to both its stability and predictive performance. We do not include results for
fast Gaussian reading here, as it performed below the performance range of the other methods.
B
Interpretation as Recurrent Mixture of Experts
ThalNet can route information from the input to the output over multiple time steps. This enables it
to trade off shallow and deep computation paths. To understand this, we view ThalNet as a smooth
mixture of experts model [16], where the modules F = (f 1 , · · · , f I ) are the recurrent experts. Each
module outputs its features to the center vector Φt . A linear combination of Φt is read at the next time
step, which effectively performs a mixing of expert outputs. Compared to the recurrent mixture of
experts model presented by Shazeer et al. [30], our model can recurrently route information through
the mixture of multiple times, increasing the number of mixture compounds.
To highlight two extreme cases, the modules could read from identical locations in the center. In this
case, the model does a wide and shallow computation over 1 time step, analogous to Graves [8]. In
the other extreme, each module reads from a different module, recovering a hierarchy of recurrent
layers. This gives a deep but narrow computation stretched over multiple time steps. In between, there
exist a spectrum of complex patterns of information flow with differing and dynamic computation
depths. This is comparable to DenseNet [15], which also blends information from paths of different
computational depth, although in a purely feed-forward model.
Using state-less modules, our model could still leverage the recurrence between the modules and the
center to store information over time. However, this bounds the number of distinct computation steps
that ThalNet could apply to an input. Using recurrent modules, the computation steps can change
over time, increasing the flexibility of the model. Recurrent modules give a stronger prior for using
feedback and shows improved performance in our experiments.
C
Comparison to Long Short-Term Memory
When viewing the Equations 1 – 4 in the model definition (Section 2), one might think how our model
compares to Long Short-Term Memory (LSTM) [14]. However, there exists only a limited similarity
between the two models. Empirically, we observed that LSTMs performed similarly to our GRU
baselines when given the same parameter budget.
LSTM’s context vector ct is processed element-wise, while ThalNet’s routing center cross-connects
modules. LSTM’s hidden output ht is a better candidate for comparison with ThalNet’s center
features Φ, which allows us to relate the recurrent weight matrix of an LSTM layer to the linear
version of our reading mechanism.
We could relate each ThalNet module to a set of multiple LSTM units. However, LSTM units perform
separate scalar computations, while our modules can learn complex interactions between multiple
features at each time step. Alternatively, we could see LSTM units as very small ThalNet modules,
reading exactly four context elements each, namely for the input and the three gates. However, the
computational capacity and local structure of individual LSTM units is not comparable to that of the
ThalNet modules used in our work.
12
| 2cs.AI
|
Towards Efficient Abstractions for Concurrent
Consensus⋆
arXiv:1304.1913v2 [cs.PL] 7 May 2013
Carlo Spaccasassi⋆⋆ and Vasileios Koutavas⋆ ⋆ ⋆
Trinity College Dublin, Ireland
{spaccasc,Vasileios.Koutavas}@scss.tcd.ie
Abstract. Consensus is an often occurring problem in concurrent and
distributed programming. We present a programming language with simple semantics and build-in support for consensus in the form of communicating transactions. We motivate the need for such a construct with a
characteristic example of generalized consensus which can be naturally
encoded in our language. We then focus on the challenges in achieving
an implementation that can efficiently run such programs. We setup an
architecture to evaluate different implementation alternatives and use it
to experimentally evaluate runtime heuristics. This is the basis for a research project on realistic programming language support for consensus.
Keywords: Concurrent programming, consensus, communicating transactions
1
Introduction
Achieving consensus between concurrent processes is a ubiquitous problem in
multicore and distributed programming [8, 6]. Among the classic instances of
consensus is leader election and synchronous multi-process communication. Programming language support for consensus, however, has been limited. For example, CML’s first-class communication primitives provide a programming language abstraction to implement two-party consensus. However, they cannot be
used to abstractly implement consensus between three or more processes [11,
Thm. 6.1]—this needs to be implemented in a case-by-case basis.
Let us consider a hypothetical scenario of generalized consensus, which we
will call the Saturday Night Out (SNO) problem. In this scenario a number of
friends are seeking partners for various activities on Saturday night. Each has
a list of desired activities to attend in a certain order, and will only agree for
a night out if there is a partner for each activity. Alice, for example, is looking
for company to go out for dinner and then a movie (not necessarily with the
same person). To find partners for these events in this order she may attempt
to synchronize on the “handshake” channels dinner and movie:
⋆
⋆⋆
⋆⋆⋆
Student project paper (primarily the work of the first author).
Supported by MSR (MRL 2011-039)
Supported by SFI project SFI 06 IN.1 1898.
2
def
Alice = sync dinner; sync movie
Here sync is a two-party synchronization operator, similar to CSP synchronization. Bob, on the other hand, wants to go for dinner and then for dancing:
def
Bob = sync dinner; sync dancing
Alice and Bob can agree on dinner but they need partners for a movie and
dancing, respectively, to commit to the night out. Their agreement is tentative.
Let Carol be another friend in this group who is only interested in dancing:
def
Carol = sync dancing
Once Bob and Carol agree on dancing they are both happy to commit to going
out. However, Alice has no movie partner and she can still cancel her agreement
with Bob. If this happens, Bob and Carol need to be notified to cancel their
agreement and everyone starts over their search of partners. An implementation
of the SNO scenario between concurrent processes would need to have a specialized way of reversing the effect of this synchronization. Suppose David is also a
participant in this set of friends.
def
David = sync dancing; sync movie
After the partial agreement between Alice, Bob, and Carol is canceled, David
together with the first two can synchronize on dinner, dancing, and movie and
agree to go out (leaving Carol at home).
Notice that when Alice raised an objection to the agreement that was forming
between her, Bob, and Carol, all three participants were forced to restart. If,
however, Carol was taken out of the agreement (even after she and Bob were
happy to commit their plans), David would have been able to take Carol’s place
and the work of Alice and Bob until the point when Carol joined in would not
need to be repeated.
Programming SNO between an arbitrary number of processes (which can
form multiple agreement groups) in CML is complicated. Especially if we consider that the participants are allowed to perform arbitrary computations between synchronizations affecting control flow, and can communicate with other
parties not directly involved in the SNO. For example, Bob may want to go
dancing only if he can agree with the babysitter to stay late:
def
Bob = sync dinner; if babysitter() then sync dancing
In this case Bob’s computation has side-effects outside of the SNO group of processes. To implement this would require code for dealing with the SNO protocol
to be written in the Babysitter (or any other) process, breaking any potential
modular implementation.
This paper shows that communicating transactions, a recently proposed mechanism for automatic error recovery in CCS processes [13], is a useful mechanism
for modularly implementing the SNO and other generalized consensus scenarios.
Previous work on communicating transactions focused on behavioral theory with
3
A ::= unit bool int A × A A → A A chan
v ::= x () true false n (v, v) fun f (x) = e c
e ::= v (e, e) e e op e let x = e in e if e then e else e
send e e recv e newChanA spawn e
atomic J e ⊲k e K commit k
P ::= e P k P
νc.P
J P ⊲k P K co k
op ::= fst snd add sub mul leq
E ::= [] (E, e) (v, E) E e v E op E let x = E in e
if E then e1 else e2 send E e send v E recv E spawn E
where n ∈ N, x ∈ Var , c ∈ Chan, k ∈ K
Fig. 1. TCML syntax.
respect to safety and liveness [13, 2]. However, the effectiveness of this construct
in a pragmatic programming language has yet to be proven. One of the main
milestones to achieve on this direction is the invention of efficient runtime implementations of communicating transactions. Here we describe the challenges
and our first results in a recently started project to investigate this research
direction.
In particular, we equip a simple concurrent functional language with communicating transactions and use it to discuss the challenges in making an efficient
implementation of such languages (Sect. 2). We also use this language to give a
modular implementation of consensus scenarios such as the SNO example. The
simple operational semantics of this language allows for the communication of
SNO processes with arbitrary other processes (such as the Babysitter process)
without the need to add code for the SNO protocol in those processes. Moreover,
the more efficient, partially aborting strategy discussed above is captured in this
semantics.
Our semantics of this language is non-deterministic, allowing different runtime scheduling strategies of processes, some more efficient than others. To study
their relative efficiency we have developed a skeleton implementation of the language which allows us to plug in and evaluate such runtime strategies (Sect. 3).
We describe several such strategies (Sect. 4) and report the results of our evaluations (Sect. 5). Finally, we summarize related work in this area and the future
directions of this project (Sect. 6).
2
The TCML Language
We study TCML, a language combining a simply-typed λ-calculus with πcalculus and communicating transactions. For this language we use the abstract
syntax shown in Fig. 1 and the usual abbreviations from the λ- and π-calculus.
Values in TCML are either constants of base type (unit, bool, and int), pairs
of values (of type A × A), recursive functions (A → A), and channels carrying
values of type A (A chan). A simple type system (with appropriate progress
4
if true then e1 else e2
if false then e1 else e2
let x = v in e
op v
fun f (x) = e v2
֒→
֒→
֒→
֒→
֒→
e1
e2
e[v/x]
δ(op, v)
e[fun f (x) = e/f ][v2 /x]
Step
E[e]
Spawn
E[spawn v]
NewChan E[newChanA ]
Atomic
E[atomic J e1 ⊲k e2 K]
Commit E[commit k]
−→
−→
−→
−→
−→
E[e′ ]
v () k E[()]
νc.E[c]
J E[e1 ] ⊲k E[e2 ] K
co k k E[()]
If-True
If-False
Let
Op
App
if e ֒→ e′
if c 6∈ fc(E[()])
Fig. 2. Sequential reductions
and preservation theorems) can be found in an accompanying technical report
[12] and is omitted here.
Source TCML programs are expressions in the functional core of the language, ranged over by e, whereas running programs are processes derived from
the syntax of P . Besides standard lambda calculus expressions, the functional
core contains the constructs send c e and recv c to synchronously send and receive a value on channel c, respectively, and newChanA to create a new channel
of type chan A. The constructs spawn and atomic, when executed, respectively
spawn a new process and transaction; commit k commits transaction k—we will
shortly describe these constructs in detail.
A simple running process can be just an expression e. It can also be constructed by the parallel composition of P and Q (P k Q). We treat free channels
as in the π-calculus, considering them to be global. Thus if a channel c is free in
both P and Q, it can be used for communication between these processes. The
construct νc.P encodes π-calculus restriction of the scope of c to process P . We
use the Barendregt convention for bound variables and channels and identify
terms up to alpha conversion. Moreover, we write fc(P ) for the free channels in
process P .
We write J P1 ⊲k P2 K for the process encoding a communicating transaction.
This can be thought of as the process P1 , the default of the transaction, which
runs until the transaction commits. If, however, the transaction aborts then P1
is discarded and the entire transaction is replaced by its alternative process P2 .
Intuitively, P2 is the continuation of the transaction in the case of an abort.
As we will explain, commits are asynchronous, requiring the addition of process
co k to the language. The name k of the transaction is bound in P1 . Thus only
the default of the transaction can potentially spawn a co k. The meta-function
ftn(P ) gives us the free transaction names in P .
Processes with no free variables can reduce using transitions of the form
P −→ Q. These transitions for the functional part of the language are shown in
Fig. 2 and are defined in terms of reductions e ֒→ e′ (where e is a redex ) and
eager, left-to-right evaluation contexts E whose grammar is given in Fig. 1. Due
to a unique decomposition lemma, an expression e can be decomposed to an
evaluation context and a redex expression in only one way. Here we use e[u/x]
5
for the standard capture-avoiding substitution, and δ(op, v) for a meta-function
returning the result of the operator op on v, when this is defined.
Rule Step lifts functional reductions to process reductions. The rest of the
reduction rules of Fig. 2 deal with the concurrent and transactional side-effects
of expressions. Rule Spawn reduces a spawn v expression at evaluation position
to the unit value, creating a new process running the application v (). The type
system of the language guarantees that value v here is a thunk. With this rule
we can derive the reductions:
spawn(λ(). send c 1); recv c −→(λ(). send c 1) () k recv c
−→ send c 1 k recv c
The resulting processes of these reductions can then communicate on channel c.
As we previously mentioned, the free channel c can also be used to communicate
with any other parallel process. Rule NewChan gives processes the ability to
create new, locally scoped channels. Thus, the following expression will result in
an input and an output process that can only communicate with each other:
let x = newChanint in (spawn (λ(). send x 1); recv x)
−→ νc. (spawn (λ(). send c 1); recv c)
−→∗ νc. (send c 1 k recv c)
Rule Atomic starts a new transaction in the current (expression-only) process, engulfing the entire process in it, and storing the abort continuation in the
alternative of the transaction. Rule Commit spawns an asynchronous commit.
Transactions can be arbitrarily nested, thus we can write:
atomic J spawn(λ(). recv c; commit k) ⊲k () K;
atomic J recv d; commit l ⊲l () K
which reduces to
J(recv c; commit k) k J recv d; commit l ⊲l () K
⊲k (); atomic J recv d; commit l ⊲l () K K
This process will commit the k-transaction after an input on channel c and the
inner l-transaction after an input on d. As we will see, if the k transaction aborts
then the inner l-transaction will be discarded (even if it has performed the input
on d) and the resulting process (the alternative of k) will restart l:
(); atomic J recv d; commit l ⊲l () K
The effect of this abort will be the rollback of the communication on d reverting
the program to a consistent state.
Process and transactional reductions are handled by the rules of Fig. 3. The
first four rules (Sync, Eq, Par, and Chan) are direct adaptations of the reduction rules of the π-calculus, which allow parallel processes to communicate, and
propagate reductions over parallel and restriction. These rules use an omitted
6
E1 [recv c] k E2 [send c v] −→ E1 [v] k E2 [()]
Eq
P ≡ P ′ −→ Q′ ≡ Q
P −→ Q
Par
P1 −→ P1′
P1 k P2 −→ P1′ k P2
Chan
P −→ P ′
νc.P −→ νc.P ′
Emb
Step
P −→ P ′
J P ⊲k P2 K −→ J P ′ ⊲k P2 K
Sync
P1 k J P2 ⊲k P3 K −→ J P1 k P2 ⊲k P1 k P3 K
Abort
Co
P1 ≡ co k k P1′
J P1 ⊲k P2 K −→ P1′ /k
J P1 ⊲k P2 K −→ P2
Fig. 3. Concurrent and Transactional reductions (omitting symmetric rules).
structural equivalence (≡) to identify terms up to the reordering of parallel processes and the extrusion of the scope of restricted channels, in the spirit of the
π-calculus semantics. Rule Step propagates reductions of default processes over
their respective transactions. The remaining rules are taken from TransCCS [13].
Rule Emb encodes the embedding of a process P1 in a parallel transaction
J P1 ⊲k P2 K. This enables the communication of P1 with P2 , the default of
k. It also keeps the current continuation of P1 in the alternative of k in case
the k-transaction aborts. To illustrate the mechanics of the embed rule, let
us consider the above nested transaction running in parallel with the process
P = send d (); send c ():
J(recv c; commit k) k J recv d; commit l ⊲l () K
⊲k (); atomic J recv d; commit l ⊲l () K K
k P
After two embedding transitions we will have
J(recv c; commit k) k
J P k recv d; commit l ⊲l P k () K
⊲k
P k ...K
⊲k
P k ...K
Now P can communicate on d with the inner transaction:
J(recv c; commit k) k
J send c () k commit l ⊲l P k () K
Next, there are (at least) two options: either commit l spawns a co l process
which causes the commit of the l-transaction, or the input on d is embedded in
the l-transaction. Let us assume that the latter occurs:
J J(recv c; commit k) k send c () k commit l
⊲l (recv c; commit k) k P k () K
⊲k P k . . . K
−→∗ J J co k k co l ⊲l . . . K ⊲k . . . K
The transactions are now ready to commit from the inner-most to the outer-most
using rule Commit. Inner-to-outer commits are necessary to guarantee that all
transactions that have communicated have reached an agreement to commit.
7
This also has the important consequence of making the following three processes behaviorally indistinguishable:
J P1 ⊲k P2 K k J Q1 ⊲l Q2 K
J P1 k J Q1 ⊲l Q2 K ⊲k P2 k J Q1 ⊲l Q2 K K
J J P1 ⊲k P2 K k Q1 ⊲l J P1 ⊲k P2 K k Q2 K
Therefore, an implementation of TCML, when dealing with the first of the three
processes can pick any of the alternative, non-deterministic mutual embeddings
of the k and l transactions without affecting the observable outcomes of the
program. In fact, when one of the transactions has no possibility of committing
or when the two transactions never communicate, an implementation can decide
never to embed the two transactions in each-other. This is crucial in creating
implementations that will only embed processes (and other transactions) only
when necessary for communication, and pick the most efficient of the available embeddings. The development of implementations with efficient embedding
strategies is one of the main challenges of our project for scaling communicating
transactions to pragmatic programming languages.
Similarly, aborts are entirely non-deterministic (Abort) and are left to the
discretion of the underlying implementation. Thus in the above example any
transaction can abort at any stage, discarding part of the computation. In such
examples there is usually a multitude of transactions that can be aborted, and
in cases where a “forward” reduction is not possible (due to deadlock) aborts are
necessary. Making the TCML programmer in charge of aborts (as we do with
commits) is not desirable since the purpose of communicating transactions is
to lift the burden of manual error prediction and handling. Minimizing aborts,
and automatically picking the aborts that will undo the fewer computation steps
while still rewinding the program back enough to reach a successful outcome is
another major challenge in our project.
The SNO scenario can be simply implemented in TCML using restarting
transactions. A restarting transaction uses recursion to re-initiate an identical
transaction in the case of an abort:
atomicrec k J e K
def
=
fun r() = atomic J e ⊲k r () K
A transactional implementation of the SNO participants we discussed in the
introduction simply wraps their code in restating transactions:
let alice = atomicrec k J sync dinner; sync movie; commit k K in
let bob
= atomicrec k J sync dinner; sync dancing; commit k K in
let carol = atomicrec k J sync dancing; commit k K in
let david = atomicrec k J sync dancing; sync movie; commit k K in
spawn alice; spawn bob; spawn carol; spawn david
Here dinner, dancing, and movie are implementations of CSP synchronization channels and sync a function to synchronize on these channels. Compared
8
Transaction
trie
Sched.
Gath.
abort,
embed,
commit
side-effect notif.
& ack
e1
e2
...
en
Fig. 4. TCML runtime architecture.
to a potential ad-hoc implementation of SNO in CML the simplicity of the above
code is evident (the version of Bob communicating with the Babysitter is just as
simple). However, as we discuss in Sect. 5, this simplicity comes with a severe
performance penalty, at least for straightforward implementations of TCML.
In essence, the above code asks from the underlying transactional implementation to solve an NP-complete satisfiability problem. Leveraging existing useful
heuristics for such problems is something we intend to pursue in future work.
In the following sections we describe an implementation where these transactional scheduling decisions can be plugged in, and a number of heuristic transactional schedulers we have developed and evaluated. Our work shows that although more advanced heuristics bring measurable performance benefits, the
exponential number of runtime choices require the development of innovative
compilation and execution techniques to make communicating transactions a
realistic solution for programmers.
3
An Extensible Implementation Architecture
We have developed an interpreter for the TCML reduction semantics in Concurrent Haskell [7, 10] to which we can plug-in different decisions about the
non-deterministic transitions of our semantics. Here we briefly explain the runtime architecture of this interpreter, shown in Fig. 4.
The main Haskell threads are shown as round nodes in the figure. Each
concurrent functional expression ei is interpreted in its own thread according to
the sequential reduction rules in Fig. 2 of the previous section. Side-effects in
an expression will be generally handled by the interpreting thread, creating new
channels, spawning new threads, and starting new transactions.
Except for new channel creation, the evaluation of all other side-effects in
an expression will cause a notification (shown as dashed arrows in Fig. 2) to be
sent to the gatherer process (Gath.). This process is responsible for maintaining
a global view of the state of the running program in a Trie data-structure. This
data-structure essentially represents the transactional structure of the program;
i.e., the logical nesting of transactions and processes inside running transactions:
data TTrie = TTrie {
9
threads
children
:: Set ThreadID,
:: Map TransactionID TTrie, ... }
A TTrie node represents a transaction, or the top-level of the program. The
main information stored in such a node is the set of threads (threads) and transactions (children) running in that transactional level. Each child transaction
has its own associated TTrie node. An invariant of the data-structure is that
each thread and transaction identifier appears only once in it. For example the
complex program we saw on page 6:
J(recv c; commit k)tid1 k J(recv d; commit l)tid2 ⊲l () K
⊲k (); atomic J recv d; commit l ⊲l () K K
k P tidP
will have an associated trie:
TTrie{threads = {tidP },
children = {k 7→ TTrie{threads = {tid1 },
children = {l 7→ TTrie{threads = {tid2 },
children = ∅}}}}}
The last ingredient of the runtime implementation is the scheduler thread
(Sched. in Fig. 4). This makes decisions about the commit, embed and abort
transitions to be performed by the expression threads, based on the information
in the trie. Once such a decision is made by the scheduler, appropriate signals
(implemented using Haskell asynchronous exceptions [10]) are sent to the running
threads, shown as dotted lines in Fig. 4. Our implementation is parametric to the
precise algorithm that makes scheduler decisions, and in the following section
we describe a number of such algorithms we have tried and evaluated.
A scheduler signal received by a thread will cause the update of the local
transactional state of the thread, affecting the future execution of the thread.
The local state of a thread is an object of the TProcess data-type:
data TProcess = TP {
expr :: Expression,
ctx :: Context,
tr
:: [Alternative] }
data Alternative = A {
tname :: TransactionID,
pr
:: TProcess }
The local state maintains the expression (expr) and evaluation context (ctx)
currently interpreted by the thread and a list of alternative processes (represented by objects of the Alternative data-type). This list contains the continuations stored when the thread was embedded in transactions. The nesting of
transactions in this list mirrors the transactional nesting in the global trie and is
thus compatible with the transactional nesting of other expression threads. Let
us go back to the example of page 6:
J(recv c; commit k)tid1 k J(recv d; commit l)tid2 ⊲l () K
⊲k (); atomic J recv d; commit l ⊲l () K K
k P tidP
where P = send d (); send c (). When P is embedded in both k and l, the thread
evaluating P will have the local state object
10
TP{expr = P , tr = [A{tname = l, pr = P }, A{tname = k, pr = P }]}
recording the fact that the thread running P is part of the l-transaction,
which in turn is inside the k-transaction. If either of these transactions aborts
then the thread will rollback to P , and the list of alternatives will be appropriately updated (the aborted transaction will be removed).
Once a transactional reconfiguration is performed by a thread, an acknowledgment is sent back to the gatherer, who, as we discussed, is responsible for updating the global transactional structure in the trie. This closes a cycle of transactional reconfigurations initiated from the process (by starting a new transaction
or thread) or the scheduler (by issuing a commit, embed, or abort).
What we described so far is a simple architecture for an interpreter of TCML.
Various improvements are possible (e.g., addressing the message bottleneck in
the gatherer) but are beyond the scope of this paper. In the following section we
discuss various policies for the scheduler which we then evaluate experimentally.
4
Transactional Scheduling Policies
Our goal here is to investigate schedulers that make decisions on transactional
reconfiguration based only on runtime heuristics. We are currently working on
more advanced schedulers, including schedulers that take advantage of static
information extracted from the program, which we leave for future work.
An important consideration when designing a scheduler is adequacy [?, Chap.
13, Sec. 4]. For a given program, an adequate scheduler is able to produce all
outcomes that the non-deterministic operational semantics can produce for that
program. However, this does not mean that the scheduler should be able to
produce all traces of the non-deterministic semantics. Many of these traces will
simply abort and restart the same computations over and over again. Previous
work on the behavioral theory of communicating transactions has shown that all
program outcomes can be reached with traces that never restart a computation
[13]. Thus a goals of our schedulers is to minimize re-computations by minimizing
the number of aborts.
Moreover, as we discussed at the end of Sect. 2, many of the exponential
number of embeddings can be avoided without altering the observable behavior
of a program. This can be done by embedding a process inside a transaction
only when this embedding is necessary to enable communication between the
process and the transaction. We take advantage of this in a communicationdriven scheduler we describe in this section.
Even after reducing the number of possible non-deterministic choices faced
by the scheduler, in most cases we are still left with a multitude of alternative
transactional reconfiguration options. Some of these are more likely to lead to
efficient traces than other. However, to preserve adequacy we cannot exclude
any of these options since the scheduler has no way to foresee their outcomes.
In these cases we assign different, non-zero probabilities to the available choices,
based on heuristics. This leads to measurable performance improvements, without violating adequacy. Of course some program outcomes might be more likely
11
to appear than others. This approach is trading measurable fairness for performance improvement.
However, the probabilistic approach is theoretically fair. Every finite trace
leading to a program outcome has a non-zero probability. Diverging traces due
to sequential reductions also have non-zero probability to occur. The only traces
with zero probability are those in the reduction semantics that have an infinite
number of non-deterministic reductions. Intuitively, these are unfair traces that
abort and restart transactions ad infinitum, even if other options are possible.
Random Scheduler (R). The very first scheduler we consider is the random
scheduler, whose policy is to simply, at each point, select one of all the nondeterministic choices with equal probability, without excluding any of these
choices. With this scheduler any abort, embed, or commit actions are equally
likely to happen. Although this naive scheduler is not particularly efficient, as
one would expect, it is an obviously adequate and fair scheduler according to the
discussion above. If a reduction transition is available infinitely often, scheduler
R will eventually select it.
This scheduler leaves much room for improvement. Suppose that a transaction k is ready to commit:
J P k co k ⊲k Q K
Since R makes no distinction between the choices of committing and aborting
k, it will often unnecessarily abort k. All processes embedded in this transaction will have to roll back and re-execute; if k was a transaction that restarts,
the transaction will also re-execute. This results to a considerable performance
penalty. Similarly, scheduler R might preemptively abort a long-running transaction that could have have committed given enough time and embeddings (for
the purpose of communication).
Staged Scheduler (S). The staged scheduler partially addresses these issues by
prioritizing its available choices. Whenever a transaction is ready to commit,
scheduler S will always decide to send a commit signal to that transaction before aborting it or embedding another process in it. This does not violate adequacy; before continuing with the algorithm of S, let us examine the adequacy
of prioritizing commits over other transactional actions with an example.
Example 1 Consider the following program in which k is ready to commit.
J P k co k ⊲k Q K k R
If embedding R in k leads to a program outcome, then that outcome can also be
reached after committing k from the residual P k R.
Alternatively, a program outcome could be reachable by aborting k (from the
process Q k R). However, the co k was spawned from one of the previous states
of the program in the current trace. In that state, transaction k necessarily had
the form: J P ′ k E[commit k] ⊲k Q K. In that state the abort of k was enabled.
Therefore, the staged interpreter indeed allows a trace leading to the program
state Q k R from which the outcome in question is reachable.
⊓
⊔
12
If no commit is possible for a transaction, the staged interpreter prioritizes
embeds into that transaction over aborting the transaction. This is again an
adequate decision because the transactions that can take an abort reduction
before an embed step have an equivalent abort reduction after that step.
When no commit nor embed options are available for a transaction, the staged
interpreter lets the transaction run with probability 0.95, giving more chances to
make progress in the current trace, and with probability 0.05 it aborts it—these
numbers have been fine-tuned with a number of experiments.
The benefit of the heuristic implemented in this scheduler is that it minimizes
unnecessary aborts improving performance. Its drawback is that it does not abort
transactions often, thus program outcomes reachable only from transactional
alternatives are less likely to appear. Moreover, this scheduler does not avoid
unnecessary embeddings.
Communication-Driven Scheduler (CD). To avoid spurious embeddings, scheduler CD improves over R by performing an embed transition only if it is necessary
for an imminent communication. For example, in the following program state the
embedding of the right-hand-side process into k will never be chosen.
J E[recv c] ⊲k Q K k ((); send c v)
However, after that process reduces to an output, its embedding into k will be
enabled. Because of the equivalence
J P ⊲k Q K k R ≡cxt J P k R ⊲k Q k R K
which we previously discussed, this scheduler is adequate.
For the implementation of this scheduler we augment the information stored
in the trie data-structure (Sect. 3) with the channel which each thread is waiting
to communicate on (if any).
As we will see in Sect. 5, this heuristic significantly boosts performance because it greatly reduces the exponential number of embedding choices.
Delayed-Aborts Scheduler (DA). The final scheduler we report is DA, which adds
a minor improvement upon scheduler CD. This scheduler keeps a timer for each
running transaction k in the transaction trie. This timer is reset whenever a communication or transactional operation happens inside k. Transaction k will only
be considered for an abort when this timer expires. This strategy benefits longrunning transactions that perform multiple communications before committing.
The CD scheduler is obviously adequate because it only adds time delays.
5
Evaluation of the Interpreters
We now report the experimental evaluation of interpreters using the preceding
Scheduling policies. The interpreters were compiled with GHC 7.0.3, and the
experiments were performed on a Windows 7 machine with Intel R CoreTM i52520M 250 GHz processor and 8Gb of RAM. We run several versions of two
programs:
13
SNO Example
Committed ops/second
Three-Way Rendezvous
102
102
101
101
100
10−1
100
10−2
10−1
10−3
5
10
5
15
Number of concurrent processes
R
10
S
15
CD
20
TA
ID
Fig. 5. Experimental Results.
1. The three-way rendezvous (3WR) in which a number of processes compete
to synchronize on a channel with two other processes, forming groups of
three which then exchange values. This is a standard example of multi-party
agreement [11, 3, 5]. In the TCML implementation of this example each process nondeterministically chooses between being a leader or follower within a
communicating transaction. If a leader and two followers communicate, they
can all exchange values and commit; any other situation leads to deadlock
and eventually to an abort of some of the transactions involved.
2. The SNO example of the introduction, as implemented in Sect. 2, with multiple instances of the Alice, Bob, Carol, and David processes.
To test the scalability of our schedulers, we tested a number of versions of the
above programs, each with a different number of competing parallel processes.
Each process in these programs continuously performs 3WR or SNO cycles and
our interpreters are instrumented to measure the number of operations in a given
time, from which we compute the mean throughput of successful 3WR or SNO
operations. The results are shown in Fig. 5.
Each graph in the figure contains the mean throughput of operations (in
logarithmic scale) as a function of the number of competing concurrent TCML
processes. The graphs contain runs with each scheduler we discussed (random
R, staged S, communication-driven, CD, and communication-driven with timed
aborts TA) as well as with an ideal non-transactional program (ID). The ideal
program in the case of 3WR is similar to the TCML, non-abstract implementation [11]. The ideal version of the SNO is running a simpler instance of the
scenario, without any Carol processes—this instance has no deadlocks and therefore needs no error handling. Ideal programs give us a performance upper bound.
As predictable, the random scheduler (R)’s performance is the worst; in many
cases R could not perform any operations in the window of measurements (30sec).
The other schedulers perform better than R by an order of magnitude. Even
just prioritizing the transactional reconfiguration choices significantly cuts down
14
the exponential number of inefficient traces. However, none of the schedulers
scale to programs with more processes; their performance deteriorates exponentially. In fact, when we go from the communication-driven (CD) to the timedaborts (TA) scheduler we see worst throughput in larger process pools. This is
because with many competing processes there is more possibility to enter a path
to deadlock; in these cases the results suggest that it is better to abort early.
The upper bound in the performance, as shown by the throughput of ID is
one order of magnitude above that of the best interpreter, when there are few
concurrent processes, and (within the range of our experiments) two orders when
there are many concurrent processes. The performance of ID is increasing with
more processes due to better utilization of the processor cores.
It is clear that in order to achieve a pragmatic implementation of TCML
we need to address the exponential nature in consensus scenarios as the ones we
tested here. Our exploration of purely runtime heuristics shows that performance
can be improved, but we need to turn to a different approach to close the gap
between ideal ad-hoc implementations and abstract TCML implementations.
6
Conclusions and Future Work
Consensus is an often occurring problem in concurrent and distributed programming. The need for developing programming language support for consensus has
already been identified in previous work on transactional events (TE) [3], communicating memory transactions (CMT) [9], transactors [4] and cJoin [1]. These
approaches propose forms of restarting communicating transactions, similar to
those described in Sect. 2. TE, CMT and Transactors can be used to implement
the instance of the Saturday Night Out (SNO) example in this paper. TE extends
CML events with a transactional sequencing operator; transactional communication is resolved at runtime by search threads which exhaustively explore all possibilities of synchronization, avoiding runtime aborts. CMT extends STM with
asynchronous communication, maintaining a directed dependency graph mirroring communication between transactions; STM abort triggers cascading aborts
to transactions that have received values from aborting transactions. Transactors
extend actor semantics with fault-tolerance primitives, enabling the composition
of systems with consistent distributed state via distributed checkpointing. The
cJoin calculus extends the Join calculus with isolated transactions which can
be merged. Merging and aborting are managed by the programmer, offering a
manual alternative to TCML’s nondeterministic transactional operations. It is
unclear to us how to write a straightforward implementation of the SNO example in cJoin. Reference implementations have been developed for TE, CMT and
cJoin. The discovery of efficient implementations for communicating transactions
could be equally beneficial for all approaches. Stabilizers [14] add transactional
support for fault-tolerance in the presence of transient faults but do not directly
address concensus scenarios such as the SNO example.
This paper presented TCML, a simple functional language with build-in support for consensus via communicating transactions. This is a construct with a
15
robust behavioral theory supporting its use as a programming language abstraction for automatic error recovery [13, 2]. TCML has a simple operational
semantics and can simplify the programming of advanced consensus scenarios;
we introduced such an example (SNO) which has a natural encoding in TCML.
The usefulness of communicating transactions in real-world applications,
however, depends on the invention of efficient implementations. This paper described the obstacles we need to overcome and our first results in a recently
started project on developing such implementations. We gave a framework to
develop and evaluate current and future runtime schedulers of communicating
transactions and used it to examine schedulers which are based solely on runtime
heuristics. We have found that some heuristics improve upon the performance
of a naive randomized implementation but do not scale to programs with significant contention, where an exponential number of alternative computation paths
lead to necessary rollbacks. It is clear that purely dynamic strategies do not lead
to sustainable performance improvements.
In future work we intend to pursue a direction based on the extraction of
information from the source code which will guide the language runtime. This
information will include an abstract model of the communication behavior of
processes that can be used to predict with high probability their future communication pattern. A promising approach to achieve this is the development of
technology in type and effect systems and static analysis. Although the scheduling of communicating transactions is theoretically computationally expensive,
realistic performance in many programming scenarios could be achievable.
References
[1] Bruni, R., Melgratti, H., Montanari, U.: cjoin: Join with communicating transactions. To appear in MSCS
[2] De Vries, E., Koutavas, V., Hennessy, M.: Liveness of communicating transactions.
pp. 392–407. APLAS’10
[3] Donnelly, K., Fluet, M.: Transactional events. pp. 124–135. ICFP ’06
[4] Field, J., Varela, C.A.: Transactors: A programming model for maintaining globally consistent distributed state in unreliable environments. POPL ’05
[5] Harris, T., Marlow, S., Jones, S.P.L., Herlihy, M.: Composable memory transactions. Commun. ACM pp. 91–100 (2008)
[6] Herlihy, M., Shavit, N.: The art of multiprocessor programming. Kaufmann (2008)
[7] Jones, S.P.L., Gordon, A.D., Finne, S.: Concurrent haskell. POPL’96
[8] Kshemkalyani, A.D., Singhal, M.: Distributed Computing: Principles, Algorithms,
and Systems. Cambridge University Press (2008)
[9] Lesani, M., Palsberg, J.: Communicating memory transactions. PPoPP ’11
[10] Marlow, S., Jones, S.P.L., Moran, A., Reppy, J.H.: Asynchronous exceptions in
haskell. PLDI’01
[11] Reppy, J.H.: Concurrent programming in ML. Cambridge University Press (1999)
[12] Spaccasassi, C.: Transactional concurrent ml. Tech. Rep. TCD-CS-2013-01
[13] de Vries, E., Koutavas, V., Hennessy, M.: Communicating transactions. pp. 569–
583. CONCUR’10
[14] Ziarek, L., Schatz, P., Jagannathan, S.: Stabilizers: a modular checkpointing abstraction for concurrent functional programs. ICFP ’06
| 6cs.PL
|
Network of Recurrent Neural Networks
Chao-Ming Wang
arXiv:1710.03414v1 [cs.NE] 10 Oct 2017
School of Software, Beijing Jiaotong University, Beijing, China
[email protected]
Abstract
We describe a class of systems theory based neural networks
called “Network Of Recurrent neural networks” (NOR),
which introduces a new structure level to RNN related models. In NOR, RNNs are viewed as the high-level neurons and
are used to build the high-level layers. More specifically,
we propose several methodologies to design different NOR
topologies according to the theory of system evolution. Then
we carry experiments on three different tasks to evaluate our
implementations. Experimental results show our models outperform simple RNN remarkably under the same number of
parameters, and sometimes achieve even better results than
GRU and LSTM.
Introduction
In recent years, Recurrent Neural Networks (RNNs) (Elman
1990) have been widely used in Natural Language Processing (NLP). Traditionally, RNNs are directly used to build the
final models. In this paper, we propose a novel idea called
“Network Of Recurrent neural networks” (NOR) which utilizes existing basic RNN layers to make the structure design
of the higher-level layers. From a standpoint of systems theory (Von Bertalanffy 1968; Von Bertalanffy 1972), a recurrent neural network is a group or an organization made up of
a number of interacting parts, and it actually is viewed as a
complex system, or a complexity. Dialectically, every system
is relative: it is not only the system of its parts, but also the
part of a larger system. In NOR structures, RNN is viewed
as the high-level neuron, and several high-level neurons are
used to build the high-level layers rather than directly used
to construct the whole models.
Conventionally, there are three levels of structure in Deep
Neural Networks (DNNs): neurons, layers and whole nets
(or called models). From a perspective of systems theory,
at each level of such increasing complexity, novel features
that do not exist at lower levels emerge (Lehn 2002). For
example, at the neurons level, single neuron is simple and
its generalization capability is very poor. But when a certain number of such neurons are accumulated into a certain
elaborate structure by certain ingenious combinations, the
layers at the higher level begin to get the unprecedented ability of classification and feature learning. More importantly,
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
such new gained capability or property is deducible from
but not reducible to constituent neurons of lower levels. It’s
not a property of the simple superposition of all constituent
neurons, and the whole is greater than the sum of the parts.
In systems theory, such kind of phenomenon is known as
whole emergence (Wierzbicki 2015). Whole emergence often comes from the evolution of the system (Arthur and others 1993), in which a system develops from the lower level
to the higher level, from simplicity to complexity. In this
paper, the motivation of NOR structures is to introduce a
new structure level to RNN-related networks by transferring
traditional RNN from the system to the agent and from the
outer dimension to the inner dimension (Fromm 2004).
In 1993, W. Brian Arthur (Arthur and others 1993) has
identified three mechanisms by which complexity tends to
grow as systems evolve:
• Mechanism 1: increase in co-evolutionary diversity. The
agent in the system seem to be a new instance of agent
class, type or species. As a result, the system seems to
have new external agent types or capabilities.
• Mechanism 2: increase in structural sophistication. The
individual system steadily accumulates increasing numbers of new systems or parts. Thus, newly formed system
seems to have new internal subsystems or capabilities.
• Mechanism 3: increase by “capturing software”. The
system capture simpler elements and learns to “program”
these as “software” to be used as its own ends.
In this paper, with the guidance of first two mechanisms,
we introduce two methodologies to NOR structures design,
which are named as aggregation and specialization. Aggregation and specialization are natural operations for increasing complexity in complex systems (Fromm 2004). The former is related to Arthur’s second mechanism, in which traditional RNNs are aggregated and accumulated into a highlevel layer in accordance with a specific structure, and the
latter is related to Arthur’s first mechanism, in which the
RNN agent in a high-level layer is specialized as the RNN
agent that performs a specific function.
We make several implementations and carry out experiments on three different tasks, including sentiment classification, question type classification and named entity recognition. Experimental results show that our models outperform constitute simple RNN remarkably with the same num-
ber of parameters and achieve even better results than GRU
and LSTM sometimes.
Background
Systems Theory Systems Theory was originally proposed
by biologist Ludwig Von Bertalanffy (Von Bertalanffy 1968;
Von Bertalanffy 1972) for biological phenomena. In biology systems, there are several different levels which begin
with the smallest units of life and reach to the largest and
most extensive category: molecule, cell, tissue, organ, organ system, organization etc. Traditionally, a system could
be decomposed into its individual components so that each
component could be analyzed as an independent entity, and
components could be added in a linear fashion to describe
the totality of the system (Walonick 1993). However, Von
Bertalanffy argued that we cannot fully comprehend a phenomenon by simply breaking it down into elementary parts
and then reforming it. We instead need to apply a global and
systematic perspective to underline its functionality (Mele,
Pels, and Polese 2010), because a system is characterized
by the interactions of its components and the nonlinearity of
those interactions (Walonick 1993).
Whole Emergence In systems theory, the phenomenon
(i.e., the whole is irreducible to its parts) is known as emergence or whole emergence (Wierzbicki 2015). Emergence
can be qualitatively described as: the whole is greater than
the sum of the parts (Upton, Janeka, and Ferraro 2014). Or
it can also be quantitatively expressed as:
n
X
W >
pi
(1)
arbitrary sequences of inputs. Formally, given a sequence
of vectors {xt }t=1···T , the equation of Simple RNN (Elman
1990) is:
ht = f (W xt + U ht−1 )
(2)
where W and U are parameter matrices, and f denotes a
nonlinearity function such as tanh or ReLU. For simplicity,
the neuron biases are omitted from the equation.
Actually, RNNs can behave chaotically. There have
been some works analysing RNNs theoretically or experimentally from the perspective of systems theory. (Sontag
1997) provided an exposition of research regarding systemtheoretic aspects of RNNs with sigmoid activation functions.
(Bertschinger and Natschläger 2004) analyzed the computation at the edge of chaos in RNNs and calculated the critical
boundary in parameter space where the transition from ordered to chaotic dynamics takes place. (Pascanu, Mikolov,
and Bengio 2013) employed a dynamical systems perspective to understand the exploding gradients and vanishing gradients problems in RNNs. In this paper, we obtain methodologies from systems theory to conduct structure designs of
RNN related models.
Network of Recurrent Neural Networks
Overall Architecture
i
where W is the whole of the system and consists of n parts,
and {pi }i=1···n is the i-th part. In 1972, Philip W. Anderson
highlighted the idea of emergence in has article “More is
Different” (Anderson 1972) in which he stated that a change
of scale very often causes a qualitative change in the behavior of the system. For example, in human brains, when one
examines a single neuron, there is nothing that suggests conscious. But a collection of millions of neurons is clearly able
to produce wonderful consciousness.
The mechanisms behind the emergence of complexity
can be used to design neural network structures. One
of the widely accepted reasons is the repeated application
and combination of two complementary forces or operations: stretching and folding (in Physics term (Thompson
and Stewart 2002)), splitting and merging (in Computer Science term (Hannebauer 2002)), or specialization and cooperation (in Sociology term). Merging or aggregating of agents
means generally a number of (sub-)agents is aggregated or
conglomerated into a single agent. Splitting or specializing
means the agents are clearly separated from each other and
each agent is constrained to a certain class or role (Fromm
2004).
Recurrent Neural Networks at the Edge of Chaos Recurrent Neural Networks (RNNs) (Werbos 1988; Elman
1990) are a class of deep neural networks that possess internal short-term memory due to recurrent feed-back connections between units, which makes them be able to process
Figure 1: Overview of NOR structure.
As the high-level illustration shown in Figure 1, NOR architecture is a three-dimensional spatial-temporal structure.
We summarize NOR architecture as four components: I, M,
S and O, in which component I (input) and O (output) control the head and tail of NOR layer, component S (subnetworks) is in charge of the spatial extension and component
M (memories) is responsible for the temporal extension of
the whole structure. We describe each component as follows:
Component I: Component I controls the head of NOR architecture. It does data preprocessing tasks and distributes
processed input data to subnetworks. At each time-step
t, the form of upcoming input data xt may be various,
such as one single vector, or several vectors with the multigranularity information, even the feature vectors with noise.
One single vector may be the simplest situation, and the
common solution is copying this vector into n duplicates and
feed each of them into one single subnetwork in the component S. In this paper, the copying method meets our needs
(a) MA-NOR layer.
(b) Another MA-NOR layer.
(c) MS-NOR layer.
(d) SS-NOR layer.
Figure 2: The sectional views of NOR layers at one time-step. “I” means the component I, “O” means the component O, and
“R” means RNN neuron.
and we formalize it as:
{xit }i=1···n
= C(xt )
(3)
in which C means copy function, and xit will be fed into i-th
subnetwork.
Component M: Component M manages all memories
over the whole layer, not only internal but also external
memories (Weston, Chopra, and Bordes 2014). But in this
paper, component M only considers internal memory and do
not apply any extra processing to the individual memory of
each RNN neuron. That is:
mjt = I(ojt−1 )
(4)
where I means identity function, the superscript j is the
identifier of j-th RNN neuron 1 , mjt is the memory of jth RNN neuron at time-step t and ojt−1 is the transformation
output of j-th RNN neuron at time-step t − 12 .
Component S: Component S is made up of several different or same subnetworks. Interaction may exist in these
subnetworks. The responsibility of component S is to manage the logic of each subnetwork and handle the interaction
between them. Suppose component S has n in-degree and
m out-degree, i.e., component S receives n inputs and produces m outputs, the k-th output is generated by necessary
inputs and memories:
skt = f ([X, M ])
(5)
where skt is the k-th output at time-step t, X and M are
needed inputs and memories, and f is the nonlinear function
which can be one-layer RNN or two-layer RNN etc.
Component O: To form a layer we need a certain amount
of neurons. So one of the NOR properties is multiple RNNs.
A natural approach to integrate multiple RNN neurons’ signals is collecting all outputs first and then using a MLP layer
to measure the weights of each outputs. Traditional neuron
outputs a single real value, so the collection method is directly arranging them into a vector. But RNN neurons is different, for each of them outputs a vector not a value. A simple method is concatenating all vectors and then connecting
1
The notation of the “subnetwork” is different from the “neuron”, for one subnetwork may be composed of several neurons. We
use superscript i as the identifier of the subnetwork. So, the input
data of i-th subnetwork at time-step t is denoted as xit .
2
In this paper, we just use simple RNN (Elman 1990) applied
with ReLU activation as our basic RNN neuron. Thus the memory
at this time-step is just the output at last time-step.
the concatenated vector to the next MLP. Another is pooling
each RNN output vector into a real value, then arranging all
these real values into a vector, which seems same as traditional neurons. In this paper, the former solution is used and
formalized as:
st = [s1t ; s2t ; · · · ; sm
t ]
ot = r(WM LP ∗ st )
(6)
(7)
where st is the concatenated vector, WM LP is the weight of
MLP and r means the ReLU activation function of MLP.
Methodology I: Aggregation
Any operation with changing a boundary can cause a emergence of complexity. The natural boundary is the agent itself, and sudden emergence of complexity is possible at this
boundary if complexity is transfered from the agent to the
system or vice versa from the system to the agent. There
are two basic operations, aggregation and specialization, that
can be used to transfer complexity between different dimensions (Fromm 2004).
According to Arthur’s second mechanism, internal complexity can be increased by aggregation and composition of
sub-agents, which means a number of RNN agents is conglomerated into a single big system. In this way, aggregation
and composition transfer traditional RNN from the outer to
the inner dimension, from the system to the agent, for the
selected RNNs are accumulated to become a part of a larger
group.
For a concrete NOR layer, suppose it is composed of n
subnetworks, and i-th subnetwork is made up of k i RNN
neurons. Then, at the time-step t, given the input xt , the
operation flow is as follows:
1. Component I: copy xt into n duplications using equation
(3), then we get x1t , x2t , · · · , xnt .
2. Component M: deliver the memory of each RNN neuron from the last time-step to the current time-step using
1,k1
equation (4), then we get memories m1,1
in
t , · · · , mt
2,1
2,k2
first subnetwork and memories mt , · · · , mt
in second subnetwork, etc.
3. Component S: for each subnetwork i, take advantage of
i,ki
the input xit and memories mi,1
to get the nont , · · · , mt
linear transformation output:
i
i,k
sit = f ([xit , mi,1
t , · · · , mt ])
then, we get s1t , s2t , · · · , snt .
(8)
4. Component O: concatenate all outputs by equation (6)
and use a MLP function to determine how much signals
in each subnetwork to flow through the component O by
equation (7).
Obviously, the number, the type and the interaction of the
aggregated RNNs determine the internal structure or inner
complexity of the newly formed layer system. Thus, we propose three kinds of topologies of NOR aggregation method.
Multi-Agent In systems theory, the natural description of
complex system is the multi-agent system created by replication and adaptation. “Replication” means to copy and reproduce a new RNN agent, and “adaptation” means they are
not totally same and some changes on weights or somewhere
else by variation can increase the diversity of the system.
As shown in Figure 2(a), there is a NOR layer (called MANOR) composed of four parallel RNNs. Figure 3 shows this
layer being unrolled into a full network. Each subnetwork
of MA-NOR layer is a one-tier RNN, thus at time-step t, the
i-th subnetwork of component S in MA-NOR is calculated
as:
sit = oit = r(Wi xit + Ui mit )
(9)
where r means ReLU activation function, Wi and Ui are parameters of corresponding RNN neuron, oit is the nonlinear
transformation output and will be delivered to next time-step
to be used as mit+1 , and sit is the output of i-th subnetwork
which is equal to oit .
Figure 4: The unfolding of MS-NOR in three time-steps.
introduce new agent type to the system and can learn sequence dependencies in different timescales.
Figure 2(c) shows a NOR layer made up of four subnetworks, in which two of them are one-tier RNNs and the others are two-tier RNNs. Two kinds of timescale dependencies are learned in component S, which are formalized as
follows:
s1t =o1t = r(Wt1 x1t + Ut1 m1t )
(12)
s2t
(13)
s3t
s4t
=o2t = r(Wt2 x2t + Ut2 m2t )
ot3,1 = r(Wt3,1 x3t + Ut3,1 m3,1
t )
3,2 3,2
3,2 3,1
3,2
=ot = r(Wt ot + Ut mt )
ot4,1 = r(Wt4,1 x4t + Ut4,1 m4,1
t )
4,2
4,2 4,1
4,2 4,2
=ot = r(Wt ot + Ut mt )
(14)
(15)
(16)
(17)
Self-Similarity The above mentioned aggregation and
composition operation lead to big RNN agent-groups. While
in turn, they can also be combined to form even bigger
group. Such repeated aggregation and high accumulation
makes the fractal and self-similar structure come into being.
Figure 3: The unfolding of MA-NOR in three time-steps.
The nonlinear function in equation (5) of each subnetwork
may be more complex. For example, Figure 2(b) shows a
NOR layer made up of three two-tier RNNs. At time-step t,
the i-th subnetwork in component S is calculated as
i,1 i
i,1 i,1
oi,1
t = r(Wt xt + Ut mt )
sit
=oi,2
t
=
r(Wti,2 xit
+
Uti,2 mi,2
t )
(10)
(11)
Multi-Scale The combination of multiple RNNs in a
Multi-Agent NOR layer makes it somewhat like an ensemble. And empirically, diversity among the members of a
group of agents is deemed to be a key issue in ensemble
structure (Kuncheva and Whitaker 2003). One way to increase the diversity is to use the Multi-Scale topology which
Figure 5: The unfolding of SS-NOR in three time-steps.
As shown in Figure 2(d), we also use three paths. But
after each path first learns its own intermediate represen-
tation, the second layers gather all intermediate representations of three paths to learn high-level abstract features.
In this way, different paths do not learn and train independently. The connections among each other helps the model
easy to share informations. Thus, it becomes possible that
the whole model learns and trains to be an organic rather
than parallel independent structure. We formalize the cooperation of component S as follows:
o1,1
= r(Wt1,1 x1t + Ut1,1 m1,1
t
t )
s1t
s2t
s3t
o2,1
t
o3,1
t
1,2
=ot
=o2,2
t
=o3,2
t
=
=
=
=
=
r(Wt2,1 x2t + Ut2,1 m2,1
t )
3,1 3
3,1 3,1
r(Wt xt + Ut mt )
2,1 3,1
r(Wt1,2 [o1,1
t , ot , ot ] +
2,1 3,1
r(Wt2,2 [o1,1
t , ot , ot ] +
2,1 3,1
r(Wt3,2 [o1,1
t , ot , ot ] +
(18)
(19)
(20)
Ut1,2 m1,2
t )
2,2 2,2
Ut mt )
Ut3,2 m3,2
t )
(21)
Figure 6: Gate Specialization.
(22)
(23)
Methodology II: Specialization
We have mentioned that the emergence of complexity is usually connected to a transfer of complexity, a transfer at the
boundary of the system. Aggregation and composition transfer complexity from the system to the agent, and from the
outer dimension to the inner dimension. Another way to be
used to cross the agent boundary is the specialization or inheritance, which transfer complexity from the agent to the
system, and from the inner dimension to the outer dimension (Fromm 2004). Specialization is related to Arthur’s first
mechanism. It increases structural sophistication outside of
the agent by adding new agent forms. Through inheritance
and specialization objects become objects of a certain class
and agents become agents of a certain type, and the more
such an agent becomes a particular class or type, the more it
needs to delegate special tasks that it can not handle alone to
other agents (Fromm 2004).
The effect of specialization is the emergence of delegation
and division of labor in the newly formed groups. Thus, the
formalization of k-th output in Component S can be rewritten as the following:
skt = g(f1 ([X, M ]), f2 ([X, M ]), · · · , fL ([X, M ])) (24)
where fl is the l-th specialized agent function, g means the
cooperation of all specialized agents, and L is the number
of specialized agents. Equation (24) denotes the function f
in equation (5) is implemented by the separated operations
f1 , f2 , · · · , fL and g.
Gate-Specialization We see gate mechanism is one of the
specialization methods. As shown in Figure 6, a general
RNN agent is separated into two specialized RNN agents,
one is for gate duty and the other is for generalization duty.
A concrete Gate-NOR is shown in Figure7. In the original
Multi-Agent NOR layer, each RNN agent is specialized as
one generalization specific RNN and one gate specific RNN.
Figure 7: The sectional views of Gate-NOR layer at one
time-step.
We formalize it as:
o1t = σ(Wt1 x1t + Ut1 m1t )
(25)
o2t
s1t
o3t
o4t
s2t
(26)
=
=
=
=
=
r(Wt2 x2t + Ut2 m2t )
o1t o2t
σ(Wt3 x3t + Ut3 m3t )
r(Wt4 x4t + Ut4 m4t )
o3t o4t
where σ denotes the sigmoid activation and
element-wise multiplication.
(27)
(28)
(29)
(30)
denotes
Relationship with LSTM and GRU We see Long Shortterm Memory (LSTM) (Hochreiter and Schmidhuber 1997)
and Gated Recurrent Unit (GRU) (Chung et al. 2014) as two
special cases of Network of Recurrent neural networks. Take
LSTM for example, at time-step t, given input xt and previous memory cell ct−1 and hidden state ht−1 , the transition
equations of standard LSTM can be expressed as the following:
i = σ(Wi xt + Ui ht−1 )
f = σ(Wi xt + Uf ht−1 )
o = σ(Wo xt + Uo ht−1 )
g = tanh(Wg xt + Ug ht−1 )
ct = ct−1 f + g i
st = tanh(ct ) o
(31)
(32)
(33)
(34)
(35)
(36)
From the perspective of NOR (Network of Recurrent Neural
Networks), LSTM is made up of four RNNs, in which three
Task
Sentiment Classification
Question classification
Named Entity Recognition
# of Params.
200 k
400 k
800 k
100 k
200 k
400 k
200 k
400 k
800 k
IRNN
212
320
468
198
319
497
197
319
497
GRU
107
166
252
86
148
244
86
148
244
LSTM
88
139
213
68
119
199
67
119
199
MA-NOR
89
136
203
74
122
193
74
122
193
MS-NOR
66
100
149
54
88
139
54
88
139
SS-NOR
61
90
132
53
83
126
53
83
126
Gate-NOR
61
97
149
45
79
133
45
79
133
Table 1: Number of hidden neurons for RNN, GRU, LSTM MA-NOR, MS-NOR, SS-NOR and Gate-NOR for each network
size specified in terms of the number of parameters (weights).
of four (i.e., i, f , o RNNs) are specialized for gate tasks to
control how much of informations let through in different
parts. Moreover, there is only a shared memory ht−1 which
can be accessed by each RNN cell in LSTM.
While in turn, LSTM and GRU can also be combined to
form even bigger group.
Experiments
In order to evaluate the performance of the presented model
structures, we design experiments on the following tasks:
sentiment classification, question type classification and
named entity recognition. We compare all models under the
comparable parameter numbers to validate the capacity of
better utilizing the parametric space. In order to verify the
effectiveness and universality of the experiments, we conduct three comparative tests under total parameters of different orders of magnitude, see Table 1. Every experiment
is repeated 20 times with different random initializations and
then we report the mean results. It’s worthy noting that our
aim here is to compare the model performance under the
same hyper-parameter settings, not to achieve best performance for one single model.
(Le, Jaitly, and Hinton 2015) showed that when initializing the recurrent weight matrix to be the identity matrix
and biases to be zero, simple RNN composed of ReLU activation function (named as IRNN) can be comparable with
even outperform LSTM. In our experiments, all basic RNN
neurons are simple RNNs applied with ReLU function. We
also keep the number of the hidden units same over all RNN
neurons in a NOR model. Obviously, our baseline model is a
single giant simple RNN (Elman 1990) applied with ReLU
activation. At the same time, two improved RNNs (GRU
(Chung et al. 2014) and LSTM (Hochreiter and Schmidhuber 1997)) have been widely and successfully used in NLP
in recent years, so we also choose them as our baseline models.
The pre-trained 300-D Glove 6B vectors3 and 300-D
Google News vectors4 were obtained for the word embeddings. During training we fix all word embeddings and
learn only the other parameters in all models. The embeddings for out-of-vocabulary words are set to zero vectors.
We pad or crop the input sentences to a fixed length. The
3
4
https://nlp.stanford.edu/projects/glove/
https://code.google.com/archive/p/word2vec/
trainings are done through stochastic gradient optimizer descent over shuffled mini-batches with the optimizer Adam
(Kingma and Ba 2014). All models are regularized by using
dropout (Srivastava et al. 2014) method. At the same time,
in order to avoid overfitting, early stopping is applied to
prevent unnecessary computation when training. More details on hyper-parameters setting can be found in our codes,
which are publicly available at ‡.
Sentiment Classification
We evaluate our models on the task of sentiment classification on the popular Stanford Sentiment Treebank (SST)
benchmark (Socher et al. 2013), which consists of 11855
movie reviews and is split into train (8544), dev (1101) and
test (2210). SST provides detailed phrase-level annotation
and all sentences along with the phrases are annotated with
5 labels: very positive, positive, neural, negative, and very
negative. In our experiments, we only use the sentence-level
annotation. One of our goals is to avoid expensive phraselevel annotation, like (Qian, Huang, and Zhu 2016). Another is, in practice, phrase-level annotation is hard to provide.
All models use the same architecture: embedding layer →
dropout layer → RNN/NOR layer → RNN/NOR layer →
max-pooling layer → dropout Layer → softmax layer. The
first layer is the word embedding layer, next are two-layer
RNN/NOR layers as the non-linear feature transformation
layer. Then a max-pooling layer max-pools all transformed
feature vectors by selecting the max value in each position
to get sentence representation. Finally, a softmax layer is
used as output layer to get the final result. To benefit from
the regularization, two dropout layers with rate of 0.5 are
added after embedding layer and before softmax layer. The
initial learning rates of all models are set to 0.0002. We
use public available 300-D Glove 840B vectors to initialize
word embeddings. Three different network sizes are tested
for each architecture, such that the number of parameters
are roughly 200 k, 400 k and 800 k (see Table 1). We set
the minibatch size as 20. Finally, we use the Cross-Entropy
criterion as loss function.
The results of the experiments are shown in Table 2. It is
obvious that NOR models get superior performances compared with IRNN baseline, especially when the network size
is big enough. All models improve with network size grows.
Among all NOR models, Gate-NOR gets the best results.
Model
IRNN
GRU
LSTM
MA-NOR
MS-NOR
SS-NOR
Gate-NOR
200k Params
44.30
47.27
46.98
45.17
44.83
44.54
44.95
400k Params
44.43
47.35
47.19
45.24
45.51
45.21
45.80
800k Params
45.28
47.65
47.37
45.37
45.59
45.42
46.06
Table 2: Accuracy (%) comparison over different experiments on SST corpus.
However, we find that LSTM and GRU get much better results in three comparative tests.
2003) which consists of 14987 sentences in the training set,
3466 sentences in the validation set and 3684 sentences in
the test set.
Model
IRNN
GRU
LSTM
MA-NOR
MS-NOR
SS-NOR
Gate-NOR
200k Params
85.07
83.97
84.95
85.96
85.51
86.06
85.67
400k Params
85.52
84.13
85.70
86.17
86.10
86.20
85.89
800k Params
85.58
85.05
86.18
86.18
86.21
86.33
86.21
Table 4: F1 (%) comparison over different experiments on
CoNLL-2003 corpus.
Question Type Classification
Question classification is an important step in a question answering system which classifies a question into a specific
type. For this task, we use TREC (Li and Roth 2002) benchmark, which divides all questions into 6 categories: location,
human, entity, abbreviation, description and numeric. TERC
provides 5452 labeled questions in the training set and 500
questions in the test. We randomly select 10% of the training
data as the validation set.
Model
IRNN
GRU
LSTM
MA-NOR
MS-NOR
SS-NOR
Gate-NOR
100k Params
92.73
91.83
91.44
92.73
92.85
93.19
92.77
200k Params
93.22
92.54
92.46
93.56
93.79
93.62
93.35
400k Params
93.62
93.64
93.10
93.73
94.04
94.12
93.81
Table 3: Accuracy (%) comparison over different experiments on TREC corpus.
All network types use the same architecture: embedding
layer → dropout layer → RNN/NOR layer → max-pooling
layer → dropout layer → softmax layer. Dropout rates are
set to 0.5. Three hidden layer sizes are chosen such that the
total number of parameters for the whole model is roughly
100 k, 200 k, 400 k, see Table 1. All networks use a learning
rate of 0.0005 and are trained to minimize the Cross Entropy
Error.
Table 3 shows the accuracy of the different networks on
the question type classification task. Here again, NOR models get better results than baseline IRNN model. Among
all NOR models, SS-NOR also gets the best result. In this
dataset, we find the performances of LSTM and GRU are
even not comparable with IRNN, which proves the validity
of results in (Le, Jaitly, and Hinton 2015).
Named Entity Recognition
Named entity recognition (NER) is a classic NLP task which
tries to identity the proper names of persons, organizations,
locations, or other entities in the given text. We experiment
on CoNLL-2003 dataset (Tjong Kim Sang and De Meulder
Recently, popular NER models are based on bidirectional
LSTM (Bi-LSTM) combined with conditional random fields
(CRF), named as Bi-LSTM-CRF (Lample et al. 2016). The
Bi-LSTM-CRF networks can effectively use past and future features via a Bi-LSTM layer and sentence level tag
information via a CRF layer. In our experiments, we also
adapt this architecture by replacing LSTM with NORs or
other variation of RNNs. So the universal architecture of all
tested models is: embedding layer → dropout layer → BiRNN/Bi-NOR layer → CRF layer. Three hidden layer sizes
are chosen such that the total number of parameters for the
whole network is roughly 200 k, 400 k and 800 k, see Table 1. We apply 50% dropout after embedding layer. Initial
learning rate is set to 0.005 and every epoch it is reduced
by factor 0.95. The size of each minibatch is 20. We train
all networks for 25 epochs and early stop the training when
there are 5 epochs no improvement on validation set.
Our results are summarized in the Table 4. Not surprisingly, all NORs perform much better than giant single RNNReLU model. As we can see, GRU performs the worst, followed by IRNN. Compared to GRU and IRNN, LSTM performs very well, especially when network size grows up. At
the same time, all NOR models get superior performances
than IRNN, GRU and LSTM. Among them, SS-NOR model
get best results.
Conclusion
In conclusion, we introduced a novel kind of systems theory
based neural networks called “Network Of Recurrent neural network” (NOR) which views existing RNNs (for example, simple RNN, GRU, LSTM) as high-level neurons
and then utilizes RNN neurons to design higher-level layers.
Then we proposed several methodologies to design different NOR topologies according to the evolution of systems
theory (Arthur and others 1993). We conducted experiments
on three kinds of tasks (including sentiment classification,
question type classification and named entity recognition) to
evaluate our proposed models. Experimental results demonstrated that NOR models get superior performances compared with single giant RNN models, and sometimes their
performances even exceed GRU and LSTM.
References
[Anderson 1972] Anderson, P. W. 1972. More is different.
Science 177 4047:393–6.
[Arthur and others 1993] Arthur, W. B., et al. 1993. On the
evolution of complexity. Technical report.
[Bertschinger and Natschläger 2004] Bertschinger, N., and
Natschläger, T. 2004. Real-time computation at the edge
of chaos in recurrent neural networks. Neural computation
16(7):1413–1436.
[Chung et al. 2014] Chung, J.; Gulcehre, C.; Cho, K.; and
Bengio, Y. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv preprint
arXiv:1412.3555.
[Elman 1990] Elman, J. L. 1990. Finding structure in time.
Cognitive science 14(2):179–211.
[Fromm 2004] Fromm, J. 2004. The emergence of complexity. Kassel university press Kassel.
[Hannebauer 2002] Hannebauer, M. 2002. Autonomous dynamic reconfiguration in multi-agent systems: improving
the quality and efficiency of collaborative problem solving.
Springer-Verlag.
[Hochreiter and Schmidhuber 1997] Hochreiter, S., and
Schmidhuber, J. 1997. Long short-term memory. Neural
computation 9(8):1735–1780.
[Kingma and Ba 2014] Kingma, D., and Ba, J. 2014. Adam:
A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
[Kuncheva and Whitaker 2003] Kuncheva, L. I., and
Whitaker, C. J. 2003. Measures of diversity in classifier ensembles and their relationship with the ensemble accuracy.
Machine learning 51(2):181–207.
[Lample et al. 2016] Lample, G.; Ballesteros, M.; Subramanian, S.; Kawakami, K.; and Dyer, C. 2016. Neural architectures for named entity recognition. arXiv preprint
arXiv:1603.01360.
[Le, Jaitly, and Hinton 2015] Le, Q. V.; Jaitly, N.; and Hinton, G. E. 2015. A simple way to initialize recurrent networks of rectified linear units. arXiv preprint
arXiv:1504.00941.
[Lehn 2002] Lehn, J.-M. 2002. Toward complex matter:
supramolecular chemistry and self-organization. Proceedings of the National Academy of Sciences 99(8):4763–4768.
[Li and Roth 2002] Li, X., and Roth, D. 2002. Learning
question classifiers. In Proceedings of the 19th international
conference on Computational linguistics-Volume 1, 1–7. Association for Computational Linguistics.
[Mele, Pels, and Polese 2010] Mele, C.; Pels, J.; and Polese,
F. 2010. A brief review of systems theories and their managerial applications. Service Science 2(1-2):126–135.
[Pascanu, Mikolov, and Bengio 2013] Pascanu, R.; Mikolov,
T.; and Bengio, Y. 2013. On the difficulty of training recurrent neural networks. In ICML.
[Qian, Huang, and Zhu 2016] Qian, Q.; Huang, M.; and Zhu,
X. 2016. Linguistically regularized lstms for sentiment classification. arXiv preprint arXiv:1611.03949.
[Socher et al. 2013] Socher, R.; Perelygin, A.; Wu, J. Y.;
Chuang, J.; Manning, C. D.; Ng, A. Y.; Potts, C.; et al. 2013.
Recursive deep models for semantic compositionality over a
sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP),
volume 1631, 1642. Citeseer.
[Sontag 1997] Sontag, E. 1997. Recurrent neural networks:
Some systems-theoretic aspects. In Dealing with Complexity: a Neural Network Approach. Citeseer.
[Srivastava et al. 2014] Srivastava, N.; Hinton, G. E.;
Krizhevsky, A.; Sutskever, I.; and Salakhutdinov, R.
2014. Dropout: a simple way to prevent neural networks
from overfitting. Journal of Machine Learning Research
15(1):1929–1958.
[Thompson and Stewart 2002] Thompson, J. M. T., and
Stewart, H. B. 2002. Nonlinear dynamics and chaos. John
Wiley & Sons.
[Tjong Kim Sang and De Meulder 2003] Tjong Kim Sang,
E. F., and De Meulder, F. 2003. Introduction to the
conll-2003 shared task: Language-independent named entity recognition. In Proceedings of the seventh conference
on Natural language learning at HLT-NAACL 2003-Volume
4, 142–147. Association for Computational Linguistics.
[Upton, Janeka, and Ferraro 2014] Upton, J.; Janeka, I.; and
Ferraro, N. 2014. The whole is more than the sum of
its parts: aristotle, metaphysical. Journal of Craniofacial
Surgery 25(1):59–63.
[Von Bertalanffy 1968] Von Bertalanffy, L. 1968. General
system theory. New York 41973(1968):40.
[Von Bertalanffy 1972] Von Bertalanffy, L. 1972. The history and status of general systems theory. Academy of Management Journal 15(4):407–426.
[Walonick 1993] Walonick, D. S.
1993.
General
systems theory.
Information on http://www. statpac.
org/walonick/systems-theory. htm.
[Werbos 1988] Werbos, P. J. 1988. Generalization of
backpropagation with application to a recurrent gas market
model. Neural Networks 1:339–356.
[Weston, Chopra, and Bordes 2014] Weston, J.; Chopra, S.;
and Bordes, A. 2014. Memory networks. arXiv preprint
arXiv:1410.3916.
[Wierzbicki 2015] Wierzbicki, A. P. 2015. Systems theory,
theory of chaos, emergence. In Technen: Elements of Recent
History of Information Technologies with Epistemological
Conclusions. Springer. 175–188.
| 9cs.NE
|
ON THE SECOND COHOMOLOGY OF NILPOTENT ORBITS IN
EXCEPTIONAL LIE ALGEBRAS
arXiv:1611.02587v1 [math.GR] 8 Nov 2016
PRALAY CHATTERJEE AND CHANDAN MAITY
Abstract. In [BC], the second de Rham cohomology groups of nilpotent orbits in all the complex
simple Lie algebras are described. In this paper we consider non-compact non-complex exceptional
Lie algebras, and compute the dimensions of the second cohomology groups for most of the nilpotent
orbits. For the rest of cases of nilpotent orbits, which are not covered in the above computations,
we obtain upper bounds for the dimensions of the second cohomology groups.
1. Introduction
Let G be a connected real simple Lie group with Lie algebra g. An element X ∈ g is called nilpotent if ad(X) : g → g is a nilpotent operator. Let OX := {Ad(g)X | g ∈ G} be the corresponding
nilpotent orbit under the adjoint action of G on g. Such nilpotent orbits form a rich class of homogeneous spaces, and they are studied at the interface of several disciplines in mathematics such
as Lie theory, symplectic geometry, representation theory, algebraic geometry. Various topological
aspects of such orbits have drawn attention over the years; see [CM], [M] and references therein
for an account. In [BC, Proposition 1.2] for a large class of semisimple Lie groups a criterion is
given for the exactness of the Kostant-Kirillov two form on arbitrary adjoint orbits which in turn
led the authors asking the natural question of describing the full second cohomology groups of
such orbits. Towards this, in [BC], the second cohomology groups of nilpotent orbits in all the
complex simple Lie algebras, under the adjoint actions of the corresponding complex groups, are
computed. In this paper we continue the program of studying the second cohomology groups of
nilpotent orbits which was initiated in [BC]. We compute the second cohomology groups for most
of the nilpotent orbits in non-compact non-complex exceptional Lie algebras, and for the rest of
the nilpotent orbits in non-compact non-complex exceptional Lie algebras we give upper bounds
of the dimensions of second cohomology groups; see Theorems 3.2, 3.3, 3.4, 3.5, 3.6, 3.7, 3.8, 3.9,
3.10, 3.11, 3.12, 3.13. In particular, our computations yield that the second cohomologies vanish
for all the nilpotent orbits in F4(−20) and E6(−26) .
2. Notation and background
In this section we fix some general notation, and mention a basic result which will be used in
this paper. A few specialized notation are defined as and when they occur later.
The center of a Lie algebra g is denoted by z(g). We denote Lie groups by capital letters, and
unless mentioned otherwise we denote their Lie algebras by the corresponding lower case German
letters. Sometimes, for convenience, the Lie algebra of a Lie group G is also denoted by Lie(G).
The connected component of a Lie group G containing the identity element is denoted by G◦ . For
a subgroup H of G and a subset S of g, the subgroup of H that fixes S point wise is called the
centralizer of S in H and is denoted by ZH (S). Similarly, for a Lie subalgebra h ⊂ g and a subset
2010 Mathematics Subject Classification. 57T15, 17B08.
Key words and phrases. Nilpotent orbits, Exceptional Lie algebras, Second cohomology.
1
2
P. CHATTERJEE AND C. MAITY
S ⊂ g, by zh (S) we will denote the subalgebra consisting elements of h that commute with every
element of S.
If G is a Lie group with Lie algebra g, then it is immediate that the coadjoint action of G◦
on z(k)∗ is trivial; in particular, one obtains a natural action of G/G◦ on z(k)∗ . We denote by
◦
[z(g)∗ ]G/G the space of fixed points of z(g)∗ under the action of G/G◦ .
For a real semisimple Lie group G, an element X ∈ g is called nilpotent if ad(X) : g → g is
a nilpotent operator. A nilpotent orbit is an orbit of a nilpotent element in g under the adjoint
representation of G; for a nilpotent element X ∈ g the corresponding nilpotent orbit Ad(G)X is
denoted by OX .
For a g be a Lie algebra over R a subset {X, H, Y } ⊂ g is said to be sl2 (R)-triple if X 6= 0,
[H, X] = 2X, [H, Y ] = −2Y and [X, Y ] = H. It is immediate that, if {X, H, Y } ⊂ g is a sl2 (R)triple then SpanR {X, H, Y } is a R-subalgebra of g which is isomorphic to the Lie algebra sl2 (R).
We now recall the well-known Jacobson-Morozov theorem (see [CM, Theorem 9.2.1]) which ensures
that if X ∈ g is a non-zero nilpotent element in a real semisimple Lie algebra g, then there exist
H, Y in g such that {X, H, Y } ⊂ g is a sl2 (R)-triple.
To facilitate the computations in §3 we need the following result.
Theorem 2.1. Let G be an algebraic group defined over R which is R-simple. Let X ∈ Lie G(R),
X 6= 0 be a nilpotent element, and OX be the orbit of X under the adjoint action of the identity
component G(R)◦ on Lie G(R). Let {X, H, Y } be a sl2 (R)-triple in Lie G(R). Let K be a maximal
compact subgroup in ZG(R)◦ (X, H, Y ), and M be a maximal compact subgroup in G(R)◦ containing
K. Then
◦
H 2 (OX , R) ≃ [(z(k) ∩ [m, m])∗ ]K/K .
In particular, dimR H 2 (OX , R) ≤ dimR z(k).
The above theorem follows from [CM, Lemma 3.7.3] and a description of the second cohomology
groups of homogeneous spaces which generalizes [BC, Theorem 3.3]. The details of the proof of
Theorem 2.1 and the generalization of [BC, Theorem 3.3], mentioned above, will appear elsewhere.
3. The second cohomology groups of nilpotent orbits
In this section we study the second cohomology of the nilpotent orbits in non-compact noncomplex exceptional Lie algebras over R. The results in this section depend on the results of [D1,
Tables VI-XV], [D2, Tables VII-VIII] and [K, Tables 1-12]. We refer to [CM, Chapter 9], [D1]
and [D2] for the generalities required in this section. We begin by recalling the parametrization of
nilpotent orbits in this set-up.
3.1. Parametrization of nilpotent orbits in exceptional Lie algebras
We follow the parametrization of nilpotent orbits in non-compact non-complex exceptional Lie
algebras as given in [D1, Tables VI-XV] and [D2, Tables VII-VIII]. We consider the nilpotent orbits
in g under the action of Int g, where g is a non-compact non-complex real exceptional Lie algebra.
We fix a semisimple algebraic group G defined over R such that g = Lie(G(R)). Here G(R) denotes
the associated real semisimple Lie group of the R-points of G. Let G(C) be the associated complex
semisimple Lie group consisting of the C-points of G. It is easy to see that orbits in g under the
action of Int g are the same as the orbits in g under the action of G(R)◦ . Thus in this set-up, for
a nilpotent element X ∈ g, we set OX := {Ad(g)X | g ∈ G(R)◦ }. Let g = m + p be a Cartan
decomposition and θ be the corresponding Cartan involution. Let gC be the Lie algebra of G(C).
Then gC can be identified with the complexification of g. Let mC , pC be the C-spans of m and p
in gC , respectively. Then gC = mC + pC . Let MC be the connected subgroup of G(C) with Lie
ON THE NILPOTENT ORBITS IN LIE ALGEBRAS
3
algebra mC . Recall that, if g is as above and g is different from both E6(−26) and E6(6) , then g is
of inner type, or equivalently, rank mC = rank gC . When g is of inner type, the nilpotent orbits
are parametrized by a finite sequence of integers of length l where l := rank mC = rank gC . When
g is not of inner type, that is, when g is either E6(−26) or E6(6) , then the nilpotent orbits are
parametrized by a finite sequence of integers of length 4.
Let X ′ ∈ g be a nonzero nilpotent element, and {X ′ , H ′ , Y ′ } ⊂ g be a sl2 (R)-triple. Then
e H,
e Ye } in g such that θ(H)
e = −H,
e
{X ′ , H ′ , Y ′ } is G(R)-conjugate to another sl2 (R)-triple {X,
e
e
e
e
e
e
e
e
e
e
θ(X) = −Y . Set E := (H − i(X + Y ))/2, F := (H + i(X + Y ))/2 and H := i(X − Y ). Then
{E, H, F } is a sl2 (R)-triple and E, F ∈ pC and H ∈ mC . The sl2 (R)-triple {E, H, F } is then called
a pC -Cayley triple associated to X ′ .
3.1.1. Parametrization in exceptional Lie algebras of inner type. We now recall from [D1, Column
2, Tables VI-XV] the parametrization of non-zero nilpotent orbits in g when g is an exceptional
Lie algebra of inner type. Let hC ⊂ mC be a Cartan subalgebra of mC such that hC ∩ m is a Cartan
subalgebra of m. As g is of inner type, hC is a Cartan subalgebra of gC . Set h := hC ∩ im. Let
R, R0 be the root systems of (gC , hC ), (mC , hC ), respectively. Let B := {α1 , . . . , αl } be a basis of
R. Let Be := B ∪ {α0 } where α0 is the negative of the highest root of (R, B). Then there exists
an unique basis of R0 , say B0 , such that B0 ⊂ Be . Let C0 be the closed Weyl chamber of R0 in
h corresponding to the basis B0 . Let l0 be the rank of [mC , mC ]. Then either l0 = l or l0 = l − 1.
If l0 = l we set B0′ := B0 . If l0 = l − 1 (in this case we have B0 ⊂ B) we set B0′ := B. Clearly,
#B0′ = l. We enumerate B0′ := {β1 , . . . , βl } as in [D1, 7, p. 506 and Table IV]. Let X ∈ g be
a nonzero nilpotent element, and {E, H, F } be a pC -Cayley triple (in gC ) associated to X. Then
Ad(MC )H ∩ C0 is a singleton set, say {H0 }. The element H0 is called the characteristic of the orbit
Ad(MC )E as it determines the orbit MC · E uniquely. Consider the map from the set of nilpotent
orbits in g to the set of integer sequences of length l, which assigns the sequence β1 (H0 ), . . . , βl (H0 )
to each nilpotent orbits OX . In view of the Kostant-Sekiguchi theorem (cf. [CM, Theorem 9.5.1]),
this gives a bijection between the set of nilpotent orbits in g and the set of finite sequences of the
form β1 (H0 ), . . . , βl (H0 ) as above. We use this parametrization while dealing with nilpotent orbits
in exceptional Lie algebras of inner type.
3.1.2. Parametrization in E6(−26) or E6(6) . We now recall from [D2, Column 1, Tables VII-VIII]
the parametrization of non-zero nilpotent orbits in g when g is either E6(−26) or E6(6) . We need a
piece of notation here : henceforth, for a Lie algebra a over C and an automorphism σ ∈ AutC a,
the Lie subalgebra consisting of the fixed points of σ in a, is denoted by aσ . Let now hC be a Cartan
subalgebra of gC (we point out the difference of our notation with that in [D2]; g and h of [D2, §1]
are denoted here by gC and hC , respectively).
Let g = E6(−26) . Let τ be the involution of gC as defined in [D2, p. 198 ] which keeps hC
invariant. Then the subalgebra gτC is of type F4 , and hτC is a Cartan subalgebra of gτC . Let G(C)τ
be the connected Lie subgroup of G(C) with Lie algebra gτC . Let {β1 , β2 , β3 , β4 } be the simple
roots of (gτC , hτC ) as defined in [D2, (1), p. 198]. Let X ∈ E6(−26) be a nonzero nilpotent element.
Let {E, H, F } be a pC -Cayley triple (in gC ) associated to X. Then H ∈ gτC and E, F ∈ g−τ
C . We
may further assume that H ∈ hτC . Then the finite sequence of integers β1 (H), β2 (H), β3 (H), β4 (H)
determine the orbit Ad(G(C)τ ) E uniquely; see [D2, p. 204].
Let g = E6(6) . Let τ ′ be the involution of gC as defined in [D2, p. 199 ] which keeps hC invariant.
′
′
′
′
Then the subalgebra gτC is of type C4 , and hτC is a Cartan subalgebra of gτC . Let G(C)τ be the
′
connected Lie subgroup of G(C) with Lie algebra gτC . Let {β0 , β1 , β2 , β3 } be the simple roots of
′
′
(gτC , hτC ) as defined in [D2, p. 199]. Let X ∈ E6(6) be a nonzero nilpotent element. Let {E, H, F }
′
−τ ′
. We may further
be a pC -Cayley triple (in gC ) associated to X. Then H ∈ gτC and E, F ∈ gC
4
P. CHATTERJEE AND C. MAITY
′
assume that H ∈ hτC . It then follows that the finite sequence of integers β0 (H), β1 (H), β2 (H), β3 (H)
′
determine the orbit Ad(G(C)τ ) E uniquely; see [D2, p. 204].
3.2. Nilpotent orbits of three types
For the sake of convenience of writing the proofs that appear in the later part §3, it will be
useful to divide the nilpotent orbits in the following three types. Let X ∈ g be a nonzero nilpotent
element, and {X, H, Y } be a sl2 (R)-triple in g. Let G be as in the beginning of §3.1. Let K be a
maximal compact subgroup in ZG(R)◦ (X, H, Y ), and M be a maximal compact subgroup in G(R)◦
containing K. A nonzero nilpotent orbit OX in g is said to be of
(1) type I if z(k) 6= 0, K/K ◦ = Id and m = [m, m];
(2) type II if either z(k) 6= 0, K/K ◦ 6= Id, m = [m, m]; or z(k) 6= 0, m 6= [m, m];
(3) type III if z(k) = 0.
In what follows we will use the next result repeatedly.
Corollary 3.1. Let g be a real simple non-compact exceptional Lie algebra. Let X ∈ g be a nonzero
nilpotent element.
(1) If the orbit OX is of type I, then dimR H 2 (OX , R) = dimR z(k).
(2) If the orbit OX is of type II, then dimR H 2 (OX , R) ≤ dimR z(k).
(3) If the orbit OX is of type III, then dimR H 2 (OX , R) = 0.
Proof. The proof of the corollary follows immediately from Theorem 2.1.
Let g be as above. In the proofs of our results in the following subsections we use the description
of a Levi factor of zg (X) for each nilpotent element X in g, as given in the last columns of [D1,
Tables VI-XV] and [D2, Tables VII-VIII]. This enables us compute the dimensions dimR z(k) easily.
We also use [K, Column 4, Tables 1-12] for the component groups for each nilpotent orbits in g.
3.3. Nilpotent orbits in the non-compact real form of G2
Recall that up to conjugation there is only one non-compact real form of G2 . We denote it by
G2(2) . There are only five nonzero nilpotent orbits in G2(2) ; see [D1, Table VI, p. 510]. Note that
in this case we have m = [m, m].
Theorem 3.2. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in G2(2) .
(1) If the parametrization of the orbit OX is given by either 1 1 or 1 3, then dimR H 2 (OX , R)
= 1.
(2) If the parametrization of the orbit OX is given by any of 2 2, 0 4, 4 8, then dimR H 2 (OX , R)
= 0.
Proof. From [D1, Column 7, Table VI, p. 510] we have dimR z(k) = 1 and from [K, Column 4,
Table 1, p. 247] we have K/K ◦ = Id for the nilpotent orbits as in (1). Thus these are of type I.
We refer to [D1, Column 7, Table VI, p. 510] for the orbits as given in (2). These orbits are of
type III as dimR z(k) = 0. In view of the Corollary 3.1 the conclusions follow.
3.4. Nilpotent orbits in non-compact real forms of F4
Recall that up to conjugation there are two non-compact real forms of F4 . They are denoted by
F4(4) and F4(−20) .
ON THE NILPOTENT ORBITS IN LIE ALGEBRAS
5
3.4.1. Nilpotent orbits in F4(4) . There are 26 nonzero nilpotent orbits in F4(4) ; see [D1, Table VII,
p. 510]. Note that in this case we have m = [m, m].
Theorem 3.3. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in F4(4) .
(1) Assume the parametrization of the orbit OX is given by any of the sequences :
001 1, 001 3, 110 2, 111 1, 131 3. Then dimR H 2 (OX , R) = 1.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
100 2, 200 0, 103 1, 111 3, 204 4. Then dimR H 2 (OX , R) ≤ 1.
(3) If the parametrization of the orbit OX is either 101 1 or 012 2, then dimR H 2 (OX , R) ≤ 2.
(4) If OX is not given by the parametrizations as in (1), (2), (3) above (# of such orbits are
14), then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra F4(4) , we can easily compute dimR z(k) from the last column of [D1,
Table VII, p. 510] and K/K ◦ from [K, Column 4, Table 2, pp. 247-248].
For the orbits OX , as in (1), we have dimR z(k) = 1 and K/K ◦ = Id. Hence these are of type I.
For the orbits OX , as in (2), we have dimR z(k) = 1 and K/K ◦ 6= Id; hence they are of type II. For
the orbits OX , as in (3), we have dimR z(k) = 2 and K/K ◦ 6= Id. Hence these are also of type II.
The rest of the 14 orbits, which are not given by the parametrizations in (1), (2), (3), are of type
III as z(k) = 0. Now the theorem follows from Corollary 3.1.
3.4.2. Nilpotent orbits in F4(−20) . There are two nonzero nilpotent orbits in F4(−20) ; see [D1, Table
VIII, p. 511].
Theorem 3.4. For all the nilpotent elements X in F4(−20) we have dimR H 2 (OX , R) = 0.
Proof. As the theorem follows trivially when X = 0 we assume that X 6= 0. We follow the
parametrization of nilpotent orbits as in §3.1.1. From the last column of [D1, Table VIII, p. 511]
we conclude that z(k) = 0. Hence the nonzero nilpotent orbits are of type III. Using Corollary 3.1
(3) we have dimR H 2 (OX , R) = 0.
3.5. Nilpotent orbits in non-compact real forms of E6
Recall that up to conjugation there are four non-compact real forms of E6 . They are denoted
by E6(6) , E6(2) , E6(−14) and E6(−26) .
3.5.1. Nilpotent orbits in E6(6) . There are 23 nonzero nilpotent orbits in E6(6) ; see [D2, Table VIII,
p. 205]. Note that in this case we have m = [m, m].
Theorem 3.5. Let the parametrization of the nilpotent orbits be as in §3.1.2. Let X be a nonzero
nilpotent element in E6(6) .
(1) If the parametrization of the orbit OX is given by either 1001 or 1101 or 1211, then
dimR H 2 (OX , R) = 1.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
0102, 0202, 1010, 2002, 1011. Then dimR H 2 (OX , R) ≤ 1.
(3) If OX is not given by the parametrizations as in (1), (2) above (# of such orbits are 15),
then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra E6(6) , we can easily compute dimR z(k) from the last column of [D2,
Table VIII, p. 205] and K/K ◦ from [K, Column 4, Table 4, p.253]. As pointed out in the 1st
paragraph of [K, p. 254], there is an error in row 5 of [D2, Table VIII, p. 205]. Thus when OX is
given by the parametrization 2000 it follows from [K, p. 254] that z(k) = 0.
6
P. CHATTERJEE AND C. MAITY
We have dimR z(k) = 1 and K/K ◦ = Id for the orbits given in (1). Thus these orbits are of type
I. For the orbits, as in (2), we have dimR z(k) = 1 and K/K ◦ = Z2 . Hence, the orbits in (2) are of
type II. For rest of the 15 nonzero nilpotent orbits, which are not given by the parametrizations of
(1), (2), are of type III as dimR z(k) = 0. Now the results follow from Corollary 3.1.
3.5.2. Nilpotent orbits in E6(2) . There are 37 nonzero nilpotent orbits in E6(2) ; see [D1, Table IX,
p. 511]. Note that in this case we have m = [m, m].
Theorem 3.6. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E6(2) .
(1) Assume the parametrization of the orbit OX is given by any of the sequences :
00000 4, 00200 2, 02020 0, 00400 8, 22222 2, 04040 4, 44044 4, 44444 8.
Then dimR H 2 (OX , R) = 0.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
10001 2, 10101 1, 21001 1, 10012 1, 11011 2, 01210 2, 10301 1, 11111 3, 22022 0.
Then dimR H 2 (OX , R) = 2.
(3) If the parametrization of the orbit OX is given by either 20002 0 or 00400 0 or 02020 4,
then dimR H 2 (OX , R) ≤ 2.
(4) If the parametrization of the orbit OX is given by 20202 2, then dimR H 2 (OX , R) ≤ 1.
(5) If OX is not given by the parametrizations as in (1), (2), (3), (4) above (# of such orbits
are 16), then we have dimR H 2 (OX , R) = 1.
Proof. For the Lie algebra E6(2) , we can easily compute dimR z(k) from the last column of [D1,
Table IX, p. 511] and K/K ◦ from [K, Column 4, Table 5, pp. 255-256].
We have z(k) = 0 for the orbits, as given in (1), and these orbits are of type III. For the orbits,
as given in (2), we have dimR z(k) = 2 and K/K ◦ = Id. Thus the orbits in (2) are of type I. For the
orbits, as given in (3), we have dimR z(k) = 2 and K/K ◦ 6= Id, hence are of type II. For the orbits,
as given in (4), we have dimR z(k) = 1 and K/K ◦ = Z2 . Thus this orbit is of type II. For the rest
of 16 orbits, which are not given in any of (1), (2), (3), (4), we have dimR z(k) = 1 and K/K ◦ = Id.
Thus these orbits are of type I. Now the conclusions follow from Corollary 3.1.
3.5.3. Nilpotent orbits in E6(−14) . There are 12 nonzero nilpotent orbits in E6(−14) ; see [D1, Table
X, p. 512]. Note that in this case m ≃ so10 ⊕ R, and hence [m, m] 6= m.
Theorem 3.7. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E6(−14) .
(1) If the parametrization of the orbit OX is given by 40000 − 2, then dimR H 2 (OX , R) = 0.
(2) If OX is not given by the above parametrization (# of such orbits are 11), then we have
dimR H 2 (OX , R) ≤ 1.
Proof. For the Lie algebra E6(−14) , we can easily compute dimR z(k) from the last column of
[D1, Table X, p. 512]. The orbit in (1) is of type III as z(k) = 0, and hence dimR H 2 (OX , R) = 0.
The other 11 orbits are of type II as dimR z(k) = 1 and m 6= [m, m]. Hence dimR H 2 (OX , R) ≤ 1.
3.5.4. Nilpotent orbits in E6(−26) . There are two nonzero nilpotent orbits in E6(−26) ; see [D2, Table
VII, p. 204].
Theorem 3.8. For all the nilpotent element X in E6(−26) we have dimR H 2 (OX , R) = 0.
ON THE NILPOTENT ORBITS IN LIE ALGEBRAS
7
Proof. As the theorem follows trivially when X = 0 we assume that X 6= 0. We follow the
parametrization of the nilpotent orbits as given in §3.1.2. The two nonzero nilpotent orbits in
E6(−26) are of type III as z(k) = 0; see last column of [D2, Table VII, p. 204]. Hence, by Corollary
3.1 (3) we conclude that dimR H 2 (OX , R) = 0.
3.6. Nilpotent orbits in non-compact real forms of E7
Recall that up to conjugation there are three non-compact real forms of E7 . They are denoted
by E7(7) , E7(−5) and E7(−25) .
3.6.1. Nilpotent orbits in E7(7) . There are 94 nonzero nilpotent orbits in E7(7) ; see [D1, Table XI,
pp. 513-514]. Note that in this case we have m = [m, m].
Theorem 3.9. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E7(7) .
(1) If the parametrization of the orbit OX is given by 1011101, then dimR H 2 (OX , R) = 3.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
1001001, 1101011, 1111010, 0101111, 2200022, 3101021, 1201013, 1211121, 2204022.
Then dimR H 2 (OX , R) = 2.
(3) Assume the parametrization of the orbit OX is given by any of the sequences :
0100010, 1100100, 0010011, 3000100, 0010003, 0102010, 0200020, 2004002, 2103101,
1013012, 2020202, 1311111, 1111131, 1310301, 1030131, 2220222, 3013131, 1313103,
3113121, 1213113, 4220224, 3413131, 1313143, 4224224.
Then dimR H 2 (OX , R) = 1.
(4) Assume the parametrization of the orbit OX is given by any of the sequences :
2000002, 0101010, 2002002, 1110111, 2020020, 0200202, 1112111, 2022020, 0202202,
2202022, 0220220. Then dimR H 2 (OX , R) ≤ 1.
(5) Assume the parametrization of the orbit OX is given by any of the sequences :
2010001, 1000102, 0120101, 1010210, 1030010, 0100301, 3013010, 0103103.
Then dimR H 2 (OX , R) ≤ 2.
(6) If the parametrization of the orbit OX is given by either 1010101 or 0020200, then
dimR H 2 (OX , R) ≤ 3.
(7) If OX is not given by the parametrizations as in (1), (2), (3), (4), (5), (6) above (# of
such orbits are 39), then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra E7(7) , we can easily compute dimR z(k) from the last column of [D1,
Table XI, pp. 513-514] and K/K ◦ from [K, Column 4, Table 8, pp. 260-264].
The orbit OX , as given in (1), is of type I as dimR z(k) = 3 and K/K ◦ = Id. For the orbits, as
given in (2), we have dimR z(k) = 2 and K/K ◦ = Id. Hence these are also of type I. For the orbits,
as given in (3), we have dimR z(k) = 1 and K/K ◦ = Id; hence they are of type I. For the orbits,
as given in (4), we have dimR z(k) = 1 and K/K ◦ = Z2 . Thus these are of type II. For the orbits,
as given in (5), we have dimR z(k) = 2 and K/K ◦ = Z2 . Hence these are also of type II. For the
orbits, as given in (6), we have dimR z(k) = 3 and K/K ◦ 6= Id, hence they are of type II. Rest of
the 39 orbits, which are not given by the parametrizations in (1), (2), (3), (4), (5), (6), are of type
III as z(k) = 0. Now the results follow from Corollary 3.1.
3.6.2. Nilpotent orbits in E7(−5) . There are 37 nonzero nilpotent orbits in E7(−5) ; see [D1, Table
XII, p. 515]. Note that in this case m = [m, m].
8
P. CHATTERJEE AND C. MAITY
Theorem 3.10. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E7(−5) .
(1) If the parametrization of the orbit OX is given by either 110001 1 or 000120 2, then
dimR H 2 (OX , R) = 2.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
000010 1, 010000 2, 000010 3, 010010 1, 200100 0, 010100 2, 000200 0, 010110 1,
010030 1, 010110 3, 201031 4, 010310 3. Then dimR H 2 (OX , R) = 1.
(3) If the parametrization of the orbit OX is given by either 020200 0 or 111110 1, then
dimR H 2 (OX , R) ≤ 2.
(4) Assume the parametrization of the orbit OX is given by any of the sequences :
020000 0, 201011 2, 040000 4, 040400 4. Then dimR H 2 (OX , R) ≤ 1.
(5) If OX is not given by the parametrizations as in (1), (2), (3), (4) above (# of such orbits
are 17), then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra E7(−5) , we can easily compute dimR z(k) from the last column of [D1,
Table XII, pp. 515] and K/K ◦ from [K, Column 4, Table 9, pp. 266-268].
For the orbit OX , as in (1), we have dimR z(k) = 2 and K/K ◦ = Id. Hence these orbits are of
type I. For the orbit OX , as in (2), we have dimR z(k) = 1 and K/K ◦ = Id. Hence these orbits are
also of type I. For the orbit OX , as in (3), we have dimR z(k) = 2 and K/K ◦ = Z2 , hence are of
type II. For the orbit OX , as in (4), we have dimR z(k) = 1 and K/K ◦ = Z2 . Hence these are also
of type II. Rest of the 17 orbits, which are not given by the parametrizations in (1), (2), (3), (4),
are of type III as z(k) = 0. Now the conclusions follow from Corollary 3.1.
3.6.3. Nilpotent orbits in E7(−25) . There are 22 nonzero nilpotent orbits in E7(−25) ; see [D1, Table
XIII, p. 516]. In this case we have m 6= [m, m].
Theorem 3.11. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E7(−25) .
(1) Assume the parametrization of the orbit OX is given by any of the sequences :
000000 2, 000000 − 2, 000002 − 2, 200000 − 2, 200002 − 2, 400000 − 2, 000004 − 6,
200002 − 6, 400004 − 6, 400004 − 10. Then dimR H 2 (OX , R) = 0.
(2) If OX is not given by any of the above parametrization (# of such orbits are 12), then we
have dimR H 2 (OX , R) ≤ 1.
Proof. Note that the parametrization of nilpotent orbits in E7(−25) as in [K, Table 10] is different
from [D1, Table X III, p. 516]. As the component group for all orbits in E7(−25) is Id; see [K,
Column 4, Table 10, pp. 269-270], it does not depend on the parametrization. We refer to the last
column of [D1, Table X III] for the orbits as given in (1). These are type III as z(k) = 0. For rest
of the 12 orbits we have dimR z(k) = 1; see last column of [D1, Table X III]. As m 6= [m, m], these
are of type II. Now the results follow from Corollary 3.1.
3.7. Nilpotent orbits in non-compact real forms of E8
Recall that up to conjugation there are two non-compact real forms of E8 . They are denoted by
E8(8) and E8(−24) .
3.7.1. Nilpotent orbits in E8(8) . There are 115 nonzero nilpotent orbits in E8(8) ; see [D1, Table
XIV, pp. 517-519]. Note that in this case we have m = [m, m].
ON THE NILPOTENT ORBITS IN LIE ALGEBRAS
9
Theorem 3.12. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E8(8) .
(1) Assume the parametrization of the orbit OX is given by any of the sequences :
10010011, 11110010, 10111011, 11110130. Then dimR H 2 (OX , R) = 2.
(2) Assume the parametrization of the orbit OX is given by any of the sequences :
01000010, 10001000, 30000001, 10010001, 01010010, 01000110, 10100100, 00100003,
11001030, 10110100, 21010100, 01020110, 30001030, 11010101, 11101011, 11010111,
11111101, 21031031, 31010211, 12111111, 13111101, 13111141, 13103041, 31131211,
13131043, 34131341. Then dimR H 2 (OX , R) = 1.
(3) If the parametrization of the orbit OX is given 00100101, then dimR H 2 (OX , R) ≤ 3.
(4) Assume the parametrization of the orbit OX is given by any of the sequences :
10001002, 10101001, 01200100, 02000200, 10101021, 10102100, 02020200, 01201031.
Then dimR H 2 (OX , R) ≤ 2.
(5) Assume the parametrization of the orbit OX is given by any of the sequences :
11000001, 20010000, 01000100, 11001010, 20100011, 01010100, 02020000, 20002000,
20100031, 10101011, 00200022, 11110110, 01011101, 01003001, 11101101, 11101121,
10300130, 04020200, 02002022, 00400040, 11121121, 30130130, 02022022, 40040040.
Then dimR H 2 (OX , R) ≤ 1.
(6) If OX is not given by the parametrizations as in (1), (2), (3), (4), (5) above (# of such
orbits are 52), then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra E8(8) , we can easily compute dimR z(k) from the last column of [D1,
Table XIV, pp. 517-519] and K/K ◦ from [K, Column 4, Table 11, pp. 271-275].
For the orbits OX , as given in (1), we have dimR z(k) = 2 and K/K ◦ = Id. Hence these orbits
are of type I. For the orbits OX , as given in (2), we have dimR z(k) = 1 and K/K ◦ = Id. Hence
these orbits are also of type I. For the orbit OX , as given in (3), we have dimR z(k) = 3 and
K/K ◦ 6= Id; hence they are of type II. For the orbits OX , as given in (4), we have dimR z(k) = 2
and K/K ◦ 6= Id. Thus these orbits are of type II. For the orbits OX , as given in (5), we have
dimR z(k) = 1 and K/K ◦ 6= Id. Hence these are of type II. Rest of the 52 orbits, which are
not given by the parametrizations of (1), (2), (3), (4), (5), are of type III as z(k) = 0. Now the
conclusions follow from Corollary 3.1.
3.7.2. Nilpotent orbits in E8(−24) . There are 36 nonzero nilpotent orbits in E8(−24) ; see [D1, Table
XV, p. 520]. Note that in this case we have m = [m, m].
Theorem 3.13. Let the parametrization of the nilpotent orbits be as in §3.1.1. Let X be a nonzero
nilpotent element in E8(−24) .
(1) Assume the parametrization of the orbit OX is given by any of the sequences :
0000001 1, 1000000 2, 0000001 3, 1000001 1, 1100000 1, 1000010 2, 0000012 2, 1000011 1,
1000011 3, 1000003 1, 0110001 2, 1010011 1, 1000031 3. Then dimR H 2 (OX , R) = 1.
(2) If the parametrization of the orbit OX is given by either 2000000 0 or 2000020 0, then
dimR H 2 (OX , R) ≤ 1.
(3) If OX is not given by the parametrizations as in (1), (2) above (# of such orbits are 21),
then we have dimR H 2 (OX , R) = 0.
Proof. For the Lie algebra E8(−24) , we can easily compute dimR z(k) from the last column of
[D1, Table XV, p. 520] and K/K ◦ from [K, Column 4, Table 12, pp. 277-278].
10
P. CHATTERJEE AND C. MAITY
For the orbits OX , as given in (1), we have dimR z(k) = 1 and K/K ◦ = Id, hence these are of
type I. For the orbits OX , as given in (2), we have dimR z(k) = 1 and K/K ◦ 6= Id. Hence these
orbits are of type II. Rest of the 21 orbits, which are not given by the parametrizations of (1), (2),
are of type III as z(k) = 0. Now the conclusions follow from Corollary 3.1.
Remark 3.14. Here we make some observations about the first cohomology groups of the nilpotent
orbits in non-compact non-complex real exceptional Lie algebras. To do this we begin by giving a
convenient description of the first cohomology groups of the nilpotent orbits. Following the set-up
of Theorem 2.1 it can be shown that
(
1 if k + [m, m] $ m
1
dimR H (OX , R) =
(3.1)
0 if k + [m, m] = m.
The proof of the above result will appear elsewhere. As a consequences of (3.1), for all the nilpotent
orbit OX in a simple Lie algebra g we have dimR H 1 (OX , R) ≤ 1. Recall that if g is a non-compact
non-complex real exceptional Lie algebra such that g 6≃ E6(−14) and g 6≃ E7(−25) then any maximal
compact subgroup of Int g is semisimple, and hence, using (3.1), it follows that dimR H 1 (OX , R) = 0
for all nilpotent orbit OX in g. We next assume g = E6(−14) or g = E7(−25) . Note that in both
the cases [m, m] $ m. We follow the parametrizations of the nilpotent orbits of g as given in [D1,
Tables X, XIII]; see §3.1.1 also. When g = E6(−14) we are able to conclude that dimR H 1 (OX , R) = 1
only for one orbit, namely, the orbit OX parametrized by 40000 − 2. In this case, from the last
column and row 9 of [D1, Table X, p. 512] one has k = [k, k]. Thus k + [m, m] $ m, and (3.1)
applies. For g = E7(−25) we obtain that dimR H 1 (OX , R) = 1 when OX is parametrized by any of
the following sequences : 000000 2, 000000 − 2, 000002 − 2, 200000 − 2, 200002 − 2, 400000 − 2,
000004 − 6, 200002 − 6, 400004 − 6, 400004 − 10. For the above orbits, from the last column of [D1,
Table XIII, p. 516] we have k = [k, k], and hence, using (3.1), analogous arguments apply.
References
[BC]
[CM]
[D1]
[D2]
[K]
[M]
I. Biswas and P. Chatterjee, On the exactness of Kostant-Kirillov form and the second cohomology of
nilpotent orbits, Internat. J. Math. 23 (2012), no. 8, 1250086, 25 pp.
D. H. Collingwood and W. M. McGovern, Nilpotent orbits in semisimple Lie algebras, Van Nostrand Reinhold Mathematics Series, Van Nostrand Reinhold Co., New York, 1993.
D.Z. Djokovic, Classification of Nilpotent Elements in Simple Exceptional Real Lie Algebras of Inner Type
and Description of Their Centralizers , J. Alg. 112 (1988), 503-524.
D.Z. Djokovic, Classification of Nilpotent Elements in Simple Exceptional Real Lie Algebras E6(6) and
E6(−26) and Description of Their Centralizers , J. Alg. 116 (1988), 196-207.
Donald King, The Component Groups of Nilpotents in Exceptional Simple Real Lie Algebras, Communications in Algebra, 20(1), 219-284 (1992).
W. M. McGovern, The adjoint representation and the adjoint action, in : Algebraic quotients. Torus actions
and cohomology. The adjoint representation and the adjoint action, 159–238, Encyclopaedia Math. Sci.,
131, Springer, Berlin, 2002.
The Institute of Mathematical Sciences, HBNI, C.I.T. Campus, Tharamani, Chennai 600113, India
E-mail address: [email protected]
The Institute of Mathematical Sciences, HBNI, C.I.T. Campus, Tharamani, Chennai 600113, India
E-mail address: [email protected]
| 4math.GR
|
arXiv:1709.05393v1 [math.AC] 15 Sep 2017
A Sheaf On The Second Spectrum Of A Module∗
Seçil Çeken and Mustafa Alkan
[email protected], [email protected]
Abstract
Let R be a commutative ring with identity and Specs (M ) denote the
set all second submodules of an R-module M . In this paper, we construct and study a sheaf of modules, denoted by O(N, M ), on Specs (M )
equipped with the dual Zariski topology of M , where N is an R-module.
We give a characterization of the sections of the sheaf O(N, M ) in terms
of the ideal transform module. We present some interrelations between
algebraic properties of N and the sections of O(N, M ). We obtain some
morphisms of sheaves induced by ring and module homomorphisms.
2010 Mathematics Subject Classification: 13C13, 13C99, 14A15,
14A05.
Keywords and phrases: Second submodule, dual Zariski topology,
sheaf of modules.
1
Introduction
Throughout this article all rings will be commutative rings with identity elements and all modules will be unital left modules. Unless otherwise stated R
will denote a ring. Given an R-module M , the annihilator of M (in R) is denoted by annR (M ) and for an ideal I of R, the annihilator of I in M is defined
as the set (0 :M I) := {m ∈ M : Im = 0}. Clearly, (0 :M I) is a submodule of
M.
Recall that a sheaf of rings (modules) F on a topological space X is an
assignment of a ring (module) F (U ) to each open subset U of X, together
with, for each inclusion of open subsets V ⊆ U , a morphism of rings (modules)
ρUV : F (U ) −→ F (V ) subject to the following conditions:
(i) F (∅) = 0.
(ii) ρUU = idF (U) .
(iii) If W ⊆ V ⊆ U , then ρUW = ρV W oρUV .
(iv) If U is an open subset of X and {Uα }α∈Λ is an open cover of U , and if
f ∈ F (U ) is an element such that ρUUα (f ) = 0 for all α ∈ Λ, then f = 0.
∗ This paper is submitted to Communications in Algebra on June 15th, 2017 for the referee
process.
1
(v) If U is an open subset of X and {Uα }α∈Λ is an open cover of U , and
if we have a collection elements fα ∈ F (Uα ) with the property that for all α,
β ∈ Λ, ρUα (Uα ∩Uβ ) (fα ) = ρUβ (Uα ∩Uβ ) (fβ ), then there is an element f ∈ F (U )
such that for all α, fα = ρUUα (f ).
If F is a sheaf on a topological space X, we refer to F (U ) as the sections of
F over the open subset U . We call the maps ρUV as the restriction maps (cf.
[19]).
The prime spectrum of a ring R, denoted by Spec(R), consists of all prime
ideals of R and it is non-empty. For each ideal I of R, the sets V (I) = {p ∈
Spec(R) : I ⊆ p}, where I is an ideal of R, satisfy the axioms for closed sets of
a topology on Spec(R), called the Zariski topology of R.
It is well-known that for any commutative ring R, there is a sheaf of rings
on Spec(R), denoted by OSpec(R) , defined as follows: For an open subset
F U ⊆
Spec(R), we define OSpec(R) (U ) to be the set of all functions s : U → p∈U Rp ,
such that s(p) ∈ Rp , for each p ∈ U , and such that for each p ∈ U , there is a
neighborhood V of p, contained in U , and elements a, f ∈ R, such that for each
q ∈ V , we have f 6∈ q and s(q) = fa in Rq (see [19]).
Let M be an R-module. A proper submodule N of M is said to be prime if
for any r ∈ R and m ∈ M with rm ∈ N , we have m ∈ N or r ∈ annR (M/N ) :=
(N : M ). If N is a prime submodule of M , then p = (N : M ) is a prime
ideal of R. In this case, N is called a p-prime submodule of M . The set of
all prime submodules of a module M is called the prime spectrum of M and
denoted by Spec(M ). For any submodule N of an R-module M , we have a set
V (N ) = {P ∈ Spec(M ) : (N : M ) ⊆ (P : M )}. Then the sets V (N ), where N is
a submodule of M satisfy the axioms for closed sets of a topology on Spec(M ),
called the Zariski topology of M . Several authors have investigated the prime
spectrum and the Zariski topology of a module over the last twenty years (see
for example [21], [22], [23], [24], [25]). Recently, some authors have investigated
a sheaf structure on the prime spectrum of a module which generalizes the sheaf
of rings OSpec(R) on the topological space Spec(R). In [26], the author obtained
an R-module OSpec(M) (U ) for each open subset U of Spec(M ) equipped with
the Zariski topology of M such that OSpec(M) is a sheaf of modules on Spec(M ).
In [20], the authors defined and studied a sheaf of modules which is denoted by
A(N, M ) on the topological space Spec(M ) equipped with the Zariski topology
of M , where M and N are two R-modules. In fact, both OSpec(M) and A(N, M )
are generalizations of the sheaf of rings OSpec(R) to modules. In [20], the authors
proved that if N = R, then A(R, M ) is a scheme on Spec(M ). This scheme
structure were investigated in [1].
Recently, a dual theory of prime submodules has been developed and extensively studied by many authors. The dual notion of prime submodules was first
introduced by S. Yassemi in [27]. A submodule N of an R-module M is said to
be a second submodule provided N 6= 0 and, for all r ∈ R, rN = 0 or rN = N .
If N is a second submodule of M , then p = annR (N ) is a prime ideal of R. In
this case, N is called a p-second submodule of M (cf. [27]). In recent years,
second submodules have attracted attention of various authors and they have
been studied in a number of papers (see for example [3], [4], [5], [7], [9], [12],
2
[13], [14], [15]).
The set of all second submodules of a module M is called the second spectrum
of M and denoted by Specs (M ). As in [11], for any submodule N of an Rmodule M we define V s∗ (N ) to be the set of all second submodules of M
contained in N . Clearly V s∗ (0) is the empty set and V s∗ (M ) is Specs (M ). Note
that for any family of submodules Ni (i ∈ I) of M , ∩i∈I V s∗ (Ni ) = V s∗ (∩i∈I Ni ).
Thus if Z s∗ (M ) denotes the collection of all subsets V s∗ (N ) of Specs (M ) where
N ≤ M , then Z s∗ (M ) contains the empty set and Specs (M ), and Z s∗ (M )
is closed under arbitrary intersections. But in general Z s∗ (M ) is not closed
under finite unions. A module M is called a cotop module if Z s∗ (M ) is closed
under finite unions. In this case Z s∗ (M ) is called the quasi-Zariski topology on
Specs (M ) (see [6]). Note that in [3] a cotop module was called a tops -module.
More information about the class of cotop modules can be found in [3] and [6].
Let M be an R-module and N be a submodule of M . We define the set
V s (N ) := {S ∈ Specs (M ) : annR (N ) ⊆ annR (S)}. In [11, Lemma 2], it
was shown that V s (N ) = V s ((0 :M annR (N ))) = V s∗ ((0 :M annR (N ))), in
particular, V s ((0 :M I)) = V s∗ ((0 :M I)) for every ideal I of R, and that the set
Z s (M ) = {V s (N ) : N ≤ M } satisfies the axioms for the closed sets. Thus there
exists a topology, say τ s , on Specs (M ) having Z s (M ) as the family of closed
subsets. This topology is called the dual Zariski topology of M (see [11, Lemma
2]). Dual Zariski topology, the second spectrum of modules, and related notions
have been investigated by some authors in recent years (see [2], [3], [6], [8], [11],
[16], [17] and [18]).
In this paper, we define and study a sheaf structure on the second spectrum
of a module. Let M be an R-module. In Section 2, we construct a sheaf,
denoted by O(N, M ), on Specs (M ) equipped with the dual Zariski topology of
M , where N is an R-module. Firstly, we find the stalk of the sheaf O(N, M )
(see Theorem 2.3). In Theorem 2.12, we give a characterization for the sections
of the sheaf O(N, M ) in terms of the ideal transform module. Let R be a
Noetherian ring and M be a faithful secondful R-module. We prove that if N is
a free, projective or flat R-module, then so is O(N, M )(Specs (M )) (see Theorem
2.18). In Section 3, we deal with a scheme structure on the second spectrum of
a module. In Theorem 3.2, we prove that O(R, M ) is a scheme when M is a
faithful secondful R-module and Specs (M ) is a T0 -space. Then we define two
morphisms of locally ringed spaces by using ring and module homomorphisms
(see Theorem 3.3 and Corollary 3.5).
2
A Sheaf Structure On The Second Spectrum
Of A Module
Throughout the rest of the paper M will be an R-module, X s will denote
Specs (M ) and we consider X s with the dual Zariski topology unless otherwise
stated. For every open subset U of X s , we set Supps (U ) = {annR (S) : S ∈ U }.
In this section, we construct a sheaf on X s and investigate some properties
3
of this sheaf.
s
Definition 2.1 Let N be an R-module. For every open subset U of
Q X we
define O(N, M )(U ) to be the set of all elements (β p )p∈Supps (U) ∈
Np
p∈Supps (U)
in which for each Q ∈ U , there is an open neighborhood W of Q with Q ∈ W ⊆ U
and there exist elements t ∈ R, m ∈ N such that for every S ∈ W , we have
t 6∈ p := annR (S) and β p = m
t ∈ Np .
Let U and V be open subsets of X s with ∅ 6= V ⊆ U and β = (β p )p∈Supps (U) ∈
O(N, M )(U ). Then it is clear that the restriction (β p )p∈Supps (V ) belongs to
O(N, M )(V ). Therefore, we have the restriction map ρUV : O(N, M )(U ) −→
O(N, M )(V ), ρUV (β) = (β p )p∈Supps (V ) for all β = (β p )p∈Supps (U) ∈ O(N, M )(U ).
We define ρU∅ : O(N, M )(U ) −→ O(N, M )(∅) = 0 to be the zero map. It is
clear from the local nature of the definition, O(N, M ) is a sheaf with the restriction maps defined above. We can define a map τ U
N : N −→ O(N, M )(U )
n
s
by τ U
)
for
all
n
∈
N
.
We
note
that
τU
(n)
=
(
N
N is an R-module
1 p∈Supp (U)
U
V
homomorphism. Clearly, ρUV o τ N = τ N .
Recall that for any r ∈ R, the set Dr := Spec(R)\V (Rr) is open in Spec(R)
and the family {Dr : r ∈ R} forms a base for the Zariski topology on Spec(R).
Let M be an R-module. For each r ∈ R, we define Yr := X s \V s ((0 :M r)). In
[11, Theorem 4], it was shown that the set B = {Yr : r ∈ R} forms a base for
the dual Zariski topology on X s .
Remark 2.2 Let r ∈ R and Q ∈ X s . Then, r 6∈ annR (Q) if and only if Q ∈ Yr .
Let Q ∈ Yr . Suppose that r ∈ annR (Q), then rQ = 0 and so Q ⊆ (0 :M r). This
implies that Q ∈ V s∗ ((0 :M r)) = V s ((0 :M r)), a contradiction. Conversely, let
r 6∈ annR (Q). Then Q 6⊆ (0 :M r) and so Q ∈ Yr . In the proof our results, we
will use this fact without any further comment.
Let F be a sheaf of modules (rings) on a topological space X and P ∈ X.
Recall that the stalk FP of F at P is defined to be the direct limit −lim
F (U )
−−→
P ∈U
of the modules (rings) F (U ) for all open subsets U of X containing P via the
restriction maps (see [19]). In the following theorem, we determine the stalk of
the sheaf O(N, M ) at a point S in X s .
Theorem 2.3 Let N be an R-module and S ∈ X s . Then, the stalk O(N, M )S
of the sheaf O(N, M ) at S is isomorphic to Np where p := annR (S).
Proof. Let S be a p-second submodule of M and m ∈ O(N, M )S = −lim
O(N, M )(U ).
−−→
P ∈U
Then there exists an open neighborhood U of S and β = (β p )p∈Supps (U) ∈
O(N, M )(U ) such that β represents m. We define φ : O(N, M )S −→ Np by
φ(m) = β p . Let V be another neighborhood of S and α = (αp )p∈Supps (V ) ∈
O(N, M )(V ) such that α also represents m. Then there exists an open set
W ⊆ U ∩ V such that S ∈ W and β |W = α|W . Since S ∈ W , we have αp = β p .
This shows that φ is a well-defined map. We claim that φ is an isomorphism.
4
Let x ∈ Np . Then x = at for some a ∈ N , t ∈ R\p. Since t 6∈ p = annR (S),
we have S ∈ Yt . Now we define β q = at in Nq for all Q ∈ Yt where t 6∈ q =
annR (Q). Then β = (β q )q∈Supps (Yt ) ∈ O(N, M )(Yt ). If m is the equivalence
class of β in O(N, M )S , then φ(m) = x. Hence φ is surjective.
Now let m ∈ O(N, M )S and φ(m) = 0. Let U be an open neeighborhood
of S and β = (β p )p∈Supps (U) ∈ O(N, M )(U ) is a representative of m. There
is an open neighborhood V ⊆ U of S and there are elements a ∈ N , t ∈ R
such that for all Q ∈ V , we have t 6∈ q := annR (Q) and β q = at ∈ Nq . Then
0 = φ(m) = β p = at in Np . So there is h ∈ R\p such that ha = 0. For all
a
Q ∈ Yth , we have β q = ha
ht = t = 0 in Nq where q = annR (Q). Thus β |Yth = 0.
Therefore, β = 0 in O(N, M )(Yth ). Consequently, m = 0. This shows that φ is
injective. Thus φ is an isomorphism.
A ringed space is a pair (X, OX ) consisting of a topological space X and a
sheaf of rings OX on X. The ringed space (X, OX ) is called a locally ringed
space if for each point P ∈ X, the stalk OX,P is a local ring (cf. [19]).
Corollary 2.4 (X s , O(R, M )) is a locally ringed space.
Example 2.5 Consider the Z-modules M = Q ⊕ Zp and N = Z, where p is a
prime number. Then Specs (M ) = {Q ⊕ 0, 0 ⊕ Zp }. By Theorem 2.3,
O(Z, Q ⊕ Zp )Q⊕0
≃
Z(0) = Q and
O(Z, Q ⊕ Zp )0⊕Zp
≃
ZpZ = Z(p) = {
a
∈ Q : a, b ∈ Z, b 6= 0, p ∤ b}.
b
Let M be an R-module. The map ψ s : Specs (M ) −→ Spec(R/annR (M ))
defined by ψ s (S) = annR (S)/annR (M ) is called the natural map of Specs (M ).
M is said to be secondful if the natural map ψ s is surjective (cf. [18]).
Let M be an R-module. The Zariski socle of a submodule N of M , denoted
by Z.soc(N ), is defined to be the sum of all members of V s (N ) and if V s (N ) = ∅,
then Z.soc(N ) is defined to be 0 (cf. [18]).
Lemma 2.6 Let R be a Noetherian ring and N be an R-module. Let M be
a secondful R-module and U = X s \V s (K) where K ≤ M . Then for each
β = (β p )p∈Supps (U) ∈ O(N, M )(U ), there exist r ∈ Z+ , s1 , ..., sr ∈ annR (K)
i
and m1 , ..., mr ∈ N such that U = ∪ri=1 Ysi and β p = m
si for all S ∈ Ysi ,
i = 1, ..., r where p = annR (S).
Proof. Since R is Noetherian, U is quasi-compact by [18, Corollary 4.4-(d)].
Thus there exist n ∈ Z+ , open subsets W1 , ..., Wn of U , t1 , ..., tn ∈ R, a1 , ..., an ∈
a
N such that U = ∪nj=1 Wj and for each j = 1, ..., n and S ∈ Wj we have β p = tjj
where tj 6∈ p := annR (S). Fix j ∈ {1, ..., n}. There is a submodule Hj of
M such that Wj = X s \V s (Hj ). Also we have V s (K) ⊆ V s (Hj ). Since R is
Noetherian, annR (Hj ) = Rbj1 + ... + Rbjnj for some bj1 , ..., bjnj ∈ R. This
implies that
Wj = X s \V s (Hj ) = X s \V s∗ ((0 :M annR (Hj )))
5
n
j
(0 :M Rbjf ))
= X s \V s∗ ((0 :M Rbj1 + ... + Rbjnj )) = X s \V s∗ (∩f =1
n
n
j
j
s
s∗
s∗
s
= X \ ∩f =1 V ((0 :M Rbjf )) = ∪f =1 (X \V ((0 :M Rbjf )))
nj
nj
nj
Ybjf .
Ybjf , i.e. Wj = ∪f =1
(X s \V s ((0 :M Rbjf ))) = ∪f =1
= ∪f =1
On the other hand, we have Z.soc(K) ⊆ Z.soc(Hj ) as V s (K) ⊆ V s (Hj ).
This implies p
that annR (Z.soc(H
p j )) ⊆ annR (Z.soc(K)). By [18, Theorem 3.5(e)], we get annR (Hj ) ⊆ annR (K). Since R is Noetherian, there exists
p
d ∈ Z+ such that ( annR (K))d ⊆ annR (K) and we have
q
p
annR (Hj )d ⊆ ( annR (Hj ))d ⊆ ( annR (K))d ⊆ annR (K)
It follows that bdjf ∈ annR (K) for each f ∈ {1, ..., nj }. Also, for each f ∈
{1, ..., nj } and for each S ∈ Ybjf , we have tj bdjf 6∈ annR (S). We conclude that
Ybjf = X s \V s ((0 :M bjf )) = X s \V s ((0 :M tj bdjf )) = Ytj bdjf
and we can write β annR (S) =
aj
tj
=
aj bd
jf
tj bd
jf
∈ NannR (S) . This completes the proof.
Let K be an R-module. For an ideal I of R, the I-torsion submodule of
K is defined to be ΓI (K) := ∪n≥1 (0 :K I n ) and K is said to be I-torsion if
K = ΓI (K) (cf. [10]).
Lemma 2.7 Let N be an R-module and U = X s \V s (K) where K ≤ M . Then
ΓannR (K) (O(N, M )(U )) = 0.
Proof. Let β = (β p )p∈Supps (U) ∈ ΓannR (K) (O(N, M )(U )). There exists n ∈ Z+
such that annR (K)n β = 0. Consider p ∈ Supps (U ). There exists a second
submodule Q ∈ U such that p = annR (Q). If annR (K)n ⊆ p, then annR (K) ⊆
p and so Q ∈ V s (K), a contradiction. Thus annR (K)n 6⊆ p. So there exists
t β
tp ∈ annR (K)n \p. Since tp β = 0, we have tp β p = 0 and β p = ptp p = 0 ∈ Np .
Hence, β = 0.
Theorem 2.8 Let R be a Noetherian ring, M be a faithful secondful R-module
and N be an R-module. Let U = X s \V s (K) where K ≤ M . Then ΓannR (K) (N ) =
U
ker(τ U
N ), and so ker(τ N ) is annR (K)-torsion.
Proof. By Lemma 2.7, we have τ U
N (ΓannR (K) (N )) ⊆ ΓannR (K) (O(N, M )(U )) =
0. So ΓannR (K) (N ) ⊆ ker(τ U
N ).
m
s
Suppose that m ∈ ker(τ U
N ). Then 1 = 0 ∈ Np for all p ∈ Supp (U ).
s
So, P
for each p ∈ Supp (U ), there is tp ∈ R\p such that tp m = 0. Put J =
Rtp . Then Jm = 0.
p∈Supps (U)
Let q ∈ V (J). We claim that annR (K) ⊆ q. Suppose on the contrary that
annR (K) 6⊆ q. Since M is faithful secondful, there is a second submodule S of
M such that q = annR (S). Therefore S ∈ U and q ∈ Supps (U ). This implies
that tq ∈ J ⊆ q. This contradicts the fact that tq ∈ R\q. Thus
√
p
T
T
p= J
annR (K) ⊆ annR (K) =
p⊆
p∈V (annR (K))
6
p∈V (J)
Since R is Noetherian, annR (K)n ⊆ J for some n ∈ Z+ . Hence annR (K)n m ⊆
Jm = 0. This shows that m ∈ ΓannR (K) (N ) and the result follows.
r
P
Lemma 2.9 Let r, n ∈ Z+ , s1 , ..., sr ∈ R. Then V s ((0 :M
r
P
i=1
Rsi )).
Rsni )) = V s ((0 :M
i=1
Proof. Let S ∈ V s ((0 :M
r
P
i=1
Rsni )). Then annR ((0 :M
r
P
Rsni )) = annR (∩ri=1 (0 :M
i=1
(Rsi )n =
Rsni ⊆ annR ((0 :M
Rsni )) ⊆ annR (S). For each i ∈ {1, ..., r}, we have
n
r
n
Rsi )) ⊆ annR (∩i=1 (0 :M Rsi )) ⊆ annR (S). Since annR (S) is a prime ideal,
r
P
Rsi ⊆ annR (S). It follows that S ⊆ (0 :M
we have Rsi ⊆ annR (S) and so
annR (S)) ⊆ (0 :M
V s ((0 :M
r
P
r
P
i=1
i=1
Rsi ). This shows that S ∈ V s∗ ((0 :M
i=1
For the other containment;
(0 :M
r
P
i=1
r
P
Rsi )) and hence V s ((0 :M
r
P
i=1
Rsni ) and hence V s ((0 :M
i=1
Rsni ⊆
r
P
i=1
Rsni )) ⊆ V s ((0 :M
r
P
r
P
r
P
Rsi )).
i=1
Rsi implies that (0 :M
i=1
Rsi )) ⊆ V s ((0 :M
r
P
i=1
Rsi )) =
i=1
r
P
i=1
Rsni )).
Rsi ) ⊆
Theorem 2.10 Let R be a Noetherian ring, M be a faithful secondful R-module,
N be an R-module and U = X s \V s (K) where K ≤ M . Let W be an open subset
of X s such that U ⊆ W . Then ker(ρW U ) = ΓannR (K) (O(N, M )(W )).
Proof. By Lemma 2.7, ρW U (ΓannR (K) (O(N, M )(W )) ⊆ ΓannR (K) (O(N, M )(U )) =
0. So ΓannR (K) (O(N, M )(W )) ⊆ ker(ρW U ).
There exists L ≤ M such that W = X s \V s (L). Let β = (β p )p∈Supps (W ) ∈
ker(ρW U ). By Lemma 2.6, there exist r ∈ Z+ , s1 , ..., sr ∈ annR (L) and
m1 , ..., mr ∈ N such that W = ∪ri=1 Ysi and for each i = 1, ..., r and each
i
S ∈ Ysi , we have β annR (S) = m
si . Since β ∈ ker(ρW U ), we have β p = 0 for all
p ∈ Supps (U ). Fix i ∈ {1, ..., r}. Set U ′ = U ∩ Ysi . Hence,
U ′ = (X s \V s (K)) ∩ (X s \V s ((0 :M si )) = X s \(V s (K) ∪ V s ((0 :M si ))) =
X s \V s (K + (0 :M si )).
s
′
U′
i
Then m
si = 0 ∈ Np for all p ∈ Supp (U ). This implies that τ N (mi ) = 0.
By Theorem 2.8, there exists hi ∈ Z+ such that annR (K + (0 :M si ))hi mi = 0.
Let h := max{h1 , ..., hr }. Now, let p ∈ Supps (W ). There exists i ∈ {1, ..., r}
such that p ∈ Supps (Ysi ). Let d ∈ annR (K)h . Since shi ∈ annR ((0 :M si ))h , we
have
dshi ∈ annR (K)h annR ((0 :M si ))h ⊆ (annR (K) ∩ annR ((0 :M si )))h =
annR (K + (0 :M si ))h .
Therefore, we conclude that dβ p =
h
h
dmi
si
=
dsh
i mi
sh+1
i
= 0 ∈ Np for all d ∈
annR (K) . This implies that annR (K) β = 0 and so β ∈ ΓannR (K) (O(N, M )(W )).
7
Theorem 2.11 Let R be a Noetherian ring, M be a faithful secondful R-module,
N be an R-module and U = X s \V s (K) where K ≤ M . Then the map τ U
N :
N −→ O(N, M )(U ) has an annR (K)-torsion cokernel.
Proof. Let β = (β p )p∈Supps (U) ∈ O(N, M )(U ). By Lemma 2.6, there exist
r
S
Ysi and for
r ∈ Z+ , s1 , ..., sr ∈ annR (K) and m1 , ..., mr ∈ N such that U =
i=1
i
each i = 1, ..., r and each S ∈ Ysi we have β annR (S) = m
si . Fix i ∈ {1, ..., r}.
si mi
Then for each S ∈ Ysi , si β annR (S) = si ∈ NannR (S) . This means that
Ys
ρUYsi (si β) = si ρUYsi (β) = τ N i (mi ) = ρUYsi (τ U
N (mi ))
Thus, si β − τ U
N (mi ) ∈ ker(ρUYsi ) = ΓannR ((0:M si )) (O(N, M )(U )) by Theorem
2.10. Hence there exists ni ∈ Z+ such that annR ((0 :M si ))ni (si β − τ U
N (mi )) =
(m
))
=
0.
Define
n
:=
max{n
,
...n
}
+
1.
Then
for all
0. Then sni i (si β − τ U
i
1
r
N
n−ni −1 ni U
i −1 ni +1
τ
(m
).
It
follows
that
i = 1, ..., r, we have sni β = sn−n
s
β
=
s
s
i
i
N
i
i
i
n−1
U
U
sni β = sin−1 τ U
(m
)
=
τ
(s
m
)
∈
τ
(N
).
By
Lemma
2.9,
we
have
i
i
N
N i
N
r
r
P
P
s
n
s
V ((0 :M
Rsi )) = V ((0 :M
Rsi ))
i=1
=V
s∗
=
=
=
s∗
i=1
r
P
((0 :M
Rsi ))
i=1
r
V (∩i=1 (0 :M Rsi ))
∩ri=1 V s∗ ((0 :M Rsi ))
∩ri=1 V s ((0 :M Rsi )).
It follows that r
P
X s \V s ((0 :M
Rsni )) = X s \ ∩ri=1 V s ((0 :M Rsi ))
i=1
= ∪ri=1 (X s \V s ((0 :M Rsi )))
= ∪ri=1 Ysi = U = X s \V s (K).
r
P
Rsni )) = V s (K). Since M is faithful secondful,
This means that V s ((0 :M
r
P
i=1
p
annR (K) ⊆
⊆ V (annR (K)). Hence annR (K) ⊆
we get that V (
i=1
s
r
P
Rsni . Since R is Noetherian, there exists h ∈ Z+ such that annR (K)h ⊆
i=1
r
r
P
P
h
n
n
Rsi . It follows that annR (K) β ⊆
Rsi β ⊆ τ U
N (N ). This completes
Rsni )
i=1
i=1
the proof.
Let K be an R-module and I be an ideal of R. Recall that the ideal transform
of K with respect to I is defined as DI (K) := −
lim
HomR (I n , K) (cf. [10]).
−→
n∈N
Theorem 2.12 Let R be a Noetherian ring, M be a faithful secondful R-module,
N be an R-module and U = X s \V s (K) where K ≤ M . Then, there is a unique
R-isomorphism
gK,N : O(N, M )(U ) −→ DannR (K) (N ) := −
lim
HomR (annR (K)n , N )
−→
n∈N
8
such that the diagram
N
τU
N
−→
ց
O(N, M )(U )
↓ gK,N
DannR (K) (N )
commutes.
Proof. By Theorems 2.8 and 2.11, both the kernel and cokernel of τ U
N are
annR (K)-torsion. Therefore, there is a unique R-homomorphism gK,N : O(N, M )(U ) −→
DannR (K) (N ) such that the given diagram commutes by [10, Corollary 2.2.13(ii)]. By Lemma 2.7, ΓannR (K) (O(N, M )(U )) = 0. So, gK,N is an isomorphism
by [10, Corollary 2.2.13-(iii)].
Corollary 2.13 Let R be a Noetherian ring, M be a faithful secondful Rmodule, N be an R-module and U = X s \V s (K) where K ≤ M . Then the
following hold.
(1) O(ΓannR (K) (N ), M )(U ) = 0
(2) O(N, M )(U ) ≃ O(N/ΓannR (K) (N ), M )(U )
(3) O(N, M )(U ) ≃ O(O(N, M )(U ), M )(U )
(4) If N is an annR (K)-torsion R-module, then O(N, M )(U ) = 0.
Proof. Parts (1), (2) and (3) follow from Theorem 2.12 and [10, Corollary
2.2.8]. Part (4) is an immediate consequence of part (1).
Example 2.14 Consider the Z-modules N = Z/9Z and M = (⊕p Zp )⊕Q where
p runs over all distinct prime numbers. Then M is a faithful secondful Z-module.
Let K = (0⊕Z3 ⊕0...)⊕0 and U = Specs (M )\V s (K). Then annR (K) = 3Z and
N is a 3Z-torsion Z-module. By Corollary 2.13-(4), O(Z/9Z, (⊕p Zp )⊕Q)(U ) =
0.
Corollary 2.15 Let R be a principal ideal domain, M be a faithful secondful
R-module, N be an R-module and U = X s \V s (K). Then there exists a ∈ R
such that O(N, M )(U ) ≃ Na , where Na is the localization of N with respect to
the multiplicative set {an : n ∈ N}.
Proof. Since R is a a principal ideal domain, there is an element a ∈ R such
that annR (K) = Ra. By Theorem 2.12 and [10, Theorem 2.2.16], we have
O(N, M )(U ) ≃ DannR (K) (N ) ≃ Na .
Theorem 2.16 Let M be a faithful secondful R-module and N be any Rmodule. For any element f ∈ R, the module O(N, M )(Yf ) is isomorphic to
the localized module Nf . In particular, O(N, M )(X s ) ≃ N .
Proof. We define the map φ : Nf −→ O(N, M )(Yf ) by φ( fam ) = ( fam )p∈Supps (Yf ) .
We claim that φ is an isomorphism. First, we show that φ is injective. Let
φ( fan ) = φ( fbm ). Then for every S ∈ Yf , fan = fbm in Np where p = annR (S).
Thus there exists h ∈ R\p such that h(f m a − f n b) = 0 in N . Let I = (0 :R
f m a − f n b). Then h ∈ I and h 6∈ p, so I 6⊆ p. This holds for any S ∈ Yf ,
9
so we deduce that Supps (Yf ) ⊆ Spec(R)\V (I). Since M is faithful secondful,
Df = Supps (Yf ) ⊆ Spec(R)\V
(I) and we get that V (I) ⊆ V (Rf ). This implies
√
√
that Rf ⊆ Rf ⊆ I. Therefore f l ∈ I for some l ∈ Z+ . Now we have
f l (f m a − f n b) = 0 which shows that fan = fbm in Nf . Thus φ is injective.
Let β = (β p )p∈Supps (Yf ) ∈ O(N, M )(Yf ). Then we can cover Yf with the
open subsets Vi on which β annR (S) is represented by agii with gi 6∈ annR (S) for
all S ∈ Vi , in other words Vi ⊆ Ygi . Since the open sets of the form Yr (r ∈ R)
form a base for the dual Zariski topology on X s , we may assume that Vi = Yhi
for some hi ∈ R. Since Yhi ⊆ Ygi , Dhi = ψ s (Yhi ) ⊆ ψ s (Ygi ) = Dg√
i by [6,
Proposition
4.1].
This
implies
that
V
(Rg
)
⊆
V
(Rh
)
and
so
Rh
⊆
Rhi ⊆
i
i
i
√
Rgi . Thus hsi ∈ Rgi for some s ∈ Z+ . So hsi = cgi for some c ∈ R and
bi
cai
cai
ai
s
gi = cgi = hsi . We see that β annR (S) is represented by ki (bi = cai , ki = hi )
on Yki and (since Yhi = Yhsi ) the Yki cover Yf . The open cover Yf = ∪Yki has
a finite subcover by [11, Theorem 4-(2)]. Suppose that Yf ⊆ Yk1 ∪ ... ∪ Ykn .For
b
1 ≤ i, j ≤ n, kbii and kjj both represent β annR (S) on Yki ∩ Ykj . By [6, Corollary
b
4.2], Yki ∩ Ykj = Yki kj , and by the injectivity of φ, we get that kbii = kjj in
Nki kj . Hence (ki kj )nij (kj bi − ki bj ) = 0 for some nij ∈ Z+ . Let m = max{nij :
1 ≤ i, j ≤ n}. Then kjm+1 (kim bi ) − kim+1 (kjm bj ) = 0. By replacing each ki by
kim+1 and bi by kim bi , we still see that β annR (S) is represented on Yki by kbii and
furthermore we have kj bi = ki bj for all i, j. Since Yf ⊆ Yk1 ∪ ... ∪ Ykn , by [6,
Proposition 4.1], we have
Df = ψ s (Yf ) ⊆ ∪ni=1 ψ s (Yki ) = ∪ni=1 Dki , i.e.,PSpec(R)\V (Rf ) ⊆ Spec(R)\∩ni=1
n
n
V (Rki ). This implies
pPn that ∩i=1 V (Rki ) = V ( i=1 Rki ) ⊆ V (Rf ) and+hence
√
are c1 , ..., cn ∈ R and t ∈ Z such
Rf ⊆ RfP⊆
i=1 Rki . So,
Pthere
n
n
that f t =
ci ki . Let a =
ci bi . Then for each j, we have kj a =
i=1
i=1
Pn
Pn
bj
a
t
i=1 ci ki bj = bj f . This implies that f t = kj on Ykj . Therei=1 ci kj bi =
fore, φ( fat ) = (β p )p∈Supps (Yf ) .
Proposition 2.17 Let K, L be R-modules and ϕ : K −→ L be an R-homomorphism.
Then ϕ induces a morphism of sheaves ϕ : O(K, M ) −→ O(L, M ). If ϕ is an
isomorphism of R-modules, then ϕ is an isomorphism of sheaves.
a
Proof. Let U be an open subset of X s and β = ( fpp )p∈Supps (U) ∈ O(K, M )(U ).
ϕ(ap )
∈ O(L, M )(U ). For each Q ∈ U , there is
We show that
fp
s
p∈Supp (U)
an open neighborhood W of Q with Q ∈ W ⊆ U and there exist elements
t ∈ R, m ∈ K such that for every S ∈ W , we have t 6∈ p := annR (S) and
ap
m
fp = t ∈ Kp . So there exists sp ∈ R\p such that sp (tap − fp m) = 0. It follows
that sp (tϕ(ap ) − fp ϕ(m)) = 0. This means that
ϕ(ap )
fp
and t 6∈ p := annR (S) for every S ∈ W . This shows
ϕ(m)
where ϕ(m) ∈
t
ϕ(ap )
that
fp
p∈Supps (U)
=
O(L, M )(U ). Thus the map ϕ(U ) : O(K, M )(U ) −→ O(L, M )(U ) defined by
ϕ(ap )
ap
.
ϕ(U ) ( )p∈Supps (U) =
fp
fp
p∈Supps (U)
10
L
∈
is well-defined. Clearly, ϕ(U ) is an R-homomorphism. Since
ap
ϕ (ap )
ρ′UV ϕ(U ) ( )p∈Supps (U)
= ρ′UV (
)p∈SuppU
fp
fp
ϕ (np )
)p∈SuppV
= (
fp
ap
= ϕ(V )ρUV ( )p∈Supps (U) ,
fp
the following diagram is commutative.
O(K, M )(U )
↓ ρUV
O(K, M )(V )
ϕ(U )
−−−→
O(L, M )(U )
↓ ρ′UV
ϕ(V ) O(L, M )(V )
−−−→
This shows that ϕ : O(K, M ) −→ O(L, M ) is a morphism of sheaves.
Now suppose
that ϕ is
We show that ϕ(U ) is injective.
anisomorphism.
ϕ(a )
ϕ(ap )
ap
s
= 0. Then fpp = 0 for every
)
=
Let ϕ(U ) ( fp p∈Supp (U)
fp
p∈Supps (U)
p ∈ Supps (U ). There exists tp ∈ R\p such that tp ϕ(ap ) = ϕ(tp ap ) = 0. Since
a
t a
ϕ is injective, tp ap = 0 for every p ∈ Supps (U ). It follows that fpp = tpp fpp = 0
a
for every p ∈ Supps (U ). This shows that ( fpp )p∈Supps (U) = 0 and so ϕ(U ) is
injective for every open subset U of X s .
b
∈ O(L, M )(U ).
Now we show that ϕ(U ) is surjective. Let tpp
s
p∈Supp (U)
s
There
exists ap ∈ K such that ϕ(ap ) = bp for each p ∈ Supp (U ). We show that
ap
∈ O(K, M )(U ). For each Q ∈ U , there is an open neighborhood
tp
p∈Supps (U)
W of Q with Q ∈ W ⊆ U and there exist elements t ∈ R, b ∈ L such that for
b
ϕ(a )
every S ∈ W , we have t ∈
6 annR (S) = p and tpp = tpp = bt . There exists
ϕ(a )
= tpp , where t 6∈ p = annR (S) for every
a ∈ K such that b = ϕ(a). So ϕ(a)
t
S ∈ W . There exists vp ∈ R\p such that vp (tp ϕ(a)−tϕ(ap )) = 0. It follows that
ϕ(vp tp a − vp tap ) = 0. Since ϕ is injective vp (tp a − tap ) = 0 for vp ∈ R\p. This
a
means that tpp = at where t 6∈ p = annR (S) for every S ∈ W . This shows that
bp
ap
ap
=
∈
O(K,
M
)(U
)
and
ϕ
.
tp
tp
tp
s
s
s
p∈Supp (U)
p∈Supp (U)
p∈Supp (U)
Thus ϕ(U ) is surjective for every open subset U of X s . Consequently, ϕ is an
isomorphism of sheaves.
Theorem 2.18 Let R be a Noetherian ring, M be a faithful secondful R-module
and N be an R-module. Then the following hold.
(1) If N is a free R-module, then O(N, M )(X s ) is a free R-module.
(2) If N is a projective R-module, then O(N, M )(X s ) is a projective Rmodule.
(3) If N is a flat R-module, then O(N, M )(X s ) is a flat R-module.
11
Proof. (1) We can write X s = X s \V s (0). Since N is a free R-module, N
is isomorphic to a direct sum of some copies of R, say N ≃ ⊕i∈Λ R for an
index set Λ. By Proposition 2.17, O(N, M )(X s ) ≃ O(⊕i∈Λ R, M )(X s ). By
Theorem 2.12, O(N, M )(X s ) ≃ DR (N ) and O(⊕i∈Λ R, M )(X s ) ≃ DR (⊕i∈Λ R).
DR commutes with direct sums by [10, 3.4.11. Corollary]. By using this fact,
Theorem 2.12 and Theorem 2.16, we get that O(N, M )(X s ) ≃ DR (⊕i∈Λ R) ≃
⊕i∈Λ DR (R) ≃ ⊕i∈Λ O(R, M )(X s ) ≃ ⊕i∈Λ R. This shows that O(N, M )(X s ) is
a free R-module.
(2) Since N is a projective R-module, there is a free R-module F and a
submodule L of F such that F ≃ N ⊕ L. By using Proposition 2.17 and
[10, 3.4.11. Corollary], we get that O(F, M )(X s ) ≃ O(N ⊕ L, M )(X s ) ≃
O(N, M )(X s ) ⊕ O(L, M )(X s ). By part (1), O(F, M )(X s ) is a free R-module.
O(N, M )(X s ) is a projective R-module as it is a direct summand of the free
R-module O(F, M )(X s ).
(3) Since every flat R-module is a direct limit of projective R-modules, N ≃
lim
P
for some projective R-modules Pi and a directed set Λ. By Proposition
−−→ i
i∈Λ
P ). DR
P , M )(X s ) ≃ DR (lim
2.17 and Theorem 2.12, O(N, M )(X s ) ≃ O(lim
−−→ i
−−→ i
i∈Λ
i∈Λ
commutes with direct limits by [10, 3.4.11. Corollary]. By using this fact,
Theorem 2.12 and Theorem 2.16, we get that O(N, M )(X s ) ≃ DR (lim
P) ≃
−−→ i
i∈Λ
lim
D (Pi ) ≃ lim
O(Pi , M )(X s ). By part (2), O(Pi , M )(X s ) is a projective and
−−→ R
−−→
i∈Λ
i∈Λ
hence a flat R-module for each i ∈ Λ. Since a direct limit of flat modules is flat,
O(N, M )(X s ) is a flat R-module.
3
A Scheme Structure On The Second Spectrum Of A Module
Recall that an affine scheme is a locally ringed space which is isomorphic to the
spectrum of some ring. A scheme is a locally ringed space (X, OX ) in which
every point has an open neighborhood U such that the topological space U ,
together with the restricted sheaf OX|U is an affine scheme. A scheme (X, OX )
is called locally Noetherian if it can be covered by open affine subsets of Spec(Ai )
where each Ai is a Noetherian ring. The scheme (X, OX ) is called Noetherian
if it is locally Noetherian and quasi-compact (cf. [19]).
A topological space X is said to be a T0 -space or a Kolmogorov space if for
every pair of distinct points x, y ∈ X, there exists open neighbourhoods U of x
and V of y such that either x ∈
/ V or y ∈
/ U . The following proposition, from
[6], gives some conditions for the dual Zariski topology of an R-module to be a
T0 -space.
Proposition 3.1 [6, Theorem 6.3] The following statements are equivalent for
an R-module M .
(1) The natural map ψ s : Specs (M ) −→ Spec(R/annR(M )) is injective.
(2) For any S1 , S2 ∈Specs (M ), if V s (S1 ) = V s (S2 ) then S1 = S2 .
12
(3) | Specsp (M ) |≤ 1 for every p ∈ Spec(R), where Specsp (M ) is the set of
all p-second submodules of M .
(4) (Specs (M ), τ s ) is a T0 -space.
Theorem 3.2 Let M be a faithful secondful R-module such that X s is a T0 space. Then (X s , O(R, M )) is a scheme. Moreover, if R is Noetherian, then
(X s , O(R, M )) is a Noetherian scheme.
Proof. Let g ∈ R. Since the natural map ψ sM : Specs (M ) −→ Spec(R) is
continuous by [6, Proposition 3.6], the restriction map ψ sM|Yg : Yg −→ ψ sM (Yg )
is also continuous. Since Yg is also a T0 -space, ψ sM|Yg is a bijection. Let E
be closed subset of Yg . Then E = Yg ∩ V s (N ) for some N ≤ M . Hence
ψ sM (E) = ψ sM (Yg ∩ V s (N )) = ψ sM (Yg ) ∩ ψ sM (V s (N )) = ψ sM (Yg ) ∩ V (annR (N ))
is a closed subset of ψ sM (Yg ). Therefore, ψ sM|Yg is a homeomorphism.
Since the sets of the form Yg (g ∈ R) form a base for the dual Zariski
topology, X s can be written as X s = ∪i∈I Ygi for some gi ∈ R. Since M
is faithful secondful and X s is a T0 -space, we have Ygi ≃ ψ sM (Ygi ) = Dgi ≃
Spec(Rgi ) for each i ∈ I. By Theorem 2.16, Ygi is an affine scheme for each
i ∈ I. This implies that (X s , O(R, M )) is a scheme.
For the last statement, we note that since R is Noetherian so is Rgi for each
i ∈ I. Hence (X s , O(R, M )) is a locally Noetherian scheme. By [11, Theorem
4], X s is quasi-compact. Therefore, (X s , O(R, M )) is a Noetherian scheme.
Theorem 3.3 Let M and N be R-modules and φ : M −→ N be a monomorphism. Then φ induces a morphism of locally ringed spaces
(f, f # ) : (Specs (N ), O(R, N )) −→ (Specs (M ), O(R, M )).
Proof. By [6, Proposition 3.14], the map f : Specs (M ) −→ Specs (N ) which
is defined by f (S) = φ(S) for every S ∈ Specs (M ), is continuous. Let U be
an open subset of Specs (N ) and β ∈ O(R, N )(U ). Suppose S ∈ f −1 (U ). Then
f (S) = φ(S) ∈ U . There exists an open neighborhood W of φ(S) with φ(S) ∈
W ⊆ U such that for each Q ∈ W , g 6∈ q = annR (Q) and β q = ag in Rq . Since
φ(S) = f (S) ∈ W , S ∈ f −1 (W ) ⊆ f −1 (U ). As f is continuous, f −1 (W ) is an
open neighborhood of S. We claim that for each Q′ ∈ f −1 (W ), g 6∈ annR (Q′ ).
Suppose on the contrary that g ∈ annR (Q′ ) for some Q′ ∈ f −1 (W ). Then
f (Q′ ) = φ(Q′ ) ∈ W . Since φ is a monomorphism, annR (Q′ ) = annR (φ(Q′ )).
So g ∈ annR (φ(Q′ )) for φ(Q′ ) ∈ W , a contradiction. Therefore for every open
subset U of Specs (N ), we can define the map
f # (U ) : O(R, N )(U ) −→ O(R, M )(f −1 (U ))
as follows. For every β ∈ O(R, N )(U ), f # (U )(β) ∈ O(R, M )(f −1 (U )) is defined
by f # (U )( β p p∈Supps (U) ) = (β annR (f (S) )S∈f −1 (U) . As we mentioned above
f # (U ) is a well-defined map an clearly it is a ring homomorphism. Now we show
that f # is a locally ringed morphism. Assume that U and V are open subsets
13
of Specs (N ) with V ⊆ U and β = β p p∈Supps (U) ∈ O(R, N )(U ). Consider the
diagram
O(R, N )(U ) f # (U ) O(R, M )(f −1 (U ))
−−−−→
↓ ρ′UV
↓ ρf −1 (U)f −1 (V )
#
O(R, N )(V ) f (V ) O(R, M )(f −1 (V )).
−−−−→
Then
ρf −1 (U)f −1 (V ) f # (U )( β p
p∈Supps (U)
) =
=
=
ρf −1 (U)f −1 (V ) ((β annR (f (S) )S∈f −1 (U) )
(β annR (f (S) )S∈f −1 (V ) = f # (V )((β p )p∈Supps (V ) )
f # (V )ρ′UV ( β p p∈Supps (U) ).
Therefore we get that ρf −1 (U)f −1 (V ) f # (U ) = f # (V )ρ′UV . Thus the above
diagram is commutative. This shows that f # : O(R, N ) −→ f∗ (O(R, M ))
is a morphism of sheaves. By Theorem 2.3, the map on the stalks, fS# :
O(R, N )f (S) −→ O(R, M )S is clearly the map of local rings RannR (f (S)) −→
RannR (S) which maps rs ∈ RannR (f (S)) to rs again. This implies that (f, f # ) :
(Specs (N ), O(R, N )) −→ (Specs (M ), O(R, M )) is a morphism of locally ringed
spaces.
Theorem 3.4 Let M , A be R-modules, φ : R −→ S be a ring homomorphism,
let N, B be S-modules and M be a secondful R-module such that Specs (M ) is
a T0 -space and annR (M ) ⊆ annR (N ). If δ : A −→ B is an R-homomorphism,
then φ induces a morphism of sheaves
h# : O(A, M ) −→ h∗ (O(B, N ))
Proof. Since annR (M ) ⊆ annR (N ), φ induces the homomorphism
Θ : R/annR (M ) −→ S/annS (N ), Θ(r + annR (M )) = φ(r) + annS (N ).
It is well-known that the maps
f : Spec(S) −→ Spec(R) defined by f (p) = φ−1 (p) and
d : Spec(S/annS (N )) −→ Spec(R/annR(M )) defined by d(p) = Θ−1 (p) and
ψ sN : Specs (N ) −→ Spec(S/annS (N )) defined by ψ sN (Q) = annS (Q)/annS (N )
for each Q ∈ Specs (N ) are continuous. Also, ψ sM : Specs (M ) −→ Spec(R/annR (M ))
is a homeomorphism by [6, Theorem 6.3]. Therefore the map h : Specs (N ) −→
Specs (M ) defined by h(Q) = (ψ sM )−1 dψ sN (Q) is continuous. Also, for each
Q ∈ Specs (N ), we get an R-homomorphism
φannS (Q) : Af (annS (Q))
a
s
−→ BannS (Q)
δ(r)
−→
φ(s)
Let U be an open subset of Specs (M ) and t = (tannR (P ) )P ∈U ∈ O(A, M )(U ).
Suppose that T ∈ h−1 (U ). Then h(T ) ∈ U and there exists an open neighborhood W of h(T ) with h(T ) ∈ W ⊆ U and elements r, g ∈ R such that
14
for each Q ∈ W , we have tannR (Q) = ag ∈ AannR (Q) where g 6∈ annR (Q),
hence g 6∈ annR (h(T )). By definition of h, annR (h(K)) = φ−1 (annS (K))
for every K ∈ h−1 (W ). So φ(g) 6∈ annS (K) for g 6∈ annR (h(K)). Thus
δ(a)
define a section O(B, N )(h−1 (U )). We can define
φannS (T ) ( ag ) = φ(g)
h# (U ) : O(A, M )(U ) −→ h∗ (O(B, N )(U )) = O(B, N )(h−1 (U ))
by h# (U )((tannR (P ) )P ∈U ) = φannS (T ) (tannR (h(T )) )
for each (tannR (P ) )P ∈U ∈
T ∈h−1 (U)
O(R, M )(U ). Assume that V ⊆ U . Consider the diagram
O(A, M )(U )
↓ ρUV
O(A, M )(V )
We see that
h# (U )
−−−−→
#
h (V )
−−−−→
O(B, N )(h−1 (U ))
↓ ρ′h−1 (U)h−1 (V )
O(B, N )(h−1 (V ))
φannS (T ) (tannR (h(T )) )
ρ′h−1 (U)h−1 (V ) h# (U )((tannR (P ) )P ∈U ) = ρ′h−1 (U)h−1 (V )
T ∈h−1 (U)
=
φannS (T ) (tannR (h(T )) )
−1
T ∈h
#
(V )
#
= h (V )((tannR (P ) )P ∈V ) = h (V )ρUV ((tannR (P ) )P ∈U )
and hence ρ′h−1 (U)h−1 (V ) h# (U ) = h# (V )ρUV for every open subset U of Specs (M ).
So the above diagram is commutative. It follows that h# : O(A, M ) −→
h∗ (O(B, N )) is a morphism of sheaves.
Corollary 3.5 Let φ : R −→ S be a ring homomorphism, let N be an Smodule and M be a secondful R-module such that Specs (M ) is a T0 -space and
annR (M ) ⊆ annR (N ). Then φ induces a morphism of locally ringed spaces
(h, h# ) : (Specs (N ), O(S, N )) −→ (Specs (M ), O(R, M ))
Proof. Taking A = R, B = S and δ = φ in Theorem 3.4, we get the morphism
of sheaves h# : O(R, M ) −→ h∗ (O(S, N )) which is defined as in the proof of
Theorem 3.4. By Theorem 2.3, the map on the stalks h#
T : O(R, M )h(T ) −→
O(S, N )T is clearly the local homomorphism
φannS (T ) : Rf (annS (T ))
r
s
−→ SannS (T )
φ(r)
−→
φ(s)
where f is the map defined in the proof of Theorem 3.4. This implies that
(h, h# ) : (Specs (N ), O(S, N )) −→ (Specs (M ), O(R, M ))
is a locally ringed spaces.
Acknowledgement
The authors would like to thank the Scientific Technological Research Council of Turkey (TUBITAK) for funding this work through the project 114F381.
The second author was supported by the Scientific Research Project Administration of Akdeniz University.
15
References
[1] A. Abbasi and D. Hassanzadeh-Lelekaami, A scheme over prime spectrum
of modules, Turkish J. Math. 37 (2013), 202–209.
[2] J. Y. Abuhlail, A dual Zariski topology for modules, Topology Appl. 158
(2011) 457–467.
[3] J.Y.Abuhlail, Zariski topologies for coprime and second submodules, Algebra Colloquium, 22 (1), (2015).
[4] H. Ansari-Toroghy, F. Farshadifar, On the dual notion of prime submodules, Algebra Colloq., 19 (1), (2012), 1109-1116.
[5] H. Ansari-Toroghy, F. Farshadifar, On the dual notion of prime submodules
II, Mediterr. J. Math., 9 (2), (2012), 327-336.
[6] H. Ansari-Toroghy and F. Farshadifar, The Zariski topology on the
second spectrum of a module, Algebra Colloquium, 671 (2014). DOI:
10.1142/S1005386714000625.
[7] H. Ansari-Toroghy, F. Farshadifar, On the dual notion of prime radicals of submodules, Asian-European J. Math, 6 (2), (2013), DOI:
10.1142/S1793557113500241.
[8] H. Ansari-Toroghy, S. Keyvani and F. Farshadifar, On the second spectrum
of a module (II), Bull. Malays. Math. Sci. Soc. 39 (3), (2016), 1089–1103.
[9] H. Ansari-Toroghy, S. S. Pourmortazavi, S. Keyvani, Strongly cotop modules, Journal of Algebra and Related Topics 3 (1), (2015), 13-29.
[10] M. P. Brodmann and R. Y. Sharp, Local cohomology: an algebraic introduction with geometric applications, Cambridge Univercity Press, 1998.
[11] S. Çeken and M. Alkan, Dual of Zariski topology for modules, Book Series:
AIP Conference Proceedings, 1389 (1), (2011), 357-360.
[12] S. Çeken, M. Alkan and P. F. Smith, Second modules over noncommutative
rings, Communications in Algebra, 41 (1), (2013), 83-98.
[13] S. Çeken, M. Alkan and P. F. Smith, The dual notion of the prime radical
of a module, Journal of Algebra, 392 (2013), 265–275.
[14] S. Çeken, M. Alkan, On graded second and coprimary modules and graded
secondary representations, Bull. Malays. Math. Sci. Soc. 38 (4), (2015),
1317–1330.
[15] S. Çeken, M. Alkan, On second submodules, Contemporary Mathematics
634, (2015), 67-77.
16
[16] S. Çeken, M. Alkan, On the second spectrum and the second classical
Zariski topology of a module, Journal of Algebra And Its Applications 14
(10), (2015), DOI: http://dx.doi.org/10.1142/S0219498815501509.
[17] S. Çeken, M.Alkan, Second Spectrum Of Modules And Spectral Spaces,
Bulletin of the Malaysian Mathematical Sciences Society, (2017), DOI:
10.1007/s40840-017-0473-0.
[18] F. Farshadifar, Modules with Noetherian second spectrum, Journal of Algebra and related topics, Vol. 1, No 1, (2013), pp 19-30.
[19] R. Hartshorne, Algebraic geometry, no. 52, Springer-Verlag New York Inc,
1977.
[20] D. Hassanzadeh-Lelekaami, H. Roshan-Shekalgourabi, Prime submodules
and a sheaf on the prime spectra of modules, Communications in Algebra
42 (7), (2014) 3063-3077.
[21] C. P. Lu, Spectra of modules, Comm. Algebra, 23 (10) (1995), 3741-3752.
[22] C.P. Lu, The Zariski topology on the prime spectrum of a module, Houston
J. Math. 25 (3) (1999) 417-432.
[23] C. P. Lu, A module whose prime spectrum has the surjective natural map,
Houston J. Math, 33 (1) (2007), 125-143.
[24] C.P. Lu, Modules with Noetherian spectrum, Comm. Algebra 38 (2010),
807-828.
[25] R. L. McCasland, M. E. Moore and P. F. Smith, On the spectrum of a
module over a commutative ring, Comm. Algebra 25 (1997), 79-103.
[26] Ü. Tekir, On the sheaf of modules, Comm. Algebra 33 (2005), no. 8, 2557–
2562.
[27] S. Yassemi, The dual notion of prime submodules, Arch. Math (Brno) 37
(2001), 273–278.
Seçil Çeken: Trakya University, Faculty of Sciences, Department of Mathematics, Edirne, Turkey.
Mustafa Alkan: Akdeniz University, Faculty of Sciences, Department of
Mathematics, Antalya, Turkey.
17
| 0math.AC
|
TOWARDS AUTOMATIC ABDOMINAL MULTI-ORGAN SEGMENTATION IN DUAL
ENERGY CT USING CASCADED 3D FULLY CONVOLUTIONAL NETWORK
Shuqing Chen1
Michael M. Lell5
Holger Roth2
Sabrina Dorn3
Matthias May4
Alexander Cavallaro4
Marc Kachelrieß3
Hirohisa Oda2
Kensaku Mori2
Andreas Maier1
arXiv:1710.05379v1 [cs.CV] 15 Oct 2017
1
Friedrich-Alexander University Erlangen-Nürnberg, Erlangen, Germany
2
Nagoya University, Nagoya , Japan
3
German Cancer Research Center (DKFZ), Heidelberg, Germany
4
Department of Radiology, University Hospital Erlangen, Erlangen, Germany
5
University Hospital Nürnberg, Paracelsus Medical University, Nürnberg, Germany
This work has been submitted to the IEEE for possible
publication. Copyright may be transferred without notice, after which this version may no longer be accessible.
ABSTRACT
Automatic multi-organ segmentation of the dual energy computed tomography (DECT) data can be beneficial for biomedical research and clinical applications. However, it is a challenging task. Recent advances in deep learning showed the
feasibility to use 3-D fully convolutional networks (FCN)
for voxel-wise dense predictions in single energy computed
tomography (SECT). In this paper, we proposed a 3D FCN
based method for automatic multi-organ segmentation in
DECT. The work was based on a cascaded FCN and a general model for the major organs trained on a large set of
SECT data. We preprocessed the DECT data by using linear
weighting and fine-tuned the model for the DECT data. The
method was evaluated using 42 torso DECT data acquired
with a clinical dual-source CT system. Four abdominal organs (liver, spleen, left and right kidneys) were evaluated.
Cross-validation was tested. Effect of the weight on the accuracy was researched. In all the tests, we achieved an average
Dice coefficient of 93% for the liver, 90% for the spleen, 91%
for the right kidney and 89% for the left kidney, respectively.
The results show our method is feasible and promising.
Index Terms— DECT, deep learning, multi-organ segmentation, U-Net
1. INTRODUCTION
The Hounsfield unit (HU) scale value depends on the inherent tissue properties, the x-ray spectrum for scanning and the
administered contrast media [1]. In a SECT image, materials having different elemental compositions can be represented by identical HU values [2]. Therefore, SECT has challenges such as limited material-specific information and beam
hardening as well as tissue characterization [1]. DECT has
been investigated to solve the challenges of SECT. In DECT,
two energy-specific image data sets are acquired at two different X-ray spectra, which are produced by different energies, simultaneously. The multi-organ segmentation in DECT
can be beneficial for biomedical research and clinical applications, such as material decomposition [3], organ-specific
context-sensitive enhanced reconstruction and display [4, 5],
and computation of bone mineral density [6]. We are aiming
at exploiting the prior anatomical information that is gained
through the multi-organ segmentation to provide an improved
context-sensitive DECT imaging [4, 5]. The novel technique
offers the possibility to present evermore complex information to the radiologists simultaneously and bears the potential
to improve the clinical routine in CT diagnosis.
Automatic multi-organ segmentation on DECT images is
a challenging task due to the inter-subject variance of human
abdomen, the complex 3-D intra-subject variance among organs, soft anatomy deformation, as well as different HU values for the same organ by different spectra. Recent researches
show the power of deep learning in medical image processing [7]. To solve the DECT segmentation problem, we use
the successful experience from multi-organ segmentation in
volumetric SECT images using deep learning [8, 9]. The proposed method is based on a cascaded 3D FCN, a two-stage,
coarse-to-fine approach [8]. The first stage is used to predict
the region of the interest (ROI) of the target organs, while the
second stage is learned to predict the final segmentation. No
organ-specific or energy-specific prior knowledge is required
in the proposed method. The cross-validation results showed
that the proposed method is promising to solve multi-organ
segmentation problem for DECT. To the best of our knowledge, this is the first study about multi-organ segmentation in
DECT images based on 3D FCNs.
2. MATERIALS AND METHODS
2.1. Network Architecture for DECT Prediction
DECT
As described by Krauss et al. [10], a mixed image display is
employed in clinical practice for the diagnose using DECT.
The mixed image is calculated by linear weighting of the images values of the two spectra:
Imix = α · Ilow + (1 − α) · Ihigh
(1)
where α is the weight of the dual energy composition, Imix
denotes the mixed image. Ilow and Ihigh are the images at low
and high kV, respectively.
We preprocessed the DECT images following Eq. 1
straightforwardly. Figure 1 illustrates the network architecture of the proposed method for the DECT multi-organ
segmentation. First of all, mixed image is calculated by combining the images at the low energy level and the high energy
level using Eq. 1. Then, a binary mask is generated by thresholding the skin contour of the mixed image. Subsequently,
the mixed image, the binary mask and the labeled image are
given into the network as multi-channel inputs. The network
consists of two stages. The first stage is applied to generate
the region of the interest (ROI) in order to reduce the search
space for the second stage. The prediction result of the first
stage is taken as the mask for the second stage. Each stage is
based on a standard 3D U-Net [11], which is a fully convolutional network including an analysis and a synthesis path. We
used the open-source implementation of two stages cascaded
network [8] developed by Roth et al. based on the 3D UNet [11] and the Caffe deep learning library [12]. A general
model was trained by Roth et al. [8] on a large set of SECT
images including some of the major organ labels. Our model
was trained by fine-tuning the general model with the mixed
DECT images. The difference between the network output
and the ground truth labels are compared using softmax with
weight voxel-wise cross-entropy loss [8, 11].
SECT
[9]
Avg.
SD
Min.
Max.
Avg.
SD
Min.
Max.
Liver
0.92
0.02
0.84
0.94
0.95
0.02
0.91
0.97
Spleen
0.84
0.08
0.62
0.95
0.90
0.06
0.75
0.95
r.Kidney
0.88
0.03
0.80
0.94
0.90
0.06
0.75
0.95
l.Kidney
0.87
0.03
0.78
0.93
0.88
0.05
0.78
0.94
Table 1:
Dice coefficients of cross-validation with
αtraining =0.6 and αtest =0.6. SD is abbreviated for standard
deviation. Notice that the methods used different data set, the
numbers are not directly comparable.
an inter-observer way. Training data, validation data, and test
data were selected randomly with the ratio 5:1:1, i.e. in each
test we used 6 images for validation, 6 images for test, and 30
images for training.
3. RESULTS
3.1. Performance Estimation with Cross-Validation
NVIDIA GeForce GTX 1080 Ti with 11 GB memory was
used for all of the experiments. The similarity between the
segmentation result and the ground truth was measured with
Dice metric by using the tool provided by VISCERAL [13].
First, the performance of the proposed method was estimated
by 8-folds cross-validation, using 0.6 as αtraining as well as
αtest . Fig. 2 shows one segmentation results in 3-D. Table1
summarizes the Dice coefficients of the segmentation results
and compares DECT results with the SECT results. The proposed method under the above weight condition yielded an
average Dice coefficient of 92% for the liver, 84% for the
spleen, 88% for the right kidney and 87% for the left kidney, respectively. Fig. 3 plots the distributions of the Dice
coefficients for different test scenarios and showed the high
robustness of the proposed method.
2.2. Experimental Setup
The proposed method was evaluated with 42 clinical torso
DECT images scanned by the department of radiology, university hospital Erlangen. All of the images were taken from
male and female adult patients who had different clinically
oriented indication justified by the radiologist. Ultravist 370
was given as contrast agent with body weight adapted volumes. The images were acquired at different X-ray tube voltage setting of 70 kV (560 mAs) and Sn 150 kV (140 mAs,
with Sn filter) using a Siemens SOMATOM Force CT system
with Stellar detector, an energy integrating detector. Each volume consists of 992-1290 slices of 512x512 pixels. The voxel
dimensions are [0.6895-0.959, 0.6895-0.959, 0.6] mm. Four
abdominal organs were tested, including liver, spleen, right
and left kidneys. Ground truth was generated by experts in
3.2. Study on the Weight α
We are aiming at exploiting the spectral information in the
DECT data. Since the α mixing results basically in pseudo
monochromatic images comparable to single energy scans,
the influence of the weight α on the accuracy was further researched. 0.3, 0.6 and 0.9 were chosen as αtraining and αtest
in this study. Fig. 4 illustrates the distributions of the Dice
coefficients with different weight combination for the testing
fold 1. Table 2 lists the average Dice coefficient. For all of
the cases, the liver had the highest accuracy (92%-93%), the
standard deviation of the dice coefficients (around 2%) was
fairly robust. The segmentation of the right kidney was usually more accurate than the left kidney. The best Dice values
per organ per training set are highlighted in Table 2. The test
Fig. 1: Cascaded network architecture for DECT multi-organ segmentation
Fig. 2: 3D rendering of one DECT segmentation with yellow
for liver, blue for spleen, green for right kidney and red for
left kidney
Fig. 4: Dice coefficients of target organs with alpha blending
for testing fold 1
with αtraining =0.9 and αtest =0.9 obtained the highest accuracy
for liver and right kidney. The test with weight combination
0.9-0.6 showed the best segmentation for spleen, the combination with 0.9-0.3 had the finest result for left kidney. The
αtest =0.9 generated better segmentation for liver, the αtest =0.6
worked better for spleen.
4. DISCUSSION AND CONCLUSION
Fig. 3: Dice coefficients of the target organs with αtraining =
0.6 and αtest = 0.6 for 8 different testing folds
We proposed a deep learning based method for automatic abdominal multi-organ segmentation in DECT. The evaluation
results show the feasibility of the proposed method. Compared to the results of the SECT images reported by Roth et
al. [9], our method is promising and robust (see Table 1). The
segmentation of liver and spleen was less accurate than the
SECT. The third testing fold had a large deviation. The reason could be that our image data were taken from patients
with different disease (liver tumor, spleen tumor, etc.). The
disease type is not considered by the data selection. Training
αtraining -αtest
0.3-0.3
0.3-0.6
0.3-0.9
0.6-0.3
0.6-0.6
0.6-0.9
0.9-0.3
0.9-0.6
0.9-0.9
Liver
0.924
0.925
0.926
0.909
0.922
0.912
0.930
0.932
0.933
Spleen
0.850
0.885
0.866
0.847
0.899
0.872
0.860
0.900
0.896
r.Kidney
0.900
0.891
0.872
0.844
0.895
0.843
0.898
0.904
0.905
l.Kidney
0.891
0.881
0.841
0.885
0.887
0.873
0.892
0.873
0.862
Table 2: Dice coefficients of different alpha for testing fold
1. Bold denotes the best organ results per training set.
and test with inconsistent symptoms could have an impact on
the accuracy.
The study on the weight can be divided into three groups
with different αtraining . α=0.9 is close to the low energy
images which have on average the best soft-tissue contrast,
αtraining =0.9 worked thus better in general. α=0.6 is close
to α=0.5 which is the optimal fusion of both images with
respect to signal-to-noise ratio (SNR), αtraining =0.6 had therefore usually the smallest deviation and showed the strongest
adaptability in the inter-group comparison. The intra-group
comparison showed that the cases with identical training and
test conditions had a higher probability to get the best segmentation result. This is expected because the mixed images
generated by the matched training and test conditions may
have the highest similarity. Furthermore, the comparison of
the case 0.3-0.9 (low-contrast model for high-contrast image)
with the case 0.9-0.3 (high-contrast model for low-contrast
image) showed that using a model trained on high-contrast
images for segmenting low-contrast test images works better. In addition, liver is well segmented in middle to high α
ranges. Spleen is segmented best at α=0.6. Kidneys work best
in matched training and test conditions. This suggests that
there is an optimal α for each organ for image segmentation.
The weight α for the mixed image calculation is currently a user-defined parameter in the preprocessing in our
approach. It can be used to augment the data for the training
in future. Also, the net could be modified with two image
inputs. Furthermore, more organs and more scans from different patients could be used.
ACKNOWLEDGMENTS This work was supported by
the German Research Foundation (DFG) through research
grant No. KA 1678/20, LE 2763/2-1 and MA 4898/5-1.
5. REFERENCES
[1] Dushyant Sahani, “Dual-energy CT: The technological
approaches,” Society of Computed Body Tomography
and Magnetic Resonance (SCBT-MR), 2012.
[2] Cynthia H. McCollough, Shuai Leng, Lifeng Yu, and
Joel G. Fletcher, “Dual- and multi-energy CT: Principles, technical approaches, and clinical applications,”
Radiology, vol. 276, no. 3, pp. 637–653, 2015.
[3] Stefan Kuchenbecker, Sebastian Faby, David Simons,
Michael Knaup, Heinz-Peter Schlemmer, Michael M.
Lell, and Marc Kachelriess, “Segmentation-assisted material decomposition in dual energy computed tomography (DECT),” in Radiological Society of North America
(RSNA), 2015.
[4] Sabrina Dorn, Shuqing Chen, Stefan Sawall, Andreas
Maier, Michael Lell, and Marc Kachelrieß, “Organspecific context-sensitive single and dual energy CT
(DECT) image reconstruction, display and analysis,” in
Radiological Society of North America, 2017.
[5] Sabrina Dorn, Shuqing Chen, Stefan Sawall, David Simons, Matthias May, Heinz-Peter Schlemmer, Andreas
Maier, Michael Lell, and Marc Kachelrieß, “Organspecific context-sensitive CT image reconstruction and
display,” in SPIE accepted, 2018.
[6] S Wesarg, M Kirschner, M Becker, M Erdt, K Kafchitsas, MF Khan, et al., “Dual-energy CT-based assessment of the trabecular bone in vertebrae,” Methods of
information in medicine, vol. 51, no. 5, pp. 398, 2012.
[7] Marc Aubreville, Miguel Goncalves, Christian Knipfer,
Nicolai Oetter, Tobias Würfl, Helmut Neumann, Florian Stelzle, Christopher Bohr, and Andreas K. Maier,
“Patch-based carcinoma detection on confocal laser endomicroscopy images - A cross-site robustness assessment,” CoRR, vol. abs/1707.08149, 2017.
[8] Holger R. Roth, Hirohisa Oda, Yuichiro Hayashi,
Masahiro Oda, Natsuki Shimizu, Michitaka Fujiwara,
Kazunari Misawa, and Kensaku Mori, “Hierarchical 3D
fully convolutional networks for multi-organ segmentation,” in arXiv preprint arXiv:1704.06382.
[9] Holger Roth, Ying Yang, Masahiro Oda, Hirohisa Oda,
Yuichiro Hayashi, Natsuki Shimizu, Takayuki Kitasaka,
Michitaka Fujiwara, Kazunari Misawa, and Kensaku
Mori, “Torso organ segmentation in CT using fine-tuned
3D fully convolutional networks,” in 36 (JAMIT), 2017.
[10] Bernhard Krauss, Bernhard Schmidt, and Thomas G.
Flohr, Dual Energy CT in Clinical Practice, chapter
Dual Source CT, Springer Berlin Heidelberg, 2011.
[11] Özgün Çiçek, Ahmed Abdulkadir, Soeren S. Lienkamp,
Thomas Brox, and Olaf Ronneberger, “3D U-Net:
Learning dense volumetric segmentation from sparse
annotation,” in Medical Image Computing and Computer Assisted Intervention (MICCAI), 2016.
[12] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey
Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell, “Caffe: Convolutional architecture for fast feature embedding,” arXiv preprint
arXiv:1408.5093, 2014.
[13] Abdel Aziz Taha and Allan Hanbury, “Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool,” BMC Medical Imaging, vol. 15, pp. 29,
August 2015.
| 1cs.CV
|
ON THE FREENESS OF RATIONAL CUSPIDAL PLANE CURVES
arXiv:1802.06688v2 [math.AG] 26 Feb 2018
ALEXANDRU DIMCA AND GABRIEL STICLARU
Abstract. We bring additional support to the conjecture saying that a rational
cuspidal plane curve is either free or nearly free. This conjecture was confirmed
for curves of even degree, and in this note we prove it for many odd degrees. In
particular, we show that this conjecture holds for the curves of degree at most 34.
1. Introduction
A plane rational cuspidal curve is a rational curve C : f = 0 in the complex
projective plane P2 , having only unibranch singularities. The study of these curves
has a long and fascinating history, some long standing conjectures, as the CoolidgeNagata conjecture being proved only recently, see [20], other conjectures, as the one
on the number of singularities of such a curve being bounded by 4, see [24], are still
open. The classification of such curves is not easy, there are a wealth of examples
even when additional strong restrictions are imposed, see [17, 18, 19, 22, 23, 27].
Free divisors, defined by a homological property of their Jacobian ideals, have
been introduced in a local analytic setting by K. Saito in [25], and then extended
to projective hypersurfaces, see [3, 29] and the references there. We have remarked
in [13] that many plane rational cuspidal curves are free. The remaining examples
of plane rational cuspidal curves in the available classification lists turned out to
satisfy a weaker homological property, which was chosen as the definition of a nearly
free curve, see [14]. Subsequently, a number of authors have establish interesting
properties of this class of curves, see [2, 21].
In view of the above remark, we have conjectured in [14, Conjecture 1.1] that any
plane rational cuspidal curve C is either free or nearly free. This conjecture was
proved in [14, Theorem 3.1] for curves C whose degree d is even, as well as for some
cases when d is odd, e.g. when d = pk , for a prime number p > 2. In this note we
take a closer look at the case d odd.
Let S = C[x, y, z] be the polynomial ring in three variables x, y, z with complex
coefficients, f ∈ S a reduced homogeneous polynomial of degree d ≥ 2, and let fx , fy
and fz be the partial derivatives of f with respect to x, y and z respectively. Consider
the graded S−submodule AR(f ) ⊂ S 3 of all relations involving the derivatives of f ,
namely
ρ = (a, b, c) ∈ AR(f )q
2010 Mathematics Subject Classification. Primary 14H45; Secondary 14B05, 14H50, 13D02,
32S35.
Key words and phrases. rational cuspidal curves, Jacobian syzygy, Tjurina number, free curves,
nearly free curves.
1
2
ALEXANDRU DIMCA AND GABRIEL STICLARU
if and only if afx + bfy + cfz = 0 and a, b, c are in Sq , the space of homogeneous
polynomials of degree q. The minimal degree of a Jacobian relation for the polynomial f ∈ Sd is the integer mdr(f ) defined to be the smallest integer m ≥ 0 such that
AR(f )m 6= 0. When mdr(f ) = 0, then C : f = 0 is a union of lines passing through
one point, and hence C is cuspidal only for d = 1.
We assume from now on in this note that mdr(f ) ≥ 1. It turns out that a rational
cuspidal curve C : f = 0 with mdr(f ) = 1 is nearly free. Indeed, this follows from
[8, Proposition 4.1]. To see this, note that the implication (1) ⇒ (2) there holds for
any d ≥ 2.
Assume from now on that d is odd, and let
(1.1)
d = pk11 · pk22 · · · pkmm
be the prime decomposition of d. We assume also that m ≥ 2, the case m = 1
of our conjecture being settled in [14, Corollary 3.2]. By changing the order of the
k
pj ’s if necessary, we can and do assume that pk11 > pj j , for any 2 ≤ j ≤ m. Set
e1 = d/pk11 . With these assumptions and notations, the main results of this note are
the following.
Theorem 1.1. Let C : f = 0 be a rational cuspidal curve of degree d = 2d′ + 1 an
odd number. Then mdr(f ) ≤ d′ and if equality holds, then C is either free or nearly
free.
Theorem 1.2. Let C : f = 0 be a rational cuspidal curve of degree d = 2d′ + 1, an
odd number as in (1.1). Then, if
d − e1
,
2
then C is either free or nearly free. In particular, the following hold.
i) If d = 3pk , with p a prime number, then C is either free or nearly free.
ii) d = 5pk , with p a prime number, pk > 3, then C is either free or nearly free,
unless mdr(f ) = d′ − 1.
mdr(f ) ≤ r0 :=
Remark 1.3. Note that, for d 6= 15, we have e1 ≤ d/7 and hence
d(1 − 71 )
d − e1
3d
r0 =
=
.
≥
2
2
7
Therefore, the only cases not covered by our results correspond to curves of odd
degree d, such that r = mdr(f ) satisfies
3d
d−3
+ 1 ≤ r0 + 1 ≤ r ≤ d ′ − 1 =
.
7
2
Corollary 1.4. A rational cuspidal curve C : f = 0 of degree d is either free or
nearly free, if one of the following holds.
(1) mdr(f ) ≤ 15, or
(2) d ≤ 90, unless we are in one of the following situations.
i) d = 35 and mdr(f ) = 16;
ON THE FREENESS OF RATIONAL CUSPIDAL PLANE CURVES
ii)
iii)
iv)
v)
vi)
vii)
d = 45
d = 55
d = 63
d = 65
d = 77
d = 85
and
and
and
and
and
and
3
mdr(f ) = 21;
mdr(f ) = 26;
mdr(f ) ∈ {29, 30};
mdr(f ) = 31.
mdr(f ) ∈ {36, 37}.
mdr(f ) = 41.
In the excluded situations, our results do not allow us to conclude.
The proof of our main results are based on a deep result by U. Walther, see [31,
Theorem 4.3] bringing into the picture the monodromy of the Milnor fiber F : f = 1
associated to the curve C : f = 0. A second ingredient is our results on the relations
between the Hodge filtration and pole order filtration on the cohomology group
H 1 (F, C), see [12, Theorem 1.2] and [16, Proposition 2.2].
The first author thanks AROMATH team at INRIA Sophia-Antipolis for excellent
working conditions, and in particular Laurent Busé for stimulating discussions.
2. Some facts about free and nearly free curves
Here we recall some basic notions on free and nearly free curves. We denote by
Jf the Jacobian ideal of f , i.e. the homogeneous ideal of S spanned by the partial
derivatives fx , fy , fz and let M(f ) = S/Jf be the corresponding graded ring, called
the Jacobian (or Milnor) algebra of f . Let If denote the saturation of the ideal Jf
with respect to the maximal ideal m = (x, y, z) in S and recall the relation with the
0-degree local cohomology
0
(M(f )).
N(f ) := If /Jf = Hm
It was shown in [9, Corollary 4.3] that the graded S-module N(f ) satisfies a Lefschetz
type property with respect to multiplication by generic linear forms. This implies in
particular the inequalities
(2.1)
0 ≤ n(f )0 ≤ n(f )1 ≤ ... ≤ n(f )[T /2] ≥ n(f )[T /2]+1 ≥ ... ≥ n(f )T ≥ 0,
where T = 3d − 6 and n(f )k = dim N(f )k for any integer k. If we set ν(f ) =
dim N(f )[T /2] , then C : f = 0 is a free curve if ν(f ) = 0. We say that C : f = 0
is a nearly free curve if ν(f ) = 1, see [2, 7, 8, 13, 14] for more details, equivalent
definitions and many examples.
Note that the curve C : f = 0 is free if and only if the graded S-module AR(f )
is free of rank 2, i.e. there is an isomorphism of graded S-modules
AR(f ) = S(−d1 ) ⊕ S(−d2 )
for some positive integers d1 ≤ d2 . When C is free, the integers d1 ≤ d2 are called
the exponents of C. They satisfy the relations
(2.2)
d1 + d2 = d − 1 and τ (C) = (d − 1)2 − d1 d2 ,
Pp
where τ (C) is the total Tjurina number of C, that is τ (C) =
i=1 τ (C, xi ), the
xi ’s being the singular points of C, and τ (C, xi ) denotes the Tjurina number of the
4
ALEXANDRU DIMCA AND GABRIEL STICLARU
isolated plane curve singularity (C, xi ), see for instance [10, 13]. In the case of a
nearly free curve, there are also the exponents d1 ≤ d2 , and this time they verify
(2.3)
d1 + d2 = d and τ (C) = (d − 1)2 − d1 (d2 − 1) − 1.
Both for a free and a nearly free curve C : f = 0, one has mdr(f ) = d1 , and hence
mdr(f ) ≤ (d − 1)/2 for a free curve C, and mdr(f ) ≤ d/2 for a nearly free curve
C. It follows that Theorem 1.1 gives a similar inequality for any rational cuspidal
curve. Our examples of rational cuspidal curves given in [11], which are also free or
nearly free, show that all the possible values of mdr(f ) do actually occur for any
fixed degree d. It follows that
If we set r = mdr(f ), then the curve C : f = 0 is free (resp. nearly free) if and
only if
(2.4)
τ (C) = τ (d, r) := (d − 1)2 − r(d − 1 − r)
(resp. τ (C) = τ (d, r) − 1), see [7].
Remark 2.1. If the equation f = 0 of the curve C is given explicitly, then one can
use a computer algebra software, for instance Singular [4], in order to compute the
integer mdr(f ). Such a computer algebra software can of course decide whether the
curve C is free or nearly free, see for instance the corresponding code on our website
http://math.unice.fr/~ dimca/singular.html
However, for large degrees d, it is much quicker to determine the integer mdr(f ).
3. The proofs
First we recall the setting used in the proof of [14, Theorem 3.1]. The key results
of U. Walther in [31, Theorem 4.3] yield the inequality
(3.1)
dim N(f )2d−2−j ≤ dim H 2 (F, C)λ ,
for j = 1, 2, ..., d, where F : f (x, y, z)−1 = 0 is the Milnor fiber in C3 associated to the
plane curve C, and the subscript λ indicates the eigenspace of the monodromy action
corresponding to the eigenvalue λ = exp(2πi(d + 1 − j)/d) = exp(−2πi(j − 1)/d).
Assume that C is a rational cuspidal curve of degree d. Denote by U the complement P2 \ C, and note that its topological Euler characteristic is given by E(U) =
E(P2 )−E(C) = 1. Since F is a cyclic d-fold covering of the complement U, it follows
that H m (F, C)1 = H m (U, C) = 0 for m = 1, 2. We have also
dim H 2 (F, C)λ − dim H 1 (F, C)λ + dim H 0 (F, C)λ = E(U) = 1,
see for instance [5, Prop. 1.21, Chapter 4] or [6, Cor. 5.1 and Remark 5.1]. For any
λ 6= 1, since clearly H 0 (F, C)λ = 0, we get
(3.2)
dim H 2 (F, C)λ = dim H 1 (F, C)λ + 1.
ON THE FREENESS OF RATIONAL CUSPIDAL PLANE CURVES
5
3.1. Proof of Theorem 1.1. Suppose now that d is odd, say d = 2d′ + 1. In
order to prove ν(f ) ≤ 1, in view of the inequality (3.1), it is enough to show that
dim H 2 (F, C)λ = 1, for λ = exp(2πid′ /(2d′ + 1)) which corresponds to j = d′ + 2.
The equation (3.2) tells us that this is equivalent to dim H 1 (F, C)λ = 0. Using [16,
Proposition 2.2], see also [15, Remark 4.4], it follows that
dim H 1 (F, C)λ = dim E21,0 (f )k + dim E21,0 (f )d−k ,
where k = j − 1 = d′ + 1. Here E21,0 (f )k and E21,0 (f )d−k denote some terms of the
second page of spectral sequences used to compute the monodromy action on the
Milnor fiber F , see [5, 12, 15, 26] for details. Note also that the weaker result in
[12, Theorem 1.2] is enough for this proof. By the construction of these spectral
sequences, it follows that, for q ≤ d, one has an identification
E21,0 (f )q = {(a, b, c) ∈ AR(f )q−2 : ax + by + cz = 0},
where ax is the partial derivative of a with respect to x and so on. It follows that
dim E21,0 (f )k + dim E21,0 (f )d−k ≤ dim AR(f )d′ −1 + dim AR(f )d′ −2 .
If mdr(f ) ≥ d′ , it follows that AR(f )d′ −1 = AR(f )d′ −2 = 0, and hence the curve C
is either free or nearly free. But this implies that mdr(f ) ≤ d′ , as explained in the
previous section.
3.2. Proof of Theorem 1.2. For the reader’s convenience, we divide this proof into
two steps.
Proposition 3.3. With the above notation, we have dim N(f )j ≤ 1, for any integer
j ≤ d − 3 + r0 and for any integer j ≥ 2d − 3 − r0 .
Proof. Let t1 = (pk11 − 1)/2 and note that
d
T
= .
2
2
We apply the inequality (3.1) with j = d + 1 − e1 t1 = d + 1 − r0 . It follows that
d − 3 + r0 = d − 3 + e1 t1 < d − 3 +
(3.3)
dim N(f )d−3−r0 ≤ dim H 2 (F, C)λ,
with
2πit1
.
λ = exp(2πir0 /d) = exp
pk11
Since this eigenvalue has order a prime power, it follows from Zariski’s Theorem,
see [1, Proposition 2.1], that H 1 (F, C)λ = 0. Using (3.2), we get dim N(f )j ≤ 1 for
j = d − 3 + r0 . The claim for j = 2d − 3 − r0 follows from the fact that the graded
module N(f ) enjoys a duality property: dim N(f )j = dim N(f )T −j , for any integer
j, see [28, 30]. The Lefschetz type property of the graded module N(f ), see (2.1),
completes the proof of this Proposition.
Proposition 3.4. A rational cuspidal curve C : f = 0 of degree d as in (1.1), and
such that r = mdr(f ) ≤ r0 , is either free or nearly free.
6
ALEXANDRU DIMCA AND GABRIEL STICLARU
Proof. We use the formulas (2.2) and (2.3) from [7] and get the following equality
dim AR(f )d−r−2 − dim N(f )2d−r−3 + dim AR(f )r−2 =
2d − r − 1
d−r
+ τ (C).
−
=3
2
2
For any curve C : f = 0, it is known that, if ρ1 ∈ AR(f )r , then the first relation
ρ ∈ AR(f )m which is not a multiple of ρ1 occurs in a degree m ≥ d − r − 1, see [29,
Lemma 1.1]. It follows that
(d − 2r)(d − 2r − 1)
d − 2r
=
dim AR(f )d−r−2 = dim Sd−2r−2 · ρ1 =
,
2
2
for any r such that 2r ≤ d. Using the obvious fact that AR(f )r−2 = 0, a direct
computation shows that
τ (C) = τ (d, r) − dim N(f )2d−r−3 .
Since r ≤ r0 , it follows that 2d − r − 3 ≥ 2d − r0 − 3, and hence dim N(f )2d−r−3 ≤ 1
by Proposition 3.3. The claim follows now using the characterization of free (resp.
nearly free) curves given above in (2.4).
It remains to prove the last claim in Theorem 1.2. If d = 3pk , we can assume
pk = 2p′ + 1 > 3 and then
d − e1
r0 =
= 3p′ .
2
On the other hand, d = 3pk = 6p′ + 3 = 2(3p′ + 1) + 1, and hence d′ = 3p′ + 1.
Since mdr(f ) ≤ d′ by Theorem 1.1, we get either mdr(f ) = d′ = 3p′ + 1, and then
we conclude by Theorem 1.1, or mdr(f ) ≤ d′ − 1 = 3p′ = r0 , and then we conclude
using Theorem 1.2.
3.5. Proof of Corollary 1.4. To prove the first claim, we have to consider the
minimal possible value of
d − e1
pk 1 − 1
= e1 1
,
2
2
when d is odd, but neither a prime power, nor of the form 3p, with p > 3 prime.
First, if p1 = 3, then k1 ≥ 2 and e1 ≥ 5, hence r0 ≥ 20. Otherwise, p1 ≥ 5, e1 ≥ 3,
but both cannot be equalites. It follows that the minimal values are obtained for
p1 = 52 and e1 = 3, or p1 = 7 and e1 = 5. In the first case we get r0 = 36, in the
second we get r0 = 15. To prove the second claim, just use Remark 1.3.
r0 =
References
[1] E. Artal Bartolo, A. Dimca, On fundamental groups of plane curve complements, Ann. Univ.
Ferrara 61(2015), 255-262. 3.2
[2] E. Artal Bartolo, L. Gorrochategui, I. Luengo, A. Melle-Hernández, On some conjectures
about free and nearly free divisors, in: Singularities and Computer Algebra, Festschrift for
Gert-Martin Greuel on the Occasion of his 70th Birthday, pp. 1–19, Springer (2017) 1, 2
[3] R.-O. Buchweitz, A. Conca, New free divisors from old, J. Commut. Algebra 5 (2013), no. 1,
17–47. 1
ON THE FREENESS OF RATIONAL CUSPIDAL PLANE CURVES
7
[4] W. Decker, G.-M. Greuel, G. Pfister and H. Schönemann. Singular 4-0-1 — A computer algebra system for polynomial computations, available at http://www.singular.uni-kl.de (2014).
2.1
[5] A. Dimca, Singularities and topology of hypersurfaces, Universitext, Springer-Verlag, New York,
1992. 3, 3.1
[6] A. Dimca, Hyperplane Arrangements: An Introduction, Universitext, Springer, 2017 3
[7] A. Dimca, Freeness versus maximal global Tjurina number for plane curves, Math. Proc.
Cambridge Phil. Soc. 163 (2017), 161–172. 2, 2, 3.2
[8] A. Dimca, On rational cuspidal plane curves, and the local cohomology of Jacobian rings,
arXiv:1707.05258. 1, 2
[9] A. Dimca, D. Popescu, Hilbert series and Lefschetz properties of dimension one almost complete
intersections, Comm. Algebra 44 (2016), 4467–4482. 2
[10] A. Dimca, E. Sernesi, Syzygies and logarithmic vector fields along plane curves, Journal de
l’École polytechnique-Mathématiques 1(2014), 247-267. 2
[11] A. Dimca, G. Sticlaru, On the exponents of free and nearly free projective plane curves, Rev.
Mat. Complut. 30(2017), 259–268. 2
[12] A. Dimca, G. Sticlaru, A computational approach to Milnor fiber cohomology, Forum Math.
29 (2017), 831– 846. 1, 3.1
[13] A. Dimca, G. Sticlaru, Free divisors and rational cuspidal plane curves, Math. Res. Lett.
24(2017), 1023–1042. 1, 2, 2
[14] A. Dimca, G. Sticlaru, Free and nearly free curves vs. rational cuspidal plane curves, Publ.
RIMS Kyoto Univ., 54:163–179 (2018). 1, 1, 2, 3
[15] A. Dimca, G. Sticlaru, Computing the monodromy and pole order filtration on Milnor fiber
cohomology of plane curves, arXiv: 1609.06818. 3.1
[16] A.Dimca, G. Sticlaru, Computing Milnor fiber monodromy for projective hypersurfaces,
arXiv:1703.07146. 1, 3.1
[17] J. Fernandez de Bobadilla, I. Luengo, A. Melle-Hernandez, A. Nemethi, Classification of rational unicuspidal projective curves whose singularities have one Puiseux pair, in: Real and
Complex Singularities Sao Carlos 2004, Trends in Mathematics, Birkhauser 2006, pp. 31-45. 1
[18] T. Fenske, Rational 1- and 2-cuspidal plane curves, Beiträge Algebra Geom., 40(1999), 309-329.
1
[19] H. Flenner, M. Zaidenberg, On a class of rational plane curves, Manuscripta Math. 89(1996),
439 - 459. 1
[20] M. Koras, K. Palka, The Coolidge–Nagata conjecture. Duke Math. J. 166 (2017), no. 16,
3085–3145. 1
[21] S. Marchesi, J. Vallès, Nearly free curves and arrangements: a vector bundle point of view,
arXiv:1712.04867. 1
[22] T. K. Moe, Rational Cuspidal Curves, arXiv:1511.02691 (139 pages, Master thesis, University
of Oslo, 2008) 1
[23] K. Palka, T. Pelka, Classification of planar rational cuspidal curves. I. C**-fibrations, Proc.
London Math. Soc. 115 (2017), 638–692. 1
[24] J. Piontkowski, On the number of cusps of rational cuspidal plane curves, Experiment. Math.,
16(2007), 251-255. 1
[25] K. Saito, Theory of logarithmic differential forms and logarithmic vector fields, J. Fac. Sci.
Univ. Tokyo Sect. IA Math. 27 (1980), no. 2, 265-291. 1
[26] M. Saito, Bernstein-Sato polynomials for projective hypersurfaces with weighted homogeneous
isolated singularities, arXiv:1609.04801. 3.1
[27] F. Sakai, K. Tono, Rational cuspidal curves of type (d,d-2) with one or two cusps, Osaka J.
Math. 37(2000), 405-415. 1
[28] E. Sernesi, The local cohomology of the jacobian ring, Documenta Mathematica, 19 (2014),
541-565. 3.2
8
ALEXANDRU DIMCA AND GABRIEL STICLARU
[29] A. Simis, S.O. Tohăneanu, Homology of homogeneous divisors, Israel J. Math. 200 (2014),
449-487. 1, 3.2
[30] D. van Straten, T. Warmt, Gorenstein duality for one-dimensional almost complete
intersections–with an application to non-isolated real singularities, Math. Proc.Cambridge Phil.
Soc.158(2015), 249–268. 3.2
[31] U. Walther, The Jacobian module, the Milnor fiber, and the D-module generated by f s , Invent.
Math. 207 (2017),1239–1287. 1, 3
Université Côte d’Azur, CNRS, LJAD and INRIA, France
E-mail address: [email protected]
Faculty of Mathematics and Informatics, Ovidius University Bd. Mamaia 124,
900527 Constanta, Romania
E-mail address: [email protected]
| 0math.AC
|
Proper affine actions: a sufficient
criterion
arXiv:1612.08942v1 [math.GR] 28 Dec 2016
Ilia Smilga
December 30, 2016
For a semisimple real Lie group G with an irreducible representation ρ on
a finite-dimensional real vector space V , we give a sufficient criterion on ρ
for existence of a group of affine transformations of V whose linear part is
Zariski-dense in ρ(G) and that is free, nonabelian and acts properly discontinuously on V . This new criterion is more general than the one given in
the author’s previous paper Proper affine actions in non-swinging representations (submitted; available at arXiv:1605.03833), insofar as it also deals
with “swinging” representations. We conjecture that it is actually a necessary
and sufficient criterion, applicable to all representations.
1 Introduction
1.1 Background and motivation
The present paper is part of a larger effort to understand discrete groups Γ of affine
transformations (subgroups of the affine group GLn (R) ⋉ Rn ) acting properly discontinuously on the affine space Rn . The case where Γ consists of isometries (in other words,
Γ ⊂ On (R) ⋉ Rn ) is well-understood: a classical theorem by Bieberbach says that such
a group always has an abelian subgroup of finite index.
We say that a group G acts properly discontinuously on a topological space X if for every compact K ⊂ X, the set {g ∈ G | gK ∩ K 6= ∅} is finite. We define a crystallographic
group to be a discrete group Γ ⊂ GLn (R) ⋉ Rn acting properly discontinuously and such
that the quotient space Rn /Γ is compact. In [Aus64], Auslander conjectured that any
crystallographic group is virtually solvable, that is, contains a solvable subgroup of finite
index. Later, Milnor [Mil77] asked whether this statement is actually true for any affine
group acting properly discontinuously. The answer turned out to be negative: Margulis
[Mar83, Mar87] gave a nonabelian free group of affine transformations with linear part
Zariski-dense in SO(2, 1), acting properly discontinuously on R3 . On the other hand,
Fried and Goldman [FG83] proved the Auslander conjecture in dimension 3 (the cases
n = 1 and 2 are easy). Recently, Abels, Margulis and Soifer [AMS] proved it in dimension
n ≤ 6. See [Abe01] for a survey of already known results.
1
Margulis’s breakthrough was soon followed by the construction of other counterexamples to Milnor’s conjecture. Most of these counterexamples have been free groups.
(Only recently, Danciger, Guéritaud and Kassel [DGK] found examples of affine groups
acting properly discontinuously that are neither virtually solvable nor virtually free.)
The author focuses on the case of free groups, asking the following question. Consider
a semisimple real Lie group G; for every representation ρ of G on a finite-dimensional
real vector space V , we may consider the affine group G ⋉ V . Does that affine group
contain a nonabelian free subgroup with linear part Zariski-dense in G and acting properly discontinuously on V ? (More precisely, for which values of G and V is the answer
positive?)
Here is a summary of previous work on this question:
• Margulis’s original work [Mar83, Mar87] gave a positive answer for G = SO+ (2, 1)
acting on V = R3 .
• Abels, Margulis and Soifer generalized this, giving [AMS02] a positive answer for
G = SO+ (2n + 2, 2n + 1) acting on V = R4n+3 (for every integer n ≥ 0).
• They showed later [AMS11] that for all other values of p and q, the answer for
G = SO+ (p, q) acting on V = Rp+q is negative.
• The author of this paper gave [Smi16] a positive answer for any noncompact
semisimple real Lie group G acting (by the adjoint representation) on its Lie algebra g.
• Recently, he gave [Smi] a simple algebraic criterion on G and V guaranteeing that
the answer is positive. However, this criterion included an additional assumption
about V (namely that the representation is “non-swinging”: see Definition 5.13.4),
which is not in fact necessary.
This paper gives a better sufficient condition on G and V for the answer to be positive.
This condition works for all representations, non-swinging and swinging. The author conjectures that the new condition is in fact necessary, thus giving a complete classification
of such counterexamples.
In order to state this condition, we need to introduce a few classical notations.
1.2 Basic notations
For the remainder of the paper, we fix a semisimple real Lie group G; let g be its Lie
algebra. Let us introduce a few classical objects related to g and G (defined for instance
in Knapp’s book [Kna96], though our terminology and notation differ slightly from his).
We choose in g:
• a Cartan involution θ. Then we have the corresponding Cartan decomposition
g = k ⊕ q, where we call k the space of fixed points of θ and q the space of fixed
points of −θ. We call K the maximal compact subgroup with Lie algebra k.
2
• a Cartan subspace a compatible with θ (that is, a maximal abelian subalgebra of g
among those contained in q). We set A := exp a.
• a system Σ+ of positive restricted roots in a∗ . Recall that a restricted root is a
nonzero element α ∈ a∗ such that the restricted root space
gα := {Y ∈ g | ∀X ∈ a, [X, Y ] = α(X)Y }
is nontrivial. They form a root system Σ; a system of positive roots Σ+ is a subset
of Σ contained in a half-space and such that Σ = Σ+ ⊔ −Σ+ . (Note that in contrast
to the situation with ordinary roots, the root system Σ need not be reduced; so in
addition to the usual types, it can also be of type BCn .)
We call Π be the set of simple restricted roots in Σ+ . We call
a++ := X ∈ a ∀α ∈ Σ+ , α(X) > 0
the (open) dominant Weyl chamber of a corresponding to Σ+ , and
a+ := X ∈ a ∀α ∈ Σ+ , α(X) ≥ 0 = a++
the closed dominant Weyl chamber.
Then we call:
• M the centralizer of a in K, m its Lie algebra.
• L the centralizer of a in G, l its Lie algebra. It is clear that l = a ⊕ m, and well
known (see e.g. [Kna96], Proposition 7.82a) that L = M A.
• n+ (resp. n− ) the sum of the restricted root spaces of Σ+ (resp. of −Σ+ ), and
N + := exp(n+ ) and N − := exp(n− ) the corresponding Lie groups.
• p+ := l ⊕ n+ and p− := l ⊕ n− the corresponding minimal parabolic subalgebras,
P + := LN + and P − := LN − the corresponding minimal parabolic subgroups.
• W := NG (A)/ZG (A) the restricted Weyl group.
• w0 the longest element of the Weyl group, that is, the unique element such that
w0 (Σ+ ) = Σ− . It is clearly an involution.
See Examples 2.3 and 2.4 in the author’s previous paper [Smi16] for working through
those definitions in the cases G = PSLn (R) and G = PSO+ (n, 1).
Finally, if ρ is a representation of G on a finite-dimensional real vector space V , we
call:
• the restricted weight space in V corresponding to a form λ ∈ a∗ the space
V λ := {v ∈ V | ∀X ∈ a, X · v = λ(X)v} ;
3
• a restricted weight of the representation ρ any form λ ∈ a∗ such that the corresponding weight space is nonzero.
Remark 1.1. The reader who is unfamiliar with the theory of noncompact semisimple
real Lie groups may focus on the case where G is split, i.e. its Cartan subspace a is
actually a Cartan subalgebra (just a maximal abelian subalgebra, without any additional
hypotheses). In that case the restricted roots are just roots, the restricted weights are just
weights, and the restricted Weyl group is just the usual Weyl group. Also the algebra m
vanishes and M is a discrete group.
The case where G is split does not actually require the full strength of this paper, in
particular because quasi-translations (see Section 7.2) then reduce to ordinary translations.
1.3 Statement of main result
Let ρ be any representation of G on a finite-dimensional real vector space V . Without
loss of generality, we may assume that G is connected and acts faithfully. We may then
identify the abstract group G with the linear group ρ(G) ⊂ GL(V ). Let VAff be the affine
space corresponding to V . The group of affine transformations of VAff whose linear part
lies in G may then be written G ⋉ V (where V stands for the group of translations). Here
is the main result of this paper.
Main Theorem. Suppose that ρ satisfies the following conditions:
(i) there exists a vector v ∈ V such that:
(a) ∀l ∈ L, l(v) = v, and
(b) w̃0 (v) 6= v, where w̃0 is any representative in G of w0 ∈ NG (A)/ZG (A);
Then there exists a subgroup Γ in the affine group G ⋉ V whose linear part is Zariskidense in G and that is free, nonabelian and acts properly discontinuously on the affine
space corresponding to V .
(Note that the choice of the representative w̃0 in (i)(b) does not matter, precisely
because by (i)(a) the vector v is fixed by L = ZG (A).)
Remark 1.2. It is sufficient to prove the theorem in the case where ρ is irreducible.
Indeed, we may decompose ρ into a direct sum of irreducible representations, and then
observe that:
• if some representation ρ1 ⊕ · · ·⊕ ρk has a vector (v1 , . . . , vk ) that satisfies conditions
(a) and (b), then at least one of the vectors vi must satisfy conditions (a) and (b);
• if V = V1 ⊕ V2 , and a subgroup Γ ⊂ G ⋉ V1 acts properly on V1 , then its image i(Γ)
by the canonical inclusion i : G ⋉ V1 → G ⋉ V still acts properly on V .
We shall start working with an arbitrary representation ρ, and gradually make stronger
and stronger hypotheses on it, introducing each one when we need it to make the construction work (so that it is at least partially motivated). Here is the complete list of
places where new assumptions on ρ are introduced:
4
• Assumption 5.5, which is a necessary condition for (i)(a);
• Assumption 7.12, which is (i)(a);
• Assumption 11.1, which is (i).
Here are a few examples. Items 1 and 2 show that all the previously-known examples do
fall under the scope of this theorem. Item 3 is the simplest previously-unknown example
that our paper brings to light:
Example 1.3.
1. For G = SO+ (2n + 2, 2n + 1), the standard representation (acting on V = R4n+3 )
satisfies these conditions (see Remark 3.9 and Examples 4.22.1.b and 10.2.1 in [Smi]
for details). So Theorem A from [AMS02] is a particular case of this theorem.
2. If the real semisimple Lie group G is noncompact, the adjoint representation satisfies these conditions (see Remark 3.9 and Examples 4.22.3 and 10.2.2 in [Smi] for
details). So the main theorem of [Smi16] is a particular case of this theorem.
3. Take G = SL3 (R) acting on V = S 3 R3 (see Example 3.7 in [Smi] for details). The
group is then split, so that l = a and the set of L-invariant vectors is precisely
the zero (restricted) weight space V 0 . The representation V has dimension 10; the
zero weight space has dimension 1: it is spanned by the vector e1 · e2 · e3 , where
(e1 , e2 , e3 ) is the canonical basis of R3 . A representative w˜0 ∈ SL3 (R) of w0 is given
by
0 0 −1
w̃0 = 0 1 0 .
1 0 0
Clearly this element acts nontrivially on the vector e1 · e2 · e3 .
Remark 1.4. When G is compact, no representation can satisfy these conditions: indeed
in that case L is the whole group G and condition (i)(a) fails. So for us, only noncompact
groups are interesting. This is not surprising: indeed, any compact group acting on a
vector space preserves a positive-definite quadratic form, and so falls under the scope of
Bieberbach’s theorem.
1.4 Strategy of the proof
A central pillar of this paper consists in the following “proposition template”:
Schema 1.5. Let g and h be two elements of “the appropriate group”, such that:
• both g and h are “regular”;
• the pair (g, h) has a “non-degenerate” geometry;
• both g and h have sufficient “contraction strength”.
5
In that case:
(i) the product gh is still “regular”;
(ii) its “attracting geometry” is close to that of g;
(iii) its “repelling geometry” is close to that of h;
(iv) its “contraction strength” is close to the product of those of g and h;
(v) its “asymptotic dynamics” is close to the sum of those of g and h.
We prove three different versions of this statement (with some slight variations, especially concerning asymptotic dynamics), all of which involve a different set of definitions
for the concepts in scare quotes (see Table 1):
• the “proximal” version is Proposition 3.5;
• the “linear” version is Proposition 4.22 (for the “main part”, i.e. points (i) to (iv) in
Schema 1.5) and Proposition 4.24 (for the asymptotic dynamics);
• the “affine” version is Proposition 8.4 (ii) and (iii) (for the main part) and Proposition 9.3 (for the asymptotic dynamics).
For the last two versions, the definitions in question also depend on a parameter fixed
once and for all (given in the second line of Table 1). To give a first intuition, very
roughly, the “geometry” of an element has to do with its eigenvectors, its “contraction
strength” has to do with its singular values, and its “asymptotic dynamics” has to do
with the moduli of its eigenvalues1 (the word “asymptotic” is explained by Gelfand’s
formula (7.13)).
Here is the relationship between these three results:
• The “proximal” version is used as a stepping stone to prove the other two versions. This result is not new: very similar lemmas were proved in [Ben96], [Ben97]
and [AMS02], and this exact statement was proved in the author’s previous papers [Smi16, Smi]. We briefly recall it in Section 3.
• The “linear” version is also used as a stepping stone to prove the affine version,
but not in a completely straightforward way. In fact, the main part of the linear
version, after being reformulated as Proposition 8.4 (i), is only involved in the proof
of the additivity of the Margulis invariant (i.e. the affine version for the asymptotic
dynamics). On the other hand, the linear version for the asymptotic dynamics
is already necessary to prove the main part of the affine version. It is a slight
generalization of a result by Benoist [Ben96]. We cover it in Section 4.
1
In the “affine” case, this is not exactly true: the translation part is also involved.
6
Group where g lives
Parameter
“Regularity”
attracting
“Geometry”:
Proximal case
Linear case
Affine case
GL (E)
(for some real
vector space E)
G
G⋉V
(none)
some subset ΠX
of simple restr.
roots (sec. 2.3)
some extreme,
generically
symmetric
X0 ∈ a+ (sec. 5)
Proximality
(Def. 3.1)
X-regularity
(Def. 4.1)
ρ-regularity
(Def. 6.12)
Egs ∈ P(E)
ygX,+ ∈ G/PX+
Ag , essentiallya ∈
G ⋉ V /PX+0 ⋉ V0≥
(Def. 3.2)
(Def. 4.3)
&
(Def. 7.5)
.
Egu ∈ P(E ∗ )
ygX,− ∈ G/PX−
Ag , essentiallya ∈
G ⋉ V /PX−0 ⋉ V0≤
“Non-degeneracy”
Def. 3.3
Def. 4.10
Def. 7.20
“Contr. strength”
s̃(g) (Def. 3.4)
~
sX (g) (Def. 4.13)
sX0 (g) (Def. 7.22)
repelling
possiblyb
“Asymp. dynamics”
a
b
c
the log
of the spectral
radius r(g)
(Def. 3.1)
roughlyc
the
Jordan
projection Jd(g)
(Def. 2.3)
roughlyc the
Jordan projection
Jd(g), AND the
Margulis
invariant M (g)
(Def. 7.19)
See the last point of Remark 7.15.
In the proximal case, there are at least two reasonable definitions of asymptotic dynamics. Another
|λ1 |
one would be the logarithm of the spectral gap |λ
, defined in the same place.
2|
Technically, the statement of Schema 1.5 only holds for a further projection of Jd(g), given by the
subset of coordinates αi (Jd(g)) for i 6∈ ΠX ; for the remaining coordinates we only get an inequality.
However Proposition 8.1 allows us to circumvent this problem. Also if ΠX0 = ∅, which is often the
case (see Example 5.14.3 and Remark 5.15.(ii)), this issue does not arise at all.
Table 1: Possible meanings of notions appearing in Schema 1.5, with references to the
corresponding definitions.
7
• The “affine” version is the key result of this paper. Just defining the affine concepts
keeps us busy for a long time (Sections 5–7), and proving the results also takes a
fair amount of work (Sections 8 and 9). Once we have it, it takes only five pages
to prove our Main Theorem.
The proof has a lot in common with the author’s previous papers [Smi16, Smi], and
ultimately builds upon an idea that was introduced in Margulis’s seminal paper [Mar83].
Let us now present a few highlights of what distinguishes this paper from the previous
works.
• The main difference lies in the treatment of the dynamical spaces. As long as
we only worked in “non-swinging” representations (see Definition 5.13.4), we could
associate to every element of G ⋉ V that is “regular” (in the appropriate sense) a
decomposition of V into three so-called dynamical subspaces
V = Vg> ⊕ Vg= ⊕ Vg< ,
all stable by g and on which g acts with eigenvalues respectively of modulus > 1,
of modulus 1 and of modulus < 1. In the general case, this is no longer possible: we need to enlarge the neutral subspace Vg= to an “approximately neutral”
subspace Vg≈ . The eigenvalues of g on that subspace are, in some rather weak
sense, “not too far” from 1, but will still grow exponentially (in the group we will
eventually construct). The decomposition now becomes
V = Vg≫ ⊕ Vg≈ ⊕ Vg≪ .
This forces us to completely change our point of view. We now define these “approximate dynamical spaces” in a purely algebraic way, by focusing on the dynamics
that g would have on them if it were conjugate to (the exponential of) some reference element X0 of the Weyl chamber a+ . For this reason, we actually call them
the ideal dynamical spaces (see Definition 7.5). Only by imposing an additional
condition on g (“asymptotic contraction”, see Definition 6.12) can we ensure that
the “ideal” dynamical spaces become indeed the “approximate” dynamical spaces.
• The linear version of Schema 1.5 never explicitly appeared in the author’s previous
papers: so far, we have been able to simply present the linear theory as a particular case of the affine theory. (Even Proposition 6.10 in [Smi], which seems almost
identical to Proposition 4.24 in the current paper, relies in fact on the affine versions of the properties). However this becomes untenable in the case of swinging
representations, as the relationship between affine contraction strength and linear
contraction strength becomes less straightforward (see Section 7.5). This led us
to develop the linear theory on its own; we think that Propositions 4.22 and 4.24
might have some interest independently of the remainder of the paper.
We must however point out that these results are not completely new. The particular case where ΠX = ∅ is due to Benoist [Ben96, Ben97]; the case where ΠX
8
is arbitrary is an easy generalization which also relies on the tools developed by
Benoist, and is well-known to experts in the field. It might seem at first sight that
the general case is actually mentioned explicitly in [Ben97], but this is not so: see
Remark 6.13 in [Smi].
• The central argument of the paper, namely the proof of Proposition 9.3, has been
completely overhauled. Though still highly technical, it is now much cleaner: it is
more symmetric, and the organization of the proof better reflects the separation
of the ideas it involves. Even if we forget the fact that it works in a more general
setting, it is definitely an improvement compared to the proof given in [Smi].
1.5 Plan of the paper
In Section 2, we give a few algebraic results and introduce some notations related to
metric properties and estimates. Most of these results are well-known.
In Section 3, we recall the definitions of the proximal versions of the key properties
and the proximal version of Schema 1.5. This is also well-known.
In Section 4, we define the linear versions of the key properties, then prove the linear version of Schema 1.5. This section clarifies and expands the results mentioned
in Section 6 of [Smi]. Ultimately, most of the definitions and ideas here are due to
Benoist [Ben96, Ben97].
In Section 5, we choose an element X0 ∈ a+ that will be used to define the affine
versions of the key properties. This generalizes Section 3 in [Smi].
In Section 6, we present some preliminary constructions involving X0 , and introduce
some (elementary) formalism that expresses affine spaces in terms of vector spaces. Part
of the material is borrowed from Sections 4.1 and 4.2 in [Smi].
In Section 7, we define the affine versions of the key properties. This generalizes
Sections 4.3–5.2 in [Smi], but introduces several new ideas.
In Section 8, we prove the main part of the affine version of Schema 1.5. This generalizes
Section 7 in [Smi].
Section 9 contains the key part of the proof. We prove the asymptotic dynamics part of
the affine version of Schema 1.5 — in other terms, approximate additivity of the Margulis
invariants. This generalizes Section 8 in [Smi] and Section 4 in [Smi16]; the proof is based
on the same idea, but has been considerably rewritten.
The very short Section 10 uses induction to extend the results of the two previous
sections to products of an arbitrary number of elements. It is a straightforward generalization of Section 9 in [Smi] and Section 5 in [Smi16]
Section 11 contains the proof of the Main Theorem. It is an almost straightforward
generalization of Section 10 in [Smi] and Section 6 in [Smi16].
1.6 Acknowledgments
I am very grateful to my Ph.D. advisor, Yves Benoist, who introduced me to this exciting
and fruitful subject, and gave invaluable help and guidance in my initial work on this
project.
9
I would also like to thank Bruno Le Floch for some interesting discussions, which in
particular helped me gain more insight about weights of representations.
2 Preliminaries
In Subsection 2.1, for any element g ∈ G, we give a formula for the eigenvalues and
singular values of the linear maps ρ(g) where ρ is an arbitrary representation of G. This
is nothing more than a reminder of Subsection 2.1 in [Smi].
In Subsection 2.2, we present some properties of restricted weights of a real finitedimensional representation of a real semisimple Lie group. This is mostly a reminder of
Subsection 2.2 in [Smi].
In Subsection 2.3, we give some basic results from the theory of parabolic subgroups
and subalgebras (not necessarily minimal). This is a development on Subsection 3.4
in [Smi].
In Subsection 2.4, we give some notation conventions related to metric properties and
estimates (mostly borrowed from the author’s earlier papers).
2.1 Eigenvalues in different representations
In this subsection, we express the eigenvalues and singular values of a given element g ∈ G
acting in a given representation ρ, exclusively in terms of the structure of g in the abstract
group G (respectively its Jordan decomposition and its Cartan decomposition). This is
just a reminder of Subsection 2.1 in [Smi]; the result itself was certainly well-known even
before that.
Proposition 2.1 (Jordan decomposition). Let g ∈ G. There exists a unique decomposition of g as a product g = gh ge gu , where:
• gh is conjugate in G to an element of A ( hyperbolic);
• ge is conjugate in G to an element of K ( elliptic);
• gu is conjugate in G to an element of N + ( unipotent);
• these three maps commute with each other.
Proof. This is well-known, and given for example in [Ebe96], Theorem 2.19.24. Note
however that the latter theorem uses representation-dependent definitions of a hyperbolic,
elliptic or unipotent element (applied to the case of the adjoint representation). To state
the theorem with the representation-agnostic definitions that we used, we need to apply
Proposition 2.19.18 and Theorem 2.19.16 from the same book.
Proposition 2.2 (Cartan decomposition). Let g ∈ G. Then there exists a decomposition
of g as a product g = k1 ak2 , with k1 , k2 ∈ K and a = exp(X) with X ∈ a+ . Moreover,
the element X is uniquely determined by g.
Proof. This is a classical result; see e.g. Theorem 7.39 in [Kna96].
10
Definition 2.3. For every element g ∈ G, we define:
• the Jordan projection of g (sometimes also known as the Lyapunov projection),
written Jd(g), to be the unique element of the closed dominant Weyl chamber a+
such that the hyperbolic part gh (from the Jordan decomposition g = gh ge gu given
above) is conjugate to exp(Jd(g));
• the Cartan projection of g, written Ct(g), to be the element X from the Cartan
decomposition given above.
To talk about singular values, we need to introduce a Euclidean structure. We are
going to use a special one.
Lemma 2.4. Let ρ∗ be some real representation of G on some space V∗ . There exists a
K-invariant positive-definite quadratic form B∗ on V∗ such that all the restricted weight
spaces are pairwise B∗ -orthogonal.
We want to reserve the plain notation ρ for the “default” representation, to be fixed
once and for all at the beginning of Section 5. We use the notation ρ∗ so as to encompass
both this representation ρ and the representations ρi defined in Proposition 2.12.
Proof. See Lemma 5.33.a) in [BQ] (for a more succinct proof) or Lemma 2.4 in [Smi] (for
a more detailed proof).
Example 2.5. If ρ∗ = Ad is the adjoint representation, then B∗ is the form Bθ given by
∀X, Y ∈ g,
Bθ (X, Y ) = −B(X, θY )
(see (6.13) in [Kna96]), where B is the Killing form and θ is the Cartan involution.
Recall that the singular values of a map g in a Euclidean space are defined as the square
roots of the eigenvalues of g∗ g, where g∗ is the adjoint map. The largest and smallest
singular values of g then give respectively the operator norm of g and the reciprocal of
the operator norm of g−1 .
Proposition 2.6. Let ρ∗ : G → GL(V∗ ) be any representation of G on some vector
space V∗ ; let λ1∗ , . . . , λd∗∗ be the list of all the restricted weights of ρ∗ , repeated according
to their multiplicities. Let g ∈ G; then:
(i) The list of the moduli of the eigenvalues of ρ∗ (g) is given by
i
eλ∗ (Jd(g))
.
1≤i≤d∗
(ii) The list of the singular values of ρ∗ (g), with respect to a K-invariant Euclidean
norm B∗ on V∗ that makes the restricted weight spaces of V∗ pairwise orthogonal
(such a norm exists by Lemma 2.4), is given by
i
eλ∗ (Ct(g))
.
1≤i≤d∗
Proof. This is also completely straightforward. See Proposition 2.6 in [Smi].
11
2.2 Properties of restricted weights
In this subsection, we introduce a few properties of restricted weights of real finitedimensional representations. (Proposition 2.7 is actually a general result about Coxeter
groups.) The corresponding theory for ordinary weights is well-known: see for example
Chapter V in [Kna96]. This is mostly a reminder of Subsection 2.2 from [Smi]; the only
addition is Lemma 2.15.
Let α1 , . . . , αr be an enumeration of the set Π of simple restricted roots generating Σ+ .
For every i, we set
(
2αi if 2αi is a restricted root
′
(2.1)
αi :=
αi
otherwise.
For every index i such that 1 ≤ i ≤ r, we define the i-th fundamental restricted weight ̟i
by the relationship
h̟i , α′j i
2
= δij
(2.2)
kα′j k2
for every j such that 1 ≤ j ≤ r.
By abuse of notation, we will often allow ourselves to write things such as “for all i in
some subset Π′ ⊂ Π, ̟i satisfies...” (tacitly identifying the set Π′ with the set of indices
of the simple restricted roots that are inside).
In the following proposition, for any subset Π′ ⊂ Π, we denote:
• by WΠ′ the Weyl subgroup of type Π′ :
WΠ′ := hsα iα∈Π′ ;
• by a+
Π′ the fundamental domain for the action of WΠ′ on a:
′
a+
Π′ := X ∈ a ∀α ∈ Π , α(X) ≥ 0 ,
(2.3)
(2.4)
which is a kind of prism whose base is the dominant Weyl chamber of WΠ′ .
Proposition 2.7. Take any Π′ ⊂ Π, and let us fix X ∈ a+
Π′ . Let Y ∈ a. Then the
following two conditions are equivalent:
(i) the vector Y is in a+
Π′ and satisfies the system of linear inequalities
(
∀i ∈ Π′ ,
̟i (Y ) ≤ ̟i (X),
′
∀i ∈ Π \ Π , ̟i (Y ) = ̟i (X);
′
(ii) the vector Y is in a+
Π′ and also in the convex hull of the orbit of X by WΠ .
Proof. For the particular case Π′ = Π, this is well known: see e.g. [Hal15], Proposition 8.44.
The general case can easily be reduced to this particular case: see [Smi], Proposition 2.7.
12
Proposition 2.8. Every restricted weight of every representation of g is a linear combination of fundamental restricted weights with integer coefficients.
Proof. This is a particular case of Proposition 5.8 in [BT65]. For a correction concerning
the proof, see also Remark 5.2 in [BT72].
Proposition 2.9. If ρ∗ is an irreducible representation of g, there is a unique restricted
weight λ∗ of ρ∗ , called its highest restricted weight, such that no element of the form
λ∗ + αi with 1 ≤ i ≤ r is a restricted weight of ρ∗ .
Remark 2.10. In contrast to the situation with non-restricted weights, the highest restricted weight is not always of multiplicity 1; nor is a representation uniquely determined
by its highest restricted weight.
Proof. The corresponding result for ordinary (non-restricted) weights is well-known: see
e.g. [Kna96], Theorem 5.5 (d).
For result for restricted weights can easily be deduced from the former: see Proposition 2.9 in [Smi].
Proposition 2.11. Let ρ∗ be an irreducible representation of g; let λ∗ be its highest
restricted weight. Let Λλ∗ be the restricted root lattice shifted by λ∗ :
Λλ∗ := {λ∗ + c1 α1 + · · · + cr αr | c1 , . . . , cr ∈ Z} .
Then the set of restricted weights of ρ∗ is exactly the intersection of the lattice Λλ∗ with
the convex hull of the orbit {w(λ∗ ) | w ∈ W } of λ∗ by the restricted Weyl group.
Proof. Once again, this follows from the corresponding result for non restricted weights
(see e.g. [Hal15], Theorem 10.1) by passing to the restriction. In the case of restricted
weights, one of the inclusions is stated in [Hel08], Proposition 4.22.
Proposition 2.12. For every index i such that 1 ≤ i ≤ r, there exists an irreducible
representation ρi of G on a space Vi whose highest restricted weight is equal to ni ̟i (for
some positive integer ni ) and has multiplicity 1.
Proof. This follows from the general Theorem 7.2 in [Tit71]. This is also stated as
Lemma 2.5.1 in [Ben97].
Example 2.13.
1. If G = SLn (R), we may take Vi = Λi Rn , so that ρi is the i-th exterior power of the
standard representation of G on Rn .
2. More generally, if G is split, then all restricted weight spaces correspond to ordinary
weight spaces, hence have dimension 1. So we may simply take every ni to be 1,
so that the ρi ’s are precisely the fundamental representations.
13
Lemma 2.14. Fix an index i such that 1 ≤ i ≤ r. Then all restricted weights of ρi other
than ni ̟i have the form
r
X
cj αj ,
ni ̟i − αi −
j=1
with cj ≥ 0 for every j.
Proof. This is the last remark in section 2.5 of [Ben97]. For a proof, see Lemma 2.13
in [Smi].
Lemma 2.15. Assume that the Lie algebra g is simple, and that its restricted root system
has a simply-laced diagram. Let ρ∗ be an irreducible representation of G; let Ω∗ be the
set of its restricted weights. Assume that we have
{0} ( Ω∗
(i.e. the representation has 0 as a restricted weight, and at least one nonzero restricted
weight). Then we have Σ ⊂ Ω∗ .
Note that the only case where 0 is the only restricted weight is when the image of G
by ρ is a compact group (for simple G, this means that either G itself is compact or ρ is
trivial.)
Proof. First let us show that Ω∗ contains at least one restricted root. By Proposition 2.11,
every restricted weight λ of ρ∗ may be written as a sum
λ = α(1) + · · · + α(l) ,
(2.5)
where α(1) , . . . , α(l) are some restricted roots. Define the level of λ to be the smallest
integer l for which such a decomposition is possible.
Let λ be an element of Ω∗ \ {0} whose level l is the smallest possible (it exists since
Ω∗ is not reduced to {0}), and consider its decomposition (2.5). Then for every i 6= l, we
have hα(i) , α(l) i ≥ 0: indeed otherwise, α(i) + α(l) would be a restricted root, so we could
combine them together to produce a decomposition of λ length l − 1. Hence we have
l
X
2hα(i) , α(l) i
2hα(l) , α(l) i
2hλ, α(l) i
=
≥
= 2,
hα(l) , α(l) i
hα(l) , α(l) i
hα(l) , α(l) i
i=1
which implies that
λ − α(l) ∈ Conv({λ, sα(l) λ}) ⊂ Conv(W λ).
From Proposition 2.11, it follows that λ − α(l) is still a restricted weight of ρ∗ ; but its
level is now l − 1. Since l is by assumption the smallest nonzero level, necessarily we have
l = 1, and λ is indeed a restricted root.
Now it is well-known (see Problem 2.11 in [Kna96]) that for a simple Lie algebra,
W acts transitively on the set of restricted roots having the same length as λ. Since
the restricted root system of g has a simply-laced diagram, all of the restricted roots
have the same length. Hence the orbit of λ by W is the whole set Σ; we conclude that
Σ ⊂ Ω∗ .
14
2.3 Parabolic subgroups and subalgebras
In this subsection we recall the well-known theory of parabolic subalgebras and subgroups. We begin by defining them, as well as the Levi subalgebra and subgroup of
a given type and the corresponding subset of Π and subgroup of W ; so far we follow
Subsection 3.4 in [Smi]. But now we go further, giving a few propositions relating these
different objects; in particular the generalized Bruhat decomposition (Lemma 2.23).
A parabolic subgroup (or subalgebra) is usually defined in terms of a subset Π′ of the
set Π of simple restricted roots. We find it more convenient however to use a slightly
different language. To every such subset corresponds a facet of the Weyl chamber, given
by intersecting the walls corresponding to elements of Π′ . We may exemplify this facet
by picking some element X in it that does not belong to any subfacet. Conversely, for
every X ∈ a+ , we define the corresponding subset
ΠX := {α ∈ Π | α(X) = 0} .
(2.6)
The parabolic subalgebras and subgroups of type ΠX can then be very conveniently
rewritten in terms of X, as follows.
Definition 2.16. For every X ∈ a+ , we define:
−
• p+
X and pX the parabolic subalgebras of type X, and lX the Levi subalgebra of
type X:
M
p+
gα ;
X := l ⊕
α(X)≥0
p−
X := l ⊕
M
gα ;
α(X)≤0
lX := l ⊕
M
gα .
α(X)=0
• PX+ and PX− the corresponding parabolic subgroups, and LX the Levi subgroup of
type X:
PX+ := NG (p+
X );
PX− := NG (p−
X );
LX := PX+ ∩ PX− .
The following statement is well-known:
Proposition 2.17. We have LX = ZG (X).
Proof. First note that by combining Propositions 7.83 (b), (e) and 7.82 (a) in [Kna96],
we get that LX = ZG (aX ), where
aX := {Y ∈ a | ∀α ∈ ΠX , α(Y ) = 0}
15
is the intersection of all walls of the Weyl chamber containing X. It remains to show
that
ZG (aX ) = ZG (X).
Clearly the Lie algebra of both groups is equal to lX ; hence their identity components
ZG (aX )e and ZG (X)e are also equal. But by combining Propositions 7.25 and 7.33
in [Kna96], it follows that ZG (X) = M ZG (X)e and similarly for ZG (aX ). The conclusion
follows.
An object closely related to these parabolic subgroups (see Corollary 2.24, the Bruhat
decomposition for parabolic subgroups) is the stabilizer of X in the Weyl group:
Definition 2.18. For any X ∈ a+ , we set
WX := {w ∈ W | wX = X} .
Remark 2.19. The group WX is also closely related to the set ΠX . Indeed, it follows
immediately that a simple restricted root α belongs to ΠX if and only if the corresponding
reflection sα belongs to WX . Conversely, it is well-known (Chevalley’s lemma, see e.g.
[Kna96], Proposition 2.72) that these reflections actually generate the group WX .
Thus WX is actually the same thing as WΠ′ (as defined before Proposition 2.7) where
we substitute Π′ = ΠX .
Example 2.20. To help understand the conventions we are taking, here are the extreme
cases:
1. If X lies in the open Weyl chamber a++ , then:
• PX+ = P + is the minimal parabolic subgroup; PX− = P − ; LX = L;
• ΠX = ∅;
• WX = {Id}.
2. If X = 0, then:
• PX+ = PX− = LX = G;
• ΠX = Π;
• WX = W .
The following result shows why WX is important.
Definition 2.21. Let Ω be a set of elements of a∗ . We say that Ξ is:
• a top-subset of Ω if
∀α ∈ Σ+ ,
(Ξ + α) ∩ Ω ⊂ Ξ;
• a bottom-subset of Ω if
∀α ∈ Σ+ ,
(Ξ − α) ∩ Ω ⊂ Ξ.
Example 2.22. In the set of restricted roots Σ, the positive restricted roots form a
top-subset and the negative restricted roots form a bottom-subset.
16
Lemma 2.23 (Generalized Bruhat decomposition). Let ρ∗ be some real representation
of G on some space V∗ ; let Ω∗ be the set of restricted weights of ρ∗ . For every subset
Ξ ⊂ Ω∗ , we set
M
V∗Ξ :=
V∗λ .
(2.7)
λ∈Ξ
Then we have:
(
StabG V∗Ξ = P + StabW (Ξ)P +
StabG V∗Ξ = P − StabW (Ξ)P −
if Ξ is a top-subset of Ω∗ ;
if Ξ is a bottom-subset of Ω∗ .
Proof. Assume that Ξ is a top-subset; the case of a bottom-subset is analogous.
The first step is to show that StabG (V∗Ξ ) contains P + . Indeed:
• The group L stabilizes every restricted weight space V∗λ . Indeed, take some λ ∈ a∗ ,
v ∈ V∗λ , l ∈ L, X ∈ a; then we have:
X · l(v) = l(Ad(l−1 )(X) · v) = l(X · v) = λ(X)l(v).
(2.8)
• For every α ∈ Σ+ and every λ ∈ Ξ, clearly we have
gα · V∗λ ⊂ V∗λ+α ⊂ V∗Ξ
by definition of a top-subset. Hence n+ stabilizes V∗Ξ .
• The statement follows as P + = L exp(n+ ).
Now take any element g ∈ G. Let us apply the Bruhat decomposition (see e.g. [Kna96],
Theorem 7.40): there exists an element w of the restricted Weyl group W such that we
may write
g = p1 w̃p2
where p1 , p2 are some elements of the minimal parabolic subgroup P + and w̃ ∈ NG (A)
is some representative of w ∈ W = NG (A)/ZG (A). From the statement that we just
proved, it follows that g stabilizes V∗Ξ iff w̃ does so. On the other hand, it is clear that
for every λ ∈ a∗ , we have
w̃V λ = V wλ
(2.9)
(the choice of the representative w̃ does not matter since as seen above (2.8), the kernel ZG (A) = L stabilizes V λ ). The conclusion follows.
The following particular case is well-known (see for example [Hum75], Theorem 29.3):
Corollary 2.24 (Bruhat decomposition for parabolic groups). We have the identities
PX+ = P + WX P + and PX− = P − WX P − .
Proof. Take ρ∗ to be the adjoint representation: then V∗ = g and Ω∗ = Σ ∪ {0}. Take
Ξ+ = {α ∈ Σ ∪ {0} | α(X) ≥ 0} ;
this is a top-subset. It is easy to show that we have StabW (Ξ+ ) = WX , and by defini+
tion gΞ = p+
X . Applying the lemma, the first identity follows. Applying the lemma to
the subset Ξ− defined analogously, the second identity follows.
17
2.4 Metric properties and estimates
In this subsection we mostly introduce some notational conventions. They were already
introduced in the beginning of Subsection 5.1 in [Smi] and the beginning of Subsection 2.6
in [Smi16].
For any linear map g acting on a Euclidean space E, we write kgk := supx6=0 kg(x)k
kxk its
operator norm.
Consider a Euclidean space E. We introduce on the projective space P(E) a metric
by setting, for every x, y ∈ P(E),
α(x, y) := arccos
|hx, yi|
∈ [0, π2 ],
kxkkyk
(2.10)
where x and y are any vectors representing respectively x and y (obviously, the value
does not depend on the choice of x and y). This measures the angle between the lines x
and y. For shortness’ sake, we will usually simply write α(x, y) with x and y some actual
vectors in E \ {0}.
For any vector subspace F ⊂ E and any radius ε > 0, we shall denote the εneighborhood of F in P(E) by:
BP (F, ε) := {x ∈ P(E) | α(x, P(F )) < ε} .
(2.11)
(You may think of it as a kind of “conical neighborhood”.)
Consider a metric space (M, δ); let X and Y be two subsets of M. We shall denote
the ordinary, minimum distance between X and Y by
(2.12)
δ(X, Y ) := inf inf δ(x, y),
x∈X y∈Y
as opposed to the Hausdorff distance, which we shall denote by
δ
Haus
(X, Y ) := max
sup δ {x}, Y , sup δ {y}, X
x∈X
y∈Y
!
.
(2.13)
Finally, we introduce the following notation. Let X and Y be two positive quantities,
and p1 , . . . , pk some parameters. Whenever we write
X .p1 ,...,pk Y,
we mean that there is a constant K, depending on nothing but p1 , . . . , pk , such that
X ≤ KY . (If we do not write any subscripts, this means of course that K is an “absolute”
constant — or at least, that it does not depend on any “local” parameters; we consider
the “global” parameters such as the choice of G and of the Euclidean norms to be fixed
once and for all.) Whenever we write
X ≍p1 ,...,pk Y,
we mean that X .p1 ,...,pk Y and Y .p1 ,...,pk X at the same time.
The following result will often be useful:
18
Lemma 2.25 ([Smi16]). Let C ≥ 1. Then any map φ ∈ GL(E) such that kφ±1 k ≤ C
induces a C 2 -Lipschitz continuous map on P(E).
Proof. See [Smi16], Lemma 2.20.
3 Proximal maps
In this section, we give the definitions of the proximal versions of the concepts from
Table 1 and state the proximal version of Schema 1.5, namely Proposition 3.5. It contains
no new results.
Let E be a Euclidean space.
Definition 3.1 (Proximal version of regularity). Let γ ∈ GL(E); let λ1 , . . . , λn be its
eigenvalues repeated according to multiplicity and ordered by nonincreasing modulus.
We define the spectral radius of γ as
r(γ) := |λ1 |
(3.1)
(we do not use “ρ(γ)” as it could be confused with the representation ρ). We say that
γ is proximal if we have
|λ2 | < |λ1 | =: r(γ),
i.e. if λ1 is the only eigenvalue with modulus r(γ) and has multiplicity 1. Equivalently,
|λ1 |
is greater than 1.
γ is proximal if and only if its spectral gap |λ
2|
Definition 3.2 (Proximal version of geometry). For every proximal map γ, we may
then decompose E into a direct sum of a line Eγs , called the attracting space of γ, and a
hyperplane Eγu , called the repelling space of γ, both stable by γ and such that:
(
γ|Eγs = λ1 Id;
for every eigenvalue λ of γ|Eγu , |λ| < |λ1 |.
Definition 3.3 (Proximal version of non-degeneracy). Consider a line E s and a hyperplane E u of E, transverse to each other. An optimal canonizing map for the pair (E s , E u )
is a map φ ∈ GL(E) satisfying
φ(E s ) ⊥ φ(E u )
and minimizing the quantity max kφk, kφ−1 k .
We define an optimal canonizing map for a proximal map γ ∈ GL(E) to be an optimal
canonizing map for the pair (Eγs , Eγu ).
Let C ≥ 1. We say that the pair formed by a line and a hyperplane (E s , E u ) (resp.
that a proximal map γ) is C-non-degenerate if it has an optimal canonizing map φ such
that φ±1 ≤ C. (This is equivalent to the angle between E s and E u being bounded
below by a constant that depends only on C.)
Now take γ1 , γ2 two proximal maps in GL(E). We say that the pair (γ1 , γ2 ) is C-nondegenerate if every one of the four possible pairs (Eγsi , Eγuj ) is C-non-degenerate.
19
Definition 3.4 (Proximal version of contraction strength). Let γ ∈ GL(E) be a proximal
map. We define the proximal contraction strength of γ by
s̃(γ) :=
γ|Eγu
γ|Eγs
=
γ|Eγu
r(γ)
(where r(γ) is the spectral radius of γ, equal to |λ1 | in the notations of the previous
definition). We say that γ is s̃-contracting if s̃(γ) ≤ s̃.
Proposition 3.5. For every C ≥ 1, there is a positive constant s̃3.5 (C) with the following
property. Take a C-non-degenerate pair of proximal maps γ1 , γ2 in GL(E), and suppose
that both γ1 and γ2 are s̃3.5 (C)-contracting. Then γ1 γ2 is proximal, and we have:
(i) α Eγs1 γ2 , Eγs1 .C s̃(γ1 );
(ii) s̃(γ1 γ2 ) .C s̃(γ1 )s̃(γ2 ).
(iii) r(γ1 γ2 ) ≍C kγ1 kkγ2 k.
(The constant s̃3.5 (C) is indexed by the number of the proposition, a scheme that we
will stick to throughout the paper.)
Similar results have appeared in the literature for a long time, see e.g. Lemma 5.7
in [AMS02], Proposition 6.4 in [Ben96] or Lemma 2.2.2 in [Ben97].
Proof. See Proposition 3.4 in [Smi16] for the proof of (i) and (ii), and Proposition 5.12
in [Smi] for the proof of (iii).
Remark 3.6. If we wanted to literally follow Schema 1.5 (taking “asymptotic dynamics”
to mean the logarithm of the spectral radius), we would need:
• to add a point (i’): α Eγu1 γ2 , Eγu2 .C s̃(γ2 );
• to replace (iii) by (iii’): r(γ1 γ2 ) ≍C r(γ1 )r(γ2 ).
However:
• The estimate (i’) will not be used in the sequel. It is nevertheless true: it follows
by considering the action of GL(E) on the dual space E ∗ and applying (i).
• The estimate (iii’) is on the contrary not strong enough for the applications we
need. It is also true; it follows by plugging the identity
r(γ) ≍C kγk
(valid for proximal, C-non-degenerate, s̃3.5 (C)-contracting γ; obtained from (iii)
by setting γ1 = γ2 = γ) into (iii) itself.
20
4 X-regular linear maps
In this section, we define the linear versions of the properties from Table 1 and state
the linear version of Schema 1.5. Most of the basic ideas, and several definitions, come
from [Ben97]; however, we use a slightly different point of view. Benoist relies most of
the time only on the proximal versions of the properties from Table 1, using the representations ρi as a proxy. We, on the other hand, clearly separate the linear versions from
the proximal versions, and establish the correspondences between linear and proximal
versions as theorems.
In Subsection 4.1, we give the definitions of these linear properties. All of them are
parametrized by some vector X ∈ a+ (or equivalently by some subset ΠX ⊂ Π; see the
discussion in the beginning of Subsection 2.3). When we will apply the linear case to the
affine case, this vector will be set to the vector X0 chosen in Section 5.
In Subsection 4.2, we examine what happens to these properties when we replace g by
its inverse g−1 .
In Subsection 4.3, we relate the linear properties with the proximal properties, and
then prove Propositions 4.22 and 4.24, which together comprise the linear version of
Schema 1.5.
4.1 Definitions
Let us fix some X ∈ a+ .
Definition 4.1 (Linear version of regularity). We say that an element g ∈ G is X-regular
if every root which does not vanish on X does not vanish on Jd(g) either:
∀α ∈ Π \ ΠX ,
α(Jd(g)) > 0.
Benoist calls such elements “elements of type θ”, where his θ is our set Π \ ΠX : see
Definition 3.2.2 in [Ben97].
Example 4.2.
1. If X ∈ a++ , then g ∈ G is X-regular if and only if Jd(g) ∈ a++ .
2. If X = 0, then the condition is vacuous: all elements of G are X-regular.
Elements g ∈ G such that Jd(g) ∈ a++ are often called R-regular or loxodromic. So
informally, being X-regular should be understood as being “partially R-regular”. (In fact,
technically we should probably say “X-R-regular” instead of “X-regular”.)
Definition 4.3 (Linear version of geometry). Let g ∈ G be an X-regular element. Let
g = gh ge gu be its Jordan decomposition, and let φ be any element of G realizing the
conjugacy
φgh φ−1 = exp(Jd(g))
(called a canonizing map for g). Then we define:
21
• the attracting X-flag of g, denoted by ygX,+ , to be the class of φ−1 in the flag variety
G/PX+ :
ygX,+ := φ−1 PX+ ∈ G/PX+ ;
• the repelling X-flag of g, denoted by ygX,− , to be the class of φ−1 in the flag variety
G/PX− :
ygX,− := φ−1 PX− ∈ G/PX− ;
• the X-geometry
of g to be the data of its attracting and repelling X-flags, i.e. the
X,+ X,−
.
pair yg , yg
Benoist defines the attracting and repelling flags in the last sentence of 3.4 in [Ben97].
Remark 4.4. Depending on context, sometimes it is the map φ itself that is more relevant
to consider, and sometimes it is its inverse. Indeed while φ is the map that “brings g
to the canonical position”, its inverse φ−1 is the map that “defines the geometry of g”,
starting from the canonical position. This is why the formulas above involve φ−1 .
We need to check that those definitions do not depend on the choice of φ. Indeed, φ−1
is unique up to multiplication on the right by an element of the centralizer of Jd(g). By
Proposition 2.17, the latter is equal to LJd(g) ; since g is X-regular, it is contained in LX ,
which in turn is contained both in PX+ and in PX− .
Definition 4.5. We say that a pair (y + , y − ) ∈ G/PX+ × G/PX− is transverse if the
intersection of y + and y − (seen as cosets in G) is nonempty, i.e. if there exists an element
φ ∈ G such that
(
y + = φPX+ ;
y − = φPX− .
In particular, the pair of flags giving the geometry of any X-regular element is transverse. (Compare this with the definition given in [Ben97] in 3.3, p. 14, l. -3.)2
Proposition 4.6. The map
ΦX : G/LX
// G/P + × G/P −
X
X
// φP + , φP −
X
X
φLX ✤
gives a canonical diffeomorphism between G/LX and the subset of G/PX+ × G/PX− formed
by transverse pairs.
From now on, we shall tacitly identify G/LX with the set of transverse pairs (which
is also known as the open G-orbit in G/PX+ × G/PX− ).
Proof. The group G acts smoothly on the manifold G/PX+ × G/PX− . The orbit of the
point (PX+ , PX− ) is precisely the set of transverse pairs, and its stabilizer is PX+ ∩ PX− =
LX .
2
By “l. -3” we mean “the third line from the end”.
22
Our map ΦX corresponds to Benoist’s “injection Zθ → Yθ × Yθ− ” introduced near the
end of 3.3 in [Ben97].
Definition 4.7. On every flag variety G/PX+ and on every flag variety G/PX− , we now
fix, once and for all, a distance coming from some Riemannian metric. All these distances
shall be denoted by δ.
Remark 4.8. Note that every flag variety G/PX+ (and G/PX− , which is isomorphic) is
compact. Indeed by the Iwasawa decomposition (see [Kna96], Theorem 6.46), the maximal compact subgroup K acts transitively on it. This means that any two Riemannian
metrics on a given flag variety are always Lipschitz-equivalent. It turns out that we will
only be interested in properties that are true up to a multiplicative constant; so the
choice of a Riemannian metric does not influence anything in the sequel.
We now introduce the notion of C-non-degeneracy, which is basically a quantitative
measure of transversality. Every transverse pair of flags is C-non-degenerate for some
constant C; but the smaller the constant gets, the “more strongly” the flags are transverse.
In [Ben97], this notion appears bundled together with contraction strength in the
concept of “(θ, ε)-proximality” (for a single element) or “(θ, ε)-Schottky-ness” (for a pair
or, more generally, a family of elements).
Definition 4.9. We fix, once and for all, a continuous proper map
ν : G → [0, +∞).
A typical example of such a map is given by
ν(g) = max kρ∗ (g)k, kρ∗ (g)−1 k ,
(4.1)
where ρ∗ can be any faithful representation of G and k • k can be any Euclidean norm on
the representation space. The specific choice of ν is not really important: see the remark
below. In practice, we will indeed find it convenient to use a very specific map ν of this
form: see (7.16).
The important property is that then the family of the preimages ν −1 ([0, C]), indexed
by C ∈ [0, +∞), is a nested family of compact sets whose union exhausts the set G.
Definition 4.10 (Linear version of non-degeneracy). Note that the last statement also
holds for the projections of these preimages onto G/LX . We may call these projections
−1
([0, C]), where we set
νX
// [0, +∞)
νX : G/LX
φLX ✤
// min ν(φ′ ).
φ′ ∈φLX
To justify that this is well-defined, notice that for every φ ∈ G, the intersection of the
coset φLX with ν −1 ([0, ν(φ)]) is compact (and nonempty), so the continuous map ν
reaches a minimum on it. Also, the map νX is still proper.
We say that an element φ ∈ G is an optimal representative of the coset φLX ∈ G/LX
if ν reaches its minimum at φ, i.e. if we have ν(φ) = νX (φLX ).
23
• We say that a transverse pair (y + , y − ) ∈ G/PX+ × G/PX− is C-non-degenerate if
νX (y + , y − ) ≤ C, or in other terms
−1
([0, C])
(y + , y − ) ∈ νX
(where we identify (y + , y − ) with a coset of G/LX by the map ΦX defined above).
• We say that an X-regular element g ∈ G is C-non-degenerate if its X-geometry is
C-non-degenerate, i.e. if
−1
ygX,+ , ygX,− ∈ νX
([0, C]).
• We say that a pair (g1 , g2 ) of X-regular elements of G is C-non-degenerate if we
have
−1
X,−
([0, C])
∈ νX
ygX,+
,
y
gj
i
for all four possible pairs (i, j) ∈ {1, 2} × {1, 2}.
Remark 4.11. Let us now explain why the choice of the function ν is not really important. Indeed, suppose we replace the function ν by another function ν ′ having the same
property. Then we will simply need to replace every constant C by some C ′ that depends
only on C.
Example 4.12. Take G = SO+ (n, 1). This group has real rank 1; so the closed Weyl
chamber a++ is a half-line, with only two facets: {0} and its interior a+ . Taking X = 0
makes everything trivial; so assume that X ∈ a+ . Let us identify G with the group of
isometries of the hyperbolic space Hn . In this case:
• An element g ∈ G is then X-regular if and only if it is loxodromic, i.e. fixes exactly
two points of the ideal boundary ∂ ∞ Hn .
• The flag variety G/PX+ canonically identifies with this ideal boundary ∂ ∞ Hn , and
so does the opposite flag variety G/PX− .
• The attracting flag ygX,+ (resp. repelling flag ygX,− ) of an X-regular element g corresponds to the attracting (resp. repelling) fixed point at infinity of the loxodromic
isometry g.
• Two flags (y + , y − ) ∈ G/PX+ × G/PX− are transverse if and only if the corresponding
points of ∂ ∞ Hn are distinct.
• One possible choice of the “non-degeneracy” function νX is as follows. Choose any
reference point x0 ∈ Hn , and let δ+ and δ− be half-lines starting at x0 and reaching
the ideal boundary at points corresponding to y + and y − respectively. Then we
may let νX (y + , y − ) be the reciprocal of the angle between δ+ and δ− .
• In that case, a pair (y + , y − ) is C-non-degenerate if and only if the corresponding
points of the ideal boundary are separated by an angle of at least C1 (when looking
from x0 ).
24
We finish this subsection by introducing the following notion.
Definition 4.13 (Linear version of contraction strength). We define the linear Xcontraction strength of an X-regular element g ∈ G to be the quantity
~sX (g) := exp − min α(Ct(g)) .
α∈Π\ΠX
It measures how far the Cartan projection of g is from the walls of the Weyl chamber,
except those containing X.
4.2 Impact of the group inverse
In this section, we examine what happens to the properties we just introduced when we
pass from an element g ∈ G to its inverse g −1 . Though slightly technical, the proof is
straightforward.
We start by observing that for every g ∈ G, we have
Jd(g−1 ) = −w0 (Jd(g))
(4.2)
Ct(g −1 ) = −w0 (Ct(g)).
(4.3)
and
(The map −w0 is Benoist’s “involution opposition” ı: compare these formulas with [Ben97],
section 2.4, p. 8, l. 10.)
The first identity immediately follows from the definitions of the Jordan projection and
of w0 . The second identity also follows from the definitions, using the well-known fact
that every element of the restricted Weyl group W (in particular w0 ) has a representative
in the maximal compact subgroup K (see [Kna96], formulas (7.84a) and (7.84b)).
Proposition 4.14.
(i) An element g ∈ G is X-regular if and only if its inverse g −1 is −w0 (X)-regular;
(ii) For every X-regular element g ∈ G, we have
−w (X),+
X,−
ι−
,
= yg−10
X yg
where ι±
X are the diffeomorphisms given by
±
ι±
X : G/PX
// G/P ∓
−w
φPX± ✤
// φw0 P ∓
−w
0 (X)
0 (X)
.
(iii) If a pair
(y + , y − ) ∈ G/PX+ × G/PX−
is C-non-degenerate, then the pair
−
+
+ +
−
(ι−
X (y ), ιX (y )) ∈ G/P−w0 (X) × G/P−w0 (X)
is C ′ -non-degenerate for some constant C ′ that depends only on C.
25
(iv) For every X-regular element g ∈ G, we have
~sX (g) = ~s−w0 (X) (g −1 ).
Remark 4.15. Starting from Section 5, we will only consider situations where X will be
symmetric, i.e. −w0 (X) = X (which simplifies the above formulas).
Proof.
(i) This is an immediate consequence of (4.2).
+
−1 −
(ii) To show that the map ι−
X is well-defined, note that we have w0 PX w0 = P−w0 (X)
(this easily follows from the definitions). It is obviously smooth. Hence the
−
map ι+
−w0 (X) is also smooth, and is clearly equal to the inverse of ιX . To show
the desired identity, note that (by (4.2)) if φ is a canonizing map for g, then w0 φ is
a canonizing map for g −1 . (Pay attention to the convention of φ versus φ−1 .)
(iii) The map φ 7→ φw0 descends to a diffeomorphism ιL
X that makes the diagram
G/LX
ΦX
G/PX+ × G/PX−
−
σ ◦ (ι+
X × ιX )
ιL
X
G/L−w0 (X)
Φ−w0 (X)
(4.4)
−
+
× G/P−w
G/P−w
0 (X)
0 (X)
commutative. Here ΦX is the embedding from Proposition 4.6, and σ denotes the
pair-switching map: (a, b) 7→ (b, a) (the crossing-over double arrow is meant to
−
suggest this graphically). So clearly the map σ ◦ (ι+
X × ιX ) preserves transversality
−1
L
([0, C]) to some
of pairs. Now for every C, the map ιX maps the preimage νX
compact subset of G/L−w0 (X) , which is in particular contained in the preimage
−1
([0, C ′ ]) for some C ′ .
ν−w
0 (X)
(iv) This is an immediate consequence of (4.3).
Remark 4.16.
• In point (i), since the choice of the metrics on the flag varieties was arbitrary,
we lose no generality in assuming that the diffeomorphisms ι±
X defined above are
actually isometries.
• In point (iii), if ν is chosen in a sufficiently natural way (for example as defined
by (7.16)), we may actually let C ′ = C.
26
4.3 Products of X-regular maps
In this subsection, we start by proving a few results that link linear properties to proximal
properties, via the representations ρi :
• Proposition 4.17.(i) for regularity;
• Lemma 4.21 for geometry;
• Proposition 4.19 for non-degeneracy;
• Proposition 4.17.(ii) for contraction strength.
Their proofs are adapted from Section 6 in [Smi].
We then prove the linear version of Schema 1.5. It consists of two parts:
• Proposition 4.22 is the “main” part. This result did not appear in the previous
paper.
• Proposition 4.24 is the “asymptotic dynamics” part. It gives the same conclusion
as Proposition 6.10 in [Smi], but uses the linear versions of the properties in its
hypotheses (which are the most natural here) rather than the affine versions.
Proposition 4.17.
(i) An element g ∈ G is X-regular if and only if for every i ∈ Π \ ΠX , the map ρi (g)
is proximal.
(ii) For every C ≥ 1, there is a constant ~s4.17 (C) with the following property. Let
g ∈ G be a C-non-degenerate X-regular map such that ~sX (g) ≤ ~s4.17 (C). Then
for every i ∈ Π \ ΠX , we have
s̃(ρi (g)) .C ~sX (g).
These two statements essentially correspond to the respective left halves of (6.6)
and (6.7) in [Smi]. The proof is also essentially the same.
Part (i) is given in Definition 3.2.2 in [Ben97].
Remark 4.18.
• Note that since all Euclidean norms on a finite-dimensional vector space are equivalent, this estimate makes sense even though we did not specify any norm on Vi .
In the course of the proof, we shall choose one that is convenient for us.
• Recall that “i ∈ Π \ ΠX ” is a notation shortcut for “i such that αi ∈ Π \ ΠX ”.
• ΠX should be thought of as a kind of “exceptional set”. In practice, it will often be
empty (see Remark 5.15 below).
Proof of Proposition 4.17.
27
(i) By Proposition 2.6 (i), the list of the moduli of the eigenvalues of ρi (g) is precisely
j
eλi (Jd(g))
,
1≤j≤di
where di is the dimension of Vi and (λji )1≤j≤di is the list of restricted weights of ρi
listed with multiplicity.
Up to reordering that list, we may suppose that λ1i = ni ̟i is the highest restricted
weight of ρi . We may also suppose that λ2i = ni ̟i − αi . Indeed we have
sαi (ni ̟i ) = sα′i (ni ̟i )
= ni ̟i − 2ni
h̟i , α′i i ′
α
hα′i , α′i i i
= ni ̟i − ni α′i
(4.5)
(recall that α′i is equal to 2αi if 2αi is a restricted root and to αi otherwise). But
by Proposition 2.11, sαi (ni ̟i ) is a restricted weight of ρi (because it is the image
of a restricted weight of ρi by an element of the Weyl group) and then ni ̟i − αi
is also a restricted weight of ρi (as a convex combination of two restricted weights
of ρi , that belongs to the restricted root lattice shifted by ni ̟i ).
Take any j > 2. Since by hypothesis, the restricted weight ni ̟i has multiplicity 1,
we have λji 6= λ1i . By Lemma 2.14, it follows that this restricted weight has the
form
λji
= ni ̟i − αi −
r
X
ci′ αi′ ,
i′ =1
with ci′ ≥ 0 for every index i′ .
Finally, since by definition Jd(g) ∈ a+ , for every index i′ we have αi′ (Jd(g)) ≥ 0.
It follows that for every j > 2, we have
λ1i (Jd(g)) ≥ λ2i (Jd(g)) ≥ λji (Jd(g)).
(4.6)
In other words, among the moduli of the eigenvalues of ρi (g), the largest is
exp λ1i (Jd(g)) = exp ni ̟i (Jd(g)) ,
and the second largest is
exp λ2i (Jd(g)) = exp ni ̟i (Jd(g)) − αi (Jd(g)) .
It follows that the spectral gap of ρi (g) is equal to
exp λ1i (Jd(g))
= exp(αi (Jd(g)));
exp λ2i (Jd(g))
the conclusion follows immediately.
28
(4.7)
(ii) Let ~s4.17 (C) be a constant small enough to satisfy all the constraints that will
appear in the course of the proof. Let us fix i ∈ Π \ ΠX , and let g ∈ G be a map
satisfying the hypotheses. Clearly it is enough to show that we have
s̃(ρi (g)) ≍C exp(αi (Ct(g)))−1
(4.8)
(indeed by definition, the right-hand side is smaller or equal than ~sX (g).)
We start with the following observation: for every C ≥ 1, the continuous map
φ 7→ max ρi (φ) , ρi (φ−1 )
(4.9)
is bounded above on the compact set ν −1 ([0, C]), by some constant Ci′ that depends
only on C (and on the choice of a norm on Vi , to be made soon). Let φ be an optimal
representative of the coset in G/LX giving the geometry of g, and let g′ = φgφ−1 ;
then we get
s̃(ρi (g)) ≍C s̃(ρi (g′ )).
(4.10)
Now let us choose, on the space Vi where the representation ρi acts, a K-invariant
Euclidean form Bi such that all the restricted weight spaces for ρi are pairwise
Bi -orthogonal (this is possible by Lemma 2.4 applied to ρi ). Then s̃(ρi (g′ )) is simply
the quotient of the two largest singular values of ρi (g ′ ). By Proposition 2.6 (ii)
(giving the singular values of an element of G in a given representation) and by a
calculation analogous to the previous point, we have
s̃(ρi (g′ )) = exp(αi (Ct(g)))−1 .
(4.11)
The desired estimate follows by combining (4.10) with (4.11).
Proposition 4.19. Let (g1 , g2 ) be a C-non-degenerate pair of X-regular elements of G.
Then for every i ∈ Π\ΠX , the pair (ρi (g1 ), ρi (g2 )) is a C ′ -non-degenerate pair of proximal
maps in GL(Vi ), where C ′ is some constant that depends only on C.
This is a straightforward generalization of Proposition 6.6 in [Smi]. Its proof relies on
the following two lemmas, which are analogous to Lemmas 6.7 and 6.9 in [Smi]. However,
the two lemmas that follow are now formulated more generally; the stronger statements
will be useful in order to prove Proposition 4.22.
Lemma 4.20. We have:
(i)
\
StabG
Vini ̟i
i∈Π\ΠX
(ii)
\
i∈Π\ΠX
StabG
M
!
λ6=ni ̟i
= PX+ ;
Viλ
!
= PX− .
Proof. For every i ∈ Π, let Ωi be the set of restricted weights of the representation ρi .
29
(i) We begin by noting that for every i, we have
!
StabG
Vini ̟i
= P + StabW (ni ̟i )P + = P + W̟i P + .
This follows from Lemma 2.23: indeed the singleton {ni ̟i } is clearly a top-subset
in Ωi . Corollary 2.24 then gives us that
!
StabG
Vini ̟i
+
= P̟
.
i
(4.12)
Now it is an easy exercise to show that for any X, Y, Z ∈ a+ , we have
+
+
PX+ ∩ PY+ = PZ+ ⇐⇒ p+
X ∩ pY = pZ
⇐⇒ WX ∩ WY = WZ
⇐⇒ ΠX ∩ ΠY = ΠZ .
(4.13)
Now obviously Π̟i = Π \ {αi }, and if we remove from Π all elements that are not
in ΠX , we are left with precisely ΠX . Thus
\
+
(4.14)
P̟
= PX+ ,
i
i∈Π\ΠX
and the conclusion follows.
(ii) Note that for every i, the complement Ωi \ {ni ̟i } is a bottom-subset of Ωi , and
has the same stabilizer in W as {ni ̟i }. We may then follow the same line of
reasoning.
In the following lemma, we identify the projective space P(Vi ) with the set of vector
lines in Vi and the projective space P(Vi∗ ) with the set of vector hyperplanes of Vi .
Lemma 4.21. There exist two smooth embeddings
Y
ΦsX : G/PX+ →
P(Vi ) and ΦuX : G/PX− →
i∈Π\ΠX
Y
P(Vi∗ )
i∈Π\ΠX
with the following properties:
(i) For every X-regular map g ∈ G, we have
ΦsX (ygX,+ ) = Eρsi (g)
;
i∈Π\ΠX
X,−
.
ΦuX (yg ) = Eρui (g)
i∈Π\ΠX
(ii) If (y + , y − ) ∈ G/PX+ × G/PX− is a transverse pair, then for every i ∈ Π \ ΠX , we
have ΦsX (y + )i 6∈ ΦuX (y − )i .
30
Our map ΦsX is the map jθ defined in the beginning of 3.4 in [Ben97]. Statement (ii)
appears a little further in the same text (together with its converse which is also true).
Proof. We define the maps ΦsX and ΦuX in the following way. For every φ ∈ G, we set
ΦsX (φPX+ ) = ρi (φ) (Vini ̟i )
;
i∈Π\Π
L
X
(4.15)
λ
.
ΦuX (φPX− ) = ρi (φ)
λ6=ni ̟i Vi
i∈Π\ΠX
These maps are well-defined and injective by Lemma 4.20, and obviously continuous.
They are by construction G-equivariant; hence to prove that they are smooth embeddings,
it is sufficient to prove that both of them have injective differential at the identity coset.
This also follows from Lemma 4.20 (by differentiating it).
To show property (i), we essentially use:
• the identities
(
Eρsi (exp(Jd(g))) = Vini ̟i ;
L
Eρui (exp(Jd(g))) = λ6=ni ̟i Viλ ,
(4.16)
which follow from the inequality (4.6) ranking the values of different restricted
weights of ρi evaluated at Jd(g);
• and the simple observation that any eigenspace of φgφ−1 is the image by φ of the
eigenspace of g with the same eigenvalue.
Property (ii) is obvious from the definitions.
Proof of Proposition 4.19. Let C ≥ 1. Then the set of C-non-degenerate transverse pairs
in G/PX+ × G/PX− is compact. On the other hand, the function
(y + , y − ) 7→ max α(ΦsX (y + )i , ΦuX (y − )i )
i∈Π\ΠX
is continuous, and (by Lemma 4.21 (ii)) takes positive values on that set. Hence it is
bounded below. So there is a constant C ′ ≥ 1, depending only on C, such that whenever
a transverse pair (y + , y − ) is C-non-degenerate, all pairs (ΦsX (y + )i , ΦuX (y − )i ) are C ′ -nondegenerate.
The conclusion then follows by Lemma 4.21 (i).
Proposition 4.22. For every C ≥ 1, there is a positive constant ~s4.22 (C) < 1 with the
following property. Take any C-non-degenerate pair (g, h) of X-regular maps; suppose
that we have ~sX (g) ≤ ~s4.22 (C) and ~sX (h) ≤ ~s4.22 (C). Then gh is still X-regular, and
we have:
X,+
δ ygh
, ygX,+ .C ~sX (g);
δ y X,− , y X,− . ~s (h).
C X
gh
h
31
Remark 4.23. We did not include the conclusion about contraction strength, namely that
~sX (gh) ≍C ~sX (g)~sX (h).
(4.17)
This statement is true (as we shall see in a moment), but will never be directly useful to
us.
Proof. Let us fix some constant ~s4.22 (C), small enough to satisfy all the constraints
that will appear in the course of the proof. Let (g, h) be a pair of maps satisfying the
hypotheses.
• By Proposition 4.17 (i), for every i ∈ Π\ΠX , the maps ρi (g) and ρi (h) are proximal.
• By Proposition 4.19, for every i ∈ Π \ ΠX , the pair (ρi (g)), (ρi (h)) is C ′ -nondegenerate (where C ′ depends only on C).
• If we choose ~s4.22 (C) ≤ ~s4.17 (C), it follows by Proposition 4.17 (ii) that for
every i ∈ Π \ ΠX , we have s̃(ρi (g)) .C ~sX (g) and s̃(ρi (h)) .C ~sX (h). If we
choose ~s4.22 (C) sufficiently small, then all the maps ρi (g) and ρi (h) are s̃3.5 (C ′ )contracting, i.e. sufficiently contracting to apply Proposition 3.5.
Thus we may apply Proposition 3.5. We obtain:
• that for every i ∈ Π\ΠX , the map ρi (gh) is proximal. Hence by Proposition 4.17 (i),
the element gh is X-regular.
• that moreover for every i ∈ Π \ ΠX , we have
α(Eρsi (gh) , Eρsi (g) ) .C s̃(ρi (g)) .C ~sX (g).
(4.18)
Q
Now let us endow the product i∈Π\ΠX P(Vi ) with the product distance δ given by
δ(y, y ′ ) := max α(yi , yi′ );
i∈Π\ΠX
then all the inequalities (4.18) may be combined together to yield
X,+
, ΦsX ygX,+ .C ~sX (g),
δ ΦsX ygh
(4.19)
(4.20)
where ΦsX is the map introduced in Lemma 4.21. Since ΦsX is a smooth embedding
of the compact manifold G/PX+ , it is in particular a bilipschitz map onto its image.
Hence we also have
X,+ X,+
.C ~sX (g).
δ ygh
, yg
This yields the first line of the conclusion. To get the second line, simply note that
by Proposition 4.14, if we replace X by −w0 (X), the pair (h−1 , g−1 ) satisfies the same
hypotheses as the pair (g, h). Applying what we did above to this pair, we get
X,−
− X,−
(y
)
.C ~sX (h);
(y
),
ι
δ ι−
X h
X gh
by Remark 4.16, the conclusion follows.
32
Proposition 4.24. For every C ≥ 1, there are positive constants ~s4.24 (C) and ε4.24 (C)
with the following property. Take any C-non-degenerate pair (g, h) of X-regular elements
of G such that ~sX (g) ≤ ~s4.24 (C) and ~sX (h) ≤ ~s4.24 (C). Then we have:
(i) ∀i ∈ Π,
̟i (Jd(gh) − Ct(g) − Ct(h)) ≤ 0;
(ii) ∀i ∈ Π \ ΠX0 ,
̟i (Jd(gh) − Ct(g) − Ct(h)) ≥ −ε4.24 (C).
Proof. The proof is completely analogous to the proof of Proposition 6.10 in [Smi].
See Figure 1 for a picture explaining both this proposition and the corollary below.
Let us also give a more palatable (though slightly weaker) reformulation.
Corollary 4.25. For every C ≥ 1, there exists a positive constant ε4.25 (C) with the
following property. For any pair (g, h) satisfying the hypotheses of the Proposition, we
have
Jd(gh) ∈ Conv WX · Ct′ (g, h) ,
(4.21)
where Conv denotes the convex hull and Ct′ (g, h) is some vector in a satisfying
k Ct′ (g, h) − Ct(g) − Ct(h)k ≤ ε4.25 (C).
(4.22)
Proof. The proof is the same as the proof of Corollary 6.12 in [Smi].
Note that we do not require that the vector Ct′ (g, h) lie in the closed dominant Weyl
chamber a+ (even though in practice, it is very close to the vector Ct(g) + Ct(h), which
does).
Remark 4.26. At first sight, one might think that by putting together Lemma 4.1
and 4.5.2 in [Ben97], we recover this result and even something stronger. However this is
not the case: in fact we recover only the particular case when ΠX = ∅. See Remark 6.13
in [Smi] for an explanation.
Remark 4.27.
• Though we shall not use it, an interesting particular case is g = h. We the obviously
have Jd(gh) = 2 Jd(g) and Ct(g)+Ct(h) = 2 Ct(g), so the statement simply reduces
to a relationship between the Jordan and Cartan projections on g.
• The Proposition, and its Corollary, also hold if we replace Jd(gh) by Ct(gh). This
immediately implies (4.17).
(Compare this to Remark 3.6, which involved a similar construction in the proximal
case.)
33
WX = {Id, sα1 }
ε4.24 (C)
X
sα1 · Ct′ (g, h)
α2
̟2
Jd(gh)
Ct(g) + Ct(h)
̟1
Ct′ (g, h)
α1
Figure 1: This picture represents the situation for G = SO+ (3, 2) acting on R5 , and X
chosen such that ΠX = {α1 } (or {1} with the usual abuse of notations). This
choice of X is not random: it satisfies the conditions that will be required
starting from Section 5. (This also corresponds to Example 3.5.3 in [Smi].)
The group WX is then generated by the single reflection sα1 . Proposition 4.24
states that Jd(gh) lies in the shaded trapezoid. Corollary 4.25 states that it
lies on the thick line segment. In any case it lies by definition in the dominant
open Weyl chamber a++ (the shaded sector).
34
5 Choice of a reference Jordan projection X0
For the remainder of the paper, we fix ρ an irreducible representation of G on a finitedimensional real vector space V . For the moment, ρ may be any representation; but in
the course of the paper, we shall gradually introduce several assumptions on ρ (namely
Assumptions 5.5, 7.12 and 11.1) that will ensure that ρ satisfies the hypotheses of the
Main Theorem.
<
We call Ω the set of restricted weights of ρ. For any X ∈ a, we call Ω>
X (resp. ΩX ,
≥
≤
Ω=
X , ΩX , ΩX ) the set of all restricted weights of ρ that take a positive (resp. negative,
zero, nonnegative, nonpositive) value on X:
Ω≥
X := {λ ∈ Ω | λ(X) ≥ 0}
Ω>
X := {λ ∈ Ω | λ(X) > 0}
Ω≤
X := {λ ∈ Ω | λ(X) ≤ 0}
Ω<
X := {λ ∈ Ω | λ(X) < 0}
Ω=
X := {λ ∈ Ω | λ(X) = 0} .
The goal of this section is to study these sets, and to choose a vector X0 ∈ a+ for which
the corresponding sets have some nice properties. This generalizes Section 3 in [Smi].
In fact, these sets are the only property of X0 that matters for us; in other terms, what
we really care about is the class of X0 with respect to the following equivalence relation:
<
<
>
Definition 5.1. We say that X and Y have the same type if Ω>
Y = ΩX and ΩY = ΩX .
Obviously this implies that the spaces Ω≥ , Ω≤ and Ω= coincide as well for X and Y .
This is an equivalence relation, which partitions a into finitely many equivalence classes.
Remark 5.2. Every such equivalence class is obviously a convex cone; taken together, the
equivalence classes decompose a into a cell complex.
Example 5.3. If ρ is the adjoint representation, two dominant vectors X, Y ∈ a+ have
the same type if and only if ΠX = ΠY . There is only one generic type, corresponding
to a++ .
See Example 3.5.3 in [Smi] for more details, and two other examples with pictures.
The motivation for the study of these five sets Ω?X is that they allow us to introduce
some “reference dynamical spaces” (see Subsection 6.1).
In Subsections 5.1 and 5.2, we define two properties that we want X0 to satisfy. Subsection 5.3 basically consists of examples, and may be safely skipped.
5.1 Generically symmetric vectors
We start by defining the property of being “generically symmetric”, which generalizes
generic and symmetric vectors as defined in Subsections 3.1 and 3.3 of [Smi]. One of the
goals is to ensure that the set Ω=
X is as small as possible.
Here is the first attempt: we say that an element X ∈ a is generic if
Ω=
X ⊂ {0}.
35
Remark 5.4. This is indeed the generic case: it happens as soon as X avoids a finite
collection of hyperplanes, namely the kernels of all nonzero restricted weights of ρ. In
fact, a vector is generic if and only if its equivalence class is open. If X is generic, its
equivalence class is just the connected component containing X in the set of generic vectors; otherwise, its equivalence class is always contained in some proper vector subspace
of a.
In fact, for generic X we actually have
Ω=
X = {0},
provided the following condition is met:
Assumption 5.5. From now on, we assume that 0 is a restricted weight of ρ:
0 ∈ Ω,
or equivalently
dim V 0 > 0.
Remark 5.6. By Proposition 2.11, this is the case if and only if the highest restricted
weight of ρ is a Z-linear combination of restricted roots.
We lose no generality in assuming this, because this assumption is necessary for condition (i)(a) of the Main Theorem (which is also Assumption 7.12, see below) to hold.
Indeed, any nonzero vector fixed by L is in particular fixed by A ⊂ L, which means that
it belongs to the zero restricted weight space.
Since we want to construct a group Γ with certain properties, it must in particular be
stable by inverse. The identity (4.2) encourages us to examine the action of −w0 on a.
Definition 5.7. We say that an element X ∈ a is symmetric if it is invariant by −w0 :
−w0 (X) = X.
Ideally, we would like our “reference” vector X0 to be both symmetric and generic.
Unfortunately, this is not always possible: indeed, every restricted weight that happens
to be invariant by w0 necessarily vanishes on every symmetric vector. Let
Ωw0 := {λ ∈ Ω | w0 λ = λ}
be the set of those restricted weights.
Definition 5.8. We say that an element X ∈ a is generically symmetric if it is symmetric
and we have
w0
Ω=
X =Ω .
In other terms, an element is generically symmetric if it is “as generic as possible” while
still being symmetric.
36
5.2 Extreme vectors
Besides WX , we are also interested in the group
Wρ,X := {w ∈ W | wX has the same type as X} ,
(5.1)
which is the stabilizer of X “up to type”. It obviously contains WX . The goal of this
subsection is to show that in every equivalence class, we can actually choose X in such
a way that both groups coincide. This generalizes Subsection 3.5 in [Smi].
Example 5.9. In Example 3.5.3 in [Smi] (G = SO+ (3, 2) acting on V = R5 ), the group
Wρ,X corresponding to any generic X is a two-element group. If we take X to be generic
not only with respect to ρ but also with respect to the adjoint representation (in other
terms if X is in an open Weyl chamber), then the group WX is trivial. If however we take
as X any element of the diagonal wall of the Weyl chamber, we have indeed WX = Wρ,X .
Definition 5.10. We call an element X ∈ a+ extreme if WX = Wρ,X , i.e. if it satisfies
the following property:
∀w ∈ W,
wX has the same type as X ⇐⇒ wX = X.
Remark 5.11. Here is an equivalent definition which is possibly more enlightening. It is
possible to show that a vector X ∈ a+ is extreme if and only if it lies in every wall of the
Weyl chamber that contains at least one vector of the same type as X. In other words,
a vector is extreme if it is in the “furthest possible corner” of its equivalence class in the
Weyl chamber a+ .
As this last statement will never be used in this paper, we have left out its proof.
Proposition 5.12. For every generically symmetric X ∈ a+ , there exists a generically
symmetric X ′ ∈ a+ that has the same type as X and that is extreme.
This is a straightforward generalization of Proposition 3.17 in [Smi]. The proof is
similar.
Proof. To construct an element that has the same type as X but has the whole group Wρ,X
as stabilizer, we simply average over the action of this group: we set
X
X′ =
wX.
(5.2)
w∈Wρ,X
(As multiplication by positive scalars does not change anything, we have written it as
a sum rather than an average for ease of manipulation.) Let us check that it has the
required properties.
• Let us show that X ′ is still symmetric. Since w0 belongs to the Weyl group, it
induces a permutation on Ω, hence we have:
>
>
<
w0 Ω>
X = Ωw0 X = Ω−X = ΩX ,
37
(5.3)
<
so that w0 swaps the sets Ω>
X and ΩX . Now by definition we have
<
Wρ,X = StabW (Ω>
X ) ∩ StabW (ΩX ),
(5.4)
hence w0 normalizes Wρ,X . Obviously the map X 7→ −X commutes with everything, so −w0 also normalizes Wρ,X . We conclude that
X
−w0 (X ′ ) =
−w0 (w(X))
w∈Wρ,X
=
X
w ′ ∈W
w′ (−w0 (X))
ρ,X
′
=X.
(5.5)
• By definition every wX for w ∈ Wρ,X has the same type as X; since the equivalence
class of X is a convex cone, their sum X ′ also has the same type as X.
=
w0
′
• In particular we have Ω=
X ′ = ΩX = Ω , so X is still generically symmetric.
• By construction whenever wX has the same type as X, we have wX ′ = X ′ ; conversely if w fixes X ′ , then wX has the same type as wX ′ = X ′ which has the same
type as X. So X ′ is extreme.
• It remains to show that X ′ ∈ a+ , i.e. that for every α ∈ Π, we have α(X ′ ) ≥ 0:
– If sα X ′ = X ′ , then obviously α(X ′ ) = 0.
– Otherwise, since X ′ is extreme, it follows that sα X ′ does not even have the
same type as X ′ . This means that there exists a restricted weight λ of ρ such
that
λ(X ′ ) ≥ 0 and sα (λ)(X ′ ) ≤ 0,
(5.6)
and (at least) one of the two inequalities is strict. In particular we have
λ − sα (λ) (X ′ ) 6= 0.
(5.7)
Since λ − sα (λ) is by definition a multiple of α, it follows that
α(X ′ ) 6= 0.
Now the same reasoning applies to any vector of the same type as X; hence α
never vanishes on the equivalence class of X. Since by hypothesis α(X) ≥ 0
and since the equivalence class is connected, we conclude that
α(X ′ ) > 0.
Note that in practice, the set ΠX for an extreme, generically symmetric X can take
only a very limited number of values: see Remark 3.19 in [Smi].
38
5.3 Simplifying assumptions
In this subsection, we discuss how the constructions of this paper may simplify in some
particular cases. These results will never be reused in the paper, which is why we do not
provide proofs. However, they can be helpful for a reader who is only interested in one
particular representation (which is likely to fit at least one of the cases outlined below).
Definition 5.13. We say that the representation ρ is:
1. limited if every restricted weight is a multiple of a restricted root:
Ω ⊂ Z · Σ := {nα | n ∈ Z, α ∈ Σ} ;
2. abundant if every restricted root is a restricted weight:
Ω ⊃ Σ ∪ {0};
3. awkward if it is neither limited nor abundant;
4. non-swinging if Ωw0 = {0}.
Example 5.14.
1. The adjoint representation is both limited and abundant, and always non-swinging.
2. The standard representation of SO(p + 1, p) on R2p+1 is limited, and non-swinging.
3. If G is simple, it seems that all but finitely many representations are abundant. If
additionally its restricted root system has a simply-laced diagram, then Lemma 2.15
says that all representations except the trivial one are abundant.
4. Among simple groups, it seems that awkward representations occur only when the
restricted root system is of type Cn or BCn , and only for n at least equal to 4.
For non-simple groups, this phenomenon is more common. See Example 6.20 for
specific examples.
5. Swinging representations occur only when −w0 is a nontrivial automorphism of the
Dynkin diagram. Among simple groups, this happens only if the restricted root
system is of type An (n ≥ 2), D2n+1 or E6 . The bad news is that for these groups,
most representations (all but finitely many) are swinging. The simplest example is
SL3 (R) acting on S 3 R3 (see Example 3.7 in [Smi]).
6. Thus no representation of a simple group is swinging and awkward at the same
time. (However this may happen for a non-simple group G1 × G2 : simply take the
tensor product of a swinging representation ρ1 of G1 and an awkward representation
ρ2 of G2 .)
All the subsequent constructions rely on the choice of a generically symmetric vector X0 ∈ a+ that is extreme; actually, only the type of X0 matters. Here is what we can
say about the choice of X0 (up to type) in these particular cases:
39
Remark 5.15.
(i) In a limited representation, there is only one type of generically symmetric vector.
So we can ignore the dependence on X0 .
(ii) In an abundant representation, every generically symmetric vector lies in particular
in a++ . In other terms we have ΠX0 = ∅, and we do not need the theory of nonminimal parabolic subgroups (as developed in section 2.3).
(iii) In both cases, we get that the type of X0 with respect to the adjoint representation
(i.e. the subset ΠX0 ) does not depend on the choice of X0 . In fact, in every nonawkward representation, for every generically symmetric and extreme X0 ∈ a+ , we
have the identity
ΠX0 = {α ∈ Π | α 6∈ Ω} .
(5.8)
(iv) Note that one of the inclusions, namely
ΠX0 ⊂ {α ∈ Π | α 6∈ Ω} ,
is obvious and holds in every representation. The other inclusion however may fail
in awkward representations, and the value of ΠX0 may then depend on the choice
of X0 : see Example 6.20.
(v) In a non-swinging representation, clearly the vector X0 is generically symmetric if
and only if it is generic and also symmetric.
6 Constructions related to X0
Definition 6.1. For the remainder of the paper, we fix some vector X0 in the closed
dominant Weyl chamber a+ that is generically symmetric and extreme.
In this section, we introduce some preliminary constructions associated to this vector X0 .
In Subsection 6.1, we introduce the “reference dynamical spaces” associated to X0 , and
find their stabilizers in G. This generalizes Subsection 4.1 in [Smi].
Subsection 6.2 consists entirely of definitions. We introduce some elementary formalism
that expresses affine spaces in terms of vector spaces, and use it to define the affine
reference dynamical spaces. We basically repeat Subsection 4.2 from [Smi].
In Subsection 6.3, we try to understand what “regularity” means in the affine context.
We introduce two different notions of regularity, and establish the relationships between
them. Previously we used only one of the two notions: see Definition 4.14 in [Smi].
Subsection 6.4 consists entirely of examples. It contains counterexamples to help understand why some statements of the previous subsection cannot be made stronger.
40
6.1 Reference dynamical spaces
Definition 6.2. We define the following subspaces of V :
L
• V0> := λ(X0 )>0 V λ , the reference expanding space;
• V0< :=
• V0= :=
• V0≥ :=
• V0≤ :=
L
L
λ(X0 )<0 V
λ,
the reference contracting space;
λ(X0 )=0 V
λ,
the reference neutral space;
λ,
the reference noncontracting space;
λ,
the reference nonexpanding space.
L
λ(X0 )≥0 V
L
λ(X0 )≤0 V
In other terms, V0≥ is the direct sum of all restricted weight spaces corresponding to
weights in Ω≥
X0 , and similarly for the other spaces.
These are precisely the dynamical spaces associated to the map exp(X0 ) (acting on V
by ρ), as defined in section 4.3 of [Smi]. See Example 4.3 in [Smi] for the case of the
adjoint representation and of G = SO+ (p, q) acting on the standard representation.
Remark 6.3. Note that by Assumption 5.5, zero is a restricted weight, so the space V0=
is always nontrivial.
Let us now determine the stabilizers in G of these subspaces.
Proposition 6.4. We have:
(i) StabG (V0≥ ) = StabG (V0> ) = PX+0 .
(ii) StabG (V0≤ ) = StabG (V0< ) = PX−0 .
This generalizes Proposition 4.4 in [Smi]; but now the proof is somewhat more involved.
Proof.
By Lemma 2.23 and Corollary 2.24, it is enough to show that
≤
>
<
StabW (Ω≥
X0 ) = StabW (ΩX0 ) = StabW (ΩX0 ) = StabW (ΩX0 ) = WX0 .
(6.1)
≤
>
<
Indeed since X0 ∈ a+ , clearly Ω≥
X0 and ΩX0 are top-subsets and ΩX0 and ΩX0 are
bottom-subsets of Ω.
Now obviously any subset of Ω and its complement always have the same stabilizer
by W (since Ω is stable by W ). On the other hand, we have
since X0 is extreme
WX0 = Wρ,X0
>
<
= StabW (ΩX0 ) ∩ StabW (ΩX0 )
by definition.
≥
<
>
So it is sufficient to show that Ω>
X0 and ΩX0 , or equivalently ΩX0 and ΩX0 , have the same
stabilizer in W . This is a consequence of Lemma 6.5 below.
41
Lemma 6.5. Every element w of the restricted Weyl group W such that
≥
wΩ>
X0 ⊂ Ω X0
(6.2)
≥
stabilizes every set of restricted weights Ξ such that Ω>
X0 ⊂ Ξ ⊂ Ω X0 .
In particular all such sets Ξ have the same stabilizer in W : indeed every w ∈ W that
≥
stabilizes Ω>
X0 , and every w ∈ W that stabilizes ΩX0 , satisfies in particular (6.2).
Proof. Let us decompose g as a sum of three pieces g = g1 ⊕ g2 ⊕ g3 , with respective
Cartan subspaces a1 , a2 and a3 , such that:
• every simple summand of g1 has a restricted root system with a nonempty, simplylaced diagram, and acts nontrivially on V ;
• the restriction of w0 to a2 is − Id;
• g3 acts trivially on V .
Then we also have the decompositions:
• Σ = Σ1 ∪ Σ2 ∪ Σ3 , where Σi is the restricted root system of gi ;
+
+
+
+
• Σ+ = Σ+
1 ∪ Σ2 ∪ Σ3 , where every Σi := Σi ∩ Σ is a system of positive restricted
roots for gi ;
• W = W1 × W2 × W3 , where Wi is the restricted Weyl group of gi .
We now prove the two contrasting statements (6.3) and (6.4).
≥
• On the one hand, we claim that for any w ∈ W such that wΩ>
X0 ⊂ ΩX0 , we have
w ∈ W2 × W3 .
(6.3)
Indeed, let us decompose w = w1 w2 w3 with wi ∈ Wi .
Applying Lemma 2.15 to every simple summand of g1 , we find that Σ1 ⊂ Ω. Since
w0 does not intersect Σ (it does not
X0 is generically symmetric, the set Ω=
1
X0 = Ω
even intersect Σ, as by definition w0 does not leave any restricted root invariant).
Since X0 is dominant, we deduce that
+
>
Ω≥
X0 ∩ Σ 1 = Ω X0 ∩ Σ 1 = Σ 1 .
It follows that
>
w(Σ+
1 ) = w(ΩX0 ∩ Σ1 )
= w(Ω>
X0 ) ∩ Σ 1
since W stabilizes Σ1
⊂ Ω≥
X0 ∩ Σ 1
=
by assumption
Σ+
1.
Thus w fixes Σ+
1 , which precisely means that w1 = Id as required.
42
• On the other hand, we claim that
w0
Ω=
⊂ a∗1
X0 = Ω
(6.4)
(the equality holds since X0 is generically symmetric). Indeed, take any λ ∈ Ωw0 ,
and let us decompose it as λ = λ1 +λ2 +λ3 with λi ∈ a∗i . The component λ3 vanishes
by definition of g3 . As for λ2 , we have by definition of g2 that w0 (λ2 ) = −λ2 , so
the component λ2 also vanishes.
Combining (6.3) with (6.4), we conclude that whenever an element w ∈ W satisfies
≥
=
wΩ>
X0 ⊂ ΩX0 , it actually fixes every element of ΩX0 . Now take such a w, and take any
element ξ ∈ Ξ. Since Ξ ⊂ Ω≥
X0 , we may distinguish two cases:
• Either ξ ∈ Ω=
X0 . Then it follows from the previous statement that w(ξ) = ξ ∈ Ξ.
=
• Or ξ ∈ Ω>
X0 . Then it follows from the previous statement that w(ξ) 6∈ ΩX0 . On the
≥
>
other hand we know that w(ξ) ∈ w(Ω>
X0 ) ⊂ ΩX0 . Thus w(ξ) ∈ ΩX0 ⊂ Ξ.
6.2 Extended affine space
Let VAff be an affine space whose underlying vector space is V .
Definition 6.6 (Extended affine space). We choose once and for all a point of VAff which
we take as an origin; we call R0 the one-dimensional vector space formally generated by
this point, and we set A := V ⊕ R0 the extended affine space corresponding to V . (We
hope that A, the extended affine space, and A, the group corresponding to the Cartan
space, occur in sufficiently different contexts that the reader will not confuse them.) Then
VAff is the affine hyperplane “at height 1” of this space, and V is the corresponding vector
hyperplane:
V = V × {0} ⊂ V × R0 ;
VAff = V × {1} ⊂ V × R0 .
Definition 6.7 (Linear and affine group). Any affine map g with linear part ℓ(g) and
translation vector v, defined on VAff by
g : x 7→ ℓ(g)(x) + v,
can be extended in a unique way to a linear map defined on A, given by the matrix
ℓ(g) v
.
0
1
From now on, we identify the abstract group G with the group ρ(G) ⊂ GL(V ), and
the corresponding affine group G ⋉ V with a subgroup of GL(A).
Definition 6.8 (Affine subspaces). We define an extended affine subspace of A to be a
vector subspace of A not contained in V . There is a one-to-one correspondence between
extended affine subspaces of A and affine subspaces of VAff of dimension one less. For
any extended affine subspace of A denoted by A1 (or A2 , A′ and so on), we denote by V1
(or V2 , V ′ and so on) the space A ∩ V (which is the linear part of the corresponding affine
space A ∩ VAff ).
43
Definition 6.9 (Translations). By abuse of terminology, elements of the normal subgroup V ⊳ G ⋉ V will still be called translations, even though we shall see them mostly
as endomorphisms of A (so that they are formally transvections). For any vector v ∈ V ,
we denote by τv the corresponding translation.
Definition 6.10 (Reference affine dynamical spaces). We now give a name for (the
vector extensions of) the affine subspaces of VAff parallel respectively to V0≥ , V0≤ and V0=
and passing through the origin: we set
≥
A≥
0 := V0 ⊕ R0 ,
the reference affine noncontracting space;
≤
≤
the reference affine nonexpanding space;
=
=
the reference affine neutral space.
A0 := V0 ⊕ R0 ,
A0 := V0 ⊕ R0 ,
These are obviously the affine dynamical spaces (in the sense of [Smi]) corresponding to
the map exp(X0 ), seen as an element of G ⋉ V by identifying G with the stabilizer of R0
in G ⋉ V .
We then have the decomposition
A≥
z }|0 {
A = V0> ⊕ A=
⊕ V <.
| 0 {z 0 }
A≤
0
(6.5)
<
This gives a hint for why we do not introduce the spaces “A>
0 ” or “A0 ”; see Remark 4.13
in [Smi] for a detailed explanation.
Definition 6.11 (Affine Jordan projection). Finally, we extend the notion of Jordan
projection to the whole group G ⋉ V , by setting
∀g ∈ G ⋉ V,
Jd(g) := Jd(ℓ(g)).
6.3 Conditions on the Jordan projection
In this subsection, we introduce two new notions of regularity of an element g ∈ G ⋉ V ,
given as conditions on its Jordan projection. We also determine the relationships between
them.
Definition 6.12 (Affine version of regularity). We say that an element g ∈ G ⋉ V is:
=
• ρ-regular if it is X0 -regular and we have Ω=
Jd(g) = ΩX0 , or in other terms:
(
∀α ∈ Π \ ΠX0 ,
∀λ ∈ Ω \ Ω=
X0 ,
α(Jd(g)) 6= 0
λ(Jd(g)) 6= 0;
(6.6)
λ< (Jd(g)) < λ≥ (Jd(g)).
(6.7)
• asymptotically contracting along X0 if
≥
≥
∀λ< ∈ Ω<
X0 , ∀λ ∈ ΩX0 ,
44
Of course both of these are really properties of Jd(g). By abuse of terminology, we say
that a vector Y ∈ a is respectively ρ-regular or asymptotically contracting along X0 if
Y itself satisfies respectively (6.6) or (6.7).
Remark 6.13. Rigorously we should talk about “(ρ, X0 )-regularity”, as the definition
depends on the choice of X0 . However the author feels that this dependence is not
significant enough to be constantly mentioned in this way. (See in particular point 3 in
the following example).
Example 6.14.
1. If the representation ρ is limited, all three properties of being X0 -regular, ρ-regular
and asymptotically contracting along X0 become equivalent. This includes the
standard representation of G = SO(p + 1, p) on R2p+1 : see Example 4.15.1 in [Smi].
2. If the representation ρ is limited and abundant at the same time (e.g. the adjoint
representation), all three properties actually reduce to ordinary R-regularity (since
X0 generic then means X0 ∈ a++ ).
3. If the representation ρ is non-awkward (either limited or abundant), the notion of
ρ-regularity does not depend on the choice of X0 : indeed ΠX0 is then uniquely
determined by (5.8), and ΩX0 = Ωw0 since X0 is by assumption generically symmetric.
4. In general (see Example 6.20), the notion of ρ-regularity does technically depend
on the choice of X0 .
5. If the representation ρ is non-swinging, it is possible to show that its set of restricted weights Ω is centrally symmetric (or, equivalently, invariant by −w0 ). (The
author must however admit that he knows no better proof of this fact than by complete enumeration of non-swinging representations.) It then follows that whenever
g is asymptotically contracting along X0 , so is g−1 ; and asymptotic contraction
along X0 is then equivalent to Jd(g) having the same type as X0 (in the sense of
Definition 5.1), which was the condition considered in [Smi].
6. In general, asymptotic contraction along X0 is a stronger condition than being of
type X0 . For a simple counterexample, take G = SL3 (R) acting on S 3 R3 (see [Smi],
Example 3.7). For a fancier counterexample, see Example 6.21.
Both of these new properties are “affine analogs” of X0 -regularity. However, they are
useful in slightly different contexts:
• The purpose of assuming that an element g ∈ G ⋉ V is ρ-regular is just to ensure that the affine ideal dynamical subspaces (introduced in section 7.1) are welldefined. This is a relatively weak property: Jd(g) is merely required to avoid a
finite collection of hyperplanes.
This is the property we will use the most often, and in particular the one that
makes the affine version of Schema 1.5 work.
45
• The purpose of assuming that an element g ∈ G ⋉ V is asymptotically contracting
along X0 is roughly to ensure that g acts “with the correct dynamics” on its ideal
dynamical subspaces. For more details, see the discussion following the definition
of the latter; for a motivation of the “asymptotically contracting” terminology, see
Proposition 7.23.(ii).
This property is not explicitly part of the hypotheses in Schema 1.5, since it is
implied by contraction strength (see Proposition 7.23.(i)) However, we will often
need it as an extra assumption in intermediate results. Actually, we will often need
to assume that both g and g −1 are asymptotically contracting.
The latter property is verified as soon as the Jordan projection of g points in a direction
that is, in a sense, “close enough” to that of X0 . More precisely:
Remark 6.15.
(i) The set
a′ρ,X0 := {Y ∈ a | Y and −w0 (Y ) are asymptotically contracting along X0 }
is a convex cone (i.e. stable by positive scaling and sum) and is open. (Indeed it is
an intersection of finitely many open vector half-spaces.)
(ii) The intersection a′ρ,X0 ∩ a+ contains X0 , and is in particular nonempty. (By the
identity (4.2), which obviously still holds in the affine case, this intersection is
precisely equal to
Jd(g) g ∈ G ⋉ V with both g and g −1 asymptotically contracting along X0 .)
(iii) Since a′ρ,X0 is open and meets the closed set a+ , it also meets its interior a++ . Thus
the intersection a′ρ,X0 ∩ a++ is a nonempty open convex cone. (The latter set might
not seem relevant at this point, but will be useful in the final proof of this paper.)
(Note the distinction between the set a′ρ,X0 introduced here and the set aρ,X0 defined
in (3.2) in [Smi], as the equivalence class of X0 . It is however true that in the non-swinging
case, the two sets coincide, as seen in Example 6.14.5.)
Here is the relationship between these two notions.
Proposition 6.16. Let g ∈ G ⋉ V . Then:
(i) if g is asymptotically contracting along X0 , then g is X0 -regular;
(ii) if both g and g−1 are asymptotically contracting along X0 , then g is ρ-regular.
Remark 6.17. As mentioned before (Example 6.14.5), if ρ is non-swinging, then we may
remove the assumption about g−1 . In general this is not true: see Example 6.21 for a
counterexample.
The proof relies on the following lemma. For the moment we only need the weak
version; the strong version will be useful later.
46
<
Lemma 6.18. For every i ∈ Π \ ΠX0 , there exists a pair of restricted weights (λ≥
i , λi )
≥
such that λi − λ<
i = αi , with:
≥
≥
<
(i) (weak version) λ<
i ∈ ΩX0 and λi ∈ ΩX0 ;
≥
<
>
(ii) (strong version) λ<
i ∈ ΩX0 and λi ∈ ΩX0 ∪ {0}.
Remark 6.19. As long as ρ is non-awkward, (5.8) provides a “super-strong version” of
this lemma: namely we can actually take λ≥
i = 0. In general this is not true: see
Example 6.20.
Proof.
(i) Let us reintroduce the decomposition g = g1 ⊕ g2 ⊕ g3 from the proof of Lemma 6.5,
together with the other notations that went along with it. We then have
Π = Π1 ⊔ Π2 ⊔ Π3 ,
where Π1,2,3 := Π ∩ Σ1,2,3 . Let i ∈ Π \ ΠX0 . We distinguish three cases:
• The case i ∈ Π3 never occurs: indeed for i ∈ Π3 , the symmetry sαi fixes Ω
pointwise, so we have i ∈ ΠX0 .
• If i ∈ Π1 , applying Lemma 2.15 to the simple summand of g1 containing αi ,
<
we find that αi ∈ Ω. Then we may simply take λ≥
i = 0 and λi = −αi .
• Finally suppose that i ∈ Π2 . Then by definition, we have sαi 6∈ WX0 ; by
Proposition 6.4, this means that sαi does not stabilize Ω≥
X0 . In other terms,
there exists a restricted weight λ of ρ such that
λ(X0 ) ≥ 0
and
(6.8)
sαi (λ)(X0 ) < 0
(compare this with (5.6), which was slightly weaker). Since λ is a restricted
weight, by Proposition 2.8, the number
nλ :=
hλ, αi i
2hαi , αi i
(6.9)
is an integer. We have, on the one hand:
nλ αi (X0 ) = (λ − sαi (λ)) (X0 ) > 0;
(6.10)
on the other hand, αi (X0 ) ≥ 0 (because X0 ∈ a+ ); hence nλ is positive.
By Proposition 2.11, every element of the sequence
λ, λ − αi , . . . , λ − nλ αi = sαi (λ)
is a restricted weight of ρ. Let λlast be the last element of this sequence that
≥
<
still lies in Ω≥
X0 ; then by construction λlast − αi ∈ ΩX0 . Taking λi = λlast
<
and λi = λlast − αi , we get the weak version.
47
(ii) The same proof works for the strong version, except in the last case (i ∈ Π2 ) if we
happen to have λlast 6∈ Ω>
X0 . But then since X0 is generically symmetric, we have
w0 . In particular, by (6.4), we have λ
∗
λlast ∈ Ω=
=
Ω
last ∈ a1 .
X0
Since λlast − αi is a restricted weight of ρ, by Proposition 2.11, the form −αi
is also a restricted weight of ρ (the latter is the average of the W1 -orbit of the
≥
former). By (6.10), we actually have −αi ∈ Ω<
X0 . Thus we may take λi = 0
and λ<
i = −αi .
Proof of Proposition 6.16.
<
(i) Let i ∈ Π \ ΠX0 , and let λ≥
i and λi be the restricted weights constructed in
Lemma 6.18 (the weak version suffices). Since g is asymptotically contracting
<
along X0 , we then have λ≥
i (Jd(g)) > λi (Jd(g)), hence αi (Jd(g)) > 0.
(ii) Suppose that both g and g−1 are asymptotically contracting along X0 . By the
previous point, we already know that g is X0 -regular. It remains to check that for
any λ 6∈ Ω=
X0 , we have λ(Jd(g)) 6= 0. We distinguish two cases:
• if λ ∈ Ω<
X0 , since g is asymptotically contracting along X0 , we have
λ(Jd(g)) < min λ≥ (Jd(g)) ≤ 0;
λ≥ ∈Ω≥
X
0
−1 is asymptotically contracting along X , we have
• if λ ∈ Ω>
0
X0 , since g
λ(Jd(g)) > max λ≤ (Jd(g)) ≥ 0.
λ≤ ∈Ω≤
X
0
Hence g is ρ-regular.
6.4 Counterexamples
Here we give three examples of “pathological” behavior, to explain why some constructions
of this paper cannot be simplified in general. All three of them are fairly far-fetched. For
the reader who wishes to focus on “non-pathological” behavior, it is probably safe to skip
this subsection.
Example 6.20. Here are two examples of awkward (i.e. neither limited nor abundant)
representations, with an explanation of how they provide a counterexample to (5.8)
and to the “super-strong version” of Lemma 6.18. (The first one is a development on
Example 3.5.4 in [Smi].)
1. Take G = PSp4 (R) (which is a split group). Hence its restricted root system
coincides with its ordinary root system, and is of type C4 . In the notations of
[Kna96], its simple (restricted) roots are e1 − e2 , e2 − e3 , e3 − e4 and 2e4 .
Take ρ to be the representation with highest weight e1 + e2 + e3 + e4 . It has:
48
• the 16 (restricted) weights of the form ±e1 ± e2 ± e3 ± e4 , with multiplicity 1;
• the 24 (restricted) weights of the form ±ei ± ej , with multiplicity 1;
• the zero (restricted) weight with multiplicity 2,
for a total dimension of 42.
For 1 ≤ i ≤ 3, we have αi = ei − ei+1 ∈ Ω and in particular −αi ∈ Ω<
X0 for any
≥
generic X0 , so obviously we may take λi = 0. For α4 = 2e4 on the other hand, we
have −α4 6∈ Ω so this is no longer the case. Now note that there are three different
types of generic elements; extreme representatives of each type are given by:
(a) X0 = (4, 2, 1, 0), with ΠX0 = {α4 };
(b) X0 = (5, 3, 2, 1), with ΠX0 = ∅;
(c) X0 = (4, 3, 2, 0), with ΠX0 = {α4 }.
In cases (a) and (c), the removal of ΠX0 excludes α4 from consideration. In case (b)
however, we have to deal with it. To wit, we then have:
Ω≥
X0 = {e1 + e2 + e3 + e4 ,
e1 + e2 + e3 − e4 ,
e1 + e2 − e3 + e4 ,
e1 + e2 − e3 − e4 ,
e1 − e2 + e3 + e4 ,
e1 − e2 + e3 − e4 ,
e1 − e2 − e3 + e4 ,
− e1 + e2 + e3 + e4 ,
ei ± ej for all i < j,
0},
and the only possible choice is
(
λ≥
4 = e1 − e2 − e3 + e4 ;
λ<
4 = e1 − e2 − e3 − e4 .
(6.11)
(6.12)
2. Take G = SO+ (3, 2) × SO+ (3, 2). Its root system (both ordinary and restricted) is
then of type B2 × B2 . Let us call
• ±e1 , ±e2 , ±e1 ± e2 the (restricted) roots of the first factor;
• ±f1 , ±f2 , ±f1 ± f2 the (restricted) roots of the second factor.
Let us order the (restricted) roots by the lexicographical order on each factor. This
gives a unique ordering of the combined root system. The simple (restricted) roots
of G are then e1 − e2 , e2 , f1 − f2 and f2 .
49
Take ρ to be the representation of G with highest weight e1 +f1 (which corresponds
to the standard action of G on R5 ⊗ R5 ). The set of its (restricted) weights is then
Ω = {−e1 , −e2 , 0, e2 , e1 } + {−f1 , −f2 , 0, f2 , f1 }
(6.13)
(where by A + B we mean {a + b | a ∈ A, b ∈ B}, i.e. the set Ω is of cardinal 25).
The set Ω contains the negative simple (restricted) roots −e2 and −f2 , but not
−e1 + e2 nor −f1 + f2 .
There are six different types of generic elements. Using on a the coordinate system
(e1 , e2 , f1 , f2 ), extreme representatives of three of the types are given by:
(a) X0 = (2, 2, 1, 1). Then ΠX0 = {e1 − e2 , f1 − f2 }: this is a “nice” case.
(b) X0 = (3, 1, 2, 2). Then ΠX0 = {f1 − f2 }, and we need to deal with e1 − e2 .
The two only possibilities are to take
(
λ≥
e1 −e2 = e1 − fi ;
(6.14)
<
λe1 −e2 = e2 − fi
for i = 1 or 2.
(c) X0 = (4, 2, 3, 1). Then ΠX0 = ∅, and we need to deal with both. We have to
take
(
λ≥
e1 −e2 = e1 − f1 ;
(6.15)
<
λe1 −e2 = e2 − f1
and
(
λ≥
f1 −f2 = −e2 + f1 ;
λ<
f1 −f2 = −e2 + f2 .
(6.16)
The three other types are obtained by exchanging e and f .
Example 6.21. Here is a counterexample to the statement of Remark 6.17: a representation ρ (necessarily swinging) and a choice of X0 such that not all elements that
asymptotic contraction along X0 does not, on its own, imply ρ-regularity. (We let the
reader check the details.)
Take G = PSL4 (R); its root system (both ordinary and restricted) is of type A3 ,
and the involution w0 maps (e1 , e2 , e3 , e4 ) to (e4 , e3 , e2 , e1 ) (in the notations of [Kna96],
appendix C).
Let ρ be a representation of G with highest weight
λ = 5̟2 + ̟3 = 4e1 − e2 − e3 − 2e2 ;
this is a representation of dimension 189, with 119 distinct (restricted) weights. Take
X0 = (10, 1, −1, −10);
this vector is generically symmetric and extreme with respect to ρ — in fact, we have
w0
Ω=
= {0, ±(e1 − e2 − e3 + e4 )}.
X0 = Ω
50
Then the vector
Y = (16, 2, −3, −15)
is asymptotically contracting along X0 , i.e.
≥
≥
∀λ< ∈ Ω<
X0 , ∀λ ∈ ΩX0 ,
λ< (Y ) < λ≥ (Y ).
However the (restricted) weight −e1 − e2 + 4e3 − 2e4 vanishes on Y , but not on X0 ; so
Y is not ρ-regular.
7 Properties of ρ-regular affine maps
The goal of this section is to define the affine versions of the remaining properties. It generalizes Subsections 4.3–5.2 in [Smi], but several constructions now become considerably
more complex.
In Subsection 7.1, we define the “ideal dynamical spaces” associated to a ρ-regular
map g ∈ G ⋉ V . The data of two of them is the affine version of the “geometry” of g, and
the remaining ones can be deduced from these two. This generalizes at the same time
Subsections 4.3 and 4.4 from [Smi], but using a different approach.
In Subsection 7.2, we study the action of a ρ-regular map on its affine ideally neutral
&
space Ag , which turns out to be a so-called quasi-translation. This generalizes Subsection 4.5 from [Smi] and Subsection 2.4 from [Smi16], with one small difference: see
Remark 7.9.
In Subsection 7.3, we introduce a groupoid of “canonical identifications” between all the
possible affine ideally neutral spaces. We then use them to define the “translation part” of
the asymptotic dynamics of g (the Margulis invariant). This is an almost straightforward
generalization of Subsection 4.6 from [Smi] and Subsection 2.5 from [Smi16].
In Subsection 7.4, we define and study the affine versions of non-degeneracy and contraction strength. We mostly follow the second half of Subsection 5.1 from [Smi] or of
Subsection 2.6 from [Smi16].
In Subsection 7.5, we study the relationships between affine and linear properties. This
generalizes Subsection 5.2 from [Smi] or Subsection 2.7 from [Smi16], but with a weaker
and more complicated statement.
7.1 Ideal dynamical spaces
The goal of this subsection is to define the ideal dynamical spaces associated to a ρ-regular
map g ∈ G ⋉ V (Definition 6.12). We start with the following particular case:
Definition 7.1. Take any element g ∈ G ⋉ V . We may then write it as
g = τv gh ge gu ,
where τv is a translation by some vector v ∈ V and gh ge gu is the Jordan decomposition
(see Proposition 2.1) of the linear part of g. We say that g is in canonical form if we
have:
51
(i) gh ∈ exp(a+ ) (i.e. gh = exp(Jd(g)));
(ii) v ∈ V0= .
If g is in canonical form, then we define its ideal dynamical spaces to be the reference
dynamical spaces introduced above. This is especially useful when g is X0 -regular, as
shown by the following property:
Proposition 7.2. If a map g ∈ G ⋉ V is in canonical form and is X0 -regular, then it
≤
≥
≤
>
=
=
stabilizes all eight reference dynamical spaces, namely A≥
0 , A0 , A0 , V0 , V0 , V0 , V0
and V0< .
Proof. First of all note that ℓ(g) commutes (by definition) with its hyperbolic part,
which is (because g is in canonical form) equal to exp(Jd(g)). Hence ℓ(g) belongs to the
centralizer of Jd(g), which is LJd(g) . Now since g is X0 -regular, we have LJd(g) ⊂ LX0 .
Finally from Proposition 6.4, it follows that the group LX0 (hence in particular ℓ(g))
stabilizes:
• the spaces V0≥ and V0> , as a subgroup of PX+0 ;
• the spaces V0≤ and V0< , as a subgroup of PX−0 ;
• the space V0= = V0≥ ∩ V0≤ , as a subgroup of both.
Now the action of the affine map g on the subspace V ⊂ V ⊕ R0 coincides with the action
of its linear part ℓ(g); so g also stabilizes these five subspaces. Finally we know (since
g is in canonical form) that v is contained in V0= , hence in V0≥ and in V0≤ ; hence g also
≥
≤
stabilizes A=
0 , A0 and A0 .
For a general g ∈ G ⋉ V , we define the ideal dynamical spaces to be the inverse images
of the reference dynamical spaces by a “canonizing” map φ such that the conjugate φgφ−1
is in canonical form. However, to ensure that the translation part of φ is well-defined,
we need g to be regular with respect to ρ. More precisely:
Proposition 7.3. Let g ∈ G ⋉ V be a ρ-regular map. Then:
(i) there exists a map φ ∈ G ⋉ V , called a canonizing map for g, such that φgφ−1 is
in canonical form;
(ii) any two such maps differ by left-multiplication by an element of LX0 ⋉ V0= .
The key point of the proof is the following lemma:
Lemma 7.4. If g ∈ G ⋉ V is in canonical form and is ρ-regular, then the linear map
ℓ(g) − Id induces an invertible linear map on the quotient space V /V0= .
52
Proof. The fact that the quotient map is well-defined follows from Proposition 7.2: indeed
ℓ(g), hence ℓ(g) − Id, stabilizes the subspace V0= .
Let us now show that the quotient map is invertible. All eigenvalues of the restriction of
the hyperbolic part exp(Jd(g)) to the subspace V0> ⊕ V0< are real and positive, and (since
g is ρ-regular) different from 1. Since the elliptic and unipotent parts of ℓ(g) commute
with the hyperbolic part of ℓ(g) and have all eigenvalues of modulus 1, it follows that
all eigenvalues of the restriction of ℓ(g) to V0> ⊕ V0< (this subspace is stable by ℓ(g) by
Proposition 7.2) are different from 1. In particular the restriction of ℓ(g) − Id to V0> ⊕ V0<
is invertible. The conclusion follows.
Proof of Proposition 7.3.
(i) Let φℓ ∈ G be a canonizing map for ℓ(g), which exists by Proposition 2.1. We then
have
′
φℓ gφ−1
ℓ = τv g ,
where g ′ ∈ G is in canonical form and v ∈ V .
We now claim that for a suitable choice of w ∈ V , the map φ := τw φℓ is a canonizing
map for g. Indeed we then have:
φgφ−1 = τw τv g′ τ−w = τv+w−g′ (w) g′ .
We already know that g′ ∈ G is in canonical form. On the other hand, by
Lemma 7.4 (here we need surjectivity of the quotient map), we may choose w ∈ V
in such a way that
v + w − g ′ (w) ∈ V0= ,
which finishes the proof.
(ii) Assume that g ∈ G ⋉ V is already in canonical form, so that
g = τv ℓ(g)
with v ∈ V0= and ℓ(g) ∈ G. It is enough to show that any φ ∈ G ⋉ V such that
φgφ−1 is still in canonical form is an element of LX0 ⋉ V0= . Indeed, let φ be such
a map; let us write
φ = τw ℓ(φ),
where w ∈ V is its translation part and ℓ(φ) ∈ G is its linear part. By Proposition 2.17, the fact that ℓ(φ) commutes with Jd(g) implies that ℓ(φ) ∈ LJd(g) ⊂ LX0 .
As for the translation part, if we have w − ℓ(g)(w) ∈ V0= , by Lemma 7.4 (here we
need injectivity of the quotient map), we have w ∈ V0= .
Definition 7.5 (Affine version of geometry). For any ρ-regular map g ∈ G ⋉ V , we
introduce the following eight spaces, called ideal dynamical spaces of g:
• Vg≫ := φ−1 (V0> ), the ideally expanding space associated to g;
53
• Vg≪ := φ−1 (V0< ), the ideally contracting space associated to g;
• Vg≈ := φ−1 (V0= ), the ideally neutral space associated to g;
&
• Vg := φ−1 (V0≥ ), the ideally noncontracting space associated to g;
.
• Vg := φ−1 (V0≤ ), the ideally nonexpanding space associated to g;
&
• Ag := φ−1 (A≥
0 ), the affine ideally noncontracting space associated to g;
• A.g := φ−1 (A≤
0 ), the affine ideally nonexpanding space associated to g;
−1
=
• A≈
g := φ (A0 ), the affine ideally neutral space associated to g,
where φ is any canonizing map of g.
Here is the idea behind this definition. Suppose first that the representation is nonswinging (so that X0 is actually generic), and that the Jordan projection of g is sufficiently
>
close to X0 . Then it actually has the same type as X0 , i.e. we have Ω>
Jd(g) = ΩX0 and
=
<
similarly for Ω , Ω . Whenever that happens, the ideal dynamical spaces of g coincide
with its actual dynamical spaces, as defined in section 4.3 of [Smi].
If X0 is not generic, this is no longer true as such. We still want to assume that Jd(g)
is sufficiently close to X0 ; but now, this can at best ensure that both g and g−1 are
asymptotically contracting along X0 . In that case, we only get that:
• the moduli of the eigenvalues of g|Vg≫ are “much larger” than 1;
• the moduli of the eigenvalues of g|Vg≪ are “much smaller” than 1;
might now differ from 1, but somehow remain
• the moduli of the eigenvalues of g|A≈
g
“not too far” from 1.
Let us finally check that this definition makes sense, and prove a few extra properties
along the way.
Proposition 7.6.
(i) The definitions above do not depend on the choice of φ;
&
&
.
.
(ii) the datum of Ag uniquely determines the spaces Vg and Vg≫ ;
(iii) the datum of Ag uniquely determines the spaces Vg and Vg≪ ;
&
(iv) the data of both Ag and A.g uniquely determine all eight ideal dynamical spaces.
&
.
The spaces Ag and Ag will play a crucial role, as they are the affine analogs of the
attracting and repelling flags ygX0 ,+ and ygX0 ,− defined in section 4.1: see Remark 7.15
below for an explanation.
54
Proof. An immediate corollary of Proposition 6.4 is that we have
≥
+
StabG⋉V (A≥
0 ) = PX0 ⋉ V0
StabG⋉V (V0≥ ) = PX+0 ⋉ V
StabG⋉V (V0> ) = PX+0 ⋉ V
≤
−
StabG⋉V (A≤
0 ) = PX0 ⋉ V0
StabG⋉V (V0≤ ) = PX−0 ⋉ V
StabG⋉V (V0< ) = PX−0 ⋉ V
=
StabG⋉V (A=
0 ) = L X 0 ⋉ V0
StabG⋉V (V0= ) = LX0 ⋉ V,
(7.1)
with the following relationships:
PX+0 ⋉ V0≥
⊂
PX+0 ⋉ V
=
PX+0 ⋉ V
PX−0 ⋉ V0≤
⊂
PX−0 ⋉ V
=
PX−0 ⋉ V
LX0 ⋉ V0=
⊂
LX0 ⋉ V.
The first two lines immediately imply (ii) and (iii).
For points (i) and (iv), note that all eight groups contain LX0 ⋉ V0= . Point (i) follows
by Proposition 7.3. Point (iv) follows using the identity
(PX+0 ⋉ V0≥ ) ∩ (PX−0 ⋉ V0≤ ) = LX0 ⋉ V0= .
7.2 Quasi-translations
Let us now investigate the action of a ρ-regular map g ∈ G ⋉ V on its affine ideally
neutral space A≈
g . The goal of this subsection is to prove that it is “almost” a translation
(Proposition 7.10).
We fix on V a Euclidean form B satisfying the conditions of Lemma 2.4 for the representation ρ.
Definition 7.7. We call quasi-translation any affine automorphism of A=
0 induced by an
element of L ⋉ V0= .
Let us explain and justify this terminology.
Proposition 7.8. Let V0t be the set of fixed points of L in V0= :
V0t := {v ∈ V0= | ∀l ∈ L, lv = v} .
Let V0a be the B-orthogonal complement of V0t in V0= . Then any quasi-translation is an
element of
GL(V0a ) ⋉ V0a × V0t .
In other words, quasi-translations are affine automorphisms of V0= that preserve the
directions of V0a and V0t and act only by translation on the V0t component. You may
think of a quasi-translation as a kind of “screw displacement”; the superscripts t and a
respectively stand for “translation” and “affine”.
Proof. We need to show that every element of L fixes V0t (pointwise) and leaves V0a
invariant (globally). The former is true by definition of V0t . For the latter, let us prove
it separately for elements of M and elements of A:
55
• By hypothesis, elements of M preserve the form B; since they leave invariant the
space V0t , they also leave invariant its B-orthogonal complement V0a .
• Let us introduce the notation
M
V0≎ :=
Vλ
(7.2)
λ∈Ω=
X \{0}
0
(the symbol “≎” is intended to represent the idea of “avoiding zero”), so that the
space V0= decomposes into the orthogonal sum
V0= = V 0 ⊕ V0≎ .
(7.3)
Then since V t ⊂ V 0 (obviously), similarly, we have an orthogonal sum
V0a = (V0a ∩ V 0 ) ⊕ V0≎ .
(7.4)
Now clearly V0≎ , being a sum of restricted weight spaces, is invariant by A. Moreover, every element of A actually fixes every element of V 0 , in particular leaves
invariant the subspace V0a ∩ V 0 . The conclusion follows.
Remark 7.9. Note that in contrast to the non-swinging case (Proposition 4.19 in [Smi]),
quasi-translations no longer have to act by isometries on V0a , as this space now comprises
the possibly nontrivial space V0≎ where A acts nontrivially.
This phenomenon is the reason why the reasoning we used in [Smi] to prove additivity
of Margulis invariants could not be reused here without major restructuring.
We now claim that any ρ-regular map acts on its affine ideally neutral space by quasitranslations:
Proposition 7.10. Let g ∈ G⋉V be a ρ-regular map, and let φ ∈ G⋉V be any canonizing
map for g. Then the restriction of the conjugate φgφ−1 to A=
0 is a quasi-translation.
Let us actually formulate an even more general result, which will have another application in the next subsection:
≤
=
Lemma 7.11. Any map f ∈ G ⋉ V stabilizing both A≥
0 and A0 acts on A0 by quasitranslation.
Proof.
−
=
• We begin by showing that any element of lX0 = p+
X0 ∩ pX0 acts on V0 in the same
way as some element of l. Recall that by definition
M
lX0 = l ⊕
gα
α(X0 )=0
and
V0= =
M
λ(X0 )=0
56
V λ.
Thus we want to show that for every restricted root α and restricted weight λ such
that α(X0 ) = λ(X0 ) = 0, we have gα · V λ = 0. It is sufficient to show that in
such a case, the sum λ + α is no longer a restricted weight. But otherwise, both λ
and λ + α would be elements of Ω=
X0 , hence they would both be fixed by w0 since
X0 is generically symmetric. This would mean that α is also fixed by w0 , which is
impossible.
• We now conclude in the same fashion as in the proof of Lemma 4.21 in [Smi], passing
±
±
from the two Lie algebras p±
X0 first to the identity components PX0 ,e of PX0 , then
to the whole groups PX±0 , then (using Proposition 6.4) to the stabilizers of A≥
0 and
≤
of A0 .
Proof of Proposition 7.10. The Proposition follows immediately by taking f = φgφ−1 .
≤
Indeed, by Proposition 7.2, the “canonized” map φgφ−1 stabilizes A≥
0 and A0 .
See Example 4.22 in [Smi] for specific examples of quasi-translations (in the nonswinging case).
We would like to treat quasi-translations a bit like translations; for this, we need to
have at least a nontrivial space V0t . So from now on, we impose the following condition
on ρ:
Assumption 7.12. The representation ρ is such that
dim V0t > 0.
This is precisely condition (i)(a) from the Main Theorem.
7.3 Canonical identifications and the Margulis invariant
The main goal of this subsection is to associate to every ρ-regular map g ∈ G ⋉ V a
vector in V0t , called its “Margulis invariant” (see Definition 7.19). The two Propositions
and the Lemma that lead up to this definition are important as well, and will be often
used subsequently.
Proposition 7.6 (iv) has shown us that the “geometry” of a ρ-regular map g (namely
the position of its ideal dynamical spaces) is entirely determined by the pair of spaces
≤
(A&g , A.g ) = φ(A≥
0 , A0 ).
In fact, such pairs of spaces play a crucial role. Let us begin with a definition; its
connection with the observation we just made will become clear after Proposition 7.14.
Definition 7.13.
• We define a parabolic space to be any subspace of V that is the image of either V0≥
or V0≤ (no matter which one, since X0 is symmetric) by some element of G.
57
• We define an affine parabolic space to be any subspace of A that is the image of A≥
0
by some element of G ⋉ V . Equivalently, a subspace A& ⊂ A is an affine parabolic
space iff it is not contained in V and its linear part V & = A& ∩ V is a parabolic
space.
• We say that two parabolic spaces (or two affine parabolic spaces) are transverse if
their intersection has the lowest possible dimension, or equivalently if their sum is
the whole space V (or A).
See Example 4.25 in [Smi].
Proposition 7.14. A pair of parabolic spaces (resp. of affine parabolic spaces) is trans≤
verse if and only if it may be sent to (V0≥ , V0≤ ) (resp. to (A≥
0 , A0 )) by some element of G
(resp. of G ⋉ V ).
&
In particular, for any ρ-regular map g ∈ G ⋉ V , the pair (Ag , A.g ) is by definition a
transverse pair of affine parabolic spaces.
This Proposition, as well as its proof, is very similar to Claim 2.8 in [Smi16] and to
Proposition 4.26 in [Smi].
Proof. Let us prove the linear version; the affine version follows immediately. Let (V1 , V2 )
be any pair of parabolic spaces. By definition, for i = 1, 2, we may write Vi = φi (V0≥ ) for
some φi ∈ G. Let us apply the Bruhat decomposition to the map φ−1
1 φ2 : we may write
φ−1
1 φ2 = p1 wp2 ,
(7.5)
where p1 , p2 belong to the minimal parabolic subgroup P + , and w is an element of the
restricted Weyl group W (or, technically, some representative thereof). Let φ := φ1 p1 =
−1
+ stabilizes V ≥ , we have
φ2 p−1
2 w ; since P
0
V1 = φ(V0≥ ) and V2 = φ(wV0≥ ).
(7.6)
Thus we have
V1 and V2 are transverse ⇐⇒ wV0≥ is transverse to V0≥
≥
⇐⇒ Ω≥
X0 ∪ wΩX0 = Ω
≥
⇐⇒ wΩ<
X0 ⊂ Ω X0
≥
⇐⇒ ww0 Ω>
X0 ⊂ Ω X0 .
(7.7)
By Lemma 6.5, this implies that ww0 stabilizes Ω>
X0 . Hence it stabilizes its comple≥
≤
ment Ω≤
,
which
means
that
wV
=
V
.
Thus
we
have
V1 = φ(V0≥ ) and V2 = φ(V0≤ ) as
0
0
X0
required.
Conversely, if those equalities hold, then V1 and V2 are obviously transverse.
Remark 7.15.
58
• It follows from Proposition 6.4 that the set of all parabolic spaces can be identified
with the flag variety G/PX+0 , by identifying every parabolic space φ(V0≥ ) with the
coset φPX+0 . For every X0 -regular element g ∈ G, this identification then matches
the ideally expanding space Vg≫ with the attracting flag ygX0 ,+ .
• Composing with the bijection ιX defined in Proposition 4.14, we may also identify
this set with the opposite flag variety G/PX−0 . For X0 -regular g ∈ G, this matches
the ideally contracting space Vg≪ with the repelling flag ygX0 ,− .
• Using the Bruhat decomposition of PX0 (see Corollary 2.24), we may then show
that two parabolic spaces V1 = φ1 (V0≥ ) and V2 = φ2 (V0≥ ) = φ2 ◦ w0 (V0≤ ) are
transverse if and only if the corresponding pair of cosets
(φ1 PX+0 , φ2 w0 PX−0 )
is transverse in the sense of Definition 4.5.
• Similarly, it follows from (7.1) that we can in principle identify the set of all affine
parabolic spaces with the “affine flag variety” G⋉ V /PX+0 ⋉ V0≥ . This would however
require very cumbersome notations.
In the linear case, it was natural to do it in order to make things representationindependent. In the affine case however, there is a “privileged” representation
anyway (namely ρ). We decided that translating everything into that “abstract”
language was not worth the trouble, but the reader may go through that exercise
if they wish so.
Consider a transverse pair of affine parabolic spaces. Their intersection may be seen
as a sort of “abstract affine neutral space”. We now introduce a family of “canonical
identifications” between those spaces. These identifications have however an inherent
ambiguity: they are only defined up to quasi-translation.
Proposition 7.16. Let (A1 , A2 ) be a pair of transverse affine parabolic spaces. Then any
≤
map φ ∈ G ⋉ V such that φ(A1 , A2 ) = (A≥
0 , A0 ) gives, by restriction, an identification of
=
the intersection A1 ∩ A2 with A0 , which is unique up to quasi-translation.
Here by φ(A1 , A2 ) we mean the pair (φ(A1 ), φ(A2 )). This generalizes Corollary 2.14
in [Smi16] and Proposition 4.28 in [Smi].
Remark 7.17. Note that if A1 ∩ A2 is obtained in another way as an intersection of
two affine parabolic spaces, the identification with A=
0 will, in general, no longer be the
same, not even up to quasi-translation: there could also be an element of the Weyl group
involved.
Proof. The existence of such a map φ follows from Proposition 7.14. Now let φ and φ′
be two such maps, and let f be the map such that
φ′ = f ◦ φ
59
(7.8)
≤
(i.e. f := φ′ ◦ φ−1 ). Then by construction f stabilizes both A≥
0 and A0 . It follows from
Lemma 7.11 that the restriction of f to A=
0 is a quasi-translation.
Let us now explain why we call these identifications “canonical”. The following lemma,
while seemingly technical, is actually crucial: it tells us that the identifications defined in
Proposition 7.16 commute (up to quasi-translation) with the projections that naturally
arise if we change one of the parabolic subspaces in the pair while fixing the other.
Lemma 7.18. Take any affine parabolic space A1 .
Let A2 and A′2 be any two affine parabolic spaces both transverse to A1 .
Let φ (resp. φ′ ) be an element of G ⋉ V that sends the pair (A1 , A2 ) (resp. (A1 , A′2 ))
≤
to (A≥
0 , A0 ); these two maps exist by Proposition 7.14.
Let W1 be inverse image of V0> by any map φ such that A1 = φ−1 (A≥
0 ) (this image is
unique by Proposition 6.4).
Let
ψ : A1 −−−−→ A1 ∩ A′2
be the projection parallel to W1 .
Then the map ψ defined by the commutative diagram
A=
0
ψ
φ′
φ
A1 ∩ A2
A=
0
ψ
A1 ∩ A′2
is a quasi-translation.
The space W1 is, in a sense, the “abstract linear expanding space” corresponding to the
“abstract affine noncontracting space” A1 : more precisely, for any ρ-regular map g ∈ G⋉V
&
such that Ag = A1 , we have Vg≫ = W1 .
≥
≤
>
>
=
The projection ψ is well-defined because A≥
0 = V0 ⊕ A0 = V0 ⊕ (A0 ∩ A0 ), and so
≥
A1 = φ′−1 (A0 ) = W1 ⊕ (A1 ∩ A′2 ).
This statement generalizes Lemma 2.18 in [Smi16] and Lemma 4.29 in [Smi].
Proof. The proof is exactly the same as the proof of Lemma 4.29 in [Smi].
Now let g be a ρ-regular map. We already know that it acts on its affine ideally neutral
space by quasi-translation; now the canonical identifications we have just introduced
allow us to compare the actions of different elements on their respective affine ideally
neutral spaces, as if they were both acting on the same space A=
0 . However there is
a catch: since the identifications are only canonical up to quasi-translation, we lose
information about what happens in V0a ; only the translation part along V0t remains.
Formally, we make the following definition. Let πt denote the projection from V0=
onto V0t parallel to V0a .
60
Definition 7.19. Let g ∈ G ⋉ V be a ρ-regular map. Take any point x in the affine
space A≈
g ∩ VAff and any map φ ∈ G that canonizes ℓ(g). Then we define the Margulis
invariant of g to be the vector
M (g) := πt (φ(g(x) − x)) ∈ V0t .
We call it the Margulis invariant of g.
This vector does not depend on the choice of x or φ: indeed, composing φ with a
quasi-translation does not change the V0t -component of the image. See Proposition 2.16
in [Smi16] for a detailed proof of this claim (for V = g).
Informally, the Margulis invariant gives the “translation part” of the asymptotic dynamics of an R-regular element g ∈ G ⋉ V (the “linear part” being given by Jd(g), just
as in the linear case). As such, it plays a central role in this paper.
7.4 Quantitative properties
In this subsection, we define and study the affine versions of non-degeneracy and contraction strength. We more or less follow the second half of Subsection 5.1 in [Smi], or
of Subsection 2.6 in [Smi16].
We endow the extended affine space A with a Euclidean norm (written simply k • k)
given by
∀(x, t) ∈ V ⊕ R0 = A,
k(x, t)k2 := B(x, x) + |t|2 ,
where B is the norm defined in Lemma 2.4. Then the subspaces V0> , V0< , V0≎ , V0a ∩ V 0 ,
V0t and R0 are pairwise orthogonal.
Definition 7.20 (Affine version of non-degeneracy). Take a pair of affine parabolic
spaces (A1 , A2 ). An optimal canonizing map for this pair is a map φ ∈ G ⋉ V satisfying
≤
φ(A1 , A2 ) = (A≥
0 , A0 )
and minimizing the quantity max kφk, kφ−1 k . By Proposition 7.14 and a compactness
argument, such a map exists iff A1 and A2 are transverse.
We define an optimal canonizing map for a ρ-regular map g ∈ G ⋉ V to be an optimal
canonizing map for the pair (A&g , A.g ).
Let C ≥ 1. We say that a pair of affine parabolic spaces (A1 , A2 ) (resp. a ρ-regular
map g) is C-non-degenerate if it has a canonizing map φ such that
kφk ≤ C
and
φ−1 ≤ C.
Now take g1 , g2 two ρ-regular maps in G ⋉ V . We say that the pair (g1 , g2 ) is C-non&
.
degenerate if every one of the four possible pairs (Agi , Agj ) is C-non-degenerate.
The point of this definition is that there are a lot of calculations in which, when we
treat a C-non-degenerate pair of spaces as if they were perpendicular, we err by no more
than a (multiplicative) constant depending on C.
61
Remark 7.21. The set of transverse pairs of extended affine spaces is characterized by two
open conditions: there is of course transversality of the spaces, but also the requirement
that each space not be contained in V . What we mean here by “degeneracy” is failure of
one of these two conditions. Thus the property of a pair (A1 , A2 ) being C-non-degenerate
actually encompasses two properties.
First, it implies that the spaces A1 and A2 are transverse in a quantitative way. More
precisely, this means that some continuous function that would vanish if the spaces were
not transversely is bounded below. An example of such a function is the smallest non
identically vanishing of the “principal angles” defined in the proof of Lemma 8.2 (iv).
Second, it implies that both A1 and A2 are “not too close” to the space V (in the same
sense). In purely affine terms, this means that the affine spaces A1 ∩ VAff and A2 ∩ VAff
contain points that are not too far from the origin.
Both conditions are necessary, and appeared in the previous literature (such as [Mar87]
and [AMS02]). However, they were initially treated separately. The idea of encompassing
both in the same concept of “C-non-degeneracy” seems to have been first introduced in
the author’s previous paper [Smi16].
Definition 7.22 (Affine version of contraction strength). Let s > 0. For a ρ-regular
map g ∈ G ⋉ V , we say that g is s-contracting along X0 if we have:
∀(x, y) ∈ Vg≪ × A&g ,
kg(y)k
kg(x)k
≤s
.
kxk
kyk
(7.9)
&
(Note that by definition the spaces Vg≪ and Ag always have the same dimensions as V0<
and A≥
0 respectively, hence they are nonzero.)
We define the affine contraction strength along X0 of g to be the smallest number
sX0 (g) such that g is sX0 (g)-contracting along X0 . In other words, we have
sX0 (g) = g|Vg≪
g−1
&
Ag
.
(7.10)
This notion is closely related to the notion of asymptotic contraction:
Proposition 7.23.
(i) If a map g ∈ G ⋉ V is ρ-regular and 1-contracting along X0 (i.e. such that
sX0 (g) < 1), then it is also asymptotically contracting along X0 .
(ii) If a map g ∈ G ⋉ V is ρ-regular and asymptotically contracting along X0 , then we
have
lim sX0 (gN ) = 0.
N →∞
Proof. Let g ∈ G ⋉ V be ρ-regular. We claim that it is asymptotically contracting
along X0 if and only if it satisfies the inequality
−1
(7.11)
r g|Vg≪ r g A& < 1
g
62
(where r(f ) denotes the spectral radius of f ). Indeed, as a porism of Proposition 2.6 (i),
we obtain that the spectrum of g|Vg≪ is
n
o
eλ(Jd(g)) λ ∈ Ω<
X0 ,
and the spectrum of g|A& is
g
n
o
eλ(Jd(g)) λ ∈ Ω≥
X0 ∪ {1}
(the “1” accounts for the affine extension). By Assumption 5.5, the eigenvalue 1 is already
contained in the spectrum of the linear part, so we may actually ignore the “∪{1}” part.
The claim follows.
The conclusion then follows from the well-known facts that for every linear map f , we
have
r(f ) ≤ kf k
(7.12)
(this gives (i)), and
log kf N k = N log r(f ) + O (log N )
(7.13)
N →∞
(also known as Gelfand’s formula; this gives (ii)).
7.5 Comparison of affine and linear properties
The goal of this subsection is to prove Proposition 7.26, that, for any ρ-regular element g ∈ G ⋉ V , relates the quantitative properties we just introduced to the corresponding properties of its linear part ℓ(g).
This is given by Lemma 2.25 in [Smi16] in the case of the adjoint representation, and
by Lemma 5.8 in [Smi] (which is a straightforward generalization) in the non-swinging
case. In the general case however, only points (i) and (ii) generalize in the obvious way.
The statement (iii) holds only in a weaker form: we now need to consider the linear
contraction strength of ℓ(g), but the affine contraction strength of g. This is basically
what forced us to develop the “purely linear theory” (section 4) in a systematic way,
rather than presenting it as a particular case of the “affine theory” as we did in the
previous papers.
In order to be able to compare the affine contraction strength with the linear one,
we begin by expressing the former in terms of the Cartan projection. The following is
a generalization of Lemma 6.5 in [Smi], formulated in a slightly more general way (the
original statement was essentially the same inequality without the absolute value.)
Proposition 7.24. For every C ≥ 1, there is a constant ε7.24 (C) with the following
property. Let g ∈ G be a C-non-degenerate ρ-regular map. Then we have
λ(Ct(g))
− log sX0 (g) − min λ(Ct(g)) − max
<
λ∈Ω≥
X
λ∈ΩX
0
≤ ε7.24 (C).
0
(Recall that ΩX0 is the set of restricted weights that take nonnegative values on X0 ,
and Ω<
X0 is its complement in Ω.)
≥
63
To make sense of this estimate, keep in mind that the quantity − log sX0 (g) is typically
positive. Also note that the “minimum” term is certainly nonpositive, as 0 ∈ Ω≥
X0 .
Proof. First of all let φ be an optimal canonizing map for g, and let g′ = φgφ−1 . Then
it is easy to see that we have
sX0 (g′ ) ≍C sX0 (g),
(7.14)
and the difference Ct(g′ ) − Ct(g) is bounded by a constant that depends only on C.
Hence we lose no generality in replacing g by g ′ . Clearly, it is enough to show that for g ′ ,
which is in canonical form, we have the equality:
min λ(Ct(g′ )) − max
λ(Ct(g′ )) = − log sX0 (g′ ).
<
λ∈Ω≥
X
0
λ∈ΩX
(7.15)
0
This is the straightforward generalization of (6.2) in [Smi], and is proved in exactly
the same fashion, mutatis mutandis.
As a stepping stone, we first need the “naïve” extension of point (ii) of Lemma 2.25
in [Smi16] or Lemma 5.8 in [Smi], giving a bound on the affine contraction strength of
the linear part of g (seen as an element of G ⋉ V by the usual embedding).
Lemma 7.25. For any ρ-regular map g ∈ G ⋉ V , we have sX0 (ℓ(g)) ≤ sX0 (g).
The proof is very similar to the one given in [Smi16].
Proof. We have
ℓ(g)−1
max g −1
max g −1
sX0 (ℓ(g)) = ℓ(g)|V ≪
ℓ(g)
= g|Vg≪
≤ g|Vg≪
= sX0 (g).
&
Aℓ(g)
&
Vg
&
Ag
,1
,1
To justify the last equality, note that if φ is a canonizing map for g, then the space
&
−1
0
Ag = φ−1 (A≥
0 ) contains the subspace φ (V ), which is nonzero by Assumption 5.5, is
−1
clearly g-invariant, and all eigenvalues of g restricted to that subspace have modulus 1;
hence g−1 A& ≥ g−1 φ−1 (V 0 ) ≥ 1.
g
We may now state and prove the appropriate generalization of Lemma 2.25 in [Smi16]
and Lemma 5.8 in [Smi], linking in all three points the purely linear properties with the
affine properties.
Proposition 7.26. For every C ≥ 1, there is a positive constant s7.26 (C) with the
following property. Let (g, h) be a C-non-degenerate pair of ρ-regular elements of G ⋉ V .
In this case:
64
(i) the pair (ℓ(g), ℓ(h)) is C ′ -non-degenerate (in the sense of Definition 4.10), for some
constant C ′ that depends only on C;
(ii) we have
~sX0 (ℓ(g)) .C sX0 (g);
(iii) moreover, if we assume that sX0 (g−1 ) ≤ s7.26 (C), then we actually have
~sX0 (ℓ(g)) g−1
.C sX0 (g).
A≈
g
Proof.
(i) By Remark 4.11, we lose no generality in specifying some particular form for the
map ν introduced in Definition 4.9. Let us define ν in a way which is consistent
with the definition of C-non-degeneracy in the affine case, namely using a particular
case of the formula (4.1):
ν(φ) := max(kρ(φ)k, kρ(φ)−1 k),
(7.16)
where ρ is our working representation, and k•k is the Euclidean norm we introduced
in the beginning of section 7.4 (or more precisely its restriction to V , which is just
the Euclidean norm introduced in Lemma 2.4). Clearly if φ is a canonizing map
for g, then ℓ(φ) is a canonizing map for ℓ(g). Since obviously for any φ ∈ G, we
have kℓ(φ)k ≤ kφk, the conclusion follows.
(ii) First note that by Lemma 7.25, we have
sX0 (ℓ(g)) ≤ sX0 (g).
(7.17)
Now apply Proposition 7.24: passing to the exponential, we have
!
sX0 (ℓ(g)) ≍C exp
max λ(Ct(ℓ(g))) exp − min λ(Ct(ℓ(g))) .
λ∈Ω<
X
λ∈Ω≥
X
0
0
(7.18)
<
Now the key point is Lemma 6.18 (i): for every i ∈ Π \ ΠX0 , we may find λ<
i ∈ Ω X0
≥
and λ≥
i ∈ ΩX0 whose difference is αi . It immediately follows that
exp
!
max λ(Ct(ℓ(g))) exp − min λ(Ct(ℓ(g))) ≥ ~sX0 (ℓ(g)).
λ∈Ω<
X
0
λ∈Ω≥
X
0
(7.19)
(iii) We proceed in two steps: we establish first (7.21) (which is straightforward), then
(7.26) (which relies on the strong version (ii) of Lemma 6.18 — in fact, this is the
only place where the strong version is needed).
65
• We have, by definition:
g−1
sX0 (g) = g|Vg≪
&
Ag
.
Let φ be an optimal canonizing map for g. Since g is C-non-degenerate (and
>
≫
=
the images φ(A≈
g ) and φ(Vg ), being respectively equal to A0 and V0 , are
orthogonal), it follows that
sX0 (g) ≍C g|Vg≪ max g−1 A≈ , g−1 V ≫ .
g
g
Clearly we have
g−1
sX0 (g−1 ) ≤ 1, we have
A≈
g
g−1
≥
g−1
φ−1 (V 0 )
= 1. On the other hand, since
≤ 1. It follows that
Vg≫
sX0 (g) ≍C
g−1
g|Vg≪
A≈
g
(7.20)
.
By methods similar to the proof of Proposition 7.24 (essentially by Proposition 2.6), we may rewrite this as
!
sX0 (g) ≍C exp
λ(Ct(ℓ(g)))
max
<
λ∈ΩX
g−1
A≈
g
.
(7.21)
0
• Note that, combining Proposition 7.24 with Lemma 7.25, we have
!
exp max λ(Ct(ℓ(g))) exp − min λ(Ct(ℓ(g))) .C sX0 (ℓ(g)−1 )
λ∈Ω≤
X
λ∈Ω>
X
0
0
≤ sX0 (g−1 ).
(7.22)
If we take s7.26 (C) equal to the inverse of the implicit constant in that inequality, we may assume that
>
>
∀λ≤ ∈ Ω≤
X0 , ∀λ ∈ ΩX0 ,
λ≤ (Ct(ℓ(g))) ≤ λ> (Ct(ℓ(g)));
(7.23)
in particular, since 0 ∈ Ω≤
X0 , we then have
∀λ ∈ Ω>
X0 ,
λ(Ct(ℓ(g))) ≥ 0.
(7.24)
Obviously this remains true for λ ∈ Ω>
X0 ∪ {0}. Now take some i ∈ Π \ ΠX0 ;
then we have
≥
−αi (Ct(ℓ(g)) = λ<
i (Ct(ℓ(g))) − λi (Ct(ℓ(g)))
≤ λ<
i (Ct(ℓ(g))),
66
(7.25)
>
<
<
where λ≥
i ∈ ΩX0 ∪ {0} and λi ∈ ΩX0 are the restricted weights introduced in
Lemma 6.18 (ii). This implies that
!
~sX0 (ℓ(g)) := exp
≤ exp
max
α∈Π\ΠX0
−α(Ct(ℓ(g)))
!
max λ(Ct(ℓ(g))) .
λ∈Ω<
X
0
(7.26)
Combining the two estimates, the conclusion follows.
8 Products of ρ-regular maps
The goal of this section is to prove Proposition 8.4, which is the “main part” of the affine
version of Schema 1.5. The general strategy is the same as in Section 7 in [Smi] or in
Section 3.2 in [Smi16]: we reduce the problem to Proposition 3.5, by considering the
action of G ⋉ V on a suitable exterior power Λp A (rather than on the spaces Vi as in
Section 4.3).
We start by proving the following result, whose role in [Smi] was played by Proposition 6.14. Even though the new version involves slightly different inequalities, the proof
is quite similar.
Proposition 8.1. For every C ≥ 1, there is a positive constant s8.1 (C) ≤ 1 with the
following property. Take any C-non-degenerate pair (g, h) of ρ-regular maps in G ⋉ V
such that sX0 (g) ≤ s8.1 (C) and sX0 (h) ≤ s8.1 (C). Then gh is asymptotically contracting
along X0 .
Proof. Let C ≥ 1, and let (g, h) be a C-non-degenerate pair of ρ-regular maps in G ⋉ V ,
such that
sX0 (g) ≤ s8.1 (C) and sX0 (h) ≤ s8.1 (C)
for some constant s8.1 (C) to be specified later.
The first thing to note is that since the property of being ρ-regular depends only on
the linear part, Lemma 7.25 and Proposition 7.26 (i) reduce the problem to the case
where g, h ∈ G.
Now Proposition 7.24 gives us
max λ(Ct(g)) − min λ(Ct(g)) ≤ log sX0 (g) + ε7.24 (C).
λ∈Ω<
X
λ∈Ω≥
X
0
(8.1)
0
Taking s8.1 (C) small enough, we may assume that
≥
≥
<
<
∀λ ∈ ΩX0 , ∀λ ∈ ΩX0 ,
(λ − λ )(Ct(g)) > max kλk ε4.25 (C).
≥
<
λ∈Ω
67
(8.2)
Of course a similar estimate holds for h:
≥
≥
<
<
∀λ ∈ ΩX0 , ∀λ ∈ ΩX0 ,
≥
<
(λ − λ )(Ct(h)) >
max kλk ε4.25 (C).
λ∈Ω
(8.3)
Let Ct′ (g, h) be the vector defined in Corollary 4.25. Then we deduce from (4.22) that
<
<
for every pair of restricted weights λ≥ ∈ Ω≥
X0 and λ ∈ ΩX0 , we have:
|(λ≥ − λ< )(Ct′ (g, h) − Ct(g) − Ct(h))| ≤ kλ≥ − λ< kk Ct′ (g, h) − Ct(g) − Ct(h)k
≤ 2 max kλk ε4.25 (C).
(8.4)
λ∈Ω
Adding together the three estimates (8.2), (8.3) and (8.4), we find that for every such
pair, we have
(λ≥ − λ< )(Ct′ (g, h)) > 0.
(8.5)
On the other hand, we have (4.21) which says that
Jd(gh) ∈ Conv(WX0 · Ct′ (g, h)).
Now take any w ∈ WX0 . From Proposition 6.4, it then follows that w−1 stabilizes
<
both Ω≥
X0 and ΩX0 ; hence we still have
(λ≥ − λ< )(w(Ct′ (g, h)) = (w−1 (λ≥ ) − w−1 (λ< ))(Ct′ (g, h))
(8.6)
> 0.
Thus the difference λ≥ −λ< takes positive values on every point of the orbit WX0 ·Ct′ (g, h);
hence it also takes positive values on every point of its convex hull. In particular, we
have
(λ≥ − λ< )(Jd(gh)) > 0.
(8.7)
We conclude that gh is indeed asymptotically contracting along X0 .
We now establish the correspondence between affine and proximal properties. We
introduce the integers:
≥
p := dim A≥
0 = dim V0 + 1;
q := dim V0< ;
(8.8)
d := dim A = dim V + 1 = q + p.
For every g ∈ G ⋉ V , we may define its exterior power Λp g : Λp A → Λp A. The Euclidean
structure of A induces in a canonical way a Euclidean structure on Λp A.
Lemma 8.2.
(i) Let g ∈ G ⋉ V be a map asymptotically contracting along X0 . Then Λp g is prox&
imal, and the attracting (resp. repelling) space of Λp g depends on nothing but Ag
≪
(resp. Vg ):
(
&
EΛs p g = Λp Ag
EΛup g = x ∈ Λp A x ∧ Λq Vg≪ = 0 .
68
(ii) For every C ≥ 1, whenever (g1 , g2 ) is a C-non-degenerate pair of ρ-regular maps
that are also asymptotically contracting along X0 , (Λp g1 , Λp g2 ) is a C p -non-degenerate
pair of proximal maps.
(iii) For every C ≥ 1, there is a constant s8.2 (C) < 1 with the following property.
For every C-non-degenerate ρ-regular map g ∈ G ⋉ V that is also asymptotically
contracting along X0 , we have
sX0 (g) .C s̃(Λp g).
(8.9)
If in addition sX0 (g) ≤ s8.2 (C), we have
sX0 (g) ≍C s̃(Λp g).
(8.10)
(Recall that s̃ and sX0 stand respectively for the proximal and affine contraction
strengths: see Definitions 7.22 and 3.4.)
(iv) For any two p-dimensional subspaces A1 and A2 of A, we have
αHaus (A1 , A2 ) ≍ α (Λp A1 , Λp A2 ) .
This is similar to Lemma 3.8 in [Smi16] and to Lemma 7.2 in [Smi]; but now, there is the
additional complication of needing to distinguish between ρ-regularity and asymptotic
contraction.
The proofs of points (ii) and (iii) however still remain very similar to the corresponding
proofs in [Smi16]. We chose to reproduce them here, in particular in order to correct a
small mistake: in [Smi16] we erroneously claimed that we could take s8.2 (C) = 1, which
stemmed from a confusion between g and its canonized version g ′ .
Proof.
(i) Let g ∈ G⋉V be a map asymptotically contracting along X0 . From Proposition 2.6
(as already noted in the proof of Proposition 7.23), it follows that
−1
r g|Vg≫ < r g−1 A&
,
(8.11)
g
i.e. every eigenvalue of g|Vg≫ is smaller in modulus than every eigenvalue of g|A& .
g
Let λ1 , . . . , λd be the eigenvalues of g (acting on A), counted with multiplicity and
ordered by nondecreasing modulus; we then have
−1
|λq | = r g|Vg≫ < r g−1 &
= |λq+1 |.
(8.12)
Ag
On the other hand, we know that the eigenvalues of Λp g counted with multiplicity
are exactly the products of the form λi1 · · · λip , where 1 ≤ i1 < · · · < ip ≤ d. As
the two largest of them (by modulus) are λq+1 · · · λd and λq λq+2 · · · λd , it follows
that Λp g is proximal.
As for the expression of E s and E u , it follows immediately by considering a basis
that trigonalizes g.
69
(ii) Take any pair (i, j) ∈ {1, 2}2 . Let φ be an optimal canonizing map for the pair
&
&
<
≪
(Agi , A.gj ). Then we have φ(Agi ) = A≥
0 and (by Proposition 7.6) φ(Vgj ) = V0 . In
≥
<
the Euclidean structure we have chosen, A0 is orthogonal to V0 ; hence Λp A≥
0 is
orthogonal to the hyperplane {x ∈ Λp A | x ∧ Λq V0< = 0}. By the previous point, it
follows that Λp φ is a canonizing map for the pair (EΛs p gi , EΛup gj ). As kΛp φk ≤ kφkp
and similarly for φ−1 , the conclusion follows.
(iii) Let C ≥ 1, and let g ∈ G ⋉ V be a C-non-degenerate ρ-regular map that is also
asymptotically contracting along X0 . Let φ be an optimal canonizing map for g,
and let g′ = φgφ−1 . Then it is clear that sX0 (g′ ) ≍C sX0 (g) and s̃(Λp g′ ) ≍C s̃(Λp g);
so it is sufficient to prove the statement for g ′ .
Let s1 ≤ . . . ≤ sp (resp. s′1 ≤ . . . ≤ s′q ) be the singular values of g′ restricted
&
<
≪
to Ag′ = A≥
0 (resp. to Vg = V0 ), so that
g′−1
A≥
0
= s−1
1 and
g ′ |V <
0
= s′q .
<
′
Since the spaces A≥
0 and V0 are stable by g and orthogonal, we get that the
singular values of g′ on the whole space A are
s′1 , . . . , s′q , s1 , . . . , sp
(note however that unless sX0 (g′ ) ≤ 1, this list may fail to be sorted in nondecreasing order.) On the other hand, we know that the singular values of Λp g′ are
products of p distinct singular values of g′ . Since EΛs p g′ is orthogonal to EΛup g′ , we
may once again analyze the singular values separately for each subspace. We know
that the singular value corresponding to E s is equal to s1 · · · sp ; we deduce that
k Λp g|E u k is equal to the maximum of the remaining singular values. In particular
it is larger than or equal to s′q · s2 · · · sp . On the other hand, if λ is the largest
eigenvalue of Λp g′ , then we have
|λ| = |λq+1 · · · λd | = det( g′
A≥
0
) = s1 · · · sp
(where λ1 , . . . , λd are the eigenvalues of g′ or equivalently of g, sorted by nondecreasing modulus). The second equality holds because g, hence g′ , is asymptotically
contracting along X0 , so that its eigenvalues are sorted in the “correct” order. It
follows that:
p ′
s̃(Λ g ) =
Λp g′ |E up
Λ g′
|λ|
s′q · s2 · · · sp
′
≥
= s′q s−1
1 = sX0 (g ),
s1 · · · sp
(8.13)
which is the first estimate we were looking for.
Now if we take s8.2 (C) small enough, we may suppose that sX0 (g′ ) ≤ 1. Then we
have s′q ≤ s1 , which means that the singular values of Λp g′ are indeed sorted in the
“correct” order. Hence s′q · s2 · · · sp is actually the largest singular value of Λp g′ |E u ,
and the inequality becomes an equality: s̃(Λp g′ ) = sX0 (g′ ). The second estimate
follows.
70
(iv) See Lemma 3.8 (iv) in [Smi16].
We also need the following technical lemma, which generalizes Lemma 3.9 in [Smi16]
and Lemma 7.3 in [Smi]:
Lemma 8.3. There is a constant ε > 0 with the following property. Let A1 , A2 be any
two affine parabolic spaces such that
(
αHaus (A1 , A≥
0) ≤ ε
≤
Haus
α
(A2 , A0 ) ≤ ε.
Then they form a 2-non-degenerate pair.
(Of course the constant 2 is arbitrary; we could replace it by any number larger than 1.)
Proof. The proof is exactly the same as the proof of Lemma 3.9 in [Smi16], mutatis
mutandis.
Proposition 8.4. For every C ≥ 1, there is a positive constant s8.4 (C) < 1 with the
following property. Take any C-non-degenerate pair (g, h) of ρ-regular maps in G ⋉ V ;
suppose that we have sX0 (g±1 ) ≤ s8.4 (C) and sX0 (h±1 ) ≤ s8.4 (C). Then gh is still
ρ-regular, 2C-non-degenerate, and we have:
&
&
αHaus Vgh
, Vg .C ~sX0 (ℓ(g))
;
(i)
αHaus V . , V . . ~s (ℓ(h))
C X0
gh
h
αHaus A&gh , A&g .C sX0 (g)
(ii)
αHaus A. , A. . s (h−1 )
C X0
gh
h
;
(iii) sX0 (gh) .C sX0 (g)sX0 (h).
Points (ii) and (iii) are a generalization of Proposition 3.6 in [Smi16] and Proposition 7.4 in [Smi]; the proof is very similar. Together they give the “main part” of the
affine version of Schema 1.5.
As for point (i), it generalizes Corollary 3.7 in [Smi16] and Corollary 7.5 in [Smi]. But
its statement is now stronger, as it involves the linear contraction strength. As such, it
can no longer be obtained as a corollary of the affine version; instead it must be proved
independently, using Proposition 4.22.
Remark 8.5. Note that point (ii) involves h−1 , but in point (i) we have simply written h
instead. In fact in the linear case, the distinction between h and h−1 becomes irrelevant,
as they both have the same linear contraction strength by Proposition 4.14 (iv). However
their affine contraction strengths can be different.
71
Proof of Proposition 8.4. Let us fix some constant s8.4 (C), small enough to satisfy all
the constraints that will appear in the course of the proof. Let (g, h) be a pair of maps
satisfying the hypotheses.
First of all note that if we assume that s8.4 (C) ≤ 1, then Proposition 7.23 (i) ensures
that g, h, g −1 and h−1 are all asymptotically contracting along X0 .
• Let us prove (i). By Proposition 6.16 (i), it follows that g and h, hence ℓ(g)
and ℓ(h), are X0 -regular. Proposition 7.26 (i) and (ii) then implies that the pair
(ℓ(g), ℓ(h)) satisfies the hypotheses of Proposition 4.22. Hence we have
X0 ,+
X0 ,+
δ yℓ(gh)
.C ~sX0 (ℓ(g));
, yℓ(g)
(8.14)
δ y X0 ,− , y X0 ,− . ~s (ℓ(h)).
C
X
0
ℓ(gh)
ℓ(h)
Now remember (Remark 7.15) that the attracting (resp. repelling) flag of a map g ∈
G carries the same information as its (linear) ideally expanding (resp. contracting)
space. Even more precisely, we may deduce from Proposition 6.4 that the orbital
map from G to the orbit of V0≥ in the (p−1)-dimensional Grassmanian of V descends
to a smooth embedding of the flag variety G/PX+0 . Since the flag variety is compact,
the embedding is in particular Lipschitz-continuous. The desired inequalities follow.
• If we take s8.4 (C) ≤ s8.1 (C), then Proposition 8.1 tells us that gh is asymptotically contracting along X0 . We may also apply Proposition 8.1 to the pair
(h−1 , g−1 ); hence (gh)−1 is also asymptotically contracting along X0 , and by Proposition 6.16 (ii), we deduce that gh is ρ-regular.
The remainder of the proof works exactly like the proof of Proposition 3.6 in [Smi16]
or of Proposition 7.4 in [Smi], namely by applying Proposition 3.5 to the maps
γ1 = Λp g and γ2 = Λp h. Let us check that γ1 and γ2 satisfy the required hypotheses:
– By Lemma 8.2 (i), γ1 and γ2 are proximal.
– By Lemma 8.2 (ii), the pair (γ1 , γ2 ) is C p -non-degenerate.
– If we choose s8.4 (C) ≤ s8.2 (C), it follows by Lemma 8.2 (iii) that s̃(γ1 ) .C
sX0 (g) and s̃(γ2 ) .C sX0 (h). If we choose s8.4 (C) sufficiently small, then γ1
and γ2 are s̃3.5 (C p )-contracting, i.e. sufficiently contracting to apply Proposition 3.5.
Thus we may apply Proposition 3.5. It remains to deduce from its conclusions the
conclusions of Proposition 8.4.
– From Proposition 3.5 (i), using Lemma 8.2 (i), (iii) and (iv), we get
&
αHaus Agh , A&g .C sX0 (g),
which shows the first line of Proposition 8.4 (ii).
– By applying Proposition 3.5 to γ2−1 γ1−1 instead of γ1 γ2 , we get in the same
way the second line of Proposition 8.4 (ii).
72
&
– Let φ be an optimal canonizing map for the pair (Ag , A.h ). By hypothesis, we
have φ±1 ≤ C. But if we take s8.4 (C) sufficiently small, the two inequalities
that we have just shown, together with Lemma 8.3, allow us to find a map φ′
with kφ′ k ≤ 2, kφ′ −1 k ≤ 2 and
≤
φ′ ◦ φ(A&gh , A.gh ) = (A≥
0 , A0 ).
It follows that the composition map gh is 2C-non-degenerate.
– The last inequality, namely Proposition 8.4 (iii), now is deduced from Proposition 3.5 (ii) by using Lemma 8.2 (iii).
9 Additivity of Margulis invariant
The goal of this section is to prove Propositions 9.1 and 9.3, which explain how the Margulis invariant behaves under group operations (respectively inverse and composition).
These two propositions are the key ingredients in the proof of the Main Theorem.
Proposition 9.1 is a generalization of Proposition 4.1 (i) in [Smi16]. The proof is
similar, and fairly straightforward. Compare it also with the results of section 4.2.
Proposition 9.3 is a generalization of Proposition 4.1 (ii) in [Smi16]; it gives the “asymptotic dynamics” part in the affine version of Schema 1.5. The proof takes up the majority
of this section.
To estimate M (gh), the idea is to introduce two vectors Mgh (g) and Mgh (h) such that
we have, by definition, M (gh) = Mgh (g) + Mgh (h). We then first find an intermediate
vector, called Mg,gh (g), that we prove to be close to M (g) (Lemma 9.7). Then we prove
that Mgh (g) is close to this intermediate vector (Lemma 9.10).
Proposition 9.1. For every ρ-regular map g ∈ G ⋉ V , we have
M (g−1 ) = −w0 (M (g)).
Remark 9.2. Note that M (g) is (by definition) an element of the space V0t , which (again
by definition) is the set of fixed points of L = ZG (A). From this, it is straightforward
to deduce that V0t is invariant by NG (A). Hence w0 induces a linear involution on V0t ,
which does not depend on the choice of a representative of w0 in G.
Proof. First of all note that if φ is a canonizing map for g, then w0 φ is a canonizing map
for g−1 . Indeed, assume that g = τv exp(Jd(g))ge gu is in canonical form; then we have
w0 (− Jd(g)) ∈ a+ (since by definition w0 exchanges the dominant Weyl chamber a+ and
its negative −a+ ), and w0 (−v) ∈ V0= (since X0 is symmetric, the action of w0 preserves
V0= ).
It remains to show that w0 , or more precisely any representative w̃0 ∈ NG (A) of w0 ,
commutes with πt . We use the well-known fact that the group W , that we defined as
the quotient NG (A)/ZG (A), is also equal to the quotient NK (A)/ZK (A) (see [Kna96],
formulas (7.84a) and (7.84b)); hence
w̃0 ∈ NG (A) = W ZG (A) = W ZK (A)A = NK (A)A ⊂ KA.
73
(9.1)
Now recall ((7.3), (7.4)) that we have an orthogonal decomposition
V0=
V0a
z
}|
{
= V0≎ ⊕ (V0a ∩ V 0 ) ⊕ V0t .
|
{z
}
V0
Let us show that all three components are invariant by w̃0 :
• For V0≎ this is obvious (since Ωw0 \ {0} is invariant by w0 ).
• For V0t this follows from Remark 9.2.
• This is obviously the case for V 0 . Now by definition the group A acts trivially
on V 0 , and by construction K acts on V 0 by orthogonal transformations (indeed
the Euclidean structure was chosen in accordance with Lemma 2.4); hence V0a ∩V 0 ,
which is the orthogonal complement of V0t in V 0 , is also invariant by w̃0 .
The desired formula now immediately follows from the definition of the Margulis invariant.
Proposition 9.3. For every C ≥ 1, there are positive constants s9.3 (C) ≤ 1 and ε9.3 (C)
with the following property. Let g, h ∈ G ⋉ V be a C-non-degenerate pair of ρ-regular
maps, with g±1 and h±1 all s9.3 (C)-contracting along X0 . Then gh is ρ-regular, and we
have:
kM (gh) − M (g) − M (h)k ≤ ε9.3 (C).
The basic idea of the proof is the same as in [Smi] (or [Smi16]). However, in the
proof of Lemma 8.6 in [Smi], a key point is that since the factors that are introduced to
construct Diagram 4, namely ℓ(ggh ) and ℓ(gg,gh ), are linear parts of quasi-translations,
they automatically have bounded norm. But in the general case, this last deduction fails:
see Remark 7.9. This issue forced us to completely reorganize the proof.
The new proof, though still technical, is more elegant. It is more symmetric: for
instance we got rid of the lopsided Diagram 4, and of the confusing series P1 , P2 , P2′
and P1′′ . It is also more structured, as it cleanly separates into:
• an “algebraic” part comprising Lemma 9.7, involving a combination of “canonical”
projections which are all bounded quasi-translations by Lemma 7.18;
• and an “analytic” part comprising Lemma 9.8, its Corollary 9.9 and Lemma 9.10,
involving projections between spaces which are “close to each other” (with the angle
controlled by the contraction strengths of g and h), hence which introduce only a
small error.
Proof. Let C ≥ 1. We choose some constant s9.3 (C) ≤ 1, small enough to satisfy all the
constraints that will appear in the course of the proof. For the remainder of this section,
we fix g, h ∈ G ⋉ V a C-non-degenerate pair of ρ-regular maps such that g ±1 and h±1
are s9.3 (C)-contracting along X0 .
The following remark will be used throughout this proof.
74
&
&
&
Remark 9.4. We may suppose that the pairs (Agh , A.gh ), (Ahg , A.hg ), (Ag , A.gh ) and
(A&hg , A.g ) are all 2C-non-degenerate. Indeed, recall that (by Proposition 8.4), we have
αHaus A&gh , A&g .C sX0 (g)
αHaus A. , A. . s (h−1 )
C X0
gh
h
and similar inequalities with g and h interchanged. On the other hand, by hypothe&
sis, (Ag , A.h ) is C-non-degenerate. If we choose s9.3 (C) sufficiently small, these four
statements then follow from Lemma 8.3.
Proof of Proposition 9.3, continued. If we take s9.3 (C) ≤ s8.4 (C), then Proposition 8.4
ensures that gh is ρ-regular.
≈
To estimate M (gh), we decompose the induced map gh : A≈
gh → Agh into a product of
several maps.
• We begin by decomposing the product gh into its factors. We have the commutative
diagram
gh
A≈
gh
A≈
hg
g
A≈
gh
h
(9.2)
≈
Indeed, since hg is the conjugate of gh by h and vice-versa, we have h(A≈
gh ) = Ahg
≈
≈
and g(Ahg ) = Agh .
≈
≈
≈
• Next we factor the map g : A≈
hg → Agh through the map g : Ag → Ag , which is
better known to us. We have the commutative diagram
A≈
gh
A≈
hg
g
πg
πg
A≈
g
g
(9.3)
A≈
g
≪
≫
where πg is the projection onto A≈
g parallel to Vg ⊕ Vg . (It commutes with g
≪
≫
≈
because Ag , Vg and Vg are all invariant by g.)
• Now we decompose again every diagonal arrow from the last diagram into two
factors. For any two ρ-regular maps u and v, we introduce the notation
&
.
A≈
u,v := Au ∩ Av .
≫
≈
We call P1 (resp. P2 ) the projection onto A≈
g,gh (resp. Ahg,g ), still parallel to Vg ⊕
≈
≪
≈
Vg . To justify this definition, we must check that Ag,gh (and similarly Ahg,g ) is
75
&
supplementary to Vg≫ ⊕ Vg≪ . Indeed, by Remark 9.4, A.gh is transverse to Ag ,
hence (by Proposition 7.14 and Proposition 7.6 (ii)) supplementary to Vg≫ ; thus
&
≈
≫
≪
≪
A&g = Vg≫ ⊕ A≈
g,gh and A = Vg ⊕ Ag = Vg ⊕ Vg ⊕ Ag,gh . Then we have the
commutative diagrams
A≈
gh
P1
A≈
g,gh
πg
A≈
g
(9.4a)
A≈
g
(9.4b)
πg
and
A≈
hg
P2
A≈
hg,g
πg
πg
• Finally (and this is new in comparison to [Smi]), we would like to replace P1 and P2
by some “better-behaved” projections.
≪
≫
– We define πg,gh to be the projection onto A≈
g,gh parallel to Vg ⊕Vgh . Obviously
≈
≈
it induces a bijection between Agh and Ag,gh .
≫
≪
– We define πhg,g to be the projection onto A≈
hg,g parallel to Vhg ⊕Vg . Obviously
≈
≈
it induces a bijection between Ahg and Ahg,g .
The reason they are “better-behaved” is that by Lemma 7.18, they actually commute with canonical identifications (see Remark 9.6 below for more details). We
then make the decompositions
−1
) ◦ πg,gh
P1 = (P1 ◦ πg,gh
(9.5a)
−1
P2 = (P2 ◦ πhg,g
) ◦ πhg,g .
(9.5b)
and
The last three steps can be repeated with h instead of g. The way to adapt the second
≈
≈
step is straightforward; for the third step, we factor πh : A≈
hg → Ah through Ah,hg and
≪
≈
≈
≈
πh : Agh → Ah through Agh,h ; for the fourth step, we project respectively along Vh≫ ⊕Vhg
≫
≪
and along Vgh ⊕ Vh .
Combining these four decompositions, we get the lower half of Diagram 2. (We left out
the expansion of h; we leave drawing the full diagram for especially brave readers.) Let
us now interpret all these maps as endomorphisms of A=
0 . To do this, we choose some
optimal canonizing maps
φg , φgh , φhg , φg,gh , φhg,g
&
&
respectively of g, of gh, of hg, of the pair (Ag , A.gh ) and of the pair (Ahg , A.g ). This
allows us to define ggh , hgh , gg,gh , g≈ , ∆1 , ∆2 , χ1 , χ2 , ψ1 , ψ2 to be the maps that make
the whole Diagram 2 commutative.
Now let us define
(
Mgh (g) := πt (ggh (x) − x)
(9.6)
Mgh (h) := πt (hgh (x) − x)
76
A=
0
A=
0
ggh
∆1
A=
0
∆2
=
=
A0
A0
gg,gh
χ1
χ2
A=
0
A=
0
ψ1
ψ2
A=
0
φgh
φgh
hgh
φg,gh
g≈
φg
A=
0
φhg,g
φg
A≈
gh
φhg
φhg
A≈
hg
g
A≈
hg
A≈
gh
πg,gh
πhg,g
P1
P2
A≈
hg,g
A≈
g,gh
πg
πg
A≈
g
g
Diagram 2
77
A≈
g
φgh
h
A≈
gh
=
=
=
:= A=
, where VAff,0
for any x ∈ VAff,0
0 ∩ VAff is the affine space parallel to V0 and
passing through the origin. Since gh is the conjugate of hg by g and vice-versa, the
elements of G ⋉ V (defined in an obvious way) whose restrictions to A=
0 are ggh and
≤
hgh stabilize the spaces A≥
g
and
h
are
thus quasiand
A
.
By
Lemma
7.11,
gh
gh
0
0
translations. It follows that these values Mgh (g) and Mgh (h) do not depend on the
choice of x. Compare this to the definition of a Margulis invariant (Definition 7.19): we
=
. It immediately follows that
have M (gh) = πt (ggh ◦ hgh (x) − x) for any x ∈ VAff,0
M (gh) = Mgh (g) + Mgh (h).
(9.7)
Thus it is enough to show that kMgh (g) − M (g)k .C 1 and kMgh (h) − M (h)k .C 1.
(Note that while the vectors Mgh (g) and Mgh (h) are elements of V0t , the maps ggh and
hgh are extended affine isometries acting on the whole subspace A=
0 .)
We shall prove the estimate for g; the proof of the estimate for h is analogous. We
proceed in two steps. First, we introduce the vector
Mg,gh (g) := πt (gg,gh (x) − x)
=
, and we show (Lemma 9.7) that it differs from Mgh (g) only by a
for any x ∈ VAff,0
bounded constant. Second, we show (Lemma 9.10) that it is very close to M (g). These
two lemmas together imply the conclusion.
Remark 9.5. In contrast to actual Margulis invariants, the values Mgh (g) and Mgh (h) do
depend on our choice of canonizing maps. Choosing other canonizing maps would force
us to subtract some constant from the former and add it to the latter.
Remark 9.6. The fourth decomposition step above, namely (9.5), is what makes the
whole proof much cleaner. In [Smi], we had a map called P1 , which was “almost a quasitranslation” (Lemma 8.9 in [Smi]). In the current proof, this map decomposes into two
pieces
P1 = χ1 ◦ ∆1 ,
(9.8)
that are much easier to deal with: χ1 is now a bounded quasi-translation just like ψ1 , and
thus falls in with the “algebraic” part; while ∆1 is now almost the identity (Lemma 9.8),
and thus stays in the “analytic” part.
Lemma 9.7. We have
kMg,gh (g) − M (g)k .C 1.
−1
Proof. By Lemma 7.18, the maps χ1 −1 , ψ1 , ψ2 and χ2 are quasi-translations. Let us
show that their norms are bounded by a constant that depends only on C; obviously this
implies the conclusion.
• Let us start with χ2 . By definition, we have:
χ2 = φhg,g ◦ πhg,g ◦ φ−1
hg
−1
◦
φ
◦
φ
= φhg,g ◦ πhg,g ◦ φ−1
hg,g
hg
hg,g
= π0 ◦ φhg,g ◦ φ−1
hg ,
78
(9.9)
>
<
where π0 is the projection onto A=
0 parallel to V0 ⊕ V0 . Now π0 is actually
an orthogonal projection, hence it has norm 1; and φhg,g ◦ φ−1
hg is bounded by
Remark 9.4.
• Similarly, we have
ψ2 = π0 ◦ φg ◦ φ−1
hg,g ,
which is bounded.
(9.10)
−1
≈
≈
≈
• To deal with ψ1 , note that πg : A≈
g,gh → Ag and πg,gh : Ag → Ag,gh are inverse
to each other. We deduce that
ψ1
−1
= φg,gh ◦ πg,gh ◦ φ−1
g
−1
= φg,gh ◦ πg,gh ◦ φ−1
g,gh ◦ φg,gh ◦ φg
= π0 ◦ φg,gh ◦ φ−1
,
g
(9.11)
and we conclude as previously.
• Similarly, we have
χ1 −1 = π0 ◦ φgh ◦ φ−1
g,gh ,
(9.12)
and we conclude in the same way.
Lemma 9.8. The estimates
(i) k∆1
−1
(ii) k∆2
−1
&
&
− Id k .C αHaus Agh , Ag ;
− Id k .C
αHaus
.
.
Ahg , Ag ,
hold, as soon as the respective right-hand sides are smaller than some constant depending
only on C.
Corollary 9.9. We have:
(i) k∆1
−1
(ii) k∆2
−1
− Id k .C sX0 (g);
− Id k .C sX0 (g−1 );
(iii) kℓ(∆1 )−1 − Id k .C ~sX0 (ℓ(g));
(iv) kℓ(∆2 )−1 − Id k .C ~sX0 (ℓ(g)).
In light of Remark 9.6, this corollary can be seen as a simpler version of Lemma 8.9
in [Smi]. Indeed the old P1 corresponds by (9.8) to the new χ1 ◦ ∆1 , while the old P1′′
corresponds to χ1 alone.
79
Proof. Points (i) and (ii) immediately follow from the Lemma combined with Proposition 8.4 (ii), provided that we take s9.3 (C) small enough.
For points (iii) and (iv), simply apply the Lemma to the pair (ℓ(g), ℓ(h)). (It is easy
to check that this pair still satisfies the hypotheses of Proposition 9.3.) We then get:
kℓ(∆1 )−1 − Id k .C αHaus V & , Vg& .C ~sX (ℓ(g));
0
gh
kℓ(∆2 )−1 − Id k .C αHaus V . , Vg. .C ~sX (ℓ(g)).
0
hg
In each case, the second inequality follows from Proposition 8.4 (i). The first inequality
is an application of the Lemma, which is licit (provided s9.3 (C) is small enough), since
Proposition 7.26 (ii) propagates the required upper bound to ~sX0 (ℓ(g)).
The proof of Lemma 9.8 might seem slightly technical, but there is an easy intuition
behind it. Essentially, the idea is that if you jump back and forth between two spaces
that almost coincide, going both times in directions whose angle with the two spaces is
not too shallow, then you can’t end up very far from your starting point.
Proof of Lemma 9.8.
(i) By Remark 9.4, we know that kφ±1
gh k ≤ 2C. Conjugating everything by this map,
it is thus sufficient to show that for every x ∈ A≈
gh , we have
P1−1 (πg,gh (x)) − x .C sX0 (g)kxk
(9.13)
≈
(where by P1−1 we mean the inverse of the bijection P1 : A≈
gh → Ag,gh ).
• Let us first estimate the quantity kπg,gh (x) − xk. To begin with, let us push
everything forward by the map φg,gh ; writing y = φg,gh (x), we have:
kφg,gh (πg,gh (x) − x)k
kπ0 (y) − yk
=
kφg,gh (x)k
kyk
= sin α(y, A=
0)
≤ α(y, A=
0 ).
By Remark 9.4, we know that kφ±1
g,gh k ≤ 2C; hence we may pull everything
back by φg,gh again:
We conclude that
kπg,gh (x) − xk
.C α(x, A≈
g,gh )
kxk
&
≤ α A≈
gh , Ag
&
≤ αHaus Agh , A&g .
&
kπg,gh (x) − xk .C αHaus Agh , A&g kxk.
80
(9.14)
(9.15)
• Let us now estimate the quantity P1−1 (πg,gh (x)) − πg,gh (x) . We introduce
the notation z = πg,gh (x), and we define ξ to be the unique linear automorphism of A satisfying
(
φg,gh (x) if x ∈ A&g
(9.16)
ξ(x) =
φg (x)
if x ∈ Vg≪
≈
(so that ξ −1 (V0> ) = Vg≫ and ξ −1 (V0< ) = Vg≈ but ξ −1 (A=
0 ) = Ag,gh ). From
±1
the inequalities kφ±1
g k ≤ C and kφg,gh k ≤ 2C, it is easy to deduce that both
−1
norms kξk and kξ k are bounded by a constant that depends only on C.
Now we obviously have ξ ◦ P1 ◦ ξ −1 = π0 , hence
kξ(P1−1 (z) − z)k
= tan α(ξ(P1−1 (z)), A=
0 ).
kξ(z)k
Now we have
−1
≈
α(ξ(P1−1 (z)), A=
0 ) .C α(P1 (z), Ag,gh )
&
.C αHaus Agh , A&g
since ξ and ξ −1 are bounded
since P1−1 (z) ∈ A≈
gh .
If the right-hand side is small enough, we may assume that α(ξ(z), A=
0 ) is
π
smaller than some fixed constant α0 (say α0 = 4 ). For every α ≤ α0 , we
obviously have
tan α .α0 α.
It follows that
kξ(P1−1 (z) − z)k
&
.C αHaus Agh , A&g .
kξ(z)k
(9.17)
Now since ξ and ξ −1 are bounded, we deduce that
kP1−1 (z) − zk .C αHaus A&gh , A&g kzk.
It remains to estimate kzk in terms of kxk. We have:
&
kzk ≤ kxk + kπg,gh (x) − xk .C kxk + αHaus Agh , A&g kxk
&
&
by (9.15). Taking αHaus Agh , Ag small enough, we may assume that kzk ≤
2kxk. We conclude that
&
(9.18)
P1−1 (πg,gh (x)) − πg,gh (x) .C αHaus Agh , A&g kxk.
Adding together (9.15) and (9.18), the desired inequality (9.13) follows.
(ii) The proof is completely analogous.
81
Lemma 9.10. We have
kMgh (g) − Mg,gh (g)k .C sX0 (g).
The proof is somewhat similar to the proof of the second half of Lemma 8.6 in [Smi].
Proof. Recall that
Mgh (g) = πt (ggh (x) − x),
=
where x can be any element of the affine space VAff,0
.
Let O be the “origin” of VAff , i.e. the intersection of the line R0 with the affine space VAff
=
(in other terms O = (0, 1) ∈ V ⊕ R0 ). By definition O is an element of VAff,0
. We then
have, for every extended affine map f ,
∀x ∈ VAff ,
f (x) = ℓ(f )(x − O) + f (O).
Let us take
x = ∆2
−1
−1
(9.19)
gg,gh −1 (O) .
◦ gg,gh ◦ ∆2 , we may then write
−1
Mgh (g) = πt ∆2 (x) − x + πt gg,gh ∆2 (x) − ∆2 (x) + πt ∆1
gg,gh ∆2 (x) − gg,gh ∆2 (x) .
−1
−1
gg,gh −1 (O)
= πt gg,gh −1 (O) − ∆2
+ πt O − gg,gh −1 (O) + πt ∆1 (O) − O .
{z
}
|
|
|
{z
}
{z
}
=: (II)
=: (I)
=: (III)
Since ggh = ∆1
The middle term (II) is then by definition equal to Mg,gh (g), so that
Mgh (g) − Mg,gh (g) = (I) + (III).
(9.20)
Now we have
k(III)k ≤ k∆1
≤ k∆1
= k∆1
−1
(O) − Ok
−1
−1
− Id kkOk
− Id k
by Corollary 9.8 (i).
.C sX0 (g)
(9.21)
It remains to estimate (I). We set
y := gg,gh −1 (O).
Let us calculate the norm of this vector:
kyk ≤
=
gg,gh −1
χ2 −1 ◦ ψ2
−1
◦ g≈ −1 ◦ ψ1 ◦ χ1
.C g≈ −1
=
φg ◦ g−1
.C
g−1
A≈
g
A≈
g
.
82
◦ φ−1
g
(9.22)
(To justify the third line, remember that we have seen in the proof of Lemma 9.7 that
−1
the four maps χ1 −1 , ψ1 , ψ2 and χ2 are all bounded.) Now we have:
−1
k(I)k = πt y − ∆2 (y)
−1
≤ ∆2 (y) − y
−1
−1
= ℓ(∆2 ) − Id (y − O) + ∆2 − Id (O)
≤ ℓ(∆2
−1
) − Id kyk + ∆2
.C ~sX0 (ℓ(g)) g−1
A≈
g
−1
− Id
by (9.19)
as kyk2 = ky − Ok2 + 1
by Corollary 9.8 (ii) and (iv) and (9.22)
+ sX0 (g)
by Proposition 7.26 (iii).
(9.23)
.C sX0 (g)
Joining together (9.21) and (9.23), the conclusion follows.
10 Margulis invariants of words
We have already studied how contraction strengths (Proposition 8.4) and Margulis invariants (Proposition 9.3) behave when we take the product of two mutually C-nondegenerate, sufficiently contracting ρ-regular maps. The goal of this section is to generalize these results to words of arbitrary length on a given set of generators. This is a
straightforward generalization of Section 9 in [Smi] and of Section 5 in [Smi16].
Definition 10.1. Take k generators g1 , . . . , gk . Consider a word g = giσ11 · · · giσll with
length l ≥ 1 on these generators and their inverses (for every m we have 1 ≤ im ≤ k
and σm = ±1). We say that g is reduced if for every m such that 1 ≤ m ≤ l − 1, we have
(im+1 , σm+1 ) 6= (im , −σm ). We say that g is cyclically reduced if it is reduced and also
satisfies (i1 , σ1 ) 6= (il , −σl ).
Proposition 10.2. For every C ≥ 1, there is a positive constant s10.2 (C) ≤ 1 with the
following property. Take any family of maps g1 , . . . , gk ∈ G ⋉ V satisfying the following
hypotheses:
(H1) Every gi is ρ-regular.
(H2) Any pair taken among the maps {g1 , . . . , gk , g1−1 , . . . , gk−1 } is C-non-degenerate, except of course if it has the form (gi , gi−1 ) for some i.
(H3) For every i, we have sX0 (gi ) ≤ s10.2 (C) and sX0 (gi−1 ) ≤ s10.2 (C).
Take any nonempty cyclically reduced word g = giσ11 · · · giσll (with 1 ≤ im ≤ k, σm = ±1
for every m). Then g is ρ-regular, 2C-non-degenerate, and we have
M (g) −
l
X
M (giσmm ) ≤ lε9.3 (2C)
m=1
(where ε9.3 (2C) is the constant introduced in Proposition 9.3).
83
The proof proceeds by induction, with Proposition 8.4 and Proposition 9.3 providing
the induction step.
Proof. The proof is exactly the same as the proof of Proposition 5.2 in [Smi16], mutatis
mutandis.
Let us just present one small improvement. The proof relies on the following lemma:
Lemma 10.3. Every cyclically reduced word g = giσ11 · · · giσll can be decomposed as a
product of two cyclically reduced subwords
g′ := giσ11 · · · giσmm
and
σ
m+1
g′′ := gim+1
· · · giσll ,
both nonempty (that is, 0 < m < l).
In [Smi16], this was proved by contradiction, in a somewhat obscure way. Let us
reformulate the proof so that it is constructive, and hopefully more comprehensible.
Proof. We may take m to be the smallest (positive) index such that
(im+1 , σm+1 ) 6= (il , −σl ).
Such an index always exists, and is at most equal to l − 1 (since the word is reduced).
Then the first subword is actually of the form
m−1
g′ = giσ11 giσll
,
and any word of this form is automatically cyclically reduced as soon as it is reduced.
As for g′′ , it is cyclically reduced by construction.
11 Construction of the group
Here we prove the Main Theorem. The reasoning is similar to that of Section 6 in [Smi16],
and almost identical to that of Section 10 in [Smi]. The main difference is the substitution
of a′ρ,X0 instead of aρ,X0 , which in particular requires us to invoke Proposition 6.16 (which
had no equivalent in [Smi]) in the final proof. Also, since we have now developed the
purely linear theory in a systematic way (Section 4), the relationship between linear
properties and affine properties becomes clearer (in particular in the second bullet point
of the final proof).
Let us recall the outline of the proof. We begin by showing (Lemma 11.2) that if we
take a group generated by a C-non-degenerate family of sufficiently contracting ρ-regular
maps that have suitable Margulis invariants, it satisfies all of the conclusions of the Main
Theorem, except Zariski-density. We then exhibit such a group that is also Zariski-dense
(and thus prove the Main Theorem).
The idea is to ensure that the Margulis invariants of all elements of the group lie
almost on the same half-line. Obviously if −w0 maps every element of V0t to its opposite,
Proposition 9.1 makes this impossible. So we now exclude this case:
84
Assumption 11.1. The representation ρ is such that the action of w0 on V0t is not
trivial.
This is precisely condition (i) from the Main Theorem. More precisely, V0t is the set of
all vectors that satisfy (i)(a), and what we say here is that some of them also satisfy (i)(b).
See Example 10.2 in [Smi] for examples of representations that do or do not satisfy this
condition.
Thanks to Assumption 11.1, we choose once and for all some nonzero vector v ∈ V0t
that is a fixed point of −w0 (which is possible since w0 is an involution). We also choose
a vector M0 collinear to v and such that kM0 k = 2ε9.3 (2C).
Lemma 11.2. Take any family g1 , . . . , gk ∈ G ⋉ V satisfying the hypotheses (H1), (H2)
and (H3) from Proposition 10.2, and also the additional condition
(H4) For every i, M (gi ) = M0 .
Then these maps generate a free group acting properly discontinuously on the affine
space VAff .
Proof. The proof is exactly the same as the proof of Lemma 6.1 in [Smi16], mutatis
mutandis. The constant denoted as µ(C) (or µ(2C)) in the earlier paper corresponds to
what we now call ε9.3 (C) (or respectively ε9.3 (2C)). The (orthogonal) projection
π̂z : ĝ → z ⊕ R0
parallel to d ⊕ n+ ⊕ n−
now becomes the (orthogonal) projection
π̂t : A → V0t ⊕ R0
parallel to V0a ⊕ V0> ⊕ V0< .
We may now finally prove the Main Theorem. We follow the same strategy as in the
proof of the Main Theorems of [Smi16] and of [Smi] (with a few additional tweaks).
Proof of Main Theorem. First note that Assumption 11.1 guarantees that ρ satisfies the
hypotheses of the Main Theorem. (The two other assumptions were “for free”: Assumption 7.12 is just the weaker condition (i)(a), and Assumption 5.5 is an even weaker
condition that follows from (i)(a).)
We find a positive constant C ≥ 1 and a family of maps g1 , . . . , gk ∈ G ⋉ V (with
k ≥ 2) that satisfy the conditions (H1) through (H4) and whose linear parts generate a
Zariski-dense subgroup of G, then we apply Lemma 11.2. We proceed in several stages.
• We begin by using a result of Benoist: we apply Lemma 4.3.a in [Ben97] to
– Γ = G;
– t = k + 1;
– Ω1 = · · · = Ωk = a′ρ,X0 ∩ a++ , as defined in Remark 6.15. (Point (iii) of that
remark assures that it is indeed a nonempty open convex cone.)
85
This gives us, for any k ≥ 2, a family of maps γ1 , . . . , γk ∈ G (which we shall see
as elements of G ⋉ V , by identifying G with the stabilizer of R0 ), such that:
(i) Every γi and every γi−1 is asymptotically contracting along X0 . In particular,
by Proposition 6.16 (ii), it is ρ-regular (this is (H1)).
(ii) For any two indices i, i′ and signs σ, σ ′ such that (i′ , σ ′ ) 6= (i, −σ), the pair
(yγXσ0 ,+ , y Xσ0′,− ) ∈ G/PX+0 × G/PX−0
i
γi′
is transverse.
(iii) Any single γi generates a Zariski-connected group.
(iv) All of the γi generate together a Zariski-dense subgroup of G.
Since in our case t is finite, Benoist’s item (v) is not relevant to us.
A comment about item (i): Benoist’s theorem only works with R-regular elements;
so it forces every γi to be not only ρ-regular, but actually R-regular (of which we
make no use).
A comment about item (ii): since we have taken Benoist’s Γ to be the whole
group G, we have θ = Π, so that YΓ is the full flag variety G/P + . So Benoist’s
theorem actually gives us the stronger property that the pair
X,−
+
× G/P −
(yγX,+
σ , y σ ′ ) ∈ G/P
i
γi′
(where X is some element of the open Weyl chamber a++ ) is transverse in the full
flag variety. Once again, we only need the weaker version.
• Using Remark 7.15, condition (ii) above may be restated in the following way: for
any two indices i, i′ and signs σ, σ ′ such that (i′ , σ ′ ) 6= (i, −σ), the pair of parabolic
spaces
&
(Vγ σ , V .σ′ )
i
γi′
is transverse. Clearly, every pair of transverse spaces is C-non-degenerate for some
finite C; and here we have a finite number of such pairs. Hence if we choose some
suitable value of C (which we fix for the rest of this proof), the hypothesis (H2)
follows.
• From condition (iii) (Zariski-connectedness), it follows that any algebraic group
containing some power γiN of some generator must actually contain the generator γi itself. This allows us to replace every γi by some power γiN without sacrificing condition (iv) (Zariski-density). Clearly, conditions (i), (ii) and (iii) are
then preserved as well. If we choose N large enough, we may suppose (thanks to
Proposition 7.23 (ii)) that the numbers sX0 (γi±1 ) are as small as we wish: this gives
us (H3). In fact, we shall suppose that for every i, we have sX0 (γi±1 ) ≤ sMain (C)
for an even smaller constant sMain (C), to be specified soon.
86
• To satisfy (H4), we replace the maps γi by the maps
gi := τφ−1 (M0 ) ◦ γi
(11.1)
i
(for 1 ≤ i ≤ k), where φi is a canonizing map for γi .
We need to check that this does not break the first three conditions. Indeed,
for every i, we have γi = ℓ(gi ); even better, since the affine map φi gi φ−1
has
i
by construction canonical form, gi has the same geometry as γi (meaning that
&
&
&
Agi = Aγi = Vγi ⊕ R0 and A.gi = A.γi = Vγ.i ⊕ R0 ). Hence the gi still satisfy the
hypotheses (H1) and (H2), but now we have M (gi ) = M0 (this is (H4)). As for
contraction strength along X0 , we have:
sX0 (gi ) = gi |Vg≪
i
= γi |Vγ≪
i
gi−1
&
Agi
(τφ−1 (M0 ) ◦ γi )−1
i
&
Ag
−1
≤ sX0 (γi ) φ−1
i ◦ τM0 ◦ φi
−1
k,
.C sMain (C)kτM
0
(11.2)
and similarly for gi−1 . Recall that kM0 k = 2ε9.3 (2C), hence the quantity
−1
k = kτ−M0 k
kτM0 k = kτM
0
depends only on C: in fact it is equal to the norm of the 2-by-2 matrix
It follows that if we choose
sMain (C) ≤ s10.2 (C)
1 2ε9.3 (2C)
0
1
−1
,
1 kM0 k
0 1
.
(11.3)
then the hypothesis (H3) is satisfied.
We conclude that the group generated by the elements g1 , . . . , gk acts properly
discontinuously (by Lemma 11.2), is free (by the same result), nonabelian (since
k ≥ 2), and has linear part Zariski-dense in G.
References
[Abe01] H. Abels. Properly discontinuous groups of affine transformations, a survey.
Geom. Dedicata, 87:309–333, 2001.
[AMS]
H. Abels, G.A. Margulis, and G.A. Soifer. The Auslander conjecture for dimension less than 7. Preprint, arXiv:1211.2525.
[AMS02] H. Abels, G.A. Margulis, and G.A. Soifer. On the Zariski closure of the linear
part of a properly discontinuous group of affine transformations. J. Differential
Geometry, 60:315–344, 2002.
87
[AMS11] H. Abels, G.A. Margulis, and G.A. Soifer. The linear part of an affine group
acting properly discontinuously and leaving a quadratic form invariant. Geom.
Dedicata, 153:1–46, 2011.
[Aus64]
L. Auslander. The structure of compact locally affine manifolds. Topology,
3:131–139, 1964.
[Ben96]
Y. Benoist. Actions propres sur les espaces homogènes réductifs. Annals of
Mathematics, 144:315–347, 1996.
[Ben97]
Y. Benoist. Propriétés asymptotiques des groupes linéaires. Geom. and funct.
anal., 7:1–47, 1997.
[BQ]
Y.
Benoist
and
J.-F.
Quint.
Random
walks
on
ductive groups.
Ergebnisse.
To appear;
available
http://www.math.u-psud.fr/~benoist/prepubli/15walk.pdf.
[BT65]
A. Borel and J. Tits. Groupes réductifs. Publications Mathématiques de l’IHÉS,
27:55–151, 1965.
[BT72]
A. Borel and J. Tits. Compléments à l’article “ Groupes réductifs ”. Publications
Mathématiques de l’IHÉS, 41:253–276, 1972.
[DGK]
J. Danciger, F. Guéritaud, and F. Kassel. Proper affine action of right-angled
Coxeter groups. In preparation.
[Ebe96]
P.B. Eberlein. Geometry of Nonpositively Curved Manifolds. University of
Chicago Press, 1996.
[FG83]
D. Fried and W.M. Goldman. Three-dimensional affine crystallographic groups.
Adv. in Math., 47:1–49, 1983.
[Hal15]
B.C. Hall. Lie Groups, Lie Algebras and Representations: An Elementary
Introduction. Springer International Publishing, second edition, 2015.
[Hel08]
S. Helgason. Geometric Analysis on Symmetric Spaces. Amer. Math. Soc.,
second edition, 2008.
reat
[Hum75] J. Humphreys. Linear Algebraic Groups. Springer-Verlag, 1975.
[Kna96] A.W. Knapp. Lie Groups Beyond an Introduction. Birkhaüser, 1996.
[Mar83] G.A. Margulis. Free properly discontinuous groups of affine transformations.
Dokl. Akad. Nauk SSSR, 272:937–940, 1983.
[Mar87] G.A. Margulis. Complete affine locally flat manifolds with a free fundamental
group. J. Soviet Math., 134:129–134, 1987.
[Mil77]
J. Milnor. On fundamental groups of complete affinely flat manifolds. Adv. in
Math., 25:178–187, 1977.
88
[Smi]
I. Smilga. Proper affine actions in non-swinging representations. Submitted;
available at arXiv:1605.03833.
[Smi16]
I. Smilga. Proper affine actions on semisimple Lie algebras. Annales de l’Institut
Fourier, 66(2):785–831, 2016.
[Tit71]
J. Tits. Représentations linéaires irréductibles d’un groupe réductif sur un
corps quelconque. Journ. Reine Angw. Math., 247:196–220, 1971.
89
| 4math.GR
|
Human-in-the-loop Artificial Intelligence
Fabio Massimo Zanzotto
University of Rome Tor Vergata
arXiv:1710.08191v1 [cs.AI] 23 Oct 2017
[email protected]
Abstract
Little by little, newspapers are revealing the bright future that Artificial Intelligence (AI) is building. Intelligent machines will help everywhere. However, this
bright future has a dark side: a dramatic job market contraction before its unpredictable transformation. Hence, in a near future, large numbers of job seekers
will need financial support while catching up with these novel unpredictable jobs.
This possible job market crisis has an antidote inside. In fact, the rise of AI is sustained by the biggest knowledge theft of the recent years. Learning AI machines are
extracting knowledge from unaware skilled or unskilled workers by analyzing their
interactions. By passionately doing their jobs, these workers are digging their own
graves.
In this paper, we propose Human-in-the-loop Artificial Intelligence (HIT-AI)
as a fairer paradigm for Artificial Intelligence systems. HIT-AI will reward aware
and unaware knowledge producers with a different scheme: decisions of AI systems
generating revenues will repay the legitimate owners of the knowledge used for taking
those decisions. As modern Robin Hoods, HIT-AI researchers should fight for a fairer
Artificial Intelligence that gives back what it steals.
1 Introduction
We are on the edge of a wonderful revolution: Artificial Intelligence (AI) is breathing life
into helpful machines, which will relieve us of our need to perform repetitive activities.
Self-driving cars [20, 23, 22] are taking their first steps in our urban environment and
their younger brothers, that is, assisted driving cars [28, 33, 13], are already a commercial
reality. Robots are vacuum cleaning and mopping the floors of our houses [34, 32, 15].
Chatbots1 [39, 37] have conquered our new window-on-the-world – our smartphones –
and, from there, they help with everyday tasks such as managing our agenda, answering
our factoid questions or being our learning companions [16, 4]. In medicine, computers
can already help in formulating diagnoses [2, 18, 10] by looking at data doctors generally
neglect. Artificial Intelligence is preparing a wonderful future where people are released
from the burden of repetitive jobs.
The bright Artificial Intelligence revolution has a dark side: a dramatic mass unemployment that will precede an unpredictable job market transformation. People and,
1
see http://houseofbots.com/
1
hence, governments are frightened. Nearly every week, newspapers all over the world
are reporting on possible futures where around one fifth of actual jobs will disappear.
Alarming reports foresee that more than one billion people will be unemployed worldwide [7]. By releasing people from repetitive jobs, intelligent machines will replace many
workers. Chatbots are slowly replacing call center agents. Self-driving trains are already
reducing the number of drivers in our trains. Self-driving cars are fighting to replace cab
drivers in our cities. Drones are expanding automation in managing delivery of goods
by drastically reducing the number of delivery people. And, these are only examples
as even more cognitive and artistic jobs are challenged. Intelligent machines may produce music jingles for commercials [5], write novels, produce news articles and so on.
Intelligent risk predictors may replace doctors [2, 18, 10]. Chatbots along with massive
open online courses may replace teachers and professors [11]. Coders risk being replaced
by machines too [25]. Nobody’s job is safe as we face this overwhelming progress of
Artificial Intelligence.
Surprisingly, the rise of Artificial Intelligence is supported by the unaware mass of
people who risk seeing their jobs replaced by machines. These people are giving away
their knowledge, which is used to train these wonderful machines. This is an enormous
and legal knowledge theft taking place in our modern era. Along with those aware
programmers and artificial intelligence researchers who set up the learning modules of
these artificial intelligent machines, an unaware mass of people is providing precious
training data by passionately doing their job or simply performing their activity on the
net. Answering an email, an interaction on a messaging service, leaving an opinion
on a hotel, and so on are all simple everyday activities people are doing. This data
is a goldmine for artificial intelligence machines. Learning systems transform these
interactions in knowledge for the artificial intelligence machines and the knowledge theft
is completed. By doing their normal everyday activity, people are digging the grave for
their own jobs.
As researchers in Artificial Intelligence, we have a tremendous responsibility: building
intelligent machines we can work with rather than intelligent machines that steal our
knowledge to do our jobs. We need to find ways to financially support job seekers as they
train to catch up with these novel unpredictable jobs. We need to prepare an antidote
as we spread this poison in the job market.
This paper propose Human-in-the-loop Artificial Intelligence (HIT-AI) as a novel
paradigm for a responsible Artificial Intelligence. This is a possible antidote to the poisoning of the job market. The idea is simple: giving the right value to the knowledge
producers. Human-in-the-loop AI is an umbrella for researchers in Artificial Intelligence
working with this underlying idea. Hence, HIT-AI promotes interpretable learning machines and, therefore, artificial intelligence systems with a clear knowledge lifecycle. For
HIT-AI systems, it will be clear whose the knowledge has been used in a specific deployment or in specific situations. This is a way to give the rightful credit and revenue to
the original knowledge producers. We need a fairer artificial intelligence.
The rest of the paper is organized as follows. Section 2 describes the enabling
paradigms of Human-in-the-loop AI. Section 3 sketches some simple proposals for a
2
better future. Then, Section 4 draws some conclusions.
2 Human-in-the-loop AI: Enabling Paradigms
2.1 Transferring Knowledge to Machines with Programming vs. with Learning from
Repeated Experience
Since the beginning of the digital era, programming is the preferred way to “teach” to
machines. Artificial non-ambiguous programming languages have been developed to have
a clear tool to tell machines what to do. According to this paradigm, whoever wants
to “teach” machines how to solve a new task or how to be useful has to master one of
these programming languages. These people, called programmers, have been teaching
machines for decades and have made these machines extremely useful. Nowadays, it
is difficult to think staying a single day without using the big network of machines
programmers have contributed to building.
As not all the tasks can be solved by programming, autonomous learning has been reinforced as an alternative way of controlling the “behavior” of machines. In autonomous
learning, machines are asked to learn from experience. With the paradigm of programming, we have asked machines to go to school before these machines have learned to walk
through trial and error. This is why machines have always been good in solving very
complex cognitive tasks but very poor in working with everyday simple problems. The
paradigm of autonomous learning has been introduced to solve this problem.
In these two paradigms, who should be paid for transferring knowledge to machines
and how should they be paid? In the programming paradigm, roles are clear: programmers are the “teachers” and machines are the “learners”. Hence, programmers could be
payed for their work. In the autonomous learning paradigm, the activity of programmers
is confined to the selection of the most appropriate learning model and of the examples
to show to these learning machines.
From the point of view of HIT-AI, programming is a fair paradigm as it keeps humans
in the loop although machines, which have been taught exactly what to do, can hardly be
called artificial intelligence. On the contrary, autonomous learning is an unfair model
of transferring knowledge as the real knowledge is extracted from data produced by
unaware people. Hence, little seems to be done by humans and machines seem to do the
whole job. Yet, knowledge is stolen without paying.
2.2 Explainable Artificial Intelligence and Explainable Machine Learning
Explaining the decisions of learning machines is a very hot topic nowadays: dedicated
workshops or specific sessions in major conferences are flourishing [1, 17]. In specific
areas of application, for example, medicine, thrust in intelligent machines cannot be
blind as final decisions can have a deep impact on humans. Hence, understanding why a
decision is taken become extremely important. However, what is exactly an explainable
machine learning model is still an open debate [21].
3
In HIT-AI, explainable machine learning can play a crucial role. In fact, seen from
another perspective, explaining machine learning decisions can keep humans in the loop
in two ways: 1) giving the last word to humans; and, 2) explaining what data sources
are responsible for the final decision. In the first case, the decision power is left in the
hand of very specialized professionals that use machines as advisers. This is a clear case
of human-in-the-loop AI. Yet, this is confined to highly specialized knowledge workers in
some specific area. The second case instead is fairly more important. In fact, machines
that take decisions or work on a task are constantly using knowledge extracted from
data. Spotting which data have been used for a specific decision or for a specific action
of the machine is very important in order to give credits to who has produced these
data. In general, data are produced by anyone and everyone, not only by knowledge
workers. Hence, understanding why a machine takes a decision may become a way to
keep everybody in the loop of artificial intelligence.
2.3 Convergence between Symbolic and Distributed Knowledge Representation
Explaining machine learning decisions is simpler in image analysis o, better, in all those
cases where the system representation is similar what is represented. In fact, for example,
neural networks interpreting images are generally interpreted by visualizing how subparts
represent salient subparts of target images. Both input images and subparts are tensors
of real numbers. Hence, these networks can be examined and understood.
However, large part of the knowledge is expressed with symbols. Both in natural
and artificial languages, combination of symbols are used to convey knowledge. In fact,
for natural languages, sounds are transformed in letters or ideograms and these symbols
are composed to produce words. Words then form sentences and sentences form texts,
discourses, dialogs, which ultimately convey knowledge, emotions, and so on. This
composition of symbols into words and of words in sentences follow rules that both the
hearer and the speaker know [6]. Hence, symbolic representations give a clear tool to
understand whose knowledge is used in specific machines.
In current Artificial Intelligence systems, symbols are fading away, erased by tensors
distributed representations. Distributed representations are pushing deep learning models
[19, 29] towards amazing results in many high-level tasks such as image recognition
[14, 31], image generation [12] and image captioning [36, 40], machine translation [3, 41],
syntactic parsing [35, 38] and even game playing at human level [30, 24].
There is a strict link between distributed representations and symbols, the first being
an approximation of the second [26, 27, 8, 9]. The representation of the input and the
output of these networks is not that different from their internal representation.
For HIT-AI, this strict link is a tremendous opportunity to track how symbolic
knowledge flows in the knowledge lifecycle. In this way, symbolic knowledge producers
can be rewarded for their unaware work.
4
3 Human-in-the-loop AI: a simple proposal for a better Future
A peasant of the late 19th century would have never imagined that after 100 years yoga
trainer, pet caretaker and ayurveda massage therapist – just to cite technology unrelated
jobs – are common jobs. It is also extremely likely that any wise politician of that period
had the same lack of imagination even though s/he had more time to spend to imagine
the future and less pressure on her/his job loss.
Today, we are in a situation similar to the end of the 19th century but we have
a complication: the speed of the AI revolution. As it was for the end-of-19th-century
peasants and politicians, we can hardly imagine what’s next on the job market. We
can see some trends but it is hard to exactly imagine what are the skills needed for
being part of the labor force of the future. Yet, the AI revolution is overwhelming and
risks elimination of many jobs in the near future. This may happen before our society
envisage a clear path for relocating workers. We urge a strategy for the immediate.
The Artificial Intelligence revolution is based on an enormous knowledge theft. Skilled
and unskilled workers do their own everyday jobs and leave important traces. These
traces are the training examples that machines can use to learn. Hence, Artificial Intelligence using machine learning is stealing these workers’ knowledge by learning from
their interactions. These unaware workers are basically digging the graves for their own
jobs.
The knowledge produced by workers and used by machines is going to produce revenues for machine owners for years. This is a major problem since only a very small
fraction of the population can benefit from this never-ending revenue source and the real
owners of the knowledge are not participating to this redistribution of wealth.
The model we propose with Human-in-the-loop Artificial Intelligence seeks to give
back part of the revenues to the unaware knowledge producers.
The key idea is that any profit-making interaction a machine does has to constantly
repay whoever has produced the original knowledge used to do that interaction. To
obtain repayment, we need to work on a major issue: determine a clear knowledge
lifecycle which performs a compete tracking of the knowledge from its initial production
to the final decision processes of the machine. Hence, we need to promote artificial intelligence models that are explainable and that track back to the initial training examples
that originated a decision. In this way, it is clear why the decision is made and who has
to be rewarded with a fraction of the profit that the decision is producing.
Managing ownership in the knowledge life-cycle poses big technological and moral
issues and it is certainly more complex than simply using knowledge while forgetting
what the source is. Each interaction has to be tracked and assigned to a specific individual. Hence, the issues are: first, a clear identification of people in the web is mandatory;
second, privacy can become an overwhelming legal issue.
Finally, to pursue HIT-AI as an ecosystem for fair artificial intelligence solutions, we
need to invest in the following enabling technologies and legal aspects:
• Explainable Artificial Intelligence which is a must because, in order to reword
knowledge producers, systems need to exactly know who is responsible for a specific
5
decision;
• Symbiotic Symbolic and Distributed Knowledge Representation Models which are
needed as a large part of knowledge is expressed with symbols;
• Trusted Technologies as the knowledge life-cycle should be clear and correctly
tracked;
• Virtual Identity Protocols and Mechanisms because systems need exactly who has
to be reworded;
• Privacy Preserving Protocols and Mechanisms as, although systems need to know
who should be rewarded and why, privacy should be preserved;
• Studying Extensions of Copyright to unaware knowledge production which can be
the legal solution to safeguard the unaware knowledge producers.
4 Conclusions
Job market contraction is the dark side of the shining future promised by Artificial
Intelligence (AI) systems. Unaware skilled and unskilled knowledge workers are digging
graves for their own jobs by passionately doing their normal, everyday work. Learning
AI are extracting knowledge from their interactions. This is a gigantic knowledge theft
of the modern era.
In this paper, we proposed Human-in-the-loop Artificial Intelligence (HIT-AI) as a
fairer AI approach. As modern Robin Hoods, HIT-AI researchers should fight for a fairer
Artificial Intelligence that gives back what it steals. As skilled and unskilled workers
are producing the knowledge which Artificial Intelligence is making profit on, we need
to give back a large part of this profit to its legitimate owners.
6
References
[1] David W. Aha, Trevor Darrell, Michael Pazzani, Darryn Reid, Claude Sammut, and
Peter Stone. Proceedings of the IJCAI 2017 Workshop on Explainable Artificial
Intelligence (XAI). August 2017.
[2] Peter C Austin, Jack V Tu, Jennifer E Ho, Daniel Levy, and Douglas S Lee. Using
methods from the data-mining and machine-learning literature for disease classification and prediction: a case study examining classification of heart failure subtypes.
Journal of clinical epidemiology, 66(4):398–407, 2013.
[3] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473,
2014.
[4] Roberta Beccaceci, Francesca Fallucchi, Cristina Giannone, Francesca Spagnoulo,
and Fabio Massimo Zanzotto. Education with” living artworks” in museums. In
CSEDU (1), pages 346–349, 2009.
[5] Jean-Pierre Briot, Gaëtan Hadjeres, and François Pachet. Deep learning techniques
for music generation-a survey. arXiv preprint arXiv:1709.01620, 2017.
[6] Naom Chomsky. Aspect of Syntax Theory. MIT Press, Cambridge, Massachussetts,
1957.
[7] Michael Chui, James Manyika, and Mehdi Miremadi. Where machines could replace
humans—and where they can’t (yet). McKinsey Quarterly, July 2016.
[8] Lorenzo Ferrone and Fabio Massimo Zanzotto. Towards syntax-aware compositional distributional semantic models. In Proceedings of COLING 2014, the 25th
International Conference on Computational Linguistics: Technical Papers, pages
721–730, Dublin, Ireland, August 2014. Dublin City University and Association for
Computational Linguistics.
[9] Lorenzo Ferrone, Fabio Massimo Zanzotto, and Xavier Carreras. Decoding distributed tree structures. In Statistical Language and Speech Processing - Third
International Conference, SLSP 2015, Budapest, Hungary, November 24-26, 2015,
Proceedings, pages 73–83, 2015.
[10] Patrizia Ferroni, Fabio Massimo Zanzotto, Noemi Scarpato, Silvia Riondino, Umberto Nanni, Mario Roselli, and Fiorella Guadagni. Risk assessment for venous
thromboembolism in chemotherapy-treated ambulatory cancer patients: a machine
learning approach. Medical Decision Making, 37(2):234–242, 2017.
[11] Chelsea Gohd. Your next teacher could be a robot, February 2017.
7
[12] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In
Advances in Neural Information Processing Systems, pages 2672–2680, 2014.
[13] Andrew Gray, Mohammad Ali, Yiqi Gao, J Hedrick, and Francesco Borrelli. Semiautonomous vehicle control for road departure and obstacle avoidance. IFAC control
of transportation systems, pages 1–6, 2012.
[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in
deep residual networks. arXiv preprint arXiv:1603.05027, 2016.
[15] Eric C Huffman and Jonathon L Miner. Robotic sweeper cleaner with dusting pad,
March 18 2008. US Patent 7,346,428.
[16] Alice Kerly, Phil Hall, and Susan Bull. Bringing chatbots into education: Towards
natural language negotiation of open learner models. Knowledge-Based Systems,
20(2):177–185, 2007.
[17] B. Kim, D. M. Malioutov, K. R. Varshney, and A. Weller. Proceedings of the
2017 ICML Workshop on Human Interpretability in Machine Learning (WHI 2017).
ArXiv e-prints, August 2017.
[18] Konstantina Kourou, Themis P Exarchos, Konstantinos P Exarchos, Michalis V
Karamouzis, and Dimitrios I Fotiadis. Machine learning applications in cancer
prognosis and prediction. Computational and structural biotechnology journal, 13:8–
17, 2015.
[19] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton.
521(7553):436–444, 2015.
Deep learning.
Nature,
[20] Hod Lipson and Melba Kurman. Driverless: Intelligent Cars and the Road Ahead.
MIT Press, 2016.
[21] Zachary Chase Lipton.
abs/1606.03490, 2016.
The mythos of model interpretability.
CoRR,
[22] Todd Litman. Autonomous vehicle implementation predictions. Victoria Transport
Policy Institute, 28, 2014.
[23] Jerome M Lutin, Alain L Kornhauser, and Eva Lerner-Lam MASCE. The revolutionary development of self-driving vehicles and implications for the transportation engineering profession. Institute of Transportation Engineers. ITE Journal,
83(7):28, 2013.
[24] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis
Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602, 2013.
8
[25] Megan Murphy. Ginni rometty on the end of programming. Bloomberg Businessweek, September 2017.
[26] T. A. Plate. Distributed Representations and Nested Compositional Structure. PhD
thesis, 1994.
[27] T. A. Plate. Holographic reduced representations. IEEE Transactions on Neural
Networks, 6(3):623–641, 1995.
[28] G Revathi and VR Sarma Dhulipala. Smart parking systems and sensors: A survey. In Computing, Communication and Applications (ICCCA), 2012 International
Conference on, pages 1–5. IEEE, 2012.
[29] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural
Networks, 61:85–117, 2015.
[30] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George
Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and
tree search. Nature, 529(7587):484–489, 2016.
[31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for
large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
[32] Charles Taylor, Andrew Parker, Shek Lau, Eric Blair, Andrew Heninger, Eric Ng,
Enrico DiBernardo, Robert Witman, Michael Stout, et al. Robot vacuum cleaner,
June 30 2005. US Patent App. 11/171,031.
[33] Miroslav Trajkovic, Antonio J Colmenarez, Srinivas Gutta, and Karen I Trovato.
Computer vision based parking assistant, January 27 2004. US Patent 6,683,539.
[34] Iwan Ulrich, Francesco Mondada, and J-D Nicoud. Autonomous vacuum cleaner.
Robotics and autonomous systems, 19(3-4):233–245, 1997.
[35] Oriol Vinyals, L ukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey
Hinton. Grammar as a foreign language. In C. Cortes, N. D. Lawrence, D. D. Lee,
M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing
Systems 28, pages 2755–2763. Curran Associates, Inc., 2015.
[36] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and
tell: A neural image caption generator. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pages 3156–3164, 2015.
[37] Richard S. Wallace. The Anatomy of A.L.I.C.E., pages 181–210. Springer Netherlands, Dordrecht, 2009.
[38] David Weiss, Chris Alberti, Michael Collins, and Slav Petrov. Structured training
for neural network transition-based parsing. arXiv preprint arXiv:1506.06158, 2015.
9
[39] Joseph Weizenbaum. Eliza—a computer program for the study of natural language
communication between man and machine. Communications of the ACM, 9(1):36–
45, 1966.
[40] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan
Salakhutdinov, Richard S Zemel, and Yoshua Bengio. Show, attend and tell: Neural
image caption generation with visual attention. arXiv preprint arXiv:1502.03044,
2(3):5, 2015.
[41] Will Y Zou, Richard Socher, Daniel M Cer, and Christopher D Manning. Bilingual
word embeddings for phrase-based machine translation. In EMNLP, pages 1393–
1398, 2013.
10
| 2cs.AI
|
TEST IDEALS IN RINGS WITH FINITELY GENERATED
ANTI-CANONICAL ALGEBRAS
arXiv:1412.6453v4 [math.AG] 31 Mar 2016
ALBERTO CHIECCHIO, FLORIAN ENESCU, LANCE EDWARD MILLER, AND KARL SCHWEDE
Abstract. Many results are known about test ideals and F -singularities for Q-Gorenstein
rings. In this paper we generalize many of these results to the case when the symbolic Rees
algebra OX ⊕ OX (−KX ) ⊕ OX (−2KX ) ⊕ . . . is finitely generated (or more generally, in the
log setting for −KX − ∆). In particular, we show that the F -jumping numbers of τ (X, at )
are discrete and rational. We show that test ideals τ (X) can be described by alterations
as in Blickle-Schwede-Tucker (and hence show that splinters are strongly F -regular in this
setting – recovering a result of Singh). We demonstrate that multiplier ideals reduce to
test ideals under reduction modulo p when the symbolic Rees algebra is finitely generated.
We prove that Hartshorne-Speiser-Lyubeznik-Gabber type stabilization still holds. We
also show that test ideals satisfy global generation properties in this setting.
1. Introduction
Test ideals were introduced by Hochster and Huneke in their theory of tight closure
[HH90] within positive characteristic commutative algebra. After it was discovered that test
ideals were closely related to multiplier ideals [Smi00, Har01], a theory of test ideals of pairs
was developed analogous to the theory of multiplier ideals [HY03, Tak04]. However, unlike
multiplier ideals, test ideals were initially defined even without the hypothesis that KX
was Q-Cartier (see [DH09] for a similar theory of multiplier ideals). But the KX Q-Cartier
hypothesis is useful for test ideals, and indeed a number of central open questions are still
unknown without it. The goal of this paper is generalize results from the hypothesis that
KX is Q-Cartier to the setting where the local section ring R(−KX ) := OX ⊕ OX (−KX ) ⊕
OX (−2KX ) ⊕ . . . (also known as the symbolic Rees algebra) is finitely generated.
Most notably, perhaps the most important open problem within tight closure theory is the
question whether weak and strong F -regularity are equivalent or more generally, whether
splinters and strong F -regularity are equivalent (from the characteristic zero perspective,
splinters, weak and strong F -regularity are competing notions of singularities analogous
to KLT singularities that are all known to coincide in the Q-Gorenstein case). These are
known to be equivalent under the KX Q-Cartier hypothesis and under some other conditions
[Sin99, LS01, LS99, AM99]. Previously A. K. Singh announced a proof that splinters with
R(−KR ) finitely generated are strongly F -regular [Sin14]. We recover a new proof of this
result and in fact show something stronger. We prove that the (big) test ideal is equal to
the image of a multiplier-ideal-like construction involving alterations.
Theorem A (Theorem 5.6, Corollary 5.7). Suppose that X is a normal F -finite integral
scheme and that ∆ on X is an effective Q-divisor such that S = R(−KX − ∆) is finitely
generated. Then there exists an alteration π : Y −→ X from a normal Y , factoring through
2010 Mathematics Subject Classification. 14J17, 13A35, 14B05.
Key words and phrases. Anticanonical, test ideal, multiplier ideal.
The fourth named author was supported in part by the NSF FRG Grant DMS #1265261/1501115, NSF
CAREER Grant DMS #1252860/1501102 and a Sloan Fellowship.
1
2
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
X ′ = Proj S so that
τ (X, ∆) = Image π∗ OY (⌈KY − π ∗ (KX + ∆)⌉) −→ OX .
◦ If X is of finite type over a perfect field, one may take Y to be regular by [dJ96].
◦ Alternately, one may take π to be a finite map (in which case Y is almost certainly
not regular).
As a consequence we obtain that
\
τ (X, ∆) =
Image π∗ OY (⌈KY − π ∗ (KX + ∆)⌉) −→ OX
π:Y −
→X
where π runs over all alterations with Y −→ X factoring through X ′ . Alternately, one can
run π over all finite maps. If additionally X is of finite type over a perfect field, then one
can run π over all regular alterations factoring through X ′ .
Actually we prove a stronger theorem allowing for triples (X, ∆, at ) but we don’t include
it here to keep the statement simpler. Note that in characteristic zero, the same intersection
over regular alterations characterized multiplier ideals by [DH09] at least after observing
Remark 2.26.
Inspired by the analog with multiplier ideals [ELSV04], there has been a lot of interest
in showing that the jumping numbers of test ideals are rational and without limit points
[BMS08, BMS09, KLZ09, BSTZ10, Bli13, ST14b, KSSZ14]. At this point, we know that
the F -jumping numbers are discrete and rational for any F -finite scheme X with KX QCartier. We also know discreteness if R(−KX ) is finitely generated and X = Spec R is
the spectrum of a graded ring [KSSZ14]. On the other hand we know that the jumping
numbers of J (X, ∆, at ) are discrete and rational if R(−KX − ∆) is finitely generated (see
Remark 2.26). We prove the following:
Theorem B (Theorem 3.5, Proposition 4.7). Suppose that (R, ∆) is a pair such that
R(−KR − ∆) is finitely generated. Then for any ideal a ⊆ R the F -jumping numbers
of τ (R, ∆, at ) are rational and without limit points.
We prove the discreteness result in two ways. First pass to the local section ring /
symbolic Rees algebra, where the pullback of −KX − ∆ is Q-Cartier. We then show that
the test ideal of the symbolic Rees algebra restricts to the test ideal of the original scheme.
Alternately, in Section 4, we prove the discreteness result for projective varieties by utilizing
the theory developed by the first author and Urbinati [CU15]. In particular, we show a
global generation result for test ideals, Theorem 4.5, which immediately implies the test
ideal result.
Another setting where the Q-Gorenstein hypothesis is used is in the study of non-F -pure
ideals. Recall that if R is F -finite, normal and Q-Gorenstein with index not divisible by p,
then it follows from [HS77, Lyu97, Gab04] that the images of the evaluation-at-1 map
HomR (F∗e R, R) −
→R
stabilize for sufficiently large e. This stable image gives a canonical scheme structure to
the non-F -pure locus of a variety. We generalize this to the case that R(−KR ) is finitely
generated (which includes the case where the index of −KR is divisible by p).
Theorem C (Corollary 3.4, Theorem 3.9, Theorem 4.8). Suppose that R is an F -finite
normal domain and that B ≥ 0 is a Q-divisor with Weil-index not divisible by p. If the
anti-log-canonical algebra R(−KR −B) is finitely generated, then the image of the evaluation
at 1 map HomR (F∗e R((pe − 1)B), R) −→ R stabilizes for e sufficiently divisible.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
3
Again we give several different proofs of this fact utilizing different strategies as above.
Finally, we also show:
Theorem D (Theorem 6.4). Suppose that X is a normal quasi-projective variety over an
algebraically closed field of characteristic zero. Further suppose that ∆ ≥ 0 is a Q-divisor
such that R(−KX − ∆) is finitely generated and also suppose that a ⊆ OX is an ideal and
t ≥ 0 is a rational number. Then
J (X, ∆, at )p = τ (Xp , ∆p , atp )
for p ≫ 0.
This should be compared with [dFDTT15] where the analogous result is shown under the
hypothesis that KX + ∆ is numerically Q-Cartier. This numerically Q-Cartier condition is
somewhat orthogonal to the finite generation of R(−KX −∆). In particular, if R(−KX −∆)
is finitely generated and −KX − ∆ is numerically Q-Cartier then it is not difficult to see
that −KX − ∆ is Q-Cartier, see also [BdFFU15].
Remark 1.1. A previous version of this paper included an incorrect statement in Lemma 2.10.
This version, which also corrects the published version, fixes the statement by making it
weaker. Fortunately, we only needed the weaker statement in all our applications.
Acknowledgements: The authors would like to thank Tommaso de Fernex, Christopher
Hacon, Nobuo Hara, Mircea Mustaţă and Anurag K. Singh for several useful discussions. We
would also like to thank (Juan) Felipe Pérez for several useful comments on a previous draft
of this paper. Finally, we would like to thank the referee for numerous valuable comments
and for pointing out a mistake in a lemma (previously Lemma 2.15, now removed).
2. Preliminaries
In this section we recall the basic properties that we will need of test ideals, local section
rings1, as well as the theory of positivity for non-Q-Cartier divisors as developed by UrbinatiChiecchio [CU15]. We conclude by stating a finite generation result for local section rings of
threefolds in positive characteristic (as a consequence of recent breakthroughs in the MMP).
Setting 2.1. Throughout this paper, all rings will be assumed to be Noetherian, of equal
characteristic p > 0, and F -finite (which implies that they are excellent and have dualizing
complexes [Kun76, Gab04]). All schemes will be assumed to be Noetherian, F -finite, separated and have dualizing complexes. For us a variety is a separated integral scheme of finite
type over an F -finite field. For any scheme X, we use F : X −
→ X to denote the absolute
Frobenius morphism. We also make the following universal assumption:
q
q
(†)
F !ω ∼
=ω .
X
X
This holds for all schemes of essentially finite type over an F -finite field (or even of essentially
of finite type over an F -finite local ring).
Frequently we will also consider divisors on schemes X. Whenever we talk about divisors
∆ on X, we make the universal assumption that X is normal and integral. In particular,
whenever we consider a pair (R, ∆) or (X, ∆), then R or X is implicitly assumed to be
normal.
We make one remark on some nonstandard notation that we use. If R is a normal domain
and D is a Weil divisor on X = Spec R, then we use R(D) to denote the fractional ideal
H 0 (X, OX (D)) ⊆ K(R).
1Also called (divisorial) symbolic Rees algebras in the commutative algebra literature.
4
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
2.1. Test ideals and F -singularities. We now recall the definitions and basic properties
of test ideals. While test ideals were introduced in [HH90], we are technically talking about
the big/non-finitistic test ideal from [LS01, Hoc07]. The particular definition of the test
ideal presented here can be found in [BS13, Definition 9.3.8] among other places.
Definition 2.2 (Test ideals). Suppose that R is an F -finite normal domain, ∆ ≥ 0 is a
Q-divisor, a ⊆ R is a non-zero ideal sheaf and t ≥ 0 is a real number. The test ideal
τ (X, ∆, at )
is the unique smallest nonzero ideal J ⊆ R such that for every e > 0 and every
φ ∈ (F∗e a⌈p
e t⌉
) · HomR (F∗e R(⌈(pe − 1)∆⌉), R) ⊆ HomR (F∗e R, R)
we have that φ(F∗e J) ⊆ J.
If ∆ = 0 then we leave it out writing τ (X, at ). If a = R or t = 0 then we write τ (R, ∆).
It is not obvious that the test ideal exists. However,
Pit can
P be shown that there exists
c ∈ R such that for each 0 6= d ∈ R, we have that c ∈ e>0 φ φ(F∗e (dR)) where φ varies
e
over (F∗e a⌈t(p −1)⌉ ) · HomR (F∗e R(⌈(pe − 1)∆⌉), R), see [Sch11, Lemma 3.21]. This element
c is then called a (big) (R, ∆, at )-test element. We then immediately obtain the following
construction of the test ideal.
Lemma 2.3. With notation as in Definition 2.2, if c is a big (R, ∆, at )-test element, then
XX
τ (R, ∆, at ) =
φ(F∗e (cR))
e≥0 φ
e
φ ranges over elements of (F∗e a⌈t(p −1)⌉ )
e
also range φ over elements of (F∗e a⌈tp ⌉ )
· HomR (F∗e R(⌈(pe − 1)∆⌉), R).
where again
One may
· HomR (F∗e R(⌈(pe − 1)∆⌉), R). Alternately, one may replace e ≥ 0 with e ≫ 0.
Finally, we also have that for any sufficiently large Cartier divisor D ≥ 0 that
X
e
Tre F∗e a⌈tp ⌉ · OX (KX − pe (KX + ∆) − D) .
τ (R, ∆, at ) =
e≥e0
Proof. For the first statement, it is easy to see that c is contained in any ideal satisfying
e
the condition φ(F∗e J) ⊆ J for all φ ∈ (F∗e a⌈t(p −1)⌉ ) · HomR (F∗e R(⌈(pe − 1)∆⌉), R). Hence
so is the sum. Thus the sum is the smallest such ideal.
P
e
e
For the second statement replacing a⌈t(p −1)⌉ with a⌈tp ⌉ obviously we have the τ ⊇
containment. Notice that if c is a test element, then so is dc for any 0 6= d ∈ R. Hence
⌈t(pe −1)⌉ ⊆ a⌈tpe ⌉ for all e. The
one can form the
P original sum with cd for some d so that da
inclusion τ ⊆
follows.
For the e ≫ 0 statement, notice that if J is the sum for e ≫ 0, then we still have
φ(F∗e J) ⊆ J. The final characterization of the test ideal follows immediately from the fact
that
e
e
a⌈tp ⌉ · OX (KX − pe (KX + ∆) − divX (c)) = a⌈tp ⌉ · OX (−D) · H omOX (F∗e OX (⌊pe ∆⌋), OX ).
We notice that any difference coming from the fact that we round down instead of round
up can be absorbed into the difference between D and divX (c).
We also recall some properties of the test ideal for later use.
Lemma 2.4. Suppose that (R, ∆, at ) is as in Definition 2.2. Then:
(a) The formation of τ (R, ∆, at ) commutes with localization and so one can define
τ (X, ∆, at ) for schemes as well.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
5
(b) If s ≥ t, then τ (R, ∆, as ) ⊆ τ (R, ∆, at ).
(c) For any t ≥ 0, there exists an ε > 0 so that if s ∈ [t, t + ε), then τ (X, ∆, at ) =
τ (X, ∆, as ).
(d) If 0 6= f ∈ R and H = V (f ) is the corresponding Cartier divisor, then
f τ (X, ∆, at ) = τ (X, ∆, at ) ⊗ OX (−H) = τ (X, ∆ + H, at ).
Proof. Part (a) follows immediately from Lemma 2.3. Part (d) follows similarly (use the
projection formula). Part (b) is obvious also from Lemma 2.3.
For part (c), this is [BS13, Exercise 9.12]. Let us quickly sketch the proof since we
do not know of a reference where this is addressed in full generality. Choose c 6= 0 a
test element. P
It is P
easy to choose c that works for all all s ∈ [t, t + 1]. We then write
τ (R, ∆, at ) = e≥0 φ φ(F∗e (cdR)) for d some element in a. This sum is a finite sum, say
for e = 0 to e = m. Let ε = p1m . Then
τ (R, ∆, at ) =
m X
X
φ(F∗e (cdR)) =
e
ψ(F∗e (cda⌈tp ⌉ )) ⊆
ψ now runs over HomR (F∗e R(⌈(pe − 1)∆⌉), R).
then see that ⌈t(pe − 1)⌉ + 1 ≥ ⌈(t + ε)pe ⌉ for e
τ (R, ∆, at ) ⊆
m X
X
m X
X
ψ(F∗e (ca⌈tp
e ⌉+1
))
e=0 ψ
e=0 ψ
e=0 φ
where
We
m X
X
≤ m and so
e
ψ(F∗e (ca⌈(t+ε)p ⌉ )) ⊆ τ (R, ∆, at+ε ).
e=0 ψ
The other containment was handled in (b).
Finally, we make one more definition related to test ideals.
Definition 2.5. A triple (X, ∆, at ) as in Definition 2.2 is called strongly F -regular if
τ (R, ∆, at ) = R.
We briefly also recall some formalities of p−e -linear maps and connections with divisors.
Lemma 2.6. Suppose that X = Spec R is an F -finite normal scheme.
(a) There is a bijection between effective divisors ∆ such that (pe − 1)(KR + ∆) ∼ 0
and elements φ of HomR (F∗e R, R) modulo pre-multiplication by units. We use the
following notation for this correspondence:
φ →
7
∆φ ,
∆ →
7
φ∆ .
(b) If φ ∈ HomR (F∗e R, R) corresponds to a Q-divisor ∆φ , then the map
ψ( ) = φ(F∗e (d ·
(c)
))
corresponds to the divisor ∆ψ = ∆φ + pe1−1 div(d).
If φ ∈ HomR (F∗e R, R) and φn := φ ◦ (F∗e φ) ◦ . . .
(n−1)e
◦ (F∗
) : F∗ne R −→ R, then
∆ φn = ∆ φ .
(d) Using the bijection of (a), if ∆ ≥ 0 is any effective R-divisor, then the elements φ
of HomR (F∗e R(⌈(pe − 1)∆⌉) ⊆ HomR (F∗e R, R) (modulo multiplication by units) are
in bijection with divisors ∆φ with ∆φ ≥ ∆ and of course with (pe − 1)(KR + ∆) ∼ 0.
Proof. (a) is just [Sch09, Theorem 3.13]. (b) is a straightforward exercise, see [BS13, Exercise 4.11]. (c) is [Sch09, Theorem 3.11(e)]. (d) is also not difficult to check, see for instance
[ST14a, Definition 6.8].
6
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
2.2. Local section rings of divisors / symbolic Rees algebras. Suppose that X is an
F -finite normal integral Noetherian scheme and Γ is a Q-divisor on X. Then one can form
M
S = R(X, Γ) =
OX (⌊iΓ⌋).
i∈Z≥0
Additionally for any integer n > 0 we use S (n) := R(X, nΓ) to denote the nth Veronese
κ
subalgebra. Note that there is a canonical map OX −
→ S and dually a map Spec S −
→ X of
schemes (note S may not be Noetherian). If S (or equivalently Y = Spec S) is Noetherian,
µ
then we also have Proj S −
→ X. These maps are very well behaved outside of codimension
µ
≥ 2. We recall that the map Proj S −
→ X is called the Q-Cartierization of Γ. It is a small
f
projective morphism Y −
→ X such that the strict transform f∗−1 Γ is Q-Cartier and f -ample.
Moreover, such a map exists if and only if S is Noetherian, [KM98, Lem. 6.2]. When S is
finitely generated, both Spec S and Proj S are normal, see for instance [Dem88] and also
see [Wat81].
Lemma 2.7. Suppose that S is finitely generated and W ⊆ X is a closed subset of codimension ≥ 2. Then κ−1 W and µ−1 W are also codimension ≥ 2 in Spec S and Proj S
respectively. Additionally µ is an isomorphism outside a closed codimension ≥ 2 subset of
X and if Γ is integral, then κ is an A1 -bundle outside a set of codimension 2.
As a consequence, if D is any Q-divisor on X, then we have canonical pullbacks µ∗ D and
∗
κ D.
Proof. Since S is a symbolic Rees algebra (of a module of rank 1), the map µ : Proj S → X
is small, [KM98, Lem. 6.2]. The case for κ can be verified locally on X.
We begin under the assumption that Γ is integral. Let U = Spec R ⊆ X and S ′ = S U .
Suppose that κ−1 W has a codimension 1 component whose support is defined by a prime
height 1 ideal Q in S. By [GHNV90, Lemma 4.3], Q ∩ R is height zero or 1. But this is
impossible since it defines a subset of W , a set of codimension 2. For the case when Γ is
not integral, we observe that the result holds for the Veronese algebra S (n) = R(X, nΓ) for
sufficiently divisible n. But S is a finite S (n) algebra (see the proof of [GHNV90, Lemma
2.4]) and the result follows.
The fact that κ is a A1 bundle, at least outside a set of codimension 2, follows immediately
from the fact that in that case, Γ is an integral Cartier divisor and so the section ring S
looks locally like OX [t] outside a set of codimension 2.
Remark 2.8. When X is separated, the pullback µ∗ Γ coincides with the pullback of de Fernex
and Hacon, [DH09]; see Remark 2.12.
We will be very interested in proving that various section rings are finitely generated and
so recall:
Lemma 2.9. With notation as defined at the start of Section 2.2,
(a) S is finitely generated if and only if S (n) is finitely generated for some (equivalently
any) n > 0.
(b) Suppose that g : Y −→ X is a finite dominant map from another normal integral
Noetherian scheme Y . Let T = R(Y, g∗ Γ). Then T is finitely generated if S is
finitely generated.
Proof. Part (a) is exactly [GHNV90, Lemma 2.4] (although it can also be found in numerous
other sources). For (b) we do not know a good reference but we sketch a proof here. By
(a), we may assume that Γ is integral. It is also harmless to assume that X = Spec A is
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
7
affine and hence so is Y = Spec B. Then we can pass to the category of commutative rings
so that S and T are actually rings (and not sheaves of rings). In particular, we suppress all
g∗ notation that we might otherwise need. We have the diagram
A
/B
/T
S
First choose a single element c ∈ A so that S[c−1 ] ∼
= A[c−1 , t] and T [c−1 ] ∼
= B[c−1 , t]
−1
−1
(here we identify the ts). It follows that S[c ] ⊆ T [c ] is finite and hence integral and
K(S) −
→ K(T ) is a finite extension as well. Let T ′ to be the integral closure of S inside T .
We want to show that T = T ′ which will complete the proof.
Recall we already assumed that Γ is integral. Let W ⊆ X be a closed set of codimension
2 outside of which Γ is Cartier. Consider the functor H 0 (X \ W, ) applied to all of
the rings (or sheaves of rings) involved. As T is a direct sum of reflexive OY -modules,
H 0 (X \ W, T ) = H 0 (Y \ g−1 (W ), T ) is just the global sections of T by Hartog’s Lemma for
reflexive sheaves [Har94]. Thus H 0 (X \ W, T ) is identified with T since X and Y are affine.
On the other hand, κ−1 W is a codimension ≥ 2 subset of Spec S, outside of which T and
′
T obviously agree. Hence H 0 (X \ W, T ) = H 0 (X \ W, T ′ ) = H 0 (Spec S \ κ−1 W, T ′ ). But
T ′ is normal and so we also have that H 0 (Spec S \ κ−1 W, T ′ ) = T ′ . We have just shown
that T = T ′ as desired.
We also will need to understand the canonical divisors of Spec S and Proj S.
Lemma 2.10. Continuing with notation from the start of Section 2.2, assuming that S is
finitely generated, then KProj S ∼ µ∗ KX . If additionally, Γ is a Weil divisor then, locally
on the base, we have that KSpec S ∼ κ∗ KX + κ∗ Γ. In particular, if Γ = −KX − B, then
KSpec S ∼ κ∗ (−B) so that KSpec S + κ∗ B ∼ 0. Thus if Γ = −KX then S is quasiGorenstein.2
Proof. Recall that κ is an A1 -bundle outside a set of codimension 2 and hence κ∗ makes
sense. The computation of KSpec S can be found in [GHNV90, Theorem 4.5]. The initial
statement that KProj S = µ∗ KX is obvious since µ is small.
2.3. Positivity for non-Q-Cartier divisors. In this section we will recall some definitions
and results of [CU15]. Let us recall that, if f : Y → X is a morphism of schemes, a coherent
sheaf F on Y is relatively globally generated, or f -globally generated, if the natural map
f ∗ f∗ F → F is surjective. If Y is a normal scheme and D is a Weil divisor on Y , it
might be that, for example, OY (D) is f -globally generated, but OY (2D) is not. To account
for such pathologies we have to work asymptotically: we will say that a Q-divisor D is
relatively asymptotically globally generated, or f -agg, if OY (mD) is f -globally generated for
all positive m sufficiently divisible.
Let f : Y → X be a projective morphism of normal Noetherian schemes. A Q-Weil
divisor D on X is f -nef if, for every f -ample Q-Cartier Q-divisor A on Y , D + A is f -agg; if
X = Spec k, we will simply say that D is nef, [CU15, Def. 2.4]. The Q-divisor D is f -ample
if, for every ample Q-Cartier Q-divisor A on Y , there exists b > 0 such that bD − A is f -nef
and the algebra of local sections R(X, D) is finitely generated; if X = Spec k, we will say
that D is ample, [CU15, Def. 2.14]. Notice that when D is Q-Cartier, these notions coincide
with the usual ones of nefness and amplitude. We remark that amplitude for Weil divisors
2Also called 1-Gorenstein.
8
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
is given by two conditions: a positivity one – which is based on the fact that the regular
ample cone is the interior of the nef cone – and a technical one – on the finite generation
of the algebra of local sections. These two conditions are independent; in particular there
are examples of Weil divisors A satisfying the positivity condition, but with algebra of local
sections R(X, A) not finitely generated, [CU15, Example 2.20].
These notions of positivity behave very much like in the Q-Cartier world: for example,
if A is an ample Weil Q-divisor, and D is a globally generated Q-Cartier divisor, D + A is
ample, [CU15, Lem. 2.18(i)].
Lemma 2.11. Let E be a Q-divisor on a normal Noetherian projective scheme X over
a field k. If the algebra of local section R(X, E) is finitely generated then there exists a
Cartier divisor L such that L + E is an ample Weil divisor.
Proof. Notice that n(L + E) is ample for some/every n > 0, if and only if L + E is ample,
[CU15, Lem. 2.18(b)]. So, without loss of generality, we can assume that R(X, E) is
generated in degree 1 and that E is integral. Let H be an ample Cartier divisor. By
definition, there exists m > 0 such that OX (mH + E) is globally generated. There is a
surjection
OX (E)⊗n ։ OX (E)·n = OX (nE),
where the last equality is a consequence of the assumption on the finite generation of the
algebra of local sections. Thus, for each n > 0, OX (n(mH+E)) is globally generated, that is,
mH + E is asymptotically globally generated. Moreover, since H is Cartier, R(X, mH + E)
is also finitely generated. By [CU15, Lem. 2.18(i)], mH + E + H = (m + 1)H + E is
ample.
The main characterization of the above positivity is in terms of their Q-Cartierization.
Let X is a normal projective Noetherian scheme over an algebraically closed field k, and
µ
let D be a Q-divisor with S = R(X, D) finitely generated. Let Proj S −
→ X. Notice that
′
D ′ = µ−1
∗ D is Q-Cartier. Then D is nef/ample if and only if so is D , [CU15, Theorems 3.3
and 3.6].
Using this characterization, Urbinati and the first author proved Fujita vanishing for
locally free sheaves, [CU15, Cor. 4.2]. Let X be a normal projective Noetherian scheme
over an algebraically closed field k, let A be an ample Q-divisor on X and let F be a locally
free coherent sheaf on X. There exists an integer m(A, F ) such that
H i (X, F ⊗ OX (mA + D)) = 0
for all positive m divisible by m(A, F ), all nef Cartier divisors D, and all i > 0.
2.4. Pullback of Weil divisors. Let f : Y → X be a proper birational morphism of
normal Noetherian separated schemes. In [DH09], de Fernex and Hacon introduced a way
of pulling back a Weil divisor on X via f .
For any Weil divisor D on X, the ♮-pullback of D along f , denoted by f ♮ D, is the Weil
divisor on Y such that
(2.11.1)
OY (−f ♮ D) = (OX (−D) · OY )∨∨ ,
[DH09, Def 2.6]. The negative sign appearing is so that, when D is effective, we are pulling
back the ideal defining it as a subscheme. The pullback of D along f is
f ∗ D = lim inf
m
f ♮ (m!D)
f ♮ (mD)
= lim
m
m
m!
coefficient-wise.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
9
The above is well-defined, the infimum limit over m is a limit over m! and an R-divisor,
[DH09, Lem. 2.8 and Def. 2.9]. Moreover, the above definition of f ∗ coincides with the
usual one whenever D is Q-Cartier, [DH09, Prop. 2.10].
Remark 2.12. If f : Y → X is a small, projective birational morphism, then f ∗ D = f∗−1 D
This notion of pullback is not quite functorial, unfortunately. Let f : Y → X and
g : V → Y be two birational morphisms of normal Noetherian separated schemes, and
D be a Weil divisor on X. The divisor (f g)♮ (D) − g♮ f ♮ (D) is effective and g-exceptional.
Moreover, if OX (−D) · OY is an invertible sheaf, (f g)♮ (D) = g♮ f ♮ (D), [DH09, Lem. 2.7].
Lemma 2.13. Let X be a normal Noetherian scheme. Let D be a Weil divisor on X,
µ
such that S = R(X, D) is finitely generated; let X ′ = Proj S −
→ X and let D ′ := µ−1
∗ D.
′
′
′
Then OX (mD) · OX = OX (mD ) for all positive m sufficiently divisible. In particular
OX (mD) · OX ′ is reflexive for m sufficiently divisible.
Proof. Since D ′ is µ-ample, see [KM98, Lemma 6.2], OY ′ (mD ′ ) is µ-globally generated for
all positive m sufficiently divisible, that is, the natural map
µ∗ µ∗ OX ′ (mD ′ ) → OX ′ (mD ′ )
is surjective for positive m sufficiently divisible. Since µ is small, µ∗ OX ′ (mD ′ ) = OX (mD)
for all integers m (this is well-known, for a proof see [Chi14, Lemma 2.8]). Thus, for all
positive m sufficiently divisible, we have a surjection
µ∗ OX (mD) ։ OX ′ (mD ′ ).
Notice that OX ′ · OX (mD) is isomorphic to the quotient of µ∗ OX (mD) by its torsion,
[Har77, Caution II.7.12.2]. Since OX ′ (mD ′ ) is torsion free, the above surjection induces a
surjection
OX ′ · OX (mD) ։ OX ′ (mD ′ ).
On the other hand, since µ is small, for all integers m, OX ′ (mD ′ ) = OX ′ · OX (mD) ∨∨ ;
since OX ′ · OX (mD) is torsion-free, we have a natural inclusion
OX ′ · OX (mD) ֒→ OX ′ (mD ′ ).
Lemma 2.14. Let X be a normal Noetherian separated scheme. Let D be a Weil divisor
µ
on X, such that R(X, −D) is finitely generated, and let Proj R(X, −D) −
→ X. Let f :
g
µ
Y → X be any birational morphism factoring as Y −
→ Proj R(X, −D) −
→ X, with Y
a normal Noetherian separated scheme. Then, for any positive m sufficiently divisible,
f ♮ (mD) = g♮ µ♮ (mD) = g∗ µ♮ (mD). Therefore, f ∗ (D) = g∗ µ∗ (D).
Proof. This is an application of Lemma 2.13 which we now explain. Consider the following chain of equalities. The first and last equalities are by definition and the third is by
Lemma 2.13 since m is sufficiently divisible:
=
=
=
=
OY (−f ♮ mD)
OX (−mD) · OY
OX (−mD) · OX ′ · OY
OX ′ (−µ♮ mD) · OY
OY (−g ♮ µ♮ mD)
This proves the first statement. The final statement is a consequence of the fact that
1 ♮mD
= µ∗ (D) for m sufficiently divisible by our finite generation hypothesis.
mµ
10
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
g
f
→Y −
→X
Lemma 2.15. Suppose we have a composition of birational morphisms f ′ : Y ′ −
∗
′∗
between normal varieties and D is a Weil on X, then f D = g∗ f D.
Proof. To check the identity it suffices to show ordE (f ∗ D) = ordE (g∗ f ′∗ D) for each prime
divisor E on Y . The generic point of each prime divisor E on Y gives rise to a DVR
(AE , mE ). For any sufficiently divisible positive integer m, set νE (mD) the power of mE
agreeing with OX (−mD)AE ; so by definition ordE (f ∗ D) = lim νE (mD)/m. A similar calculation computes f ′∗ D on Y ′ . Finally, as g is birational g∗ f ′∗ D keeps the same coefficients
on divisors that are not contracted by g, thus, ordE (f ∗ D) = ordE (g∗ f ′∗ D) for each prime
divisor E on Y, as desired.
We define multiplier ideals in a way which is a slight generalization of the one of [DH09].
Definition 2.16. Let X be a normal quasi-projective variety
an algebraically closed
Q aover
k
field of characteristic zero, ∆ an R-Weil divisor, and I = Jk a formal R-linear product
of non-zero fractional ideal sheaves. The collection of the data of X, ∆ and I will be called
a triple,
Q and it will be denoted by (X, ∆, I). We say that the triple is effective if ∆ ≥ 0,
I = Jkak where all the Jk s are ideals and ak ≥ 0 for all k.
Remark 2.17. Notice that we do not assume that KX + ∆ is R-Cartier.
Remark 2.18. A triple is effective if and only if (X, ∆ + I) is an effective pair in the sense
of [DH09, Definition 4.3].
Definition 2.19. Let (X, ∆, I) be an effective triple, and let m be a positive integer. Let
f : Y → X be a log resolution Q
of the pair (X, OX (−m(KX + ∆)) + I), [DH09, Definition
4.1 and Theorem 4.2]. Let I = Jkak be a formal product, and OY · Jk = OX (−Gk ). We
define the sheaf
X
1
ak Gk ⌉).
Jm (X, ∆, I) := f∗ OY (⌈KY − f ♮ (m(KX + ∆)) −
m
Remark 2.20. The reason for this new notation is that our notation is slightly more general
than the one of [DH09]. In particular, de Fernex and Hacon did not include a boundary
divisor term.
This might cause some confusion since the reader might think one could absorb the divisor
∆ into the ideal I (indeed, what is a divisor but a formal combination of height 1 ideals).
Unfortunately, this does not yield the same object (and in particular, does not yield the
usual multiplier ideal even when KX + ∆ is Q-Cartier). The difference is that asymptotics
are already built into Q-Cartier KX + ∆ whereas no asymptotics are built into I in [DH09].
In particular, let I (∆) denote the formal product of ideals corresponding to ∆ in the
obvious way. Then in general, Jm (X, ∆, I) ⊇ Jm (X, I · I (∆)): we have
X
1
ak Gk ⌉)
Jm (X, I · I (∆)) = f∗ OY (⌈KY − f ♮ (mKX ) − f ♮ ∆ −
m
X
1
1
⊆ f∗ OY (⌈KY − f ♮ (mKX ) − f ♮ (m∆) −
ak Gk ⌉)
m
m
X
1
ak Gk ⌉)
⊆ f∗ OY (⌈KY − f ♮ (mKX + m∆) −
m
= Jm (X, ∆, I).
Here the first containment is [DH09, Lemma 2.8] and the second is a consequence of [DH09,
Remark 2.11].
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
11
Lemma 2.21. Let (X, ∆, I) be an effective triple. The sheaf Jm (X, ∆, I) is a (coherent)
sheaf of ideals on X, and its definition is independent of the choice of f .
Proof. The proof proceeds as in the proof of [DH09, 4.4].
Lemma 2.22. Let (X, ∆, I) be an effective triple. The set of ideal sheaves {Jm (X, ∆, I)}m≥1
has a unique maximal element.
Proof. For any positive integers m, q, Jm (X, ∆, I) ⊆ Jmq (X, ∆, I) by [DH09, Remark 3.3]
(by the previous lemma the two ideals can be computed on a common resolution). The
unique maximal ideal exists by Noetherianity.
Definition 2.23. Let (X, ∆, I) be an effective triple. We will call the unique maximal
element of {Jm (X, ∆, I)}m≥1 the multiplier ideal of the triple (X, ∆, I), and we will denote
it by J (X, ∆, I).
Remark 2.24. In the case when ∆ = 0 we write J (X, I) = J (X, 0, I) and then our definition
agrees with the one in [DH09].
Corollary 2.25. Working in characteristic zero, suppose R(−KX −∆) is finitely generated
and X ′ = Proj R(−KX − ∆) just as before. Let a be any ideal sheaf on X, then we have
J (X, ∆, at ) = µ∗ J (X ′ , µ∗ ∆, (a · OX ′ )t ).
Proof. Let π : ProjX R(X, D) = X ′ → X and ∆′ := π∗−1 ∆. It is enough to show that, for
every m satisfying the result of Lemma 2.13,
π∗ J ((X ′ , ∆′ ); tI · OX ′ ) = Jm (X, ∆, tI).
Q
Let f : Y → X be a log resolution of (X, m∆ + I) factoring through X ′ . Let I = Jkak
and let Jk · OY = O(−Gk ); since OY · I = OY · OX ′ · I, f is a log resolution of ((X ′ , ∆′ ); I).
Let g : Y → X ′ . Since (X ′ , ∆′ ) is a log pair, the multiplier ideal J ((X ′ , ∆′ ); tI · OX ′ ) is
X
tak Gk ⌉).
J ((X ′ , ∆′ ); tI · OX ′ ) = g∗ OY (⌈KY − g ∗ (KX ′ + ∆′ ) −
On the other hand, for each m ≥ 1, the multiplier ideal Jm (X, ∆; tI) is
X
1
Jm (X, ∆, tI) = f∗ OY (⌈KY − f ♮ (m(KX + ∆)) −
tak Gk ⌉).
m
For each m satisfying OX (m(−KX − ∆)) · OX ′ = OX ′ (m(−KX ′ − ∆′ )), by Lemma 2.14,
f ♮(m(KX + ∆)) = g ∗ π ♮ (m(KX + ∆)) = g ∗ (m(KX ′ + ∆′ )) = mg ∗ (KX ′ + ∆′ ).
Therefore, for each m satisfying OX (m(−KX − ∆)) · OX ′ = OX ′ (m(−KX ′ − ∆′ )),
X
1
tak Gk ⌉) =
Jm (X, ∆, tI) = f∗ OY (⌈KY − f ♮ (m(KX + ∆)) −
m
X
tak Gk ⌉) =
= f∗ OY (⌈KY − g∗ (KX ′ + ∆′ ) −
X
tak Gk ⌉) =
= π∗ g∗ OY (⌈KY − g∗ (KX ′ + ∆′ ) −
= π∗ J ((X ′ , ∆′ ); tI · OX ′ ).
Remark 2.26. With the assumptions of Corollary 2.25, it follows immediately that the
jumping numbers of J (X, ∆, at ) are rational and without limit points. Recall that the
12
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
jumping numbers are real numbers t0 ≥ 0 such that J (X, ∆, at0 −ε ) 6= J (X, ∆, at0 ) for any
ε > 0. It also follows that
\
TrY /X
J (X, ∆, at ) =
Image π∗ OY (⌈KY − π ∗ (KX + ∆) − tG⌉) −−−−→ OX
π:Y −
→X
where π runs over all alterations factoring through µ : X ′ −
→ X such that aOY = OY (−G)
is invertible.
2.5. Finite generation of local section rings for threefolds in characteristic p > 5.
Of course, one might ask how often it even happens that a section ring R(X, D) is finitely
generated. For (pseudo-)rational surface singularities of any characteristic, it is known that
D is always locally torsion in the divisor class group, see [Lip69, Theorem 17.4], and so
obviously R(D) is finitely generated. However, for threefolds, rational singularities are not
enough by an example of Cutkosky, [Cut88], even if they are additionally log canonical. Of
course, in characteristic zero, the finite generation of these section rings holds for KLT X of
any dimension by the minimal model program [Kol08, Theorem 92] (and of course is closely
linked with the existence of flips).
Using the recent breakthroughs on the minimal model program for threefolds in characteristic p > 5 [Bir13, Xu15, CTX13, HX15, BW14] one can prove finite generation of
R(X, D) in some important cases (again in dimension 3, characteristic p > 5). The proof
is essentially the same as it is in characteristic zero, see [Kol08, Exercises 108 and 109], but
we reproduce it here for the reader’s convenience.
Theorem 2.27. Let (X, ∆) be a KLT pair of dimension 3 with KX + ∆ Q-Cartier over an
algebraically closed field k of char p > 5. Then, for any Q-divisor D, the algebra R(X, D)
is finitely generated.
b−
Proof. Let φ : X
→ X be a small Q-factorialization of X, which exists by [Bir13, Theorem
b
b to be the strict transforms of ∆ and D on X.
b Then we notice that
1.6]. Set ∆ and D
b + (1/m)D
b is
b ∆
b + (1/m)D)
b is KLT for m ≫ 0. By [Bir13, Theorem 1.3], since K b + ∆
(X,
X
big over X, we see that
M
b + (1/m)D))
b
φ∗ OXb (n(KXb + ∆
n≥0
is finitely generated. Since φ is small, this implies that
M
OX (n(KX + ∆ + (1/m)D))
n≥0
is finitely generated as well (since the algebras are the same). However, by taking a high
Veronese, and recalling that KX + L
∆ is Q-Cartier (and so locally, contributes nothing to
finite generation) we conclude that n≥0 OX (nD) is finitely generated as desired.
Of course this also implies that strongly F -regular pairs have finitely generated local
section algebras since they are always KLT for an appropriate boundary by [SS10].
3. Stabilization, discreteness and rationality via Rees algebras
In this section we aim to prove discreteness and rationality of jumping numbers of test
ideals as well as Hartshorne-Speiser-Lyubeznik-Gabber-type stabilization results under the
hypothesis that the anti-canonical algebra S = R(−KR ) is finitely generated. We first
notice that we can extend p−e -linear maps on R to p−e -linear maps on S. Note that this
argument is substantially simpler than what the fourth author and K. Tucker did to obtain
similar results for finite maps in [ST14a].
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
13
Lemma 3.1. Suppose that R is an F -finite normal domain, D is a Weil divisor on Spec R
with associated algebra S := R(D). Then for any R-linear map φ : F∗e R −→ R we have an
induced S-linear map φS : F∗e S −→ S and a commutative diagram
F∗e R
φ
φS
/ R
_
_
F∗e S
/S
F∗e ρ
ρ
F∗e R
/R
φ
where ρ is the projection map onto degree zero S −→ R.
Proof. First note that we give F∗e S a Z[ p1e ]-graded structure so that our induced map φS
will be homogeneous. The idea is then simple, given an integer i ≥ 0, [F∗e S]i = F∗e R(ipe D).
We want to show that
φ(F∗e R(ipe D)) ⊆ R(iD).
(3.1.1)
But this is obvious since it holds in codimension 1 and all the sheaves are reflexive. Finally,
we simply have φS send [F∗e S]i/pe to zero if i is not divisible by pe . This completes the
proof.
In fact, it is not difficult to see that every homogeneous p−e -linear map on S comes from
R in this way.
Lemma 3.2. Suppose R is an F -finite normal domain, D a Weil divisor on Spec R, and
S = R(D), suppose we have a homogeneous map φS : F∗e S −→ S (again we give F∗e S the
Z[1/pe ]-grading). Then φS is induced from φS |R = φ : F∗e R −→ R as in Lemma 3.1.
Proof. Choose F∗e z ∈ [F∗e S]i = F∗e R(pe iD), invert an element u ∈ S0 = R to make D Cartier
e
and principal and then z = (f /um )y p where y generates R(iD)[u−1 ] and the element
f /um ∈ R[u−1 ] ⊆ S[u−1 ] is in degree zero in S[u−1 ]. We see that
e
ump z = f um(p
e −1)
e
y p ∈ F∗e R(pe iD).
e
e
e
e
Hence um φS (F∗e z) = φS (F∗e ump z) = φS (F∗e f um(p −1) y p ) = yφS (F∗e f um(p −1) ). The point
is that we can choose the same y regardless of the choice of z. Hence φS is completely
determined by φS |R .
Lemma 3.1 is key in the following proposition which lets us relate p−e -linear maps in
general on R and S.
Proposition 3.3. Suppose that R is an F -finite normal domain, D is a Weil divisor, and
the algebra S = R(D) is finitely generated (and in particular an F -finite Noetherian ring).
Further suppose that G is an effective Weil divisor on Spec R with pullback h∗ G = GS on
Spec S. Then we have a commutative diagram
HomS (F∗e S(GS ), S)
/ HomS (F e S, S)
∗
ES =eval@1
/S
γ
ν
HomR (F∗e R(G), R)
ρ
/ HomR (F e R, R)
∗
ER =eval@1
/ R.
14
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Here the map ρ is projection onto the 0th coordinate [S]0 = R and γ is the map which
restricts φ ∈ HomS (F∗e S, S) to F∗e R = [F∗e S]0 and then projects onto [S]0 = R. Furthermore
the maps γ and ν are surjective.
Proof. We first handle the commutativity. Given ψ ∈ HomS (F∗e S, S), we see that ρ(ES (ψ)) =
ρ(ψ(1)). On the other hand ER (γ(ψ)) = ρ(ψ(1)) as well. Hence we have commutativity of
the right square. The commutativity of the left square is obvious since GS is pulled back
from Spec R
To see that γ is surjective, for any φ ∈ HomR (F∗e R, R) construct φS as in Lemma 3.1.
Obviously γ(φS ) = φ. Similarly, Lemma 3.1 implies the surjectivity of the map ν.
As an immediate corollary we obtain a stabilization result similar to Hartshorne-SpeiserLyubeznik-Gabber.
Corollary 3.4. Suppose that R is an F -finite normal domain and that B ≥ 0 is a Weil
divisor. If the anti-log-canonical algebra R(−KR − B) is finitely generated, then the image
of the evaluation-at-1 map HomR (F∗e R((pe − 1)B), R) −→ R stabilizes for e ≫ 0.
Proof. Set S = R(−KR − B) and consider the diagram of Proposition 3.3. Since KS + h∗ B
is Cartier, we see that the images of
eval@1
ESe : HomS (F∗e S((pe − 1)h∗ B), S) −−−−→ S
stabilize, see for instance [FST11]. But then since ν, as in Proposition 3.3, surjects, we see
that the image of
(3.4.1)
ν
eval@1
HomS (F∗e S((pe − 1)B), R) −
→ F∗e HomR (F∗e R((pe − 1)B), R) −−−−→ R
Ee
R
R. But the image of (3.4.1) is the
coincides with that of HomR (F∗e R((pe − 1)B), R) −−→
same as the image of
Ee
ρ
S
HomS (F∗e S((pe − 1)h∗ B), S) −−→
S−
→ R.
However the ESe have stable image as we have already observed and the result follows.
Later in Theorem 3.9, we will obtain the same result for Q-divisors whose Weil-index is
not divisible by p. For now though, we move on to discreteness and rationality of F -jumping
numbers, generalizing [KSSZ14, Theorem 6.4] from the case of a graded ring R.
Theorem 3.5. Suppose that R is a normal domain and ∆ ≥ 0 is an effective Q-divisor
such that R(−KR − ∆) is finitely generated. Then for any ideal a ⊆ R the F -jumping
numbers of τ (R, ∆, at ) are rational and without limit points.
Proof. First let R ⊆ R′ be a separable extension of normal F -finite domains corresponding
ν
→ X = Spec R such that ν ∗ ∆ is an integral divisor (this
to a map of schemes Spec R′ = X ′ −
is easy, the idea is to simply take roots of generators of DVRs, if one has to take a pth root,
′
use Artin-Schreyer type equations see [BST15, Lemma 4.5]). Let Tr : K(X
→ K(X)
) −
′
∗
′
t
be the trace map and then recall that Tr ν∗ τ (X , ν ∆ − Ramν , (aR ) ) = τ (X, ∆, at )
by the main result of [ST14a]. It immediately follows that if the F -jumping numbers of
test ideal τ (X ′ , ν ∗ ∆ − Ramν , at ) are discrete and rational, so are the F -jumping numbers
of τ (X, ∆, at ). Additionally, by adding a Cartier divisor H to ∆, we can assume that
ν ∗ ∆ − Ramν is effective since τ (X, ∆ + H, at ) = τ (X, ∆, at ) ⊗ OX (−H) by Lemma 2.4(d).
Finally, note that
−KR′ − ν ∗ ∆ + Ramν = ν ∗ (−KR − ∆)
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
15
and so R(−KR′ − ν ∗ ∆ + Ramν ) is finitely generated by Lemma 2.9. The upshot of this
entire paragraph is of course that we may now without loss of generality assume that ∆ is
an integral effective divisor.
Next choose c ∈ R that is a test element for both ((R, ∆) and S := R(−KR − ∆). The
choice of such a c is easy, simply choose a test element so that additionally −KR − ∆ is
Cartier on X \ V (c). Away from V (c), S looks locally like R[t] which will certainly be
strongly F -regular over wherever R is strongly F -regular [HH94]. Let H be the Cartier
divisor corresponding to c and consider the commutative diagram.
(F∗e (aS)⌈t(p
e
−1)⌉ )
· HomS (F∗e S(h∗ (pe − 1)∆ + h∗ H), S)
/
e
−1)⌉ )
eval@1
/
S
ρ
γ
ν
(F∗e a⌈t(p
HomS (F∗e S, S)
· HomR (F∗e R((pe − 1)∆ + H), R)
/
HomR (F∗e R, R)
eval@1
/
R.
The sum over e > 0 of the images of the bottom rows is equal to τ (R, ∆, at ) and the
sum over e > 0 of the images of the top row is equal to τ (S, h∗ ∆, (aS)t ). Since ν surjects by Proposition 3.3, we immediately see that ρ(τ (S, h∗ ∆, (aS)t )) = τ (R, ∆, at ). But
now observe that KS + h∗ ∆ ∼ 0 by Lemma 2.10. But then the F -jumping numbers of
τ (S, h∗ ∆, (aS)t ) are discrete and rational by [ST14b]. The result follows.
We immediately obtain the following using the aforementioned breakthroughs in the
MMP.
Corollary 3.6. Suppose that R is strongly F -regular, of dimension 3, and of finite type
over an algebraically closed field of characteristic p > 5. Then the F -jumping numbers
τ (R, ∆, at ) are rational and without limit points for any choice of Q-divisor ∆ and ideal a.
Proof. Since R is strongly F -regular, there exists a divisor Γ ≥ 0 so that KR + Γ is QCartier and so that (R, Γ) is KLT by [SS10]. The result then follows from Theorem 2.27
and Theorem 3.5.
Of course, we also obtain discreteness and rationality of F -jumping numbers τ (R, ∆, at )
for any R, a 3-dimensional ring of finite type over an algebraically closed field k of characteristic p > 5 such that there exists a Γ ≥ 0 so that (R, Γ) is KLT.
3.1. A more general Hartshorne-Speiser type result. In Corollary 3.4, we used a
compatibility of the formation of Rees algebras to prove that the images of HomR (F∗e R, R) −
→
R stabilize for large e if S := R(−KR ) is finitely generated. In this short section, we generalize this result to the case of Q-divisors, at least whose Weil-index is not divisible by p. As
an alternate strategy, one could try to prove compatibilities analogous to Proposition 3.3
for Rees algebras of Q-divisors. Unfortunately this gets quite messy. Instead we take a
different approach utilizing Proj S. We first prove the result for Q-Gorenstein varieties and
then we handle the finitely generated case via the small map µ : X ′ −
→ X.
We do restrict ourselves to the case where the Weil-index of KX + ∆ is not divisible by
p. We realize that the methods we discuss here can apply to more general situations but
there are then several potential competing definitions for what the stable image should be.
Proposition 3.7. Suppose that (R, ∆ ≥ 0) is a pair such that KR +∆ is Q-Cartier. Suppose
that the Weil index of KR + ∆ is not divisible by p and that (pe − 1)(KR + ∆) is an integral
Weil divisor. Then
E ne =eval@1
σne (R, ∆) := Image HomR (F∗ne R((pne − 1)∆), R) −−−−−−−−→ R
stabilizes for large n.
16
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Proof. Fix m > 0 so that m(KR + ∆) is a Cartier divisor. The main idea is that module
HomR (F∗ne R((pne −1)∆), R) only takes on the values of finitely many sheaves, at least up to
twisting by line bundles (in particular, multiples of R(m(KR + ∆)). We also take advantage
of the fact that it is sufficient to show that the images stabilize partially up the chain.
Claim 3.8. Fix n0 > 0 and consider n ≥ n0 . Then
E ne =eval@1
HomR (F∗ne R((pne − 1)∆), R) −−−−−−−−→ R
factors through
E n0 e =eval@1
HomR (F∗n0 e R((pn0 e − 1)∆), R) −−−−−−−−→ R.
Hence it is sufficient to show that the images of HomR (F∗ne R((pne − 1)∆), R) in
HomR (F∗n0 e R((pn0 e − 1)∆), R)
stabilize.
Proof of claim. One simply notices that
p(n−n0 )e (pn0 e − 1)∆ ≤ (pne − 1)∆
ne
and hence (R((pne − 1)∆))1/p contains (R((pn0 e − 1)∆))1/p
ization occurs simply by restriction of scalars.
n0 e
. Thus the claimed factor
We continue on with the main proof. Note that pt (mod m) is eventually periodic. Then
choose a linear function θ(a) = ca+r, for c > 0, r ≥ 0, c, r ∈ Z, such that (pθ(a)e −1) (mod m)
is constant. Set
M = R(((1 − pre ) (mod m))(KR + ∆)) = R(((1 − pθ(0)e ) (mod m))(KR + ∆))
and note that for any a ≥ 0,
θ(a)e
∼
=
∼
=
∼
=
R((pθ(a)e − 1)∆), R)
HomR (F∗
θ(a)e
R((1 − pθ(a)e )(KR + ∆))
F∗
θ(a)e
θ(a)e
F∗
R(((1 − pθ(a)e ) (mod m))(KR + ∆)) ⊗ R(⌊ 1−pm ⌋m(KR + ∆))
θ(a)e
θ(a)e
M ⊗ R(⌊ 1−pm ⌋m(KR + ∆))
F∗
By inverting an element of R if necessary, we may assume that m(KR + ∆) ∼ 0. Thus
by utilizing this, we have maps
t
θ(a)e
a
. . . −→
F∗
ta−1
θ(a−1)e
M −−−→ F∗
t
θ(1)e
1
M−
→ ... −
→
F∗
t
θ(0)e
0
M−
→
F∗
M.
If these maps are Frobenius pushforwards of each other, i.e. F∗ce ta−1 = ta (or at least up
to a unit), then we can apply the standard Hartshorne-Speiser-Lyubeznik-Gabber theorem
θ(0)e
M . But this may be checked in
[Gab04] to conclude that the images stabilize in F∗
codimension 1 (since all sheaves are reflexive and so maps between them are determined in
codimension 1). However, after localizing to reduce to codimension 1, all the complicated
twisting we have done is irrelevant. Furthermore, in codimension 1, R is Gorenstein with
KR ∼ 0 and ∆ is Q-Cartier with index not divisible by p > 0 (since its Weil index was not
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
17
divisible by p > 0). Our chain of maps then just turns into
θ(a)e
HomR (F∗
/ HomR (F∗θ(a−1)e R((pθ(a−1)e − 1)∆), R) −
→ ...
O
R((pθ(a)e − 1)∆), R)
O
∼
∼
θ(a)e
F∗
ωR (−pθ(a)e (KR ) − (pθ(a)e − 1)∆)
O
Tr
/ F∗θ(a−1)e ωR (−pθ(a−1)e (KR ) − (pθ(a−1)e − 1)∆)
O
∼
(ca+r)e
F∗
∼
ωR (−(p(ca+r)e − 1)∆)
Tr
/ F∗(c(a−1)+r)e ωR (−(p(c(a−1)+r)e − 1)∆).
The bottom horizontal map is then obtained via
(ca+r)e
ωR (−(p(ca+r)e − 1)∆)
F∗
(ca+r)e
ωR (−pce (p(c(a−1)+r)e − 1)∆)
֒→ F∗
Tr
(c(a−1)+r)e
−→ F∗
ωR (−(p(c(a−1)+r)e − 1)∆).
Note the inclusion ֒→ can be identified with multiplication by a defining equation for
−pce (p(c(a−1)+r)e − 1)∆ + (p(ca+r)e − 1)∆ = (pce − 1)∆.
This is independent of a and so the maps in our chain are really the same, up to pushforward,
as claimed.
Note that this completes the proof. Even though we only proved stabilization of images
for a subset of ne > 0, these images are descending and our subset is infinite.
Now we are in a position to prove Corollary 3.4 in the more general situation.
Theorem 3.9. Suppose that R is an F -finite normal domain and that B ≥ 0 is a Q-divisor
with Weil-index not divisible by p. If the anti-log-canonical algebra R(−KR − B) is finitely
generated, then the image of the evaluation at 1 map HomR (F∗e R((pe − 1)B), R) −→ R
stabilizes for e sufficiently divisible.
Proof. For this proof, we will phrase our evaluation-at-1 maps in terms of the trace
F∗e R((1 − pe )(KR + B)) −
→ R.
We thus fix an e > 0 so that (pe − 1)(KR + B) is an integral Weil divisor.
Let µ : X ′ = Proj R(−KR − B) −
→ X = Spec R be as before. We observe that
∗
µ (−KR − B) is Q-Cartier and also still has Weil-index not divisible by p by Lemma 2.9.
Hence the images
(3.9.1)
... −
→
−
→
−
→
−
→
−
→
(n+1)e
OX ′ ((1 − p(n+1)e )(KX ′ + µ∗ B))
F∗
ne
F∗ OX ′ ((1 − pne )(KX ′ + µ∗ B))
...
F∗e OX ′ ((1 − pe )(KX ′ + µ∗ B))
OX ′
stabilize by Proposition 3.7. In fact, the same argument even shows that the images
even stabilize in any finite stage, such as in F∗ne OX ′ ((1 − pne )(KX ′ + µ∗ B)). However, the
terms and maps in this chain take on finitely many values up to twisting by (large) Cartier
multiples of −(KX ′ + µ∗ B) (as argued in Proposition 3.7). Our goal is to thus show that
these images stabilize after pushing forward by µ.
18
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Claim 3.10. If one applies µ∗ to (3.9.1), obtaining
. . . −→ µ∗ F∗ne OX ′ ((1−pne )(KX ′ +µ∗ B)) −→ . . . −→ µ∗ F∗e OX ′ ((1−pe )(KX ′ +µ∗ B)) −→ µ∗ OX ′
then the chain of images in OX = µ∗ OX ′ still stabilizes.
Proof of claim. Choose d > 0 so that
(n+d)e
→ F∗ne OX ′ ((1−pne )(KX ′ +µ∗ B))
OX ′ ((1−p(n+d)e )(KX ′ +µ∗ B)) −
(3.10.1) Image F∗
is equal to the stable image, which we denote by F∗ne σne , for all n ≥ 0 (note by Proposition 3.7
there are finitely many conditions). Observe that there are only finitely many σne up to
twisting by (large) multiples of −(KX ′ +µ∗ B). The fact that −(KX ′ +µ∗ B) is ample implies
that there exists an n0 ≥ 0 so that for any n ≥ n0 ,
F∗ne µ∗ σne
(n+d)e
−
→ F∗ne µ∗ OX ′ ((1 − pne )(KX ′ + µ∗ B))
µ∗ OX ′ ((1 − p(n+d)e )(KX ′ + µ∗ B))
= Image F∗
(n+d)e
−
→ F∗ne µ∗ σne
µ∗ OX ′ ((1 − p(n+d)e )(KX ′ + µ∗ B))
= Image F∗
(n+2d)e
→ F∗ne µ∗ OX ′ ((1 − pne )(KX ′ + µ∗ B)) .
µ∗ OX ′ ((1 − p(n+2d)e )(KX ′ + µ∗ B)) −
= Image F∗
(n+2d)e
But the map F∗
factors through
(n+2d)e
F∗
→ F∗ne µ∗ OX ′ ((1−pne )(KX ′ +µ∗ B))
µ∗ OX ′ ((1−p(n+2d)e )(KX ′ +µ∗ B)) −
(n+d)e
→ F∗
µ∗ OX ′ ((1 − p(n+2d)e )(KX ′ + µ∗ B)) −
µ∗ OX ′ ((1 − p(n+d)e )(KX ′ + µ∗ B))
which has image µ∗ σ(n+d)e by our assumption that (3.10.1) is the stable image, applied
(n+d)e
to the choice of n = n + d. It follows that F∗
(n+c)e
n and so by composition, F∗
µ∗ σ(n+d)e −
→ F∗ne µ∗ σne surjects for all
µ∗ σ(n+c)e −
→ F∗ne µ∗ σne surjects for every n ≥ n0 and
(n+c)e
every c ≥ d (note n does not depend on c). Thus since F∗
(n+d+c)e
OX ′ ((1
µ∗ F∗
−
p(n+d+c)e )(KX ′
+
µ∗ B)),
µ∗ σ(n+c)e is the image of
we see that
(n+c)e
→ F∗ne µ∗ OX ′ ((1 − pne )(KX ′ + µ∗ B))
µ∗ OX ′ ((1 − p(n+c)e )(KX ′ + µ∗ B)) −
Image F∗
(n+c)e
→ F∗ne OX ′ ((1 − pne )(KX ′ + µ∗ B))
OX ′ ((1 − p(n+c)e )(KX ′ + µ∗ B)) −
= µ∗ Image F∗
= µ∗ σne
for all n ≥ n0 and all c ≥ 2d. This clearly proves the desired stabilization.
We return to the proof of theorem. But this is trivial once we observe that
F∗ne µ∗ OX ′ ((1 − pne )(KX ′ + µ∗ B)) = F∗ne OX ((1 − pne )(KX + B))
since µ is small. Hence the proof is complete.
4. Stabilization and discreteness via positivity
In the previous section we showed the discreteness and rationality of F -jumping numbers
via passing to the local section algebra (i.e. a symbolic Rees algebra) where we already
knew discreteness and rationality. In this section, we recover the same discreteness result in
the projective setting by using the methods of [CU15] which allow us to apply asymptotic
vanishing theorems to Weil divisors. Indeed we first prove global generation results for test
ideals by employing similar methods to [Mus13].
Setting 4.1. Let X be a normal projective variety of characteristic p > 0, ∆ ≥ 0 an
effective Weil Q-divisor, a an ideal sheaf on X and t ∈ Q. We make no assumptions about
KX + ∆ being Q-Cartier.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
19
Assume G is a line bundle such that there are global sections x1 , . . . , xm ∈ H 0 (X, a ⊗ G )
which globally generate a ⊗ G and then let Symc (x1 , . . . , xm ) denote the cth symmetric
power of the vector space hx1 , . . . , xm i. Observe that Symc (x1 , . . . , xl ) ⊆ H 0 (X, ac ⊗ G c )
globally generates ac ⊗ G c . Thus we have a surjection of sheaves
Symc (x1 , . . . , xt ) ⊗k G −c −
→ ac .
Lemma 4.2. If the Weil-index of ∆ is not divisible by p, and t = a/b, with p not dividing
b, then there is a Cartier divisor H and a finite set of integers e1 , . . . , es ≫ 0 such that
(pei − 1)∆ is integral, (pei − 1)t ∈ Z and τ (X, ∆, at ) equals
s
X
Image
i=1
e
F∗ i Symt(p
ei
−1)
(x1 , . . . , xm ) ⊗k G −t(p
F∗ei at(p
ei
−1)
ei
−1)
⊗ OX ((1 − pei )(KX + ∆) − H)
/
❢❢❢3
❢❢❢❢❢
❢
❢
❢
❢
❢
❢❢❢❢❢
❢❢❢❢❢ Trei
❢❢❢❢❢
OX
· OX ((1 − pei )(KX + ∆) − H)
.
At some level this result is obvious. The only technicalities involve showing that the
various rounding choices we make all give the same result in the end (since we can absorb
any differences into the test element – a local generator of H). We include a complete
proof but we invite the reader to skip over it if they are already familiar with this type of
argument.
Proof. The statement in the end is local and so trivializing G it suffices to show that
τ (X, ∆, at ) =
s
X
Trei F∗ei at(p
i=1
ei −1)
· OX ((1 − pei )(KX + ∆) − H) .
Pick an effective Cartier divisor H0 corresponding to the vanishing locus of a test element
so that for any integer e0 > 0
X
e
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H⌉)
τ (X, ∆, at ) =
e≥e0
for all Cartier H ≥ H0 , c.f., [ST14b, Prop. 3.6], [BSTZ10, Def.-Prop. 3.3]. This equality also
holds for any Q-divisor H ≥ H0 as one can always pick a Cartier H ′ so that H ′ ≥ H ≥ H0
and one obtains inclusions
X
e
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H0 ⌉)
τ (X, ∆, at ) =
e≥e0
⊃
X
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H⌉)
X
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H ′ ⌉) = τ (X, ∆, at ).
e
e≥e0
⊃
e
e≥e0
Next consider the claim which will allow us restrict to those e which are multiples of e0 .
Claim 4.3. For any Weil divisor H ≥ H0 , there exists a Cartier divisor G such that for
any integer 0 ≤ b ≤ e0 − 1 and for any integer m > 0 we have
e m+b
⌉ O (⌈K − pe0 m+b (K + ∆) − H − G⌉)
Tre0 m+b F∗e0 m+b a⌈tp 0
X
X
X
e
m
⊆ Tre0 m F∗e0 m a⌈tp 0 ⌉ OX (⌈KX − pe0 m (KX + ∆) − H⌉⌉).
20
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Proof. To prove the claim, first note that by [Sch09, Lemma 4.6] (among many other places),
b
e
if d ∈ alp 0 ⊆ al(p −1) then
da⌈tp
e0 m+b ⌉
⊆ a⌈tp
e0 m ⌉+l(pb −1)
⊆ (a⌈tp
e0 m ⌉
b
)[p ] .
where again the l is an upper bound for the number of generators of a (note that d works
for any b ≤ e0 − 1). Set then G = div(d) + pe0 H and notice that
e m+b
⊆
⊆
⊆
⊆
⌉ O (⌈K − pe0 m+b (K + ∆) − H − G⌉)
Trb F∗b a⌈tp 0
X
X
X
e0 m+b ⌉
b b
⌈tp
Tr F∗ da
OX (⌈KX − pe0 m+b (KX + ∆)⌉) ⊗ OX (−pb H)
e m
b
Trb F∗b (a⌈tp 0 ⌉ )[p ] OX (⌈KX − pe0 m+b (KX + ∆)⌉) ⊗ OX (−pb H)
e m
a⌈tp 0 ⌉ OX (⌈KX − pe0 m (KX + ∆)⌉) ⊗ OX (−H)
e m
a⌈tp 0 ⌉ OX (⌈KX − pe0 m (KX + ∆) − H⌉)
Now applying Tre0 m F∗e0 m proves the claim.
Now we return to the proof of the lemma. The claim and our previous work implies that
for a sufficiently large Cartier divisor H ≥ H0 and G depending on H, we have that
P
e
τ (X, ∆, at ) =
Tre F e a⌈tp ⌉ O (⌈KX − pe (KX + ∆) − H − G⌉)
Pe≥e0 e e∗ ⌈tpe ⌉ X
OX (⌈KX − pe (KX + ∆) − H⌉)
⊆
Tr F∗ a
Pe0 |e
e
e e ⌈tp ⌉
OX (⌈KX − pe (KX + ∆) − H⌉)
⊆
e≥e0 Tr F∗ a
t
= τ (X, ∆, a )
P
e
and therefore that τ (X, ∆, at ) = e0 |e Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H⌉).
Pick a Cartier divisor H ′ so that H ′ ≥ H0 + ∆ + ⌈t⌉div(d) ≥ H0 where d ∈ a. We have
that
X
e
τ (X, ∆, at ) =
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H ′ ⌉)
e0 |e
⊆
X
Tre F∗e a⌈t(p
e −1)⌉
OX (⌈KX − pe (KX + ∆) + ∆ − H ′ ⌉)
=
X
Tre F∗e a⌈t(p
e −1)⌉
OX (⌈(1 − pe )(KX + ∆) − H ′ ⌉)
⊆
X
Tre F∗e a⌈t(p
e −1)⌉
OX (⌈(1 − pe )(KX + ∆) − H0 − ∆ − ⌈t⌉div(d)⌉)
⊆
X
Tre F∗e a⌈t(p
e −1)⌉+⌈t⌉
⊆
X
Tre F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − H0 ⌉)
e0 |e
e0 |e
e0 |e
OX (⌈KX − pe (KX + ∆) − H0 ⌉)
e0 |e
e
e0 |e
⊆ τ (X, ∆, at ).
In particular
τ (X, ∆, at ) =
X
Tre F∗e a⌈t(p
e −1)⌉
OX (⌈(1 − pe )(KX + ∆) − H ′ ⌉).
e0 |e
Since the Weil-index of ∆ is not divisible by p, (1 − pe )(Kx + ∆ + tD) is an integral divisor
for e sufficiently divisible. Hence by choosing our e0 sufficiently divisible and noting that
our scheme is Noetherian and so the above sum is finite, we obtain our desired result.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
21
Remark 4.4. While it is certainly possible to generalize this to handle t ∈ R or to handle ∆
such that (pe − 1)(KX + ∆) is not integral, those generalizations are not the ones we need.
In particular we will need a power of KX + ∆ times a locally free sheaf.
Theorem 4.5. Suppose X is normal and projective, R(X, −KX − ∆) is finitely generated
and ∆ has Weil-index not divisible by p and fix t > 0. There exists a Cartier divisor3 L
such that
τ (X, ∆, aw ) ⊗ OX (L)
is globally generated when 0 ≤ w = a/b ≤ t, with p not dividing b.
Proof. Choose a line bundle G = OX (G) such that a ⊗ G is globally generated by sections
x1 , . . . , xm ∈ H 0 (X, a ⊗ G ). By Lemma 4.2 there is a Cartier divisor H and integers
e1 , . . . , es ≫ 0 such that the test ideal τ (X, ∆, aw ) is equal to
s
X
e
T ei
Image F∗ei Symc (x1 , . . . , xm ) ⊗k G −w(p i −1) ⊗ OX ((1 − pei )(KX + ∆) − H) −−→ OX
i=1
which is globally generated if each summand is. Fix now A a globally generated ample
Cartier divisor. We claim it suffices to find a Cartier divisor L1 , such that
F∗ei OX ((1 − pei )(KX + ∆ + wG) − H) ⊗ OX (L1 + (d + 1)A)
= F∗ei OX ((pei − 1)(L1 − ∆ − KX − wG) + L1 − H + pei (d + 1)A)
is globally generated (the equality in the displayed equation follows from the projection
formula). Indeed, assuming this global generation choose L = L1 + (d+ 1)A with d = dim X
and note that the image of a globally generated sheaf is still globally generated. We will
find a single L1 that works for all 0 ≤ w ≤ t.
Since R(X, −∆ − KX ) is finitely generated, we can use Lemma 2.11 to find a Cartier
divisor M so that (M − ∆ − KX ) is an ample Weil divisor. Moreover, we can find an ample
Cartier divisor N such that N − tG is ample. Notice that, for all 0 ≤ w ≤ t,
w
w
N + (N − tG) is ample.
N − wG = 1 −
t
t
This observation is what lets us replace tG with wG. Set L1 := M + N . By [CU15, Lem.
2.18(i)], L1 − KX − ∆ − wG = (M − KX − ∆) + (N − wG) is an ample Weil divisor.
Fix 0 ≤ w ≤ t. We now show that the Castelnuovo-Mumford regularity, with respect to
A, of
F∗e OX ((pe − 1)(L1 − ∆ − KX − wG) + L1 − H + pe (d + 1)A)
is zero for each e = ei ≫ 0, which guarantees by Mumford’s theorem [Laz04, Thm. 1.8.3]
the desired global generation. It suffices now to show that
(4.5.1) H i (X, OX (−iA)⊗OX F∗e OX ((pe −1)(L1 −∆−KX −wG)+L1 −H +pe(d+1)A)) = 0
which by the projection formula and the fact that F∗e doesn’t change the underlying sheaves
of Abelian groups, is the same as showing
H i (X, OX ((pe − 1)(L1 − ∆ − KX − wG) + L1 − H + pe (d + 1 − i)A)) = 0
for 0 < i ≤ d and d = dim X. Since we may assume that e ≫ 0, L1 − H + pe (d + 1 − i)A
is nef. Therefore because L1 − ∆ − KX − wG is ample Weil and L − H + pe (d + 1 − i)A is
nef and Cartier, we may apply the version of Fujita vanishing [CU15, Thm. 4.1] to obtain
the vanishing desired in (4.5.1). This completes the proof.
3See Remark 4.6 for a discussion of how to choose L effectively.
22
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Remark 4.6. Indeed, it is not hard to choose L effectively. Summarizing the proof above,
fix an ample Cartier G so that a ⊗ OX (G) is globally generated, fix A to be a globally
generated ample Cartier divisor, and fix M Cartier so that M − ∆ − KX is ample and
choose an ample Cartier N so that N − tG is ample. Then we can take
L = (d + a)A + M + N.
We now turn to the promised results on discreteness and rationality.
Proposition 4.7. Suppose now that X is normal and projective and R(X, −KX − ∆) is
finitely generated. Then for any ideal sheaf a on X, the jumping numbers of τ (X, ∆, at ) are
without limit points.
Proof. First assume that ∆ has Weil-index not divisible by p. It follows from an appropriately generalized version of the argument of [BSTZ10, Lemma 3.23] that τ (X, ∆, at ) =
τ (X, ∆, at+ε ) for all 0 < ε ≪ 1. Hence for every real number t ∈ R≥0 , there is a rational number w = a/b with p not dividing p with τ (X, ∆, at ) = τ (X, ∆, aw ). Now
fix t0 > 0. It follows from Theorem 4.5 that there exists a Cartier divisor L such that
τ (X, ∆, aw ) ⊗ OX (L) is globally generated for every w < t0 with w = a/b and where p does
not divide b. But then by our previous discussion, we also see that τ (X, ∆, at ) ⊗ OX (L)
is globally generated for every t < t0 . The discreteness follows since now for 0 ≤ t < t0 ,
H 0 (X, τ (X, ∆, at ) ⊗ OX (L)) ⊆ H 0 (X, OX (L)) form a decreasing chain of subspaces of a
finite dimensional vector space H 0 (X, OX (L)), and of course by the global generation hypothesis, if H 0 (X, τ (X, ∆, at2 )⊗OX (L)) = H 0 (X, τ (X, ∆, at1 )⊗OX (L)) then τ (X, ∆, at2 ) =
τ (X, ∆, at1 ). This proves the result when ∆ has Weil-index not divisible by p.
→ OX (KX )
Next assume that pd ∆ has Weil-index divisible by p. Fix a map F∗d OX (KX ) −
inducing a map on the fraction fields T : F∗d K(X) −
→ K(X). As in [ST14a, Theorem 6.25],
this map induces a possibly non-effective Weil divisor RT ∼ (1 − pd )KX with
d
(4.7.1)
T τ (X, pd ∆ − RT , (a[p ] )t ) = τ (X, ∆, at )
Choose a Cartier divisor G so that pd ∆ − RT + pd G is effective and notice that it also has
Weil-index not divisible by p. Next observe that
−KX − pd ∆ + RT − pd G
∼ −KX − pd ∆ + (1 − pd )KX − pd G
= −pd (KX + ∆) − pd G
= F d∗ (−KX − ∆ − G)
and hence that R(X, −KX − pd ∆ + RT − pd G) is finitely generated (note that the −G is
Cartier and thus harmless so we are really taking the pd th Veronese of R(X, −KX − ∆).
d
Hence by what we have already shown, the F -jumping numbers of τ (X, pd ∆ − RT , (a[p ] )t )
have no limit points. Therefore by applying T via (4.7.1), we see that the F -jumping
numbers of τ (X, ∆ + G, at ) also have no limit points. But then by [BSTZ10, 3.26] the
F -jumping numbers of τ (X, ∆, at ) have no limit points proving the theorem.
4.1. Global generation and stabilization of σ. We now give another proof of Corollary 3.4
in the projective setting.
Theorem 4.8. Suppose that (X, ∆) is a projective pair such that R(X, −KX −∆) is finitely
generated and KX + ∆ has Weil-index not divisible by p. Then the images
Image F∗e OX ((1 − pe )(KX + ∆)) −→ OX
stabilize for e sufficiently large and divisible. We use σ(X, ∆) to denote this stable image.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
23
Proof. Choose a globally generated ample Cartier divisor A and a Cartier divisor L such
that L − KX − ∆ is an ample Weil divisor by 2.11. For each e such that (pe − 1)(KX + ∆)
is integral, set
σe (X, ∆) = Image F∗e OX ((1 − pe )(KX + ∆)) −
→ OX .
Then fixing d = dim X
σe (X, ∆) ⊗ OX (dA + L) = Image (F∗e OX (pe dA + L + (pe − 1)(L − KX − ∆)) −
→ OX ) .
We immediately notice that F∗e OX (pe dA + L + (pe − 1)(L − KX − ∆)) is 0-regular with
respect to A and hence its image σe (X, ∆) ⊗ OX (dA + L) is globally generated. As the
global generating sections all lie in H 0 (X, OX (dA + L)) which is finite dimensional, and
as the σe form a descending chain of ideals as e increases, we see that σe stabilizes for e
sufficiently large and divisible as claimed.
As an immediate corollary of the proof, we obtain:
Corollary 4.9. Suppose again that (X, ∆) is a projective pair of dimension d such that
R(X, −KX − ∆) is finitely generated and that KX + ∆ has Weil-index not divisible by p. If
L is a Cartier divisor such that L − KX − ∆ is an ample Weil divisor and if A is a globally
generated ample Cartier divisor, then σ(X, ∆) ⊗ OX (dA + L) is globally generated.
5. Alterations
In this section we give a description of the test ideal τ (X, ∆, at ) under the assumption that
R(−KX − ∆) is finitely generated. This generalizes [BST15] from the case that −KX − ∆ is
Q-Cartier. As a consequence we obtain a generalization of a result of A. K. Singh’s [Sin14]
4
, also compare with [Sin99].
Before starting in on this, let us fix notation and recall the following from Section 2.
Setting 5.1. Suppose that ∆ ≥ 0 is a Q-divisor on an F -finite normal scheme X, R(−KX −
µ
∆) is finitely generated with X ′ = Proj R(−KX − ∆) −
→ X. Suppose that a is an ideal
sheaf on X and t ≥ 0 is a real number.
We have already seen that we can pullback −KX − ∆ to X ′ by µ where it becomes a
Q-Cartier divisor, see Lemma 2.10. Suppose further that π : Y −
→ X is any alteration that
factors through X ′ as
Y
ψ
/ X′
µ
/ X.
?
π
Then we define π ∗ (KX + ∆) = ψ ∗ µ∗ (KX + ∆), or equivalently we define π ∗ is as in [DH09]
(even though π is not birational), see Section 2.4. Recall, of course, that if π : Y −
→ X
is a small alteration (meaning that the non-finite-to-1 locus of π has codimension ≥ 2 in
Y ), then this coincides with the obvious pullback operation. More generally, if π : Y −
→X
ξ
π
µ
→ X, and ξ is
→ Y −
→ X factors through both π and X ′ −
is any alteration and ρ : Y ′ −
∗
birational, then we define π (KX + ∆) to be ξ∗ ρ∗ (−KX − ∆).
In the next lemma and later in the section, we use the notion and notation of parameter
test modules τ (ωX , KX + ∆, at ) := τ (X, ∆, at ). For a concise introduction and more about
their relation to test ideals, please see [ST14b, Section 4].
4This result was announced years ago, but has not been distributed.
24
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
Lemma 5.2. Working in Setting 5.1, if m ∈ Z>0 is such that tm ∈ Z, that m∆ is integral,
and such that the m-th Veronese of the symbolic Rees algebra R(X, −KX − ∆) is generated
in degree 1, then
1
1
τ (X, ∆, at ) = τ (ωX , (OX (−m(KX + ∆))atm ) m ) = τ (X, −KX , (OX (−m(KX + ∆))atm ) m ).
Proof. We know by Lemma 2.3 that for any sufficiently large Cartier D ≥ 0 and any e0 ≥ 0
that
X
e
TreX F∗e a⌈tp ⌉ OX (⌈KX − pe (KX + ∆) − D⌉) .
τ (X, ∆, at ) =
e≥e0
Choose 0 <
D′
≤
D ′′
divisors such that D ′′ is Cartier and KX − D ′ is Cartier. Since
τ (X, ∆, at )
P
e
e ⌈tpe ⌉ · O (⌈K − pe (K + ∆) − D − D ′′ ⌉)
=
X
X
X
e≥e0 TrX F∗ a
P
e
e ⌈tpe ⌉ · O (⌈K − D ′ − pe (K + ∆) − D⌉)
⊆
X
X
X
e≥e0 TrX F∗ a
P
e
e a⌈tpe ⌉ · O (K − D ′ ) · O (⌈−pe (K + ∆) − D⌉)
=
Tr
F
X
X
X
X
X
e≥e0
∗
P
e⌉
e
⌈tp
e
e
a
·
O
(K
)
·
O
(⌈−p
(K
+
∆)
−
D⌉)
F
⊆
Tr
X
X
X
X
X
e≥e0
∗
P
e⌉
e
e
⌈tp
e
F
a
·
O
(⌈K
−
p
(K
+
∆)
−
D⌉)
⊆
Tr
X
X
X
∗
X
e≥e0
= τ (X, ∆, at )
P
e
we see that τ (X, ∆, at ) = e≥e0 TreX F∗e a⌈tp ⌉ · OX (KX ) · OX (⌈−pe (KX + ∆) − D⌉) .
This is already very close.
Claim 5.3. We can choose a Cartier D3 > 0 so that
⌈ 1 ·pe ⌉
e
OX (−D3 ) · a⌈tp ⌉ · OX (⌈−pe (KX + ∆)⌉) ⊆ (atm ) · OX (−m(KX + ∆)) m
for all e ≥ 0.
e
1
e
Proof of claim. Checking this assertion is easy, we can certainly knock a⌈tp ⌉ into (atm )⌈ m p ⌉
by multiplication by a Cartier divisor. Handling the other multiplication is a little tricke
ier. Likewise certainly we can multiply OX (⌈−pe (KX + ∆)⌉) into OX (−⌈ pm ⌉m(KX + ∆)).
But then notice that OX (−am(KX + ∆)) = OX (−m(KX + ∆))a by our finite generation
hypothesis. This proves the claim.
Returning to the proof, we see that
τ (X, ∆, at )
P
e
e a⌈tpe ⌉ · O (K ) · O (⌈−pe (K + ∆) − D − D ⌉)
F
=
Tr
X
X
X
X
3
X
e≥e0
∗
P
1 e
1 e
e
p ⌉
⌈m
p ⌉
e (O (K ) · (atm )⌈ m
F
⊆
Tr
·
O
(−m(K
+
∆))
·
O
(−D))
X
X
X
X
X
X
e≥e0
∗
P
1 e
1 e
p ⌉
⌈m
p ⌉
e
e (O (K − pe (K − K )) · (atm )⌈ m
·
O
(−m(K
+
∆))
·
O
(−D))
F
=
Tr
X
X
X
X
X
X
X
∗
X
e≥e0
1
= τ (X, −KX , (OX (−m(KX + ∆))atm ) m )
tm 1
= τ (ωX , (OX (−m(K
X + ∆))a ) m )
P
1 e
p ⌉
e
e (K + ∆) − D⌉)
e (atm )⌈ m
·
O
(⌈K
−
p
F
⊆
Tr
X
X
X
∗
X
e≥e0
= τ (X, ∆, at )
which proves the Lemma.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
25
Remark 5.4. It is tempting try to use Lemma 5.2 to give another proof of discreteness and
rationality of F -jumping numbers by appealing to [ST14b]. However, this doesn’t seem to
work. In particular in [ST14b] the authors did not prove discreteness and rationality of F jumping numbers for τ (X, ∆, bs at ) (as t-varies) – mixed test ideals were not handled. One
could probably easily recover discreteness of F -jumping numbers via the usual arguments
of gauge boundedness for Cartier algebras [Bli13] at least in the case when X is finite type
over a field. For additional reading on mixed test ideals (and their pathologies) we invite
the reader to look at [Pér13].
The really convenient thing about Lemma 5.2 for our purposes is the following.
Lemma 5.5. Using the notation of Lemma 5.2, suppose that π : Y −→ X is an alteration
from a normal Y where if we write b = OX (−m(KX + ∆)) then b · OY = OY (−TY ) is an
invertible sheaf. Then TY = mπ ∗ (KX + ∆) where π ∗ (KX + ∆) is defined as in the text
below Setting 5.1.
Proof. This is easy. Indeed, we already know that π factors through the normalized blowup
of b by the universal property of blowups. On the other hand OY (−mπ ∗ (KX + ∆)) =
b · OY .
As a result, we immediately obtain the following.
Theorem 5.6. Suppose that X is a normal F -finite integral scheme and that ∆ on X is an
effective Q-divisor such that S = R(−KX − ∆) is finitely generated. Suppose also that a is
an ideal sheaf and t ≥ 0 is a rational number. Then there exists an alteration π : Y −→ X
from a normal Y , factoring through X ′ = Proj S and with G = divY (a), so that
(5.6.1)
τ (X, ∆, at ) = Image π∗ OY (⌈KY − π ∗ (KX + ∆) − tG⌉) −→ OX .
This π may be taken independently of t ≥ 0 if desired. If a is locally principal (for instance
if a = OX ), then one may take π to be a small alteration if desired. Alternately, if X is
essentially of finite type over a perfect field, then one may take Y regular by [dJ96].
As a consequence we obtain that
\
τ (X, ∆, at ) =
Image π∗ OY (⌈KY − π ∗ (KX + ∆ + tG)⌉) −→ OX
π:Y −
→X
where π runs over all alterations with a · OY = OY (−G) is invertible (or all such regular
alterations if X is of finite type over a perfect field).
Proof. Most of the result follows immediately from [ST14b, Theorem A] combined with
Lemma 5.2 and Lemma 5.5. Indeed, simply choose m such that tm is an integer and the
condition of Lemma 5.2 is satisfied, then apply [ST14b, Theorem A] to find an alteration
1
tm m
such that the image of the above map is τ (X, −K
TX , (OX (−m(KX + ∆))a ) ). Consider
the alterations π that occur in the intersection π:Y −
→X . Following [ST14b, Theorem A]
might seem to require that we only consider π that factor through the normalized blowup
of a · OX (−m(KX + ∆)) (for m-sufficiently divisible). However, it is easy to see that other
Y ’s can be dominated by those that factor through this blowup and the further blowups
certainly have smaller images.
One also must handle the case of varying t which is not quite done in [ST14b, Theorem A]
in our generality (there the authors treated τ (X, ∆, at ), while here we need τ (X, −KX , bs at )).
However, the argument there essentially goes through verbatim (alternately, this is the same
argument as in [STZ12]).
The only remaining part of the statement that doesn’t follow immediately is the assertion
in the case when a is locally principal. However, in the proof of [ST14b, Theorem A], the
26
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
alteration needed can always be taken to be a finite cover of the normalized blowup of the
ideal (in this case, the normalized blowup of OX (−m(KX + ∆))atm which coincides with
the normalized blowup of OX (−m(KX + ∆))). This normalized blowup is of course X ′ in
our setting.
In the above proof, our constructed Y was definitely not finite over X. This is different
from [BST15] where the simplest constructed Y definitely was finite over X. Fortunately,
we can reduce to the case of a finite Y at least when a = OX .
Corollary 5.7. Suppose that X is a normal F -finite integral scheme and that ∆ on X is
an effective Q-divisor such that S = R(−KX − ∆) is finitely generated. Then there exists
a finite map φ : Y −→ X from a normal Y , factoring through X ′ = Proj S such that
τ (X, ∆) = Image φ∗ OY (⌈KY − φ∗ (KX + ∆)⌉) −→ OX .
Proof. Let π : Y ′ −
→ X be a small alteration satisfying (5.6.1) from Theorem 5.6. Additionα
φ
→Y −
→X
ally assume that π ∗ (KX + ∆) is integral (for simplicity of notation). Next let Y ′ −
be the Stein factorization of π. Since α : Y ′ −
→ Y is small, we see that
π∗ OY ′ (KY ′ − π ∗ (KX + ∆)) = φ∗ OY (KY − φ∗ (KX + ∆))
and the result follows.
Question 5.8. Can one limit oneself to separable alterations in Theorem 5.6? In particular,
is there always a separable alteration π : Y −
→ X with
τ (X, ∆) = Image π∗ OY (⌈KY − π ∗ (KX + ∆)⌉) −
→ OX ?
The analogous result is known if KX + ∆ is Q-Cartier by [BST15]. However in our proof,
π is definitely not separable because we rely on [ST14b, Theorem A], which uses Frobenius
to induce certain vanishing results. It is possible that this could be replaced by cohomology
killing arguments as in for instance [BST15, Bha12, HL07, HH92].
As a special case, we recover a result of Anurag K. Singh (that was announced years ago)
[Sin14].
Corollary 5.9 (Singh). Suppose that X is an F -finite splinter and R(−KX ) is finitely
generated. Then X is strongly F -regular.
Proof. Indeed, if X is a splinter then for any finite morphism φ : Y −
→ X, the evaluation-at1 map H omOX (φ∗ OY , OX ) −
→ OX surjects. However, H omOX (φ∗ OY , OX ) ∼
= φ∗ OY (KY −
φ∗ KX ) and the trace map to OX is identified with the evaluation-at-1 map. Hence using
Corollary 5.7 we see that τ (X) = τ (X, 0) = OX . Since for us τ (X) always denotes the big
test ideal, this proves that X is strongly F -regular.
It would be natural to try to use the above to show that splinters are strongly F -regular
for 3-dimensional varieties of characteristic p > 5, using the fact that such KLT varieties
satisfy finite generation of their anticanonical rings Theorem 2.27. The gap is the following:
Question 5.10. Suppose that R is a normal F -finite domain that is also a splinter. Does
there exist a Q-divisor ∆ ≥ 0 on Spec R such that KX +∆ is Q-Cartier and that (Spec R, ∆)
is KLT?
The analogous result on the existence of ∆ for strongly F -regular varieties was shown in
[SS10]. Of course, the fact that splinters are in fact derived splinters in characteristic p > 0
[Bha12] would likely be useful.
In particular, we do obtain the following.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
27
Corollary 5.11. Suppose that R is an F -finite three dimensional splinter which is also
KLT (for an appropriate ∆ ≥ 0) and that R is finite type over an algebraically closed field
of characteristic p > 5. Then R is strongly F -regular.
6. Reduction from characteristic zero
The goal of this section is to show that multiplier ideals J (X, ∆, at ) reduce to test
ideals τ (Xp , ∆p , atp ) after reduction to characteristic p ≫ 0 at least if R(−KX − ∆) is
finitely generated. We begin with some preliminaries on the reduction process.
Let X be a scheme of finite type over an algebraically closed field k of characteristic zero,
∆ a Q-divisor, and a ⊂ OX an ideal sheaf. One may choose a subring A ⊂ k which is
finitely generated over Z over which X, ∆, and a are all defined. Denote by XA , ∆A , and
aA ⊂ OXA the models of X, ∆, and a over A. For any closed point s ∈ Spec A we denote
the corresponding reductions Xs , ∆s , and as ⊂ OXs defined over the residue field k(s)
which is necessarily finite. In the simple case where A = Z, if XA = Spec Z[x1 , . . . , xn ]/I
for I = (f1 , . . . , fm ) and p ∈ Z is prime, the scheme Xp = Spec Fp [x1 , . . . , xn ]/(f1 mod
p, . . . , fm mod p).
Warning 6.1. In what follows, we abuse terminology in the following way. By p ≫ 0, we
actually mean the set of closed points of an open dense set U ⊆ Spec A. Furthermore, if we
start with X as above, by Xp for p ≫ 0 we actually mean some Xs for some closed point s
in the aforementioned U . This is a common abuse of notation and we do not expect it will
cause any confusion. It does substantially shorten statements of theorems.
Lemma 6.2. Suppose X is a normal quasi-projective variety over an algebraically closed
field k of characteristic zero. For any Q-divisor Γ so that R(X, Γ) is finitely generated we
have R(X, Γ)p ∼
= R(Xp , Γp ) for p ≫ 0.
In particular, if ∆ is a Q-divisor and R(X, −KX − ∆) is finitely generated, setting
X ′ = Proj R(X, −KX − ∆) and µ : X ′ −→ X we have
R(X, −KX − ∆)p = R(Xp , (−KX − ∆)p ) = R(Xp , −KXp − ∆p )
and so this ring is also finitely generated. This means (X ′ )p = (Xp )′ . We denote both by
Xp′ .
Proof. Note that R(X, Γ)p makes sense for p ≫ 0 as R(X, Γ) is finitely generated and
R(X, Γ)p is naturally finitely generated by the reduction of the generators of R(X, Γ). The
problem is that potentially
L the algebra R(X, Γ)p may not be the symbolic Rees algebra
(local section algebra) n≥0 OXp (nΓp ). Throughout this proof we will constantly need to
choose p ≫ 0 (or technically, restrict to a smaller open subset U of Spec A). First we record
a claim that is certainly well known to experts.
Claim 6.3. For any Weil divisor D and prime p ≫ 0, potentially depending on D, we
claim that OX (D)p ∼
= OXp (Dp ) as sheaves of OXp -modules.
Proof of claim. To see this we prove that OX (D)p is reflexive and agrees outside a codimension 2 subset with OXp (Dp ). Of course, since OX (D) is reflexive,
OX (D) −
→ H omX (H omX (OX (D), OX ), OX )
is an isomorphism. But this isomorphism is certainly preserved via reduction to characteristic p so (OX (D))p is reflexive at least for p ≫ 0. Choose a closed set Z ⊆ X of codimension
2, defined with no additional coefficients (other than the ones already needed to define D
and X) so that D|X\Z is Cartier. Note that Dp |Xp \Zp is Cartier and so OXp \Zp (Dp ) is
locally free and agrees with (OX\Z (D))p , the claim follows.
28
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
We return to the proof of the lemma. Next define X ′ to be the blowup of OX (mΓ) for some
m ≫ 0 sufficiently divisible. Then since µ : X ′ −
→ X is small, so is µp : Xp′ −
→ Xp , and note
′
that Xp is still the blowup of (OX (mΓ))p = OXp (mΓp ) by the claim. Since mΓ′ = mπ∗−1 Γ
was Cartier in characteristic zero, mΓ′p is Cartier after reduction to characteristic p ≫ 0
as well. Now Xp′ −
→ Xp is still small and notice that Γ′p is relatively ample (since Xp′ was
obtained by blowing up OXp (mΓp )). Hence
M
M
(µp )∗ OXp′ (nΓ′p ) =
OXp (nΓp )
n≥0
n≥0
is finitely generated and has Proj equal to Xp′ . The lemma follows immediately.
Armed with this lemma, the proof of the main theorem for this section is easy.
Theorem 6.4. Suppose that X is a normal quasi-projective variety over an algebraically
closed field of characteristic zero. Further suppose that ∆ ≥ 0 is a Q-divisor such that
R(−KX − ∆) is finitely generated and also suppose that a ⊆ OX is an ideal and t ≥ 0 is a
rational number. Then
J (X, ∆, at )p = τ (Xp , ∆p , atp )
for p ≫ 0.
1 ♮
π m(KX +∆)−tG⌉) for some sufficiently
Proof. We know that J (X, ∆, at ) = π∗ OXe (⌈KXe − m
divisible m and sufficiently large log resolution of singularities, by Definition 2.19 (here we
need that OX (−m(KX + ∆)) · OXe = OXe (−A) and a · OXe = OXe (−B) are invertible). We
rewrite this multiplier ideal as
1
1
π∗ OXe (⌈KXe − π ♮ m(KX + ∆) − tG⌉) = π∗ OXe (⌈KXe − π ♮ (KX − KX ) − A − tB⌉)
m
m
1
and observe it is equal to J (X, −KX , OX (−m(KX +∆)) m ·at ). Note that since R(−KX −∆)
e is independent of the choice of m (at least for m
is finitely generated, the choice of X
sufficiently divisible).
Since KX − KX is Cartier, we know that
1
1
J (X, −KX , OX (−m(KX + ∆)) m · at )p = τ (Xp , −KXp , OXp (−m(KXp + ∆p )) m )
for p ≫ 0 by [Tak04, HY03]. But Lemma 5.2 shows that τ (Xp , −KXp , OXp (−m(KXp +
1
∆p )) m ) = τ (Xp , ∆p , atp ). Combining these equalities proves the result.
Remark 6.5. Theorem 6.4 also implies that if Xp is strongly F -regular for all p ≫ 0 and
R(−KX ) is finitely generated, then X is KLT. Nobuo Hara gave a talk about this result at
the conference in honor of Mel Hochster’s 65th Birthday in 2008 but the result was never
published.
Corollary 6.6. Suppose X is a variety over an algebraically closed field of characteristic
zero that is KLT in the sense of [DH09]. Then for any Q-divisor ∆ ≥ 0 and any ideal sheaf
a and rational t ≥ 0, we have that
J (X, ∆, at )p = τ (Xp , ∆p , atp )
for p ≫ 0.
Proof. It follows from the minimal model program [BCHM10] and in particular, [Kol08,
Theorem 92] that R(−KX − ∆) is finitely generated, the result follows immediately from
Theorem 6.4.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
29
References
[AM99]
I. M. Aberbach and B. MacCrimmon: Some results on test elements, Proc. Edinburgh Math.
Soc. (2) 42 (1999), no. 3, 541–549. MR1721770 (2000i:13005)
[Bha12]
B. Bhatt: Derived splinters in positive characteristic, Compos. Math. 148 (2012), no. 6, 1757–
1786. 2999303
[Bir13]
C. Birkar: Existence of flips and minimal models for 3-folds in char p, arXiv:1311.3098, to
appear in Annales Scientifiques de l’ENS.
[BCHM10] C. Birkar, P. Cascini, C. D. Hacon, and J. McKernan: Existence of minimal models for
varieties of log general type, J. Amer. Math. Soc. 23 (2010), no. 2, 405–468. 2601039 (2011f:14023)
[BW14]
C. Birkar and J. Waldron: Existence of Mori fibre spaces for 3-folds in char p,
arXiv:1410.4511.
[Bli13]
M. Blickle: Test ideals via algebras of p−e -linear maps, J. Algebraic Geom. 22 (2013), no. 1,
49–83. 2993047
[BMS08]
M. Blickle, M. Mustaţǎ, and K. E. Smith: Discreteness and rationality of F -thresholds,
Michigan Math. J. 57 (2008), 43–61, Special volume in honor of Melvin Hochster. 2492440
(2010c:13003)
[BMS09]
M. Blickle, M. Mustaţă, and K. E. Smith: F -thresholds of hypersurfaces, Trans. Amer.
Math. Soc. 361 (2009), no. 12, 6549–6565. 2538604 (2011a:13006)
[BS13]
M. Blickle and K. Schwede: p−1 -linear maps in algebra and geometry, Commutative algebra,
Springer, New York, 2013, pp. 123–205. 3051373
[BSTZ10] M. Blickle, K. Schwede, S. Takagi, and W. Zhang: Discreteness and rationality of F jumping numbers on singular varieties, Math. Ann. 347 (2010), no. 4, 917–949. 2658149
[BST15]
M. Blickle, K. Schwede, and K. Tucker: F -singularities via alterations, Amer. J. Math.
137 (2015), no. 1, 61–109. 3318087
[BdFFU15] S. Boucksom, T. de Fernex, C. Favre, and S. Urbinati: Valuation spaces and multiplier
ideals on singular varieties, 29–51. 3380442
[CTX13]
P. Cascini, H. Tanaka, and C. Xu: On base point freeness in positive characteristic,
1305.3502.
[Chi14]
A. Chiecchio: About a minimal model program without flips, arXiv:1410.6230.
[CU15]
A. Chiecchio and S. Urbinati: Ample Weil divisors, J. Algebra 437 (2015), 202–221. 3351963
[Cut88]
S. Cutkosky: Weil divisors and symbolic algebras, Duke Math. J. 57 (1988), no. 1, 175–183.
952230 (89e:14003)
[dFDTT15] T. de Fernex, R. Docampo, S. Takagi, and K. Tucker: Comparing multiplier ideals to test
ideals on numerically Q-Gorenstein varieties, Bull. Lond. Math. Soc. 47 (2015), no. 2, 359–369.
3335128
[DH09]
T. De Fernex and C. Hacon: Singularities on normal varieties, Compos. Math. 145 (2009),
no. 2, 393–414.
[dJ96]
A. J. de Jong: Smoothness, semi-stability and alterations, Inst. Hautes Études Sci. Publ. Math.
(1996), no. 83, 51–93. 1423020 (98e:14011)
[Dem88]
M. Demazure: Anneaux gradués normaux, Introduction à la théorie des singularités, II,
Travaux en Cours, vol. 37, Hermann, Paris, 1988, pp. 35–68. MR1074589 (91k:14004)
[ELSV04] L. Ein, R. Lazarsfeld, K. E. Smith, and D. Varolin: Jumping coefficients of multiplier
ideals, Duke Math. J. 123 (2004), no. 3, 469–506. MR2068967 (2005k:14004)
[FST11]
O. Fujino, K. Schwede, and S. Takagi: Supplements to non-lc ideal sheaves, Higher Dimensional Algebraic Geometry, RIMS Kôkyûroku Bessatsu, B24, Res. Inst. Math. Sci. (RIMS),
Kyoto, 2011, pp. 1–47.
[Gab04]
O. Gabber: Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter
de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711–734.
[GHNV90] S. Goto, M. Herrmann, K. Nishida, and O. Villamayor: On the structure of Noetherian
symbolic Rees algebras, Manuscripta Math. 67 (1990), no. 2, 197–225. 1042238 (91a:13006)
[HX15]
C. D. Hacon and C. Xu: On the three dimensional minimal model program in positive characteristic, J. Amer. Math. Soc. 28 (2015), no. 3, 711–744. 3327534
[Har01]
N. Hara: Geometric interpretation of tight closure and test ideals, Trans. Amer. Math. Soc.
353 (2001), no. 5, 1885–1906 (electronic). MR1813597 (2001m:13009)
[HY03]
N. Hara and K.-I. Yoshida: A generalization of tight closure and multiplier ideals, Trans.
Amer. Math. Soc. 355 (2003), no. 8, 3143–3174 (electronic). MR1974679 (2004i:13003)
30
[Har77]
[Har94]
[HS77]
[Hoc07]
[HH90]
[HH92]
[HH94]
[HL07]
[KLZ09]
[KSSZ14]
[Kol08]
[KM98]
[Kun76]
CHIECCHIO, ENESCU, MILLER, AND SCHWEDE
R. Hartshorne: Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in
Mathematics, No. 52. MR0463157 (57 #3116)
R. Hartshorne: Generalized divisors on Gorenstein schemes, Proceedings of Conference on
Algebraic Geometry and Ring Theory in honor of Michael Artin, Part III (Antwerp, 1992),
vol. 8, 1994, pp. 287–339. MR1291023 (95k:14008)
R. Hartshorne and R. Speiser: Local cohomological dimension in characteristic p, Ann. of
Math. (2) 105 (1977), no. 1, 45–79. MR0441962 (56 #353)
M. Hochster: Foundations of tight closure theory, lecture notes from a course taught on the
University of Michigan Fall 2007 (2007).
M. Hochster and C. Huneke: Tight closure, invariant theory, and the Briançon-Skoda theorem, J. Amer. Math. Soc. 3 (1990), no. 1, 31–116. MR1017784 (91g:13010)
M. Hochster and C. Huneke: Infinite integral extensions and big Cohen-Macaulay algebras,
Ann. of Math. (2) 135 (1992), no. 1, 53–89. 1147957 (92m:13023)
M. Hochster and C. Huneke: F -regularity, test elements, and smooth base change, Trans.
Amer. Math. Soc. 346 (1994), no. 1, 1–62. MR1273534 (95d:13007)
C. Huneke and G. Lyubeznik: Absolute integral closure in positive characteristic, Adv. Math.
210 (2007), no. 2, 498–504. 2303230 (2008d:13005)
M. Katzman, G. Lyubeznik, and W. Zhang: On the discreteness and rationality of F jumping coefficients, J. Algebra 322 (2009), no. 9, 3238–3247. 2567418 (2011c:13005)
M. Katzman, K. Schwede, A. K. Singh, and W. Zhang: Rings of Frobenius operators,
Math. Proc. Cambridge Philos. Soc. 157 (2014), no. 1, 151–167. 3211813
J. Kollár: Exercises in the birational geometry of algebraic varieties, arXiv:0809.2579.
J. Kollár and S. Mori: Birational geometry of algebraic varieties, Cambridge Tracts in Mathematics, vol. 134, Cambridge University Press, Cambridge, 1998, With the collaboration of C.
H. Clemens and A. Corti, Translated from the 1998 Japanese original. MR1658959 (2000b:14018)
E. Kunz: On Noetherian rings of characteristic p, Amer. J. Math. 98 (1976), no. 4, 999–1013.
MR0432625 (55 #5612)
[Laz04]
[Lip69]
[Lyu97]
[LS99]
[LS01]
R. Lazarsfeld: Positivity in algebraic geometry. I, Ergebnisse der Mathematik und ihrer
Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics
and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 48, SpringerVerlag, Berlin, 2004, Classical setting: line bundles and linear series. MR2095471 (2005k:14001a)
J. Lipman: Rational singularities, with applications to algebraic surfaces and unique factorization, Inst. Hautes Études Sci. Publ. Math. (1969), no. 36, 195–279. MR0276239 (43 #1986)
G. Lyubeznik: F -modules: applications to local cohomology and D-modules in characteristic
p > 0, J. Reine Angew. Math. 491 (1997), 65–130. MR1476089 (99c:13005)
G. Lyubeznik and K. E. Smith: Strong and weak F -regularity are equivalent for graded rings,
Amer. J. Math. 121 (1999), no. 6, 1279–1290. MR1719806 (2000m:13006)
G. Lyubeznik and K. E. Smith: On the commutation of the test ideal with localization and
completion, Trans. Amer. Math. Soc. 353 (2001), no. 8, 3149–3180 (electronic). MR1828602
(2002f:13010)
[Mus13]
[Pér13]
[Sch09]
M. Mustaţă: The non-nef locus in positive characteristic, A celebration of algebraic geometry,
Clay Math. Proc., vol. 18, Amer. Math. Soc., Providence, RI, 2013, pp. 535–551. 3114955
F. Pérez: On the constancy regions for mixed test ideals, J. Algebra 396 (2013), 82–97. 3108073
K. Schwede: F -adjunction, Algebra Number Theory 3 (2009), no. 8, 907–950. 2587408
(2011b:14006)
[Sch11]
[SS10]
[ST14a]
[ST14b]
[STZ12]
K. Schwede: Test ideals in non-Q-Gorenstein rings, Trans. Amer. Math. Soc. 363 (2011),
no. 11, 5925–5941.
K. Schwede and K. E. Smith: Globally F -regular and log Fano varieties, Adv. Math. 224
(2010), no. 3, 863–894. 2628797 (2011e:14076)
K. Schwede and K. Tucker: On the behavior of test ideals under finite morphisms, J. Algebraic Geom. 23 (2014), no. 3, 399–443. 3205587
K. Schwede and K. Tucker: Test ideals of non-principal ideals: Computations, jumping
numbers, alterations and division theorems, J. Math. Pures Appl. (9) 102 (2014), no. 5, 891–
929. 3271293
K. Schwede, K. Tucker, and W. Zhang: Test ideals via a single alteration and discreteness
and rationality of f -jumping numbers, Math. Res. Lett. 19 (2012), no. 01, 191–197.
TEST IDEALS AND ANTI-CANONICAL ALGEBRAS
[Sin99]
[Sin14]
[Smi00]
[Tak04]
[Wat81]
[Xu15]
31
A. K. Singh: Q-Gorenstein splinter rings of characteristic p are F-regular, Math. Proc. Cambridge Philos. Soc. 127 (1999), no. 2, 201–205. 1735920 (2000j:13006)
A. K. Singh: Private communication.
K. E. Smith: The multiplier ideal is a universal test ideal, Comm. Algebra 28 (2000), no. 12,
5915–5929, Special issue in honor of Robin Hartshorne. MR1808611 (2002d:13008)
S. Takagi: An interpretation of multiplier ideals via tight closure, J. Algebraic Geom. 13 (2004),
no. 2, 393–415. MR2047704 (2005c:13002)
K. Watanabe: Some remarks concerning Demazure’s construction of normal graded rings,
Nagoya Math. J. 83 (1981), 203–211. 632654 (83g:13016)
C. Xu: On the base-point-free theorem of 3-folds in positive characteristic, J. Inst. Math. Jussieu
14 (2015), no. 3, 577–588. 3352529
TASIS in Dorado
E-mail address: [email protected]
Department of Mathematics and Statistics, Georgia State University, Atlanta GA 30303
USA
E-mail address: [email protected]
Department of Mathematical Sciences, University of Arkansas, Fayetteville, AR 72701
E-mail address: [email protected]
Department of Mathematics, University of Utah, 155 S 1400 E Room 233, Salt Lake City,
UT, 84112
E-mail address: [email protected]
| 0math.AC
|
arXiv:1602.04689v1 [math.GR] 15 Feb 2016
CHARACTER DEGREES OF SOME p-GROUPS
Avinoam Mann
It was shown by I.M.Isaacs [Is2], that any set of powers of a prime p which
includes 1, can occur as the set of character degrees of some p-group. It becomes
then of interest to see to what extent that remains true if we consider a particular
class of p-groups. Isaacs’ construction yields groups of nilpotency class 2. Here
we consider the other extreme. Recall that a group of order pn ≥ p3 is said to
be of maximal class, if its nilpotency class cl(G) is n − 1 (see [LGM] for the well
developed theory of these groups). Such a group has a non-abelian factor group of
order p3 , and therefore it has irreducible characters of degree p. It was suggested in
the last section of [Sl1], that there are further restrictions on the possible character
degrees set of a group of maximal class. The present note verifies that conjecture,
indeed under weaker assumptions than maximal class, for some results it suffices to
assume that the derived subgroup has index p2 . This last assumption has quite a
few consequences for the structure of the given group, and a secondary aim of this
note is to derive some of them (see Theorem 4 and Propositions 5 and 6. These
results are going to be applied in [Ma2]). About character degrees our results are:
Theorem 1. Let G be a p-group in which |G : G′ | = p2 . If G has an irreducible
p+1
character of degree > p, then it has such a character of degree at most p 2 .
We remark that 2-groups satisfying |G : G′ | = 4 are of maximal class, and their
character degrees are 1 and 2. For all odd primes, however, there exist groups
satisfying |G : G′ | = p2 which are not of maximal class, and which have irreducible
characters of degrees > p.
M.Slattery [Sl2] has constructed, for all primes p ≥ 5, groups of maximal class
p+1
whose character degrees are 1, p, p 2 , showing that the bound of Theorem 1 is best
possible. It is also easy to see that there exist 3-groups of that type.
Theorem 2. Let G be a p-group of maximal class. If G has irreducible characters
p+3
of degree higher than p2 , then it has such characters of degree at most p 2 .
Theorem 3. Let G be a p-group satisfying |G : G′ | = p2 . If |G| ≥ p6 , and
|γ3 (G) : γ4 (G)| > p (equivalently, if |G : γ4 (G)| = p5 ), then G has irreducible
characters of degree p2 .
For the proofs we first quote some of the theory of groups of maximal class, the
proofs of which can be found in [LGM] and in [Hu, III.14]. Let X be a p-group
of maximal class, of class c, say. We write Xi = γi (X) for the terms of the lower
central series, for 2 ≤ i ≤ c, and X1 = CX (X2 /X4 ). These notations will be
applied to each group of maximal class that we will encounter below, e.g. if that
group will be denoted by H, we will let Hi denote the corresponding subgroups
Typeset by AMS-TEX
1
2
AVINOAM MANN
of H, etc. Returning to X, we have X1 = CX (Xi /Xi+2 ), for 2 ≤ i < c − 1. We
call X1 the major centralizer of X. It is a regular p-group, therefore, if A, B ⊳ X1 ,
then [A, B p ] = [A, B]p (see [Hu, III.10] for the theory of regular p-groups). If
|X| ≥ pp+2 , then Xip = Xi+p−1 . This holds also for order pp+1 , except for one
group, the wreath product of two groups of order p, in which Xip = 1 for all i. If
|X| ≤ pp , then Xip = 1 for i ≥ 2. If X1 6= CX (Xc−1 ), then X is termed exceptional.
If X is metabelian, it is not exceptional. In a non-exceptional group, all maximal
subgroups different from X1 are of maximal class. Finally, X is of maximal class
iff it contains an element x ∈ X such that |CX (x)| = p2 .
Next, about groups with a derived subgroup of index p2 . Since all normal subgroups of index p2 contain the commutator subgroup, it follows that G′ is the only
normal subgroup of that index. The factor group G/G′ is elementary abelian, and if
x and y, say, are elements of G that are independent modulo G′ , then they generate
G, and the factor group G′ /γ3 (G) is generated by the image of [x, y]. Therefore
γ3 (G) has index p3 in G and it is the only normal subgroup of that index. Similarly,
γ3 (G)/γ4 (G) is generated by the images of [x, y, x] and [x, y, y], therefore γ4 (G) has
index either p4 or p5 , and in the first case it is the only normal subgroup of that
index.
Recall that d(X) denotes the minimal number of generators of the group X.
Theorem 4. Let G be a p-group such that |G| ≥ p3 and |G : G′ | = p2 , and let H
be a maximal subgroup of G. Then
(a) G/H ′ is of maximal class, and d(H) ≤ p.
(b) If d(H) < p − 1, then |Φ(H) : H ′ | ≤ p, hence |H : H ′ | equals either pd(H)+1
or pd(H) .
(c) If d(H) = p, then G/H ′ ∼
= Cp wrCp , and H/H ′ is elementary abelian of order
p
p .
(d) If |γ3 (G) : γ4 (G)| = p, then |γ4 (G) : γ5 (G)| = p, and all maximal subgroups
H of G, save one, satisfy H ′ = γ3 (G).
(e) If |γ3 (G) : γ4 (G)| = p2 , then H ′ 6= γ3 (G) and H ′ 6≤ γ4 (G). If K is another
maximal subgroup, then H ′ 6≤ K ′ , and in particular H ′ 6= K ′ .
(f ) If G contains a 2-generator maximal subgroup H, then either H is metacyclic,
or G contains at most one maximal subgroup K 6= H such that |K : K ′ | ≥ pp+1 .
(g) If γ3 (G) 6= 1, then Gp ≤ γ3 (G).
Note. The first claim in (a) was already pointed out in Exercise 1.4 of [Be].
Proof. For (a), we may as well assume that H ′ = 1. Then G = hH, xi for some
element x, and commutation with x induces on H an endomorphism with image
[G, H] = G′ and kernel Z(G). Since |H : G′ | = p, we have |Z(G)| = p. Moreover,
the same argument shows that all non-abelian quotients of G have centres of order
p, hence G is of maximal class.
If |H : H ′ | > pp−1 , then K := G/H ′ is a p-group of maximal class and order
at least pp+1 . Since L := H/H ′ is abelian, it is the major centralizer of K. It is
known that in such groups K p = γp (K), a subgroup of index pp , and that usually
Lp = K p , the only exception to this equality occurring when K = Cp wrCp , when
L is elementary abelian of order pp . In the other cases we have d(H) = d(L) =
logp |L : Φ(L)| = logp |L : Lp | = p − 1.
If |H : H ′ | ≤ pp−1 , then d(H) ≤ p − 1, and by the above, this is the only case in
which strict inequality is possible. Moreover, in that case |K| ≤ pp , and it is known
that groups of maximal class of these orders satisfy |K p | ≤ p, which proves (b).
CHARACTER DEGREES OF SOME p-GROUPS
3
When |γ3 (G) : γ4 (G)| = p, write C = CG (G′ /γ4 (G)). This is a maximal subgroup, and C ′ = [C, G′ ] ≤ γ4 (G). Let H be another maximal subgroup. By
Lemma 30 of [Ma1], |G′ : C ′ H ′ | ≤ p, i.e. |G : C ′ H ′ | ≤ p3 . But C ′ H ′ ≤ γ3 (G), and
|G : γ3 (G)| = p3 . Therefore C ′ H ′ = γ3 (G). Since C ′ ≤ γ4 (G), this is possible only
if H ′ = γ3 (G).
Since d(G) = 2, we have G′′ ≤ γ5 (G) [Hu, III.2.12(b)], and since |G : γ4 (G)| = p4 ,
an argument of R.Brandl implies that G/G′′ is of maximal class (see Theorem C of
[Ma1]), which shows that |γ4 (G) : γ5 (G)| = p.
If |γ3 (G) : γ4 (G)| = p2 , and H is a maximal subgroup, then H ′ 6≤ γ4 (G), because
G/H ′ is of maximal class. Let M and N be two subgroups lying properly between
γ3 (G) and γ4 (G), and let C = CG (G′ /M ) and D = CG (G′ /N ). Then M and N
are normal and C and D are maximal. If C = D, then C ′ = [C, G′ ] ≤ γ4 (G),
contradiction. Thus the p + 1 subgroups like M and N determine p + 1 distinct
maximal subgroups, i.e. each maximal subgroup is obtained in this way. Moreover,
C ′ ≤ M , thus C ′ 6= γ3 (G). But C ′ D′ = γ3 (G), therefore D′ 6≤ C ′ .
To prove (f) we suppose that G contains the 2-generator maximal subgroup H,
and two other maximal subgroups, K, and L, whose commutator factor groups
have orders at least pp+1 . If |H : H p | ≤ p2 , then H is either cyclic or metacyclic.
Thus we may assume that |H : H p | ≥ p3 . We write N := H p γ3 (H) (if p = 3
then N = H 3 ). Then |H : N | = p3 . On the other hand, since G/K ′ is a group of
maximal class of order at least pp+2 , its maximal subgroup H/K ′ is also of maximal
class and of order at least pp+1 , and this implies that |H : K ′ N | = p3 . Therefore
N = K ′ N , i.e. K ′ ≤ N . Similarly, L′ ≤ N , which leads to the contradiction
K ′ L′ ≤ N < γ3 (G).
Finally, if γ3 (G) 6= 1, let H = G/γ4 (G). Then Z(H) = γ3 (G)/γ4 (G), and
H/Z(H) is one of the two non-abelian groups of order p3 . But of these two, the
one of exponent p2 cannot be the central factor group of any group (is incapable),
therefore H/Z(H) has exponent p, which is our claim.
For p = 3 we have p − 1 = 2, hence
Proposition 5. Let G be a 3-group such that |G : G′ | = 9 and |G| ≥ 81. Then
G3 = γ3 (G) and that subgroup has index 27. All maximal subgroups of G have
at most three generators, and either one of them is metacyclic, or G contains a
maximal subgroup H such that |H : H ′ | ≤ 27. If |G : γ4 (G)| = 34 , then G is of
maximal class.
Proof. Most of this is either stated in Theorem 4, or was derived during the
proof of part (f) there. The last claim is the case p = 3 of the fact, that if in a
p-group G the factor group G/γp+1 (G) is of maximal class, then so is G [B1, 3.9].
Proposition 6. Let G be a p-group, of order p5 at least, in which |G : G′ | ≤ p3 .
Then all maximal subgroups of G, with at most one exception, have irreducible
characters of degree p. If an exceptional maximal subgroup exists, then all other
maximal subgroups M satisfy |M : M ′ | ≤ |G : G′ |.
Proof. Since |G : G′ | ≤ p3 , G has irreducible characters of degree p. By [Ma1,
Proposition 14(2)], if some maximal subgroup H does not have such characters,
and M is another maximal subgroup, then |G′ : M ′ | ≤ p. That means that |M :
M ′ | ≤ |G : G′ |, and this shows that M has irreducible characters of degree p. QED
An exceptional maximal subgroup may or may not exist. If G is a p-group of
maximal class, which is metabelian, but in which the subgroup G1 is not abelian,
4
AVINOAM MANN
then all irreducible characters of all maximal subgroups of G have degrees 1 or p.
On the other hand, in the groups of maximal class and order p2r (p > 3, 2r =
6, 8, ..., p + 1) constructed by N.Blackburn [B1, pp. 61-62, or Hu, III.14.24], or the
ones of order pp+2 constructed by Slattery in [Sl2] (p > 3), it is not difficult to show
that the maximal subgroups H (in [B1]) or E (in [Sl2]) have only irreducibles of
p−1
degrees 1 and pr−1 , or 1 and p 2 .
In the proofs of Theorems 1 to 3 we separate between groups of maximal class
and others.
Proof of Theorem 1 for groups of maximal class, and of Theorem 2. We
may assume that |G| ≥ pp+2 . By a famous result of Isaacs-Passman [Is1, 12.11],
a p-group G has characters of degrees 1 and p only, iff either |G : Z(G)| ≤ p3
or G contains an abelian maximal subgroup. For groups of maximal class the
first possibility means that |G| ≤ p4 , and then anyway G has an abelian maximal
subgroup. Moreover, if G has an abelian maximal subgroup, that subgroup must
be G1 . Thus the assumption that G has characters of degree deg(χ) > p means
that G′1 6= 1. Now [G1 , Gp ] = [G1 , Gp1 ] = [G1 , G1 ]p is properly contained in G′1 . In
H = G/[G1 , Gp ] we have H1′ 6= 1, therefore H has irreducibles of degree exceeding
p. But Hp ≤ Z(H1 ), and |H1 : Hp | = pp−1 , therefore the irreducible characters of
p−1
p+1
H1 have degrees at most p 2 , and those of H have degrees at most p 2 .
Next, let G have characters of degree deg(χ) > p2 . If |G| ≤ p8 , then |G : Z(G)| ≤
7
p , and all irreducibles of G have degree p3 at most. Thus to prove our claim we
may assume that |G| ≥ p9 . According to [Pa], a p-group X has only characters of
degree at most p2 , for p odd, iff one of the following four possibilities occur:
(i) G contains an abelian subgroup of index ≤ p2 ,
(ii) G contains a maximal subgroup H such that |H : Z(H)| ≤ p3 ,
(iii) |G : Z(G)| ≤ p5 ,
(iv) |G : Z(G)| = p6 , and if H ≥ Z(G) is a maximal subgroup of G, then
Z(H) = Z(G).
In a group of maximal class (iv) is impossible, because CG (Z2 (G)) is a maximal
subgroup, and (iii) means that |G| ≤ p6 . (ii) means that either G is an exceptional
group of order at most p6 , or that [G1 , G4 ] = 1, and (i) means that G′ is abelian.
Thus if G has an irreducible character of degree at least p3 , then G′′ 6= 1 6= [G1 , G4 ].
Note that by our assumption G8 6= 1, and then G/G7 is non-exceptional, which
implies that [G1 , G4 ] ≤ G6 . There are two indices i, j, such that G′′ = Gi and
[G1 , G4 ] = Gj . First assume that i > j, and let H = G/Gi+1 . Then H ′′ 6=
1 6= [H1 , H4 ]. The last inequality shows that H6 6= 1, therefore |H| ≥ p7 . Thus
H violates (i)-(iv), and it has an irreducible character of degree at least p3 . But
1 = (H ′′ )p = [H ′ , (H ′ )p ] = [H2 , Hp+1 ]. Thus Hp+1 ≤ Z(H ′ ), and again the
p−1
characters of H ′ have degrees at most p 2 and the characters of H have degrees at
p+3
most p 2 . If i ≤ j, we take H = G/Gj+1 , and obtain 1 = [H1 , H4 ]p = [H1 , Hp+3 ],
with |H1 : Hp+3 | = pp+2 , and proceed as before.
Proof of the rest of Theorem 1, and of Theorem 3. Since G is not of
maximal class, neither is G/γp+1 (G) [Bl, 3.9]. Let i ≥ 2 be the first index such that
|γi (G)/γi+1 (G)| > p. Thus 3 ≤ i ≤ p. Then |γi (G)/γi+1 (G)| = p2 . Theorem 4(a)
shows that if H is a maximal subgroup of G, then H ′ 6≤ γi+1 (G). The fact that G
has characters of degree > p implies that |G : Z(G)| > p3 , therefore γ3 (G) 6≤ Z(G),
i.e. γ4 (G) 6= 1. Assume first that i + 1 6= 4, and let K = G/γi+1 (G). Then in
CHARACTER DEGREES OF SOME p-GROUPS
5
K = G/γi+1 (G) no maximal subgroup is abelian, and the centre has index > p3 ,
therefore K has an irreducible character χ such that deg(χ) > p. On the other
p+1
hand, |K| = pi+2 ≤ pp+2 , therefore deg(χ) ≤ p 2 .
If i = 3, then γi+1 (G) = γ4 (G) 6= 1. Let N ⊳ G have index p in γi+1 (G), let
K = G/N , and proceed as before.
If |G : γ4 (G)| = p5 , then we saw in Theorem 4 that if C and D are two maximal
subgroups, then C ′ 6≤ D′ . This implies that C ′ 6≤ G′′ , and thus L := G/G′′ does
not have abelian maximal subgroups. And |L : Z(L)| > p3 , because [γ3 (G), G] =
γ4 (G) 6≤ G′′ , and γ3 (G) is the only normal subgroup of index p3 . It follows that
G/G′′ has irreducible characters of degree bigger that p, and that degree must be
p2 , because G′ /G′′ is an abelian subgroup of index p2 . QED
References
Be. Y.Berkovich, Groups of Prime Power Order, vol. 1, de Gruyter, Berlin
2008.
Bl. N.Blackburn, On a special class of p-groups, Acta Math. 100 (1958), 45-92.
Hu. B.Huppert, Endliche Gruppen I, Springer, Berlin 1967.
Is1. I.M.Isaacs, Character Theory of Finite Groups, Academic Press, San Diego
1976.
Is2. I.M.Isaacs, Sets of p-powers as irreducible character degrees, Proc. Amer.
Math. Soc. 96 (1986), 551-552.
LGM. C.R.Leedham-Green and S.McKay, The Structure of Groups of Prime
Power Order, Oxford University Press, Oxford 2002.
Ma1. A.Mann, Minimal characters of p-groups, J. Gp. Th. 2 (1999), 225-250.
Ma2. A.Mann, More on normally monomial p-groups, in preparation.
Pa. D.S.Passman, Groups whose irreducible representations have degrees dividing p2 , Pac. J. Math. 17 (1966), 475-496.
Sl1. M.C.Slattery, Character degrees of normally monomial maximal class 5groups, in Character Theory of Finite Groups (the Isaacs conference), Contemporary Mathematics 524, American Mathematical Society, Providence 2010.
Sl2. M.C.Slattery, Maximal class p-groups with large character degree gaps,
preprint.
| 4math.GR
|
PREPRINT VERSION
An agent-driven semantical identifier using radial basis neural
networks and reinforcement learning
C. Napoli, G. Pappalardo, and E. Tramontana
arXiv:1409.8484v1 [cs.NE] 30 Sep 2014
PUBLISHED ON: Proceedings of the XV Workshop ”Dagli Oggetti agli Agenti”
BIBITEX:
@inproceedings{Napoli2014Anagent
year={2014},
issn={1613-0073},
url={http://ceur-ws.org/Vol-1260/},
booktitle={Proceedings of the XV Workshop ”Dagli Oggetti agli Agenti”},
title={An agent-driven semantical identifier using radial basis neural networks and reinforcement learning},
publisher={CEUR-WS},
volume={1260},
author={Napoli, Christian and Pappalardo, Giuseppe and Tramontana, Emiliano},
}
Published version copyright c 2014
UPLOADED UNDER SELF-ARCHIVING POLICIES
An agent-driven semantical identifier using radial
basis neural networks and reinforcement learning
Christian Napoli, Giuseppe Pappalardo, and Emiliano Tramontana
Department of Mathematics and Informatics
University of Catania, Viale A. Doria 6, 95125 Catania, Italy
{napoli, pappalardo, tramontana}@dmi.unict.it
Abstract—Due to the huge availability of documents in digital
form, and the deception possibility raise bound to the essence of
digital documents and the way they are spread, the authorship
attribution problem has constantly increased its relevance. Nowadays, authorship attribution, for both information retrieval and
analysis, has gained great importance in the context of security,
trust and copyright preservation.
This work proposes an innovative multi-agent driven machine
learning technique that has been developed for authorship attribution. By means of a preprocessing for word-grouping and timeperiod related analysis of the common lexicon, we determine a
bias reference level for the recurrence frequency of the words
within analysed texts, and then train a Radial Basis Neural
Networks (RBPNN)-based classifier to identify the correct author.
The main advantage of the proposed approach lies in the generality of the semantic analysis, which can be applied to different
contexts and lexical domains, without requiring any modification.
Moreover, the proposed system is able to incorporate an external
input, meant to tune the classifier, and then self-adjust by means
of continuous learning reinforcement.
I.
I NTRODUCTION
Nowadays, the automatic attribution of a text to an author,
assisting both information retrieval and analysis, has become
an important issue, e.g. in the context of security, trust and
copyright preservation. This results from the availability of
documents in digital form, and the raising deception possibilities bound to the essence of the digital reproducible contents,
as well as the need for new mechanical methods that can
organise the constantly increasing amount of digital texts.
During the last decade only, the field of text classification
and attribution has undergone new developement due to the
novel availability of computational intelligence techniques,
such as natural language processing, advanced data mining
and information retrieval systems, machine learning and artificial intelligence techniques, agent oriented programming,
etc. Among such techniques, Computer Intelligence (CI) and
Evolutionary Computation (EC) methods have been largely
used for optimisation and positioning problems [1], [2]. In [3],
agent driven clustering has been used as an advanced solution
for some optimal management problems, whereas in [4] such
problems are solved for mechatronical module controls. Agent
driven artificial intelligence is often used in combination with
advanced data analysis techniques in order to create intelligent
control systems [5], [6] by means of multi resolution analysis [7]. CI and parallel analysis systems have been proposed
in order to support developers, as in [8], [9], [10], [11],
where such a classification and analysis was applied to assist
refactoring in large software systems [12], [13], [14], [15].
Moreover, CI and techniques like neural networks (NNs)
have been used in [16], [17] in order to model electrical
networks and the related controls starting by classification
strategies, as well as for other complex physical systems
by using several kinds of hybrid NN-based approaches [18],
[19], [20], [21]. All the said works use different forms of
agent-based modeling and clustering for recognition purposes,
and these methods efficiently perform very challenging tasks,
where other common computational methods failed or had
low efficiency or, simply, resulted as inapplicable due to
complicated model underlying the case study. In general,
agent-driven machine learning has been proven as a promising
field of research for the purpose of text classification, since
it allows building classification rules by means of automatic
learning, taking as a basis a set of known texts and trying to
generalise for unknown ones.
While machine learning and NNs are a very promising
field, the effectiveness of such approaches often lies on the
correct and precise preprocessing of data, i.e. the definition
of semantic categories, affinities and rules used to generate a
set of numbers characterising a text sample, to be successively
given as input to a classifier. Typical text classification, e.g.
by using NNs, takes advantage of topics recognition, however
results are seldom appropriate when it comes to classify people
belonging to the same social group or who are involved
in a similar business (e.g. the classification of: texts from
different scientists in the same field of research, the politicians
belonging to the same party, texts authored by different people
using the same technical jargon).
In our approach we devise a solution for extracting from the
analysed texts some characteristics that can express the style of
a specific author. Obtaining this kind of information abstraction
is crucial in order to create a precise and correct classification
system. On the other hand, while data abound in the context
of text analysis, a robust classifier should rely on input sets
that are compact enough to be apt to the training process.
Therefore, some data have to reflect averaged evaluations that
concern some anthropological aspects such as the historical
period, or the ethnicity, etc. This work satisfies the above
conditions of extracting compact data from texts since we
use a preprocessing tool for word-grouping and time-period
related analysis of the common lexicon. Such a tool computes
a bias reference system for the recurrence frequency of the
word used in the analysed texts. The main advantage of this
choice lies in the generality of the implemented semantical
REFERENCE
DATABASE
TEXT
DATABASE
TRAINING
SET
KNOWN
PREPROCESSING
BIASING
NEW
DATA
UNKNOWN
RBPNN
with
reinforcement
learning
LOCAL
EXTERNAL
TEXT
DATABASE
WORDNET
LEXICON
Fig. 1. A general schema of the data flow through the agents of the developed
system.
identifier, which can be then applied to different contexts and
lexical domains without requiring any modification. Moreover,
in order to have continuous updates or complete renewals of
the reference data, a statically trained NN would not suffice
to the purpose of the work. For these reasons, the developed
system is able to self-correct by means of continuous learning
reinforcement. The proposed architecture also diminishes the
human intervention over time thanks to its self-adaption properties. Our solution comprises three main collaborating agents:
the first for preprocessing, i.e. to extract meaningful data from
texts; the second for classification by means of a proper Radial
Basis NN (RBNN); and finally, one for adapting by means of
a feedforward NN.
The rest of this paper is as follows. Section II gives
the details of the implemented preprocessing agent based on
lexicon analysis. Section III describes the proposed classifier
agent based on RBNNs, our introduced modifications and
the structure of the reinforcement learning agent. Section IV
reports on the performed experiments and the related results.
Finally, Section V gives a background of the existing related
works, while Section VI draws our conclusions.
Algorithm 1: Find the group a word belongs to and count
occurrences
Start,
Import a speech into T ext,
Load dictionary into W ords,
Load group database into Groups,
thisW ord = Text.get();
while thisW ord do
thisGroup = Words.search(thisW ord);
if !thisGroup then
Load a different Lexicon;
if Lexicon.exist(thisW ord) then
W ords.update();
Groups.update();
end
else
break;
end
end
while Text.search(thisW ord) do
Groups.count(thisGroup);
end
thisW ord = Text.get();
end
Export W ords and Groups,
Stop.
The fundamental steps of the said analysis (see also Algorithm 1) are the followings:
1)
2)
3)
4)
5)
6)
II.
E XTRACTING SEMANTICS FROM LEXICON
Figure 1 shows the agents for our developed system: a
preprocessing agent extracts characteristics from given text
parts (see text database in the Figure), according to a known
set of words organised into groups (see reference database);
a RBPNN agent takes as input the extracted characteristics,
properly organised, and performs the identification on new
data, after appropriate training. An additional agent, dubbed
adaptive critic, shown in Figure 6, dynamically adapts the
behaviour of the RBPNN agent when new data are available.
Firstly, preprocessing agent analyses a text given as input
by counting the words that belong to a priori known groups of
mutually related words. Such groups contain words that pertain
to a given concern, and have been built ad-hoc and according
to the semantic relations between words, hence e.g. assisted
by the WordNet lexicon1 .
1 http://wordnet.princeton.edu
7)
8)
9)
import a single text file containing the speech;
import word groups from a predefined database, the
set containing all words from each group is called
dictionary;
compare each word on the text with words on the
dictionary;
if the word exists on the dictionary then the relevant
group is returned;
if the word has not been found then search the
available lexicon;
if the word exists on the lexicon then the related
group is identified;
if the word is unkown, then a new lexicon is loaded
and if the word is found then dictionary and groups
are updated;
search all the occurrences of the word in the text;
when an occurrence has been found, then remove it
from the text and increase the group counter.
Figure 2 shows the UML class diagram for the software
system performing the above analysis. Class Text holds a text
to be analysed; class Words represents the known dictionary,
i.e. all the known words, which are organised into groups given
by class Groups; class Lexicon holds several dictionaries.
III.
T HE RBPNN CLASSIFIER AGENT
For the work proposed here, we use a variation on Radial
Basis Neural Networks (RBNN). RBNNs have a topology similar to common FeedForward Neural Networks (FFNN) with
BackPropagation Training Algorithms (BPTA): the primary
Lexicon
Text
get()
exist()
get()
Words
search()
update()
Groups
Filter
service()
search()
count()
update()
FFNN
RBNN
PNN
Our RBPNN
Fig. 2. UML class diagram for handling groups and counting words belonging
to a group.
difference only lies in the activation function that, instead of
being a sigmoid function or a similar activation function, is a
statistical distribution or a statistically significant mathematical
function. The selection of transfer functions is indeed decisive
for the speed of convergence in approximation and classification problems [22]. The kinds of activation functions used
for Probabilistic Neural Networks (PNNs) have to meet some
important properties to preserve the generalisation abilities of
the ANNs. In addition, these functions have to preserve the
decision boundaries of the probabilistic neural networks. The
selected RBPNN architecture is shown in Figure 4 and takes
advantage from both the PNN topology and the Radial Basis
Neural Networks (RBNN) used in [23].
Each neuron performs a weighted sum of its inputs and
passes it through a transfer function f to produce an output.
This occurs for each neural layer in a FFNN. The network
can be perceived as a model connecting inputs and outputs,
with the weights and thresholds being free parameters of the
model, which are modified by the training algorithm. Such
networks can model functions of almost arbitrary complexity
with the number of layers and the number of units in each
layer determining the function complexity. A FFNN is capable
to generalise the model, and to separate the input space in
various classes (e.g. in a 2D variable space it is equivalent to
the separation of the different semi-planes). In any case, such
a FFNN can only create a general model of the entire variable
space, while can not insert single set of inputs into categories.
On the other hand, a RBNN is capable of clustering the inputs
by fitting each class by means of a radial basis function [24],
while the model is not general for the entire variable space, it
is capable to act on the single variables (e.g. in a 2D variable
space it locates closed subspaces, without any inference on the
remaining space outside such subspaces).
Another interesting topology is provided by PNNs, which
are mainly FFNNs also functioning as Bayesian networks
with Fisher Kernels [25]. By replacing the sigmoid activation
function often used in neural networks with an exponential
function, a PNN can compute nonlinear decision boundaries
approaching the Bayes optimal classification [26]. Moreover,
a PNN generates accurate predicted target probability scores
with a probabilistic meaning (e.g. in the 2D space it is
equivalent to attribute a probabilistic score to some chosen
points, which in Figure 3 are represented as the size of the
points).
Finally, in the presented approach we decided to combine
the advantages of both RBNN and PNN using the so called
Fig. 3. A comparison between results of several types of NNs, Our RBPNN
includes the maximum probability selector module
RBPNN. The RBPNN architecture, while preserving the capabilities of a PNN, due to its topology, then being capable
of statistical inference, is also capable of clustering since the
standard activation functions of a PNN are substituted by radial
basis functions still verifying the Fisher kernel conditions
required for a PNN (e.g. such an architecture in the 2D variable
space can both locate subspace of points and give to them a
probabilistic score). Figure 3 shows a representation of the
behaviour for each network topology presented above.
A. The RBPNN structure and topology
In a RBPNN both the input and the first hidden layer
exactly match the PNN architecture: the input neurones are
used as distribution units that supply the same input values
to all the neurones in the first hidden layer that, for historical
reasons, are called pattern units. In a PNN, each pattern unit
performs the dot product of the input pattern vector v by a
weight vector W(0) , and then performs a nonlinear operation
on the result. This nonlinear operation gives output x(1) that
is then provided to the following summation layer. While a
common sigmoid function is used for a standard FFNN with
BPTA, in a PNN the activation function is an exponential, such
that, for the j-esime neurone the output is
||W(0) · v||
(1)
(1)
xj ∝ exp
2σ 2
where σ represents the statistical distribution spread.
The given activation function can be modified or substituted
while the condition of Parzen (window function) is still satisfied for the estimator N̂ . In order to satisfy such a condition
some rules must be verified for the chosen window function in
order to obtain the expected estimate, which can be expressed
as a Parzen window estimate p(x) by means of the kernel K
of f in the d-dimensional space S d
n
P
x−xi
1
pn (x) = n1
K
d
hn
h
i=1
R
Sd
K(x)dx = 1
n
(2)
P1
Σ
υ
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Σ
.
.
.
.
.
.
P*
Pn
Σ
Fig. 4. A representation of a Radial Basis Probabilistic Neural Network with
maximum probability selector module
where hn ∈ N is called window width or bandwidth parameter
and corresponds to the width of the kernel. In general hn ∈ N
depends on the number of available sample data n for the
estimator pn (x). Since the estimator pn (x) converges in mean
square to the expected value p(x) if
lim hpn (x)i = p(x)
n→∞
(3)
lim var (pn (x))
n→∞
=
0
where hpn (x)i represents the mean estimator values and
var (pn (x)) the variance of the estimated output with respect
to the expected values, the Parzen condition states that such
convergence holds within the following conditions:
sup K(x) < ∞
x
lim xK(x)
=
0
lim hnd
=
0
|x|→∞
Fig. 5.
Setup values for the proposed RBPNN: NF is the number of
considered lexical groups, NS the number of analysed texts, and NG is the
number of people that can possibly be recognised as authors.
units work as in the neurones of a linear perceptron network.
The training for the output layer is performed as in a RBNN,
however since the number of summation units is very small
and in general remarkably less than in a RBNN, the training
is simplified and the speed greatly increased [27].
The output of the RBPNN (as shown in Figure 4) is given
to the maximum probability selector module, which effectively
acts as a one-neuron output layer. This selector receives as
input the probability score generated by the RBPNN and
attributes to one author only the analysed text, by selecting
the most probable author, i.e. the one having the maximum
input probability score. Note that the links to this selector are
weighted (with weights adjusted during the training), hence the
actual input is the product between the weight and the output
of the summation layer of the RBPNN.
(4)
n→∞
lim nhnd
n→∞
= ∞
In this case, while preserving the PNN topology, to obtain
the RBPNN capabilities, the activation function is substituted
with a radial basis function (RBF); an RBF still verifies all
the conditions stated before. It then follows the equivalence
between the W(0) vector of weights and the centroids vector of
a radial basis neural network, which, in this case, are computed
as the statistical centroids of all the input sets given to the
network.
We name f the chosen radial basis function, then the new
output of the first hidden layer for the j-esime neurone is
||v − W(0) ||
(1)
(5)
xj , f
β
where β is a parameter that is intended to control the distribution shape, quite similar to the σ used in (1).
The second hidden layer in a RBPNN is identical to a PNN,
it just computes weighted sums of the received values from the
preceding neurones. This second hidden layer is called indeed
summation layer: the output of the k-esime summation unit is
X
(2)
(1)
xk =
Wjk xj
(6)
j
where Wjk represents the weight matrix. Such weight matrix
consists of a weight value for each connection from the j-esime
pattern units to the k-esime summation unit. These summation
B. Layer size for a RBPNN
The devised topology enables us to distribute to different
layers of the network different parts of the classification task.
While the pattern layer is just a nonlinear processing layer, the
summation layer selectively sums the output of the first hidden
layer. The output layer fullfills the nonlinear mapping such as
classification, approximation and prediction. In fact, the first
hidden layer of the RBPNN has the responsibility to perform
the fundamental task expected from a neural network [28]. In
order to have a proper classification of the input dataset, i.e. of
analysed texts to be attributed to authors, the size of the input
layer should match the exact number NF of different lexical
groups given to the RBPNN, whereas the size of the pattern
units should match the number of samples, i.e. analysed texts,
NS . The number of the summation units in the second hidden
layer is equal to the number of output units, these should match
the number of people NG we are interested in for the correct
recognition of the speakers (Figure 5).
C. Reinforcement learning
In order to continuously update the reference database for
our system, a statically trained NN would not suffice for the
purpose of the work. Since the aim of the presented system is
having an expanding database of text samples for classification
and recognition purpose, the agent driven identification should
dynamically follow the changes in such a database. When
a new entry is made then the related feature set and biases
change, it implies that also the RBPNN should be properly
managed in order to ensure a continuous adaptive control for
reinforcement learning. Moreover, for the considered domain
it is desirable that a human supervisor supply suggestions, expecially when the system starts working. The human activities
are related to the supply of new entries into the text sample
database, and to the removal of misclassifications made by the
RBPNN.
We used a supervised control configuration (see Figure 6),
where the external control is provided by the actions and
choices of a human operator. While the RBPNN is trained with
a classical backpropagation learning algorithm, it is also embedded into an actor-critic reinforcement learning architecture,
which back propagates learning by evaluating the correctness
of the RBPNN-made choices with respect to the real word.
Let ξ be the error function, i.e. ξ = 0 for the results
supported by human verification, or the vectorial deviance for
the results not supported by a positive human response. This
assessment is made by an agent named Critic. We consider
the filtering step for the RBPNN output, to be both: Critic,
i.e. a human supervisor acknowledging or rejecting RBPNN
classifications; or Adaptive critic, i.e. an agent embedding a
NN that in the long run simulates the control activity made by
the human Critic, hence decreasing human control over time.
Adaptive critic needs to learn, and this learning is obtained by
a modified backpropagation algorithm using just ξ as error
function. Hence, Adaptive critic has been implemented by
a simple feedforward NN trained, by means of a traditional
gradient descent algorithm so that the weight modification
∆wij is
∆wij = −µ
∂ξ
∂ξ ∂ f˜i ∂ ũi
= −µ
∂wij
∂ f˜i ∂ ũi ∂wij
(7)
The f˜i is the activation of i-esime neuron, ũi is the i-esime
input to the neurone weighted as
X
ũi =
wij f˜j (ξi )
(8)
j
The result of the adaptive control determines whether to
continue the training of the RBPNN with new data, and
whether the last training results should be saved or discarded.
At runtime this process results in a continuous adaptive learning, hence avoiding the classical problem of NN polarisation
and overfitting. Figure 6 shows the developed learning system
reinforcement. According to the literature [29], [30], [31], [32],
straight lines represent the data flow, i.e. training data fed
to the RBPNN, then new data inserted by a supervisor, and
the output of the RBPNN sent to the Critic modules also
by means of a delay operator z −1 . Functional modifications
operated within the system are represented as slanting arrows,
i.e. the choices made by a human supervisor (Critic) modify
the Adaptive critic, which adjust the weight of its NN; the
combined output of Critic and Adaptive critic determines
whether the RBPNN should undergo more training epochs and
so modify its weights.
z -1
Critic
Features
DATA
RBPNN
Fig. 6. The adopted supervised learning model reinforcement. Slanting arrows
represent internal commands supplied in order to control or change the status
of the modules, straight arrows represent the data flow along the model, z −1
represents a time delay module which provides 1-step delayed outputs.
characteristics (see Section II), then such results have been
given to the classification agent. The total number of text
samples was 344, and we used 258 of them for training the
classification agent and 86 for validation. The text samples,
both for training and validation, were from different persons
that have given a speech (from A. Einstein to G. Lewis), as
shown in Figure 7.
Given the flexible structure of the implemented learning
model, the word groups are not fixed and can be modified,
added or removed over time by an external tuning activity. By
using the count of words in a group, instead of a word-by-word
counts, the multi-agent system realises a statistically driven
classifier that identifies the main semantic concerns regarding
the text samples, and then, attributes such concerns to the most
probable person.
The relevant information useful in order to recognise the
author of the speech is usually largely spread over a certain
number of word groups that could be indication of the cultural
extraction, heritage, field of study, professional category etc.
This implies that we can not exclude any word group, a priori,
while the RBPNN could learn to automatically enhance the
relevant information in order to classify the speeches.
Figure 7-left shows an example of the classifier performances for results generated by the RBPNN (before the filter
implemened by the probabilistic selector). Since the RBPNN
results have a probability between 0 and 1, then the shown
performance is 0 when a text was correctly attributed (or
not attributed) to a specific person. Figure 7-right shows the
performances of the system when including the probabilistic
selector. For this case, a boolean selection is involved, then
correct identifications are represented as 0, false positive
identifications as −1 (black marks), and missed identifications
as +1 (white marks).
For validation purposes, Figure 7-(left and right) shows
results according to e:
e = y − ỹ
IV.
E XPERIMENTAL SETUP
The proposed RBPNN architecture has been tested using
several text samples collected from public speeches of different
people both from the present and the past era. Each text sample
has been given to the preprocessing agent that extract some
Adaptive
critic
(9)
where e identifies the performance, ỹ the classification result,
and y the expected result. Lower e (negative values) identify
an excess of confidence in the attribution of a text to a person,
while greater e (positive values) identify a lack of confidence
in that sense.
Fig. 7. The obtained performance for our classification system before (left) and after (right) the maximum probability selector choice. The mean grey color
represents the correct classifications, while white color represents missed classification and black color false classifications.
The system was able to correctly attribute the text to the
proper author with only a 20% of missing assignments.
V.
R ELATED W ORKS
Several generative models can be used to characterise
datasets that determine properties and allow grouping data
into classes. Generative models are based on stochastic block
structures [33], or on ‘Infinite Hidden Relational Models’ [34],
and ‘Mixed Membership Stochastic Blockmodel’ [35]. The
main issue of class-based models is the type of relational
structure that such solutions are capable to describe. Since
the definition of a class is attribute-dependent, generally the
reported models risk to replicate the existing classes for each
new attribute added. E.g. such models would be unable to
efficiently organise similarities between the classes ‘cats’ and
‘dogs’ as child classes of the more general class ‘mammals’.
Such attribute-dependent classes would have to be replicated as
the classification generates two different classes of ‘mammals’:
the class ‘mammals as cats’ and the class ‘mammals as dogs’.
Consequently, in order to distinguish between the different
races of cats and dogs, it would be necessary to further
multiply the ‘mammals’ class for each one of the identified
race. Therefore, such models quickly lead to an explosion of
classes. In addition, we would either have to add another class
to handle each specific use or a mixed membership model, as
for crossbred species.
Another paradigm concerns the ’Non-Parametric Latent
Feature Relational Model’ [36], i.e. a Bayesian nonparametric
model in which each entity has boolean valued latent features
that influence the model’s relations. Such relations depend on
well-known covariant sets, which are neither explicit or known
in our case study at the moment of the initial analysis.
In [37], the authors propose a sequential forward feature
selection method to find the subset of features that are relevant
to a classification task. This approach uses novel estimation of
the conditional mutual information between candidate feature
and classes, given a subset of already selected features used as
a classifier independent criterion for evaluating feature subsets.
In [38], data from the charge-discharge simulation of
lithium-ions battery energy storage are used for classification
purposes with recurrent NNs and PNNs by means of a theoretical framework based on signal theory.
While showing the effectiveness of the neural network
based approaches, in our case study classification results are
given by means of a probability, hence the use of a RBPNN,
and an on-line training achieved by reinforcement learning.
VI.
C ONCLUSION
This work has presented a multi-agent system, in which an
agent analyses fragments of texts and another agent consisting
of a RBPNN classifier, performs probabilistic clustering. The
system has successfully managed to identify the most probable
author among a given list for the examined text samples. The
provided identification can be used in order to complement
and integrate a comprehensive verification system, or other
kinds of software systems trying to automatically identify
the author of a written text. The RBPNN classifier agent
is continuously trained by means of reinforcement learning
techniques in order to follow a potential correction provided by
an human supervisor, or an agent that learns about supervision.
The developed system was also able to cope with new data
that are continuously fed into the database, for the adaptation
abilities of its collaborating agents and their reasoning based
on NNs.
ACKNOWLEDGMENT
This work has been supported by project PRIME funded
within POR FESR Sicilia 2007-2013 framework and project
PRISMA PON04a2 A/F funded by the Italian Ministry of
University and Research within PON 2007-2013 framework.
R EFERENCES
[1]
C. Napoli, G. Pappalardo, E. Tramontana, Z. Marszałek, D. Połap,
and M. Woźniak, “Simplified firefly algorithm for 2d image key-points
search,” in IEEE Symposium Series on Computational Intelligence.
IEEE, 2014.
[2]
[3]
M. Gabryel, M. Woźniak, and R. K. Nowicki, “Creating learning sets
for control systems using an evolutionary method,” in Proceedings of
Artificial Intelligence and Soft Computing (ICAISC), ser. LNCS, vol.
7269. Springer, 2012, pp. 206–213.
F. Bonanno, G. Capizzi, A. Gagliano, and C. Napoli, “Optimal management of various renewable energy sources by a new forecasting
method,” in Proceedings of International Symposium on Power Electronics, Electrical Drives, Automation and Motion (SPEEDAM). IEEE,
2012, pp. 934–940.
[20]
[21]
[4]
A. Nowak and M. Woźniak, “Analysis of the active module mechatronical systems,” in Proceedings of Mechanika - ICM. Kaunas, Lietuva:
Kaunas University of Technology Press, 2008, pp. 371–376.
[22]
[5]
C. Napoli, G. Pappalardo, and E. Tramontana, “A hybrid neuro–wavelet
predictor for qos control and stability,” in Proceedings of AI*IA:
Advances in Artificial Intelligence. Springer, 2013, pp. 527–538.
[23]
[6]
F. Bonanno, G. Capizzi, G. L. Sciuto, C. Napoli, G. Pappalardo, and
E. Tramontana, “A novel cloud-distributed toolbox for optimal energy
dispatch management from renewables in igss by using wrnn predictors
and gpu parallel solutions,” in Power Electronics, Electrical Drives,
Automation and Motion (SPEEDAM), 2014 International Symposium
on. IEEE, 2014, pp. 1077–1084.
[7]
A. Nowak and M. Woźniak, “Multiresolution derives analysis of module
mechatronical systems,” Mechanika, vol. 6, no. 74, pp. 45–51, 2008.
[8]
C. Napoli, G. Pappalardo, and E. Tramontana, “Using modularity metrics to assist move method refactoring of large systems,” in Proceedings
of International Conference on Complex, Intelligent, and Software
Intensive Systems (CISIS). IEEE, 2013, pp. 529–534.
[9]
G. Pappalardo and E. Tramontana, “Suggesting extract class refactoring
opportunities by measuring strength of method interactions,” in Proceedings of Asia Pacific Software Engineering Conference (APSEC).
IEEE, December 2013.
[10]
E. Tramontana, “Automatically characterising components with concerns and reducing tangling,” in Proceedings of Computer Software
and Applications Conference (COMPSAC) Workshop QUORS. IEEE,
July 2013. DOI: 10.1109/COMPSACW.2013.114, pp. 499–504.
[11]
C. Napoli, G. Papplardo, and E. Tramontana, “Improving files availability for bittorrent using a diffusion model,” in IEEE 23nd International
Workshop on Enabling Technologies: Infrastructure for Collaborative
Enterprises - WETICE 2014, June 2014, pp. 191–196.
[12]
R. Giunta, G. Pappalardo, and E. Tramontana, “Aspects and annotations
for controlling the roles application classes play for design patterns,” in
Proceedings of Asia Pacific Software Engineering Conference (APSEC).
IEEE, December 2011, pp. 306–314.
[13]
A. Calvagna and E. Tramontana, “Delivering dependable
reusable components by expressing and enforcing design
decisions,” in Proceedings of Computer Software and Applications
Conference (COMPSAC) Workshop QUORS.
IEEE, July 2013.
DOI: 10.1109/COMPSACW.2013.113, pp. 493–498.
[14]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
R. Giunta, G. Pappalardo, and E. Tramontana, “AODP: refactoring code
to provide advanced aspect-oriented modularization of design patterns,”
in Proceedings of Symposium on Applied Computing (SAC). ACM,
2012.
[34]
[15]
E. Tramontana, “Detecting extra relationships for design patterns roles,”
in Proceedings of AsianPlop, March 2014.
[35]
[16]
G. Capizzi, C. Napoli, and L. Paternò, “An innovative hybrid neurowavelet method for reconstruction of missing data in astronomical
photometric surveys,” in Proceedings of Artificial Intelligence and Soft
Computing (ICAISC). Springer, 2012, pp. 21–29.
[36]
[17]
F. Bonanno, G. Capizzi, G. L. Sciuto, C. Napoli, G. Pappalardo, and
E. Tramontana, “A cascade neural network architecture investigating
surface plasmon polaritons propagation for thin metals in openmp,” in
Proceedings of Artificial Intelligence and Soft Computing (ICAISC), ser.
LNCS, vol. 8467. Springer, 2014, pp. 22–33.
[18]
C. Napoli, F. Bonanno, and G. Capizzi, “Exploiting solar wind time
series correlation with magnetospheric response by using an hybrid
neuro-wavelet approach,” Proceedings of the International Astronomical
Union, vol. 6, no. S274, pp. 156–158, 2010.
[19]
G. Capizzi, F. Bonanno, and C. Napoli, “Hybrid neural networks
architectures for soc and voltage prediction of new generation batteries
[37]
[38]
storage,” in Proceedings of International Conference on Clean Electrical Power (ICCEP). IEEE, 2011, pp. 341–344.
C. Napoli, F. Bonanno, and G. Capizzi, “An hybrid neuro-wavelet
approach for long-term prediction of solar wind,” IAU Symposium, no.
274, pp. 247–249, 2010.
G. Capizzi, F. Bonanno, and C. Napoli, “A new approach for lead-acid
batteries modeling by local cosine,” in Power Electronics Electrical
Drives Automation and Motion (SPEEDAM), 2010 International Symposium on, June 2010, pp. 1074–1079.
W. Duch, “Towards comprehensive foundations of computational intelligence,” in Challenges for Computational Intelligence. Springer,
2007, pp. 261–316.
G. Capizzi, F. Bonanno, and C. Napoli, “Recurrent neural networkbased control strategy for battery energy storage in generation systems
with intermittent renewable energy sources,” in Proceedings of International Conference on Clean Electrical Power (ICCEP). IEEE, 2011,
pp. 336–340.
S. Haykin, Neural Networks - A comprehensive foundation. Prentice
Hall, 2004.
S. Mika, G. Ratsch, W. Jason, B. Scholkopft, and K.-R. Muller,
“Fisher discriminant analysis with kernels,” in Proceedings of the Signal
Processing Society Workshop. Neural networks for signal processing IX.
IEEE, 1999.
D. F. Specht, “Probabilistic neural networks,” Neural networks, vol. 3,
no. 1, pp. 109–108, 1990.
H. Deshuang and M. Songde, “A new radial basis probabilistic neural
network model,” in Proceedings of Conference on Signal Processing,
vol. 2. IEEE, 1996.
W. Zhao, D.-S. Huang, and L. Guo, “Optimizing radial basis probabilistic neural networks using recursive orthogonal least squares algorithms
combined with micro-genetic algorithms,” in Proceedings of Neural
Networks, vol. 3. IEEE, 2003.
D. V. Prokhorov, R. A. Santiago, and D. C. W. II, “Adaptive
critic designs: A case study for neurocontrol,” Neural Networks,
vol. 8, no. 9, pp. 1367 – 1372, 1995. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/0893608095000429
H. Javaherian, D. Liu, and O. Kovalenko, “Automotive engine torque
and air-fuel ratio control using dual heuristic dynamic programming,”
in Proceedings of International Joint Conference on Neural Networks
(IJCNN), 2006, pp. 518–525.
B. Widrow and M. Lehr, “30 years of adaptive neural networks:
perceptron, madaline, and backpropagation,” Proceedings of the IEEE,
vol. 78, no. 9, pp. 1415–1442, Sep 1990.
J.-W. Park, R. Harley, and G. Venayagamoorthy, “Adaptive-criticbased optimal neurocontrol for synchronous generators in a power
system using mlp/rbf neural networks,” IEEE Transactions on Industry
Applications, vol. 39, no. 5, pp. 1529–1540, Sept 2003.
K. Nowicki and T. A. B. Snijders, “Estimation and prediction for
stochastic blockstructures,” Journal of the American Statistical Association, vol. 96, no. 455, pp. 1077–1087, 2001.
Z. Xu, V. Tresp, K. Yu, and H. peter Kriegel, “Infinite hidden relational
models,” in In Proceedings of International Conference on Uncertainity
in Artificial Intelligence (UAI), 2006.
E. M. Airoldi, D. M. Blei, E. P. Xing, and S. E. Fienberg, “Mixed
membership stochastic block models,” in Advances in Neural Information Processing Systems (NIPS). Curran Associates, 2009.
K. Miller, T. Griffiths, and M. Jordan, “Nonparametric latent feature
models for link prediction,” in Advances in neural information processing systems (NIPS). Curran Associates, Inc., 2009, vol. 22, pp.
1276–1284.
J. Novovičová, P. Somol, M. Haindl, and P. Pudil, “Conditional mutual
information based feature selection for classification task,” in Progress
in Pattern Recognition, Image Analysis and Applications. Springer,
2007, pp. 417–426.
F. Bonanno, G. Capizzi, and C. Napoli, “Some remarks on the application of rnn and prnn for the charge-discharge simulation of advanced
lithium-ions battery energy storage,” in Proceedings of International
Symposium on Power Electronics, Electrical Drives, Automation and
Motion (SPEEDAM). IEEE, 2012, pp. 941–945.
| 9cs.NE
|
From Traces To Proofs: Proving Concurrent
Programs Safe
arXiv:1506.07635v3 [cs.PL] 28 Apr 2016
Chinmay Narayan∗, Subodh Sharma† , Shibashis Guha‡ and S.Arun-Kumar§
Department of Computer Science and Engineering,
Indian Institute of Technology Delhi
Email: ∗ [email protected], † [email protected], ‡ [email protected], § [email protected]
flagi = f alse, flagj = f alse, t = 0;
Abstract—Nondeterminism in scheduling is the cardinal reason
for difficulty in proving correctness of concurrent programs.
A powerful proof strategy was recently proposed [6] to show
the correctness of such programs. The approach captured dataflow dependencies among the instructions of an interleaved and
error-free execution of threads. These data-flow dependencies
were represented by an inductive data-flow graph (iDFG), which,
in a nutshell, denotes a set of executions of the concurrent
program that gave rise to the discovered data-flow dependencies.
The iDFGs were further transformed in to alternative finite
automatons (AFAs) in order to utilize efficient automata-theoretic
tools to solve the problem. In this paper, we give a novel and
efficient algorithm to directly construct AFAs that capture the
data-flow dependencies in a concurrent program execution. We
implemented the algorithm in a tool called ProofTraPar to
prove the correctness of finite state cyclic programs under the
sequentially consistent memory model. Our results are encouranging and compare favorably to existing state-of-the-art tools.
Pi
W hile(true){
1. flagi ∶=true;
2. t∶=j;
3. while(flag j = true & t = j);
4. //Critical Section
5. flagi ∶=false;
}
Pj
W hile(true){
6. flagj ∶=true;
7. t∶=i;
8. while(flagi = true & t = i);
9. //Critical Section
10. flagj ∶=false;
}
Fig. 1: Peterson’s algorithm for two processes Pi and Pj
shown to hold true on unbounded number of traces (trace is “a
sequence of events corresponding to an interleaved execution
of processes in the program”[9]) generated due to unbounded
number of unfoldings of the loops. Notice that events at control
locations 3 and 8 are data-dependent on events from control
locations 2, 6, 7 and 1, 2, 7, respectively. In any finite prefix
of a trace of Pi ∥Pj (interleaved execution of Pi and Pj ) up
to the events corresponding to control location 3 or 8, the
last instance of event at control location 2, lst2, and the last
instance of event at control location 7, lst7, can be ordered
in only one of the following two ways; either lst2 appears
before lst7 or lst2 appears after lst7. This has resulted in
partitioning of an unbounded set of traces to a set with mere
two traces.
When lst2 appears before lst7, then the final value of the
variable t is i, thus making the condition at control location 8
to be true. In the other case, when lst2 appears after lst7,
the final value of the variable t is j, thereby making the
condition at control location 3 evaluate to true. Hence, in
no trace both the conditions are false simultaneously. This
informal reasoning indicates that both processes can never
simultaneously enter in their critical sections. Thus, proof of
correctness for Peterson’s algorithm can be demonstrated by
picking two traces, as mentioned above, from the set of infinite
traces and proving them correct. In general, the intuition is that
a proof for a single trace of a program can result in pruning
of a large set of traces from consideration. To convert this
intuition to a feasible verification method, there is a need to
construct a formal structure from a proof of a trace σ such
that the semantics of this structure includes a set of all those
traces that have proof arguments equivalent to proof of σ.
Inductive Data Flow Graphs (iDFG) was proposed in [6] to
capture data-dependencies among the events of a trace and to
perform trace partitioning. All traces that have the same iDFG
I. I NTRODUCTION
The problem of checking whether or not a correctness
property (specification) is violated by the program (implementation) is already known to be challenging in a sequential
set-up, let alone when programs are implemented exploiting concurrency. The central reason for greater complexity
in verification of concurrent implementations is due to the
exponential increase in the number of executions. A concurrent
program with n threads and k instructions per thread can
have (nk)!/(k!)n executions under a sequentially consistent
(SC)[12] memory model. A common approach to address the
complexity due to the exponential number of executions is
trace partitioning.
In [6], a powerful proof strategy was presented which
utilized the notion of trace partitioning. Let us take Peterson’s
algorithm in Figure 1 to convey the central idea behind the
trace partitioning approach. In this algorithm, two processes,
Pi and Pj , coordinate to achieve an exclusive access to a
critical section (CS) using shared variables. A process Pi will
busy-wait if Pj has expressed interest to enter its CS and t is
j.
In order to prove the mutual exclusion (ME) property of
Peterson’s algorithm, we must consider the boolean conditions
of the while loops at control locations 3 and 8. the ME property
is established only when at most one of these conditions is
false under every execution of the program, i.e., ME must be
1
d
the lines of [6].
While [6] allowed the use of any sequential verification
method to construct a proof of a given trace, the paper
does not comment on the performance and the feasibility
of their approach due to the lack of an implementation.
The second contribution of this paper is an implementation in the form of a tool, ProofTraPar. We compare
our implementation against other state-of-the-art tools
in this domain, such as THREADER [10] and LazyCSeq [11] (winners in the concurrency category of the
software verification competitions held in 2013, 2014, and
2015). ProofTraPar, on average, performed an order
of magnitude better than THREADER and 3 times better
than Lazy-CSeq.
The paper is organized as follows: Section II covers the
notations, definitions and programming model used in this
paper; Section III presents our approach with the help of an
example to convey the overall idea and describes in detail
the algorithms for constructing the proposed alternating finite
automaton along with their correctness proofs. This section
ends with the overall verification algorithm with the proof
of its soundness and completeness for finite state concurrent
programs. Section IV presents the experimental results and
comparison with existing tools namely THREADER [10] and
Lazy-CSeq [11]. Section V presents the related work and
Section VI concludes with possible future directions.
w>3
●
init s0
true
r ∶= w + 1 s1
r>w
a. y ∶= w
b. r ∶= w + 1
c. t ∶= x − 1
(a)
w>3
y ∶= w s2
y>3
t ∶= x − 1 s3
y>3 ∧ t<x
∧ r>w
{abc, bac}
(b)
{abc, bac, acb, cab, bca, cba}
d
(d)
(c)
Fig. 2: Comparison with [6]
must have the same proof of correctness. In every iteration of
their approach, a trace is picked from the set of all traces that
is yet to be covered by the iDFG. An iDFG is constructed
from its proof. The process is repeated until all the traces are
either covered in the iDFG or a counter-example is found.
An intervening step is involved where the iDFG is converted
to an alternating finite automaton (AFA). While we explain
AFA in later sections, it suffices to understand at this stage
that the language accepted by this AFA and the set of traces
captured by the corresponding iDFG is the same. Their reason
for this conversion is to leverage the use of automata-theoretic
operations such as subtraction, complement etc., on the set of
traces.
Though the goal of paper [6] is verification of concurrent
programs which is the same as in this work, our work has some
crucial differences: (i) An AFA is constructed directly from the
proof of a trace without requiring the iDFG construction, (ii)
the verification procedure built on directly constructed AFA is
shown to be sound and complete (weakest-preconditions are
used to obtain the proof of correctness of a trace), (iii) to the
best of our knowledge, we provide the first implementation of
the proof strategy discussed in [6]. The example trace of Figure
2(a) highlights the key difference between iDFG to AFA conversion of [6] and the direct approach presented in this work.
Note that all three events a, b, and c are data independent,
hence every resulting trace after permuting the events in abc
also satisfies the same set of pre- and post-conditions. For a
Hoare triple {w > 3} abc {y > 3 ∧ t < x ∧ r > w} , Figure
2(b) shows the set of traces admitted by an AFA (obtained
from iDFG shown in Figure 2(d)) after the first iteration, as
computed by [6]. This set clearly does not represent every
permutation of abc; consequently, more iterations are required
to converge to an AFA that represents all traces admissible
under the same set of pre- and post-conditions. In contrast,
the AFA that is constructed directly by our approach from the
Hoare triple {w > 3} abc {y > 3 ∧ t < x ∧ r > w} , admits
the set of traces shown in Figure 2(c). Hence, on this example,
our strategy terminates in a single iteration.
To summarize, the contributions of this work are as follows:
●
II. P RELIMINARIES
A. Program Model
We consider shared-memory concurrent programs composed of a fixed number of deterministic sequential processes
and a finite set of shared variables SV. A concurrent program
is a quadruple P = (P, A, I, D) where P is a finite set of
processes, A = {Ap ∣ p ∈ P } is a set of automata, one for each
process specifying their behaviour, D is a finite set of constants
appearing in the syntax of processes and I is a function from
variables to their initial values. Each process p ∈ P has a
disjoint set of local variables LVp . Let Expp (BExpp ) denote
the set of expressions (boolean expressions), ranged over by
exp (φ) and constructed using shared variables, local variables,
D, and standard mathematical operators. Each specification
automaton Ap is a quadruple ⟨Qp , qpinit , δp , Assrnp ⟩ where Qp
is a finite set of control states, qpinit is the initial state, and
Assrnp ⊆ Qp ×BExpp is a relation specifying the assertions that
must hold at some control state. Each transition in δp is of the
form (q, opp , q ′ ) where opp ∈ {x∶=exp, assume(φ), lock(x)}.
Here x∶=exp evaluates exp in the current state and assigns
the value to x where x ∈ SV ∪ LVp . assume(φ) is a blocking
operation that suspends the execution if the boolean expression
φ evaluates to false otherwise it acts as nop. This instruction is
used to encode control path conditions of a program. lock(x),
where x ∈ SV, is a blocking operation that suspends the execution if the value of x is not equal to 0 otherwise it assigns 1 to
x. Operation unlock is achieved by assigning 0 to this shared
variable. Each of these operations are deterministic in nature,
we present a novel algorithm to directly construct an AFA
from a proof of a sequential trace of a finite state (possibly
cyclic) concurrent program. This construction is used to
give a sound and complete verification procedure along
2
p
a
qq
qb
b
qc
e
⎧
⎪
φ
⎪
⎪
⎪
⎪
⎪
⎪
⎪
φ[x/exp]
⎪
⎪
⎪
⎪
⎪
⎪
⎪φ ∧ φ′
wp(op, φ) = ⎨
⎪
⎪
wp(assume(x = 0), wp(x∶=1, φ))
⎪
⎪
⎪
⎪
⎪wp(op , wp(op , φ))
⎪
⎪
1
2
⎪
⎪
⎪ ′
⎪
⎪
⎪
⎩φ ⇒ φ
qp
qa
t
a.flag1 ∶=true
b. turn∶=2
A. assume(¬φ)
B. assume(φ)
c. res∶=1
d. ℓ1 ∶=res
e. flag1 ∶=false
q
B
qr
A
Q
P
qd
qs
c
r
qe
p. flag2 ∶=true
q. turn∶=1
P. assume(¬φ′ )
Q. assume(φ′ )
r. res∶=2
s. ℓ2 ∶=res
t.flag2 ∶=false
qu
def
def
AssrnP1 (qf ) = (ℓ1 = 1), AssrnP2 (qu ) = (ℓ2 = 2)
def
if op = x∶=exp
if op = assert(φ′ )
def
def
if op = lock(x)
def
if op = op1 .op2
if op = assume(φ′ )
def
of the operation op terminates and the resulting program state
s′ satisfies φ.
Given a formula φ, variable X and expression e, let
φ[X/e] denote the formula obtained after substituting all free
occurrences of X by e in φ. We assume an equality operator
over formulae that represents syntactic equality. Every formula
is assumed to be normalized in a conjunctive normal form
(CNF). We use true (false) to syntactically represent a logically
valid (unsatisfiable) formula. Weakest precondition axioms for
different program statements are shown in Figure 4. Here
empty sequence of statements is denote by skip. We have the
following properties about weakest preconditions.
Property 1: If wp(op, φ1 ) = ψ1 and wp(op, φ2 ) = ψ2 then,
● wp(op, φ1 ∧ φ2 ) = ψ1 ∧ ψ2 , and
● wp(op, φ1 ∨ φ2 ) = ψ1 ∨ ψ2 . Note that this property
holds only when S is a deterministic operation which is
true in our programming model.
Property 2: Let φ1 and φ2 be the formulas such that φ1
logically implies φ2 then for every operation op, the formula
wp(op, φ1 ) logically implies wp(op, φ2 ).
We say that a formula φ is stable with respect to a statement
S if wp(S, φ) is logically equivalent to φ. In this paper, we use
weakest preconditions to check the correctness of a trace with
respect to some safety assertion. A trace σ reaching up to a
safety assertion φ is safe if the execution of σ starting from the
initial state I either 1) blocks (does not terminate) because of
not satisfying some path conditions, or 2) terminates and the
resulting state satisfies φ. The following lemmas clearly define
the conditions, using weakest precondition axioms, for declaring a trace σ either safe or unsafe. Detailed proofs of these
are given in Appendix A and in B. Here σ[assume/assert]
denote the trace obtained by replacing every instruction of the
form assume(φ) by assert(φ) in σ.
Lemma 1: For a trace σ, an initial program state I and
a safety property φ, if wp(σ[assume/assert], ¬φ) ∧ I is
unsatisfiable then the execution of σ, starting from I, either
does not terminate or terminates in a state satisfying φ.
Lemma 2: For a trace σ, an initial program state I and
a safety property φ, if wp(σ[assume/assert], ¬φ) ∧ I is
satisfiable then the execution of σ, starting from I, terminates
in a state not satisfying φ.
s
qf
def
Fig. 4: Weakest precondition axioms
qt
d
def
if op = skip
′ def
φ = flag2 = true && turn = 2, φ = flag1 = true && turn = 1
Fig. 3: Specification of Peterson’s algorithm
i.e. execution of any two same operations from the same
states always give the same behaviour. In all examples of this
paper, we use symbolic labels to succinctly represent program
operations. For example, Figure 3 shows the specification of
two processes in Peterson’s algorithm. Labels {a, b, p, q⋯}
denote operations in the program. Variable res is introduced
to specify the mutual exclusion property as a safety property.
A process Pi sets this variable to i inside its critical section.
Assertions assert(res = i) is checked in Pi before leaving its
critical section. If these assertions hold in every execution of
these two processes then the mutual exclusion property holds.
These assertions are shown as AssrnP1 (qf ) and AssrnP2 (qu )
in Figure 3 and they need to be checked at state qf and qu
respectively. A tuple, say t, of n elements can be represented
as a function such that t[k] returns the k th element of this
tuple. Given a function f un, f un[a ← b] denotes another
function same as of f un except at a where it returns b.
a) Parallel Composition in the SC memory model:
Given a concurrent program P = (P, A, I, D) consisting of
n processes P = {p1 , ⋯, pn } we define an automaton A(P) =
(Q, q init , δ, Assrn) to represent the parallel composition of P
in the SC memory model. Here Q = Qp1 × ⋯ × Qpn is the set
of states ranged over by q, q init = (qpinit
, ⋯, qpinit
) is the initial
1
n
state, and transition relation δ models the interleaving semantics. Formally, (q, opj , q ′ ) ∈ δ iff there exists a j ∈ {1⋯n} such
that q[j] = qpj , q ′ = q[j ← qp′ j ] and (qpj , opj , qp′ j ) ∈ δpj . For a
state q, let T (q) = {Assrnpi (q[i]) ∣ i ∈ {1⋯n}}. If T (q) is not
empty then Assrn(q) is the conjunction of assertions in the
set T (q). Relation Assrn captures the assertions which need
to be checked in the interleaved traces of P . As our interest
lies in analyzing those traces which reach those control points
where assertions are specified, we mark all those states where
the relation Assrn is defined as accepting states. Every word
accepted by A(P) represents one SC execution leading to a
control location where at least one assertion is to be checked.
C. Alternating Finite Automata (AFA)
B. Weakest Precondition
Alternating finite automata [1], [3] are a generalization of
nondeterministic finite automata (NFA). An NFA is a five tuple
⟨S, Σ ∪ {ǫ}, δ, s0 , SF ⟩ with a set of states S, ranged over by s,
an initial state s0 , a set of accepting states SF and a transition
function δ ∶ S×Σ∪{ǫ} → P(S). For any state s of this NFA, the
Given an operation op ∈ OP(P ) and a postcondition
formula φ, the weakest precondition of op with respect to φ,
denoted by wp(op, φ), is the weakest formula ψ such that,
starting from any program state s that satisfies ψ, the execution
3
δ(s, op) =
set of words accepted by s is inductively defined as acc(s) =
{a.σ ∣ a ∈ Σ ∪ {ǫ}, σ ∈ Σ∗ , ∃s′ . s′ ∈ δ(s, a). σ ∈ acc(s′ )}
where ǫ ∈ acc(s) for all s ∈ SF . Here, the existential quantifier
represents the fact that there should exist at least one outgoing
transition from s along which a.σ gets accepted. An AFA is
a six tuple ⟨S∀ , S∃ , Σ ∪ {ǫ}, δ, s0 , SF ⟩ with Σ, s0 and SF ⊆ S
denoting the alphabet, initial state and the set of accepting
states respectively. S = S∀ ⋃ S∃ is the set of all states, ranged
over by s and δ ∶ S ×Σ∪{ǫ} → P(S) is the transition function.
The set of words accepted by a state of an AFA depends on
whether that state is an existential state (from the set S∃ ) or
a universal state (from the set S∀ ). For an existential state
s ∈ S∃ , the set of accepted words is inductively defined in
the same way as in NFA. For a universal state s ∈ S∀ the
set of accepted words are acc(s) = {a.σ ∣ a ∈ Σ ∪ {ǫ}, ∀s′ ∈
δ(s, a). σ ∈ acc(s′ )} with ǫ ∈ acc(s) for all s ∈ SF . Notice
the change in the quantifier from ∃ to ∀. In the diagrams of
AFA used in this paper, we annotate universal states with ∀
symbol and existential states with ∃ symbol. For a state s, let
succ(s, a) = {S ∣ (s, a, S) ∈ δ} be the set of a-successors of
s. For an automaton A, let L(A) be the language accepted by
the initial state of that automaton. For any σ ∈ Σ∗ ∣σ∣ denote
the length of σ and rev(σ) denote the reverse of σ.
1.AMap(s′ ) = wp(op[assume/assert], AMap(s)),
2.s is an existential state, and
3.RMap(s) = RMap(s′ ).op.σ ′′
where σ ′′ is the longest sequence s.t.
wp(σ ′′ [assume/assert], AMap(s)) = AMap(s)
(L ITERAL -A SSN)
{s}
if {
{s1 , ⋯, sk }
⎧
⎪
⎪
⎪
⎪
⎪
if ⎨
⎪
⎪
⎪
⎪
⎪
⎩
1.AMap(s) = wp(op[assume/assert], AMap(s)), and
2.s is an existential state
(L ITERAL -S ELF -A SSN)
{}
1.AMap(s) = ⋀k φk or AMap(s) = ⋁k φk ,
2.AMap(sk ) = φk ,
3.∀k, RMap(s) = RMap(sk ),
4.op = ǫ
(C OMPOUND -A SSN)
otherwise
Fig. 5: Transition function used in the Definition 1
1) (OP ǫ = OP ∪ {ǫ}) is the alphabet ranged over by op.
Here OP is the set of instructions used in program P .
Symbol ǫ acts as an identity element of concatenation
and wp(ǫ, φ) = φ.
2) S = S∀ ⋃ S∃ is the largest set of states, ranged over by
s s.t.
a) Every state is annotated with a formula and a prefix
of σ denoted by AMap(s) and RMap(s) respectively.
State s0 is the initial state such that AMap(s0 ) = φ,
RMap(s0 ) = σ.
b) s′ ∈ S iff either of the following two conditions
hold,
● ∃s
∈
S such that AMap(s′ ) is
wp(op[assume/assert], AMap(s)),
RMap(s) = RMap(s′ ).op.σ ′ and σ ′ is the largest
suffix of RMap(s) such that formula AMap(s) is
stable with respect to σ ′ [assume/assert].
● ∃s ∈ S such that AMap(s) = ⋀{φ1 , ⋯, φk } or
AMap(s) = ⋁{φ1 , ⋯, φk }, RMap(s) = RMap(s′ ),
AMap(s′ ) = φ′ and φ′ ∈ {φ1 , ⋯, φk }.
c) A state s ∈ S is an existential state (universal state)
iff AMap(s) is a literal (compound formula).
3) SF ⊆ S is a set of accepting states such that s ∈
SF iff wp(RMap(s)[assume/assert], AMap(s)) is same
as AMap(s), i.e. AMap(s) is stable with respect to
RMap(s)[assume/assert], and
4) Function δ ∶S ×OP ǫ →P(S) is defined in Figure 5.
Following Point 2b, any state added to S is either annotated
with a smaller RMap or a smaller formula compared to the
states already present in S. Further, every formula and trace
σ is of finite length. Hence the set of states S is finite. By
Point 2c of this construction, a state s where AMap(s) is a
compound formula, is always a universal state irrespective of
whether AMap(s) is a conjunction or a disjunction of clauses.
The reason behind this decision will be clear shortly when
we will use this AFA to inductively construct the weakest
precondition wp(σ[assume/assert], φ). Note that we assume
every formula is normalized in CNF.
Figure 7 shows an example trace σ = abApqPrcs of
Peterson’s algorithm. This trace is picked from the Peterson’s
III. O UR A PPROACH
The overall approach of this paper can be described in the
following steps: (i) Given a concurrent program P, construct
all its interleaved traces represented by automaton A(P), as
defined in Subsection II-A; (ii) Pick a trace σ and a safety
property, say φ, to prove for this trace; (iii) Prove σ correct
with respect to φ using Lemma 1 and Lemma 2 and generate
a set of traces which are also provably correct. Let us call this
set T r′ ; (iv) Remove set T r′ from the set of traces represented
by A(P) and repeat from Step (ii) until either all the traces
in P are proved correct or an erroneous trace is found.
Step (iii) of this procedure, correctness of σ, can be achieved
by checking the unsatisfiability of wp(σ[assume/assert],
¬φ) ∧ I. However, we are not only interested in checking the
correctness of σ but also in constructing a set of traces which
have a similar reasoning as of σ. Therefore, instead of computing wp(σ[assume/assert], ¬φ) directly from the weakest
precondition axioms of Figure 4, we construct an AFA from
σ and ¬φ. Step (iv) is then achieved by applying automatatheoretic operations such as complementation and subtraction
on this AFA. Notion of universal and existential states of
AFA helps us in finding a set of sufficient dependencies used
in the weakest precondition computation so that any other
trace satisfying those dependencies gets captured by AFA.
Subsequent subsections covers the construction, properties and
use of this AFA in detail.
A. Constructing the AFA from a Trace and a Formula
Definition 1: An
AFA
constructed
from
trace σ of a Program P and a formula φ
Âσ,φ = ⟨S∀ , S∃ , OPǫ , s0 , SF , δ, AMap, RMap⟩, where,
{s′ }
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
if ⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
a
is
4
Σ ∖ {a, e, A, P}
Σ
abApqPrcs
∃
¬(ℓ2 = 2)
s0
s
Σ ∖ {s, A, P}
true ∃
s2
Σ ∖ {A, P}
abApqPr
P
flag1 = false
ǫ
abApq
flag1 = false
∨ turn = 2
s3
false ∃
s12
flag1 = false
s8
ǫ
Σ ∖ {a, e, A, P}
s4
Σ ∖ {c, r, A, P}
¬(res = 2) ∃
s1
c
abApqPrc
a
∃
abApq
flag1 = false∧
(flag2 = false
∨ turn = 1)
A
s7
Σ
abApq
s5 turn = 2 ∃
Σ ∖ {b, q, A, P}
q
ab
flag2 = false
s10
ǫ
ǫ
flag2 = false
∨ turn = 1
∀
∃
ab
∀
ab
s9
ǫ
ab
∀
ǫ
∃
Σ ∖ {p, t, A, P}
turn = 1 ∃ ab
Σ ∖ {b, q, A, P}
s11
b
false ∃ a
s13
Σ
false ∃ abAp
s6
Fig. 6: AFA of trace given in Figure 6(b) and φ = ¬(ℓ2 = 2)
a. flag1 ∶=true
b. turn∶=2
flag2 = false
A. assume(
)
∣∣ turn = 1
p. flag2 ∶=true
q. turn∶=1
flag1 = false
P. assume(
)
∣∣ turn = 2
r. res∶=2
c. res∶=1
s. ℓ2 ∶=res
Fig. 7: A trace from Peterson’s algorithm
HMap(s) =
specification in Figure 3. To prove σ correct with respect to the
def
safety formula φ = (ℓ2 = 2) we first construct Âσ,¬φ which
will later help us to derive wp(σ[assume/assert], ¬φ). This
AFA is shown in Figure 6. For a state s, AMap(s) is written
inside the rectangle representing that state and RMap(s) is
written inside an ellipse next to that state. We show here some
of the steps illustrating this construction.
1) By Definition 1, we have AMap(s0 ) = ¬(ℓ2 = 2) and
RMap(s0 ) = σ = abApqPrcs for initial state s0 .
2) In a transition δ(s, op) = {s′ } created by Rule
L ITERAL -A SSN the state s′ is annotated with the weakest precondition of an operation op, taken from RMap(s),
with respect to AMap(s). Operation op is picked in
such a way that AMap(s) is stable with respect to
every other operation present after op in RMap(s).
Such transitions capture the inductive construction of
the weakest precondition for a given φ and trace σ.
Transition δ(s0 , s) = {s1 } in Figure 6 is created by this
rule as wp(s[assume/assert], AMap(s0 )) = AMap(s1 ),
and RMap(s0 ) = RMap(s1 ).s.
3) In any transition created by Rule C OMPOUND -A SSN ,
say from s to s1 , ⋯, sk , the states s1 , ⋯sk are annotated
with the subformulae of AMap(s). For example, transitions δ(s3 , ǫ) = {s4 , s5 } and δ(s7 , ǫ) = {s8 , s9 }.
4) Transition δ(s8 , a) = {s12 } follows from the rule
L ITERAL -A SSN. Note that RMap(s12 ) is empty and
hence by Point 3 of Definition 1, s12 is an accepting
state. Following the same reasoning, states s6 , s10 and
s13 are also set as accepting states.
5) Rule L ITERAL -S ELF -A SSN adds a self transition at a
state s on a symbol op ∈ OP ǫ such that AMap(s) is stable with respect to op[assume/assert]. For example,
transitions δ(s0 , op) = {s0 } where op ∈ OP ǫ ∖ {s, A, P}.
The following lemma relates RMap(s) at any state to the set
of words accepted by s in this AFA.
Lemma 3: Given a σ ∈ L(A(P)) and φ, let Âσ,φ be the
AFA satisfying Definition 1. For every state s of this AFA,
the condition rev(RMap(s)) ∈ acc(s) holds.
A detailed proof of this lemma is given in Appendix C. This
lemma uses the reverse of RMap(s) in its statement because
if s ∈ SF
AMap(s)
⋀ HMap(sk )
k
⋁ HMap(sk )
k
HMap(s )
′
(BASE - CASE)
if δ(s, ǫ) = {s1 , ⋯, sk } and AMap(s) = ⋀ AMap(sk )
k
(C ONJ - CASE)
if δ(s, ǫ) = {s1 , ⋯, sk } and AMap(s) = ⋁ AMap(sk )
k
(D ISJ - CASE)
if (s, op, {s′ }) ∈ δ
(L IT- CASE)
Fig. 8: Rules for HMap construction
the weakest precondition of a sequence is constructed by
scanning it from the end. This can be seen in the transition
rule L ITERAL -A SSN. As a corollary, rev(σ) is also accepted
by this AFA because by Definition 1, RMap(s0 ) is σ.
B. Constructing the weakest precondition from Âσ,φ
After constructing Âσ,φ the rules given in Figure 8 are used
to inductively construct and assign a formula, HMap(s), to
every state s of Âσ,φ . Figure 9 shows the AFA of Figure
6 where states are annotated with formula HMap(s). This
formula is shown in the ellipse beside every state. For better
readability we do not show RMap(s) in this figure.
Following Rule BASE - CASE , HMap of s6 , s12 , and s13
are set to false whereas HMap(s10 ) is set to flag2 = false.
By Rule L IT- CASE, HMap of s5 , s8 and s11 are also set to
false. After applying Rule D ISJ - CASE for transition δ(s9 , ǫ) =
{s10 , s11 }, HMap(s9 ) is set to flag2 = false. Similarly, using
Rule C ONJ - CASE we get HMap(s7 ) as false. Finally, HMap(s0 )
is also set to false. HMap constructed inductively in this manner
satisfies the following property;
Lemma 4: Let  be an AFA constructed from a trace and a
post condition as in Definition 1 then for every state s of this
AFA and for every word σ accepted by state s, HMap(s) is logically equivalent to wp(rev(σ)[assume/assert], AMap(s)).
Here we present the proof outline. Detailed proof is given
in Appendix E. First consider the accepting states of Â.
For example, states s6 , s10 , s12 and s13 of Figure 9.
Following the definition of an accepting state and by the
self-loop adding transition rule L ITERAL -S ELF -A SSN, ev5
Σ ∖ {a, e, A, P}
false
Σ
¬(ℓ2 = 2) ∃
s0
s
¬(res = 2)
s1
c
false
true ∃
Σ ∖ {A, P}
flag1 = false
s8
ǫ
s2
∃
Σ ∖ {s, A, P}
false
flag2 = false
s10
ǫ
ǫ
flag2 = false
∨ turn = 1
∀
s7
Σ ∖ {c, r, A, P}
flag1 = false
∨ turn = 2
s3
s4 flag1 = false ∃
false
Σ ∖ {a, e, A, P}
ǫ
Σ
∀
ǫ
Σ ∖ {b, q, A, P}
flag2 = false
∀
flag2 = false
s9
false ∃ false
s13
Σ
false
s5 turn = 2 ∃
ǫ
∃
Σ ∖ {b, q, A, P}
turn = 1 ∃
s11
b
false
A
false
P
∃
flag1 = false∧
(flag2 = false
∨ turn = 1)
false
false
false
a
false ∃
s12
Σ ∖ {p, t, A, P}
q
false ∃ false
s6
Algorithm 1: Converting universal to existential states while
preserving Lemma 4
Data: Input AFA ⟨S∀ , S∃ , OP ∪ {ǫ}, s0 , SF , AMap, RMap⟩
Result: Modified AFA
1 Let s be a state in AFA such that s ∈ S∀ ,
δ(s, ǫ) = {s1 , ⋯, sk }, HMap(s) is unsatisfiable, and
AMap(s) = ⋀k AMap(sk );
2 Let Unsatcore(s) ⊆ P({s1 , ⋯, sk }) such that
{s′1 , ⋯, s′n } ∈ Unsatcore(s) iff {HMap(s′1 ), ⋯, HMap(s′n )}
is a minimal unsat core of ⋀k HMap(sk ) ;
3 Create an empty set U;
′
′
4 foreach {s1 , ⋯, sn } ∈ Unsatcore(s) do
5
create a new universal state su ∈ S∀ and add it to the
set U;
6
Set AMap(su ) = ⋀i AMap(s′i ) ;
7
Set HMap(su ) = ⋀i HMap(s′i );
8
Add a transition by setting δ(su , ǫ) = {s′1 , ⋯, s′n };
9 end
10 Remove transition δ(s, ǫ) = {s1 , ⋯, sk };
11 Convert s to an existential state;
12 Add a transition from s on ǫ by setting δ(s, ǫ) = U where U
is the set of universal states created one for each element
of Unsatcore(s);
Fig. 9: HMap construction for the running example
C. Enlarging the set of words accepted by Âσ,φ
ery word σ accepted by such an accepting state s satisfies
wp(rev(σ)[assume/assert], AMap(s)) = AMap(s). Therefore, setting HMap(s) as AMap(s) for these accepting states, as
done in Rule BASE - CASE completes the proof for accepting
states.
Converting Universal States to Existential States Figure
10 shows an example trace σ = abcde obtained from the
parallel composition of some program P . Figure 11 shows the
AFA constructed for σ and φ as S < t ∧ z < x. From Lemma
4 we get wp(σ, φ) as false. Note that the wp(σ, S < t) and
wp(σ, z < x) are unsatisfiable, i.e. we have two ways to derive
the unsatisfiability of wp(σ, φ); one is due to the operation d,
and the other is due to the operation a followed by operation
e. In this example, any word that enforces either of these
two ways will derive false as the weakest precondition. For
example, the sequence σ ′ = adcbe is not accepted by the
AFA of Figure 11 but the condition wp(rev(σ ′ ), ¬φ) = false
follows from wp(d, ¬φ) = false which is already captured in
the AFA of Figure 11. Note that states s1 and s2 in Figure
11 are annotated with unsatisfiable HMap assertion. It seems
sufficient to take any one of these branches to argue the
unsatisfiability of HMap(s0 ) because HMap(s0 ), by definition,
is a conjunction of HMap(s1 ) and HMap(s2 ). Therefore, if we
convert s0 , a universal state, to an existential state then the
modified AFA will accept adcbe. Let us look at Algorithm 1
to see the steps involved in this transformation. This algorithm
picks a universal state s such that AMap(s) is a conjunction
of clauses and only a subset of its successors are sufficient to
make HMap(s) unsatisfiable. State s0 of Figure 11 is one such
state. For each such minimal subsets of its successors, this
algorithm creates a universal state, as shown in Line 5 of this
algorithm. It is easy to see that HMap(su ) is also unsatisfiable.
Before adding δ(su , ǫ) = {s′1 , ⋯, s′n } transition in AFA this
algorithm sets AMap(su ) as ⋀i AMap(s′i ). By construction,
every word accepted by su must be accepted by s′1 , ⋯, s′n .
Each of these states s′1 , ⋯, s′n satisfy Lemma 4. Hence Lemma
4 continues to hold for these newly created universal states as
well. Now consider a newly created transition (s, ǫ, U) in Line
12. For any state s′′ ∈ U, AMap(s) logically implies AMap(s′′ )
because s′′ represents a subset of the original successors
Now consider a state s with transition δ(s, ǫ) = {s1 , ⋯, sk },
created using Rule C OMPOUND -A SSN , and let σ be a word
accepted by s. By construction, s must be a universal state
and hence σ must be accepted by each of s1 , ⋯, sk as well.
Using this lemma inductively on successor states s1 , ⋯, sk (induction on the formula size) we get wp(σ[assume/assert],
AMap(si )) = HMap(si ) for all i ∈ {1⋯k}. Now we
can apply Property 1 depending on whether AMap(s) is
a conjunction or a disjunction of AMap(sk ). By replacing
AMap(s) with ⋁k AMap(sk )(⋀k AMap(sk )) and HMap(s) with
⋁k HMap(sk )(⋀k HMap(sk )) completes the proof. Note that,
making s as a universal state when AMap(s) is either a
conjunction or a disjunction allowed us to use Property 1 in
this proof. Otherwise, if we make s an existential state when
AMap(s) is a disjunction of formulae then we can not prove
this lemma for states where HMap(s) is constructed using Rule
D ISJ - CASE.
This lemma serves two purposes. First, it checks the correctness of a trace σ w.r.t. a safety property for which this AFA
was constructed. If HMap(s0 ) ∧ I is unsatisfiable, as in our
Peterson’s example trace, then σ is declared as correct. Second,
it guarantees that every trace accepted by this AFA, that is
present in the set of all traces of P, is also safe and hence we
can skip proving their correctness altogether. Removing such
traces is equivalent to subtracting the language of this AFA
from the language representing the set of all traces. Then a
natural question to ask is if we can increase the set of accepted
words of this AFA while preserving Lemma 4.
6
Σ ∖ {a, e, A, P}, P
false
Σ
false
false
false
s1 S < t ∃ d
false ∃
ǫ
s3
a, b, c, e
a, b, c, d, e
false
S<t
∧ z<x
false
ǫ
z < x∃
e
s2
Y∶=x + 1
W∶=t
z∶=W
S∶=t + 1
z∶=Y
a, b, d a
a, b, c, d, e
Fig. 10: Example Trace
false
∀
s0
a.
b.
c.
d.
e.
¬(ℓ2 = 2) ∃
s0
s
Y < x ∃ false
s4
b, c, d, e
true ∃
s2
Σ ∖ {A, P}, A
false
false ∃
false
¬(res = 2) ∃
s1
c
P
Σ ∖ {c, r, A, P}
Fig. 11: AFA for σ given in Figure 10
δ(s, op) = δ(s, op) ∪ {s′ } iff
s4 flag1 = false ∃
ǫ
false
s3
false
s5
false ∃
s12
ǫ
Σ ∖ {a, e, A, P}
Σ ∖ {s, A, P}
flag1 = false
∨ turn = 2
a
false
false ∃
s6
∃
flag1 = false
su
ǫ
∀
flag1 = false∧
(flag2 = false
∨ turn = 1)
∀
Σ
A
flag1 = false
s8
ǫ
ǫ
q
false
s7
false
false
∃
false
turn = 2 ∃
s5
Σ ∖ {b, q, A, P}
Fig. 12: AFA of Figure 9 after Modification
Algorithm 2: Algorithm to check the safety assertions of
a concurrent program P
Input: A concurrent program P = {p1 , ⋯, pn } with safety
property map Assrn
Result: yes, if program is safe else a counterexample
1 Let A(P) bet the automaton that represents the set of all
the SC executions of P (as defined in Section II);
̂));
2 Set tmp ∶= L(A(P
3 while tmp is not empty do
4
Let σ ∈ tmp with φ as a safety assertion to be
checked;
5
Let Âσ,¬φ be the AFA constructed from σ and ¬φ ;
6
if I ∧ HMap(s0 ) is satisfiable then
7
σ is a valid counterexample violating φ;
8
return (σ);
9
else
10
Let Â′ be the AFA modified by proposed
transformations;
11
tmp ∶= tmp ∖ Rev, where
Rev = {rev(σ) ∣ σ ∈ L(Â′ )};
12
end
13 end
14 return (yes);
HMap(s) and HMap(s′ ) are unsatisfiable,
(s) is a literal, and
op[assume/assert]AMap(s) ⇒ AMap(s′ )
(RULE -U NSAT)
OR
HMap(s) and HMap(s′ ) are valid
(s) is a literal, and
(s′ ) ⇒ wp(op[assume/assert], AMap(s))
(RULE -VALID)
Fig. 13: Rules for adding more edges
of s, viz. s1 , ⋯, sk . As s is now an existential state, any
word accepted by s, say σ ′ , is accepted by at least one
state in U, say s′ . Using Lemma 4 on s′ , HMap(s′ ) is logically equivalent to wp(rev(σ ′ )[assume/assert], AMap(s′ )).
Using unsatisfiability of HMap(s) and HMap(s′ ) and the
monotonicity property of the weakest precondition, Property 2, we get that HMap(s) is logically equivalent to
wp(rev(σ ′ )[assume/assert], AMap(s)). This transformation
is formally proved correct in Appendix E.
Adding More transitions to Âσ,φ using the Monotonicity
Property of the Weakest Precondition We further modify
Âσ,φ by adding more transitions. For any two states s and
s′ such that AMap(s) and AMap(s′ ) are literals, both HMap(s)
and HMap(s′ ) are unsatisfiable, and there exists a symbol a
(can be ǫ as well) such that wp(a[assume/assert], AMap(s))
logically implies AMap(s′ ), an edge labeled a is added from s
to s′ . This transformation also preserves Lemma 4 following
the same monotonicity property, Property 2 used in the previous transformation. Similar argument holds when HMap(s)
and HMap(s′ ) are valid and AMap(s′ ) ⇒ wp(a, AMap(s)) holds.
The rules of adding edges are shown in Figure 13.
Figure 12 shows the AFA of Figure 9 modified by above
transformations. Rule RULE -U NSAT adds an edge from s4
to s8 on symbol ǫ because HMap(s4 ) and HMap(s8 ) are
unsatisfiable and wp(ǫ, AMap(s4 )) logically implies AMap(s8 ).
Same rule also adds a self loop at s8 on operation P and a
self loop at s2 on operation A. Transformation by Algorithm
1 removes the transition from s7 to s9 and all other states
reachable from s9 . Now consider a trace rev(abpqPArcs)
that is accepted by this modified AFA in Figure 12 but was
not accepted by the original AFA of Figure 9. Note that
wp(abpqPArcs, ¬(ℓ2 = 2)) is unsatisfiable and this is a direct
consequence of Lemma 4. Because of the transformations
presented in this sub-section we do not need to reason about
this trace separately. This transformation is formally proved
correct in Appendix F.
D. Putting All Things Together For Safety Verification
In Algorithm 2, all the above steps are combined to check
if all the SC executions of a concurrent program P satisfy the
safety properties specified as assertions. Proof of the following
theorem is given in Appendix G.
Theorem 1: Let P = (p1 , ⋯, pn ) be a finite state program (with or without loops) with associated assertion maps
Assrnpi . All assertions of this program hold iff Algorithm 2
returns yes. If the algorithm returns a word σ then at least
one assertion fails in the execution of σ.
7
Program
Peterson.safe
Dekker.safe
Lamport.safe
Szymanksi.safe
TimeVarMutex.safe
RWLock.safe (2R+2W)
RWLock.unsafe (2R+2W)
Qrcu.safe (2R+1W)
Qrcu.unsafe (2R+1W)
ProofTraPar
THREADER[10]
Lazy-CSeq[11]
0.3
1.1
2.4
3
0.76
8.8
3.8
20
13.8
3.2
1.7
47
12.8
8.56
140
153
–
76
3.1
4.2
5.1
4
4.2
6.7
0.7
41
1.1
handle larger number of interleavings. These optimizations
also selectively check a representative set of traces among the
set of all interleavings. POR based methods were traditionally
used in bug finding but recently they have been extended
efficiently, using abstraction and interpolants, for proving
programs correct [14]. The technique presented in this paper,
using AFA, can possibly be used to keep track of partial orders
in POR based methods. In [15], a formalism called concurrent
trace program (CTP) is defined to capture a set of interleavings
corresponding to a concurrent trace. CTP captures the partial
orders encoded in that trace. Corresponding to a CTP, a
formula φctp is defined such that φctp is satisfiable iff there is
a feasible linearization of the partial orders encoded in CTP
that violates the given property. Our AFA is also constructed
from a trace but unlike CTP it only captures those different
interleavings which guarantee the same proof outline. Recently
in [9], a formalism called HB-formula has been proposed
to capture the set of happens-before relations in a set of
executions. This relation is then used for multiple tasks such as
synchronization synthesis[2], bug summarization and predicate
refinement. Since the AFA constructed by our algorithm can
also be represented as a boolean formula (universal states
correspond to conjunction and existential states correspond
to disjunction) that encodes the ordering relations among the
participating events, it will be interesting to explore other
usages of this AFA along the lines of [9].
Fig. 14: Comparison with THREADER[10], and Lazy-CSeq [11] (Time in seconds)
IV. E XPERIMENTAL E VALUATION
We implemented our approach in a prototype tool,
ProofTraPar. This tool reads the input program written
in a custom format. In future, we plan to use off-the-shelf
parsers such as CIL or LLVM to remove this dependency.
Individual processes are represented using finite state automata. We use an automata library, libFAUDES [5] to carry
out operations on automata. As this library does not provide
operations on AFA, mainly complementation and intersection,
we implemented them in our tool. After constructing the AFA
from a trace we first remove ǫ transitions from this AFA.
This is followed by adding additional edges in AFA using
proposed transformations. Instead of reversing this AFA (as
in Line 11 of Algorithm 2) we subtract it with an NFA that
represents the reversed language of the set of all traces. This
avoids the need of reversing an AFA. Note that we do not
convert our AFA to NFA but rather carry out intersection and
complementation operations (needed for language subtraction
operation) directly on AFA. Our tool uses the Z3 [4] theorem
prover to check the validity of formulae during AFA construction. ProofTraPar can be accessed from the repository
https://github.com/chinuhub/ProofTraPar.git.
Figure 14 tabulates the result of verifying pthreadatomic category of SV-COMP benchmarks using our tool,
THREADER [10] and Lazy-CSeq [11]. These tools were
the winners in the concurrency category of the software
verification competition of 2013 (THREADER), 2014 and
2015 (Lazy-CSeq). Dash (–) denotes that the tool did not
finish the analysis within 15 minutes. Numbers in bold text
denote the best time of that experiment. Safe/Unsafe versions
of these programs are labeled with .safe/.unsafe. Except on
Reader-Writer Lock and on unsafe version of QRCU(Quick
Read Copy Update), our tool performed better than the other
two tools. On unsafe versions, our approach took more time
to find out an erroneous trace as compared to Lazy-CSeq
[11]. Context-bounded exploration by Lazy-CSeq [11] and the
presence of bugs at a shallow depth seem to be a possible reason behind this performance difference. Introducing priorities
while picking traces in order to make our approach efficient
in bug-finding is left open for future work.
VI. C ONCLUSION
AND
F UTURE W ORK
We presented a trace partitioning based approach for verifying safety properties of a concurrent program. To this
end, we introduced a novel construction of an alternating
finite automaton to capture the proof of correctness of a
trace in a program. We also presented an implementation of
our algorithm which compared competitively with existing
state-of-the-art tools. We plan to extend this approach for
parameterized programs and programs under relaxed memory
models. We also plan to investigate the use of interpolants
with weakest precondition axioms to incorporate abstraction
for handling infinite state programs.
R EFERENCES
[1] J.A. Brzozowski and L. L. Ernst. On equations for regular languages,
finite automata, and sequential networks. TCS, 10:19–35, 1980.
[2] P. Cerný, E.M. Clarke, T. A. Henzinger, A. Radhakrishna, L. Ryzhyk,
R. Samanta, and T. Tarrach. From non-preemptive to preemptive
scheduling using synchronization synthesis. In CAV, 2015.
[3] A. K. Chandra, D. C. Kozen, and L. J. Stockmeyer. Alternation. J.
ACM, 28(1):114–133, January 1981.
[4] L. De Moura and N. Bjørner. Z3: An efficient SMT solver. In TACAS,
pages 337–340. Springer-Verlag, 2008.
[5] Bernd Opitz et al. libfaudes-discrete event system library.
[6] A. Farzan, Z. Kincaid, and A. Podelski. Inductive data flow graphs. In
POPL, pages 129–142, 2013.
[7] C. Flanagan and P. Godefroid. Dynamic partial-order reduction for
model checking software. In POPL, pages 110–121, 2005.
[8] P. Godefroid. Partial-Order Methods for the Verification of Concurrent
Systems: An Approach to the State-Explosion Problem. Springer, 1996.
[9] A. Gupta, T. A. Henzinger, A. Radhakrishna, R. Samanta, and T. Tarrach.
Succinct representation of concurrent trace sets. In POPL, 2015.
[10] A. Gupta, C. Popeea, and A. Rybalchenko. Threader: A constraint-based
verifier for multi-threaded programs. In CAV, pages 412–417, 2011.
V. R ELATED W ORK
Verifying the safety properties of a concurrent program
is a well studied area. Automated verification tools which
use model checking based approaches employ optimizations
such as Partial Order Reductions (POR) [13], [8], [7] to
8
A PPENDIX
[11] O. Inverso, E. Tomasco, B. Fischer, S. La Torre, and G. Parlato. Bounded
model checking of multi-threaded C programs via lazy sequentialization.
In CAV, volume 8559 of LNCS, pages 585–602. Springer, 2014.
[12] L. Lamport. How to make a multiprocessor computer that correctly
executes multiprocess programs. IEEE Trans. Comput., September 1979.
[13] D. Peled. All from one, one for all: On model checking using
representatives. In CAV, pages 409–423. Springer-Verlag, 1993.
[14] B. Wachter, D. Kroening, and J. Ouaknine. Verifying multi-threaded
software with Impact. In FMCAD, pages 210–217. IEEE, 2013.
[15] C. Wang, S. Kundu, M. Ganai, and A. Gupta. Symbolic predictive
analysis for concurrent programs. In Ana Cavalcanti and DennisR.
Dams, editors, FM 2009: Formal Methods, volume 5850 of LNCS, pages
256–272. Springer Berlin Heidelberg, 2009.
A. Proof of Lemma 1
We prove it by induction on n.
1) Base case ∣σ∣ = 0: If ∣σ∣ = 0 then wp(σ[assume/assert],
¬φ) = ¬φ. If ¬φ ∧ I is unsatisfiable then I satisfies φ.
Hence proved.
2) Induction step, ∣σ∣ = n + 1: Let σ = σ ′ .a. If
wp(σ ′ .a[assume/assert], ¬φ) ∧ I is unsatisfiable then
following cases can happen based on a.
′
● a ∶ x ∶= E:- If wp(σ .a[assume/assert], ¬φ) ∧ I
is unsatisfiable then wp(σ ′ [assume/assert],
wp(a, ¬φ)) ∧ I is also unsatisfiable. By
substituting wp(a, ¬φ) with ¬φ[E/x] we get
that wp(σ ′ [assume/assert], ¬φ[E/x]) ∧ I is
unsatisfiable. Using IH on σ ′ it implies that after
executing σ ′ from I the resultant state either does
not terminate or terminates in a state satisfying
φ[E/x]. If σ ′ does not terminate then so does the
execuction of σ starting from I. If σ ′ terminates in
a state satisfying φ[E/x] then by the definition of
the weakest precondition, execution of a from this
state will satisfy φ. Hence proved.
● a
∶ assume(φ′ ):-If wp(σ ′ .a[assume/assert],
¬φ)
∧
I
is
unsatisfiable
then
wp(σ ′ [assume/assert], wp(a, ¬φ))
∧
I is
also unsatisfiable. By substituting wp(a, ¬φ) with
φ′ ∧ ¬φ we get that wp(σ ′ [assume/assert],
φ′ ∧ ¬φ) ∧ I is unsatisfiable. Using IH on σ ′ it
implies that after executing σ ′ from I the resultant
state either does not terminate or terminates in a
state satisfying ¬φ ∨ φ′ . If σ ′ does not terminate
then the execution of σ from I does not terminate
as well. If σ ′ terminates in a state satisfying ¬φ
then the execution of a blocks and hence the
execution of σ does not terminate. If σ ′ terminates
in a state satisfying φ′ but ¬φ does not hold then
φ ∧ φ′ must hold. Execution of assume(φ′ ) acts as
nop instruction and the resultant state satisfies φ.
hence proved.
● a ∶ lock(x):- As weakest precondition of lock(x)
is obtained from the weakest precondition of assignment and assume instruction hence the similar
reasoning works for this case.
B. Proof of Lemma 2
Proof: Let us prove it by induction on the length of σ.
1) Base case, ∣σ∣ = 0: When the length of σ is 0 and I ∧ ¬φ
is satisfiable then I does not satisfy φ. Hence proved.
2) Induction Step, ∣σ∣ = n + 1: Let σ = σ ′ .a. Following case
can happen based on the type of a.
● a
∶ x ∶= E:- If wp(σ[assume/assert],
¬φ) ∧ I is satisfiable then wp(σ ′ [assume/assert],
wp(a, ¬φ)) ∧ I is also satisfiable. By substituting wp(a, ¬φ) = ¬φ[E/x] we get that
1
●
●
wp(σ ′ [assume/assert], ¬φ[E/x]) ∧ I is satisfiable. By IH on σ ′ , execution of σ ′ from I terminates
in a state not satisfying φ[E/x]. By definition of
the weakest precondition, the state reached after
executing a from this state does not satisfy φ. Hence
proved.
a
∶
assume(φ′ ):-If wp(σ[assume/assert],
¬φ) ∧ I is satisfiable then wp(σ ′ [assume/assert],
wp(assume(φ′ )[assume/assert], ¬φ))
∧
I
is
also
satisfiable.
By
substituting
wp(assume(φ′ )[assume/assert], ¬φ) = φ′ ∧ ¬φ
we
get
that
wp(σ ′ [assume/assert],
′
¬(¬φ ∨ φ)) ∧ I is satisfiable. By IH on
σ ′ , execution of σ ′ from I terminates in a state
not satisfying ¬φ′ ∨ φ. In other words, φ′ and
¬φ holds in the state reached after executing σ ′
from I. Therefore, after executing assume(φ′ ), the
resultant state satisfies ¬φ and hence proved.
a ∶ lock(x):-Similar to the combination of above
two cases.
an accepting state. It should have a successor state
s′ such that (s, op, {s′ }) is a transition. By transition rule L ITERAL -A SSN RMap(s) = RMap(s′ ).op.σ ′′
such that wp(σ ′′ [assume/assert], AMap(s)) =
AMap(s). By transition rule L ITERAL -S ELF -A SSN,
s will have self loop transitions on all symbols in
σ ′′ (*). Applying IH on s′ gives that rev(RMap(s′ )) ∈
acc(s′ )(#). Because of the transition (s, op, {s′ }),
op.acc(s′ ) ⊆ acc(s). This along with (#) gives
us op.rev(RMap(s′ )) ∈ acc(s)(**). Rearranging this
and using (*) we get rev(RMap(s′ ).op.σ ′′ ) ∈ acc(s)
or equivalently rev(RMap(s)) ∈ acc(s). Hence
proved.
D. Proof of Lemma 4
Proof: We use induction for this proof. Same as in
the previous proof, let us use the following ordering on
the states of Â. For any two states s and s′ , s < s′ if
∣RMap(s)∣ < ∣RMap(s′ )∣ or if lengths are same then AMap(s)
is a sub formula of AMap(s′ ). Any two states which are not
related by this order, put them in any order to make < as a total
order. It is clear that the smallest state in this total order must
be one of the accepting state. Now we are ready to proceed
by induction using this total order.
● Base case, By definition of the accepting state in AFA
construction, Point 3 of Definition 1, and the self loop
transition rule, Rule L ITERAL -S ELF -A SSN, we know that
for every word σ ′ ∈ acc(s), wp(σ ′ [assume/assert],
AMap(s)) = AMap(s). Rule BASE - CASE of Figure 8 sets
HMap(s) same as AMap(s) for such states hence the
statement of this lemma follows for the accepting states.
● Induction step; we pick a state s such that one of the
following holds,
1) s is a universal state;By construction, there should
be states s1 , ⋯, sk such that (s, ǫ, {s1 , ⋯, sk }) is
a transition. Let σ be a word accepted by s then
by the definition of accepting set of words of a
universal states, σ must be accepted by each of
s1 , ⋯sk . By our induction ordering, s1 , ⋯, sk are
smaller than s and hence we apply IH on them to
get that wp(rev(σ)[assume/assert], AMap(si )) =
HMap(si ) for i ∈ {1⋯k}. Two cases arise based on
whether
– AMap(s) is a conjunction of AMap(si ) for
i ∈ {1⋯k}; Following Rule C ONJ - CASE
we set HMap(s)
=
⋀i HMap(si ) and
wp(rev(σ)[assume/assert], AMap(s))
=
HMap(s) then follows from the Property 1, using
conjunction, of the weakest precondition.
– AMap(s) is a disjunction of AMap(si ) for
i ∈ {1⋯k}; Following Rule C ONJ - CASE
we set HMap(s)
=
⋀i HMap(si ) and
wp(rev(σ)[assume/assert], AMap(s))
=
HMap(s) then follows from the Property 1, using
disjunction, of the weakest precondition.
C. Proof of Lemma 3
Proof: We use induction for this proof. Let us use the
following ordering on the states of Âσ φ. For any two states s
and s′ , s < s′ if ∣RMap(s)∣ < ∣RMap(s′ )∣ or if lengths are same
then AMap(s) is a sub formula of AMap(s′ ). Any two states
which are not related by this order, put them in any order to
make < as a total order. It is clear that the smallest state in
this total order must be one of the accepting state. Now we
are ready to proceed by induction using this total order.
● Base case; For every accepting state s ∈ SF , by Point 3
of Definition 1, the condition wp(op, AMap(s)) = AMap(s)
holds for every op ∈ EL(RMap(s)). Further, By transition
rule L ITERAL -S ELF -A SSN of this AFA, a self transition
must be there for all such op ∈ EL(RMap(s)) and hence
the condition rev(RMap(s)) ∈ acc(s) holds (because
these transitions can be taken in any order to construct
the required word).
● Induction step; Following possibilities exist for the state
s,
– s is a universal state; By construction, there should
be states s1 , ⋯, sk such that (s, ǫ, {s1 , ⋯, sk }) is a
transition. By our induction ordering, s1 , ⋯, sk are
smaller than s and hence we apply IH on them to
get that rev(RMap(si )) ∈ acc(si ) for i ∈ {1⋯k}.
However, by the transition rule C OMPOUND -A SSN ,
RMap(s) = RMap(s1 ) = ⋯ = RMap(sk ) and hence
rev(RMap(s) ∈ acc(si ) for i ∈ {1⋯k}. By the
definition of acc(s) for a universal state, acc(s) is
intersection of the sets acc(si ) for i ∈ {1⋯k} and
hence we get the required result, viz. rev(RMap(s)) ∈
acc(s).
– s is an existential state; If s is an accepting state then
Base case holds here. Consider the case when s is not
2
2) s is an existential state; If s is an accepting state then
the same argument as used in the Base case holds.
If s is not an accepting state then the only outgoing
transition from s is of the form (s, op, {s′ }), By
rule L ITERAL -A SSN(*). Now consider a word
σ ∈ acc(s). σ must be of the form σ ′′ .op.σ ′ where
wp(σ ′′ [assume/assert], AMap(s)) = AMap(s)(*)
(because of the self transitions constructed from
Rule L ITERAL -S ELF -A SSN) and σ ′ ∈ acc(s′ ).
Therefore, wp(rev(σ)[assume/assert], AMap(s))
=
=wp(rev(σ ′′ .op.σ ′ )[assume/assert], AMap(s))
=wp(rev(σ ′ ).op.rev(σ ′′ )[assume/assert],
AMap(s))
=wp(rev(σ ′ ).op[assume/assert], AMap(s)) (using
(*))
=wp(rev(σ ′ )[assume/assert],
wp(op[assume/assert], AMap(s))) (using weakest
precondition definition)
=wp(rev(σ ′ )[assume/assert], AMap(s′ )) (using
Transition rule L ITERAL -A SSN)
As σ ′ ∈ acc(s′ ) this is same as HMap(s′ ) by
applying IH on s′ . As HMap(s) is same as
HMap(s′ ), as done in Rule L IT- CASE, we prove
this case as well.
accepted by state s, HMap(s) is logically equivalent to
wp(rev(σ)[assume/assert], AMap(s)).
Proof: As a result of adding edges in this transformation,
we can not use the ordering among states as done for earlier
proofs. This is because, now a transition (s, op, S) does not
guarantee that the states in the set S are smaller then s and
hence it will not be possible to apply IH directly. Therefore
in this proof we apply induction on the length of σ ′ accepted
by some state s.
● Induction step; Let s ∈ Â and σ ∈ acc(s) such that
∣σ∣ = m + 1. Either s ∈ S∃ or s ∈ S∀ . If s ∈ S∃
and σ ∈ acc(s) then there exists a state s′ such that
(s, op, {s′ }) ∈ δ and σ ′ ∈ acc(s′ ), where σ = σ ′′ .op.σ ′
and wp(σ ′′ [assume/assert], AMap(s)) = AMap(s)(**).
Based on this transition (s, op, {s′ }) ∈ δ we have the
following sub-cases,
– (s, op, {s′ }) was added by the this transformation
virtue of one of the following conditions,
∗ HMap(s) and HMap(s′ ) are unsatisfiable and
wp(op[assume/assert], AMap(s)) ⇒ AMap(s′ )
(Rule RULE -U NSAT); By IH on σ ′ we have
wp(rev(σ ′ )[assume/assert], AMap(s′ ))
is
logically
equivalent
to
HMap(s′ ).
Using Property 2 (conjunction part) and
the
assumption
wp(op[assume/assert],
AMap(s))
⇒
AMap(s′ )
we
get
′
wp(rev(σ )[assume/assert], wp(op, AMap(s)))
is unsatisfiable and same as HMap(s).
Using
(**),
wp(rev(σ ′ )[assume/assert],
′′
wp(op.rev(σ ), AMap(s))) is unsatisfiable and
same as HMap(s). By replacing σ = σ ′′ .op.σ ′ we
get the required proof.
∗ HMap(s) and HMap(s′ ) are valid and
AMap(s′ ) ⇒ wp(op[assume/assert], AMap(s))
(Rule RULE -VALID); By IH on σ ′ we have
wp(rev(σ ′ )[assume/assert], AMap(s′ ))
is
logically equivalent to HMap(s′ ). Using property
2 (disjunction part) and the assumption
AMap(s′ )
⇒
wp(op[assume/assert],
AMap(s)) we get wp(rev[assume/assert](σ ′ ),
wp(op[assume/assert], AMap(s))) is valid and
same as HMap(s). Using (**) and or replacing
σ = σ ′′ .op.σ ′ we get the required result and hence
proved.
– If this transition was already in δ; we can use the
same reasoning as used in the proof of Lemma 4 to
show that wp(rev(σ)[assume/assert], AMap(s)) is
logically equivalent to HMap(s)
● If s ∈ S∀ then similar argument goes as in the proof
of Lemma 4 because no new transition gets added from
these states as a result of this transformation.
E. Proof of Correctness of Transformation-I
Lemma 5: Let  be an automaton constructed from a trace
and a post condition as defined in Definition 1 and further
modified by Algorithm 1 then for every state s of this AFA
and for every word σ accepted by state s, HMap(s) is logically
equivalent to wp(rev(σ)[assume/assert], AMap(s)).
Proof: Proof of this lemma is very similar to the proof
of Lemma 4 given in Appendix . Here we only highlight the
changes in the proof. Note that this transformation converts
some universal states to existential states. Let s be one such
state that was converted from universal to existential state. Let
(s, ǫ, {s1 , ⋯, sk }) was the original transition in the AFA which
got modified to (s, ǫ, {su1 , ⋯, sun } where sui are newly created universal states in Line 5 of Algorithm 1. By construction,
HMap(sui ) is unsatisfiable for each of these su1 , ⋯, sun (*). Let
σ be a word accepted by s after converting it to existential
state. By acceptance conditions, σ must be accepted by at least
one state, say sum in the set {su1 , ⋯, sun }. By IH on sum we
get wp(σ[assume/assert], AMap(sum )) = HMap(sum )(**).
Further, by construction AMap(s) implies AMap(sum ). This
fact, along with the monotonicity property of the weakest
precondition, Property 2, we get that wp(σ[assume/assert],
AMap(s)) is unsatisfiable and hence same as HMap(s).
F. Proof of Correctness of Transformation-II
Lemma 6: Let  be an
trace and a post condition
further modified by adding
for every state s of this
automaton constructed from a
as defined in Definition 1 and
edges as discussed above then
AFA and for every word σ
G. Proof of Theorem 1
Proof:
3
●
●
Let us first prove that this algorithm terminates for finite
state programs. For finite state programs the number of
possible assertions used in the construction of AFA are
finite and hence only a finite number of different AFA
are possible. It implies the termination of this algorithm.
Following Lemma 4 and the fact that AMap(s0 ) = ¬φ,
every word σ ′ accepted by this AFA, equivalently written
as σ ′ ∈ acc(s0 ), satisfies wp(rev(σ ′ )[assume/assert],
¬φ) = HMap(s0 )(*). By Lemma 3 and the fact that
RMap(s0 ) = σ we get rev(σ) ∈ acc(s0 )(**). Combining
(**) and (*), we get wp(rev(rev(σ))[assume/assert],
¬φ) = HMap(s0 ) or equivalently wp(σ[assume/assert],
¬φ) = HMap(s0 ).
– If I ∧ HMap(s0 ) is satisfiable (Line 6) then
I ∧ wp(σ[assume/assert], ¬φ) is satisfiable as
well. Following Lemma 2 we got a valid error trace
which is returned in Line 8.
– If I ∧ HMap(s0 ) is unsatisfiable then by Lemma 1
this trace is provably correct. Now we apply transformations of Section III-C on the AFA to increase the
set of words accepted by it. The final AFA is then
reversed and subtracted from the set of executions
seen so far. Lemma 4 ensures that for all such
words σ ′ the condition I ⇏ wp(σ ′ , ¬φ) holds and
therefore none of them violate φ starting from the
initial state. Therefore in every iteration only correct
set of executions are being removed from the set of
all executions. Therefore when this loop terminates
then all the executions have been proved as correct.
4
| 6cs.PL
|
NON-ASYMPTOTIC RESULTS FOR CORNISH-FISHER
EXPANSIONS
arXiv:1604.00539v1 [math.ST] 2 Apr 2016
V.V. ULYANOV, M. AOSHIMA, AND Y. FUJIKOSHI
Abstract. We get the computable error bounds for generalized
Cornish-Fisher expansions for quantiles of statistics provided that
the computable error bounds for Edgeworth-Chebyshev type expansions for distributions of these statistics are known. The results
are illustrated by examples.
1. Introduction and main results
In statistical inference it is of fundamental importance to obtain
the sampling distribution of statistics. However, we often encounter
situations, where the exact distribution cannot be obtained in closed
form, or even if it is obtained, it might be of little use because of its
complexity. One practical way of getting around the problem is to
provide reasonable approximations of the distribution function and its
quantiles, along with extra information on their possible errors. It can
be made with help of Edgeworth–Chebyshev and Cornish–Fisher type
expansions. Recently the interest for Cornish–Fisher type expansions
stirred up because of intensive study of VaR (Value at Risk) models
in financial mathematics and financial risk management (see, e.g. [14]
and [15]).
Mainly, it is studied the asymptotic behavior of the expansions mentioned above. It means that accuracy of approximation for distribution
of statistics or its quantiles is given as O(·), that is in the form of order
with respect to some parameter(s) (usually, w.r.t. n as a number of
observations and/or p as dimension of observations). In this paper we
construct non-asymptotic error bounds, in other words – computable
error bounds, for Cornish–Fisher type expansions, that is for an error of approximation we prove upper bounds with closed-form dependence on n and/or p and, perhaps, on some moment characteristics of
observations. We get these bounds under condition that similar nonasymptotic results are already known for accuracy of approximation of
distributions of statistics by Edgeworth–Chebyshev type expansions.
Let X be a univariate random variable with a continuous distribution
function F . For α : 0 < α < 1, there exists x such that F (x) = α,
Key words and phrases. computable bounds, non-asymptotic results, CornishFisher expansions.
This work was supported by RSCF grant No. 141100364.
1
2
V.V. ULYANOV, M. AOSHIMA, AND Y. FUJIKOSHI
which is called the (lower) 100α% point of F . If F is strictly increasing,
the inverse function F −1 (·) is well defined and the 100α% point is
uniquely determined. We also speak of “quantiles” without reference
to particular values of α meaning the values given by F −1 (·). Even
in the general case, when F (x) is not necessarily continuous nor is it
strictly increasing, we can define its inverse function by formula
F −1 (u) = inf{x; F (x) > u}.
This is a right-continuous nondecreasing function defined on the interval (0, 1) and F (x0 ) ≥ u0 if x0 = F −1 (u0 ).
Let Fn (x) be a sequence of distribution functions and let each Fn
admit the Edgeworth-Chebyshev type expansion (ECE) in the powers
of ǫ = n−1/2 or n−1 :
(1)
Fn (x) = Gk,n (x) + Rk (x) with Rk (x) = O(ǫk ) and
Gk,n (x) = G(x) + ǫa1 (x) + · · · + ǫk−1 ak−1 (x) g(x),
where g(x) is a density function of the limiting distribution function
G(x). An important approach to the problem of approximating the
quantiles of Fn is to use their asymptotic relation to those of G’s. Let
x and u be the corresponding quantiles of Fn and G, respectively. Then
we have
(2)
Fn (x) = G(u).
Write x(u) and u(x) to denote the solutions of (2) for x in terms of u
and u in terms of x, respectively [i.e. u(x) = G−1 (Fn (x)) and x(u) =
Fn−1 (G(u))]. Then we can use the ECE (1) to obtain formal solutions
x(u) and u(x) in the form
(3)
and
x(u) = u + ǫb1 (u) + ǫ2 b2 (u) + · · ·
(4)
u(x) = x + ǫc1 (x) + ǫ2 c2 (x) + · · · .
Cornish and Fisher in [3] and [6] obtained the first few terms of these
expansions when G is the standard normal distribution function (i.e.,
G = Φ). Both (3) and (4) are called the Cornish–Fisher expansions,
(CFE). Concerning CFE for random variables obeying limit laws from
the family of Pearson distributions see, e.g., [1]. Hill and Davis in [13]
gave a general algorithm for obtaining each term of CFE when G is an
analytic function.
Usually the CFE are applied in the following form with k = 1, 2 or
3:
(5)
xk (u) = u +
k−1
X
j=1
ǫj bj (u) + R̂k (u) with R̂k (u) = O(ǫk ).
NON-ASYMPTOTIC RESULTS FOR CORNISH-FISHER EXPANSIONS
3
It is known (see, e.g., [15]) how to find the explicit expressions for b1 (u)
and b2 (u) as soon as we have (1). By Taylor’s expansions for G, g, and
a1 , we obtain
b1 = −a1 (u),
(6)
1
b2 = {g ′(u)/g(u)}a21(u) − a2 (u) + a′1 (u)a1 (u),
2
provided that g and a1 are smooth enough functions.
In the following Theorems we show how xk (u) from (5) could be
expressed in terms of u. Moreover, we show what kind of bounds we
can get for R̂k (x) as soon as we have some bounds for Rk (x) from (1).
Theorem 1. Suppose that for the distribution function of a statistic
U we have
(7)
F (x) ≡ Pr{U ≤ x} = G(x) + R1 (x),
where for remainder term R1 (x) there exists a constant c1 such that
|R1 (x)| ≤ d1 ≡ c1 ǫ.
Let xα and uα be the upper 100α% points of F and G respectively, that
is
(8)
Pr{U ≤ xα } = G(uα ) = 1 − α.
Then for any α such that 1 − c1 ǫ > α > c1 ǫ > 0 we have
(i): uα+d1 ≤ xα ≤ uα−d1 .
(ii): |xα − uα | ≤ c1 ǫ/g(u(1) ), where g is the density function of
the limiting distribution G and
g(u(1) ) =
min
g(u).
u∈[uα+d1 ,uα−d1 ]
Theorem 2. In the notation of Theorem 1 we assume that
F (x) ≡ Pr{U ≤ x} = G(x) + ǫg(x)a(x) + R2 (x),
where for remainder term R2 (x) there exists a constant c2 such that
|R2 (x)| ≤ d2 ≡ c2 ǫ2 .
Let T = T (u) be a monotone increasing transform such that
Pr{T (U) ≤ x} = G(x) + R̃2 (x) with |R̃2 (x)| ≤ d˜2 ≡ c̃2 ǫ2 .
Let x̃α and uα be the upper 100α% points of Pr{T (U) ≤ x} and G,
respectively. Then for any α such that
we have
(9)
where
1 − c̃2 ǫ2 > α > c̃2 ǫ2 > 0,
|x̃α − uα | ≤ c̃2 ǫ2 /g(u(2) ),
g(u(2) ) =
min
u∈[uα+d̃ ,uα−d̃ ]
2
2
g(u).
4
V.V. ULYANOV, M. AOSHIMA, AND Y. FUJIKOSHI
Theorem 3. We use the notation of Theorem 2. Let b(x) be a function
inverse to T , i.e. b(T (x)) = x. Then xα = b(x̃α ) and for α such that
1 − c̃2 ǫ2 > α > c̃2 ǫ2 we have
|b′ (u∗ )| 2
(10)
|xα − b(uα )| ≤ c̃2
ǫ,
g(u(2) )
where
|b′ (u∗ )| =
max
u∈[uα+d̃ ,uα−d̃ ]
2
2
|b′ (u)|.
Moreover,
(11)
b(x) = x − ǫa(x) + O(ǫ2 ).
Remark 1. The main assumption of the Theorems is that for distributions of statistics and for distributions of transformed statistics we
have some approximations with computable error bounds. There are
not many papers with this kind of non-asymptotic results because it
requires technique which is different from the asymptotic results methods (cf., e.g., [10] and [20]). In series of papers [7], [8], [10], [11], [2],
[16], [18], [19] we got non-asymptotic results for wide class of statistics
including multivariate scale mixtures and MANOVA tests. We considered as well the case of high dimensions, that is the case when the
dimension of observations and sample size are comparable. The results
were included in the book [9]. See also [5].
Remark 2. The results of Theorems 1–3 could not be extended
to the whole range of α ∈ (0, 1). It follows from the fact that the
Cornish-Fisher expansion does not converge uniformly in 0 < α < 1.
See corresponding example in Section 2.5 of [12].
Remark 3. In Theorem 2 we required the existence of a monotone
increasing transform T (z) such that distribution of transformed statistic T (U) is approximated by some limit distribution G(x) in better way
than the distribution of original statistic U. We call this transformation T (z) the Bartlett type correction. See corresponding examples in
Section 3.
Remark 4. According to (10) and (11) the function b(uα ) in Theorem 3 could be considered as an ”asymptotic expansion” for xα up to
order O(ǫ2 ).
2. Proofs of main results
Proof of Theorem 1. By the mean value theorem,
|G(xα ) − G(uα )| ≥ |xα − uα | min g(uα + θ(xα − uα )).
0<θ<1
From (7) and the definition of xα and uα in (8), we get
|G(xα ) − G(uα )| = |G(xα ) − Pr{U ≤ xα }|
= |R1 (xα )| ≤ d1 .
NON-ASYMPTOTIC RESULTS FOR CORNISH-FISHER EXPANSIONS
5
Therefore,
(12)
|xα − uα | ≤
d1
.
min0<θ<1 g(uα + θ(xα − uα )
On the other hand, it follows from (7) that
G(xα ) = 1 − α − R1 (α)
≤ 1 − (α − d1 ) = G(uα−d1 ).
This implies that xα ≤ uα−d1 . Similarly, we have uα+d1 ≤ xα . Therefore, we proved Theorem 1 (i).
It follows from Theorem 1 (i) that
min
u∈[uα+d1 ,uα−d1 ]
g(u) ≤ min g(uα + θ(xα − uα )).
0<θ<1
Thus, using (12) we get statement of Theorem 1 (ii).
Proof of Theorem 2. It is easy to see that it is sufficient to apply Theorem 1 (ii) to the transformed statistic T (U).
Proof of Theorem 3.
we obtain
(13)
Using now (9) and the mean value theorem
x̃α − uα = b−1 (xα ) − b−1 (b(uα )) = (b−1 )′ (x∗ ) xα − b(uα ) ,
where x∗ is a point on the interval min{xα , b(uα )} , max{xα , b(uα )} .
By Theorem 1 (i) we have
uα+d˜2 ≤ x̃α ≤ uα−d˜2 .
Therefore, for xα = b(x̃α ) we get
(14)
min{b−1 (xα ), uα } , max{b−1 (xα ), uα } ⊆ uα+d˜2 , uα−d˜2 .
Since by properties of derivatives of inverse functions
(b−1 )′ (z) = 1/ b′ (b−1 (z)) = 1/b′ (y)
for z = b(y), the relations (13) and (14) imply (10).
Representation (11) for b(x) follows from (6) and (10).
3. Examples
In [17] we gave sufficient conditions for transformation T (x) to be
the Bartlett type correction (see Remark 3 above) for wide class of
statistics U allowing the following represantion
(15)
Pr{U ≤ x} = Gq (x) +
k
1 X
aj Gq+2j (x) + R2k ,
n j=0
where R2k = O(n−2 ) and Gq (x) is the distribution function of chisquared distribution with q degrees of freedom and coefficients aj ’s
P
satisfy the relation kj=0 aj = 0. Some examples of the statistic U are
6
V.V. ULYANOV, M. AOSHIMA, AND Y. FUJIKOSHI
as follows: for k = 1, the likelihood ratio test statistic; for k = 2, the
Lawley-Hotelling trace criterion and the Bartlett-Nanda-Pillai trace
criterion, which are test statistics for multivariate linear hypothesis
under normality; for k = 3, the score test statistic and Hotelling’s T 2 statistic under nonnormality. The results of [17] were extended in [4]
and [5].
In [10] we were interested in the null distribution of Hotelling’s generalized T02 statistic defined by
T02 = n trSh Se−1 ,
where Sh and Se are independently distributed as Wishart distributions
Wp (q, Ip ) and Wp (n, Ip ) with identity operator Ip in Rp , respectively.
In Theorem 4.1 (ii) in [10] we proved (15) for all n ≥ p with k = 3 and
computable error bound:
r
|Pr(T02 ≤ x) − Gr (x) −
{(q − p − 1)Gr (x)
4n
−2qGr+2 (x) + (q + p + 1)Gr+4 (x)}|
cp,q
≤ 2,
n
where r = pq and for constant cp,q we gave expicit formula with dependence on p and q.
Therefore, according to [17] we can take in this case the Bartlett
type correction T (z) as
s
2
z
a−1
a−1
+
T (z) =
+ ,
2b
2b
b
where
a=
1
p(q − p − 1),
2n
1
p(q + p + 1)(q + 2) −1.
2n
It is clear that T (z) is invertable and we can apply Theorem 3.
Other examples and numerical calculations and comparisons of approximation accuracy see in [4] and [5].
One more example is connected with sample correlation coefficient.
~ = (X1 , ..., Xn )T , and Y~ = (Y1 , ..., Yn )T be two vectors from an
Let X
n-dimensional normal distribution N(0, In ) with zero mean, identity
covariance matrix In and the sample correlation coefficient
Pn
Xk Y k
~
~
.
R = R(X, Y ) = pPn k=1 2 Pn
2
k=1 Yk
k=1 Xk
b=
In [2] it was proved for n ≥ 7 and N = n − 2.5:
√
Bn
x3 ϕ(x)
supx Pr
≤ 2,
N R ≤ x − Φ(x) −
4N
N
NON-ASYMPTOTIC RESULTS FOR CORNISH-FISHER EXPANSIONS
7
with Bn ≤ 2.2. It is easy to see that we can take T (z) as the Bartlett
type correction in the form T (z) = z + z 3 /(4N). Then the inverse
function b(z) = T −1 (z) is defined by formula
1/3
p
2
3
b(z) = 2 N z + (2Nz) + (4N/3)
1/3
p
− − 2 N z + (2Nz)2 + (4N/3)3
z3
3 z5
+
+ O(N −3 ).
4N
16 N 2
Now we can apply Theorem 3.
= z−
References
[1] L.N. Bol’shev, ”Asymptotically Pearson transformations”, Theor. Probab.
Appl., 8, 121–146 (1963).
[2] G. Christoph, V.V. Ulyanov, and Y. Fujikoshi, ”Accurate approximation of
correlation coefficients by short Edgeworth-Chebyshev expansion and its statistical applications”, Springer Proceedings in Mathematics and Statistics, 33,
239–260 (2013).
[3] E.A. Cornish and R.A. Fisher, ”Moments and cumulants in the specification
of distributions”, Rev. Inst. Internat. Statist., 4, 307–320 (1937).
[4] H. Enoki and M. Aoshima, ”Transformations with improved chi-squared approximations”, Proc. Symp., Res. Inst. Math. Sci., Kyoto Univ., 1380, 160–
181 (2004).
[5] H. Enoki and M. Aoshima, ”Transformations with improved asymptotic approximations and their accuracy”, SUT Journal of Mathematics, 42, no.1,
97–122 (2006).
[6] R.A. Fisher and E.A. Cornish, ”The percentile points of distributions having
known cumulants”, J. Amer. Statist. Assoc., 80, 915–922 (1946).
[7] Y. Fujikoshi, and V.V. Ulyanov, ”Error bounds for asymptotic expansions of
Wilks’ lambda distribution”, Journal of Multivariate Analysis, 97, no.9, 1941–
1957 (2006).
[8] Y. Fujikoshi, and V.V. Ulyanov, ”On accuracy of approximations for location
and scale mixtures”, Journal of Mathematical Sciences, 138, no.1, 5390–5395
(2006).
[9] Y. Fujikoshi, V.V. Ulyanov, and R. Shimizu, Multivariate Statistics : HighDimensional and Large-Sample Approximations, Wiley Series in Probability
and Statistics, John Wiley & Sons, Hoboken, N.J., (2010).
[10] Y. Fujikoshi, V.V. Ulyanov, and R. Shimizu, ”L1 -norm error bounds for asymptotic expansions of multivariate scale mixtures and their applications to
Hotelling’s generalized T02 ”, Journal of Multivariate Analysis, 96, no.1, 1–19
(2005).
[11] Y. Fujikoshi, V.V. Ulyanov, and R. Shimizu, ”Error bounds for asymptotic
expansions of the distribution of multivariate scale mixture”, Hiroshima Mathematical Journal, 35, no.3, 453–469 (2005).
[12] P. Hall, The Bootstrap and Edgeworth Expansion, Springer-Verlag, New York
(1992).
[13] G.W. Hill and A.W. Davis, ”Generalized asymptotic expansions of CornishFisher type”, Ann. Math. Statist., 39, 1264–1273 (1968).
[14] S. Jaschke, ”The Cornish-Fisher-expansion in the context of delta-gammanormal approximations”, J. Risk, 4, no.4, 33–52 (2002).
8
V.V. ULYANOV, M. AOSHIMA, AND Y. FUJIKOSHI
[15] V.V. Ulyanov, ”Cornish-Fisher Expansions”. In International Encyclopedia of
Statistical Science, Ed. M.Lovric, Springer-Verlag, Berlin–Heidelberg, 312–315
(2011).
[16] V.V. Ulyanov, G.Christoph, and Y. Fujikoshi, ”On approximations of transformed chi-squared distributions in statistical applications”, Siberian Mathematical Journal, 47, no.6, 1154–1166 (2006).
[17] V.V. Ulyanov and Y. Fujikoshi, ”On accuracy of improved χ2 -approximations”,
Georgian Mathematical Journal, 8, no.2, 401–414 (2001).
[18] V.V. Ulyanov, Y. Fujikoshi, and R. Shimizu, ”Nonuniform error bounds in asymptotic expansions for scale mixtures under mild moment conditions”, Journal of Mathematical Sciences, 93, no.4, 600–608 (1999).
[19] V.V. Ulyanov, H. Wakaki, and Y. Fujikoshi, ”Berry-Esseen bound for high dimensional asymptotic approximation of Wilks’ lambda distribution”, Statistics
and Probability Letters, 76, no.12, 1191–1200 (2006).
[20] H. Wakaki, Y. Fujikoshi, and V.V. Ulyanov ”Asymptotic Expansions of the
Distributions of MANOVA Test Statistics when the Dimension is Large”, Hiroshima Mathematical Journal, 44, no.3, 247–259 (2014).
V.V. Ulyanov, Faculty of Computational Mathematics and Cybernetics, Moscow State University, Moscow, 119991, Russia, and
National Research University Higher School of Economics (HSE),
Moscow, 101000, Russia
E-mail address: [email protected]
M. Aoshima, Institute of Mathematics, University of Tsukuba,
Tsukuba, Ibaraki 305-8571, Japan
E-mail address: [email protected]
Y. Fujikoshi, Department of Mathematics, Hiroshima University,
Higashi-Hiroshima, 739-8526, Japan
E-mail address: fujikoshi [email protected]
| 10math.ST
|
Distance-based Confidence Score for Neural Network Classifiers
arXiv:1709.09844v1 [cs.AI] 28 Sep 2017
Amit Mandelbaum
School of Computer Science and Engineering
Hebrew University of Jerusalem, Israel
1
Daphna Weinshall
School of Computer Science and Engineering
Hebrew University of Jerusalem, Israel
Abstract
scores are used as proxies, as for example the margin
in SVM classifiers.
The reliable measurement of confidence in
classifiers’ predictions is very important for
many applications, and is therefore an important part of classifier design. Yet, although deep learning has received tremendous attention in recent years, not much
progress has been made in quantifying the
prediction confidence of neural network classifiers. Bayesian models offer a mathematically
grounded framework to reason about model
uncertainty, but usually come with prohibitive
computational costs. In this paper we propose a simple, scalable method to achieve a
reliable confidence score, based on the data
embedding derived from the penultimate layer
of the network. We investigate two ways to
achieve desirable embeddings, by using either
a distance-based loss or Adversarial Training. We then test the benefits of our method
when used for classification error prediction,
weighting an ensemble of classifiers, and novelty detection. In all tasks we show significant
improvement over traditional, commonly used
confidence scores.
When trying to evaluate the confidence of neural network (NN) classifiers, a number of scores are commonly
used. One is the strength of the most activated output
unit followed by softmax normalization, or the closely
related ratio between the activities of the strongest
and second strongest units. Another is the (negative)
entropy of the output units, which is minimal when
all units are equally probable. Often, however, these
scores do not provide a reliable measure of confidence.
Introduction
Classification confidence scores are designed to measure the accuracy of the model when predicting class
assignment (rather than the uncertainty inherent in the
data). Most generative classification models are probabilistic in nature, and therefore provide such confidence
scores directly. Most discriminative models, on the
other hand, do not have direct access to the probability
of each prediction. Instead, related non-probabilistic
Why is it important to reliably measure prediction confidence? In various contexts such as medical diagnosis
and decision support systems, it is important to know
the prediction confidence in order to decide how to act
upon it. For example, if the confidence in a certain
prediction is too low, the involvement of a human expert in the decision process may be called for. Another
important aspect of real world applications is the ability to recognize samples that do not belong to any of
the known classes, which can also be improved with a
reliable confidence score. But even irrespective of the
application context, reliable prediction confidence can
be used to boost the classifier performance via such
methods as self-training or ensemble classification. In
this context a better confidence score can improve the
final performance of the classifier. The derivation of a
good confidence score should therefore be part of the
classifier’s design, as important as any other component
of classifiers’ design.
In order to derive a reliable confidence score for NN
classifiers, we focus our attention on an empirical observation concerning neural networks trained for classification, which have been shown to demonstrate in parallel
useful embedding properties. Specifically, a common
practice these days is to treat one of the upstream
layers of a pre-trained network as a representation (or
embedding) layer. This layer activation is then used for
representing similar objects and train simpler classifiers
(such as SVM, or shallower NNs) to perform different
tasks, related but not identical to the original task the
network had been trained on.
Distance-based Confidence Score for Neural Network Classifiers
In computer vision such embeddings are commonly
obtained by training a deep network on the recognition
of a very large database (typically ImageNet (Deng
et al., 2009)). These embeddings have been shown to
provide better semantic representations of images (as
compared to more traditional image features) in a number of related tasks, including the classification of small
datasets (Sharif Razavian et al., 2014), image annotation (Donahue et al., 2015) and structured predictions
(Hu et al., 2016). Given this semantic representation,
one can compute a natural multi-class probability distribution as described in Section 2.1, by estimating
local density in the embedding space. This estimated
density can be used to assign a confidence score to each
test point, using its likelihood to belong to the assigned
class.
We note, however, that the commonly used embedding
discussed above is associated with a network trained
for classification only, which may impede its suitability
to measure confidence reliably. In fact, when training neural networks, metric learning is often used to
achieve desirable embeddings (e.g., Weston et al. (2012);
Schroff et al. (2015); Hoffer & Ailon (2015); Tadmor
et al. (2016)). Since our goal is to improve the probabilistic interpretation of the embedding, which is essentially based on local point density estimation (or
the distance between points), we may wish to modify
the loss function and add a term which penalizes for
the violation of pairwise constraints as in Hadsell et al.
(2006). Our experiments show that the modified network indeed produces a better confidence score, with
comparable classification performance. Surprisingly,
while not directly designed for this purpose, we show
that networks which are trained with adversarial examples following the Adversarial Training paradigm
(Szegedy et al., 2013; Goodfellow et al., 2014), also
provide a suitable embedding for the new confidence
score.
Our first contribution, therefore, is a new prediction
confidence score which is based on local density estimation in the embedding space of the neural network.
This score can be computed for every network, but in
order for this score to achieve superior performance, it
is necessary to slightly change the training procedure.
In our second contribution we show that suitable embedding can be achieved by either augmenting the loss
function of the trained network with a term which penalizes for distance-based similarity loss (as in Eq. (2)
below), or by using Adversarial Training. The importance of the latter contribution is two fold: Firstly, we
are the first to show that the density of image embeddings is improved with indirect Adversarial Training
perturbations, in addition to the improved word embedding quality shown in Miyato et al. (2016) by direct
Adversarial Training perturbations. Secondly, we show
in Section 3 that Adversarial Training improves the
results while imposing a much lighter burden of hyperparameters to tune as compared to the distance-based
loss.
The new confidence score is evaluated in comparison
to other scores, using the following tasks: (i) Performance in the binary classification task of identifying
each class prediction as correct or incorrect (see Section 2.1). (ii) Training an ensemble of NN classifiers,
where each classifier’s prediction is weighted by the
new confidence score (see Section 2.4). (iii) Novelty
detection, where confidence is used to predict whether
a test point belongs to one of the known classes from
the train set (see Section 2.5).
The empirical evaluation of our method is described
in Section 3, using a few datasets and different network architectures which have been used in previous
work when using these specific datasets. Our method
achieves significant improvement in all 3 tasks. When
compared with a more recent method which had been
shown to improve traditional measures of classification
confidence - MC dropout (Gal & Ghahramani, 2015),
our distance-based score achieves better results while
also maintaining lower computational costs.
Prior Work
The Bayesian approach seeks to compute a posterior
distribution over the parameters of the neural network
which is used to estimate prediction uncertainty, as in
MacKay (1992) and Neal (2012). However, Bayesian
neural networks are not always practical to implement,
and the computational cost involved it typically high.
In accordance, in a method which is referred to below
as MC-Dropout, Gal & Ghahramani (2015) proposed
to use dropout during test time as a Bayesian approximation of the neural network, providing a cheap proxy
to Bayesian Neural Networks. Lakshminarayanan et al.
(2016) proposed to use Adversarial Training to improve
the uncertainty measure of the entropy score of the
neural network.
Still, the most basic and one of the most common
confidence scores for neural networks can be derived
from the strength of the most activated output unit, or
rather its normalized version (also called softmax output
or max margin). A confidence score that handles better
a situation where there is no one class which is most
probable, is the (negative) entropy of the normalized
network’s output. Zaragoza & d’Alché Buc (1998)
compared these scores, as well as some more complex
ones (e.g. Tibshirani (1996)), demonstrating somewhat
surprisingly the empirical superiority of the two most
basic methods described in the previous paragraph.
Amit Mandelbaum, Daphna Weinshall
Ensembles of models have been used to improve the
overall performance of the final classifier (see reviews
in Dietterich (2000) and Li et al. (2017)). There are
many ways to train an ensemble, such as boosting or
bagging. There are also many ways to integrate the
predictions of the classifiers in the ensemble, including
the average prediction or voting discussed by Bauer
& Kohavi (1999). Some ensemble methods use the
confidence score to either weight the predictions of the
different classifiers (average weighting) or for confidence
voting.
Novelty detection, where the task is to determine
whether a test point belongs to a known class label
or not, is another problem which becomes more relevant with the ever increasing availability of very large
datasets, see reviews in Markou & Singh (2003), Pimentel et al. (2014) and the recent work in Vinokurov
& Weinshall (2016). This task is also highly relevant
in real world applications, where the classifier is usually exposed to many samples which do not belong to
a known class. Note that novelty detection is quite
different from the learning of classes with no examples,
as in zero shot learning (Palatucci et al., 2009).
2
New Confidence Score
We propose next a new confidence score. We then
discuss how it can be used to boost classification performance with ensemble methods, or when dealing with
novelty detection.
2.1
New Confidence Score for Neural
Network Classifiers
Our confidence score is based on the estimation of local
density as induced by the network, when points are
represented using the effective embedding created by
the trained network in one of its upstream layers. Local
density at a point is estimated based on the Euclidean
distance in the embedded space between the point and
its k nearest neighbors in the training set.
Specifically, let f (x) denote the embedding of x as
defined by the trained neural network classifier. Let
A(x) = {xjtrain }kj=1 denote the set of k-nearest neighbors of x in the training set, based on the Euclidean
distance in the embedded space, and let {y j }kj=1 denote
the corresponding class labels of the points in A(x). A
probability space is constructed (as is customary) by
assuming that the likelihood that two points belong to
the same class is proportional to the exponential of the
negative Euclidean distance between them. In accordance, the local probability that a point x belongs to
class c is proportional to the probability that it belongs
to the same class as the subset of points in A(x) that
belong to class c.
Based on this local probability, the confidence score
D(x) for the assignment of point x to class ŷ is defined
as follows:
Pk
−||f (x)−f (xjtrain )||2
j=1,y j =ŷ e
D(x) = Pk
(1)
−||f (x)−f (xjtrain )||2
j=1 e
D(x) is a score between 0 to 1, which is monotonically
related to the local density of similarly labeled train
points in the neighborhood of x. Henceforth (1) is
referred to as Distance score.1 We note here that while
intuitively it might be beneficial to add a scaling factor
to the distance in (1), such as the mean distance, we
found it to have a deteriorating effect, in line with
related work such as Salakhutdinov & Hinton (2007).
Two ways to achieve effective embedding: As
mentioned is Section 1, in order to achieve an effective
embedding it helps to modify the training procedure of
the neural network classifier. The simplest modification
augments the network’s loss function during training
with an additional term. The resulting loss function is
a linear combination of two terms, one for classification
denoted Lclass (X, Y ), and another pairwise loss for
the embedding denoted Ldist (X, Y ). This is defined as
follows:
L(X, Y ) = Lclass (X, Y ) + αLdist (X, Y )
Ldist (X, Y ) =
1
P
P
X
(2)
Ldist (xp1 , xp2 )
p=1
where Ldist (xi , xj ) is defined as
(
||f (xi ) − f (xj )||2
if y i = y j
max{0, (m − ||f (xi ) − f (xj )||2 )} if y i 6= y j
A desirable embedding can also be achieved by Adversarial Training, using the fast gradient method suggested in Goodfellow et al. (2014). In this method,
given an input x with target y, and a neural network
with parameters θ, adversarial examples are generated
using:
x0 = x + sign(5x Lclass (θ, x, y))
(3)
In each step an adversarial example is generated for
each point x in the batch and the current parameters
of the network, and classification loss is minimized for
both the regular and adversarial examples. Although
originally designed to improve robustness, this method
1
Related measures of density, such as a count of the
"correct" neighbors or the inverse of the distance, behave
similarly and perform comparably.
Distance-based Confidence Score for Neural Network Classifiers
seems to improve the network’s embedding for the purpose of density estimation, possibly because along the
way it increases the distance between pairs of adjacent
points with different labels.
Implementation details: In (2) Ldist is defined by
all pairs of points, denoted (xp1 , xp2 ). For each training
minibatch, this set is sampled with no replacement from
the training points in the minibatch, with half as many
pairs as the size of the minibatch. In our experiments,
Lclass (X, Y ) is the regular cross entropy loss. We note
here that we also tried distance-based loss functions
which do not limit the distance between points of the
same class to be exactly 0 (such as those in Hoffer
& Ailon (2015) and Tadmor et al. (2016)). However,
those functions produced worse results, especially when
the dataset had many classes. Finally we note that we
have tried using the distance-based loss and adversarial
training together while training the network, but this
has also produced worse results.
2.2
Alternative confidence scores
Given a trained network, two measure are usually used
to evaluate classification confidence:
Max margin: the maximal activation, after normalization, in the output layer of the network.
Entropy: the (negative) entropy of the activations in
the output layer of the network.
As noted above, the empirical study in Zaragoza &
d’Alché Buc (1998) showed that these two measures
are typically as good as any other existing method for
the evaluation of classification confidence.
Two recent methods have been shown to improve the reliability of the confidence score based on Entropy: MCDropout (Gal & Ghahramani, 2015) and Adversarial
Training (Lakshminarayanan et al., 2016; Goodfellow
et al., 2014). In terms of computational cost, adversarial training can increase (and sometimes double) the
training time, due to the computation of additional
gradients and the addition of the adversarial examples
to the training set. MC-Dropout, on the other hand,
does not change the training time but increases the test
time by orders of magnitude (typically 100-fold). Both
methods are complementary to our approach, in that
they focus on modifications to the actual computation
of the network during either train or test time. After
all is done, they both evaluate confidence using the
Entropy score. As we show in our experiments, adversarial training combined with our proposed confidence
score improves the final results significantly.
2.3
Our method: computational analysis
Unlike the two methods described above, MC-Dropout
and Adversarial Training, our distance-based confidence score takes an existing network and computes
a new confidence score from the network’s embedding
and output activation. It can use any network, with
or without adversarial training or MC dropout. If the
loss function of the network is suitably augmented (see
discussion above), empirical results in Section 3 show
that our score always improves results over the Entropy
score of the given network.
Train and test computational complexity: Considering the distance-based loss, Tadmor et al. (2016)
showed that computing distances during the training
of neural networks have negligible effect on training
time. Alternatively, when using adversarial training,
additional computational cost is incurred as mentioned
above, while on the other hand fewer hyper parameters are left for tuning. During test time, our method
requires carrying over the embeddings of the training data and also the computation of the k nearest
neighbors for each sample.
Nearest neighbor classification has been studied extensively in the past 50 years, and consequently there are
many methods to perform either precise or approximate k-nn with reduced time and space complexity (see
Gunadi (2011) for a recent empirical comparison of the
main methods). In our experiments, while using either
Condensed Nearest Neighbours (Hart, 1968) or Density
Preserving Sampling (Budka & Gabrys, 2013), we were
able to reduce the memory requirements of the train set
to 5% of its original size without affecting performance.
At this point the additional storage required for the
nearest neighbor step was much smaller than the size
of the networks used for classification, and the increase
in space complexity became insignificant.
With regards to time complexity, recent studies have
shown how modern GPU’s can be used to speed up
nearest neighbor computation by orders of magnitude
(Garcia et al., 2008; Arefin et al., 2012). Hyvönen
et al. (2015) also showed that k-nn approximation with
99% recall can be accomplished 10-100 times faster as
compared to precise k-nn.
Combining such reductions in both space and time, we
note that even for a very large dataset, including for
example 1M images embedded in a 1K dimensional
space, the computation complexity of the k nearest
neighbors for each test sample requires at most 5M
floating-point operations. This is comparable and even
much faster than a single forward run of this test sample
through a modern, relatively small, ResNets (He et al.,
2016) with 2-30M parameters. Thus, our method scales
Amit Mandelbaum, Daphna Weinshall
well even for very large datasets.
2.4
Ensembles of Classifiers
There are many ways to define ensembles of classifiers,
and different ways to put them together. Here we focus
on ensembles which are obtained when using different
training parameters with a single training method. This
specifically means that we train several neural networks
using random initialization of the network parameters,
along with random shuffling of the train points.
Henceforth Regular Networks will refer to networks
that were trained only for classification with the regular cross-entropy loss, Distance Networks will refer
to networks that were trained with the loss function
defined in (2), and AT Networks will refer to networks
that were trained with adversarial examples as defined
in (3).
Ensemble methods differ in how they weigh the predictions of different classifiers in the ensemble. A number
of options are in common use (see Li et al. (2017)
for a recent review), and in accordance are used for
comparison in the experimental evaluation section: 1)
softmax average, 2) simple voting, 3) weighted softmax
average (where each softmax vector is multiplied by
its related prediction confidence score), 4) confidence
voting (where the most confident network gets n2 votes),
and 5) dictator voting (the decision of the most confident network prevails). We evaluate methods 3 − 5
with weights defined by either the Entropy score or the
Distance score defined in (1).
2.5
Novelty Detection
Novelty detection seeks to identify points in the test set
which belong to classes not present in the train set. To
evaluate performance in this task we train a network
with a known benchmark dataset, while augmenting
the test set with test points from another dataset that
includes different classes. Each confidence score is used
to differentiate between known and unknown samples.
This is a binary classification task, and therefore we
can evaluate performance using ROC curves.
3
Experimental Evaluation
2010) (32 × 32 version) and SVHN (Netzer et al., 2011).
In all cases, as is commonly done, the data was preprocessed using global contrast normalization and ZCA
whitening. No other method of data augmentation was
used for CIFAR-100 and SVHN, while for SVHN we
also did not use the additional 5̃00K labeled images2 .
For STL-10, on the other hand, cropping and flipping
were used for STL-10 to check the robustness of our
method to heavy data augmentation.
In our experiments, all networks used ELU (Clevert
et al., 2015) for non-linear activation. For CIFAR-100
and STL-10 we used the network suggested in Clevert
et al. (2015) with the following architecture:
C(192, 5) ⇒ P (2) ⇒ C(192, 1) ⇒ C(240, 3) ⇒
P (2) ⇒ C(240, 1) ⇒ C(260, 3) ⇒ P (2) ⇒
C(260, 1) ⇒ C(280, 2) ⇒ P (2) ⇒ C(280, 1) ⇒
C(300, 2) ⇒ P (2) ⇒ C(300, 1) ⇒ F C(100)
C(n, k) denotes a convolution layer with n kernels of
size k × k and stride 1. P (k) denotes a max-pooling
layer with window size k × k and stride 2, and F C(n)
denotes a fully connected layer with n output units. For
STL-10 the last layer was replaced by FC(10). During
training (only) we applied dropout (Srivastava et al.,
2014) before each max pooling layer (excluding the first)
and after the last convolution, with the corresponding
drop probabilities of [0.1, 0.2, 0.3, 0.4, 0.5].
With the SVHN dataset we used the following architecture:
C(32, 3) ⇒ C(32, 3) ⇒ C(32, 3) ⇒ P (2) ⇒
C(64, 3) ⇒ C(64, 3) ⇒ C(64, 3) ⇒ P (2) ⇒
C(128, 3) ⇒ C(128, 3) ⇒ C(128, 3) ⇒ P (2) ⇒
F C(128) ⇒ F C(10)
For the networks trained with distance loss, for each
batch, we randomly picked pairs of points so that at
least 20% of the batch included pairs of points from
the same class. The margin m in (2) was set to 25 in
all cases, and the parameter α in (2) was set to 0.2.
The rest of the training parameters can be found in the
supplementary material. For the distance score we observed that the number of k nearest neighbors could be
set to the maximum value, which is the number of samples in each class in the train data. We also observed
that smaller numbers (even k = 50) often worked as
2
In this section we empirically evaluate the benefits of
our proposed approach, comparing the performance
of the new confidence score with alternative existing
scores in the 3 different tasks described above.
3.1
Experimental Settings
For evaluation we used 3 data sets: CIFAR-100
(Krizhevsky & Hinton, 2009), STL-10 (Coates et al.,
Note that reported results denoted as "state-of-the-art"
for these datasets often involve heavy augmentation. In our
study, in order to be able to do the exhaustive comparisons
described below, we opted for the un-augmented scenario
as more flexible and yet informative enough for the purpose
of comparison between different methods. Therefore our
numerical results should be compared to empirical studies
which used similar un-augmented settings. We specifically
selected commonly used architectures that achieve good
performance, close to the results of modern ResNets, and
yet flexible enough for extensive evaluations.
Distance-based Confidence Score for Neural Network Classifiers
Table 1: AUC Results of Correct Classification.
Conf.
Score
Margin
Entropy
Distance
CIFAR-100
(Classifier acccuracy: 60%)
Reg.
Dist.
AT
MCD
0.828 0.834
0.844 0.836
0.833 0.837
0.851 0.833
0.789 0.853 0.843 0.726
STL-10
(Classifier acccuracy: 70.5%)
Reg.
Dist.
AT
MCD
0.804 0.775 0.812
0.804
0.806 0.786 0.816
0.810
0.798 0.824 0.863 0.671
SVHN
(Accuracy: 93.5%)
Reg.
Dist.
AT
0.904 0.896
0.909
0.916 0.907
0.918
0.916 0.925 0.925
Table 1, legend. Leftmost column: Margin and Entropy denote the commonly used confidence scores described in
Section 2.2. Distance denotes our proposed method described in Section 2.1. Second line: Reg. denotes networks trained
with the entropy loss, Dist. denotes networks trained with the distance loss defined in (2), AT denotes networks trained
with adversarial training as defined in (3), and MCD denotes MC-Dropout when applied to networks normally trained
with the entropy loss. Since the network trained for SVHN was trained without dropout,MCD was not applicable.
Table 2: AUC Results of Correct classification - Ensemble of 2 Networks.
Confidence
Score
Max margin
Entropy
Distance (1)
Distance (2)
CIFAR-100
Reg.
0.840
0.844
0.775
0.876
Dist.
0.846
0.839
0.863
0.872
AT
0.856
0.857
0.862
0.879
STL-10
Reg.
0.802
0.807
0.815
0.833
Dist.
0.792
0.798
0.834
0.832
SVHN
AT
0.814
0.816
0.866
0.859
Reg.
0.909
0.918
0.916
0.918
Dist.
0.901
0.912
0.924
0.929
AT
0.911
0.920
0.926
0.927
Table 2, legend. Notations are similar to those described in the legend of Table 1, with one distinction: Distance (1) now
denotes the regular architecture where the distance score is computed independently for each network in the pair using its
own embedding, while Distance (2) denotes the hybrid architecture where one network in the pair is fixed to be a Distance
network, and its embedding is used to compute the distance score for the prediction of the second network in the pair.
well. In general, the results reported below are not
sensitive to the specific values of the hyper-parameters
as listed above; we observed only minor changes when
changing the values of k, α and the margin m.
MC-Dropout: as proposed in Gal & Ghahramani (2015),
we used MC dropout in the following manner. We
trained each network as usual, but computed the predictions while using dropout during test. This was
repeated 100 times for each test example, and the
average activation was delivered as output.
Adversarial Training: we used (3) following Goodfellow
et al. (2014), fixing = 0.1 in all the experiments.
3.2
Error Prediction of Multi-class Labels
We first compare the performance of our confidence
score in the binary task of evaluating whether the network’s predicted classification label is correct or not.
While our results are independent of the actual accuracy, we note that the accuracy is comparable to those
achieved with ResNets when not using augmentation
for CIFAR-100 or when using only the regular training
data for SVHN (see Huang et al. (2016) for example).
Performance in this binary task is evaluated using ROC
curves computed separately for each confidence score.
Results on all three datasets can be seen in Table 1. In
all cases our proposed distance score, when computed
on a suitably trained network, achieves significant improvement over the alternative scores, even when those
are enhanced by using either Adversarial Training or
MC-Dropout.
To further test our distance score we evaluate performance over an ensemble of two networks. Results are
shown in Table 2. Here too, the distance score achieves
significant improvement over all other methods. We
also note that the difference between the distance score
computed over Distance networks and the entropy score
computed over adversarially trained networks, is now
much higher as compared to this difference when using
only one network. As we show in Section 3.3, adversarial training typically leads to a decreased performance
when using an ensemble of networks and relying only
on the entropy score (probably due to a decrease in variance among the classifiers). This observation further
supports the added value of our proposed confidence
score.
As a final note, we also used a hybrid architecture using
a matched pair of one classification network (of any
kind) and a second Distance network. The embedding
defined by the Distance network is used to compute
the distance score for the predictions of the first classification network. Surprisingly, this method achieves
Amit Mandelbaum, Daphna Weinshall
Figure 1: Accuracy when using an ensemble of networks with CIFAR-100 (top left), STL-10 (top right) and
SVNH (bottom). The X-axis denotes the number of networks in the ensemble. Absolute accuracy (marked on
the left Y -axis) is shown for the 2 most successful ensemble methods among all the methods we had evaluated
(blue and yellow solid lines, see text), and 2 methods which did not use our distance score including the best
performing method in this set (red dotted line, denoted baseline). Differences in accuracy between the two top
performers and the top baseline method are shown using a bar plot (marked on the right Y -axis), with standard
deviation of the difference over (at least) 5 repetitions.
the best results in both CIFAR-100 and SVHN while
being comparable to the best result in STL-10. This
method is used later in Section 3.3 to improve accuracy
when running an ensemble of networks. Further investigation of this phenomenon lies beyond of the scope
of the current study.
3.3
Ensemble Methods
In order to evaluate the improvement in performance
when using our confidence score to direct the integration of classifiers in an ensemble, we used a few common
ways to define the integration procedure, and a few
ways to construct the ensemble itself. In all comparisons the number of networks in the ensemble remained
fixed at n. Our experiments included the following
ensemble compositions: (a) n regular networks, (b) n
distance networks, (c) n AT (Adversarially Trained)
networks, and (d-f) n networks such that n2 networks
belong to one kind of networks (regular, distance or
AT ) and the remaining n2 networks belong to another
kind, spanning all 3 combinations.
As described in Section 2.4, the predictions of classifiers
in an ensemble can be integrated using different criteria.
In general we found that all the methods which did not
use our distance score (1), including methods which
used any of the other confidence score for prediction
weighting, performed less well than a simple average
of the softmax activation (method 1 in Section 2.4).
Otherwise the best performance was obtained when
using a weighted average (method 3 in Section 2.4)
with weights defined by our distance score (1). With
variants (d-f) we also checked two options of obtaining
the distance score: (i) Each network defined its own
confidence score; (ii) in light of the advantage demonstrated by hybrid networks as shown in Section 3.2
and for each pair of networks from different kinds, the
distance score for both was computed while using the
embedding of only one of networks in the pair. MCDropout was not used in this section due to its high
computational cost.
While our experiments included all variants and all
weighting options, only 4 cases are shown in the following description of the results in order to improve
Distance-based Confidence Score for Neural Network Classifiers
readability: 1) the combination achieving best performance; 2) the combination achieving best performance
when not using adversarial training (as AT entails
additional computational load at train time); 3) the
ensemble variant achieving best performance without
using the distance score (baseline), 4) ensemble average
when using adversarial training without distance score.
Additional results for most of the other conditions we
tested can be found in the supplementary material. To
gain a better statistical significance, each experiment
was repeated at least 5 times, with no overlap between
the networks.
CIFAR-100 and STL-10: Fig. 1 shows the ensemble accuracy for the methods mentioned above when
using these datasets. It can be clearly seen that weighting predictions based on the distance score from (1)
improves results significantly. The best results are
achieved when combining Distance Networks and Adversarial Networks, with significant improvement over
an ensemble of only one kind of networks (not shown
in the graph). Still, we note importantly that the distance score is used to weight both kind of networks.
Since Adversarial Training is not always applicable
due to its computational cost at train time, we show
that the combination of Distance networks and Regular
networks can also lead to significant improvement in
performance when using the distance score and the
hybrid architecture described in Section 3.2. Finally
we note that Adversarial networks alone achieve very
poor results when using the original ensemble average, further demonstrating the value of the distance
score in improving the performance of an ensemble of
Adversarial networks alone.
SVHN: Results for this dataset are also shown in Fig.1.
While not as significant as those in the other datasets
(partly due to the high initial accuracy), they are still
consistent with them, demonstrating again the power
and robustness of the distance score.
3.4
Novelty Detection
Finally, we compare the performance of the different
confidence scores in the task of novelty detection. In
this task the confidence score is used to decide another
binary classification problem: does the test example belong to the set of classes the networks had been trained
on, or rather to some unknown class? Performance in
this binary classification task is evaluated using the
corresponding ROC curve of each confidence score.
We used two contrived datasets to evaluate performance
in this task, following the experimental construction
suggested in Lakshminarayanan et al. (2016). In the
first experiment, we trained the network on the STL-10
dataset, and then tested it on both STL-10 and SVHN
test sets. In the second experiment we switched be-
tween the datasets (and changed the trained network)
making SVHN the known dataset and STL-10 the
novel one. The task requires to discriminate between
the known and the novel datasets. For comparison we
computed novelty, as one often does, with a one-class
SVM classifier while using the same embeddings. Novelty thus computed showed much poorer performance,
possibly because this dataset involves many classes
(one class SVM is typically used with a single class),
and therefore these results are not included here.
Table 3: AUC Results for Novelty Detection.
Confide.
Score
Max
Margin
Entropy
Distance
STL-10/SVHN
SVHN/STL-10
Reg.
Dist.
AT
Reg.
Dist.
AT
.808
.849
.860
.912
.922
.985
.810
.798
.857
.870
.870
.901
.917
.904
.933
.934
.992
.996
Table 3, legend. Left: STL-10 (known) and SVHN (novel).
Right: SVHN (known) and STL-10 (novel).
Results are shown in Table 3. Adversarial training,
which was designed to handle this sort of challenge,
is not surprisingly the best performer. Nevertheless,
we see that our proposed confidence score improves
the results even further, again demonstrating its added
value.
4
Conclusions
We proposed a new confidence score for multi-class
neural network classifiers. The method we proposed
to compute this score is scalable, simple to implement,
and can fit any kind of neural network. This method
is different from other commonly used methods as it is
based on measuring the point density in the effective
embedding space of the network, thus providing a more
coherent statistical measure for the distribution of the
network’s predictions.
We also showed that suitable embeddings can be
achieved by using either a distance-based loss or, somewhat unexpectedly, Adversarial Training. We demonstrated the superiority of the new score in a number
of tasks. Those tasks were evaluated using a number
of different datasets and with task-appropriate network architectures. In all tasks our proposed method
achieved the best results when compared to traditional
confidence scores.
References
Arefin, Ahmed Shamsul, Riveros, Carlos, Berretta,
Regina, and Moscato, Pablo. Gpu-fs-k nn: A soft-
Amit Mandelbaum, Daphna Weinshall
ware tool for fast and scalable k nn computation
using gpus. PloS one, 7(8):e44000, 2012.
Bauer, Eric and Kohavi, Ron. An empirical comparison
of voting classification algorithms: Bagging, boosting,
and variants. Machine learning, 36(1-2):105–139,
1999.
Budka, Marcin and Gabrys, Bogdan.
Densitypreserving sampling: robust and efficient alternative
to cross-validation for error estimation. IEEE transactions on neural networks and learning systems, 24
(1):22–34, 2013.
Clevert, Djork-Arné, Unterthiner, Thomas, and Hochreiter, Sepp. Fast and accurate deep network learning
by exponential linear units (elus). arXiv preprint
arXiv:1511.07289, 2015.
Coates, Adam, Lee, Honglak, and Ng, Andrew Y. An
analysis of single-layer networks in unsupervised feature learning. Ann Arbor, 1001(48109):2, 2010.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and
Fei-Fei, L. ImageNet: A Large-Scale Hierarchical
Image Database. In CVPR09, 2009.
Dietterich, Thomas G. Ensemble methods in machine
learning. In International workshop on multiple classifier systems, pp. 1–15. Springer, 2000.
Donahue, Jeffrey, Anne Hendricks, Lisa, Guadarrama,
Sergio, Rohrbach, Marcus, Venugopalan, Subhashini,
Saenko, Kate, and Darrell, Trevor. Long-term recurrent convolutional networks for visual recognition
and description. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
2625–2634, 2015.
Gal, Yarin and Ghahramani, Zoubin. Dropout as
a bayesian approximation: Representing model
uncertainty in deep learning.
arXiv preprint
arXiv:1506.02142, 2, 2015.
Garcia, Vincent, Debreuve, Eric, and Barlaud, Michel.
Fast k nearest neighbor search using gpu. In Computer Vision and Pattern Recognition Workshops,
2008. CVPRW’08. IEEE Computer Society Conference on, pp. 1–6. IEEE, 2008.
Goodfellow, Ian J, Shlens, Jonathon, and Szegedy,
Christian. Explaining and harnessing adversarial
examples. arXiv preprint arXiv:1412.6572, 2014.
Gunadi, Hendra. Comparing nearest neighbor algorithms in high-dimensional space. 2011.
Hadsell, Raia, Chopra, Sumit, and LeCun, Yann. Dimensionality reduction by learning an invariant mapping. In Computer vision and pattern recognition,
2006 IEEE computer society conference on, volume 2,
pp. 1735–1742. IEEE, 2006.
Hart, Peter. The condensed nearest neighbor rule
(corresp.). IEEE transactions on information theory,
14(3):515–516, 1968.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Deep residual learning for image recognition.
In Proceedings of the IEEE conference on computer
vision and pattern recognition, pp. 770–778, 2016.
Hoffer, Elad and Ailon, Nir. Deep metric learning
using triplet network. In International Workshop
on Similarity-Based Pattern Recognition, pp. 84–92.
Springer, 2015.
Hu, Hexiang, Zhou, Guang-Tong, Deng, Zhiwei, Liao,
Zicheng, and Mori, Greg. Learning structured inference neural networks with label relations. In Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pp. 2960–2968, 2016.
Huang, Gao, Sun, Yu, Liu, Zhuang, Sedra, Daniel, and
Weinberger, Kilian Q. Deep networks with stochastic
depth. In European Conference on Computer Vision,
pp. 646–661. Springer, 2016.
Hyvönen, Ville, Pitkänen, Teemu, Tasoulis, Sotiris,
Jääsaari, Elias, Tuomainen, Risto, Wang, Liang,
Corander, Jukka, and Roos, Teemu. Fast k-nn search.
arXiv preprint arXiv:1509.06957, 2015.
Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple layers of features from tiny images. 2009.
Lakshminarayanan, Balaji, Pritzel, Alexander, and
Blundell, Charles. Simple and scalable predictive
uncertainty estimation using deep ensembles. arXiv
preprint arXiv:1612.01474, 2016.
Li, Hui, Wang, Xuesong, and Ding, Shifei. Research
and development of neural network ensembles: a
survey. Artificial Intelligence Review, pp. 1–25, 2017.
MacKay, David JC. Bayesian methods for adaptive
models. PhD thesis, California Institute of Technology, 1992.
Markou, Markos and Singh, Sameer. Novelty detection:
a review—part 2:: neural network based approaches.
83(12):2499–2521, 2003.
Miyato, Takeru, Dai, Andrew M, and Goodfellow, Ian.
Adversarial training methods for semi-supervised
text classification. arXiv preprint arXiv:1605.07725,
2016.
Neal, Radford M. Bayesian learning for neural networks, volume 118. Springer Science & Business
Media, 2012.
Netzer, Yuval, Wang, Tao, Coates, Adam, Bissacco,
Alessandro, Wu, Bo, and Ng, Andrew Y. Reading
digits in natural images with unsupervised feature
learning. In NIPS workshop on deep learning and
unsupervised feature learning, volume 2011, pp. 5,
2011.
Distance-based Confidence Score for Neural Network Classifiers
Palatucci, Mark, Pomerleau, Dean, Hinton, Geoffrey E,
and Mitchell, Tom M. Zero-shot learning with semantic output codes. In Advances in neural information
processing systems, pp. 1410–1418, 2009.
Pimentel, Marco AF, Clifton, David A, Clifton, Lei,
and Tarassenko, Lionel. A review of novelty detection.
Signal Processing, 99:215–249, 2014.
Salakhutdinov, Ruslan and Hinton, Geoffrey E. Learning a nonlinear embedding by preserving class neighbourhood structure. In AISTATS, volume 11, 2007.
Schroff, Florian, Kalenichenko, Dmitry, and Philbin,
James. Facenet: A unified embedding for face recognition and clustering. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, pp. 815–823, 2015.
Sharif Razavian, Ali, Azizpour, Hossein, Sullivan,
Josephine, and Carlsson, Stefan. Cnn features offthe-shelf: an astounding baseline for recognition. In
Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition Workshops, pp. 806–
813, 2014.
Srivastava, Nitish, Hinton, Geoffrey E, Krizhevsky,
Alex, Sutskever, Ilya, and Salakhutdinov, Ruslan.
Dropout: a simple way to prevent neural networks
from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014.
Szegedy, Christian, Zaremba, Wojciech, Sutskever, Ilya,
Bruna, Joan, Erhan, Dumitru, Goodfellow, Ian, and
Fergus, Rob. Intriguing properties of neural networks.
arXiv preprint arXiv:1312.6199, 2013.
Tadmor, Oren, Rosenwein, Tal, Shalev-Shwartz, Shai,
Wexler, Yonatan, and Shashua, Amnon. Learning
a metric embedding for face recognition using the
multibatch method. In Advances In Neural Information Processing Systems, pp. 1388–1389, 2016.
Tibshirani, Robert. A comparison of some error estimates for neural network models. Neural Computation, 8(1):152–163, 1996.
Vinokurov, Nomi and Weinshall, Daphna. Novelty
detection in multiclass scenarios with incomplete
set of class labels. arXiv preprint arXiv:1604.06242,
2016.
Weston, Jason, Ratle, Frédéric, Mobahi, Hossein, and
Collobert, Ronan. Deep learning via semi-supervised
embedding. In Neural Networks: Tricks of the Trade,
pp. 639–655. Springer, 2012.
Zaragoza, Hugo and d’Alché Buc, Florence. Confidence
measures for neural network classifiers. In Proceedings of the Seventh Int. Conf. Information Processing
and Management of Uncertainty in Knowlegde Based
Systems, 1998.
| 2cs.AI
|
arXiv:1205.4691v2 [cs.LO] 22 May 2012
On the discriminating power of tests in resource
λ-calculus
January 20, 2018
Abstract
Since its discovery, differential linear logic (DLL) inspired numerous
domains. In denotational semantics, categorical models of DLL are now
commune, and the simplest one is Rel, the category of sets and relations.
In proof theory this naturally gave birth to differential proof nets that are
full and complete for DLL. In turn, these tools can naturally be translated
to their intuitionistic counterpart. By taking the co-Kleisly category associated to the ! comonad, Rel becomes MRel, a model of the λ-calculus that
contains a notion of differentiation. Proof nets can be used naturally to
extend the λ-calculus into the lambda calculus with resources, a calculus
that contains notions of linearity and differentiations. Of course MRel is
a model of the λ-calculus with resources, and it has been proved adequate,
but is it fully abstract?
That was a strong conjecture of Bucciarelli, Carraro, Ehrhard and
Manzonetto in [4]. However, in this paper we exhibit a counter-example.
Moreover, to give more intuition on the essence of the counter-example
and to look for more generality, we will use an extension of the resource
λ-calculus also introduced by Bucciarelli et al in [4] for which M∞ is fully
abstract, the tests.
1
Introduction
The first extension of the λ-calculus with resources, by Boudol in 1993 [1], was
introducing a special resource sensitive application that may involve multisets
of affine arguments (each one has to be used at most one time). This was a
natural way to export resource sensitiveness to the functional setting. However,
gathering no known and interesting properties (confluence, linearity...), it was
not fully explored.
Later on, Ehrhard and Regnier, working on functional interpretation of differential proof nets, discovered a calculus similar to Boudol’s one, named differential λ-calculus [10]. By adding to the λ-calculus a derivative operation
∂M
(N ), which syntactically corresponds to a linear substitution of x by N in
x
M , it recovers the resource-sensitiveness. This is done through the translation
1
(n)
(λx.M ) [N1 , ..., Nn ; N ! ] ≃ (λx. ∂ xM (N1,..., Nn)) N where Ni are the linear arguments and N is the non linear one. This more semantical view even allow
for the generalisation of the operation and recover excellent semantical properties (confluence, Taylor expansion...). We will adopt the syntax of [13] that
re-implements improvements from differential λ-calculus into Boudol’s calculus,
and we will call it ∂λ-calculus.
The category Rel of set and relations is known to model the linear logic,
and, despite its high degree of degeneration (Relop = Rel), it is a very natural
construction. Indeed, what appeared to be a degeneration is in reality a natural
choice that preserves all proofs, i.e. the interpretation function from proof to
MRel is injective up to isomorphism ([8]). But our principal interest for this
category is that it models the differential linear logic, and of known such category
it is the simplest and more natural.
As for every categorical model of linear logic, the interpretation of the ! induced a comonad. From that comonad we can construct the co-kleisly category.
In the case of Rel, this new category, MRel, corresponds to the category of sets
with, as morphisms from A to B, the relations from Mf (A) (the finite multisets
over A) to B. It is then a model of the λ-calculus and of the ∂λ-calculus. This
construction being the most natural we can do, MRel is, a priori, one of the
most natural models of the ∂λ-calculus (even if non well pointed [6]).
It is only natural, then, to question on the depth of the link among the
reflexive elements of MRel and the ∂λ-calculus. And more precisely among
MRel’s canonical reflexive element M∞ and the ∂λ-calculus. Until now we knew
that M∞ ([6]) was adequate for the λ-calculus, i.e. that two terms carrying the
same interpretations in MRel behave the same way in all contexts. But we did
not know anything about the counterpart, named full abstraction.
This question has been thoroughly studied, however, since M∞ has been
proved (resp. in [12], [5] and [4]) fully abstract not only for both of the principal
sub-calculi of ∂λ-calculus, namely the usual λ-calculus and Kfoury’s linear calculus of [11], but also for the extension with tests of [4] (denoted τ ∂λ-calculus).
Therefore Bucciarelli et al emit in [4] a strong conjecture of full abstraction for
the ∂λ-calculus.
However, and it is our purpose here, a counter example can be found. In
order to exhibit this counter example, we will take an unusual shortcut using
full abstraction result for τ ∂λ-calculus. Indeed, we will prove a slightly more
general theorem: the failure of full abstraction for ∂λ-calculus of any model that
is fully abstract for τ ∂λ-calculus. Due to this generalization we will not have to
introduce the full description of M∞ in the core of the article (it is available in
annexes).
Additionally to be considerably easier and more intuitive than the direct and
usual method, this way of proceeding is part of a larger study of full abstraction.
Indeed, we are looking for a mechanical way to tackle full abstraction problems
in two steps. First we extend the calculus with well chosen semantical objects
in order to reach the definability of compact elements. Then we study the full
abstraction question indirectly via the link between the operational equivalence
2
of the original calculus and of its artificial extension. This reduces our mix of
semantic and syntactic question to a purely syntactic one, allowing us to use
powerful syntactic constructions.
Tests where introduced in [2] to have a full abstraction theorem for Boudol
λ-calculus with resources. Later on, the principle was improved in [4], implementing semantic objects in the syntax in order to get full abstraction of M∞ ,
following an idea of [9]. This extension can, in our context, be compared to a
basic exception mechanism. The term τ̄ (Q) is raising the exception (or test)
Q, absorbing all its non resources-limited applications and the exception τ (M )
is catching any exception in M by annihilating all the head-λ-abstraction. The
most important here being the scope of the τ (M ) that act as an infinite application over M .
Notations: λx̄n will be used for λx1 , ..., xn (n is not specified when it can be
any integer) and I will denotes the identity λx.x.
2
2.1
Background
∂λ-calculus
As explained, this article is directly following [4]. For this reason we need to
introduce the ∂λ-calculus and then the tests. In the ∂λ-calculus , the notion of
linearity is capital. Any term in linear position will never suffer any duplication
or erasing regardless the reduction strategy. Linear subterms are subterms that
are either the first subterm of a lambda abstraction in linear position, the left
side of an application that is in linear position, or in the linear part of its
right side. The last case is the real improvement and asks for arguments to be
separated in linear and non linear arguments. Therefore, the right side of the
applications will be replaced by a new kind of expression different from terms,
the “bags”. Bags are multisets containing some linear (non banged) arguments
and exactly one non linear (banged) argument:
(terms) M, N ∶ λx.M ∣ M B ∣ M +N ∣ 0
(bags)
B, C ∶ [M1 , ..., Mn ; M ! ] ∣ B + C ∣ 0
This is the syntax of [4] modulo the macro M +N = (λx.x) [{M +N }! ].
For convenience, the finite sums will be denoted Σi Qi and the different 0’s are
just the neutral elements of the different sums. This demonic sum had to be
implemented since we want the calculus to be resource sensitive and confluent, thus there is no other choice than to considere the sum of all the possible
outcomes. Sums distribute with any linear context:
λx.(Σi Mi ) = Σi (λx .Mi )
(Σi Mi ) (Σj Nj ) = Σi,j (Mi Nj )
([(Σi1 ≤k1 Mi11 ), ..., (Σin ≤kn Minn ); M ! ]) = Σ(ij )j ≤(kj )j [Mii1 , ..., Minn ; M ! ]
In the application, each linear argument will replace one and only one occurrence
of the variable, thus the need of two kinds of substitutions, the usual one,
denoted {.}, and the linear one, denoted ⟨.⟩. This last will act like a derivation
(N ):
M ⟨N /x⟩ ∼ ∂M
∂x
3
x⟨N /x⟩ = N
x⟨N /y⟩ = 0
(λy.M )⟨N /x⟩ = λy.(M ⟨N /x⟩)
([M1 , ..., Mn ; M ! ])⟨N /x⟩ = (Σni=1 [M1 , .., Mi ⟨N /x⟩, ...Mn ; M ! ]) + [M1 , ..., Mn , M ⟨N /x⟩; M ! ]
(M P )⟨N /x⟩ = (M ⟨N /x⟩ P ) + (M P ⟨N /x⟩)
This enables us to describe the β-reduction:
(λx.M )[N1 , ..., Nn ; N ! ] → M ⟨N1 /x⟩⋯⟨Nn /x⟩{N ! /x}
(β)
In other words (λx.M )[N1 , ..., Nn ; N ! ] →
2.2
∂nM
(N1 , ..., Nn )(N )
∂nx
τ∂λ-calculus
In differential proof nets the 0-ary tensor and the 0-ary par can be added freely
in the sense that we still have a natural interpretation in MRel and M∞ . These
operations can be translated in our calculus as an exception mechanism. With
on one side a τ (Q) that “raises” the exception (or test) Q by burning its applicative context (whenever these applications do not have any linear component, otherwise it diverges). And with on the other side a τ̄ (M ) that “catch”
the exceptions in M by burning the abstraction context of M (whenever this
abstraction is dummy). The main difference with a usual exception system is
the divergence of the catch if no exception are raised.
We introduce a new operators and a new kind of expression that will play the
role of exception, the tests:
(terms) M, N ∶ τ̄ (Q)
(test)
ǫ ∣ Q∣R ∣ τ (M ) ∣ Q+R ∣ 0
Q, R ∶
New operator immediately imply new distribution rules for the sum and the
linear substitution:
τ (Σi Mi ) = Σi τ (Mi )
τ̄ (Σi Qi ) = Σi τ̄ (Qi )
∥j Σi Qi(j) = Σi ∥j Qi(j)
τ (M )⟨N /x⟩ = τ (M ⟨N /x⟩)
τ̄ (Q)⟨N /x⟩ = τ̄ (Q⟨N /x⟩)
(Q+R)⟨N /x⟩ = Q⟨N /x⟩+R⟨N /x⟩
(Q∣R)⟨N /x⟩ = Q⟨N /x⟩∣R⟨N /x⟩
Here is the corresponding operational semantics:
(γ)
τ [τ̄ (Q)]
(τ )
(τ̄1 )
τ (λx.M )
(τ̄ (Q)) [M ! ]
(τ̄2 )
(ǫ)
(τ̄ (Q)) [M1 , ..., Mn≥1 ; M ! ]
ǫ∣ǫ
→ Q
→ τ (M {0/x})
→ τ̄ (Q)
→ 0
→ ǫ
The intuition of τ̄ (Q) is an operator that take a test (a Boolean value), compute
it and returns an infinite λ-abstraction with no occurrence of the abstracted
variables. The test τ (M ) is taking a term and returns a successful test if the
term is converging in a context that consists of an infinite empty application.
4
2.3
Observational order and full abstraction
In order to ask for full abstraction, one has to specify a reduction strategy. A
natural choice would be the head reduction, but this would make (λx.x [0, 0]) a
normal form while no applicative instantiation of x allow the convergence of this
term. Therefore, the reduction strategy we are considering will not be headreduction, but the outer-head-reduction. This reduction will reduce subterms
in linear position after the subterms in head positions ([14]). The corresponding
normal forms are terms and tests of the form:
M + λx̄.y [N1,1 , ..., N1,k1 ; L!1 ] ⋯ [Nn,1 , ..., Nn,kn ; L!n ]
M + λx̄.τ̄ (Q)
Q + (∥τ (Ni ))
Where every N and Q must be in outer-head normal forms and can’t be a sum
(but the Li are of any kind).
Definition 1 M is observationally below N , if for all context CL.M, we have
CLN M which is outer-head-converging whenever CLM M is outer-head-converging.
They are observationally equivalent if moreover N is observationally below M
In the particular case of the τ∂λ-calculus , we can easily restrict contexts to
test-contexts, which is contexts whose output is a tests. This will be applied
systematically for simplification.
We will denote ≤τ ∂ and ≡τ ∂ the observational order and equivalence of the τ∂λcalculus and ≤∂ and ≡∂ those of the ∂λ-calculus .
Bucciarelli, Carraro, Ehrhard and Manzonetto were then able to prove a
strong theorem relating the model to the calculus:
Theorem 1 M∞ is fully abstract for the Λ-calculus with resources and tests:
for all closed terms M, N with resources ans tests,
JM K = JN K ⇔ M ≡τ ∂ N
3
The counter-example
In order to exhibit our counter-example we will use the following property:
Fact 1 Let B a calculus and A a super-calculus. Let M a model that is fully
abstract for A. M is fully abstract for B iff the operational equivalences of B
and A are equal on their domain intersection.
In our context it means that, in order to prove the non full abstraction for the
∂λ-calculus , it is sufficient to find two terms of the ∂λ-calculus that cannot be
separated by any context of the ∂λ-calculus but that are separated by a context
of the τ∂λ-calculus . This makes the research and the proof quite easier when
the terms of the ∂λ-calculus involved are complex but not the context of the
τ∂λ-calculus .
5
We are firstly exhibiting a term A of the ∂λ-calculus that is observationally above
the identity in the ∂λ-calculus ’, but not in the τ∂λ-calculus ’s observational order:
A = Θ [λuvw.w [I [v ! ]] , (λuvw.u [(v [w! ])! ])! ]
(1)
where Θ is the Turing fix point combinator:
(λgu.u[(g [(g [u! ])! ])! ]) [(λgu.u[(g [(g [u! ])! ])! ])! ]
This term seems quite complex, but, modulo η-equivalence, A reduces exactly to
any of the elements of the following sum, and thus can be think as an equivalent:
Σn≥1 Bn = Σn≥1 λūn w.w [I [u!1 ] [u!2 ] ⋯ [u!n ]]
(2)
This is due to the following property:
Lemma 1 If Ai = λx̄i+1 .A (x1 [x!2 ] ⋯ [x!i+1 ]) then A0 →η A and for all i,
Ai → Ai+1 + Bi+1
Proof. Simple reduction unfolding the Θ once.
◻
In absence of tests, this term has a comportment similar to ǫ0 in the sense that
it will converges in any applicative context provided that these applications do
not carry linear components. In particular it converges more often than the
identity:
Lemma 2 For all context CL.M of the ∂λ-calculus , if CLIM converges then CLAM
converges, i.e. I ≤∂ A
Proof. Let CL.M a context that converge on I
With the context lemma ([14]), and since neither I nor A has free variables, we
can assume that CL.M = L.M P1 ⋯ Pk , thus by lemma 1, we have A →∗ U + Bk
and
CLAM →∗ U ′ + λw.w [I P1 ⋯ Pk ] = U ′ + λw.w [CLM M]
converges.
◻
But in presence of real tests, its comportment appeared to be different that ǫ0
in the sense that it diverges under a τ̄ . In particular it is not observationally
above the identity in τ∂λ-calculus :
Lemma 3 In the τ∂λ-calculus , τ (A [ǫ!0 ]) diverges, while τ (I [ǫ!0 ]) converges,
i.e. I ≤/ τ ∂ A
Proof. For all i, τ (Ai [ǫ!0 ]) diverges since, by co-induction:
τ (Ai [ǫ0 ]) → τ ((Ai+1 + (λx̄i+1 y.y [I [x!1 ] ⋯ [x!i+1 ]])) [ǫ0 ])
=
→
τ (Ai+1 [ǫ0 ]) + τ ((λx̄i+1 y.y [I [x!1 ] ⋯ [x!i+1 ]]) [ǫ0 ])
τ (Ai+1 [ǫ0 ]) + τ (λx̄i y.y [ǫ0 [x!1 ] ⋯ [x!i+1 ]])
→∗
τ (Ai+1 [ǫ0 ]) + τ (0 [ǫ0 [0! ] ⋯ [0! ]])
The non outer-head convergence comes with the co-induction hypothesis for the
first term, and is trivial for the second.
◻
Hence, we have broken the conjecture concerning the equality between the observational and denotational orders. Let’s break the whole conjecture:
Theorem 2 M∞ is not fully abstract for the λ-calculus with resources
6
4
Further works
First a diligent reader will remark that we have a critical use of the demonic
sum which is very powerful in this calculus. And an even more diligent one will
remark that an arbitrary choice have been made concerning this sum: we could
differentiate terms and reduced of terms and remove sums from the original syntax (they just have to appear in reductions of terms). The choice we made here
corresponds to the one of [4] and carries an understandable counter-example.
But we claim that another equivalent counter-example arises for the case with
limited sum. This counter-example is, however, a little more complicated and
make it necessary to rework the material of [4] (even if everything works exactly
the same way).
Our counter-example can be translated to some related cases. In particular,
to prove non full abstraction of Scott’s D∞ for the λ-calculus with angelic and
demonic sums (conjectured in [7]). For this calculus the extension with tests
exists and is fully abstract for D∞ , this is a trivial modification of the tests
of [3] (using general demonic and angelic sums). In this framework the term
θ(λxy.x + y) plays exactly the role of A in our example with the same output.
In the end, from a unique object that is DLL, we exhibit two natural constructions, one in the semantical world, the other in the syntactical one, but
they appeared do not respect full abstraction. One would say that they are not
that natural and that more natural one may be found. But this would be to
easy, from the state of art we don’t known more natural construction. The misunderstanding comes with the concept of naturality, it seems that the syntactic
idea of “convergence” does not really correspond to the equivalent in semantical word. One being a lowest fix point and the second a largest one. This
difference appears when working with the demonic sum that allow to check the
convergence in unbounded applicative context.
Finally we presented tests as a general tool whose importance is above the
role we gave them here. This result is interesting and important as it presents
tests as useful tools to verify that full abstraction fails. But it remains a negative
result that does not justify alone any real interest for them. Further works will
then focus on presenting positive proofs of full abstractions that are using tests.
Following this way we already submitted a revisited proof of full abstraction of
the Scott’s D∞ for the usual λ-calculus [3].
References
[1] Gerard Boudol (1993): The lambda-calculus with multiplicities. INRIA
Research Report 2025.
[2] P. Boudol, P.-L. Curien & C. Lavatelli (1999): A semantics for lambda
calculi with resources. Mathematical Structures in Comput. Sci. (MSCS)
Vol. 9, pp. 437–482.
7
[3] Flavien Breuvart (2012): A new proof of the Hyland/Wadsworth full abstraction theorem. Submited.
[4] Antonio Bucciarelli, Alberto Carraro, Thomas Ehrhard & Giulio Manzonetto (2011): Full Abstraction for Resource Calculus with Tests. In
Marc Bezem, editor: Computer Science Logic (CSL’11) - 25th International
Workshop/20th Annual Conference of the EACSL, Leibniz International
Proceedings in Informatics (LIPIcs) 12, Schloss Dagstuhl–Leibniz-Zentrum
fuer Informatik, Dagstuhl, Germany, pp. 97–111.
[5] Antonio Bucciarelli, Alberto Carraro, Thomas Ehrhard & Giulio Manzonetto (2012): Full Abstraction for Resource Lambda Calculus with Tests,
throught Taylor expansion. Accepted.
[6] Antonio Bucciarelli, Thomas Ehrhard & Giulio Manzonetto (2007): Not
Enough Points Is Enough. In Jacques Duparc & Thomas A. Henzinger,
editors: CSL’07: Proceedings of 16th Computer Science Logic, Lecture
Notes in Computer Science 4646, Springer, pp. 268–282.
[7] Antonio Bucciarelli, Thomas Ehrhard & Giulio Manzonetto (2009): A
relational model of a parallel and non-deterministic lambda-calculus. In
Sergei N. Artëmov & Anil Nerode, editors: Logical Foundations of Computer Science, International Symposium, LFCS 2009, Lecture Notes in
Computer Science 5407, pp. 107–121.
[8] Daniel de Carvalho & Lorenzo Tortora de Falco (2010): The relational
model is injective for Multiplicative Exponential Linear Logic (without
weakenings). CoRR abs/1002.3131.
[9] T. Ehrhard & O. Laurent (2010): Interpreting a finitary pi-calculus in
differential interaction nets. Inf. Comput. 208(6), pp. 606–633.
[10] Thomas Ehrhard & Laurent Regnier (2004): The differential lambdacalculus. Theoretical Computer Science, Elsevier.
[11] A. J. Kfoury (2000): A linearization of the Lambda-calculus and
consequences.
J. Log. Comput. 10(3), pp. 411–436. Available at
http://dx.doi.org/10.1093/logcom/10.3.411.
[12] Giulio Manzonetto (2009): A general class of models of H⋆ . In: Mathematical Foundations of Computer Science (MFCS’09), Lecture Notes in
Computer Science 5734, Springer, pp. 574–586.
[13] M. Pagani & Tranquilli P. (2009): Parallel reduction in resource lambdacalculus. APLAS’09 5904 LNCS, pp. 226–242.
[14] Michele Pagani & Simona Ronchi Della Rocca (2010): Linearity, Nondeterminism and Solvability. Fundamenta Informaticae 103(1-4), pp. 173–
202. Available at http://dx.doi.org/10.3233/FI-2010-324.
8
A
A.1
The model M∞
Categorical model
The category Rel of sets and relations is known to be a model of linear logic
as it is a Seely category (we are giving the interpretation, but we will let the
comutative diagrams to the reader since their comutations are trivial or like):
It is monoidal with tensor functor given by A ⊗ B = A × B, f ⊗ g =
{((u, x), (v, y))∣(u, v) ∈ f, (x, y) ∈ g} and with the arbitrary unit 1 = {∗}. It
is symetric monoidal close with (A ⊸ B) = A × B if we take the evaluation
ev = {(((a, b), a), b)∣a ∈ A, b ∈ B} ∈ Rel((A ⊸ B) ⊗ A, B). So it is star autonomus with 1 as dualising object (for a trvial duality).
This give us the interpretation of multiplicatives: A⊗B ˘
= A`B = A ⊸ B = A×B
The category is cartesian, with catesian product
Ai = {(i, x)∣i ˘
∈ I, x ∈
si }, projections πi = {((i, a), a)∣a ∈ Ai } and product of morphisms
i fi =
{(b, (i, a))∣(b, a) ∈ fi }. The terminal object is ⊺ = ∅. ˘
This give us the interpretation of additives: ⊕i∈I Ai = i∈I Ai = {(i, a)∣i ∈ I, a ∈
Ai }
We can add a comonade (!, d, p) where the functor is define by !A = Mf (A),
!f = {([a1 , ..., ak ], [b1 , ..., bk )∣∀i ≤ k, (ai , bi ) ∈ f }, the deriliction by dA =
{([a], a)∣a ∈ A} and the digging by pA = {(m1 +⋯+mk , [m1 , ..., mk ])∣m1 , ..., mk ∈
MF (a)}.
This give us the interpretation of exponentials: !P =?P = Mf (P )
This is a Seely category and a model of linear logic since the isomorphismes
1 ≃ [] is trivial and !A⊗!B ≃!(A & B) is defined by ([a1 , ..., al ], [b1 , ..., br ]) ≃
[(1, a1 ), ..., (1, al ), (2, b1 ), ..., (2, br )].
But it can even be seen as a categorical model of differential linear logic. By
defining the co-dereliction natural transforamtion d¯A = {(a, [a])∣a ∈ A} ∈ A →
!A, we are fixing the contraction cA = {(l+r, (l, r))∣l, r ∈!A} the co-contraction
c̄A = {((l, r), l+r)∣l, r ∈!A}, the weakening wA = {([], ∗)} and the co-weakening
w̄A = {(∗, [])}. So that we can define the derivative ∂X = (id!X ⊗ dX ) ○ cX ∶!X →
!X ⊗X and the co-derivative ∂¯X = c̄X ○(id!X ⊗ d¯D ) ∶!X ⊗X →!X. This derivatives
are Taylor, i.e. if two morphisms f1 , f2 ∶!A → B are such that f1 ○ ∂¯A = f2 ○ ∂¯A
then f1 + (f2 ○ w̄X ○ wX ) = (f1 ○ w̄X ○ wX ) + f2 . Finally the exponential acept
anti-derivatives, since it is bi-comutative and JA = IdA + ∂¯A ∂A = IdA is an
isomorphism. For more detail about models of DLL see ??.
As for every categorical model of linear logic, the exponential is a comonade
and induced a coKleisly M Rel = Rel! whose objects are the set and whose
morphisms from P to Q are the relations between !P and Q. The identities are
the relations digP = {({x}, x)∣x ∈ P } and the composition f ○g = {(X, z)∣∃(Y, z) ∈
f, ∀y ∈ Y, (X, y) ∈ g}.
A.2
Algebraic model
In order to have an algebraic model of ∂λ-calculus we only need a reflexive
object, i.e. a triplet (M, app, abs) where M is an object of MRel, app ∶ (M →
9
(!M ⊸ M ) = (Mf (M ) × M )) and aps ∶ ((Mf (M ) × M ) → M ) such that
app ○ abs = Id. Such an object can a priory found by taking the lower fix
point of M ↦ M ⇒ M =?M ⊸ M = Mf (M ) × M . But this will just leads to
the trivial empty model. We will then resolve the more complicated fix point
M ↦ (?M )&N = N × Mf (M ) where the exponent represent an infinit tensor
product. The lower fix point will be called M∞ .
An other way to see the fixpoint is to say that M have to be equal to
the set of quazi everywhere empty lists of finite substets of itself. Its element
are the recursively defined as being either ∗, the list of empty elements, or
a∶∶α with a ∈ Mf (M∞ ) and α ∈ M∞ . The coresponding app and abs arise
imediatly from the functoriality: app = {(a∶∶α, (a, α))∣(a, α) ∈ M∞ } and abs =
{((a, α), a∶∶α)∣a, α ∈ M∞ }
In order to be understandable, we are presenting the interpretation of terms
via a type system with types living in M∞ . The usual presentation of the
interpretation can be recoverd from the type system:
JM Kx̄ = {(ā, α)∣x̄ ∶ ā ⊢ M ∶ α}
JQKx̄ = {ā∣x̄ ∶ ā ⊢ Q}
The type system is the following:
Γ, x ∶ v ⊢ M ∶ α
Γ⊢∆
Γ⊢M ∶α
Γ⊢N ∶α
x
∶
[α]
⊢
x
∶
α
Γ ⊢ M +N ∶ α
Γ ⊢ M +N ∶ α
x ∶ [], Γ ⊢ ∆
Γ ⊢ λx.M ∶ v∶∶α
⋀i≥n Γi ⊢ L ∶ βi
⋀j≤n Γj ⊢ Lj ∶ βj
Γ ⊢ M ∶ w∶∶α
Γ′ ⊢ B ∶ w
!
Γ + Γ′ ⊢ M B ∶ α
(Σn+m
r=1 Γr ) ⊢ [L1 , ..., Ln ; L ] ∶ [β1 , ..., βn+m ]
′
Γ⊢Q
Γ⊢Q
Γ ⊢R
Γ⊢M ∶∗
⊢ǫ
Γ ⊢ τ [M ]
Γ ⊢ τ̄ (Q) ∶ ∗
Γ + Γ′ ⊢ Q∣R
10
| 6cs.PL
|
arXiv:1603.02639v4 [math.MG] 8 Dec 2016
PLIABILITY, OR THE WHITNEY EXTENSION THEOREM FOR
CURVES IN CARNOT GROUPS
NICOLAS JUILLET AND MARIO SIGALOTTI
Abstract. The Whitney extension theorem is a classical result in analysis giving a necessary
and sufficient condition for a function defined on a closed set to be extendable to the whole
space with a given class of regularity. It has been adapted to several settings, among which
the one of Carnot groups. However, the target space has generally been assumed to be equal
to Rd for some d ≥ 1.
We focus here on the extendability problem for general ordered pairs (G1 , G2 ) (with G2
non-Abelian). We analyze in particular the case G1 = R and characterize the groups G2 for
which the Whitney extension property holds, in terms of a newly introduced notion that we
call pliability. Pliability happens to be related to rigidity as defined by Bryant an Hsu. We
exploit this relation in order to provide examples of non-pliable Carnot groups, that is, Carnot
groups so that the Whitney extension property does not hold. We use geometric control theory
results on the accessibility of control affine systems in order to test the pliability of a Carnot
group. In particular, we recover some recent results by Le Donne, Speight and Zimmermann
about Lusin approximation in Carnot groups of step 2 and Whitney extension in Heisenberg
groups. We extend such results to all pliable Carnot groups, and we show that the latter may
be of arbitrarily large step.
1. Introduction
Extending functions is a basic but fundamental tool in analysis. Fundamental is in particular
the extension theorem established by H. Whitney in 1934, which guarantees the existence of an
extension of a function defined on a closed set of a finite-dimensional vector space to a function
of class C k , provided that the minimal obstruction imposed by Taylor series is satisfied. The
Whitney extension theorem plays a significative part in the study of ideals of differentiable
functions (see [26]) and its variants are still an active research topic of classical analysis (see
for instance [11]).
Analysis on Carnot groups with a homogeneous distance like the Carnot–Carathéodory
distance, as presented in Folland and Stein’s monograph [13], is nowadays a classical topic too.
Carnot groups provide a generalization of finite-dimensional vector spaces that is both close to
the original model and radically different. This is why Carnot groups provide a wonderful field
of investigation in many branches of mathematics. Not only the setting is elegant and rich but
it is at the natural crossroad between different fields of mathematics, as for instance analysis
of PDEs or geometric control theory (see for instance [5] for a contemporary account). It is
2010 Mathematics Subject Classification. 22E25, 41A05, 53C17, 54C20, 58C25.
Key words and phrases. Whitney extension theorem, Carnot group, rigid curve, horizontal curve.
1
2
NICOLAS JUILLET AND MARIO SIGALOTTI
therefore natural to recast the Whitney extension theorem in the context of Carnot groups. As
far as we know, the first generalization of a Whitney extension theorem to Carnot groups can
be found in [14, 15], where De Giorgi’s result on sets of finite perimeter is adapted first to the
Heisenberg group and then to any Carnot group of step 2. This generalization is used in [19],
where the authors stress the difference between intrinsic regular hypersurfaces and classical C 1
hypersurfaces in the Heisenberg group. The recent paper [34] gives a final statement for the
Whitney extension theorem for scalar-valued functions on Carnot groups: The most natural
generalization that one can imagine holds in its full strength (for more details, see Section 2).
The study of the Whitney extension property for Carnot groups is however not closed. Following a suggestion by Serra Cassano in [29], one might consider maps between Carnot groups
instead of solely scalar-valued functions on Carnot groups. The new question presents richer
geometrical features and echoes classical topics of metric geometry. We think in particular of
the classification of Lipschitz embeddings for metric spaces and of the related question of the
extension of Lipschitz maps between metric spaces. We refer to [3, 36, 28, 4] for the corresponding results for the most usual Carnot groups: Abelian groups Rm or Heisenberg groups
Hn (of topological dimension 2n + 1). In view of Pansu–Rademacher theorem on Lipschitz
maps (see Theorem 2.1), the most directly related Whitney extension problem is the one for
1
CH
-maps, the so-called horizontal maps of class C 1 defined on Carnot groups. This is the
framework of our paper.
Simple pieces of argument show that the Whitney extension theorem does not generalize to
every ordered pair of Carnot groups. Basic facts in contact geometry suggest that the extension does not hold for (Rn+1 , Hn ), i.e., for maps from Rn+1 to Hn . It is actually known that
local algebraic constraints of first order make n the maximal dimension for a Legendrian submanifold in a contact manifold of dimension 2n + 1. In fact if the derivative of a differentiable
map has range in the kernel of the contact form, the range of the map has dimension at most
1
n. A map from Rn+1 to Hn is CH
if it is C 1 with horizontal derivatives, i.e., if its derivatives
1
take value in the kernel of the canonical contact form. In particular, a CH
-map defined on
n+1
R
is nowhere of maximal rank. Moreover, it is a consequence of the Pansu–Rademacher
theorem that a Lipschitz map from Rn+1 to Hn is derivable at almost every point with only
horizontal derivatives. Again n is their maximal rank. In order to contradict the extendability
of Lipschitz maps, it is enough to define a function on a subset whose topological constraints
force any possible extension to have maximal rank at some point. Let us sketch a concrete example that provides a constraint for the Lipschitz extension problem: It is known that Rn can
be isometrically embedded in Hn with the exponential map (for the Euclidean and Carnot–
Carathéodory distances). One can also consider two ‘parallel’ copies of Rn in Rn+1 mapped to
parallel images in Hn : the second is obtained from the first by a vertical translation. Aiming
for a contradiction, suppose that there exists an extending Lipschitz map F . It provides on
Rn × [0, 1] a Lipschitz homotopy between F (Rn × {0}) and F (Rn × {1}). Using the definition
of a Lipschitz map and some topology, the topological dimension of the range is at least n + 1
and its (n + 1)-Hausdorff measure is positive. This is not possible because of the dimensional
constraints explained above. See [3] for a more rigorous proof using a different set as a domain
for the function to be extended. The proof in [3] is formulated in terms of index theory and
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
3
purely (n + 1)-unrectifiability of Hn . The latter property means that the (n + 1)-Hausdorff
measure of the range of a Lipschitz map is zero. Probably this construction and some other
ideas from the works on the Lipschitz extension problem [3, 36, 28, 4] can be adapted to
the Whitney extension problem. It is not really our concern in the present article to list the
similarities between the two problems, but rather to exhibit a class of ordered pairs of Carnot
groups for which the validity of the Whitney extension problem depends on the geometry of
the groups. Note that a different type of counterexample to the Whitney extension theorem,
involving groups which are neither Euclidean spaces nor Heisenberg groups, has been obtained
by A. Khozhevnikov in [20]. It is described in Example 2.6.
Our work is motivated by F. Serra Cassano’s suggestion in his Paris’ lecture notes at the
Institut Henri Poincaré in 2014 [29]. He proposes — (i) to choose general Carnot groups G
1
as target space, (ii) to look at CH
curves only, i.e., C 1 maps from R to G with horizontal
derivatives. As we will see, the problem is very different from the Lipschitz extension problem
for (R, G) and from the Whitney extension problem for (G, R). Indeed, both such problems can
be solved for every G, while the answer to the extendibility question asked by Serra Cassano
depends on the choice of G. More precisely, we provide a geometric characterization of those
1
G for which the CH
-Whitney extension problem for (R, G) can always be solved. We say in
1
this case that the pair (R, G) has the CH
extension property. Examples of target non-Abelian
1
Carnot groups for which CH extendibility is possible have been identified by S. Zimmerman in
1
[39], where it is proved that for every n ∈ N the pair (R, Hn ) has the CH
extension property.
1
The main component of the characterization of Carnot groups G for which (R, G) has the CH
extension property is the notion of pliable horizontal vector. A horizontal vector X (identified
with a left-invariant vector field) is pliable if for every p ∈ G and every neighborhood Ω of
1
X in the horizontal layer of G, the support of all CH
curves with derivative in Ω starting
from p in the direction X form a neighborhood of the integral curve of X starting from p (for
details, see Definition 3.4 and Proposition 3.7). This notion is close but not equivalent to the
property of the integral curves of X not to be rigid in the sense introduced by Bryant and
Hsu in [9], as we illustrate by an example (Example 3.5). We say that a Carnot group G is
pliable if all its horizontal vectors are pliable. Since any rigid integral curve of a horizontal
vector X is not pliable, it is not hard to show that there exist non-pliable Carnot groups of
any dimension larger than 3 and of any step larger than 2 (see Example 3.3). On the other
hand, we give some criteria ensuring the pliability of a Carnot group, notably the fact that it
has step 2 (Theorem 6.5). We also prove the existence of pliable groups of any positive step
(Proposition 6.6).
Our main theorem is the following.
1
Theorem 1.1. The pair (R, G) has the CH
extension property if and only if G is pliable.
The paper is organized as follows: in Section 2 we recall some basic facts about Carnot
1
groups and we present the CH
-Whitney condition in the light of the Pansu–Rademacher Theorem. In Section 3 we introduce the notion of pliability, we discuss its relation with rigidity,
1
and we show that pliability of G is necessary for the CH
extension property to hold for (R, G)
(Theorem 3.8). The proof of this result goes by assuming that a non-pliable horizontal vector
4
NICOLAS JUILLET AND MARIO SIGALOTTI
1
exists and using it to provide an explicit construction of a CH
map defined on a closed subset
of R which cannot be extended on R. Section 4 is devoted to proving that pliability is also a
sufficient condition (Theorem 4.4). In Section 5 we use our result to extend some Lusin-like
theorem proved recently by G. Speight for Heisenberg groups [30] (see also [39] for an alternative proof). More precisely, it is proved in [22] that an absolutely continuous curve in a group
1
of step 2 coincides on a set of arbitrarily small complement with a CH
curve. We show that
this is the case for pliable Carnot groups (Proposition 5.2). Finally, in Section 6 we give some
criteria for testing the pliability of a Carnot group. We first show that the zero horizontal
vector is always pliable (Proposition 6.1). Then, by applying some results of control theory
providing criteria under which the endpoint mapping is open, we show that G is pliable if its
step is equal to 2.
2. Whitney condition in Carnot groups
A nilpotent Lie group G is said to be a Carnot group if it is stratified, in the sense that its
Lie algebra G admits a direct sum decomposition
G1 ⊕ · · · ⊕ Gs ,
called stratification, such that [Gi , Gj ] = Gi+j for every i, j ∈ N∗ with i + j ≤ s and [Gi , Gj ] =
{0} if i + j > s. We recall that [Gi , Gj ] denotes the linear space spanned by {[X, Y ] ∈ G |
X ∈ Gi , Y ∈ Gj }. The subspace G1 is called the horizontal layer and it is also denoted
by GH . We say that s is the step of G if Gs 6= {0}. The group product of two elements
x1 , x2 ∈ G is denoted by x1 · x2 . Given X ∈ G we write adX : G → G for the operator defined
by adX Y = [X, Y ].
The Lie algebra G can be identified with the family of left-invariant vector fields on G. The
exponential is the application that maps a vector X of G into the end-point at time 1 of the
integral curve of the vector field X starting from the identity of G, denoted by 0G . That is, if
γ(0) = 0G
γ̇(t) = X ◦ γ(t)
then γ(1) = exp(X). We also denote by etX : G → G the flow of the left-invariant vector field
X at time t. Notice that etX (p) = p · exp(tX). Integral curves of left-invariant vector fields
are said to be straight curves.
P
The Lie group G is diffeomorphic to RN with N = sk=1 dim(Gk ). A usual way to identify
G and RN through a global system of coordinates is to pull-back by exp the group structure
from G to G, where it can be expressed by the Baker–Campbell–Hausdorff formula. In this
way exp becomes a mapping of G = G onto itself that is simply the identity.
For any λ ∈ R we introduce the dilation ∆λ : G → G uniquely characterized by
∆λ ([X, Y ]) = [∆λ (X), ∆λ (Y )] for any X, Y ∈ G,
∆λ (X) = λX
for any X ∈ G1 .
Using the decomposition X = X1 + · · · + Xs with Xk ∈ Gk , it holds ∆λ (X) =
For any λ ∈ R we also define on G the dilation δλ = exp ◦∆λ ◦ exp−1 .
Ps
k=1
λk Xk .
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
5
Given an absolutely continuous curve γ : [a, b] → G, the velocity γ̇(t), which exists from
almost every t ∈ [a, b], is identified with the element of G whose associated left-invariant
vector field, evaluated at γ(t), is equal to γ̇(t). An absolutely continuous curve γ is said to
1
be horizontal if γ̇(t) ∈ GH for almost every t. For any interval I of R, we denote by CH
(I, G)
1
the space of all curves φ ∈ C (I, G) such that φ̇(t) ∈ GH for every t ∈ I.
Assume that the horizontal layer GH of the algebra is endowed with a quadratic norm
k · kGH . The Carnot–Carathéodory distance dG (p, q) between two points p, q ∈ G is then
defined as the minimal length of a horizontal curve connecting p and q, i.e.,
nZ b
dG (p, q) = inf
kγ̇(t)kGH dt |γ : [a, b] → G horizontal,
a
o
γ(a) = p, γ(b) = q .
Note that dG is left-invariant. It is known that dG provides the same topology as the usual
one on G. Moreover it is homogeneous, i.e., dG (δλ p, δλ q) = |λ|dG (p, q) for any λ ∈ R.
Observe that the Carnot–Carathéodory distance depends on the norm k · kGH considered
on GH . However all Carnot–Carathéodory distances are in fact metrically equivalent. They
are even equivalent with any left-invariant homogeneous distance [13] in a very similar way as
all norms on a finite-dimensional vector space are equivalent.
Notice that dG (p, ·) can be seen as the value function of the optimal control problem
m
X
γ̇ =
ui Xi (γ), (u1 , . . . , um ) ∈ Rm ,
i=1
γ(a) = p,
Z bp
u1 (t)2 + · · · + um (t)2 dt → min,
a
where X1 , . . . , Xm is a k · kGH -orthonormal basis of GH .
1
Finally, the space CH
([a, b], G) of horizontal curves of class C 1 can be endowed with a natural
1
C metric associated with (dG , k · kGH ) as follows: the distance between two curves γ1 and γ2
1
in CH
([a, b], G) is
!
max
sup dG (γ1 (t), γ2 (t)), sup kγ̇2 (t) − γ̇1 (t)kGH
t∈[a,b]
.
t∈[a,b]
In the following, we will write kγ̇2 − γ̇1 k∞,GH to denote the quantity supt∈[a,b] kγ̇2 (t) −
γ̇1 (t)kGH .
2.1. Whitney condition. A homogeneous homomorphism between two Carnot groups G1
and G2 is a group morphism L : G1 → G2 with L ◦ δλG1 = δλG2 ◦ L for any λ ∈ R. Moreover
L is a homogeneous homomorphism if and only if exp−1
G2 ◦L ◦ expG1 is a homogeneous Lie
algebra morphism. It is in particular a linear map on G1 identified with GG1 . The first layer
6
NICOLAS JUILLET AND MARIO SIGALOTTI
is mapped on the first layer so that a homogeneous homomorphism from R to G2 has the form
G2
L(t) = expG2 (tX), where X ∈ GH
.
Proposition 2.1 (Pansu–Rademacher Theorem). Let f be a locally Lipschitz map from an
open subset U of G1 into G2 . Then for almost every p ∈ U, there exists a homogeneous
homomorphism Lp such that
G2
G1 ∋ q 7→ δ1/r
(1)
f (p)−1 · f (p · δrG1 (q))
tends to Lp uniformly on every compact set K ⊂ G1 as r goes to zero.
Note that in Proposition 2.1 the map Lp is uniquely determined. It is called the Pansu
derivative of f at p and denoted by Dfp .
1
We denote by CH
(G1 , G2 ) the space of functions f such that (1) holds at every point p ∈ G1
and p 7→ Dfp is continuous for the usual topology. For G1 = R this coincides with the
1
definition of CH
(I, G2 ) given earlier. We have the following.
1
Proposition 2.2 (Taylor expansion). Let f ∈ CH
(G1 , G2 ) where G1 and G2 are Carnot
groups. Let K ⊂ G1 be compact. Then there exists a function ω from R+ to R+ with ω(t) =
o(t) at 0+ such that for any p, q ∈ K,
dG2 f (q), f (p) · Dfp (p−1 · q) ≤ ω(dG1 (p, q)),
where Dfp is the Pansu derivative.
Proof. This is a direct consequence the “mean value inequality” by Magnani contained in [25,
Theorem 1.2].
The above proposition hints at the suitable formulation of the C 1 -Whitney condition for
Carnot groups. This generalization already appeared in the literature in the paper [34] by
Vodop’yanov and Pupyshev.
1
Definition 2.3 (CH
-Whitney condition).
• Let K be a compact subset of G1 and consider f : K → G2 and a map L which associates with any p ∈ K a homogeneous group
1
homomorphism L(p). We say that the CH
-Whitney condition holds for (f, L) on K if
L is continuous and there exists a function ω from R+ to R+ with ω(t) = o(t) at 0+
such that for any p, q ∈ K,
(2)
dG2 f (q), f (p) · L(p)(p−1 · q) ≤ ω(dG1 (p, q)).
• Let K0 be a closed set of G1 , and f : K0 → G2 , and L such that K0 ∋ p 7→ L(p) is
1
continuous. We say that the CH
-Whitney condition holds for (f, L) on K0 if for any
compact set K ⊂ K0 it holds for the restriction of (f, L) to K.
1
Of course, according to Proposition 2.2, if f ∈ CH
(G1 , G2 ), then the restriction of (f, Df )
1
to any closed K0 satisfies the CH -Whitney condition on K0 .
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
7
In this paper we focus on the case G1 = R. The condition on a compact set K reads
rK,η → 0 as η → 0, where
(3)
rK,η =
dG2 (f (t), f (τ ) · exp[(t − τ )X(τ )])
,
|τ − t|
τ,t∈K, 0<|τ −t|<η
sup
2
because for every τ ∈ R one has [L(τ )](h) = exp(hX(τ )) for some X(τ ) ∈ GG
H and every
1
h ∈ R. With a slight abuse of terminology, we say that the CH -Whitney condition holds for
(f, X) on K.
In the classical setting the Whitney condition is equivalent to the existence of a C 1 map
f¯ : Rn1 → Rn2 such that f¯ and D f¯ have respectively restrictions f and L on K. This property
is usually known as the C 1 -Whitney extension theorem or simply Whitney extension theorem
(as for instance in [10]), even though the original theorem by Whitney is more general and in
particular includes higher order extensions [37, 38] and considers the extension f → f¯ as a
linear operator. This theorem is of broad use in analysis and is still the subject of dedicated
research. See for instance [8, 11, 12] and the references therein.
1
Definition 2.4. We say that the pair (G1 , G2 ) has the CH
extension property if for every
1
1
(f, L) satisfying the CH -Whitney condition on some closed set K0 there exists f¯ ∈ CH
(G1 , G2 )
¯
which extends f on G1 and such that D fp = L(p) for every p ∈ K0 .
1
We now state the CH
-extension theorem that Franchi, Serapioni, and Serra Cassano proved
in [15, Theorem 2.14]. It has been generalised by Vodop’yanov and Pupyshev in [34, 35] in a
form closer to the original Whitney’s result including higher order extensions and the linearity
of the operator f 7→ f¯.
Theorem 2.5 (Franchi, Serapioni, Serra Cassano). For any Carnot group G1 and any d ∈ N,
1
the pair (G1 , Rd ) has the CH
extension property.
The proof proposed by Franchi, Serapioni, and Serra Cassano is established for Carnot
groups of step two only, but is identical for general Carnot groups. It is inspired by the proof
in [10], that corresponds to the special case G1 = Rn1 for n1 ≥ 1.
Let us mention an example from the literature of non-extension with G1 6= R. This remarkable fact was explained to us by A. Kozhevnikov.
Example 2.6. If G1 and G2 are the ultrarigid Carnot groups of dimension 17 and 16 respectively, presented in [21] and analysed in Lemma A.2.1 of [20], one can construct an example
1
(f, L) satisfying the CH
-Whitney condition on some compact K without any possible exten1
¯
¯
sion (f , D f ) on G1 . For this, one exploits the rarity of CH
maps of maximal rank in ultrarigid
Carnot groups. The definition of ultrarigid from [21, Definition 3.1] is that all quasimorphisms
are Carnot similitudes, i.e., a composition of dilations and left-translations. We do not use
here directly the definition of ultrarigid groups but just the result stated in Lemma A.2.1 of
[20] for G1 and G2 . Concretely, let us set
K = {(p1 , . . . , p17 ) ∈ G1 | p2 = · · · = p16 = 0, p1 ∈ [−1, 1], p17 = p1 }.
8
NICOLAS JUILLET AND MARIO SIGALOTTI
Let the map f be constantly equal to 0 on K and L be the constant projection Λ : G1 ∋
(q1 , . . . , q17 ) 7→ (q1 , . . . , q16 ) ∈ G2 . Lemma A.2.1 in [20] applied at the point 0G1 implies that
the only possible extension of f is the projection L(0) = Λ. But this map vanishes only on
{p ∈ G1 | p1 = · · · = p16 = 0}, which does not contain K. It remains us to prove that
Whitney’s condition holds. In fact for two points p = (x, 0R15 , x) and q = (y, 0R15 , y) in K, we
look at the distance from f (x) = 0G2 to
f (p) · L(0)(p−1 · q) = L(0)((x, 0R15 , x)−1 · (y, 0R15 , y)) = (y − x, 0R15 )
on the one side and from p to q on the other side. The first one is |y −x|, up to a multiplicative
constant, and when |y − x| goes to zero the second one is c|y − x|1/3 for some constant c > 0.
1
This proves the CH
-Whitney condition for (f, L) on K.
In the present paper we provide examples of ordered pairs (G1 , G2 ) with G1 = R such that
1
the CH
extension property does or does not hold, depending on the geometry of G2 . We do
not address the problem of Whitney extensions for orders larger than 1. A preliminary step
for considering higher-order extensions would be to provide a suitable Taylor expansion for
m
CH
-functions from R to G2 , in the spirit of what recalled for m = 1 in Proposition 2.2.
1
extension property holds for some
Let us conclude the section by assuming that the CH
ordered pair (G1 , G2 ) of Carnot groups and by showing how to deduce it for other pairs. We
describe here below three such possible implications.
(1) Let S1 be a homogeneous subgroup of G1 that admits a complementary group K in the
sense of [29, Section 4.1.2]: both S1 and K are homogeneous Lie groups and the intersection is
reduced to {0}. Assume moreover that S1 is a Carnot group and K is normal, so that one can
define canonically a projection π : G1 → S1 that is a homogeneous homomorphism. Moreover
π is Lipschitz continuous (see [29, Proposition 4.13]). For the rest of the section, we say that
1
S1 is an appropriate Carnot subgroup of G1 . It can be easily proved that (S1 , G2 ) has the CH
1
extension property. In particular, according to [29, Example 4.6], for every k ≤ dim(GG
H )
k
k
1
the vector space R is an appropriate Carnot subgroup of G1 . Therefore (R , G2 ) has the CH
extension property.
(2) Assume now that S2 is an appropriate Carnot subgroup of G2 . Using the Lipschitz
1
continuity of the projection π : G2 → S2 , one easily deduces from the definition of CH
-Whitney
1
condition that (G1 , S2 ) has the CH extension property.
1
(3) Finally assume that (G1 , G′2 ) has the CH
extension property, where G′2 is a Carnot group.
Then one checks without difficulty that the same is true for (G1 , G2 × G′2 ).
As a consequence of Theorem 1.1, we can use these three implications to infer pliability
1
statements. Namely, a Carnot group G is pliable — i) if (G0 , G) has the CH
extension property
for some Carnot group G0 of positive dimension, ii) if G is the appropriate Carnot subgroup
of a pliable Carnot group, iii) if G is the product of two pliable Carnot groups.
1
3. Rigidity, Necessary condition for the CH
extension property
Let us first adapt to the case of horizontal curves on Carnot groups the notion of rigid curve
introduced by Bryant and Hsu in [9]. We will show in the following that the existence of rigid
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
9
1
curves in a Carnot group G can be used to identify obstructions to the validity of the CH
extension property for (R, G).
1
Definition 3.1 (Bryant, Hsu). Let γ ∈ CH
([a, b], G). We say that γ is rigid if there exists a
1
neighborhood V of γ in the space CH ([a, b], G) such that if β ∈ V and γ(a) = β(a), γ(b) = β(b)
then β is a reparametrization of γ.
A vector X ∈ GH is said to be rigid if the curve [0, 1] ∋ t 7→ exp(tX) is rigid.
A celebrated existence result of rigid curves for general sub-Riemannian manifolds has been
obtained by Bryant and Hsu in [9] and further improved in [23] and [2]. Examples of Carnot
groups with rigid curves have been illustrated in [16] and extended in [18], where it is shown
that, for any N ≥ 6 there exists a Carnot group of topological dimension N having rigid
curves. Nevertheless, such curves need not be straight. Actually, the construction proposed
in [18] produces curves which are necessarily not straight.
Following [2] (see also [27]), and focusing on rigid straight curves in Carnot groups, we
can formulate Theorem 3.2 below. In order to state it, let π : T ∗ G → G be the canonical
projection and recall that a curve p : I → T ∗ G is said to be an abnormal path if π ◦ p : I → G
is a horizontal curve, p(t) 6= 0 and p(t)X = 0 for every t ∈ I and X ∈ GH , and, moreover, for
every Y ∈ G and almost every t ∈ I,
d
p(t)Y = p(t)[Z(t), Y ],
dt
(4)
where Z(t) =
d
π
dt
◦ p(t) ∈ GH .
Theorem 3.2. Let X ∈ GH and assume that p : [0, 1] → T ∗ G is an abnormal path with
π ◦ p(t) = exp(tX).
If t 7→ exp(tX) is rigid, then p(t)[V, W ] = 0 for every V, W ∈ GH and every t ∈ [0, 1].
Moreover, denoting by Qp(t) the quadratic form Qp(t) (V ) = p(t)[V, [X, V ]] defined on {V ∈
GH | V ⊥ X}, we have that Qp(t) ≥ 0 for every t ∈ [0, 1].
Conversely, if p(t)[V, W ] = 0 for every V, W ∈ GH and every t ∈ [0, 1] and Qp(t) > 0 for
every t ∈ [0, 1] then t 7→ exp(tX) is rigid.
Example 3.3. An example of Carnot structure having rigid straight curves is the standard
Engel structure. In this case s = 3, dim G1 = 2, dim G2 = dim G3 = 1 and one can can pick
two generators X, Y of the horizontal distribution whose only nontrivial bracket relations are
[X, Y ] = W1 and [Y, W1 ] = W2 , where W1 and W2 span G2 and G3 respectively.
Let us illustrate how the existence of rigid straight curves can be deduced from Theorem 3.2
(one could also prove rigidity by direct computations of the same type as those of Example 3.5
below).
One immediately checks that p with p(t)X = p(t)Y = p(t)W1 = 0 and p(t)W2 = 1 is an
abnormal path such that π ◦ p(t) = exp(tX). The rigidity of t 7→ exp(tX) then follows from
Theorem 3.2, thanks to the relation Qp (Y ) = 1.
An extension of the previous construction can be used to exhibit, for every N ≥ 4, a Carnot
group of topological dimension N and step N − 1 having straight rigid curves. It suffices
10
NICOLAS JUILLET AND MARIO SIGALOTTI
to consider the N-dimensional Carnot group with Goursat distribution, that is, the group
such that dim G1 = 2, dim Gi = 1 for i = 2, . . . , N − 1, and there exist two generators
X, Y of G1 whose only nontrivial bracket relations are [X, Y ] = W1 and [Y, Wi ] = Wi+1 for
i = 1, . . . , N − 3, where Gi+1 = Span(Wi ) for i = 1, . . . , N − 2.
The following definition introduces the notion of pliable horizontal curve, in contrast to a
rigid one.
1
Definition 3.4. We say that a curve γ ∈ CH
([a, b], G) is pliable if for every neighborhood V
1
of γ in CH ([a, b], G) the set
{(β(b), β̇(b)) | β ∈ V, (β, β̇)(a) = (γ, γ̇)(a)}
is a neighborhood of (γ(b), γ̇(b)) in G × GH .
A vector X ∈ GH is said to be pliable if the curve [0, 1] ∋ t 7→ exp(tX) is pliable.
We say that G is pliable if every vector X ∈ GH is pliable.
By metric equivalence of all Carnot–Carathéodory distances, it follows that the pliability
of a horizontal vector does not depend on the norm k · kGH considered on GH .
1
Notice that, by definition of pliability, in every CH
neighborhood of a pliable curve γ :
[a, b] → G there exists a curve β with β(a) = γ(a), (β, β̇)(b) = (γ(b), W ), and W 6= γ̇(b). This
shows that pliable curves are not rigid. It should be noticed, however, that the converse is
not true in general, as will be discussed in Example 3.5. In this example we show that there
exist horizontal straight curves that are neither rigid nor pliable.
Example 3.5. We consider the 6-dimensional Carnot algebra G of step 3 that is spaned by
X, Y, Z, [X, Z], [Y, Z], [Y, [Y, Z]] where X, Y, Z is a basis of G1 and except from permutations
all brackets different from the ones above are zero.
According to [7, Chapter 4] there is a group structure on R6 with coordinates (x, y, z, z1 , z2 , z3 )
isomorphic to the corresponding Carnot group G such that the vectors of G1 are the leftinvariant vector fields
X = ∂x ,
Y = ∂y ,
Z = ∂z + x∂z1 + y∂z2 + y 2∂z3 .
Consider the straight curve [0, 1] ∋ t 7→ γ(t) = exp(tZ) ∈ G. First notice that γ is not
pliable, since for all horizontal curves in a small enough C 1 neighbourhood of γ the component
of the derivative along Z is positive, which implies that the coordinate z3 is nondecreasing.
No endpoint of a horizontal curve starting from 0G and belonging to a small enough C 1 neighbourhood of γ can have negative z3 component.
Let us now show that γ is not rigid either. Consider the solution β of
β̇(t) = Z(β(t)) + u(t)X(β(t)),
β(0) = 0G .
Notice that the y component of β is identically equal to zero. As a consequence, the same is
true
R t for the components
R t R τ z2 and z3 , while the x, z and z1 components of β(t) are, respectively,
u(τ )dτ , t, and 0 0 u(θ)dθdτ . In order to disprove the rigidity, it is then sufficient to take
0
R1
R1Rτ
a nontrivial continuous u : [0, 1] → R such that 0 u(τ )dτ = 0 = 0 0 u(θ)dθdτ .
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
11
Let us list some useful manipulations which transform horizontal curves into horizontal
curves. Let γ be a horizontal curve defined on [0, 1] and such that γ(0) = 0G .
(T1) For every λ > 0, the curve t ∈ [0, λ] 7→ δλ ◦ γ(λ−1 t) is horizontal and its velocity at
time t is γ̇(λ−1 t).
(T2) For every λ < 0, the curve t ∈ [0, |λ|] 7→ δλ ◦ γ(|λ|−1t) is horizontal and its velocity at
time t is −γ̇(|λ|−1 t).
(T3) The curve γ̄ defined by γ̄(t) = γ(1)−1 · γ(1 − t) is horizontal. It starts in 0G and finishes
in γ −1 (1). Its velocity at time t is −γ̇(1 − t).
(T4) If one composes the (commuting) transformations (T2) with λ = −1 and (T3), one
obtains a curve with derivative γ̇(1 − t) at time t.
(T5) It is possible to define the concatenation of two curves γ1 : [0, t1 ] → G and γ2 : [0, t2 ] →
G both starting from 0G as follows: the concatenated curve γ̃ : [0, t1 + t2 ] → G satisfies
γ̃(0) = 0G , has the same velocity as γ1 on [0, t1 ] and the velocity of γ2 (· − t1 ) on
[t1 , t1 + t2 ]. We have γ̃(t1 + t2 ) = γ1 (t1 ) · γ(t2 ) as a consequence of the invariance of
the the Lie algebra for the left-translation.
A consequence of (T1) and (T2) is that X ∈ GH \ {0} is rigid if and only if λX is rigid
for every λ ∈ R \ {0}. Similarly, X ∈ GH is pliable if and only if λX is pliable for every
λ ∈ R \ {0}.
Proposition 3.7 below gives a characterization of pliable horizontal vectors in terms of a
condition which is apriori easier to check than the one appearing in Definition 3.4. Before
proving the proposition, let us give a technical lemma. From now on, we write BG (x, r) to
denote the ball of center x and radius r in G for the distance dG and, similarly, BGH (x, r) to
denote the ball of center x and radius r in GH for the norm k · kGH .
Lemma 3.6. For any x ∈ G and 0 < r < R, there exists ε > 0 such that if y, z ∈ G and
ρ ≥ 0 satisfy dG (y, 0G ), dG (z, 0G ), ρ ≤ ε, then
BG (x, r) ⊂ y · δ1−ρ (BG (x, R)) · z.
Proof. Assume, by contradiction, that for every n ∈ N there exist xn ∈ BG (x, r), yn , zn ∈
BG (0G , 1/n) and ρn ∈ [0, 1/n] such that
xn 6∈ yn · δ1−ρn (BG (x, R)) · zn .
Equivalently,
δ(1−ρn )−1 (yn−1 · xn · zn−1 ) 6∈ BG (x, R).
However, lim supn→∞ dG (x, δ(1−ρn )−1 (yn−1 · xn · zn−1 )) ≤ r, leading to a contradiction.
Proposition 3.7. A vector V ∈ GH is pliable if and only if for every neighborhood V of the
1
curve [0, 1] ∋ t 7→ exp(tV ) in the space CH
([0, 1], G), the set
{β(1) | β ∈ V, (β, β̇)(0) = (0G , V )}
is a neighborhood of exp(V ).
12
NICOLAS JUILLET AND MARIO SIGALOTTI
1
Proof. Let F : CH
([0, 1], G) ∋ β 7→ (β, β̇)(1) ∈ G × GH and denote by π : G × GH → G the
canonical projection.
One direction of the equivalence being trivial, let us take ε > 0 and assume that π ◦ F (Uε )
is a neighbourhood of exp(V ) in G, where
1
Uε = {β ∈ CH
([0, 1], G) | (β, β̇)(0) = (0G , V ), kβ̇ − V k∞,GH < ε}.
We should prove that F (Uε ) is a neighborhood of (exp(V ), V ) in G × GH .
Step 1: As an intermediate step, we first prove that there exists η > 0 such that BG (exp(V ), η)×
{V } is contained in F (Uε ).
Let ρ be a real parameter in (0, 1). Using the transformations among horizontal curves
1
described earlier in this section, let us define a map Tρ : Uε → CH
([0, 1], G) associating with
a curve γ ∈ Uε the concatenation (transformation (T5)) of γ1 : t 7→ δρ ◦ γ(ρ−1 t) on [0, ρ]
obtained by transformation (T1) and a curve γ2 defined as follows. Consider γ2,1 : [0, 1 − ρ] ∋
t 7→ δ1−ρ ◦ γ((1 − ρ)−1 t) (again (T1)). The curve γ2 is defined from γ2,1 by
γ2 (t) = γ1 (ρ) · γ2,1 (1 − ρ)−1 · δ−1 ◦ γ2,1 ((1 − ρ) − t)
(see transformation (T4)). The derivative of Tρ (γ) at time t ∈ [0, ρ) is γ̇(ρ−1 t). Its derivative
at time ρ + t is γ̇(1 − (1 − ρ)−1 t) for t ∈ (0, 1 − ρ]. Hence Tρ (γ) is continuous and has derivative
1
γ̇(1) at limit times ρ− and ρ+ , i.e., is a well-defined map from Uε into CH
([0, 1], G). Moreover,
Tρ (γ) has the same derivative V = γ̇(0) at times 0 and 1 and its derivative at any time in
[0, 1] is in the set of the derivatives of γ. In particular, Tρ (Uε ) ⊂ Uε .
Notice now that, by construction, the endpoint Tρ (γ)(1) of the curve Tρ (γ) is a function of
γ(1) and ρ only. It is actually equal to
Fρ (x) = δρ (x) · δρ−1 (x)−1 ,
where x = γ(1) (see (T1) and (T4)) . Let x0 = exp(V ) and γ0 : t 7→ exp(tV ). We have
Fρ (x0 ) = x0 because Tρ (γ0 ) = γ0 , both curves having derivative constantly equal to V . We
prove now that for ρ close enough to 1, the differential of Fρ at x0 is invertible. Let us use
the coordinate identification of G with RN . For every y ∈ G, the limits of δρ (y) and δ1−ρ (y)
as ρ tends to 1 are y and 0G respectively, while Dδρ (y) and Dδρ−1 (y) converge to Id and
0 respectively. One can check (see, e.g., [7, Proposition 2.2.22]) that the inverse function
has derivative −Id at 0G . Finally the left and right translations are global diffeomorphisms.
Collecting these informations and applying the chain rule, we get that DFρ (x0 ) tends to an
invertible operator as ρ goes to 1. Hence for ρ great enough, Fρ (x0 ) is a local diffeomorphism.
We know by assumption on V that, for any ε > 0, the endpoints of the curves of Uε form a
neighborhood of x0 . We have shown that this is also the case if we replace Uε by Tρ (Uε ), for ρ
close to 1. The curves of Tρ (Uε ) are in Uε and have, moreover, derivative V at time 1. He have
thus proved that for every ε > 0 there exists η > 0 such that BG (x0 , η) × {V } is contained in
F (Uε ).
Step 2: Let us now prove that F (Uε ) is a neighborhood of (x0 , V ) in G × GH .
Let β be a curve in Uε with β̇(1) = V and consider for every W ∈ BGH (V, ε) and every ρ ∈
(0, 1) the curve αρ,W defined as follows: αρ,W = δ1−ρ ◦β((1−ρ)−1 t) on [0, 1−ρ] (transformation
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
13
(T1)) and α̇ρ,W is the linear interpolation between V and W on [1 − ρ, 1]. Notice that αρ,W is
in Uε .
Let u ∈ G be the endpoint at time ρ of the curve in G starting at 0G whose derivative is the
linear interpolation between V and W on [0, ρ]. Then (αρ,W , α̇ρ,W )(1) = (δ1−ρ (β(1)) · u, W )
and u depends only on V , ρ and W , and not on the curve β. Moreover, u tends to 0G as ρ
goes to 1, uniformly with respect to W ∈ BGH (V, ε). Lemma 3.6 implies that for ρ sufficiently
close to 1, for every W ∈ BGH (V, ε), it holds δ1−ρ (BG (x0 , η)) · u ⊃ BG (x0 , η/2). We proved
that BG (x0 , η/2) × BGH (V, ε) ⊂ F (Uε ), concluding the proof of the proposition.
The main result of this section is the following theorem, which constitutes the necessity
1
part of the characterization of CH
extendability stated in Theorem 1.1.
1
Theorem 3.8. Let G be a Carnot group. If (R, G) has the CH
extension property, then G is
pliable.
Proof. Suppose, by contradiction, that there exists V ∈ GH which is not pliable. We are going
1
to prove that (R, G) has not the CH
extension property.
Let γ(t) = exp(tV ) for t ∈ [0, 1]. Since V is not pliable, it follows from Proposition 3.7
1
that there exist a neighborhood V of γ in the space CH
([0, 1], G) and a sequence (xn )n≥1
converging to 0G such that for every n ≥ 1 no curve β in V satisfies (β(0), β̇(0)) = (0G , V )
and β(1) = γ(1) · xn . In particular, there exists a neighborhood Ω of V in GH such that for
1
every β ∈ CH
([0, 1], G) with (β(0), β̇(0)) = (0G , V ) and
β̇(t) ∈ Ω,
∀t ∈ [0, 1],
we have (β(1), β̇(1)) 6= (γ(1) · xn , V ) for every n ∈ N. Since limn→∞ xn = 0G , we can assume
without loss of generality that, for every n ≥ 1,
(5)
max{d(δρ (xn ) · exp(tV ), exp(tV )) | ρ ∈ [0, 1], t ∈ [−1, 1]} ≤ 2−n .
By homogeneity and left-invariance, we deduce that for every y ∈ G and every ρ > 0, for
every β ∈ C 1 ([0, ρ], G) with (β(0), β̇(0)) = (y, V ) and
β̇(t) ∈ Ω,
∀t ∈ [0, ρ],
we have (β(ρ), β̇(ρ)) 6= (y · γ(ρ) · δρ (xn ), V ) for every n ∈ N.
1
1
Define ρn = n1 − n+1
= n(n+1)
and x̃n = δρn (xn ) for every n ∈ N. It follows from (5) that
(6)
max{d(x̃n · exp(tV ), exp(tV )) | t ∈ [−1, 1]} ≤ 2−n ,
∀n ≥ 1.
We introduce the sequence defined recursively by y0 = 0G and
(7)
yn+1 = yn · γ(ρn ) · x̃n .
Notice that (yn )n≥1 is a Cauchy sequence and denote by y∞ its limit as n → ∞.
1
By construction, for every n ∈ N and every β ∈ CH
([0, ρn ], G) with (β(0), β̇(0)) = (yn , V )
and β̇(t) ∈ Ω for all t ∈ [0, ρn ], we have (β(ρn ), β̇(ρn )) 6= (yn+1 , V ). The proof that the (R, G)
14
NICOLAS JUILLET AND MARIO SIGALOTTI
1
1
has not the CH
extension property is then concluded if we show that the CH
-Whitney condition
holds for (f, X) on K, where
[
1
∞
{1},
K = ∪n=1 1 −
n
and f : K → G and X : K → GH are defined by
f (1 − n−1 ) = yn ,
X(1 − n−1 ) = V,
n ∈ N∗ ∪ {∞}.
For i, j ∈ N∗ ∪ {∞}, let
D(i, j) = dG (f (1 − i−1 ), f (1 − j −1 ) · exp[(j −1 − i−1 )X(1 − j −1 )])
= dG (yi , yj · exp[(j −1 − i−1 )V ]).
We have to prove that
D(i, j) = o(j −1 − i−1 )
as i, j → ∞, that is, for every ε > 0, there exists iε ∈ N∗ such that D(i, j) < ε|j −1 − i−1 | for
i, j ∈ N∗ ∪ {∞} with i, j > iε .
By triangular inequality we have
max(i,j)−1
D(i, j) ≤
X
dG (yk+1 · exp[((k + 1)−1 − i−1 )V ], yk · exp[(k −1 − i−1 )V ]).
k=min(i,j)
Notice that
dG (yk+1 · exp[((k + 1)−1 − i−1 )V ], yk · exp[(k −1 − i−1 )V ])
= dG (yk+1 · exp[((k + 1)−1 − i−1 )V ], [yk · γ(ρk )] · exp[((k + 1)−1 − i−1 )V ])
= dG (x̃k · exp[((k + 1)−1 − i−1 )V ], exp[((k + 1)−1 − i−1 )V ]),
where the last equality follows from (7) and the invariance of dG by left-multiplication. Thanks
to (6), one then concludes that
dG (yk+1 · exp[((k + 1)−1 − i−1 )V ], yk · exp[(k −1 − i−1 )V ]) ≤ 2−k .
Pmax(i,j)−1
Hence, D(i, j) ≤ k=min(i,j) 2−k = o(j −1 −i−1 ) and this concludes the proof of Theorem 3.8.
1
4. Sufficient condition for the CH
extension property
We have seen in the previous section that, differently from the classical case, for a general
Carnot group G the suitable Whitney condition for (f, X) on K is not sufficient for the
existence of an extension (f, f˙) of (f, X) on R. More precisely, it follows from Theorem 3.8
that if G has horizontal vectors which are not pliable, then there exist triples (K, f, X) such
1
1
that the CH
-Whitney condition holds for (f, X) on K but there is not a CH
-extension of (f, X).
1
In this next section we prove the converse to the result above, showing that the CH
extension
property holds when all horizontal vectors are pliable, i.e., when G is pliable.
We start by introducing the notion of locally uniformly pliable horizontal vector.
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
15
Definition 4.1. A horizontal vector X is called locally uniformly pliable if there exists a
neighborhood U of X in GH such that for every ε > 0, there exists η > 0 so that for every
W ∈U
1
{(γ, γ̇)(1) | γ ∈ CH
([0, 1], G), (γ, γ̇)(0) = (0G , W ), kγ̇ − W k∞,GH ≤ ε}
⊃ BG (exp(W ), η) × BGH (W, η).
Remark 4.2. As it happens for pliability, if X is locally uniformly pliable then, for every
λ ∈ R \ {0}, λX is locally uniformly pliable.
We are going to see in the following (Remark 6.2) that pliability and local uniform pliability
are not equivalent properties. The following proposition, however, establishes the equivalence
between pliability and local uniform pliability of all horizontal vectors.
Proposition 4.3. If G is pliable, then all horizontal vectors are locally uniformly pliable.
Proof. Assume that G is pliable. For every V ∈ GH and ε > 0 denote by η(V, ε) a positive
constant such that
1
{(γ, γ̇)(1) | γ ∈ CH
([0, 1], G), (γ, γ̇)(0) = (0G , V ), kγ̇ − V k∞,GH ≤ ε}
⊃ BG (exp(V ), η(V, ε)) × BGH (V, η(V, ε)).
We are going to show that there exists ν(V, ε) > 0 such that for every W ∈ BGH (V, ν(V, ε))
(8)
1
{(γ, γ̇)(1) | γ ∈CH
([0, 1], G), (γ, γ̇)(0) = (0G , W ), kγ̇ − W k∞,GH ≤ ε}
!
!
η V, 2ε
η V, 2ε
× BGH W,
.
⊃ BG exp(W ),
4
4
The proof of the local uniform pliability of any horizontal vector X is then concluded by
simple compactness arguments (taking any compact neighborhood U of X, using the notation
of Definition 4.1).
First fix ν̄(V, ε) > 0 in such a way that
exp(W ) ∈ BG (exp(V ), η(V, ε/2)/4)
for every W ∈ BGH (V, ν̄(V, ε)).
1
For every W ∈ GH , every ρ ∈ (0, 1), and every curve γ ∈ CH
([0, 1], G) such that (γ, γ̇)(0) =
1
(0G , V ), define γW,ρ ∈ CH ([0, 1], G) as follows: γW,ρ(0) = 0G , γ̇W,ρ(t) = (t/ρ)V + ((ρ − t)/ρ)W
for t ∈ [0, ρ], γ̇W,ρ(ρ + (1 − ρ)t) = γ̇(t) for t ∈ [0, 1]. In particular
γW,ρ(1) = γW,ρ(ρ) · δ1−ρ (γ(1)),
γ̇W,ρ (1) = γ̇(1),
and
If kV − W kGH
kγ̇W,ρ − W k∞,GH ≤ kγ̇ − V k∞,GH + kW − V kGH .
≤ ε/2, we then have
ε
kγ̇W,ρ − W k∞,GH ≤ ε ∀γ such that kγ̇ − V k∞,GH ≤ .
2
16
NICOLAS JUILLET AND MARIO SIGALOTTI
Since γW,ρ(ρ) depends on V, W , and ρ, but not on γ, we conclude that, for every W ∈
BGH (V, ε/2),
1
{(β,β̇)(1) | β ∈ CH
([0, 1], G), (β, β̇)(0) = (0G , W ), kβ̇ − W k∞,GH ≤ ε}
ε
ε
× BGH V, η V,
.
⊃ γW,ρ(ρ) · δ1−ρ BG exp(V ), η V,
2
2
Notice that dG (0G , γW,ρ(ρ)) ≤ ρ max(kV kGH , kW kGH ). Thanks to Lemma 3.6, for ρ sufficiently small,
!
ε
η V, 2ε
γW,ρ(ρ) · δ1−ρ BG exp(V ), η V,
⊃ BG exp(V ),
.
2
2
Now,
BG
η V, 2ε
exp(V ),
2
!
⊃ BG
whenever W ∈ BGH (V, ν̄(V, ε)).
Similarly,
ε
⊃ BGH
BGH V, η V,
2
η V, 2ε
exp(W ),
4
η V, 2ε
W,
4
!
!
,
provided that kV − W kGH ≤ 3η(V, ε/2)/4. The proof of (8) is concluded by taking ν(V, ε) =
min(ν̄(V, ε), ε/2, 3η(V, ε/2)/4).
We are now ready to prove the converse of Theorem 3.8, concluding the proof of Theorem 1.1.
1
Theorem 4.4. Let G be a pliable Carnot group. Then (R, G) has the CH
extension property.
Proof. By Proposition 4.3, we can assume that all vectors in GH are locally uniformly pliable.
Note moreover that it is enough to prove the extension for maps defined on compact sets K.
The generalisation to closed sets K0 is immediate because the source Carnot group is R. Let
1
(f, X) satisfy the CH
-Whitney condition on K where K is compact. We have to define f¯
on the complementary (open) set R \ K, which is the countable and disjoint union of open
intervals. For the unbounded components of R \ K, we simply define f¯ as the curve with
constant speed X(i) or X(j) where i = min(K) and j = max(K). For the finite components
(a, b) we proceed as follows. We consider y = δ1/(b−a) (f (a)−1 · f (b)). We let ε be the smallest
number such that
1
{(γ, γ̇)(1) | γ ∈ CH
([0, 1], G), (γ, γ̇)(0) = (0G , X(a)), kγ̇ − X(a)k∞,GH ≤ ε′ }
1
contains (y, X(b)) for every ε′ > ε. We consider an extension f¯ ∈ CH
of f on [a, b] such that
˙
˙
˙
1
f¯(a) = X(a), f¯(b) = X(b), and kf¯ − X(a)k∞,GH ≤ 2ε. By definition of the CH
-Whitney
+
+
condition, there exists a function ω : R → R tending to 0 at 0 such that R(a, b) =
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
17
dG (f (b), f (a)·exp[(b−a)X(a)]) is smaller than ω(b−a)(b−a) and kX(b)−X(a)kGH ≤ ω(b−a).
Since R(a, b) is equal to (b − a)dG (exp(X(a)), y), we can conclude that
(9)
dG (exp(X(a)), y) ≤ ω(b − a).
Using the corresponding estimates for R(b, a), we deduce that
(10)
dG (exp(−X(b)), y −1 ) ≤ ω(b − a).
1
By construction f¯ extends f and f¯˙ = X on the interior of K. We prove now that f¯ is CH
and
˙
1
¯
¯
that f = X on the boundary ∂K of K. It is clear that f is CH on R \ ∂K. In order to conclude
the proof we are left to pick x ∈ ∂K, let xn tend to x, and we must show that f¯(xn ) and f¯˙(xn )
tend to f (x) and X(x) respectively. As f and X are continuous on K, we can assume without
loss of generality that each xn is in R \ K. Assume for now that xn < x for every n. The
connected component (an , bn ) of R \ K containing xn is either constant for n large (in this case
x = bn ) or its length goes to zero as n → ∞. In the first case we simply notice that f¯|[an ,bn ] is
C 1 by construction. In the second case we can assume that an < xn < bn and bn − an goes to
zero. As f and X are continuous, f (an ) and X(an ) converge to f (x) and X(x) respectively.
Inequality (9) guarantees that dG (exp(X(an )), δ1/(bn −an ) [f (an )−1 · f (bn )]) ≤ ω(bn − an ) → 0 as
n → ∞ and the local uniform pliability of X(x) implies that kf¯˙|[an ,bn ] − X(an )k∞,GH goes to
zero as n → ∞. It follows that kf¯˙(xn ) − X(an )kGH and dG (f¯(xn ), f (an )) go to zero, proving
that f¯(xn ) and f¯˙(xn ) tend to f (x) and X(x) respectively. The situation where xn > x for
infinitely many n can be handled similarly replacing (9) by (10).
5. Application to the Lusin approximation of an absolutely continuous
curve
In a recent paper, E. Le Donne and G. Speight prove the following result ([22, Theorem
1.2]).
Proposition 5.1 (Le Donne–Speight). Let G be a Carnot group of step 2 and consider a
1
horizontal curve γ : [a, b] → G. For any ε > 0, there exist K ⊂ [a, b] and a CH
-curve
γ1 : [a, b] → G such that L([a, b] \ K) < ε and γ = γ1 on K.
In the case in which G is equal to the n-th Heisenberg group Hn , such result had already been
proved in [30, Theorem 2.2] (see also [39, Corollary 3.8]). In [30] G. Speight also identifies a
horizontal curve on the Engel group such that the statement of Proposition 5.1 is not satisfied
([30, Theorem 3.2]).
The name “Lusin approximation” for the property stated in Proposition 5.1 comes from
the use of the classical theorem of Lusin [24] in the proof. Let us sketch a proof when G is
replaced by a vector space Rn . The derivative γ̇ of an absolutely continuous curve γ is an
integrable function. Lusin’s theorem states that γ̇ coincides with a continuous vector-valued
function X : K → Rn on a set K of measure arbitrarily close to b − a. Thanks to the inner
continuity of the Lebesgue measure, one can assume that K is compact. Moreover K can be
18
NICOLAS JUILLET AND MARIO SIGALOTTI
chosen so that the Whitney condition is satisfied by (γ|K , X) on K. This is a consequence of
the mean value inequality
(11)
kγ(x + h) − γ(x) − hγ̇(x)k ≤ o(h),
where o(h) depends on x ∈ K. By usual arguments of measure theory, inequality (11) can
be made uniform with respect to x if one slightly reduces the measure of K. The (classical)
Whitney extension theorem provides a C 1 -curve γ1 defined on [a, b] with γ1 = γ and γ˙1 = X
on K.
The proof in [22] (and also in [30]) follows the same scheme as the one sketched above. We
show here below how the same scheme can be adapted to any pliable Carnot group. The fact
that all Carnot groups of step 2 are pliable and that not all pliable Carnot groups are of step
1 or 2 is proved in the next section (Theorem 6.5 and Proposition 6.6), so that our paper
actually provides a nontrivial generalization of Proposition 5.1. The novelty of our approach
with respect to those in [22, 30, 39] is to replace the classical Rademacher differentiablility
theorem for Lipschitz or absolutely continuous curves from R to Rn by the more adapted
Pansu–Rademacher theorem.
Proposition 5.2 (Lusin approximation of a horizontal curve). Let G be a pliable Carnot
group and γ : [a, b] → G be a horizontal curve. Then for any ε > 0 there exist K ⊂ [a, b] with
1
L([a, b] \ K) < ε and a curve γ1 : [a, b] → G of class CH
such that the curves γ and γ1 coincide
on K.
Proof. We are going to prove that for any ε > 0 there exists a compact set K ⊂ [a, b] with
L([a, b] \ K) < ε such the three following conditions are satisfied:
(1) γ̇(t) exists and it is a horizontal vector at every t ∈ K;
(2) γ̇|K is uniformly continuous;
(3) For every ε > 0 there exists η > 0 such that, for every t ∈ K and |h| ≤ η with
t + h ∈ [a, b], it holds dG (γ(t + h), γ(t) · exp(hγ̇(t))) ≤ ηε.
1
With these conditions the CH
-Whitney condition holds for (γ, γ̇|K ) on K. Since G is pliable,
1
according to Theorem 4.4 the CH
extension property holds for (R, G), yielding γ1 as in the
statement of Proposition 5.2.
Case 1: γ is Lipschitz continuous. Let γ be a Lipschitz curve from [a, b] to G. The Pansu–
Rademacher theorem (Proposition 2.1) states that there exists A ⊂ [a, b] of full measure such
that, for any t ∈ A, the curve γ admits a derivative at t and it holds
dG (γ(t + h), γ(t) · exp(hγ̇(t))) = o(h),
as h goes to zero. Let ε be positive. By Lusin’s theorem, one can restrict A to a compact set
K1 ⊂ A such that t 7→ γ̇(t) is uniformly continuous on K1 and L(A \ K1 ) < ε/2. Moreover by
classical arguments of measure theory, the functions h 7→ |h|−1 dG (γ(t + h), γ(t) · exp(hγ̇(t0 )))
can be bounded by a function that is o(1) as h goes to zero, uniformly in t on some compact
set K2 with L(A \ K2 ) < ε/2 . In other words for every ε > 0 there exists η such that for
t ∈ K2 and h ∈ [t − η, t + η] it holds
dG (γ(t + h), γ(t) · exp(hγ̇(t))) ≤ ε|h|.
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
19
With K = K1 ∩ K2 , the three conditions (1), (2), (3) listed above hold true.
Case 2: γ general horizontal curve. Let γ be absolutely continuous on [a, b]. It admits a
pathlength parametrisation, i.e., there exists a Lipschitz continuous curve ϕ : [0, T ] → G and
a function F : [a, b] → [0, T ], absolutely continuous and non-decreasing, such that γ = ϕ ◦ F .
Moreover ϕ̇ has norm 1 at almost every time. As F is absolutely continuous, for every ε > 0
there exists η such that, for any measurable K, the inequality L([0, T ] \ K) < η implies
L([a, b] \ F −1 (K)) < ε.
Let ε be positive and let η be a number corresponding to ε/2 in the previous sentence.
Applying to F the scheme of proof sketched after Proposition 5.1 for n = 1, there exists
a compact set KF ⊂ [a, b] with L([a, b] \ KF ) < ε/2 such that F is differentiable with a
continuous derivative on KF and the bound in the mean value inequality is uniform on KF .
For the Lipschitz curve ϕ and for every η > 0, Case 1 provides a compact set Kϕ ⊂ [0, T ]
with the listed properties with ε/2 in place ε.
Let K be the compact KF ∩ F −1 (Kϕ ) and note that L([a, b] \ K) < ε. For t ∈ K it holds
|F (t + h) − F (t) − hF ′ (t)| = o(h)
and
dG (ϕ(F (t) + H), ϕ(F (t)) · exp(H ϕ̇(F (t)))) = o(H),
as h and H go to zero, uniformly with respect to t ∈ K. We also know that t 7→ F ′ (t) and
t 7→ ϕ̇(F (t)) ∈ GH exist and are continuous on K. It is a simple exercise to compose the two
Taylor expansions and obtain the wanted conditions for γ = ϕ ◦ F . Note that the derivative
of γ on K is F ′ (t)ϕ̇(F (t)), which is continuous on K.
Remark 5.3. A set E ⊂ Rn is said 1-countably rectifiable if there exists a countable family
of Lipschitz curves fk : R → Rn such that
!
[
H1 E \ fk (R) = 0.
k
The usual Lusin approximation of curves in Rn permits one to replace Lipschitz by C 1 in
this classical definition of rectifiability. When Rn is replaced by a pliable Carnot group the two
definitions still make sense and, according to Proposition 5.2, are still equivalent. Rectifiability
in metric spaces and Carnot groups is a very active research topic in geometric measure theory
(see [22] for references).
6. Conditions ensuring pliability
The goal of this section is to identify conditions ensuring that G is pliable. Let us first focus
on the pliability of the zero vector.
Proposition 6.1. For every Carnot group G, the vector 0 ∈ G is pliable.
Proof. According to Proposition 3.7, we should prove that for every ε > 0 the set
1
{β(1) ∈ G | β ∈ CH
([0, 1], G), kβ̇k∞,GH < ε, (β, β̇)(0) = (0G , 0)}
20
NICOLAS JUILLET AND MARIO SIGALOTTI
is a neighborhood of 0G in G.
Recall that there exist k ∈ N, V1 , . . . , Vk ∈ GH and t1 , . . . , tk > 0 such that the map
φ : (τ1 , . . . , τk ) 7→ eτk Vk ◦ · · · ◦ eτ1 V1 (0G )
has rank equal to dim(G) at (τ1 , . . . , τk ) = (t1 , . . . , tk ) and satisfies φ(t1 , . . . , tk ) = 0G (see
[31]). Notice that for every ν > 0, the function
φν : (τ1 , . . . , τk ) 7→eντk Vk ◦ · · · ◦ eντ1 V1 (0G )
=e
ν 2 τk
ν
Vk
◦ · · · ◦ eν
2 τ1 V
ν 1
τ
τk
1
,...,
(0G ) = δν 2 φ
ν
ν
has also rank equal to dim(G) at (τ1 , . . . , τk ) = (νt1 , . . . , νtk ) and satisfies φν (νt1 , . . . , νtk ) =
0G . Hence, up to replacing tj by νtj and Vj by ν 2 Vj for j = 1, . . . , k and ν small enough, we
can assume that t1 + · · · + tk < 1 and kVj kGH < ε for j = 1, . . . , k.
Let O be a neighborhood of (t1 , . . . , tk ) such that for every (τ1 , . . . , τk ) ∈ O we have
τ1 , . . . , τk > 0 and τ1 + · · · + τk < 1. Notice that {φ(τ1 , . . . , τk ) | (τ1 , . . . , τk ) ∈ O} is a
neighborhood of 0G in G.
We complete the proof of the proposition by constructing, for every τ = (τ1 , . . . , τk ) ∈ O a
1
curve βτ ∈ CH
([0, 1], G) such that
(12)
kβ̇τ k∞,GH < ε,
(βτ , β̇τ )(0) = (0G , 0),
βτ (1) = φ(τ ).
1
For every X ∈ GH , p ∈ G and r > 0 let us exhibit a curve γ ∈ CH
([0, r], G) such that γ(0) =
γ(r) = p, γ̇(0) = 0, γ̇(r) = X, and kγ̇k∞,GH = kXkGH . The curve γ can be constructed by
imposing γ̇(r/2) = −X/2 and by extending γ̇ on [0, r/2] and [r/2, r] by convex interpolation. It
is also possible to reverse such a curve by transformation (T4) and connect on any segment [0, r]
1
the point-with-velocity (p, X) with the point-with-velocity (p, 0) by a CH
curve γ respecting
moreover kγ̇k∞,GH = kXkGH . Finally just concatenating (transformation (T5)) curves of
this type it is possible, for every r > 0, to connect (p, X) and (p, Y ) on [0, r] with a curve
1
γr,X,Y ∈ CH
([0, r], G) with kγ̇r,X,Y k∞,GH = max(kXkGH , kY kGH ).
P
We then construct βτ as follows: we fix r = (1 − kj=1 τj )/k, we impose βτ (0) = 0G and
we define β̇τ to be the concatenation of the following 2k continuous curves in GH : first take
γ̇r,0,V1 , then the constant equal to V1 for a time τ1 , then γ̇r,V1 ,V2 , then the constant equal to V2
for a time τ2 , and so on up to γ̇r,Vk−1 ,Vk and finally the constant equal to Vk for a time τk . By
1
construction, βτ ∈ CH
([0, 1], G) and satisfies (12).
Remark 6.2. Let us show that, as a consequence of the previous proposition, pliability and
local uniform pliability are not equivalent properties (albeit we know from Proposition 4.3
that pliability of all horizontal vectors is equivalent to local uniform pliability of all horizontal
vectors).
Recall that local uniform pliability of a horizontal vector X implies pliability of all horizontal
vectors in a neighborhood of X (cf. Definition 4.1). Therefore, if 0 is locally uniformly pliable
for a Carnot group G then every horizontal vector of G is pliable (Remark 4.2). Hence 0
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
21
cannot be locally uniformly pliable if G is not pliable. The remark is concluded by recalling
that non-pliable Carnot groups exist (see Examples 3.3 and 3.5).
Let G be a Carnot group and let X1 , . . . , Xm be an orthonormal basis of GH . Let us consider
the control system in G × Rm given by
m
X
γ̇ =
ui Xi (γ),
(13)
i=1
u̇ = v,
where both u = (u1 , . . . , um) and the control v = (v1 , . . . , vm ) vary in Rm .
Let us rewrite x = (γ, u),
Pm
ui Xi (γ)
0
i=1
,
Fi (x) =
for i = 1, . . . , m,
F0 (x) =
0
ei
where e1 , . . . , em denotes the canonical basis of Rm . System (13) can then be rewritten as
m
X
(14)
ẋ = F0 (x) +
vi Fi (x).
i=1
For every ū ∈ Rm , let Fū : L1 ([0, 1], Rm ) → G×Rm be the endpoint map at time 1 for system
(14) with initial condition (0G , ū). Notice that if x(·) = (γ(·), u(·)) is a solution of (14) with
1
m
1
initial condition
Pm (0G , ū) corresponding to a control v ∈ L ([0, 1], R ), then γ ∈ CH ([0, 1], G)
and kγ̇ − i=1 ūi Xi k∞,GH ≤ kvk1 .
We can then state the following criterium for pliability.
Proposition
6.3. If the map Fū : L1 ([0, 1], Rm ) → G × Rm is open at 0, then the horizontal
Pm
vector i=1 ūi Xi is pliable.
As a consequence, if the restriction of Fū to
L∞ ([0, 1], Rm ) is open at 0, when the L∞
P
m
topology is considered on L∞ ([0, 1], Rm ), then i=1 ūi Xi is pliable. We deduce the following
∗
property: if a straight
Pmcurve is not pliable, then it admits an abnormal lift in T G. Indeed, if
a horizontal vector i=1 ūi Xi is not pliable, then the differential of Fū |L∞ ([0,1],Rm ) at 0 must
be singular. Hence (see, for instance, [1, Section 20.3] or [33, Proposition 5.3.3]), there exist
pγ : [0, 1] → T ∗ G and pu : [0, 1] → (Rm )∗ with (pγ , pu ) 6= 0 such that
∂
(15)
ṗγ (t) = − H(γ(t), ū, pγ (t), pu (t), 0),
∂γ
∂
(16)
ṗu (t) = − H(γ(t), ū, pγ (t), pu (t), 0),
∂u
∂
0=
(17)
H(γ(t), ū, pγ (t), pu (t), 0),
∂v
P
for t ∈ [0, 1], where γ(t) = exp(t m
i=1 ūi Xi ) and
m
X
H(γ, u, pγ , pu , v) = pγ
ui Xi (γ) + pu v.
i=1
22
NICOLAS JUILLET AND MARIO SIGALOTTI
From (17) it follows that pu (t) = 0 for all t ∈ [0, 1]. Equation (16) then implies that
pγ (t)Xi (γ(t)) = 0 for every i = 1, . . . , m and every t ∈ [0, 1]. Moreover, pγ must be different from zero. Comparing (4) and (15), it follows that pγ is an abnormal path.
The control literature proposes several criteria for testing the openness at 0 of an endpoint
map of the type Fū |L∞ ([0,1],Rm ) . The test presented here below, taken from [6], generalizes
previous criteria obtained in [17] and [32].
Theorem 6.4 (Bianchini and Stefani [6, Corollary 1.2]). Let M be a C ∞ manifold and
V0 , V1 , . . . , Vm be C ∞ vector fields on M. Assume that the family of vector fields J = {adkV0 Vj |
k ≥ 0, j = 1, . . . , m} is Lie bracket generating. Denote by H the iterated brackets of elements
in J and recall that the length of an element of H is the sum of the number of times that
each of the elements V0 , . . . , Vm appears in its expression. Assume that every element of H
in whose expression each of the vector fields V1 , . . . , Vm appears an even number of times is
equal, at every q ∈ M, to the linear combination of elements of H of smaller length, evaluated
at q. Fix q0 ∈ M and a neighborhood Ω of 0 in P
Rm . Let U ⊂ L∞ ([0, 1], Ω) be the set of those
controls v such that the solution of q̇ = V0 (q) + m
i=1 vi Vi (q), q(0) = q0 , is defined up to time
1 and denote by Φ(v) the endpoint q(1) of such a solution. Then Φ(U) is a neighborhood of
eV0 (q0 ).
The following two results show how to apply Theorem 6.4 to guarantee that a Carnot group
1
G is pliable and, hence, that (R, G) has the CH
extension property.
1
Theorem 6.5. Let G be a Carnot group of step 2. Then G is pliable and (R, G) has the CH
extension property.
Proof.
6.4 in order to prove that for every horizontal vector
Pm We are going to apply Theorem
∞
u
X
the
endpoint
map
F
:
L
([0,
1], Rm) → G × Rm is open at zero.
u
i=1 i i
Notice that
Xi (γ)
[F0 , Fi ](γ, w) = −
,
i = 1, . . . , m,
0
and
Pm
w
[X
,
X
](γ)
j
i
j
j=1
[F0 , [F0 , Fi ]](γ, w) =
,
i = 1, . . . , m.
0
Moreover for every i, j = 1, . . . , m,
[Xi , Xj ](γ)
[[F0 , Fi ], Fj ] = 0, [[F0 , Fi ], [F0 , Fj ]](γ, w) =
,
0
and all other Lie bracket in and between elements of J = {adkF0 Fi | k ≥ 0, i = 1, . . . , m} is
zero since G is of step 2.
In particular all Lie brackets between elements of J in which each of the vector fields
F1 , . . . , Fm appears an even number of times is zero.
According to Theorem 6.4, we are left to prove that J is Lie bracket generating. This is
clearly true, since
Span{Fi (γ, w), [F0, Fi ](γ, w), [[F0 , Fi ], [F0 , Fj ]](γ, w) | i, j = 1, . . . , m}
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
is equal to T(γ,w) (G × Rm ) for every (γ, w) ∈ G × Rm .
23
We conclude the paper by showing how to construct pliable Carnot groups of arbitrarily
large step.
Proposition 6.6. For every s ≥ 1 there exists a pliable Carnot group of step s.
Proof. Fix s ≥ 1 and consider the free nilpotent stratified Lie algebra A of step s generated
by s elements Z1 , . . . , Zs .
For every i = 1, . . . , s, denote by Ii the ideal of A generated by Zi and by Ji the ideal [Ii , Ii ].
Then J = ⊕si=1 Ji is also an ideal of A.
Then the factor algebra G = A/J is nilpotent and inherits the stratification of A. Denote by
G the Carnot group generated by G. Let X1 , . . . , Xs be the elements of GH obtained by projecting Z1 , . . . , Zs . By construction, every bracket of X1 , . . . , Xs in G in which at least one of
the Xi ’s appears more than once is zero. Moreover, G has step s, since [X1 , [X2 , [ · · · , Xs ], · · · ]
is different from zero.
Let us now apply Theorem 6.4 in order to prove that for every X ∈ GH thePendpoint map
Fu : L∞ ([0, 1], Rs ) → G × Rs is open at zero, where u ∈ Rs is such that X = si=1 ui Xi .
Following the same computations as in the proof of Theorem 6.5,
k
adX Xi (γ)
k+1
adF0 Fi (γ, u) =
,
k ≥ 0, i = 1, . . . , s.
0
In particular the family J = {adkF0 Fi | k ≥ 0, i = 1, . . . , s} is Lie bracket generating.
Moreover, every Lie bracket of elements of Jb = {adk+1
F0 Fi | k ≥ 0, i = 1, . . . , s} in which at
least one of the elements F1 , . . . , Fs appears more than once is zero.
Consider now a Lie bracket W between h ≥ 2 elements of J . Let k1 , . . . , ks be the number
of times in which each of the elements F1 , . . . , Fs appears in W . Let us prove by induction
on h that W is the linear combination of brackets between elements of Jb in which each Fi
appears ki times, i = 1, . . . , s. Consider the case h = 2. Any bracket of the type [adkF0 Fi , Fj ],
k ≥ 0, i, j = 1, . . . , s, is the linear combination of brackets between elements of Jb in which
Fi and Fj appear once, as it can easily be proved by induction on k, thanks to the Jacobi
identity. The induction step on h also follows directly from the Jacobi identity.
We can therefore conclude that every Lie bracket of elements of J in which at least one of
the elements F1 , . . . , Fs appears more than once is zero. This implies in particular that the
hypotheses of Theorem 6.4 are satisfied, concluding the proof that G is pliable.
Acknowledgment. We warmly thank Frédéric Chapoton and Gwénaël Massuyeau for the
suggestions leading us to Proposition 6.6. We are also grateful to Artem Kozhevnikov, Dario
Prandi, Luca Rizzi and Andrei Agrachev for many stimulating discussions. This work has been
initiated during the IHP trimester “Geometry, analysis and dynamics on sub-Riemannian
manifolds” and we wish to thank the Institut Henri Poincaré and the Fondation Sciences
Mathématiques de Paris for the welcoming working conditions.
24
NICOLAS JUILLET AND MARIO SIGALOTTI
The second author has been supported by the European Research Council, ERC StG 2009
“GeCoMethods”, contract number 239748, by the Grant ANR-15-CE40-0018 of the ANR and
by the FMJH Program Gaspard Monge in optimization and operation research.
References
[1] A. A. Agrachev and Y. L. Sachkov. Control theory from the geometric viewpoint, volume 87 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, Berlin, 2004. Control Theory and Optimization,
II.
[2] A. A. Agrachev and A. V. Sarychev. Abnormal sub-Riemannian geodesics: Morse index and rigidity. Ann.
Inst. H. Poincaré Anal. Non Linéaire, 13(6):635–690, 1996.
[3] Z. M. Balogh and K. S. Fässler. Rectifiability and Lipschitz extensions into the Heisenberg group. Math.
Z., 263(3):673–683, 2009.
[4] Z. M. Balogh, U. Lang, and P. Pansu. Lipschitz extensions of maps between Heisenberg groups. Ann.
Inst. Fourier (Grenoble), 66(4):1653–1665, 2016.
[5] D. Barilari, U. Boscain, and M. Sigalotti, editors. Dynamics, geometry and analysis on sub-Riemannian
manifolds, Volumes I-II. EMS Series of Lectures in Mathematics. European Mathematical Society (EMS),
Zürich, 2016.
[6] R. M. Bianchini and G. Stefani. Graded approximations and controllability along a trajectory. SIAM J.
Control Optim., 28(4):903–924, 1990.
[7] A. Bonfiglioli, E. Lanconelli, and F. Uguzzoni. Stratified Lie groups and potential theory for their subLaplacians. Springer Monographs in Mathematics. Springer, Berlin, 2007.
[8] Y. Brudnyi and P. Shvartsman. Generalizations of Whitney’s extension theorem. Int. Math. Res. Not.,
1994(3):129–139, 1994.
[9] R. L. Bryant and L. Hsu. Rigidity of integral curves of rank 2 distributions. Invent. Math., 114(2):435–461,
1993.
[10] L. C. Evans and R. F. Gariepy. Measure theory and fine properties of functions. Textbooks in Mathematics.
CRC Press, Boca Raton, FL, revised edition, 2015.
[11] C. L. Fefferman. A sharp form of Whitney’s extension theorem. Ann. of Math. (2), 161(1):509–577, 2005.
[12] C. L. Fefferman, A. Israel, and G. K. Luli. Sobolev extension by linear operators. J. Amer. Math. Soc.,
27(1):69–145, 2014.
[13] G. B. Folland and E. M. Stein. Hardy spaces on homogeneous groups, volume 28 of Mathematical Notes.
Princeton University Press, Princeton, N.J., 1982.
[14] B. Franchi, R. Serapioni, and F. Serra Cassano. Rectifiability and perimeter in the Heisenberg group.
Math. Ann., 321(3):479–531, 2001.
[15] B. Franchi, R. Serapioni, and F. Serra Cassano. On the structure of finite perimeter sets in step 2 Carnot
groups. J. Geom. Anal., 13(3):421–466, 2003.
[16] C. Golé and R. Karidi. A note on Carnot geodesics in nilpotent Lie groups. J. Dynam. Control Systems,
1(4):535–549, 1995.
[17] H. Hermes. Control systems which generate decomposable Lie algebras. J. Differential Equations,
44(2):166–187, 1982. Special issue dedicated to J. P. LaSalle.
[18] T. Huang and X. Yang. Extremals in some classes of Carnot groups. Sci. China Math., 55(3):633–646,
2012.
[19] B. Kirchheim and F. Serra Cassano. Rectifiability and parameterization of intrinsic regular surfaces in
the Heisenberg group. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5), 3(4):871–896, 2004.
[20] A. Kozhevnikov. Metric properties of level sets of differentiable maps on Carnot groups. Doctoral thesis,
Université Paris Sud - Paris XI, May 2015.
WHITNEY THEOREM FOR CURVES IN CARNOT GROUPS
25
[21] E. Le Donne, A. Ottazzi, and B. Warhurst. Ultrarigid tangents of sub-Riemannian nilpotent groups. Ann.
Inst. Fourier (Grenoble), 64(6):2265–2282, 2014.
[22] E. Le Donne and G. Speight. Lusin approximation for horizontal curves in step 2 Carnot groups. Calc.
Var. Partial Differential Equations, 55(5):111, 2016.
[23] W. Liu and H. J. Sussman. Shortest paths for sub-Riemannian metrics on rank-two distributions. Mem.
Amer. Math. Soc., 118(564):x+104, 1995.
[24] N. Lusin. Sur les propriétés des fonctions mesurables. C. R. Acad. Sci., Paris, 154:1688–1690, 1912.
[25] V. Magnani. Towards differential calculus in stratified groups. J. Aust. Math. Soc., 95(1):76–128, 2013.
[26] B. Malgrange. Ideals of differentiable functions. Tata Institute of Fundamental Research Studies in Mathematics, No. 3. Tata Institute of Fundamental Research, Bombay; Oxford University Press, London,
1967.
[27] R. Montgomery. A survey of singular curves in sub-Riemannian geometry. J. Dynam. Control Systems,
1(1):49–90, 1995.
[28] S. Rigot and S. Wenger. Lipschitz non-extension theorems into jet space Carnot groups. Int. Math. Res.
Not. IMRN, (18):3633–3648, 2010.
[29] F. Serra Cassano. Some topics of geometric measure theory in Carnot groups. In Dynamics, geometry
and analysis on sub-Riemannian manifolds, Volume I, EMS Series of Lectures in Mathematics. European
Mathematical Society (EMS), Zürich, 2016.
[30] G. Speight. Lusin approximation and horizontal curves in Carnot groups. To appear in Revista Matematica
Iberoamericana.
[31] H. J. Sussmann. Some properties of vector field systems that are not altered by small perturbations. J.
Differential Equations, 20(2):292–315, 1976.
[32] H. J. Sussmann. A general theorem on local controllability. SIAM J. Control Optim., 25(1):158–194, 1987.
[33] E. Trélat. Contrôle optimal. Mathématiques Concrètes. [Concrete Mathematics]. Vuibert, Paris, 2005.
Théorie & applications. [Theory and applications].
[34] S. K. Vodop’yanov and I. M. Pupyshev. Whitney-type theorems on the extension of functions on Carnot
groups. Sibirsk. Mat. Zh., 47(4):731–752, 2006.
[35] S. K. Vodop’yanov and I. M. Pupyshev. Whitney-type theorems on the extension of functions on the
Carnot group. Dokl. Akad. Nauk, 406(5):586–590, 2006.
[36] S. Wenger and R. Young. Lipschitz extensions into jet space Carnot groups. Math. Res. Lett., 17(6):1137–
1149, 2010.
[37] H. Whitney. Analytic extensions of differentiable functions defined in closed sets. Trans. Amer. Math.
Soc., 36(1):63–89, 1934.
[38] H. Whitney. Differentiable functions defined in closed sets. I. Trans. Amer. Math. Soc., 36(2):369–387,
1934.
[39] S. Zimmerman. The Whitney extension theorem for C 1 , horizontal curves in Hn . J. Geom. Anal., to
appear.
Institut de Recherche Mathématique Avancée, UMR 7501, Université de Strasbourg et
CNRS, 7 rue René Descartes, 67 000 Strasbourg, France
E-mail address: [email protected]
Inria, team GECO & CMAP, École Polytechnique, CNRS, Université Paris-Saclay, Palaiseau,
France
E-mail address: [email protected]
| 4math.GR
|
arXiv:1310.2350v1 [cs.AI] 9 Oct 2013
The Generalized Traveling Salesman Problem
solved with Ant Algorithms
Camelia-M. Pintea, Petrică C. Pop, Camelia Chira
North University Baia Mare, Babes-Bolyai University, Romania
[email protected], pop [email protected], [email protected]
Abstract
A well known N P-hard problem called the Generalized Traveling Salesman
Problem (GTSP) is considered. In GTSP the nodes of a complete undirected
graph are partitioned into clusters. The objective is to find a minimum cost
tour passing through exactly one node from each cluster.
An exact exponential time algorithm and an effective meta-heuristic algorithm for the problem are presented. The meta-heuristic proposed is a
modified Ant Colony System (ACS) algorithm called Reinforcing Ant Colony
System (RACS) which introduces new correction rules in the ACS algorithm.
Computational results are reported for many standard test problems. The
proposed algorithm is competitive with the other already proposed heuristics
for the GTSP in both solution quality and computational time.
1
Introduction
Many combinatorial optimization problems are N P-hard, and the theory of N P-completeness
has reduced hopes that N P-hard problems can be solved within polynomial bounded computation times. Nevertheless, sub-optimal solutions are sometimes easy to find. Consequently, there is much interest in approximation and heuristic algorithms that can find near
optimal solutions within reasonable running time. Heuristic algorithms are typically among
the best strategies in terms of efficiency and solution quality for problems of realistic size
and complexity.
In contrast to individual heuristic algorithms that are designed to solve a specific
problem, meta-heuristics are strategic problem solving frameworks that can be adapted
to solve a wide variety of problems. Meta-heuristic algorithms are widely recognized as
one of the most practical approaches for combinatorial optimization problems. The most
representative meta-heuristics include genetic algorithms, simulated annealing, tabu search
and ant colony. Useful references regarding meta-heuristic methods can be found in [3].
The Generalized Traveling Salesman Problem (GTSP) has been introduced in [6] and
[7]. The GTSP has several applications to location and telecommunication problems. More
information on these problems and their applications can be found in [4, 5, 6].
Several approaches were considered for solving the GTSP: a branch-and-cut algorithm
for Symmetric GTSP is described and analyzed in [5], in [7] is given a Lagrangian-based
approach for Asymmetric GTSP, in [11] is described a random-key genetic algorithm for
the GTSP, in [10] it is proposed an efficient composite heuristic for the Symmetric GTSP
etc.
1
The aim of this paper is to provide an exact algorithm for the GTSP as well as
an effective meta-heuristic algorithm for the problem. The proposed meta-heuristic is
a modified version of Ant Colony System (ACS). Introduced in ([1, 2]), Ant System is
a heuristic algorithm inspired by the observation of real ant colonies. ACS is used to
solve hard combinatorial optimization problems including the Traveling Salesman Problem
(TSP).
2
Definition and complexity of the GTSP
Let G = (V, E) be an n-node undirected graph whose edges are associated with nonnegative costs. We will assume w.l.o.g. that G is a complete graph (if there is no edge
between two nodes, we can add it with an infinite cost).
Let V1 , ..., Vp be a partition of V into p subsets called clusters (i.e. V = V1 ∪V2 ∪...∪Vp
and Vl ∩ Vk = ∅ for all l, k ∈ {1, ..., p}). We denote the cost of an edge e = {i, j} ∈ E
by cij .
The generalized traveling salesman problem (GTSP) asks for finding a minimum-cost
tour H spanning a subset of nodes such that H contains exactly one node from each
cluster Vi , i ∈ {1, ..., p}. The problem involves two related decisions: choosing a node
subset S ⊆ V , such that |S ∩ Vk | = 1, for all k = 1, ..., p and finding a minimum cost
Hamiltonian cycle in the subgraph of G induced by S.
Such a cycle is called a Hamiltonian tour. The GTSP is called symmetric if and
only if the equality c(i, j) = c(j, i) holds for every i, j ∈ V , where c is the cost function
associated to the edges of G.
3
An exact algorithm for the GTSP
In this section, we present an algorithm that finds an exact solution to the GTSP.
Given a sequence (Vk1 , ..., Vkp ) in which the clusters are visited, we want to find the
best feasible Hamiltonian tour H ∗ (w.r.t cost minimization), visiting the clusters according
to the given sequence. This can be done in polynomial time by solving |Vk1 | shortest path
problems as described below.
We construct a layered network, denoted by LN, having p + 1 layers corresponding to
the clusters Vk1 , ..., Vkp and in addition we duplicate the cluster Vk1 . The layered network
contains all the nodes of G plus some extra nodes v ′ for each v ∈ Vk1 . There is an arc
(i, j) for each i ∈ Vkl and j ∈ Vkl+1 (l = 1, ..., p − 1), having the cost cij and an arc
(i, h), i, h ∈ Vkl , (l = 2, ..., p) having cost cih . Moreover, there is an arc (i, j ′ ) for each
i ∈ Vkp and j ′ ∈ Vk1 having cost cij ′ .
For any given v ∈ Vk1 , we consider paths from v to w′ , w′ ∈ Vk1 , that visits exactly
two nodes from each cluster Vk2 , ..., Vkp , hence it gives a feasible Hamiltonian tour.
Conversely, every Hamiltonian tour visiting the clusters according to the sequence
(Vk1 , ..., Vkp ) corresponds to a path in the layered network from a certain node v ∈ Vk1
to w′ ∈ Vk1 .
Therefore the best (w.r.t cost minimization) Hamiltonian tour H ∗ visiting the clusters
in a given sequence can be found by determining all the shortest paths from each v ∈ Vk1
to each w′ ∈ Vk1 with the property that visits exactly one node from cluster.
The overall time complexity is then |Vk1 |O(m + n log n), i.e. O(nm + nlogn) in
the worst case. We can reduce the time by choosing |Vk1 | as the cluster with minimum
cardinality.
2
It should be noted that the above procedure leads to an O((p − 1)!(nm + nlogn)) time
exact algorithm for the GTSP, obtained by trying all the (p−1)! possible cluster sequences.
Therefore we have established the following result: the above procedure provides an exact
solution to the GSTP in O((p − 1)!(nm + nlogn)) time, where n is the number of nodes,
m is the number of edges and p is the number of clusters in the input graph. Clearly, the
algorithm presented is an exponential time algorithm unless the number of clusters p is
fixed.
4
Ant Colony System
Ant System proposed in [1, 2] is a multi-agent approach used for various combinatorial optimization problems. The algorithms were inspired by the observation of real ant colonies.
An ant can find shortest paths between food sources and a nest. While walking from
food sources to the nest and vice versa, ants deposit on the ground a substance called
pheromone, forming a pheromone trail. Ants can smell pheromone and, when choosing
their way, they tend to choose paths marked by stronger pheromone concentrations. It
has been shown that this pheromone trail following behavior employed by a colony of ants
can lead to the emergence of shortest paths.
When an obstacle breaks the path ants try to get around the obstacle randomly choosing either way. If the two paths encircling the obstacle have the different length, more ants
pass the shorter route on their continuous pendulum motion between the nest points in particular time interval. While each ant keeps marking its way by pheromone the shorter route
attracts more pheromone concentrations and consequently more and more ants choose this
route. This feedback finally leads to a stage where the entire ant colony uses the shortest
path. There are many variations of the ant colony optimization applied on various classical
problems. Ant System make use of simple agents called ants which iterative construct candidate solution to a combinatorial optimization problem. The ants solution construction
is guided by pheromone trails and problem dependent heuristic information.
An individual ant constructs candidate solutions by starting with an empty solution and
then iterative adding solution components until a complete candidate solution is generated.
Each point at which an ant has to decide which solution component to add to its current
partial solution is called a choice point. After the solution construction is completed, the
ants give feedback on the solutions they have constructed by depositing pheromone on
solution components which they have used in their solution. Solution components which
are part of better solutions or are used by many ants will receive a higher amount of
pheromone and, hence, will more likely be used by the ants in future iterations of the
algorithm. To avoid the search getting stuck, typically before the pheromone trails get
reinforced, all pheromone trails are decreased by a factor.
Ant Colony System (ACS) was developed to improve Ant System, making it more
efficient and robust. Ant Colony System works as follows: m ants are initially positioned
on n nodes chosen according to some initialization rule, for example randomly. Each ant
builds a tour by repeatedly applying a stochastic greedy rule - the state transition rule.
While constructing its tour, an ant also modifies the amount of pheromone on the visited
edges by applying the local updating rule. Once all ants have terminated their tour, the
amount of pheromone on edges is modified again by applying the global updating rule.
As was the case in ant system, ants are guided, in building their tours by both heuristic
information and by pheromone information: an edge with a high amount of pheromone is
a very desirable choice. The pheromone updating rules are designed so that they tend to
give more pheromone to edges which should be visited by ants.
3
The ants solutions are not guaranteed to be optimal with respect to local changes and
hence may be further improved using local search methods. Based on this observation,
the best performance are obtained using hybrid algorithms combining probabilistic solution
construction by a colony of ants with local search algorithms as 2-3 opt, tabu-search etc.
In such hybrid algorithms, the ants can be seen as guiding the local search by constructing
promising initial solutions, because ants preferably use solution components which, earlier
in the search, have been contained in good locally optimal solutions.
5
Reinforcing Ant Colony System for GTSP
An Ant Colony System for the GTSP it is introduced. In order to enforces the construction
of a valid solution used in ACS a new algorithm called Reinforcing Ant Colony System
(RACS) it is elaborated with a new pheromone rule as in [8] and pheromone evaporation
technique as in [12].
Let Vk (y) denote the node y from the cluster Vk . The RACS algorithm for the GTSP
works as follows:
• Initially the ants are placed in the nodes of the graph, choosing randomly the clusters
and also a random node from the chosen cluster
• At iteration t + 1 every ant moves to a new node from an unvisited cluster and the
parameters controlling the algorithm are updated.
• Each edge is labeled by a trail intensity. Let τij (t) represent the trail intensity of the
edge (i, j) at time t. An ant decides which node is the next move with a probability
that is based on the distance to that node (i.e. cost of the edge) and the amount
of trail intensity on the connecting edge. The inverse of distance from a node to
the next node is known as the visibility, ηij = c1ij .
• At each time unit evaporation takes place. This is to stop the intensity trails
increasing unbounded. The rate evaporation is denoted by ρ, and its value is between
0 and 1. In order to stop ants visiting the same cluster in the same tour a tabu list
is maintained. This prevents ants visiting clusters they have previously visited. The
ant tabu list is cleared after each completed tour.
• To favor the selection of an edge that has a high pheromone value, τ , and high
visibility value, η a probability function pk iu is considered. J k i are the unvisited
neighbors of node i by ant k and u ∈ J k i , u = Vk (y), being the node y from the
unvisited cluster Vk . This probability function is defined as follows:
pk iu (t) =
[τiu (t)][ηiu (t)]β
,
Σo∈J k i [τio (t)][ηio (t)]β
(1)
where β is a parameter used for tuning the relative importance of edge cost in
selecting the next node. pk iu is the probability of choosing j = u, where u = Vk (y)
is the next node, if q > q0 (the current node is i). If q ≤ q0 the next node j is
chosen as follows:
j = argmaxu∈J k {τiu (t)[ηiu (t)]β },
(2)
i
where q is a random variable uniformly distributed over [0, 1] and q0 is a parameter
similar to the temperature in simulated annealing, 0 ≤ q0 ≤ 1.
4
• After each transition the trail intensity is updated using the correction rule from [8]:
τij (t + 1) = (1 − ρ)τij (t) + ρ
1
.
n · L+
(3)
where L+ is the cost of the best tour.
• In Ant Colony System only the ant that generate the best tour is allowed to globally
update the pheromone. The global update rule is applied to the edges belonging to
the best tour. The correction rule is
τij (t + 1) = (1 − ρ)τij (t) + ρ∆τij (t),
(4)
where ∆τij (t) is the inverse cost of the best tour.
• In order to avoid stagnation we used the pheromone evaporation technique introduced in [12]. When the pheromone trail is over an upper bound τmax , the
pheromone trail is re-initialized. The pheromone evaporation is used after the global
pheromone update rule.
The RACS algorithm computes for a given time timemax a sub-optimal solution, the
optimal solution if it is possible and can be stated as follows in the pseudo-code description.
6
Representation and computational results
A graphic representation of Reinforcing Ant Colony System for solving GTSP is show in
Fig. 2. At the beginning, the ants are in their nest and will start to search food in a specific
area. Assuming that each cluster has specific food and the ants are capable to recognize
this, they will choose each time a different cluster. The pheromone trails will guide the
ants to the shorter path, a solution of GTSP, as in Fig. 2.
To evaluate the performance of the proposed algorithm, the RACS was compared to the
basic ACS algorithm for GTSP and furthermore to other heuristics from literature: Nearest
Neighbor (NN), a composite heuristic GI 3 and a random key-Genetic Algorithm [10, 11].
The numerical experiments that compare RACS with other heuristics used problems from
TSP library [9]. TSPLIB provides optimal objective values for each of the problems. Several
problems with Euclidean distances have been considered. The exact algorithm proposed in
section 3, is clearly outperformed by heuristics including RACS, because his running time
is exponential, while heuristics including RACS are polynomial time algorithms and provide
good sub-optimal solution for reasonable sizes of the problem.
To divide the set of nodes into subsets we used the procedure proposed in [4]. This
procedure sets the number of clusters m = [n/5], identifies the m farthest nodes from each
other, called centers, and assigns each remaining node to its nearest center. Obviously,
some real world problems may have different cluster structures, but the solution procedure
presented in this paper is able to handle any cluster structure.
The initial value of all pheromone trails, τ0 = 1/(n · Lnn ), where Lnn is the result
of Nearest Neighbor, (NN) algorithm. In NN algorithm the rule is always to go next to
the nearest as-yet-unvisited location. The corresponding tour traverses the nodes in the
constructed order. For the pheromone evaporation phase, let denote the upper bound with
τmax = 1/((1 − ρ)Lnn ).
The decimal values can be treated as parameters and can be changed if it is necessary.
The parameters for the algorithm are critical as in all other ant systems.
5
Figure 1: Pseudo-code: the Reinforcing Ant Colony System (RACS).
6
Figure 2: A graphic representation of the Generalized Traveling Salesman Problem
(GTSP) solved with an ant-based heuristic called Reinforcing Ant Colony System (RACS)
is illustrated. The first picture shows an ant starting from the nest to find food, going once
through each cluster and returning to the nest; all the ways are initialized with the same
τ0 pheromone quantity; after several iterations performed by each ant from the nest, the
solution is visible. The second picture shows a solution of Generalized Traveling Salesman
Problem (GTSP) represented by the largest pheromone trail (thick lines); the pheromone
is evaporating on all the other trails (gray lines).
Currently there is no mathematical analysis developed to give the optimal parameter in
each situation. In the ACS and RACS algorithm the values of the parameters were chosen
as follows: β = 5, ρ = 0.5, q0 = 0.5.
In table from Figure 3 we compare the computational results for solving the GTSP
using the ACS and RACS algorithm with the computational results obtained using NN,
GI 3 and random key-Genetic Algorithm mentioned above.
The columns in table from Figure 3 are as follows: Problem- the name of the test
problem; the first digits give the number of clusters (nc)and the last ones give the number
of nodes (n); Opt.val.-the optimal objective value for the problem [11]; ACS, RACS,
NN, GI 3 , GA- the objective value returned by ACS, RACS, NN, GI 3 and random-key
Genetic Algorithm.
The best solutions are in Table 3 in the bold format. All the solutions of ACS
and RACS are the average of five successively run of the algorithm, for each problem.
Termination criteria for ACS and RACS is given by the timemax the maximal computing
time set by the user; in this case ten minutes.
Table 3 shows that Reinforcing Ant Colony System performed well finding the optimal
solution in many cases. The results of RACS are better than the results of ACS. The
RACS algorithm for the Generalized Traveling Salesman Problem can be improved if more
appropriate values for the parameters are used. Also, an efficient combination with other
algorithms can potentially improve the results.
7
Figure 3: Reinforcing Ant Colony System (RACS) versus ACS, NN, GI 3 and GA.
8
7
Conclusion
The basic idea of ACS is that of simulating the behavior of a set of agents that cooperate
to solve an optimization problem by means of simple communications. The algorithm
introduced to solve the Generalized Traveling Salesman Problem, called Reinforcing Ant
Colony System, is an ACS-based algorithm with new correction rules.
The computational results of the proposed RACS algorithm are good and competitive
in both solution quality and computational time with the existing heuristics from the
literature [10, 11]. The RACS results can be improved by considering better values for the
parameters or combining the RACS algorithm with other optimization algorithms. Some
disadvantages have also been identified and they refer the multiple parameters used for
the algorithm and the high hardware resources requirements.
References
[1] Colorni, A., Dorigo, F., Maniezzo, V.: “Distributed Optimization by Ant Colonies”;
Proc. of ECAL-91-Euro. Conf. on Artif. Life, Paris, France, Elsevier Publishing,
(1991), 134–142.
[2] Dorigo, M.: “Optimization, Learning and Natural Algorithms”; (in Italian) Ph.D
thesis, Dipart. di Elettronica, Politecnico di Milano, Italy, (1992).
[3] Glover, F.W., Kochenberger, G.A.: “Handbook of Metaheuristics”; Kluwer, (2002).
[4] Fischetti, M., Gonzales, J.J.S, Toth, P.: “A Branch-and-Cut Algorithm for the Symmetric Generalized Travelling Salesman Problem”; Oper. Res. 45, 3 (1997), 378-394.
[5] Fischetti, M., Gonzales, J.J.S, Toth, P.: “The Generalized Traveling Salesman and
Orienteering Problem”; Kluwer, (2002).
[6] Laporte, G., Nobert, Y.: “Generalized traveling salesman problem through n sets of
nodes: An integer programming approach”; INFOR 21, 1 (1983), 61–75.
[7] Noon, C.E., Bean, J.C.: “A Lagrangian based approach for the asymmetric generalized traveling salesman problem”; Oper. Res. 39 (1991), 623–632.
[8] Pintea, C-M., Dumitrescu, D.:“Improving ant systems using a local updating rule.“
Proceedings. IEEE Computer Society. International Symposium on Symbolic and Numeric Algorithms for Scientific Computing (SYNASC) (2005) 295–299.
[9] Bixby, B., Reinelt, G.: “TSPLIB–a library of travelling salesman and related problem
instances”; (1995) http://nhse.cs.rice.edu/softlib/catalog/tsplib.html
[10] Renaud, J., Boctor, F.F.: “An efficient composite heuristic for the Symmetric Generalized Traveling Salesman Problem”; Euro. J. Oper.Res., 108, 3 (1998), 571–584.
[11] Snyder, L.V., Daskin, M.S.: “A Random-Key Genetic Algorithm for the Generalized
Traveling Salesman Problem”; INFORMS, San Antonio, TX (Nov.2000).
[12] Stützle, T., Hoos, H.H.: “The Max-Min Ant System and local search for the traveling
salesman problem”; Proc. Int. Conf. on Evol. Comp., IEEE Press, Piscataway, NJ,
(1997) 309–314.
9
| 9cs.NE
|
Notes on Pure Dataflow Matrix Machines
arXiv:1610.00831v1 [cs.PL] 4 Oct 2016
Programming with Self-referential Matrix Transformations
Michael Bukatin
Steve Matthews
Andrey Radul
HERE North America LLC
Burlington, Massachusetts, USA
[email protected]
Department of Computer Science
University of Warwick
Coventry, UK
[email protected]
Project Fluid
Cambridge, Massachusetts, USA
[email protected]
Abstract
the streams associated with all neuron outputs using the matrix controlling the DMM. This computation is linear and is potentially
quite global, as any neuron output in the net can contribute to any
neuron input in the net.
DMMs described in the literature are heavily typed. One normally defines a finite collection of allowed kinds of linear streams,
and a finite collection of allowed types of neurons. These two collections are called the DMM signature. One considers a particular
fixed signature. Then one assumes the address space accommodating a countable number of neurons of each type, and then a DMM
is determined by a countable-sized matrix of connectivity weights
(one normally assumes that only a finite number of those weights
are non-zero at any given moment of time).
In particular, DMMs can be equipped with powerful reflection
facilities. Include in the signature the kind of streams of matrices
shaped in such a fashion as to be capable of describing a DMM over
this signature. Then designate a particular neuron, Self, working
as an accumulator of matrices of this shape, and agree that the most
recent output of this neuron will be used at the “down movement”
of each step as the matrix controlling the calculations of all neuron
inputs from all neuron outputs.
Dataflow matrix machines are self-referential generalized recurrent
neural nets. The self-referential mechanism is provided via a stream
of matrices defining the connectivity and weights of the network
in question. A natural question is: what should play the role of
untyped lambda-calculus for this programming architecture? The
proposed answer is a discipline of programming with only one kind
of streams, namely the streams of appropriately shaped matrices.
This yields pure dataflow matrix machines which are networks
of transformers of streams of matrices capable of defining a pure
dataflow matrix machine.
Categories and Subject Descriptors
guages
General Terms
Keywords
software
D [3]: 2—Data-flow lan-
higher-order programming, dataflow
continuous deformation of software, self-referential
1. Introduction
The purpose of these notes is to contribute to the theoretical understanding of dataflow matrix machines.
Dataflow matrix machines (DMMs) arise in the context of synchronous dataflow programming with linear streams, i.e. streams
equipped with an operation of taking a linear combinations of several streams [2].
This is a new general-purpose programming architecture with
interesting properties. One of these properties is that large classes
of programs are parametrized by matrices of numbers. In this aspect
DMMs are similar to recurrent neural nets and, in fact, they can be
considered to be a very powerful generalization of recurrent neural
nets [3, 4].
Just like recurrent neural nets, DMMs are essentially “twostroke engines”. On the “up movement” the built-in neuron transformations compute the next elements of the streams associated
with the neuron outputs from the streams associated with neuron
inputs. This computation is local to the neuron in question and is
generally nonlinear. On the “down movement”, the next elements
of the streams associated with all neuron inputs are computed from
Pure Dataflow Matrix Machines - Version 1.0
1.1 One Kind of Streams
DMMs seem to be a powerful programming platform. In particular,
it is convenient to manually write general-purpose software as
DMMs. At the same time the options to automatically synthesize
DMMs by synthesizing the matrices in question are available.
However, DMMs are a bit too unwieldy for a theoretical investigation.
From the theoretical viewpoint, it is inconvenient that there are
many kinds of streams. It is also inconvenient that one needs to fix
a signature, and that the parametrization by matrices is valid only
for this fixed signature.
So a question naturally arises: What would be the equivalent of
untyped lambda-calculus for dataflow matrix machines?
One of the principles of untyped lambda-calculus: one data type
is enough, namely the type of programs. All data can be expressed
as programs.
The equivalent of this principle for DMMs would be to have
only one kind of streams: streams of matrices, where a matrix is so
shaped as to be able to define a DMM which would be a network
of transformers of streams of matrices (see Section 4 for details).
Instead of string rewriting, a number of streams of matrices are
unfolding in time in this approach.
So all data are to be expressed as countably-infinite matrices of
numbers under this approach (see Section 5), just like all data must
be expressed as lambda-terms in the untyped lambda-calculus.
1
2018/1/13
1.2 One Signature
order constructions continuous, in particular, in making spaces of
programs continuous. Denotationally, the continuous domains representing the meaning of programs are common. But operationally,
we tend to fall back onto discrete schemas.
Dataflow matrix machines are seeking to change that and to provide programming facilities using continuous programs and continuous deformations of programs on the level of operational semantics and of implementation. This can be done both for discrete time
and discrete index spaces (countably-sized matrices of computational elements), and, potentially, for continuous time and continuous index spaces for computational elements.
The oldest electronic continuous platform is electronic analog
computers. The analog program itself, however, is very discrete,
because this kind of machine has a number of single-contact sockets and for every pair of such sockets there is an option to connect
them via a patch cord, or not to connect them.
Among dataflow architectures oriented towards handling the
streams of continuous data one might mention LabVIEW [8] and
Pure Data (e.g. [5]). In both cases, the programs themselves are
quite discrete.
The computational platform which should be discussed in more
details in this context is recurrent neural networks. Turing universality of recurrent neural networks is known for at least 30
years [16, 18].
However, together with many other useful and elegant Turinguniversal computational systems, recurrent neural networks do not
constitute a convenient general-purpose programming platform, but
belong to the class of esoteric programming languages (see [3, 4]
for detailed discussion of that).
Interestingly enough, whether recurrent neural networks understood as programs are discrete or continuous depends on how one
approaches the representation of network topology. If one treats the
network connectivity as a graph, and thinks about this graph as a
discrete data structure, then recurrent neural networks themselves
are discrete.
If one states instead that the network connectivity is always the
complete graph, and that the topology is defined by some of the
weights being zeros, then recurrent neural networks themselves are
continuous.
The most frequent case is borderline. One considers a recurrent
neural net to be defined by the matrix of weights, and therefore to
be continuous, however there are auxiliary discrete structures, e.g.
the matrix of weights is often a sparse matrix, and so a dictionary
of nonzero weights comes into play. Also a language used in describing the network or its implementation comes into play as an
auxiliary discrete structure.
Dataflow matrix machines belong to this borderline case. In particular, the use of sparse matrices is inevitable, because the matrices in question are countable-sized matrices with finite number of
nonzero elements.
Choosing a fixed selection of types of neurons seems too difficult at
the moment. For the time being we would like to retain the ability
to add arbitrary types of neurons to our DMMs.
So instead of selecting a fixed canonical signature, we assume
that there is an underlying language allowing to describe countable
collection of neuron types in such a fashion that all neuron types of
interest can be expressed in that language.
Then assume that all neuron types described by all neuron
type expressions in the underlying language are in the signature.
Assume that our address space is structured in such a way as
to accommodate countable number of neurons for each type of
neurons (see Section 4). Since we have a countable collection
of expressions describing neuron types, our overall collection of
neurons is still countable, and the matrix describing the rules to
recompute neuron inputs from the neuron outputs is still countable.
So, now we have a parametrization by countable matrices of
numbers across all DMMs, and not just across DMMs with a
particular fixed signature.
1.3 Accumulators Revised
The notion of accumulator plays a key role in a number of DMM
constructions including the reflection facility Self.
The most standard version is a neuron performing an identity
transform of its vector input, x, to its vector output, y, of the same
kind. One sets the weight of the recurrent connection from y to x to
1, and then the neuron accumulates contributions of other neurons
connected to x with nonzero weights.
So, at each step the accumulator neuron in effect performs
v := v + ∆v operation. However, it is somewhat of abuse of the
system of kinds of streams to consider v and ∆v as belonging to
the same space, and we’ll see evidence that to do so is a suboptimal
convention later in the paper.
So, what we do, first of all, is that we equip the accumulator
neuron with another input, where ∆v is collected. Then the body
of the neuron computes the sum of v and ∆v, instead of just
performing the identity transform (see Section 6 for more details).
In the situations, where one has multiple kinds of linear streams,
one would often want to assign different kinds to v and to ∆v
(although in other situations one would still use the same kind for
the both of them, effectively considering ∆v to be 0 + ∆v).
1.4 Structure of the Paper
In Section 2 we discuss continuous models of computation and
their higher-order aspects. In Section 3 we juxtapose string rewriting with stream-based approaches to higher-order programming. In
Section 4 we discuss the language of indexes of the network matrix
and how to accommodate countable number of neuron types within
one signature. In Section 5 we discuss representation of constants
and vectors as matrices.
Section 6 provides two examples where it is natural to split the
accumulator input into v and ∆v. One such example comes from
the neuron Self controlling the network matrix. Another example
(Section 6.1) is more involved and requires us to revisit domain
theory in the context of linear models of computation. This is
a bitopological setting, more specifically, bi-continuous domains,
allowing for both monotonic and anti-monotonic inference, and
this is the setting where approximations spaces tend to become
embedded into vector spaces, which is where the connection with
linear models of computation comes into play.
3. Higher-order Programming: String Rewriting
vs. Stream-based Approach
There are several approaches to higher-order stream-based programming. The most popular approach starts with standard higherorder functional programming and focuses on integrating streambased programming into that standard paradigm. The theoretical
underpinning of this approach is lambda-calculus and string rewriting (e.g. [7]).
The dataflow community produced purely stream-based approaches to higher-order programming. One of those approaches
which should be mentioned is an approach based on multidimensional streams [19].
2. Continuous Models of Computation
The history of continuous models of computation is actually quite
long. Where the progress was more limited was in making higher-
Pure Dataflow Matrix Machines - Version 1.0
2
2018/1/13
For any field name Iti , the concatenation t+ @ +Iti + \ +s
is the name of the corresponding neuron input. For any field
name Otj , the concatenation t+ @ +Otj + % +s is the name of
the corresponding neuron output. For every such pair of indices,
(t1 + @ +It1 i + \ +s1 , t2 + @ +Ot2 j + % +s2 ), there is a matrix
element in our matrices under consideration.
To summarize: in this approach the class of pure dataflow matrix
machines is implicitly parametrized by a sufficiently universal language LT describing all types of neurons taken to be of potential
interest together with their associated built-in stream transformations.
For details of DMM functioning see Sections 3.3 and 3.4 of [3].
The approach which we adopt in this paper is based on the
notion of streams of programs. An early work which should be
mentioned in connection with this approach is [15]. An argument
in favor of this approach for programming with linear streams was
presented in Section 3 of [2].
Among recent papers exploring various aspects of the approach
based on the notion of streams of programs are [3, 4, 13]. One of
the goals of the present paper is to show that this approach can
play the role in synchronous dataflow programming with linear
streams comparable to the role played by untyped lambda-calculus
in functional programming.
4. DMM Address Space: Language of Indices
5. Constants and Vectors as Matrices
When one has a countable-sized matrix, it is often more convenient
to index its rows and columns by finite strings over a fixed finite
alphabet than by numbers. There is no principal difference, but
this choice discourages focusing on an arbitrary chosen order, and
encourages semantically meaningful names for the indices.
Here we explain how the construction promised in Section 1.2
works.
To implement the program outlined in Section 1.1 one needs to express the most important linear streams, such as streams of numbers
(scalars), streams of matrix rows and streams of matrix columns,
and other frequently used streams of vectors as streams of matrices. As indicated in [3, 4], one of the key uses of scalars and also
of matrix rows and columns is their use as multiplicative masks.
The ability to use scalars as multiplicative masks needs to be
preserved when those scalars are represented by matrices. For example, if we have a neuron which takes an input stream of scalars a
and an input stream of matrices M , and produces an output stream
of matrices a ∗ M , then we still need to be able to reproduce this
functionality when scalars a are represented by matrices of the
same shape as matrix M .
The most straightforward way to do this is to have a neuron which takes two input streams of matrices and performs their
element-wise multiplication (Hadamard product, sometimes also
called the Schur product). If we chose the Hadamard product as
our main bilinear operation on matrices, then the scalar x must be
represented by the matrix all elements of which are equal to x.
4.1 Neuron Types
Define the notion of a type of neurons following the outline presented in Section 3.1 of [3] for multiple kinds of linear streams. We
only have one kind of linear streams in the present paper, so the
definition is simplified. A neuron type consists of a non-negative
integer input arity M , a positive integer output arity N , and a transformation describing how to map M input streams of matrices into
N output streams of matrices.
Namely associate with the neuron type in question a transformation F taking as inputs M streams of length t − 1 and producing
as outputs N streams of length t for integer time t > 0. Require
the obvious prefix condition that when F is applied to streams of
length t > 0, the first t elements of the output streams of length
t + 1 are the elements which F produces when applied to the prefixes of length t − 1 of the input streams. The most typical situation
is when for t > 1 the t’s elements of the output streams are produced solely on the basis of elements number t − 1 of the input
streams, but our definition also allows neurons to accumulate unlimited history, if necessary.
5.1 Matrices Admitting Finite Descriptions
One particular feature of this approach is that we can no longer
limit ourselves by matrices containing finite number of non-zero
elements, but we also need at least some infinite matrices admitting
finite descriptions.
This means that one needs a convention of what should be done
in case of incorrect operations, such as taking a scalar product
of two infinite vectors of all ones (or adding a matrix consisting
of all ones to Self). It seems likely that the technically easiest
convention in such cases would be to output zeros (or to reset the
network matrix to all zeros).
On the other hand, it is of interest to consider and study the
limits of sequences of finitely describable matrices, and a network
might be computing such a limit when t → ∞.
4.2 Language LT
In this section we are going to use several alphabets. Assume that
the following special symbols don’t belong to any of the other
alphabets: Σ∗ = {\%@}.
Assume that there is a language LT over alphabet ΣT , such that
finite strings from LTT ⊂ LT describe all neuron types of interest.
Call a string t the name of the neuron type it defines (we are not
worried about uniqueness of names for a type here). Assume that
the input arity of the type in question is Mt and the output arity
of the type in question is Nt . That for every integer i such that
0 < i ≤ Mt associate field name Iti from LT and for every integer
j such that 0 < j ≤ Nt associate field name Otj from LT , so that
i1 6= i2 implies Iti1 6= Iti2 and j1 6= j2 implies Otj1 6= Otj2 .
Also assume that there is an alphabet Σ0 with more than one
letter in it and any finite string s over Σ0 is a valid simple name.
5.2 Representing Matrix Rows and Columns as Matrices
Streams of matrix rows and streams of matrix columns also play
important roles in [3, 4]. Represent element y of a row by the corresponding matrix column all elements of which equal y. Represent element z of a column by the corresponding matrix row all
elements of which equal z.
Hence, rows are represented by matrices with equal values
along each column, and columns are represented by matrices with
equal values along each row.
Given matrix row α, denote by (↑ α) its representation as a
matrix. Given matrix column β, denote by (β → ) its representation
as a matrix. Given scalar x, denote by (↑ x→ ) its representation as
a matrix.
Respecting the MATLAB convention to denote the Hadamard
product by .*, we denote the Hadamard product of two matrices
4.3 Language of Indices
The following convention describes the address space for a countable number of neurons for each of the countable number of neuron
types of interest. The indexes are expressed by strings over the alphabet Σ∗ ∪ ΣT ∪ Σ0 . For any name of neuron type t ∈ LTT and
for any simple name s, the concatenation t+ @ +s is a name of a
neuron.
Pure Dataflow Matrix Machines - Version 1.0
3
2018/1/13
˙ while omitting the infix for matrix multiplication, AT B
by A∗B,
or AB T .
Note that because matrix rows correspond to neuron inputs and
matrix columns correspond to neuron outputs, one should always
think about these matrices as rectangular, and not as square matrices, so the transposition is always needed when performing the
standard matrix multiplication on these matrices.
In [3] a standard matrix update operation generalized from
several natural examples is proposed. Given a row α, two columns
β and γ (with the constraint that both β and γ have finite number
of nonzero elements),
P the matrix is updated by the formula aij :=
aij + γi ∗ αj ∗ k βk akj .
In terms of matrix representations what gets added to the net˙ ↑ α)∗(
˙ ↑ (β T A)).
work matrix A is (γ → )∗(
In Section 4 of [4] matrix rows and columns are used for subgraph selection. Consider a subset of neurons, and take α to be a
row with values 1 at the positions corresponding to the neuron outputs of the subset in question and zeros elsewhere, and take β to be
a column with values 1 at the positions corresponding to the neuron inputs of the subset in question and zeros elsewhere. Denote
˙
the element-wise matrix maximum as A⊔B.
The overall connectivity of the subgraph in question is ex˙ → ))∗A,
˙ while the internal connecpressed by the matrix ((↑ α)⊔(β
˙ → )∗A.
˙
tivity of this subgraph is (↑ α)∗(β
6.1 Partial Inconsistency Landscape and Warmus Numbers
Another example of why it is natural to have separate inputs for
v and ∆v in an accumulator comes from considering a scheme of
computation with Warmus numbers.
We have to explain first what are Warmus numbers and why
considering them and a particular scheme of computation in question is natural in this context.
6.1.1 Partial Inconsistency and Vector Semantics
In the presence of partial inconsistency approximation spaces tends
to become embedded into vector spaces. One well-known example
of this phenomenon is that if one allows negative values for probabilities, then probabilistic powerdomain is embedded into the space
of signed measures which is a natural setting for denotational semantics of probabilistic programs [12].
6.1.2 Warmus Numbers
Another example involves algebraic extension of interval numbers
with respect to addition. Interval numbers don’t form a group with
respect to addition. However one can extend them with pseudosegments [b, a] with the contradictory property a < b. For example,
[3,2] is a pseudosegment expressing an interval number with the
contradictory constraint that x ≥ 3 and at the same time x ≤ 2.
The so extended space of interval numbers is a group and a 2D
vector space over reals.
The first discovery of this construction known to us was made
by Mieczysław Warmus [20]. Since then it was rediscovered many
times. For a rather extensive bibliography related to those rediscoveries see [17].
5.3 Other Vectors as Matrices.
The most straightforward way to represent other finite-dimensional
vectors or countable-dimensional vectors with finite number of
nonzero elements in this setup is to represent them as matrix rows
as well. This means reserving a finite or countable number of
appropriately typed neurons to represent coordinates.
For example, to describe vectors representing characters in the
“1-of-N ” encoding which is standard in neural nets [10] one would
need to reserve neurons to represent the letters of the alphabet in
question.
6.1.3 Partial Inconsistency Landscape
There are a number of common motives which appear multiple
times in various studies of partial inconsistency, in particular,
bilattices, bitopology, bicontinuous domains, facilities for nonmonotonic and anti-monotonic inference, order-reversing involutions, etc. Together, these motives serve as focal elements of the
field of study which has been named the partial inconsistency landscape in [2].
In particular, the following situation is typical in the context of
bitopological groups. The two topologies, T and T −1 , are group
dual of each other (that is, the group inverse induces a bijection between the respective systems of open sets), and the antimonotonic group inverse is an order-reversing involution, which
is a bicontinuous map from (X, T, T −1 ) to its bitopological dual,
(X, T −1 , T ) [1].
Because approximation domains tend to become embedded
into vector spaces in this context, the setting of bicontinuous domains [11] equipped with two Scott topologies which tend to be
group dual of each other seems to be natural for semantic studies
of computations with linear streams.
6. Accumulators Revised
Here we continue the line of thought started in Section 1.3.
We give a couple of examples illustrating why it is natural to
have separate inputs for v and ∆v in an accumulator.
The main example is the neuron Self itself, producing the matrix controlling the network on the output, and taking additive updates to that matrix on the input. This is a countable-sized matrix
with finite number of nonzero elements, so it has to be represented
as a sparse matrix, e.g. via a dictionary of nonzero elements. A
typical situation is that the additive update on each time step is
small compared to the matrix itself (more specifically, the update is
typically small in the sense that the number of affected matrix elements is small compared to the overall number of nonzero matrix
elements).
So it does not make much sense to actually copy the output of
Self to its input of Self and perform the additive update there,
which is what should be done if the non-optimized definition of an
accumulator with one input is to be taken literally.
What should be done instead is that additive updates should be
added together at an input of Self, and then on the “up movement”
the Self should add the sum of those updates to the matrix it
accumulates.
So instead of hiding this logic as “implementation details”, it
makes sense to split the inputs of Self into x (with the output of
Self connected to x with weight 1, nothing else connected to x
with non-zero weight, and the copying of the output of Self to x
being a no-op) and ∆x accumulating the additive updates to Self.
Pure Dataflow Matrix Machines - Version 1.0
6.1.4 Computing with Warmus Numbers
Section 4 of [2] provides a detailed overview of the partial inconsistency landscape, including the bitopological and bilattice properties of Warmus numbers. It turns out that Warmus numbers play a
fundamental role in mathematics of partial inconsistency. In particular, Section 4.14 of that paper proposes a schema of computation
via monotonic evolution punctuated by order-reversing involutive
steps.
Computations with Warmus extension of interval numbers via
monotonic evolution punctuated by involutive steps are a good example of why the accumulators should have the asymmetry between v and ∆v.
If an accumulator neuron is to accumulate a monotonically
evolving Warmus number by accepting additive updates to that
4
2018/1/13
number, then the ∆x cannot be an arbitrary Warmus number, but it
must be a pseudosegment [b, a], such that a ≤ 0 ≤ b (the case of
a = b = 0 is allowed). Given that there is a constraint of this kind,
it is natural to want to accumulate ∆x contributions at a separate
input on the “down movement”, and to let the accumulator enforce
the constraint on the “up movement” (e.g. by ignoring requests for
non-monotonic updates). Yet another input might be added to trigger involutive steps (an involutive step in this context transforms
[d, c] into [c, d]). Alternatively, requests for non-monotonic updates
might trigger the involutions. Normally, the involution would be
triggered only if the accumulated number is already a pseudosegment, in which case the involution is an anti-monotonic step.
[10] A. Karpathy. The unreasonable effectiveness of recurrent neural networks, 2015.
http://karpathy.github.io/2015/05/21/rnn-effectiveness/.
[11] K. Keimel. Bicontinuous domains and some old problems in domain
theory. Electronic Notes in Theoretical Computer Science, 257:35–54,
2009.
[12] D. Kozen. Semantics of probabilistic programs. Journal of Computer
and System Sciences, 22(3):328–350, 1981.
[13] N. Krishnaswami. Higher-order reactive programming without spacetime leaks. ACM SIGPLAN Notices, 48(9):221–232, 2013.
[14] J. Lawson. Stably compact spaces. Mathematical Structures in
Computer Science, 21(1):125–169, 2011.
[15] S. Matthews. Adding second order functions to Kahn data flow.
Technical Report Research Report 189, University of Warwick, 1991.
http://www.dcs.warwick.ac.uk/report/pdfs/cs-rr-189.pdf.
[16] J. Pollack. On connectionist models of natural language processing.
PhD thesis, University of Illinois at Urbana-Champaign, 1987.
Chapter 4 is available at
http://www.demo.cs.brandeis.edu/papers/neuring.pdf .
6.1.5 Open Problem: Bicontinuous Reflexive Domains
Despite impressive progress in studies of bicontinuity and bitopology in the context of partial inconsistency landscape [9, 11, 12, 14],
the issues related to reflexive domains and solutions of recursive
domain equation in the context of bicontinuous domains and vector semantics don’t seem to be well understood.
Given that dataflow matrix machines equipped with self-referential facilities work directly on the level of vector spaces, one would
hope that the gap between operational and denotational descriptions would be more narrow in this case than for more traditional
situations such as untyped lambda-calculus.
[17] E. Popova. The arithmetic on proper & improper intervals (a repository of literature on interval algebraic extensions), 1996-2013.
http://www.math.bas.bg/~ epopova/directed.html.
[18] H. Siegelmann and E. Sontag. On the computational power of neural
nets. Journal of Computer and System Sciences, 50:132–150, 1995.
[19] W. Wadge. Higher-order Lucid. In R. Jagannathan, editor, Proceedings of the 4th International Symposium on Lucid and Intensional Programming, pages 62–69, 1991.
http://www.cse.unsw.edu.au/~ plaice/archive/WWW/1991/U-ISLIP91-LucidHigher.pdf .
7. Conclusion
Dataflow matrix machines work with arbitrary linear streams. In
this paper, we focus on the case of pure dataflow matrix machines,
which work with the single kind of linear streams, namely the
streams of matrices defining the connectivity patterns and weights
in pure DMMs themselves.
This allows us to pinpoint the key difference between pure
DMMs and recurrent neural networks: instead of working with
streams of numbers, pure dataflow matrix machines work with
streams of programs, with programs being represented as network
connectivity matrices.
[20] M. Warmus. Calculus of approximations. Bull. Acad. Pol. Sci., Cl. III,
4(5):253–259, 1956.
http://www.cs.utep.edu/interval-comp/warmus.pdf.
[21] G. Zhou, J. Wu, C. Zhang, and Z. Zhou. Minimal Gated Unit for Recurrent Neural Networks, 2016. http://arxiv.org/abs/1603.09420
A. Monotonic Evolution by Additions: Warmus
Numbers vs. Conventional Interval Numbers
Consider a sequence x ⊑ (x + x1 ) ⊑ (x + x1 + x2 ) ⊑ . . . of
elements, which are monotonically increasing and are obtained by
additive corrections from previous elements of the sequence.
S. Andima, R. Kopperman, and P. Nickolas. An asymmetric Ellis
If these are conventional interval numbers, this situation is only
theorem. Topology and Its Applications, 155:146–160, 2007.
possible for the trivial case of 0 = x1 = x2 = . . . , as addition
M. Bukatin and S. Matthews. Linear models of computation and
cannot reduce the degree of imprecision (self-distance) for conprogram learning. In G. Gottlob et al., editors, GCAI 2015, EasyChair
ventional interval numbers. It is not possible to perform nontrivial
Proceedings in Computing, vol. 36, pages 66–78, 2015.
http://easychair.org/publications/download/Linear_Models_of_Computation_and_Program_Learning . monotonic evolution of conventional interval numbers by adding
other interval numbers to previous elements of the sequence in
M. Bukatin, S. Matthews, and A. Radul. Dataflow matrix machines as
question.
programmable, dynamically expandable, self-referential generalized
For Warmus numbers, monotonic evolution by additive correcrecurrent neural networks, 2016. http://arxiv.org/abs/1605.05296.
tions is possible, provided that every additive correction summand
M. Bukatin, S. Matthews, and A. Radul. Programming patterns in
xi = [ai , bi ] is a pseudo-segment anti-approximating zero:
dataflow matrix machines and generalized recurrent neural nets, 2016.
[0, 0] ⊑ [ai , bi ], that is bi ≤ 0 ≤ ai .
http://arxiv.org/abs/1606.09470 .
References
[1]
[2]
[3]
[4]
[5] A. Farnell. Designing Sound. MIT Press, 2010.
B. Rectifiers and Quasi-metrics
[6] Fluid: Project “Fluid” github repository,
https://github.com/anhinga/fluid
Rectified linear unit (ReLU) is a neuron with the activation function
f (x) = max(0, x).
In the recent years, ReLU became the most popular neuron in
the context of non-recurrent deep networks. Whether it is equally
good for recurrent networks remains to be seen.
The activation function max(0, x) is an integral of the Heaviside step function. Lack of smoothness at 0 does not seem to interfere with gradient methods used during neural net training.
Interestingly enough, the standard quasi-metrics on reals associated with upper and lower topologies on reals are closely related
to ReLU: q1 (x, y) = f (x − y) = q2 (y, x).
[7] N. Goodman, V. Mansinghka, D. Roy, K. Bonawitz, and J. Tenenbaum. Church: a language for generative models. In Proc. of Uncertainty in Artificial Intelligence, 2008.
http://danroy.org/papers/church_GooManRoyBonTen-UAI-2008.pdf .
[8] W. Johnston, J. Hanna, and R. Millar. Advances in dataflow programming languages. ACM Computing Surveys, 36:1–34, 2004.
[9] A. Jung and M. Moshier.
On the bitopological nature
of Stone duality.
Technical Report CSR-06-13, School
of Computer Science, University of Birmingham, 2006.
http://www.cs.bham.ac.uk/∼axj/pub/papers/Jung-Moshier-2006-On-the-bitopological-nature-of-Stone-duality.pdf.
Pure Dataflow Matrix Machines - Version 1.0
5
2018/1/13
C. Linear and Bilinear Neurons in LSTM and
Gated Recurrent Unit Networks
bers, but a vector of M matrices M × N . This is what accounts for
factoring M 2 × N 2 dimension out.
Various schemas of recurrent networks with gates and memory
were found to be useful in overcoming the problem of vanishing
gradients in the training of recurrent neural networks, starting with
LSTM in 1997 and now including a variety of other schemas.
For a convenient compact overview of LSTM, gated recurrent
units networks, and related schemas see Section 2 of [21].
The standard way to describe LSTM and gated recurrent unit
networks is to think about them as networks of sigmoid neurons
augmented with external memory and gating mechanisms.
However, it is long understood (and is used in the present paper) that neurons with linear activation functions can be used as
accumulators to implement memory.
It is also known for at least 30 years that bilinear neurons (such
as neurons multiplying two inputs, each of those inputs accumulating linear combinations of output signals of other neurons) can
be used to modulate signals via multiplicative masks (gates) and to
implement conditional constructions in this fashion [16] (see also
Section 1.3.2 of [4]).
Looking at the formulas for LTSM and gated recurrent unit networks in Table 1 of [21] one can observe that instead of thinking
about these networks as networks of sigmoid neurons augmented
with external memory and gating mechanisms, one can describe
them simply as recurrent neural networks built from sigmoid neurons, linear neurons, and bilinear neurons, without any external
mechanisms.
When LTSM and gated recurrent unit networks are built as
recurrent neural networks from sigmoid neurons, linear neurons,
and bilinear neurons, some weights are variable and subject to
training, and some weights are fixed as zeros or ones to establish a
particular network topology.
D.2 Software Prototypes
We prototyped lightweight pure DMMs in Processing 2.2.1 in the
Lightweight Pure DMMs directory of Project Fluid, which is our
open source project dedicated to experiments with the computational architectures based on linear streams [6].
For simplicity we used numbers to index rows and columns of
the matrices, instead of using semantically meaningful strings we
recommend to use as indices for non-lightweight work.
In particular, we demonstrated during those experiments that
it is enough to consider a set of several constant update matrices together with our self-referential network update mechanism
described in the present paper to create oscillations of network
weights and waves of network connectivity patterns.
D.2.1 The aug 26 16 experiment directory
Assume that the neuron Self adds matrices X 0 and X 1 on the
“up movement” to obtain matrix Y 0 . Assume that at the starting
0
0
0
moment t = 0, Y0,0
= 1, Y0,j
= 0 for all j 6= 0, Y1,1
= 1,
0
Y1,j = 0 for all j 6= 1.
1
Assume that Y 1 is a constant matrix, such that Y0,j
= 0 for all
1
1
j, Y1.1 = −2, Y1,j = 0 for all j 6= 1.
The network starts with a “down movement”. After the first
“down movement”, X 0 becomes a copy of Y 0 , X 1 becomes a
copy of Y 1 , and after the first “up movement” at the time t = 1
0
0
Y1,1
changes sign: Y1,1
= −1.
After the second “down movement”, X 1 becomes minus Y 1 ,
1
X1,1
= 2, and after the second “up movement” at the time t = 2
0
0
Y1,1 changes sign again: Y1,1
= 1, etc.
Here we have obtained a simple oscillation of a network weight,
0
Y1,1
(the network matrix is Y 0 at any given moment of time).
D. Lightweight Pure Dataflow Matrix Machines
D.2.2 The aug 27 16 experiment directory
Pure dataflow matrix machines are countable-sized networks with
a finite part of the network being active at any given moment of
time. They process streams of countable-sized matrices with finite
number of non-zero elements,
Sometimes it is convenient to consider the case of networks
of finite size, with fixed number of inputs, M , and fixed number
of outputs, N . If we still would like those networks to process
streams of matrices describing network weights and topology, those
matrices would be finite rectangular matrices M × N .
We call the resulting class of networks Lightweight Pure
DMMs. If we work with reals of limited precision and consider
fixed values of M and N , the resulting class is not Turing-universal,
as its memory space is finite. However, it is often useful to consider
this class for didactic purposes, as both theoretical constructions
and software prototypes tend to be simpler in this case, while many
computational effects can already be illustrated in this generality.
Here instead of Y 1 we take a collection of constant update matrices,
Y j1 , . . . Y jn . Just like in the previous example, make sure that the
first rows (indexed by 0) of those matrices are 0. For the second
j1
j1
rows (indexed by 1), take Y1,j
= −1, Y1,j
= 1,
1
2
j
j
j2
j2
n−1
= 1,
Y1,j2 = −1, Y1,j3 = 1, . . . , Y1,jn−1 = −1, Y1,jn−1
n
jn
jn
Y1,j
=
−1,
Y
=
1,
and
the
rest
of
the
elements
of the second
1,j1
n
rows of these matrices are 0.
Start at t = 0 with Y 0 matrix having the first row as before, and the second row containing the only non-zero element
0
Y1,j
= 1. Then one can easily see (or verify by downloading
1
and running under Processing 2.2.1 the open source software in
the Lightweight Pure DMMs/aug 27 16 experiment directory
of the Project Fluid [6]) that at the moment t = 1 the only non-zero
0
element in the second row of Y 0 is Y1,j
= 1, at the moment t = 2
2
0
the only non-zero element in the second row of Y 0 is Y1,j
= 1,
3
and so on until at the moment t = n this wave of network connec0
tivity pattern loops back to Y1,j
= 1, and then continues looping
1
indefinitely through these n states.
D.1 Dimension of the Network Operators
The network has N outputs, each of which is a matrix M × N ,
hence the overall dimension of the output space is M × N 2 .
The network has M inputs, each of which is a matrix M × N ,
hence the overall dimension of the input space is M 2 × N .
So, overall the dimension of space of all possible linear operators from outputs to inputs (which could potentially be used during
the “down movement”) is M 3 × N 3 . However, our model actually
uses matrices of the dimension M × N during the “down movement”, so only a subspace of dimension M × N of the overall
space of all possible linear operators of the dimension M 3 × N 3
is allowed. The matrix is applied not to a vector of numbers, but to
a vector of N matrices M × N , and yields not a vector of num-
Pure Dataflow Matrix Machines - Version 1.0
D.2.3 Final remarks
A. The actual implementation of Self in the prototype enforces
0
0
the constraint that Y0,0
= 1, Y0,j
= 0 for all j 6= 0.
B. By making the update matrices dynamically dependent upon
e.g. input symbols, one could embed an arbitrary deterministic
finite automaton into this control mechanism in this fashion.
6
2018/1/13
| 6cs.PL
|
Accepted as a workshop contribution at ICLR 2015
L EARNING L ONGER M EMORY IN
R ECURRENT N EURAL N ETWORKS
arXiv:1412.7753v2 [cs.NE] 16 Apr 2015
Tomas Mikolov, Armand Joulin, Sumit Chopra, Michael Mathieu & Marc’Aurelio Ranzato
Facebook Artificial Intelligence Research
770 Broadway
New York City, NY, 10003, USA
{tmikolov,ajoulin,spchopra,myrhev,ranzato}@fb.com
A BSTRACT
Recurrent neural network is a powerful model that learns temporal patterns in sequential data. For a long time, it was believed that recurrent networks are difficult
to train using simple optimizers, such as stochastic gradient descent, due to the
so-called vanishing gradient problem. In this paper, we show that learning longer
term patterns in real data, such as in natural language, is perfectly possible using
gradient descent. This is achieved by using a slight structural modification of the
simple recurrent neural network architecture. We encourage some of the hidden
units to change their state slowly by making part of the recurrent weight matrix
close to identity, thus forming a kind of longer term memory. We evaluate our
model on language modeling tasks on benchmark datasets, where we obtain similar performance to the much more complex Long Short Term Memory (LSTM)
networks (Hochreiter & Schmidhuber, 1997).
1
I NTRODUCTION
Models of sequential data, such as natural language, speech and video, are the core of many machine
learning applications. This has been widely studied in the past with approaches taking their roots in
a variety of fields (Goodman, 2001b; Young et al., 1997; Koehn et al., 2007). In particular, models
based on neural networks have been very successful recently, obtaining state-of-the-art performances
in automatic speech recognition (Dahl et al., 2012), language modeling (Mikolov, 2012) and video
classification (Simonyan & Zisserman, 2014). These models are mostly based on two families of
neural networks: feedforward neural networks and recurrent neural networks.
Feedforward architectures such as time-delayed neural networks usually represent time explicitly
with a fixed-length window of the recent history (Rumelhart et al., 1985). While this type of models
work well in practice, fixing the window size makes long-term dependency harder to learn and can
only be done at the cost of a linear increase of the number of parameters.
The recurrent architectures, on the other hand, represent time recursively. For example, in the simple
recurrent network (SRN) (Elman, 1990), the state of the hidden layer at a given time is conditioned
on its previous state. This recursion allows the model to store complex signals for arbitrarily long
time periods, as the state of the hidden layer can be seen as the memory of the model. In theory, this
architecture could even encode a “perfect” memory by simply copying the state of the hidden layer
over time.
While theoretically powerful, these recurrent models were widely considered to be hard to train due
to the so-called vanishing and exploding gradient problems (Hochreiter, 1998; Bengio et al., 1994).
Mikolov (2012) showed how to avoid the exploding gradient problem by using simple, yet efficient
strategy of gradient clipping. This allowed to efficiently train these models on large datasets by using
only simple tools such as stochastic gradient descent and back-propagation through time (Williams
& Zipser, 1995; Werbos, 1988).
Nevertheless, simple recurrent networks still suffer from the vanishing gradient problem: as gradients are propagated back through time, their magnitude will almost always exponentially shrink.
This makes memory of the SRNs focused only on short term patterns, practically ignoring longer
1
Accepted as a workshop contribution at ICLR 2015
term dependencies. There are two reasons why this happens. First, standard nonlinearities such as
the sigmoid function have a gradient which is close to zero almost everywhere. This problem has
been partially solved in deep networks by using the rectified linear units (ReLu) (Nair & Hinton,
2010). Second, as the gradient is backpropagated through time, its magnitude is multiplied over
and over by the recurrent matrix. If the eigenvalues of this matrix are small (i.e., less than one), the
gradient will converge to zero rapidly. Empirically, gradients are usually close to zero after 5 - 10
steps of backpropagation. This makes it hard for simple recurrent neural networks to learn any long
term patterns.
Many architectures were proposed to deal with the vanishing gradients. Among those, the long short
term memory (LSTM) recurrent neural network (Hochreiter & Schmidhuber, 1997) is a modified
version of simple recurrent network which has obtained promising results on hand writing recognition (Graves & Schmidhuber, 2009) and phoneme classification (Graves & Schmidhuber, 2005).
LSTM relies on a fairly sophisticated structure made of gates which control flow of information to
hidden neurons. This allows the network to potentially remember information for longer periods.
Another interesting direction which was considered is to exploit the structure of the Hessian matrix
with respect to the parameters to avoid vanishing gradients. This can be achieved by using secondorder methods designed for non-convex objective functions (see section 7 in LeCun et al. (1998)).
Unfortunately, there is no clear theoretical justification why using the Hessian matrix would help,
nor there is, to the best of our knowledge, any conclusive thorough empirical study on this topic.
In this paper, we propose a simple modification of the SRN to partially solve the vanishing gradient
problem. In Section 2, we demonstrate that by simply constraining a part of the recurrent matrix
to be close to identity, we can drive some hidden units, called context units to behave like a cache
model which can capture long term information similar to the topic of a text (Kuhn & De Mori,
1990). In Section 3, we show that our model can obtain competitive performance compared to the
state-of-the-art sequence prediction model, LSTM, on language modeling datasets.
2
2.1
M ODEL
S IMPLE RECURRENT NETWORK
yt
yt
R
U
ht
U
R
V
P
ht
A
A
st
↵
B
xt
xt
(a)
(b)
Figure 1: (a) Simple recurrent network. (b) Recurrent network with context features.
We consider sequential data that comes in the form of discrete tokens, such as characters or words.
We assume a fixed dictionary containing d tokens. Our goal is to design a model which is able to
predict the next token in the sequence given its past. In this section, we describe the simple recurrent
network (SRN) model popularized by Elman (1990), and which is the cornerstone of this work.
A SRN consists of an input layer, a hidden layer with a recurrent connection and an output layer
(see Figure 1-a). The recurrent connection allows the propagation through time of information
about the state of the hidden layer. Given a sequence of tokens, a SRN takes as input the one-hot
encoding xt of the current token and predicts the probability yt of next one. Between the current
token representation and the prediction, there is a hidden layer with m units which store additional
information about the previous tokens seen in the sequence. More precisely, at each time t, the state
2
Accepted as a workshop contribution at ICLR 2015
of the hidden layer ht is updated based on its previous state ht−1 and the encoding xt of the current
token, according to the following equation:
ht = σ (Axt + Rht−1 ) ,
(1)
where σ(x) = 1/(1 + exp(x)) is the sigmoid function applied coordinate wise, A is the d × m
token embedding matrix and R is the m × m matrix of recurrent weights. Given the state of these
hidden units, the network then outputs the probability vector yt of the next token, according to the
following equation:
yt = f (U ht ) ,
(2)
where f is the soft-max function and U is the m × d output matrix. In some cases, the size d of the
dictionary can be significant (e.g., more than 100K tokens for standard language modeling tasks)
and computing the normalization term of the soft-max function is often the bottle-neck of this type
of architecture. A common trick introduced in Goodman (2001a) is to replace the soft-max function
√
by a hierarchical soft-max. We use a simple hierarchy with two levels, by binning the tokens into d
clusters with same cumulative word frequency (Mikolov√et al., 2011). This reduces the complexity
of computing the soft-max from O(hd) to about O(h d), but at the cost of lower performance
(around 10% loss in perplexity). We will mention explicitly when we use this approximation in the
experiments.
The model is trained by using stochastic gradient descent method with back-propagation through
time (Rumelhart et al., 1985; Williams & Zipser, 1995; Werbos, 1988). We use gradient renormalization to avoid gradient explosion. In practice, this strategy is equivalent to gradient clipping since
gradient explosions happen very rarely when reasonable hyper-parameters are used. The details of
the implementation are given in the experiment section.
It is generally believed that using a strong nonlinearity is necessary to capture complex patterns
appearing in real-world data. In particular, the class of mapping that a neural network can learn
between the input space and the output space depends directly on these nonlinearities (along with
the number of hidden layers and their sizes). However, these nonlinearities also introduce the socalled vanishing gradient problem in recurrent networks. The vanishing gradient problem states
that as the gradients get propagated back in time, their magnitude quickly shrinks close to zero.
This makes learning longer term patterns difficult, resulting in models which fail to capture the
surrounding context. In the next section, we propose a simple extension of SRN to circumvent this
problem, yielding a model that can retain information about longer context.
2.2
C ONTEXT FEATURES
In this section, we propose an extension of SRN by adding a hidden layer specifically designed to
capture longer term dependencies. We design this layer following two observations: (1) the nonlinearity can cause gradients to vanish, (2) a fully connected hidden layer changes its state completely
at every time step.
SRN uses a fully connected recurrent matrix which allows complex patterns to be propagated
through time but suffers from the fact that the state of the hidden units changes rapidly at every
time step. On the other hand, using a recurrent matrix equal to identity and removing the nonlinearity would keep the state of the hidden layer constant, and every change of the state would have to
come from external inputs. This should allow to retain information for longer period of time. More
precisely, the rule would be:
st = st−1 + Bxt ,
(3)
where B is the d × s context embedding matrix. This solution leads to a model which cannot be
trained efficiently. Indeed, the gradient of the recurrent matrix would never vanish, which would
require propagation of the gradients up to the beginning of the training set.
Many variations around this type of memory have been studied in the past (see Mozer (1993) for
an overview of existing models). Most of these models are based on SRN with no off-diagonal
3
Accepted as a workshop contribution at ICLR 2015
recurrent connections between the hidden units. They differ in how the diagonal weights of the
recurrent matrix are constrained. Recently, Pachitariu & Sahani (2013) showed that this type of
architecture can achieve performance similar to a full SRN when the size of the dataset and of
the model are small. This type of architecture can potentially retain information about longer term
statistics, such as the topic of a text, but it does not scale well to larger datasets (Pachitariu & Sahani,
2013). Besides, it can been argued that purely linear SRNs with learned self-recurrent weights
will perform very similarly to a combination of cache models with different rates of information
decay (Kuhn & De Mori, 1990). Cache models compute probability of the next token given a bagof-words (unordered) representation of longer history. They are well known to perform strongly
on small datasets (Goodman, 2001b). Mikolov & Zweig (2012) show that using such contextual
features as additional inputs to the hidden layer leads to a significant improvement in performance
over the regular SRN. However in their work, the contextual features are pre-trained using standard
NLP techniques and not learned as part of the recurrent model.
In this work, we propose a model which learns the contextual features using stochastic gradient
descent. These features are the state of a hidden layer associated with a diagonal recurrent matrix
similar to the one presented in Mozer (1993). In other words, our model possesses both a fully connected recurrent matrix to produce a set of quickly changing hidden units, and a diagonal matrix that
that encourages the state of the context units to change slowly (see the detailed model in Figure 1b). The fast layer (called hidden layer in the rest of this paper) can learn representations similar to
n-gram models, while the slowly changing layer (called context layer) can learn topic information,
similar to cache models. More precisely, denoting by st the state of the p context units at time t, the
update rules of the model are:
st
ht
yt
= (1 − α)Bxt + αst−1 ,
= σ (P st + Axt + Rht−1 ) ,
= f (U ht + V st )
(4)
(5)
(6)
where α is a parameter in (0, 1) and P is a p×m matrix. Note that there is no nonlinearity applied to
the state of the context units. The contextual hidden units can be seen as an exponentially decaying
bag of words representation of the history. This exponential trace memory (as denoted by Mozer
(1993)) has been already proposed in the context of simple recurrent networks (Jordan, 1987; Mozer,
1989).
A close idea to our work is to use so-called ”leaky integration” neurons (Jaeger et al., 2007), which
also forces the neurons to change their state slowly, however without the structural constraint of
SCRN. It was evaluated on the same dataset as we use further (Penn Treebank) by Bengio et al.
(2013). Interestingly, the results we observed in our experiments show much bigger gains over
stronger baseline using our model, as will be shown later.
Alternative Model Interpretation. If we consider the context units as additional hidden units
(with no activation function), we can see our model as a SRN with a constrained recurrent matrix
M on both hidden and context units:
M=
R
0
P
αIp
,
(7)
where Ip is the identity matrix and M is a square matrix of size m + p, i.e., the sum of the number
of hidden and context units. This reformulation shows explicitly our structural modification of the
Elman SRN (Elman, 1990): we constrain a diagonal block of the recurrent matrix to be equal to a
reweighed identity, and keep an off-diagonal block equal to 0. For this reason, we call our model
Structurally Constrained Recurrent Network (SCRN).
Adaptive Context Features. Fixing the weight α to be constant in Eq. (4) forces the hidden units
to capture information on the same time scale. On the other hand, if we allow this weight to be
learned for each unit, we can potentially capture context from different time delays (Pachitariu &
Sahani, 2013). More precisely, we denote by Q the recurrent matrix of the contextual hidden layer,
and we consider the following update rule for the state of the contextual hidden layer st :
4
Accepted as a workshop contribution at ICLR 2015
st = (I − Q)Bxt + Qst−1 ,
(8)
where Q is a diagonal matrix with diagonal elements in (0, 1). We suppose that these diagonal
elements are obtained by applying a sigmoid transformation to a parameter vector β, i.e., diag(Q) =
σ(β). This parametrization naturally forces the diagonal weights to stay strictly between 0 and 1.
We study in the following section in what situations does learning of the weights help. Interestingly,
we show that learning of the self-recurrent weights does not seem to be important, as long as one
uses also the standard hidden layer in the model.
3
E XPERIMENTS
We evaluate our model on the language modeling task for two datasets. The first dataset is the Penn
Treebank Corpus, which consists of 930K words in the training set. The pre-processing of data and
division to training, validation and test parts are the same as in (Mikolov et al., 2011). The state-ofthe-art performance on this dataset has been achieved by Zaremba et al. (2014), using combination
of many big, regularized LSTM recurrent neural network language models. The LSTM networks
were first introduced to language modeling by Sundermeyer et al. (2012).
The second dataset, which is moderately sized, is called Text8. It is composed of a pre-processed
version of the first 100 million characters from Wikipedia dump. We did split it into training part
(first 99M characters) and development set (last 1M characters) that we use to report performance.
After that, we constructed the vocabulary and replaced all words that occur less than 5 times by
<UNK> token. The resulting vocabulary size is about 44K. To simplify reproducibility of our
results, we released both the SCRN code and the scripts which construct the datasets 1 .
In this section we compare the performance of our proposed model against standard SRNs, and
LSTM RNNs which are becoming the architecture of choice for modeling sequential data with
long-term dependencies.
3.1
I MPLEMENTATION D ETAILS .
We used Torch library and implemented our proposed model following the graph given in Figure
1-b. Note that following the alternative interpretation of our model with the recurrent matrix defined
in Eq. 8, our model could be simply implemented by modifying a standard SRN. We fix α at 0.95
unless stated otherwise. The number of backpropagation through time (BPTT) steps is set to 50 for
our model and was chosen by parameter search on the validation set. For normal SRN, we use just
10 BPTT steps because the gradients vanish faster. We do a stochastic gradient descent after every
5 forward steps. Our model is trained with a batch gradient descent of size 32, and a learning rate of
0.05. We divide the learning rate by 1.5 after each training epoch when the validation error does not
decrease.
3.2
R ESULTS ON P ENN T REEBANK C ORPUS .
We first report results on the Penn Treebank Corpus using both small and moderately sized models
(with respect to the number of hidden units). Table 1 shows that our structurally constrained recurrent network (SCRN) model can achieve performance comparable with LSTM models on small
datasets with relatively small numbers of parameters. It should be noted that the LSTM models have
significantly more parameters for the same size of hidden layer, making the comparison somewhat
unfair - with the input, forget and output gates, the LSTM has about 4x more parameters than SRN
with the same size of hidden layer.
Comparison to ”leaky neurons” is also in favor of SCRN: Bengio et al. (2013) report perplexity
reduction from 136 (SRN) to 129 (SRN + leaky neurons), while for the same dataset, we observed
much bigger improvement, going from perplexity 129 (SRN) down to 115 (SCRN).
Table 1 also shows that SCRN outperforms the SRN architecture even with much less parameters.
This can be seen by comparing performance of SCRN with 40 hidden and 10 contextual units (test
1
The SCRN code can be downloaded at http://github.com/facebook/SCRNNs
5
Accepted as a workshop contribution at ICLR 2015
perplexity 127) versus SRN with 300 hidden units (perplexity 129). This suggests that imposing
a structure on the recurrent matrix allows the learning algorithm to capture additional information.
To obtain further evidence that this additional information is of a longer term character, we did
further run experiments on the Text8 dataset that contains various topics, and thus the longer term
information affects the performance on this dataset much more.
Model
Ngram
Ngram + cache
SRN
SRN
SRN
LSTM
LSTM
LSTM
SCRN
SCRN
SCRN
SCRN
#hidden
50
100
300
50
100
300
40
90
100
300
#context
10
10
40
40
Validation Perplexity
153
137
133
129
120
123
133
124
120
120
Test Perplexity
141
125
144
129
129
123
115
119
127
119
115
115
Table 1: Results on Penn Treebank Corpus: n-gram baseline, simple recurrent nets (SRN), long
short term memory RNNs (LSTM) and structurally constrained recurrent nets (SCRN). Note that
LSTM models have 4x more parameters than SRNs for the same size of hidden layer.
3.2.1
L EARNING S ELF -R ECURRENT W EIGHTS .
We evaluate influence of learning the diagonal weights of the recurrent matrix for the contextual
layer. For the following experiments, we used a hierarchical soft-max with 100 frequency-based
classes on the Penn Treebank Corpus to speedup the experiments. In Table 2, we show that when
the size of the hidden layer is small, learning the diagonal weights is crucial. This result confirms
the findings in Pachitariu & Sahani (2013). However, as we increase the size of our model and use
sufficient number of hidden units, learning of the self-recurrent weights does not give any significant
improvement. This indicates that learning the weights of the contextual units allows these units to
be used as multi-scale representation of the history, i.e., some contextual units can specialize on the
very recent history (for example, for α close to 0, the contextual units would be part of a simple
bigram language model). With various learned self-recurrent weights, the model can be seen as a
combination of cache and bigram models. When the number of standard hidden units is enough to
capture short term patterns, learning the self-recurrent weights does not seem crucial anymore.
Keeping this observation in mind we fixed the diagonal weights when working with the Text8 corpus.
Model
SCRN
SCRN
SCRN
SCRN
SCRN
SCRN
#hidden
50
25
0
140
100
0
#context
0
25
50
0
40
140
Fixed weights
156
150
344
140
127
334
Adaptive weights
156
145
157
140
127
147
Table 2: Perplexity on the test set of Penn Treebank Corpus with and without learning the weights
of the contextual features. Note that in these experiments we used a hierarchical soft-max.
3.3
R ESULTS ON T EXT 8.
Our next experiment involves the Text8 corpus which is significantly larger than the Penn Treebank.
As this dataset contains various articles from Wikipedia, the longer term information (such as current
topic) plays bigger role than in the previous experiments. This is illustrated by the gains when cache
is added to the baseline 5-gram model: the perplexity drops from 309 to 229 (26% reduction).
6
Accepted as a workshop contribution at ICLR 2015
We report experiments with a range of model configurations, with different number of hidden units.
In Table 3, we show that increasing the capacity of standard SRNs by adding the contextual features
results in better performance. For example, when we add 40 contextual units to SRN with 100
hidden units, the perplexity drops from 245 to 189 (23% reduction). Such model is also much better
than SRN with 300 hidden units (perplexity 202).
Model
SCRN
SCRN
SCRN
#hidden
100
300
500
context = 0
245
202
184
context = 10
215
182
177
context = 20
201
172
166
context = 40
189
165
162
context = 80
184
164
161
Table 3: Structurally constrained recurrent nets: perplexity for various sizes of the contextual layer,
reported on the development set of Text8 dataset.
In Table 4, we see that when the number of hidden units is small, our model is better than LSTM.
Despite the LSTM model with 100 hidden units being larger, the SCRN with 100 hidden and 80
contextual features achieves better performance. On the other hand, as the size of the models increase, we see that the best LSTM model is slightly better than the best SCRN (perplexity 156 versus
161). As the perplexity gains for both LSTM and SCRN over SRN are much more significant than
in the Penn Treebank experiments, it seems likely that both models actually model the same kind of
patterns in language.
Model
SRN
SRN
SRN
LSTM
LSTM
LSTM
SCRN
SCRN
SCRN
#hidden
100
300
500
100
300
500
100
300
500
#context
80
80
80
Perplexity on development set
245
202
184
193
159
156
184
164
161
Table 4: Comparison of various recurrent network architectures, evaluated on the development set
of Text8.
4
C ONCLUSION
In this paper, we have shown that learning longer term patterns in real data using recurrent networks is perfectly doable using standard stochastic gradient descent, just by introducing structural
constraint on the recurrent weight matrix. The model can then be interpreted as having quickly
changing hidden layer that focuses on short term patterns, and slowly updating context layer that
retains longer term information.
Empirical comparison of SCRN to Long Short Term Memory (LSTM) recurrent network shows very
similar behavior in two language modeling tasks, with similar gains over simple recurrent network
when all models are tuned for the best accuracy. Moreover, SCRN shines in cases when the size
of models is constrained, and with similar number of parameters it often outperforms LSTM by a
large margin. This can be especially useful in cases when the amount of training data is practically
unlimited, and even models with thousands of hidden neurons severely underfit the training datasets.
We believe these findings will help researchers to better understand the problem of learning longer
term memory in sequential data. Our model greatly simplifies analysis and implementation of recurrent networks that are capable of learning longer term patterns. Further, we published the code
that allows to easily reproduce experiments described in this paper.
At the same time, it should be noted that none of the above models is capable of learning truly long
term memory, which has a different nature. For example, if we would want to build a model that can
7
Accepted as a workshop contribution at ICLR 2015
store arbitrarily long sequences of symbols and reproduce these later, it would become obvious that
this is not doable with models that have finite capacity. A possible solution is to use the recurrent
net as a controller of an external memory which has unlimited capacity. For example in (Joulin
& Mikolov, 2015), a stack-based memory is used for such task. However, a lot of research needs
to be done in this direction before we will develop models that can successfully learn to grow in
complexity and size when solving increasingly more difficult tasks.
R EFERENCES
Bengio, Yoshua, Simard, Patrice, and Frasconi, Paolo. Learning long-term dependencies with gradient descent is difficult. Neural Networks, IEEE Transactions on, 5(2):157–166, 1994.
Bengio, Yoshua, Boulanger-Lewandowski, Nicolas, and Pascanu, Razvan. Advances in optimizing
recurrent networks. In ICASSP, 2013.
Dahl, George E, Yu, Dong, Deng, Li, and Acero, Alex. Context-dependent pre-trained deep neural
networks for large-vocabulary speech recognition. Audio, Speech, and Language Processing,
IEEE Transactions on, 20(1):30–42, 2012.
Elman, Jeffrey L. Finding structure in time. Cognitive science, 14(2):179–211, 1990.
Goodman, Joshua. Classes for fast maximum entropy training. In Acoustics, Speech, and Signal
Processing, 2001. Proceedings.(ICASSP’01). 2001 IEEE International Conference on, volume 1,
pp. 561–564. IEEE, 2001a.
Goodman, Joshua T. A bit of progress in language modeling. Computer Speech & Language, 15(4):
403–434, 2001b.
Graves, Alex and Schmidhuber, Juergen. Offline handwriting recognition with multidimensional
recurrent neural networks. In Advances in Neural Information Processing Systems, pp. 545–552,
2009.
Graves, Alex and Schmidhuber, Jürgen. Framewise phoneme classification with bidirectional lstm
and other neural network architectures. Neural Networks, 18(5):602–610, 2005.
Hochreiter, Sepp. The vanishing gradient problem during learning recurrent neural nets and problem
solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02):
107–116, 1998.
Hochreiter, Sepp and Schmidhuber, Jürgen. Long short-term memory. Neural computation, 9(8):
1735–1780, 1997.
Jaeger, Herbert, Lukoševičius, Mantas, Popovici, Dan, and Siewert, Udo. Optimization and applications of echo state networks with leaky-integrator neurons. Neural Networks, 20(3):335–352,
2007.
Jordan, Michael I. Attractor dynamics and parallelism in a connectionist sequential machine. Proceedings of the Eighth Annual Conference of the Cognitive Science Society, pp. 531–546, 1987.
Joulin, Armand and Mikolov, Tomas. Inferring algorithmic patterns with stack-augmented recurrent
nets. arXiv preprint arXiv:1503.01007, 2015.
Koehn, Philipp, Hoang, Hieu, Birch, Alexandra, Callison-Burch, Chris, Federico, Marcello,
Bertoldi, Nicola, Cowan, Brooke, Shen, Wade, Moran, Christine, Zens, Richard, et al. Moses:
Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the ACL on Interactive Poster and Demonstration Sessions, pp. 177–180. Association for
Computational Linguistics, 2007.
Kuhn, Roland and De Mori, Renato. A cache-based natural language model for speech recognition.
Pattern Analysis and Machine Intelligence, IEEE Transactions on, 12(6):570–583, 1990.
LeCun, Yann, Bottou, Leon, Orr, Genevieve, and Müller, Klaus. Efficient backprop. Neural Networks: Tricks of the Trade, pp. 546–546, 1998.
8
Accepted as a workshop contribution at ICLR 2015
Mikolov, Tomáš. Statistical language models based on neural networks. PhD thesis, Ph. D. thesis,
Brno University of Technology, 2012.
Mikolov, Tomas and Zweig, Geoffrey. Context dependent recurrent neural network language model.
In SLT, pp. 234–239, 2012.
Mikolov, Tomas, Kombrink, Stefan, Burget, Lukas, Cernocky, JH, and Khudanpur, Sanjeev. Extensions of recurrent neural network language model. In Acoustics, Speech and Signal Processing
(ICASSP), 2011 IEEE International Conference on, pp. 5528–5531. IEEE, 2011.
Mozer, Michael C. A focused back-propagation algorithm for temporal pattern recognition. Complex
systems, 3(4):349–381, 1989.
Mozer, Michael C. Neural net architectures for temporal sequence processing. In Santa Fe Institute
Studies in The Sciences of Complexity, volume 15, pp. 243–243. Addison-Wessley Publishing
CO, 1993.
Nair, Vinod and Hinton, Geoffrey E. Rectified linear units improve restricted boltzmann machines.
In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp. 807–
814, 2010.
Pachitariu, Marius and Sahani, Maneesh. Regularization and nonlinearities for neural language
models: when are they needed? arXiv preprint arXiv:1301.5650, 2013.
Rumelhart, David E, Hinton, Geoffrey E, and Williams, Ronald J. Learning internal representations
by error propagation. Technical report, DTIC Document, 1985.
Simonyan, Karen and Zisserman, Andrew. Two-stream convolutional networks for action recognition in videos. In Advances in Neural Information Processing Systems, pp. 568–576, 2014.
Sundermeyer, Martin, Schlüter, Ralf, and Ney, Hermann. Lstm neural networks for language modeling. In INTERSPEECH, 2012.
Werbos, Paul J. Generalization of backpropagation with application to a recurrent gas market model.
Neural Networks, 1(4):339–356, 1988.
Williams, Ronald J and Zipser, David. Gradient-based learning algorithms for recurrent networks
and their computational complexity. Back-propagation: Theory, architectures and applications,
pp. 433–486, 1995.
Young, Steve, Evermann, Gunnar, Gales, Mark, Hain, Thomas, Kershaw, Dan, Liu, Xunying,
Moore, Gareth, Odell, Julian, Ollason, Dave, Povey, Dan, et al. The HTK book, volume 2. Entropic Cambridge Research Laboratory Cambridge, 1997.
Zaremba, Wojciech, Sutskever, Ilya, and Vinyals, Oriol. Recurrent neural network regularization.
arXiv preprint arXiv:1409.2329, 2014.
9
| 9cs.NE
|
Convex Regularization for High-Dimensional
arXiv:1512.01215v2 [math.ST] 13 Apr 2017
Multi-Response Tensor Regression
Garvesh Raskutti∗ , Ming Yuan† and Han Chen†
University of Wisconsin-Madison
Abstract
In this paper we present a general convex optimization approach for solving highdimensional multiple response tensor regression problems under low-dimensional structural assumptions. We consider using convex and weakly decomposable regularizers
assuming that the underlying tensor lies in an unknown low-dimensional subspace.
Within our framework, we derive general risk bounds of the resulting estimate under
fairly general dependence structure among covariates. Our framework leads to upper
bounds in terms of two very simple quantities, the Gaussian width of a convex set in
tensor space and the intrinsic dimension of the low-dimensional tensor subspace. To
the best of our knowledge, this is the first general framework that applies to multiple response problems. These general bounds provide useful upper bounds on rates
of convergence for a number of fundamental statistical models of interest including
multi-response regression, vector auto-regressive models, low-rank tensor models and
pairwise interaction models. Moreover, in many of these settings we prove that the
resulting estimates are minimax optimal. We also provide a numerical study that both
validates our theoretical guarantees and demonstrates the breadth of our framework.
∗
Departments of Statistics and Computer Science, and Optimization Group at Wisconsin Institute for
Discovery, University of Wisconsin-Madison, 1300 University Avenue, Madison, WI 53706. The research of
Garvesh Raskutti is supported in part by NSF Grant DMS-1407028
†
Morgridge Institute for Research and Department of Statistics, University of Wisconsin-Madison, 1300
University Avenue, Madison, WI 53706. The research of Ming Yuan and Han Chen was supported in part
by NSF FRG Grant DMS-1265202, and NIH Grant 1-U54AI117924-01.
1
1
Introduction
Many modern scientific problems involve solving high-dimensional statistical problems where
the sample size is small relative to the ambient dimension of the underlying parameter to be
estimated. Over the past few decades there has been a large amount of work on solving such
problems by imposing low-dimensional structure on the parameter of interest. In particular
sparsity, low-rankness and other low-dimensional subspace assumptions have been studied
extensively both in terms of the development of fast algorithms and theoretical guarantees.
See, e.g., Buhlmann and van de Geer (2011) and Hastie et al. (2015), for an overview.
Most of the prior work has focussed on scenarios in which the parameter of interest is
a vector or matrix. Increasingly common in practice, however, the parameter or object
to be estimated naturally has a higher order tensor structure. Examples include hyperspectral image analysis (Li and Li, 2010), multi-energy computed tomography (Semerci et al.,
2014), radar signal processing (Sidiropoulos and Nion, 2010), audio classification (Mesgarani
et al., 2006) and text mining (Cohen and Collins, 2012) among numerous others. It is much
less clear how the low dimensional structures inherent to these problems can be effectively
accounted for. The main purpose of this article is to fill in this void and provide a general
and unifying framework for doing so.
Consider a general tensor regression problem where covariate tensors X (i) ∈ Rd1 ×···×dM
and response tensors Y (i) ∈ RdM +1 ×···×dN are related through:
Y (i) = hX (i) , T i + (i) ,
i = 1, 2, . . . , n.
(1)
Here T ∈ Rd1 ×···×dN is an unknown parameter of interest, and (i) s are independent and
identically distributed noise tensors whose entries are independent and identically distributed
centred normal random variables with variance σ 2 . Further, for simplicity we assume the
covariates (X (i) )ni=1 are Gaussian, but with fairly general dependence assumptions. The
notation h·, ·i will refer throughout this paper to the standard inner product taken over
appropriate Euclidean spaces. Hence, for A ∈ Rd1 ×···×dM and B ∈ Rd1 ×···×dN :
hA, Bi =
d1
X
j1 =1
···
dM
X
Aj1 ,...,jM Bj1 ,...,jM ∈ R
jM =1
is the usual inner product if M = N ; and if M < N , then hA, Bi ∈ RdM +1 ×···×dN such that
2
its (jM +1 , . . . , jN ) entry is given by
(hA, Bi)jM +1 ,...,jN =
d1
X
j1 =1
···
dM
X
Aj1 ,...,jM Bj1 ,...,jM ,jM +1 ,...,jN .
jM =1
The goal of tensor regression is to estimate the coefficient tensor T based on observations
{(X (i) , Y (i) ) : 1 ≤ i ≤ n}. In addition to the canonical example of tensor regression with
Y a scalar response (i.e., M = N ), many other commonly encountered regression problems
are also special cases of the general tensor regression model (1). Multi-response regression
(see, e.g., Anderson, 1984), vector autoregressive model (see, e.g., Lütkepolhl, 2006), and
pairwise interaction tensor model (see, e.g., Rendle and Schmidt-Thieme, 2010) are some
of the notable examples. In this article, we provide a general treatment to these seemingly
different problems.
Our main focus here is on situations where the dimensionality dk s are large when compared with the sample size n. In many practical settings, the true regression coefficient tensor
T may have certain types of low-dimensional structure. Because of the high ambient dimension of a regression coefficient tensor, it is essential to account for such a low-dimensional
structure when estimating it. Sparsity and low-rankness are the most common examples of
such low dimensional structures. In the case of tensors, sparsity could occur at the entry-wise
level, fiber-wise level, or slice-wise level, depending on the context and leading to different
interpretations. There are also multiple ways in which low-rankness may be present when it
comes to higher order tensors, either at the original tensor level or at the matricized tensor
level.
In this article, we consider a general class of convex regularization techniques to exploit
either type of low-dimensional structure. In particular, we consider the standard convex
regularization framework:
(
Tb ∈
arg min
A∈Rd1 ×···×dN
)
n
1 X (i)
kY − hA, X (i) ik2F + λR(A) ,
2n i=1
(2)
where the regularizer R(·) is a norm on Rd1 ×···×dN , and λ > 0 is a tuning parameter. Hereafter, for a tensor A, kAkF = hA, Ai1/2 . We derive general risk bounds for a family of so-called
weakly decomposable regularizers under fairly general dependence structure among the covariates. These general upper bounds apply to a number of concrete statistical inference
3
problems including the aforementioned multi-response regression, high-dimensional vector
auto-regressive models, low-rank tensor models, and pairwise interaction tensors where we
show that they are typically optimal in the minimax sense.
In developing these general results, we make several contributions to a fast growing literature on high dimensional tensor estimation. First of all, we provide a unified and principled
approach to exploit the low dimensional structure in these tensor problems. In doing so, we
incorporate an extension of the notion of decomposability originally introduced by Negahban et al. (2012) for vector and matrix models to weak decomposability previously introduced
in van de Geer (2014) which allows us to handle more delicate tensor models such as the
nuclear norm regularization for low-rank tensor models. Moreover, we provide, for the regularized least squared estimate given by (2), a general risk bound under an easily interpretable
condition on the design tensor. The risk bound we derive is presented in terms of merely two
geometric quantities, the Gaussian width which depends on the choice of regularization and
the intrinsic dimension of the subspace that the tensor T lies in. We believe this is the first
general framework that applies to multiple responses and general dependence structure for
the covariate tensor X. Finally, our general results lead to novel upper bounds for several
important regression problems involving high-dimensional tensors: multi-response regression, multi-variate auto-regressive models and pairwise interaction models, for which we also
prove that the resulting estimates are minimiax rate optimal with appropriate choices of
regularizers.
Our framework incorporates both tensor structure and multiple responses which present
a number of challenges compared to previous approaches. These challenges manifest themselves both in terms of the choice of regularizer R and the technical challenges in the proof
of the main result. Firstly since the notion of low-dimensional is more generic for tensors
meaning there are a number of choices of convex regularizer R and these must satisfy a
form of weak decomposability and provide optimal rates. Multiple responses and the flexible
dependence structure among the covariates also present significant technical challenges for
proving restricted strong convexity, a key technical tool for establishing rates of convergence.
In particular, a one-sided uniform law (Lemma 12) is required instead of classical techniques
as developed in e.g. Negahban and Wainwright (2012); Raskutti et al. (2010) that only apply
4
to univariate responses.
The remainder of the paper is organized as follows: In Section 2 we introduce the general framework of using weakly decomposable regularizers for exploiting low-dimensional
structures in high dimensional tensor regression. In Section 3 we present a general upper
bound for weakly decomposable regularizers and discuss specific risk bounds for commonly
used sparsity or low-rankness regularizers for tensors. In Section 4 we apply our general
result to three specific statistical problems, namely, multi-response regression, multivariate
autoregressive model, and the pairwise interaction model. We show that in each of the three
examples appropriately chosen weakly decomposable regularizers leads to minimax optimal
estimation of the unknown parameters. Numerical experiments are presented in Section 5
to further demonstrate the merits and breadth of our approach. Proofs are provided in
Section 6.
2
Methodology
Recall that the regularized least-squares estimate is given by
)
(
n
X
1
kY (i) − hA, X (i) ik2F + λR(A) .
Tb = arg min
2n
d
×···×d
N
A∈R 1
i=1
For brevity, we assume implicitly hereafter that the minimizer on its left hand side is uniquely
defined. Our development here actually applies to the more general case where Tb can be
taken as an arbitrary element from the set of the minimizers. Of particularly interest here is
the so-called weakly decomposable convex regularizers, extending a similar concept introduced
by Negahban et al. (2012) for vectors and matrices.
Let A be an arbitrary linear subspace of Rd1 ×···×dN and A⊥ its orthogonal complement:
A⊥ := {A ∈ Rd1 ×···×dN | hA, Bi = 0 for all B ∈ A}.
We call a regularizer R(·) weakly decomposable with respect to a pair (A, B) where B ⊆ A
if there exist a constant 0 < cR ≤ 1 such that for any A ∈ A⊥ and B ∈ B,
R(A + B) ≥ R(A) + cR R(B).
5
(3)
In particular, if (3) holds for any B ∈ B = A, we say R(·) is weakly decomposable with
respect to A. A more general version of this concept was first introduced in van de Geer
(2014). Because R is a norm, by triangular inequality, we also have
R(A + B) ≤ R(A) + R(B).
Many of the commonly used regularizers for tensors are weakly decomposable, or decomposable for short. When cR = 1, our definition of decomposability naturally extends from
similar notion for vectors (N = 1) and matrices (N = 2) introduced by Negahban et al.
(2012). We also allow for more general choices of cR here to ensure a wider applicability. For
example as we shall see the popular tensor nuclear norm regularizer is decomposable with
respect to appropriate linear subspaces with cR = 1/2, but not decomposable if cR = 1.
We have now described a catalogue of commonly used regularizers for tensors and argue
that they are all decomposable with respect to appropriately chosen subspaces of Rd1 ×···×dN .
To fix ideas, we shall focus in what follows on estimating a third-order tensor T , that is
N = 3, although our discussion can be straightforwardly extended to higher-order tensors.
2.1
Sparsity regularizers
An obvious way to encourage entry-wise sparsity is to impose the vector `1 penalty on the
entries of A:
R(A) :=
d1 X
d2 X
d3
X
|Aj1 j2 j3 |,
(4)
j1 =1 j2 =1 j3 =1
following the same idea as the Lasso for linear regression (see, e.g., Tibshirani, 1996). This is
a canonical example of decomposable regularizers. For any fixed I ⊂ [d1 ] × [d2 ] × [d3 ] where
[d] = {1, 2, . . . , d}, write
A(I) = B(I) = A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = 0 for all (j1 , j2 , j3 ) ∈
/I .
It is clear that
A⊥ (I) = A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = 0 for all (j1 , j2 , j3 ) ∈ I ,
and R(A) defined by (4) is decomposable with respect to A with cR = 1.
6
(5)
In many applications, sparsity arises with a more structured fashion for tensors. For
example, a fiber or a slice of a tensor is likely to be zero simultaneously. Mode-1 fibers of a
tensor A ∈ Rd1 ×d2 ×d3 are the collection of d1 -dimensional vectors
A·j2 j3 = (A1j2 j3 , . . . , Ad1 j2 j3 )> : 1 ≤ j2 ≤ d2 , 1 ≤ j3 ≤ d3 .
Mode-2 and -3 fibers can be defined in the same fashion. To fix ideas, we focus on mode-1
fibers. Sparsity among mode-1 fibers can be exploited using the group-based `1 regularizer:
R(A) =
d3
d2 X
X
kA·j2 j3 k`2 ,
(6)
j2 =1 j3 =1
similar to the group Lasso (see, e.g., Yuan and Lin, 2006), where k · k`2 stands for the usual
vector `2 norm. Similar to the vector `1 regularizer, the group `1 -based regularizer is also
decomposable. For any fixed I ⊂ [d2 ] × [d3 ], write
A(I) = B(I) = A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = 0 for all (j2 , j3 ) ∈
/I .
(7)
It is clear that
A⊥ (I) = A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = 0 for all (j2 , j3 ) ∈ I ,
and R(A) defined by (6) is decomposable with respect to A with cR = 1. Note that in
defining the regularizer in (6), instead of vector `2 norm, other `q (q > 1) norms could also
be used. See, e.g., Turlach et al. (2005).
Sparsity could also occur at the slice level. The (1, 2) slices of a tensor A ∈ Rd1 ×d2 ×d3 are
the collection of d1 × d2 matrices
{A··j3 = (Aj1 j2 j3 )1≤j1 ≤d1 ,1≤j2 ≤d2 : 1 ≤ j3 ≤ d3 } .
Let k · k be an arbitrary norm on d1 × d2 matrices. Then the following group regularizer can
be considered:
R(A) =
d3
X
kA··j3 k.
(8)
j3 =1
Typical examples of the matrix norm that can be used in (8) include Frobenius norm and
nuclear norm among others. In the case when k · kF is used, R(·) is again a decomposable
regularizer with respect to
/I .
A(I) = B(I) = A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = 0 for all j3 ∈
7
(9)
for any I ⊂ [d3 ].
Now consider the case when we use the matrix nuclear norm k · k∗ in (8). Let P1j and
P2j , j = 1, . . . , d3 be two sequences of projection matrices on Rd1 and Rd2 respectively. Let
⊥
⊥
A··j P2j
= 0, j = 1, . . . , d3 ,
A(P1j , P2j : 1 ≤ j ≤ d3 ) = A ∈ Rd1 ×d2 ×d3 : P1j
(10)
and
B(P1j , P2j : 1 ≤ j ≤ d3 ) = A ∈ Rd1 ×d2 ×d3 : A··j = P1j A··j P2j , j = 1, . . . , d3 .
(11)
By pinching inequality (see, e.g., Bhatia, 1997), it can be derived that R(·) is decomposable
with respect to A(P1j , P2j : 1 ≤ j ≤ d3 ) and B(P1j , P2j : 1 ≤ j ≤ d3 ).
2.2
Low-rankness regularizers
In addition to sparsity, one may also consider tensors with low-rank. There are multiple
notions of rank for higher-order tensors. See, e.g., Koldar and Bader (2009), for a recent
review. In particular, the so-called CP rank is defined as the smallest number r of rank-one
tensors needed to represent a tensor A ∈ Rd1 ×d2 ×d3 :
A=
r
X
uk ⊗ vk ⊗ wk
(12)
k=1
where uk ∈ Rd1 , vk ∈ Rd2 and wk ∈ Rd3 . To encourage a low rank estimate, we can consider
the nuclear norm regularization. Following Yuan and Zhang (2014), we define the nuclear
norm of A through its dual norm. More specifically, let the spectral norm of A be given by
kAks =
hA, u ⊗ v ⊗ wi.
max
kuk`2 ,kvk`2 ,kwk`2 ≤1
Then its nuclear norm is defined as
kAk∗ = max hA, Bi.
kBks ≤1
We shall then consider the regularizer:
R(A) = kAk∗ .
We now show this is also a weakly decomposable regularizer.
8
(13)
Let Pk be a projection matrix in Rdk . Define
(P1 ⊗ P2 ⊗ P3 )A =
r
X
P1 uk ⊗ P2 vk ⊗ P3 wk .
k=1
Write
Q = P1 ⊗ P2 ⊗ P3 + P1⊥ ⊗ P2 ⊗ P3 + P1 ⊗ P2⊥ ⊗ P3 + P1 ⊗ P2 ⊗ P3⊥ ,
and
Q⊥ = P1⊥ ⊗ P2⊥ ⊗ P3⊥ + P1⊥ ⊗ P2⊥ ⊗ P3 + P1 ⊗ P2⊥ ⊗ P3⊥ + P1⊥ ⊗ P2 ⊗ P3⊥ ,
where Pk⊥ = I − Pk .
Lemma 1. For any A ∈ Rd1 ×d2 ×d3 and projection matrices Pk in Rdk , k = 1, 2, 3, we have
1
kAk∗ ≥ k(P1 ⊗ P2 ⊗ P3 )Ak∗ + kQ⊥ Ak∗ .
2
Lemma 1 is a direct consequence from the characterization of sub-differential for tensor
nuclear norm given by Yuan and Zhang (2014), and can be viewed as a tensor version of the
pinching inequality for matrices.
Write
A(P1 , P2 , P3 ) = A ∈ Rd1 ×d2 ×d3 : QA = A ,
(14)
B(P1 , P2 , P3 ) = A ∈ Rd1 ×d2 ×d3 : (P1 ⊗ P2 ⊗ P3 )A = A .
(15)
and
By Lemma 1, R(·) defined by (13) is weakly decomposable with respect to A(P1 , P2 , P3 )
and B(P1 , P2 , P3 ) with cR = 1/2. We note that a counterexample is also given by Yuan and
Zhang (2014) which shows that, for the tensor nuclear norm, we cannot take cR = 1.
Another popular way to define tensor rank is through the so-called Tucker decomposition.
Recall that the Tucker decomposition of a tensor A ∈ Rd1 ×d2 ×d3 is of the form:
Aj1 j2 j3 =
r1 X
r2 X
r3
X
Sk1 k2 k3 Uj1 k1 Vj2 k2 Wj3 k3
(16)
k1 =1 k2 =1 k3 =1
so that U , V and W are orthogonal matrices, and the so-called core tensor S = (Sk1 k2 k3 )k1 ,k2 ,k3
is such that any two slices of S are orthogonal. The triplet (r1 , r2 , r3 ) are referred to as
the Tucker ranks of A. It is not hard to see that if (12) holds, then the Tucker ranks
9
(r1 , r2 , r3 ) can be equivalently interpreted as the dimensionality of the linear spaces spanned
by {uk : 1 ≤ k ≤ r}, {vk : 1 ≤ k ≤ r}, and {wk : 1 ≤ k ≤ r} respectively. The following
relationship holds between CP rank and Tucker ranks:
max{r1 , r2 , r3 } ≤ r ≤ min{r1 r2 , r2 r3 , r1 r3 }.
A convenient way to encourage low Tucker ranks in a tensor is through matricization.
Let M1 (·) denote the mode-1 matricization of a tensor. That is M1 (A) is a d1 × (d2 d3 )
matrix whose column vectors are the the mode-1 fibers of A ∈ Rd1 ×d2 ×d3 . M2 (·) and M3 (·)
can also be defined in the same fashion. It is clear
rank(Mk (A)) = rk (A).
A natural way to encourage low-rankness is therefore through nuclear norm regularization:
3
1X
kMk (A)k∗ .
R(A) =
3 k=1
(17)
By the pinching inequality for matrices, R(·) defined by (17) is also decomposable with
respect to A(P1 , P2 , P3 ) and B(P1 , P2 , P3 ) with cR = 1.
3
Risk Bounds for Decomposable Regularizers
We now establish risk bounds for general decomposable regularizers. In particular, our
bounds are given in terms of the Gaussian width of a suitable set of tensors. Recall that the
Gaussian width of a set S ⊂ Rd1 ×d2 ×...×dN is given by
wG (S) := E suphA, Gi ,
A∈S
where G ∈ Rd1 ×d2 ×...×dN is a tensor whose entries are independent N (0, 1) random variables.
See, e.g. Gordon (1988) for more details on Gaussian width.
Note that the Gaussian width is a geometric measure of the volume of the set S and can
be related to other volumetric characterizations (see, e.g., Pisier, 1989).We also define the
unit ball for the norm-regularizer R(.) as follows:
BR (1) := {A ∈ Rd1 ×d2 ×...×dN | R(A) ≤ 1}.
10
We impose the mild assumption that kAkF ≤ R(A) which ensures that the regularizer R(·)
encourages low-dimensional structure.
Now we define a quantity that relates the size of the norm R(A) to the Frobenius norm
kAkF over the the low-dimensional subspace A. Following Negahban et al. (2012), for a
subspace A of Rd1 ×···×dN , define its compatibility constant s(A) as
s(A) :=
R2 (A)
,
2
A∈A/{0} kAkF
sup
which can be interpreted as a notion of intrinsic dimensionality of A.
Now we turn our attention to the covariate tensor. Denote by X (i) = vec(X (i) ) the
vectorized covariate from the ith sample. With slight abuse of notation, write
X = vec((X (1) )> , . . . , (X (n) )> ) ∈ Rn.d1 d2 ···dM
the concatenated covariates from all n samples. For convenience let DM = d1 d2 · · · dM .
Further for brevity we assume a Gaussian design so that
X ∼ N (0, Σ)
where
Σ = cov(X, X) ∈ RnDM ×nDM .
With more technical work our results may be extended beyond Gaussian designs. We note
that we do not require that the sample tensors X (i) be independent.
We shall assume that Σ has bounded eigenvalues which we later verify for a number of
statistical examples. Let λmin (·) and λmax (·) represent the smallest and largest eigenvalues
of a matrix, respectively. In what follows, we shall assume that
c2` ≤ λmin (Σ) ≤ λmax (Σ) ≤ c2u ,
(18)
for some constants 0 < c` ≤ cu < ∞.
Note that in particular if all covariates {X (i) : i = 1, . . . , n} are independent and identically distributed, then Σ has a block diagonal structure, and (18) boils down to similar
conditions on cov(X (i) , X (i) ). However (18) is more general and applicable to settings in
which the X (i) ’s may be dependent such as time-series models, which we shall discuss in
further detail in Section 4.
11
We are now in position to state our main result on the risk bounds in terms of both
Frobenius norm k · kF and the empirical norm k · kn where for a tensor A ∈ Rd1 ×···×dN , which
we define as:
n
kAk2n
1X
khA, X (i) ik2F .
:=
n i=1
The main reason we focus on random Gaussian design is so that we can prove a one-sided
uniform law that relates the empirical norm defined above to the Frobenius norm of a tensor in A (see Lemma 12). Lemma 12 is analogous to restricted strong convexity defined
in Negahban et al. (2012) but since we are dealing with multiple responses a more refined
technique is required to prove Lemma 12.
Theorem 1. Suppose that (1) holds for a tensor T from a linear subspace A0 ⊂ Rd1 ×···×dN
where (18) holds. Let Tb be defined by (2) where the regularizer R(·) is decomposable with
respect to A and A0 for some linear subspace A ⊇ A0 . If
λ≥
2σcu (3 + cR )
√
wG [BR (1)],
cR n
(19)
2
then there exists a constant c > 0 such that with probability at least 1 − exp{−cwG
[BR (1)]},
n
o 6(1 + c ) 9c2
R
u
max kTb − T k2n , kTb − T k2F ≤
s(A)λ2 ,
3 + cR c2`
(20)
when n is sufficiently large, assuming that the right hand side converges to zero as n increases.
As stated in Theorem 1, our upper bound boils down to bounding two quantities, s(A)
and wG [BR (1)] which are both purely geometric quantities. To provide some intuition,
wG [BR (1)] captures how large the R(·) norm is relative to the k · kF norm and s(A) captures
the low dimension of the subspace A.
Several technical remarks are in order. Note that wG [BR (1)] can be expressed as expectation of the dual norm of G. According to R (see, e.g., Rockafellar, 1970, for details), the
dual norm R∗ (·) is given by:
R∗ (B) :=
sup hA, Bi,
A∈BR (1)
where the supremum is taken over tensors of the same dimensions as B. It is straightforward
to see that wG [BR (1)] = E[R∗ (G)].
12
To the best of our knowledge, this is the first general result that applies to multiple
responses. As mentioned earlier, incorporating multiple responses presents a technical challenge (see Lemma 12) which is a one-sided uniform law analogous to restricted strong convexity. While Theorem 1 focusses on Gaussian design, results can be extended to random
sub-Gaussian design using more sophisticated techniques (see e.g., Mendelson, 2014; Zhou,
2009) or for fixed design by assuming covariates deterministically satisfy the condition in
Lemma 12. Since the focus of this paper is on general dependence structure, we assume
random Gaussian design.
One important practical challenge is that σ 2 , cu and c` are typically unknown and these
clearly influence the choice of λ. This is a common challenge for high-dimensional statistical
inference and we don’t address this issue in this paper. In practice, λ is typically chosen
through cross-validation. A more sophisticated choice of λ based on estimation of σ 2 and
other constants remains an open question. Another important and open question is for what
choices of A0 is the upper bound optimal (up to a constant). In Section 4 we provide specific
examples in which we provide minimax lower bounds which match the upper bounds up to
constant. However as we see for low-rank tensor regression for low-rank tensor regression
discussed in Section 3.2, we are not aware of a convex regularizer that matches the minimax
lower bound.
Now we develop upper bounds on both quantities in different scenarios. As in the previous
section, we shall focus on third order tensor in the rest of the section for the ease of exposition.
3.1
Sparsity regularizers
We first consider sparsity regularizers described in the previous section.
3.1.1
Entry-wise and fiber-wise sparsity
Recall that vectorized `1 regularizer:
R1 (A) =
d1 X
d2 X
d3
X
j1 =1 j2 =1 j3 =1
13
|Aj1 j2 j3 |,
could be used to exploit entry-wise sparsity. Clearly,
R∗1 (A) = max |Aj1 j2 j3 |.
j1 ,j2 ,j3
It can then be shown that:
Lemma 2. There exists a constant 0 < c < ∞ such that
p
wG [BR1 (1)] ≤ c log(d1 d2 d3 ).
Let
(
Θ1 (s) =
A ∈ Rd1 ×d2 ×d3 :
d3
d2 X
d1 X
X
(21)
)
I(Aj1 j2 j3 6= 0) ≤ s .
j1 =1 j2 =1 j3 =1
For an arbitrary A ∈ Θ1 (s), write
I(A) = {(j1 , j2 , j3 ) ∈ [d1 ] × [d2 ] × [d3 ] : Aj1 j2 j3 6= 0} .
Then R1 (·) is decomposable with respect to A(I(A)) as defined by (5). It is easy to verify
that for any A ∈ Θ1 (s),
s1 (A(I)) =
R21 (B)
≤ s.
2
B∈A(I(A))/{0} kBkF
sup
(22)
In light of (22) and (21), Theorem 1 implies that
n
o s log(d d d )
1 2 3
2
2
b
b
sup max kT1 − T kn , kT1 − T kF .
,
n
T ∈Θ1 (s)
with high probability by taking
r
λ
log(d1 d2 d3 )
,
n
where Tb1 is the regularized least squares estimate defined by (2) when using regularizer R1 (·).
A similar argument can also be applied to fiber-wise sparsity. To fix ideas, we consider
here only sparsity among mode-1 fibers. In this case, we use a group Lasso type of regularizer:
R2 (A) =
d2 X
d3
X
kA·j2 j3 k`2 .
j2 =1 j3 =1
Then
R∗2 (A) = max kA·j2 j3 k`2 .
j2 ,j3
14
Lemma 3. There exists a constant 0 < c < ∞ such that
wG [BR2 (1)] ≤ c
Let
p
(
Θ2 (s) =
A ∈ Rd1 ×d2 ×d3 :
max{d1 , log(d2 d3 )}.
d3
d2 X
X
(23)
)
I(A·j2 j3 6= 0) ≤ s .
j2 =1 j3 =1
Similar to the previous case, for an arbitrary A ∈ Θ1 (s), write
I(A) = {(j2 , j3 ) ∈ [d2 ] × [d3 ] : A·j2 j3 6= 0}.
Then R2 (·) is decomposable with respect to A(I(A)) as defined by (7). It is easy to verify
that for any A ∈ Θ2 (s),
s2 (A(I)) =
R22 (B)
≤ s.
2
B∈A(I(A))/{0} kBkF
sup
(24)
In light of (24) and (30), Theorem 1 implies that
n
o s max{d , log(d d )}
1
2 3
2
2
b
b
sup max kT2 − T kn , kT2 − T kF .
,
n
T ∈Θ2 (s)
with high probability by taking
r
λ
max{d1 , log(d2 d3 )}
,
n
where Tb2 is the regularized least squares estimate defined by (2) when using regularizer R2 (·).
Comparing with the rates for entry-wise and fiber-wise sparsity regularization, we can
see the benefit of using group Lasso type of regularizer R2 when sparsity is likely to occur
at the fiber level. More specifically, consider the case when there are a total of s1 nonzero
entries from s2 nonzero fibers. If an entry-wise `1 regularization is applied, we can achieve
the risk bound:
s1 log(d1 d2 d3 )
.
n
On the other hand, if fiber-wise group `1 regularization is applied, then the risk bound
kTb1 − T k2F .
becomes:
s2 max{d1 , log(d2 d3 )}
.
n
When nonzero entries are clustered in fibers, we may expect s1 s2 d1 . In this case, Tb2 enjoys
performance superior to that of Tb1 since s2 d1 log(d1 d2 d3 ) is larger than s2 max{d1 , log(d2 d3 )}.
kTb2 − T k2F .
15
3.1.2
Slice-wise sparsity and low-rank structure
Now we consider slice-wise sparsity and low-rank structure. Again, to fix ideas, we consider
here only sparsity among (1, 2) slices. As discussed in the previous section, two specific types
of regularizers could be employed:
R3 (A) =
d3
X
kA··j3 kF ,
j3 =1
and
R4 (A) =
d3
X
kA··j3 k∗ ,
j3 =1
where recall that k.k∗ denotes the nuclear norm of a matrix, that is the sum of all singular
values.
Note that
R∗3 (A) = max kA··j3 kF .
1≤j3 ≤d3
Then we have the following result:
Lemma 4. There exists a constant 0 < c < ∞ such that
wG [BR3 (1)] ≤ c
Let
p
max{d1 d2 , log(d3 )}.
(
Θ3 (s) =
A ∈ Rd1 ×d2 ×d3 :
d3
X
(25)
)
I(A··j3 6= 0) ≤ s .
j3 =1
For an arbitrary A ∈ Θ1 (s), write
I(A) = {j3 ∈ [d3 ] : A··j3 6= 0}.
Then R3 (·) is decomposable with respect to A(I(A)) as defined by (9). It is easy to verify
that for any A ∈ Θ3 (s),
s3 (A(I(A))) =
R23 (B)
≤ s.
2
B∈A(I(A))/{0} kBkF
sup
Based on (26) and (25), Theorem 1 implies that
n
o s max{d d , log(d )}
1 2
3
sup max kTb3 − T k2n , kTb3 − T k2F .
,
n
T ∈Θ3 (s)
16
(26)
with high probability by taking
r
λ
max{d1 d2 , log(d3 )}
,
n
where Tb3 is the regularized least squares estimate defined by (2) when using regularizer R3 (·).
Alternatively, for R4 (·),
R∗4 (A) = max kA··j3 ks ,
j3
we have the following:
Lemma 5. There exists a constant 0 < c < ∞ such that
wG [BR4 (1)] ≤ c
Now consider
p
max{d1 , d2 , log(d3 )}.
(
Θ4 (r) =
A ∈ Rd1 ×d2 ×d3 :
d3
X
(27)
)
rank(A··j3 ) ≤ r .
j3 =1
For an arbitrary A ∈ Θ4 (r), denote by P1j and P2j the projection onto the row and column
space of A··j respectively. It is clear that A ∈ B(P1j , P2j : 1 ≤ j ≤ d3 ) as defined by (11).
In addition, recall that R4 is decomposable with respect to B(P1j , P2j : 1 ≤ j ≤ d3 ) and
A(P1j , P2j : 1 ≤ j ≤ d3 ) as defined by (10). It is not hard to see that for any A ∈ Θ4 (r),
A(P1j , P2j : 1 ≤ j ≤ d3 ) ⊂ Θ4 (2r), from which we can derive that:
Lemma 6. For any A ∈ Θ4 (r),
s4 (A(P1j , P2j : 1 ≤ j ≤ d3 )) ≤
R24 (B)
≤ 2r.
2
B∈A/{0} kBkF
sup
(28)
In light of (28) and (27), Theorem 1 implies that
n
o r max{d , d , log(d )}
1 2
3
sup max kTb4 − T k2n , kTb4 − T k2F .
,
n
T ∈Θ4 (r)
with high probability by taking
r
λ
max{d1 , d2 , log(d3 )}
,
n
where Tb4 is the regularized least squares estimate defined by (2) when using regularizer R4 (·).
17
Comparing with the rates for estimates with regularizers R3 and R4 , we can see the
benefit of using R4 when the nonzero slices are likely to be of low-rank. In particular,
consider the case when there are s1 nonzero slices and each nonzero slice has rank up to r.
Then applying R3 leads to risk bound:
kTb3 − T k2F .
s1 max{d1 d2 , log(d3 )}
,
n
whereas applying R4 leads to:
kTb4 − T k2F .
s1 r max{d1 , d2 , log(d3 )}
.
n
It is clear that Tb4 is a better estimator when r d1 = d2 = d3 .
3.2
Low-rankness regularizers
We now consider regularizers that encourages low rank estimates. We begin with the tensor
nuclear norm regularization:
R5 (A) = kAk∗ .
Recall that R∗5 (A) = kAks .
Lemma 7. There exists a constant 0 < c < ∞ such that
p
wG [BR5 (1)] ≤ c (d1 + d2 + d3 ).
(29)
Now let
Θ5 (r) = A ∈ Rd1 ×d2 ×d3 : max{r1 (A), r2 (A), r3 (A)} ≤ r .
For an arbitrary A ∈ Θ5 (r), denote by P1 , P2 , P3 the projection onto the linear space spanned
by the mode-1, -2 and -3 fibers respectively. As we argued in the previous section, R5 (·) is
weakly decomposable with respect to A(P1 , P2 , P3 ) and B(P1 , P2 , P3 ), and A ∈ B(P1 , P2 , P3 )
where A(P1 , P2 , P3 ) and B(P1 , P2 , P3 ) are defined by (14) and (15) respectively.
Lemma 8. For any A ∈ Θ5 (r),
R25 (B)
s5 (A(P1 , P2 , P3 )) =
sup
≤ r2 .
2
B∈A(P1 ,P2 ,P3 )/{0} kBkF
18
Lemmas 7 and 8 show that
n
o r2 (d + d + d )
1
2
3
2
2
b
b
sup max kT5 − T kn , kT5 − T kF .
,
n
T ∈Θ5 (r)
with high probability by taking
r
λ
d1 + d2 + d3
,
n
where Tb5 is the regularized least squares estimate defined by (2) when using regularizer R5 (·).
Next we consider the low-rankness regularization via matricization:
R6 (A) =
1
(kM1 (A)k∗ + kM2 (A)k∗ + kM3 (A)k∗ ) .
3
It is not hard to see that
R∗6 (A) = 3 max {kM1 (A)ks , kM2 (A)ks , kM3 (A)ks } .
Lemma 9. There exists a constant 0 < c < ∞ such that
p
wG [BR6 (1)] ≤ c max{d1 d2 , d2 d3 , d1 d3 }.
(30)
On the other hand,
Lemma 10. For any A ∈ Θ5 (r),
s6 (A(P1 , P2 , P3 )) =
R26 (B)
≤ r.
2
B∈A(P1 ,P2 ,P3 )/{0} kBkF
sup
Lemmas 9 and 10 suggest that
n
o r max{d d , d d , d d }
1 2 2 3 1 3
sup max kTb6 − T k2n , kTb6 − T k2F .
,
n
T ∈Θ5 (r)
with high probability by taking
r
λ
max{d1 d2 , d2 d3 , d1 d3 }
.
n
where Tb6 is the regularized least squares estimate defined by (2) when using regularizer R6 (·).
19
Comparing with the rates for estimates with regularizers R5 and R6 , we can see the
benefit of using R5 . For any T ∈ Θ5 (r), If we apply regularizer R5 , then
kTb5 − T k2F .
r2 (d1 + d2 + d3 )
.
n
This is to be compared with the risk bound for matricized regularization:
kTb6 − T k2F .
r max{d1 d2 , d2 d3 , d1 d3 }
.
n
Obviously Tb5 always outperform Tb6 since r ≤ min{d1 , d2 , d3 }. The advantage of Tb5 is typically
rather significant since in general r min{d1 , d2 , d3 }. On the other hand, Tb6 is more
amenable for computation.
Both upper bounds on Frobenius error on Tb5 and Tb6 are novel results and complement
the existing results on tensor completion Gandy et al. (2011); Mu et al. (2014) and Yuan and
Zhang (2014). Neither Tb5 nor Tb6 is minimax optimal and remains an interesting question as
to whether there exists a convex regularization approach that is minimax optimal.
4
Specific Statistical Problems
In this section, we apply our results to several concrete examples where we are attempting to
estimate a tensor under certain sparse or low rank constraints, and show that the regularized
least squares estimate Tb is typically minimiax rate optimal with appropriate choices of
regularizers. In particular we focus on the multi-response aspect of the general framework
to provide novel upper bounds and matching minimax lower bounds.
4.1
Multi-Response regression with large p
The first example we consider is the multi-response regression model:
(i)
Yk
=
p
m
X
X
(i)
(i)
Xj` Tj`k + k ,
j=1 `=1
where 1 ≤ i ≤ n represents the index for each sample, 1 ≤ k ≤ m represents the index for
each response and 1 ≤ j ≤ p represents the index for each feature. For the multi-response
20
regression problem we have N = 3, M = 2, d1 = d2 = m which represents the total number
of responses and d3 = p, which represent the total number of parameters.
Since we are in the setting where p is large but only a small number s are relevant, we
define the subspace:
p
X
(
T1 =
A∈R
m×m×p
|
)
I(kA··j kF 6= 0) ≤ s .
j=1
Furthermore for each i we assume X (i) ∈ Rm×p where each entry of X (i) , [X (i) ]k,j , corresponds
to the j th feature for the k th response. For simplicity, we assume the X (i) ’s are independent
e ∈ Rmp×mp . The penalty function we are considering is:
Gaussian with covariance Σ
R(A) =
p
X
kA··j kF ,
(31)
j=1
and the corresponding dual function applied to the i.i.d. Gaussian tensor G is:
R∗ (G) = max kG..j kF .
1≤j≤p
Theorem 2. Under the multi-response regression model with T ∈ T1 and independent Gause ≤ λmax (Σ)
e ≤ c2 , if
sian design where c2 ≤ λmin (Σ)
u
`
r
λ ≥ 3σcu
such that
max{m2 , log p}
,
n
√
sλ converges to zero as n increases, then there exist some constants c1 , c2 > 0
such that with probability at least 1 − p−c1
n
o c c2
2
max kTb − T k2n , kTb − T k2F ≤ 2u sλ2 ,
c`
when n is sufficiently large, where Tb is the regularized least squares estimate defined by (2)
with regularizer given by (31). In addition,
c3 σ 2 s max{m2 , log p/s}
min max kTe − T k2F ≥
,
c2u n
Te T ∈T1
for some constant c3 > 0, with probability at least 1/2, where the minimum is taken over all
estimators Te based on data {(X (i) , Y (i) ) : 1 ≤ i ≤ n}.
21
Theorem 2 shows that when taking
r
λ
max{m2 , log p}
,
n
the regularized least squares estimate defined by (2) with regularizer given by (31) achieves
minimax optimal rate of convergence over the parameter space T1 .
Alternatively, there are settings where the effect of covariates on the multiple tasks may
be of low rank structure. In such a situation, we may consider
(
)
p
X
m×m×p
T2 = A ∈ R
|
rank(A..j ) ≤ r .
j=1
An appropriate penalty function in this case is:
R(A) =
p
X
kA..j k∗ ,
(32)
j=1
and the corresponding dual function applied to G is:
R∗ (G) = max kG..j ks .
1≤j≤p
Theorem 3. Under the multi-response regression model with T ∈ T2 and independent Gause ≤ λmax (Σ)
e ≤ c2u , if
sian design where c2 ≤ λmin (Σ)
`
r
λ ≥ 3σcu
such that
√
max{m, log p}
,
n
rλ converges to zero as n increases, then there exist some constants c1 , c2 > 0
such that with probability at least 1 − p−c1 ,
n
o c c2
2
max kTb − T k2n , kTb − T k2F ≤ 2u rλ2
c`
when n is sufficiently large, where Tb is the regularized least squares estimate defined by (2)
with regularizer given by (32). In addition,
c3 σ 2 r max{m, log(p/r)}
min max kTe − T k2F ≥
,
c2u n
Te T ∈T2
for some constant c3 > 0, with probability at least 1/2, where the minimum is taken over all
estimators Te based on data {(X (i) , Y (i) ) : 1 ≤ i ≤ n}.
22
Again Theorem 3 shows that by taking
r
max{m, log p}
λ
,
n
the regularized least squares estimate defined by (2) with regularizer given by (32) achieves
minimax optimal rate of convergence over the parameter space T2 . Comparing with optimal
rates for estimating a tensor from T1 , one can see the benefit and importance to take advantage of the extra low rankness if the true coefficient tensor is indeed from T2 . As far as
we are aware, these are the first results that provide upper bounds and matching minimax
lower bounds for high-dimensional multi-response regression with sparse or low-rank slices.
As pointed out earlier, the challenge in going from scalar to multiple response is proving
Lemma 12 which is an analog of restricted strong convexity.
4.2
Multivariate sparse auto-regressive models
Now we consider the setting of vector auto-regressive models. In this case, our generative
model is:
X
(t+p)
=
p
X
Aj X (t+p−j) + (t) ,
(33)
j=1
where 1 ≤ t ≤ n represents the time index, 1 ≤ j ≤ p represents the lag index, {X (t) }n+p
t=0
is an m-dimensional vector, (t) ∼ N (0, σ 2 Im×m ) represents the additive noise. Note that
the parameter tensor T is an m × m × p tensor so that T··j = Aj , and Tk`j represents the
co-efficient of the k th variable on the `th variable at lag j. This model is studied by Basu and
Michailidis (2015) where p is relatively small (to avoid introducing long-range dependence)
and m is large. Our main results allow more general structure and regularization schemes
than those considered in Basu and Michailidis (2015).
Since we assume the number of series m is large, and there are m2 possible interactions
between the series we assume there are only s m2 interactions in total.
(
)
m X
m
X
T3 = A ∈ Rm×m×p |
I(Ak`· 6= 0) ≤ s .
(34)
k=1 `=1
The penalty function we are considering is:
R(A) =
m X
m
X
k=1 `=1
23
kAk`· k`2 ,
(35)
and the corresponding dual function applied to G is:
R∗ (G) = max kGk,`,. k`2 .
1≤k,`≤m
The challenge in this setting is that the X’s are highly dependent and we use the results
developed in Basu and Michailidis (2015) to prove that (18) is satisfied.
Prior to presenting the main results, we introduce concepts developed in Basu and Michailidis (2015) that play a role in determining the constants c2u and c2` which relate to the stability of the auto-regressive processes. A p-variate Gaussian time series is defined by its
auto-covariance matrix function
ΓX (h) = Cov(X (t) , X (t+h) ),
for all t, h ∈ Z. Further, we define the spectral density function:
∞
1 X
ΓX (`)e−i`θ , θ ∈ [−π, π].
fX (θ) :=
2π `=−∞
To ensure the spectral density is bounded, we make the following assumption:
M(fX ) := ess sup Λmax (fX (θ)) < ∞.
θ
Further, we define the matrix polynomial
A(z) = Im×m −
p
X
Aj z j
j=1
where {Aj }pj=1 denote the back-shift matrices, and z represents any point on the complex
plane. Note that for a stable, invertible AR(p) process,
fX (θ) =
1 −1 −iθ −1 −iθ
A (e )A (e ).
2π
We also define the lower extremum of the spectral density:
m(fX ) := ess inf Λmin (fX (θ)).
θ
Note that m(fX ) and M(fX ) satisfy the following bounds:
m(fX ) ≥
1
2πµmax (A)
,
and
24
M(fX ) ≤
1
2πµmin (A)
,
where
µmin (A) := min Λmin (A(z)A(z))
|z|=1
and
µmax (A) := max Λmax (A(z)A(z)).
|z|=1
From a straightforward calculation, we have that for any fixed ∆:
1
µmax
k∆k2F ≤ E k∆k2n ≤
1
µmin
k∆k2`2 .
(36)
Hence c2u = 1/µmin and c2` = 1/µmax . Now we state our main result for auto-regressive
models.
Theorem 4. Under the vector auto-regressive model defined by (33) with T ∈ T3 , if
s
max{p, 2 log m}
λ ≥ 3σ
,
nµmin
√
such that sλ converges to zero as n increases, then there exist some constants c1 , c2 > 0
such that with probability at least 1 − m−c1 ,
n
o cµ
2 max
2
2
b
b
max kT − T kn , kT − T kF ≤
sλ2 ,
µmin
when n is sufficiently large, where Tb is the regularized least squares estimators defined by (2)
with regularizer given by (35). In addition,
√
2
2 s max{p, log(m/ s)}
e
min max kT − T kF ≥ c3 µmin σ
,
n
Te T ∈T3
for some constant c3 > 0, with probability at least 1/2, where the minimum is taken over all
estimators Te based on data {X (t) : t = 0, . . . , n + p}.
Theorem 4 provides, to our best knowledge, the only lower bound result for multivariate
time series. The upper bound is also novel and is different from Proposition 4.1 in Basu and
Michailidis (2015) since we impose sparsity only on the large m directions and not over the p
lags, whereas Basu and Michailidis (2015) impose sparsity through vectorization. Note that
Proposition 4.1 in Basu and Michailidis (2015) follows directly from Lemma 2 with d1 = p
and d2 = d3 = m. Using the sparsity regularizer Basu and Michailidis (2015) vectorize
the problem and prove restricted strong convexity whereas since we leave the problem as a
multi-response problem, we requried the more refined technique used for proving Lemma 12.
25
4.3
Pairwise interaction tensor models
Finally, we consider the tensor regression (1) where T follows a pairwise interaction model.
More specifically, (X (i) , Y (i) ), i = 1, 2, . . . , n are independent copies of a random couple
X ∈ Rd1 ×d2 ×d3 and Y ∈ R such that
Y = hX, T i +
and
(13)
(12)
(23)
Tj1 j2 j3 = Aj1 j2 + Aj1 j3 + Aj2 j3 .
Here A(k1 ,k2 ) ∈ Rdk1 ×dk2 such that
A(k1 ,k2 ) 1 = 0,
(A(k1 ,k2 ) )> 1 = 0.
and
The pairwise interaction was used originally by Rendle et al. (2009); Rendle and SchmidtThieme (2010) for personalized tag recommendation, and later analyzed in Chen et al. (2013).
Hoff (2003) briefly introduced a single index additive model (amongst other tensor models)
which is a sub-class of the pairwise interaction model. The regularizer we consider is:
R(A) = kA(12) k∗ + kA(13) k∗ + kA(23) k∗ .
(37)
It is not hard to see that R defined above is decomposable with respect to A(P1 , P2 , P3 ) for
any projection matrices.
Let
(12)
(13)
(23)
T4 = {A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = Aj1 j2 + Aj1 j3 + Aj2 j3 , A(k1 ,k2 ) ∈ Rdk1 ×dk2 ,
A(k1 ,k2 ) 1 = 0,
and
(A(k1 ,k2 ) )> 1 = 0
max rank(A(k1 ,k2 ) ) ≤ r}.
k1 ,k2
For simplicity, we assume i.i.d. Gaussian design so c2` = c2u = 1.
Theorem 5. Under the pairwise interaction model with T ∈ T4 , if
r
max{d1 , d2 , d3 }
λ ≥ 3σ
,
n
26
such that
√
rλ converges to zero as n increases, then there exist constants c1 , c2 > 0 such
that with probability at least 1 − min{d1 , d2 , d3 }−c1 ,
n
o
2
2
b
b
max kT − T kn , kT − T kF ≤ c2 rλ2 ,
when n is sufficiently large, where Tb is the regularized least squares estimate defined by (2)
with regularizer given by (37). In addition,
c3 σ 2 r max{d1 , d2 , d3 }
min max kTe − T k2F ≥
,
n
Te T ∈T4
for some constant c3 > 0, with probability at least 1/2, where the minimum is taken over all
estimate Te based on data {(X (i) , Y (i) ) : 1 ≤ i ≤ n}.
As in the other settings, Theorem 5 establishes the minimax optimality of the regularized
least squares estimate (2) when using an appropriate convex decomposable regularizer. Since
this is single response and the norm involves matricization, this result is a straightforward
extension to earlier results.
5
Numerical Experiments
In this section, we provide a series of numerical experiments that both support our theoretical results and display the flexibility of our general framework. In particular, we consider several different models including: third-order tensor regression with a scalar response
(Section 5.1); fourth-order tensor regression (Section 5.2); matrix-response regression with
both group sparsity and low-rankness regularizers (Section 5.3); multi-variate sparse autoregressive models (Section 5.4); and pairwise interaction models (Section 5.5). To perform
the simulations in a computationally tractable way, we adapt the block coordinate descent
approaches in multi-response case developed by Simon et al. (2013), and those developed by
Qin et al. (2013) for univariate response settings, to capture group sparsity and low-rankness
regularizers.
To fix ideas, in all numerical experiments, the covariate tensors X (i) s were independent
standard Gaussian ensembles (except for the multivariate auto-regressive models); and the
noise (i) s are i.i.d. random tensors with elements following N (0, σ 2 ) independently. As to
27
the choice of tuning parameter, we adopt grid search on λ to find the one with the least
estimation error (in terms of mean squared error) in all our numerical examples.
5.1
Third-order tensor regression
First we consider a third-order tensor regression model:
Y (i) = hB, X (i) i + (i)
where B ∈ Rd×d×d ,
Y (i) , (i) ∈ R,
X (i) ∈ Rd×d×d . The regression coefficient tensor B
was generated as follows: the first s slices B··1 , . . . B··s are i.i.d standard normal ensembles;
and the remaining slices B··s+1 , . . . B··d3 are set to be zero. Naturally, we consider here the
group-sparsity regularizer:
(
min
A∈Rd×d×d
n
d
X
1 X (i)
(i) 2
kY − hA, X ikF + λ
kA··j3 kF
2n i=1
j =1
)
.
3
Figure 1 shows the mean squared error of the estimate averaged over 50 runs (with standard
20
1.5
0.3
15
n
5000
10
4000
mse
4000
8
mse
mse
n
d
3000
1.0
4500
10
0.2
5000
12
0.5
5
0.1
0
0.0
8
10
12
d
14
16
3000
3500
4000
n
4500
5000
2
3
4
s
5
6
Figure 1: Mean squared error of the group-sparsity regularization for third order tensor
regression. The plot was based on 50 simulation runs and the error bars in each panel
represent ± one standard deviation.
deviation) versus d, n and s respectively. In the left and middle panels, we set s = 2,
whereas in the right panel, we fixed d = 16. As we can observe, the mean squared error
28
increases approximately according to d2 , s, and 1/n which agrees with the risk bound given
in Lemma 4.
We also considered a setting where B is slice-wise low-rank. More specifically, the s
nonzero slices B··1 , . . . B··s were random rank-r matrices. In this case, the slice-wise lowrankness regularizer can be employed:
)
(
d3
n
X
1 X (i)
kA··j3 k∗ .
kY − hA, X (i) ik2F + λ
min
2n i=1
A∈Rd1 ×d2 ×d3
j =1
3
The performance of the estimate, averaged over 50 simulation runs, is summarized by Figure
5
3
0.6
4
2
n
d
0.4
12
4000
4000
mse
3000
n
3
10
mse
mse
2000
4500
14
5000
2
1
0.2
1
0
8
10
12
d
14
16
3000
3500
4000
n
4500
5000
2
3
4
r
5
6
Figure 2: Mean squared error for third order tensor regression with low-rank slices tensor
coefficients. The plot was based on 50 simulation runs and the error bars in each panel
represent ± one standard deviation.
2 where in the left and middle panels r = 2, and in the right panel, d = 16. Once again, our
results are consistent with our theoretical results.
5.2
Fourth-order tensor regression
Although we have focused on third order tensors for brevity, our treatment applies to higher
order tensors as well. As an illustration, we now consider fourth order models where B ∈
Rd×d×d×d ,
Y (i) , (i) ∈ R,
X (i) ∈ Rd×d×d×d .
29
To generate low-rank fourth-order tensors, we impose low CP rank as follows: generate
four independent groups of r independent random vectors of unit length, {uk,1 }rk=1 , {uk,2 }rk=1 ,
{uk,3 }rk=1 and {uk,4 }rk=1 via performing an SVD of Gaussian random matrix two times and
keeping the r pairs of leading singular vectors, and then compute the outer-product yielding
P
a rank-r tensor B = rk=1 uk,1 ⊗ uk,2 ⊗ uk,3 ⊗ uk,4 .
We consider two different regularization schemes. First we impose low-rank structure
through mode-1 matricization:
(
min
A∈Rd×d×d×d
n
1 X (i)
kY − hA, X (i) ik2F + λkM1 (A)k∗
2n i=1
)
.
Secondly we use the square matricization as follows:
(
)
n
1 X (i)
kY − hA, X (i) ik2F + λkM12 (A)k∗ ,
min
2n i=1
A∈Rd×d×d×d
where M12 (·) reshape a fourth order tensor into a d2 × d2 matrix by collapsing its first two
indices, and last two indices respectively. Table 1 shows the average root-mean-square error
(RMSE, for short) for both approaches. As we can see, the 2-by-2 approach appears superior
to the 1-by-3 approach which is also predicted by the theory.
n
d
r
2000
7
2000
7
σ
SNR
RMSE (Mode-1 Matricization)
RMSE (Square Matricization)
5 10 3.0 (0.1)
0.53 (0.01)
0.51 (0.01)
3 10 1.5 (0.1)
0.58 (0.02)
0.49 (0.02)
4000 10 3 10 1.7 (0.1)
0.67 (0.01)
0.51 (0.02)
Table 1: Tensor regression with fourth order tensor covariates and scale response based on
matricization: RMSE were computed based on 50 simulations runs. Numbers in parentheses
are standard errors.
5.3
Matrix-response regression
Our general framework can handle matrix-responses in a seamless fashion. For demonstration, we consider here matrix-response regression with both group sparsity and low-rankness
regularizer. More specifically, the following model was considered:
Y (i) = hB, X (i) i + (i)
30
where B ∈ Rd×d×d ,
Y (i) , (i) ∈ Rd×d ,
X (i) ∈ Rd . As before, to impose group sparsity, the
first s slices of B were generated as Gaussian ensembles and the remaining slices were set to
zero.
For both the group sparsity and low-rankness regularizers, we used the matrix-version
algorithm for group-penalized multi-response regression in Simon et al. (2013). For each
block of the coordinate descent, the sub-problem with both `1 and nuclear norm penalty
have closed-form solutions.
6
6
7.5
5
4
n
2000
n
1000
20
mse
mse
4
d
1000
25
mse
5.0
2000
3
3000
2.5
3000
30
2
2
1
0.0
10
15
20
d
25
30
1000
1500
2000
n
2500
3000
2
4
6
s
8
10
Figure 3: Matrix response regression with sparse slices tensor coefficients. The plot was
based on 50 simulation runs and the error bars in each panel represent ± one standard
deviation.
Figure 3 shows the average (with standard deviation) mean squared error over 50 runs
versus the d, n and s parameter. (Here d1 = d2 = d3 = d). As we observe, the mean-squared
error increase approximately according to log d, s, and 1/n which supports our upper bound
in Theorem 2.
We also generated low-rank B in the same fashion as before. Figure 4 plots the average
(with standard deviation) mean squared error against d, n and r respectively. These results
are consistent with the main result in Theorem 3.
31
10.0
10.0
6
7.5
7.5
n
d
2000
25
5.0
3000
1000
mse
5.0
n
20
mse
mse
1000
2000
4
30
3000
2.5
2
2.5
0.0
10
15
20
d
25
30
1000
1500
2000
n
2500
3000
2
4
6
r
8
10
Figure 4: Matrix response regression with low-rank slices tensor coefficients. The plot was
based on 50 simulation runs and the error bars in each panel represent ± one standard
deviation.
5.4
Multivariate sparse auto-regressive models
Now we consider dependent covariates and responses through the multivariate auto-regressive
model. Recall that the generative model is:
X
(t+p)
=
p
X
B··j X (t+p−j) + (t) ,
j=1
(n+p)
where 1 ≤ t ≤ n represents the time index, 1 ≤ j ≤ p represents the lag index, {X (t) }t=0
is an m-dimensional vector, (t) ∼ σN (0, Im×m ) represents the additive noise.
We consider four different low-dimensional structures for B and we choose the entries of
B to be sufficiently small to ensure the time series is stable.
• Slice-wise sparsity: B··1 , . . . B··s are s non-zero slices of diagonal matrix, where diagonal
elements are constants ρ with ρ = 2. B··s+1 , . . . B··d3 are zero slices.
• Sparse low-rank slices: B··1 , . . . B··s are s non-zero slices, which are independent random
rank-r matrix (truncated matrix with i.i.d elements from N (0, τ 2 )). B··s+1 , . . . B··d3 are
zero slices. Here τ = 1/150 for m = 10 and τ = 1/500 for m = 20.
32
• Group Sparsity by lag (sparse normal slices):B··1 , . . . B··s are s non-zero slices, where
elements follow i.i.d N (0, τ 2 ) with τ = 0.05. B··s+1 , . . . B··d3 are zero slices.
• Group sparsity by coordinate (sparse normal fibers):
Bs1 s2 · is a vector of i.i.d normal elements following N (0, τ 2 ) (τ = 0.1) when (s1 , s2 ) ∈ S,
which is a random sample of size s from {1, . . . , m} × {1, . . . , m}, and zero otherwise.
Table 2 shows the average rmse for 50 runs of each case as a function of m, p, s and r.
In general, the smaller the n is, or the larger the m(or p) is, the harder it is to recover the
coefficient B. These findings are consistent with our theoretical developments.
5.5
Pairwise interaction tensor models
Finally we consider the so-called pairwise interaction tensor models as described in Section 4.3. To implement this regularization scheme we kept iterating among the matrix slices
A1,2 , A1,3 and A2,3 and updating one of the three at a time while assuming the other two
components are fixed. For the update of Ak1 ,k2 , we conducted an approximated projection
onto the zero-row-sum/zero-column-sum subspace after each generalized gradient descent
(soft thresholding) step.
(i+1)
(i)
Ak1 ,k2 = P̂ (P̂λη (Ak1 ,k2 − η∇f ))
where η is the step size for the gradient step, ∇f is the gradient of the least square objective
function, P̂λη is the singular space soft-thresholding operator with threshold λη, and P̂ is
the approximated projection operator that make any given matrix have zero row sums (by
shifting rows) and zero column sums (by shifting columns). We simulated independent
random low-rank matrix Ck1 ,k2 s and make them have zero column sums and row sums by
Bk1 ,k2 = P̂ (Ck1 ,k2 ).
Table 3 shows the average (with standard deviation) rmse under different r, d, n combinations under 50 runs. In general, the rmse in estimating the tensor coefficient increases as
s and d increases.
33
34
s diagonal slices
s diagonal slices
s rank-r slices
s rank-r slices
s Gaussian slices
s Gaussian slices
s Gaussian fibers
s Gaussian fibers
Vectorized
Sparsity
Low-rank
Slices
Group Sparsity
(by lag)
Group Sparsity
(by coordinate)
m
p
2000 20 20
2000 10 10
2000 20 20
2000 10 10
2000 20 20
2000 10 10
2000 20 20
2000 10 10
n
2
2
σ
s = 10
s = 10
s=5
s=5
0.02
0.02
0.2
0.2
s = 10, r = 5 0.05
s = 10, r = 5 0.05
s=5
s=5
s/r
1.2 (0.1)
0.93 (0.2)
0.42 (0.01)
0.42 (0.02)
1.2 (0.1)
1.0 (0.1)
2.9 (0.7)
5.5 (1.4)
SNR (sd)
0.72 (0.07)
0.35 (0.04)
0.64 (0.02)
0.40 (0.01)
0.82 (0.02)
0.37 (0.01)
0.25 (0.04)
0.43 (0.15)
rmse (sd )
simulations runs. Numbers in parentheses are standard errors.
Table 2: Multivariate auto-regressive model with various sparsity/low-rankness: RMSE were computed based on 50
coefficient tensor
Regularizer
n
d1
d2
d3
s
σ
RMSE
SNR
2000 40 40 40
5
10 0.54 (0.02)
2.4 (0.1)
2000 40 40 40
10 10 0.70 (0.01)
3.4 (0.1)
2000 20 20 20
5
10 0.39 (0.01)
1.7 (0.1)
2000 20 20 20
10 10 0.37 (0.01)
2.3 (0.1)
1000 20 20 20
5
10 0.58 (0.02)
1.6 (0.1)
1000 20 20 20
10 10 0.63 (0.02)
2.3 (0.1)
Table 3: Pairwise interaction model: RMSE were computed based on 50 simulations runs.
Numbers in parentheses are standard errors.
6
Proofs
In this section, we present the proofs to our main results. We begin with the proof of
Theorem 1.
6.1
Proof of Theorem 1
Our proof involves the following main steps:
• In the initial step, we use an argument similar to those developed in Negahban et al.
(2012) to exploit weak decomposability and properties of the empirical risk minimizer
and convex duals to upper bound kTb − T kn in terms of R(Tb − T ), and λ.
• Next, we use properties of Gaussian random variables and supremum of Gaussian
processes to express the lower bound on λ in terms of the Gaussian width E[R∗ (G)]
(Lemma 11 below).
• The final and most challenging step involves proving a one-sided uniform law relating
kTb − T kn to kTb − T kF (Lemma 12 below) which is analogous to restricted strong
convexity. The proof for Lemma 12 uses a novel truncation argument and is similar
in spirit to that of Lemma 4 in Raskutti et al. (2012). Lemma 12 is necessary to
incorporate multiple responses as existing results relating the k · kn to the population
35
k · kF norm (e.g., Dasgupta and Gupta, 2010; Raskutti et al., 2010; van de Geer, 2000)
only apply to univariate functions.
Throughout R(A) refers to the weakly decomposable regularizer over the tensor A. For a
tensor A, we shall write A0 and A⊥ as its projections onto A0 and A⊥ with respect to the
Frobenius norm, respectively.
Since Tb is the empirical minimizer,
n
n
1 X (i)
1 X (i)
kY − hX (i) , Tbik2F + λR(Tb) ≤
kY − hX (i) , T ik2F + λR(T ).
2n i=1
2n i=1
Substituting Y (i) = hX (i) , T i + (i) and ∆ = Tb − T ,
n
n
1 X
1 X (i)
(i)
2
khX , ∆ikF ≤
h ⊗ X (i) , ∆i + λ(R(T ) − R(Tb))
2n i=1
n i=1
!
n
X
1
≤ R∗
(i) ⊗ X (i) R(∆) + λ(R(T ) − R(Tb0 ) − cR R(Tb⊥ ))
n i=1
!
n
X
1
(i) ⊗ X (i) R(∆) + λ(R(∆0 ) − cR R(∆⊥ )),
≤ R∗
n i=1
where the second inequality follows from the decomposability and the last one follows from
triangular inequality.
Let G ∈ Rd1 ×d2 ×...×dN be an tensor where each entry is i.i.d. N (0, 1). Recall the definition
of Gaussian width:
wG [BR (1)] = E[R∗ (G)].
For simplicity let
ηR =
3 + cR
2cR
and recall that λ ≥ 2cu ηR n−1/2 E[R∗ (G)]. We have the following Lemma:
Lemma 11. If λ ≥ 2σcu ηR n−1/2 E[R∗ (G)], then
λ ≥ ηR R∗
!
n
1 X (i)
⊗ X (i) ,
n i=1
2
with probability at least 1 − exp{−ηR
E[R∗ (G)]2 /4}
36
The proof relies on Gaussian comparison inequalities and concentration inequalities.
Proof of Lemma 11. Recall that we have set:
λ≥
2σηR cu
√
E[R∗ (G)].
n
First we show that λ ≥ σcu ηR n−1/2 R∗ (G) with high probability using concentration of
Lipschitz functions for Gaussian random variables (see Theorem 7 in Appendix A). First we
prove that f (G) = R∗ (G) = supA∈BR (1) hG, Ai is a 1-Lipschitz function in terms of G. In
particular note that:
f (G) − f (G0 ) =
sup hG0 , Ai.
sup hG, Ai −
A:R(A)≤1
A:R(A)≤1
e := arg maxA:R(A)≤1 hG, Ai. Then
Let A
sup hG, Ai −
A:R(A)≤1
e − sup hG0 , Ai
sup hG0 , Ai = hG, Ai
R(A)≤1
A:R(A)≤1
e − hG0 , Ai
e
≤ hG, Ai
e
≤ hG − G0 , Ai
sup hG − G0 , Ai
≤
A:R(A)≤1
sup hG − G0 , Ai
≤
A:kAkF ≤1
0
≤ kG − G kF ,
where recall that kAkF ≤ R(A) which implies the second last inequality. Therefore f (G) is a
1-Lipschitz function with respect to the Frobenius norm. Therefore, by applying Theorem 7
in Appendix A,
(
P
)
sup hG, Ai − E[ sup hG, Ai] > wG (BR (1))
A∈BR (1)
A∈BR (1)
Therefore
ηR σcu
λ ≥ √ R∗ (G)
n
2
with probability at least 1 − 2 exp{−wG
[BR (1)]/2}.
37
1 2
≤ 2 exp − wG [BR (1)] .
2
To complete the proof, we use a Gaussian comparison inequality between the supremum
P
of the process cu n−1/2 hG, Ai and n−1 ni=1 h(i) ⊗ X (i) , Ai over the set BR (1). Recall that:
!
*
+
n
n
X
X
1
1
(i) ⊗ X (i) = sup
A,
(i) ⊗ X (i) .
R∗
n i=1
n
A∈BR (1)
i=1
Recall that each (i) ∈ RdM +1 ×dM +2 ×...×dN is an i.i.d. standard centered Gaussian tensor
with each entry having variance σ 2 and vec(X) ∈ Rnd1 d2 ···dM is a Gaussian vector covariance
Σ ∈ R(nDM )×(nDM ) . First we condition on all the (i) ’s, which are indepedendent of the
X (i) ’s. Further let {w(i) : i = 1, . . . , n} be i.i.d. standard normal Gaussian tensors where
w(i) ∈ Rd1 ×d2 ×...×dM . First we condition on all the (i) ’s, which are indepedendent of the X (i) ’s.
Assuming (18) and using a standard Gaussian comparison inequality due to Lemma 14 in
Appendix A proven earlier in Anderson (1955), if we condition on the (i) ’s we get
)
(
)
(
n
n
X
1
x
1 X (i)
h ⊗ X (i) , Ai > x ≤ P
sup
h(i) ⊗ w(i) , Ai >
,
P
sup
n
n
cu
A:R(A)≤1
A:R(A)≤1
i=1
i=1
since
Cov(vec(X)) = Σ c2u I(nDM )×(nDM ) .
Now we define the wj ∈ Rn as the standard random vector where 1 ≤ j ≤ DM and wj =
(1)
(n)
(wj , ..., wj ). Conditioning on the w(i) ’s and dealing with the randomness in the (i) ’s,
n
1 X (i)
kwj k`2 σ
√ √ hG, Ai,
h ⊗ w(i) , Ai ≤ max
1≤j≤DM
n i=1
n
n
where G ∈ Rd1 ×d2 ×...×dN is an i.i.d. standard normal tensor, since the (i) ’s are i.i.d. standard
normal. Now we upper bound
max
1≤j≤DM
kwj k`2
√ ,
n
using standard χ2 tail bounds. Since kwj k2`2 is a χ2 random variable with n degrees of
freedom, for each j,
P
kwj k2`2
≥ 4 ≤ exp(−n),
n
using χ2 tail bounds provided in Appendix A presented in Laurent and Massart (1998). Now
taking the union bound over DM ,
kwj k`2
√
P
max
≥ 2 ≤ exp(log DM − n),
1≤j≤DM
n
38
and provided n ≥ 2 log DM , it follows that with probability greater than 1 − exp(−n/2),
max1≤j≤DM
kwj k`2
√
n
≤ 2. Therefore, with probability at least 1 − exp(−n/2),
n
1 X (i)
2σ
h ⊗ w(i) , Ai ≤ √ hG, Ai.
n i=1
n
Now we apply Slepian’s lemma (Slepian, 1962) to complete the proof. Slepian’s lemma
is stated in Appendix A. Applying Slepian’s lemma (Lemma 15 in Appendix A),
)
(
(
)
n
2σ
1 X (i)
(i)
h ⊗ w , Ai > x ≤ P
sup √ hG, Ai > x ,
P
sup
n
R(A)≤1
R(A)≤1 n i=1
for all x > 0. Substituting x by x/cu means that
!
)
(
n
X
2σcu ∗
1
(i)
(i)
∗
h ⊗ w , Ai > x ≤ P √ R (G) > x ,
P R
n i=1
n
for any x > 0. This completes the proof.
In light of Lemma 11, for the remainder of the proof, we can condition on the event that
!
n
X
1
λ ≥ ηR R∗
(i) ⊗ X (i) .
n i=1
Under this event,
n
1 X
1
λR(∆) + λ(R(∆0 ) − cR R(∆⊥ ))
khX (i) , ∆ik2F ≤
2n i=1
ηR
1
1
≤
1+
λR(∆0 ) − cR −
λR(∆⊥ ).
ηR
ηR
Since
n
1 X
kh∆, X (i) ik2F ≥ 0,
2n i=1
we get
R(∆⊥ ) ≤
3
R(∆0 ).
cR
Hence we define the cone
C = ∆ | R(∆⊥ ) ≤ 3c−1
R R(∆0 ) ,
39
and know that ∆ ∈ C. Hence
n
1 X
3(1 + cR )
3(1 + cR ) p
khX (i) , ∆ik2F ≤
λR(∆0 ) ≤
s(A)λk∆kF .
2n i=1
3 + cR
3 + cR
Recall that
n
1X
khX (i) , ∆ik2F = k∆k2n .
n i=1
Thus,
k∆k2n ≤
6(1 + cR ) p
s(A)λk∆kF .
3 + cR
For convenience, in the remainder of this proof let
δn :=
6(1 + cR ) p
s(A)λ.
3 + cR
Now we split into three cases. (i) If k∆kn ≥ k∆kF , then
max{k∆kn , k∆kF } ≤ δn .
On the other hand if (ii) k∆kn ≤ k∆kF and k∆kF ≤
max{k∆kn , k∆kF } ≤
cu
δ ,
c` n
then
cu
δn .
c`
Hence the only case we need to consider is (iii) k∆kn ≤ k∆kF and k∆kF ≥ cu c−1
` δn . Now we
follow a similar proof technique to the proof for Theorem 1 in Raskutti et al. (2012).
Let us define the following set:
C(δn ) := ∆ ∈ Rd1 ×d2 ×···×dN | R(∆⊥ ) ≤ 3c−1
R R(∆0 ), k∆kn ≤ k∆kF .
Further, let us define the event:
1
cu
2
2
E(δn ) := k∆kn ≥ k∆kF | ∆ ∈ C(δn ), k∆kF ≥ δn .
4
c`
Let us define the alternative event:
1
cu
E 0 (δn ) := {k∆k2n ≥ k∆k2F | ∆ ∈ C(δn ), k∆kF = δn }.
4
c`
40
We claim that it suffices to show that E 0 (δn ) holds with probability at least 1 − exp(−cn)
for some constant c > 0. In particular, given an arbitrary non-zero ∆ ∈ C(δn ), consider the
re-scaled tensor
e = c u δn ∆ .
∆
c` k∆kF
e ∈ C(δn ) and k∆k
e F = cu c−1 δn by
Since ∆ ∈ C(δn ) and C(δn ) is star-shaped, we have ∆
`
construction. Consequently, it is sufficient to prove that E 0 (δn ) holds with high probability.
Lemma 12. Assume that for any c > 0, there exists an n such that
√
sλ ≤ c. Then there
exists a c̃ > 0 such that
P E 0 (δn ) ≥ 1 − exp(−c̃n).
Proof of Lemma 12. Denote by DN = d1 d2 · · · dN and DM = d1 d2 · · · dM . Now we define the
random variable
(
Zn (C(δn )) = sup
∆∈C(δn )
n
c2u 2 1 X
δ −
kh∆, X (i) ik2F
c2` n n i=1
then it suffices to show that
Zn (C(δn )) ≤
Recall that the norm
)
,
c2u δn2
.
2c2`
n
k∆k2n
1X
kh∆, X (i) ik2F .
=
n i=1
To expand this out, reacall [M ] = {1, 2, ..., M } and define an extension of the standard
matricization
˜ := M[M ] (∆) ∈ RDM ×DN /DM ,
∆
which groups together the first M modes. With a slight abuse of notation, it follows that
k∆k2n
n DN /DM
1X X ˜
h∆m , vec(X (i) )i2 ,
=
n i=1 m=1
˜ m ∈ RDM , and clearly vec(X (i) ) ∈ RDM . In order to complete the proof we make use
where ∆
of a truncation argument. For a constant τ > 0 to be chosen later, consider the truncated
quadratic function
φτ (u) = min{u2 , τ 2 },
41
and define
q
˜ m , vec(X)i) φτ (h∆
˜ m , vec(X)i)
∆m,τ (X) = sign h∆
where X ∈ Rd1 ×d2 ×...×dM is an input tensor. Further let
∆τ (X) = (∆1,τ (X), ∆2,τ (X), ..., ∆DN /DM ,τ (X))
and
k∆m,τ k2F = E[∆2m,τ (X)],
k∆τ k2F =
PDN /DM
m=1
k∆m,τ k2F , and similarly
n
k∆m,τ k2n
k∆τ k2n =
PDN /DM
m=1
1X 2
∆m,τ (X (i) ),
=
n i=1
k∆m,τ k2n . By construction, for any ∆ ∈ C(δn ), k∆k2n ≥ k∆τ k2n and hence
k∆k2n ≥ k∆τ k2F − sup |k∆τ k2n − k∆τ k2F |
∆∈C(δn )
The remainder of the proof consists of showing that for a suitable of τ ,
3
k∆τ k2F ≥ k∆k2F , for all ∆ ∈ C(δn )
4
and
c2
P Zn ≥ u2 δn2
4c`
2
≤ c1 e−c2 nδn ,
where Zn := sup∆∈C(δn ) |k∆τ k2n − k∆τ k2F |. By definition
˜ m , vec(X)i2 1[|∆
˜ m , vec(X)i| ≥ τ ]
k∆m k2F − k∆m,τ k2F ≤ E h∆
q
q
4
˜
˜ m , vec(X)i| ≥ τ ]
≤
E[h∆m , vec(X)i ] P[|∆
˜ m , vec(X)i4 ]
E[h∆
≤
,
τ2
where the second last inequality follows from the Cauchy-Schwartz inequality and the fi˜ m , vec(X)i is a Gaussian random
nal inequality follows from Markov’s inequality. Since h∆
variable,
˜ m , vec(X)i4 ] ≤ 3E[h∆
˜ m , vec(X)i2 ] = 3k∆m k2 .
E[h∆
F
42
Therefore
k∆m k2F
−
≤9
.
τ2
Setting τ 2 = 36, k∆m k2F − k∆m,τ k2F ≤ 14 k∆m k2F . Summing over m implies
k∆m k2F
k∆m,τ k2F
1
k∆k2F − k∆τ k2F ≤ k∆k2F ,
4
which implies k∆τ k2F ≥ 34 k∆k2F . Now to prove the high probability bound on Zn by first upper
bounding E[Zn ]. A standard symmetrization argument (see e.g., Pollard, 1984), shows that
/DM
n DN
X
X
1
(i)
˜ m , vec(X (i) )i) ,
φτ (h∆
EX [Zn ] ≤ 2EX,z sup
zm
n
∆∈C(δn )
i=1 m=1
(i)
(i)
(i)
where zm are i.i.d. Rademacher random variables (that is P (zm = +1) = P (zm = −1) =
˜ m , vec(X (i) )i) is a Lipschitz function with Lipschitz constant 2τ , the
1/2). Since φτ (h∆
Ledoux-Talagrand contrtaction inequality (Ledoux and Talagrand, 1991) implies that ,
DN /DM
n
1 X X (i)
˜ m , vec(X (i) )i)
EX [Zn ] ≤ 2EX,z sup
zm φτ (h∆
∆∈C(δn ) n i=1 m=1
/DM
n DN
X
X
1
(i) ˜
≤ 8τ EX,z sup
zm
h∆m , vec(X (i) )i .
n
∆∈C(δn )
i=1 m=1
Using standard comparisons between Rademacher and Guassian complexities (see, e.g.,
Lemma 4 of Bartlett and Mendelson, 2002), there exists a C > 0 such that
/DM
n DN
X
X
1
(i) ˜
zm
h∆m , vec(X (i) )i
EX,z sup
∆∈C(δn ) n i=1 m=1
DN /DM
n
1X X
(i) ˜
wm
h∆m , vec(X (i) )i ,
≤ CEX,w sup
∆∈C(δn ) n i=1 m=1
(i)
where wm s (1 ≤ i ≤ n and 1 ≤ m ≤ DN /DM ) are independent standard normal random
variables.
Next we upper bound the Gaussian complexity
Ew
!
n
1 X (i)
sup
hw ⊗ X (i) , ∆i .
∆∈C(δn ) n i=1
43
Clearly,
n
1 X (i)
hw ⊗ X (i) , ∆i ≤ R∗
n i=1
!
n
X
1
λ
R(∆).
w(i) ⊗ X (i) R(∆) ≤
n i=1
ηR
by the definition of λ and our earlier argument. Since ∆ ∈ C(δn ),
p
p
λ
λ(1 + 3c−1
λ(1 + 3c−1
cu (1 + 3c−1
R )
R )
R )
R(∆) ≤
R(∆0 ) ≤
δn s(A)λ.
s(A)k∆0 kF ≤
ηR
ηR
ηR
c` ηR
Therefore,
Ew
!
n
p
cu (1 + 3c−1
1 X (i)
R )
δn s(A)λ.
hw ⊗ X (i) , ∆i ≤
sup
c` ηR
∆∈C(δn ) n i=1
Since
we have
Ew
p
cu (1 + 3c−1
2cu
R )
s(A)λ =
δn ,
c` ηR
c`
!
n
X
1
cu 2
sup
hw(i) ⊗ X (i) , ∆i ≤
δ .
c` ηR n
∆∈C(δn ) n i=1
Finally we need a concentration bound to show that
c2u 2
2
P Zn ≥ 2 δn ≤ c1 e−c2 nδn .
4c`
In particular using Talagrand’s theorem for empirical processes (Talagrand, 1996). By construction φτ (.) ≤ τ 2 = 36 and
DN /DM
X
˜ m , vec(X)i) ≤
Var
φτ (h∆
DN /DM
X
m=1
˜ m , vec(X)i)]
E[φ4τ (h∆
m=1
DN /DM
≤ τ
2
X
˜ m k2` ≤
k∆
2
m=1
36c2u δn2
.
c2`
Consequently Talagrand’s inequality implies that
c2 nu2
P Zn ≥ E[Zn ] + u ≤ c1 exp − 2
.
3δn + 9u2
Since E[Zn ] ≤
cu
δ2 ,
c` ηR n
the claim follows by setting u =
cu
δ2 .
c` η R n
Finally we return to the main proof. On the event E(δn ), it now follows easily that,
max{k∆k2`2 , k∆k2n } ≤
This completes the proof for Theorem 1.
44
ηR c2u
s(A)λ2 .
c2`
6.2
Proof of other results in Section 3
In this section we present proofs for the other main results from Section 3, deferring the
more technical parts to the appendix.
Proof of Lemmas 2, 3 and 4. We prove these three lemmas together since the proofs follow
a very similar argument. First let S ⊂ {1, 2, 3} denote the directions in which sparsity is
Q
applied and DS = k∈S dk denote the total dimension in all these directions. For example,
in Lemma 2 S = {1, 2, 3} and DS = d1 d2 d3 , for Lemma 3, S = {2, 3} and DS = d2 d3 and for
Lemma 4, S = {1} and DS = d1 . Recall N = {1, 2, 3} and DN = d1 d2 d3 .
Note that R∗ (G) can be represented by the variational form:
R∗ (G) =
dS1 ×...×dS|S|
where u ∈ R
and v ∈ R
hG, u ⊗ vi,
sup
kvec(u)k`1 ≤1,kvkF ≤1
dS c ×...×dS c
1
N −|S|
. Now we express the supremum of this
Gaussian process as:
sup vec(u)> MS (G)vec(v),
(u,v)∈V
where recall MS is the matricization involving either slice or fiber S. The remainder of the
proof follows from Lemma 16 in Appendix B.
Proof of Lemma 5. Recall that
R∗ (G) := max kG..j3 ks .
1≤j3 ≤d3
For each 1 ≤ j3 ≤ d3 , Lemma 17 in Appendix B with N = 2 satisfies the concentration
inequality
E[kG..j3 ks ] ≤
p
6(d1 + d2 ).
Applying standard bounds on the maximum of functions of independent Gaussian random
variables,
E[ max kG..j3 ks ] ≤
p
1≤j3 ≤d3
This completes the proof.
45
6(d1 + d2 + log d3 ).
Proof of Lemma 6. Using the standard nuclear norm upper bound for a matrix in terms of
rank and Frobenius norm:
R24 (A)
d3
X
=
!2
kA··j3 k∗
j3 =1
d3 q
X
rank(A··j3 )kA··j3 kF
≤
!2
j3 =1
≤
d3
X
rank(A··j3 )
j3 =1
d3
X
kA··j3 k2F =
j3 =1
d3
X
rank(A··j3 )kAk2F ,
j3 =1
where the final inequality follows from the Cauchy-Schwarz inequality. Finally, note that for
any A ∈ Θ4 (r)/{0},
d3
X
rank(A··j3 ) ≤ r,
j3 =1
which completes the proof.
Proof of Lemma 7. Note that R∗ (G) = kGks , we can directly apply Lemma 17 with N = 3
from Appendix B.
Proof of Lemma 8. From Tucker decomposition (16), it is clear that for any A ∈ Θ5 (r), we
can find sets of vectors {uk : k = 1, . . . , r2 }, {vk : k = 1, . . . , r2 } and {wk : k = 1, . . . , r2 }
such that
2
A=
r
X
uk ⊗ vk ⊗ w k ,
k=1
and in addition,
>
>
u>
k uk0 = (vk vk0 )(wk wk0 ) = 0
for any k 6= k 0 . It is not hard to see that
2
kAk2F
=
r
X
kuk k2`2 kvk k2`2 kwk k2`2 .
k=1
46
On the other hand, as shown by Yuan and Zhang (2014),
2
kAk∗ =
r
X
(kuk k`2 kvk k`2 kwk k`2 ) .
k=1
The claim then follows from an application of Cauchy-Schwartz inequality.
Proof of Lemma 9. Recall that we are considering the regularizer
R∗6 (A) = 3 max {kM1 (A)ks , kM2 (A)ks , kM3 (A)ks } ,
and our goal is to upper bound
R∗6 (G) = 3 max kMk (G)ks .
1≤k≤3
Once again apply Lemma 17 in Appendix B with N = 2 for each matricization implies
p p p
E[R∗6 (G)] ≤ 4 max( d1 , d2 , d3 ).
Proof of Lemma 10. It is not hard to see that
1
(kM1 (A)k∗ + kM2 (A)k∗ + kM3 (A)k∗ )2
9
√
√
1 √
( r1 + r2 + r3 )2 kAk2F
≤
9
≤ max{r1 (A), r2 (A), r3 (A)}kAk2F ,
R6 (A)2 =
which completes the proof.
6.3
Proof of results in Section 4
In this section we prove the results in Section 4. First we provide a general minimax lower
result that we apply to our main results. Let T ⊂ Rd1 ×d2 ×···×dN be an arbitrary subspace of
order-N tensors.
47
Theorem 6. Assume that (18) holds and there exists a finite set {A1 , A2 , . . . , Am } ∈ T of
tensors such that log m ≥ 128nδ 2 , such that
2 −2 2
`2 2
2
`1
nσ 2 c−2
u δ ≤ kA − A kF ≤ 8nσ cu δ ,
for all `1 6= `2 ∈ [m] and all δ > 0. Then
2
min max kTe − T k2F ≥ cσ 2 c−2
u δ ,
Te
T ∈T
with probability at least 1/2 for some c > 0.
Proof. We use standard information-theoretic techniques developed in Ibragimov and Has’minskii
(1981) and extended in Yang and Barron (1999). Let {A1 , A2 , . . . , Am } be a set such that
2
kA`1 − A`2 k2F ≥ nσ 2 c−2
u δ
for all `1 6= `2 , and let m
e be a random variable uniformly distributed over the index set
[m] = {1, 2, . . . , m}.
Now we use a standard argument which allows us to provide a minimax lower bound in
terms of the probability of error in a multiple hypothesis testing problem (see, e.g., Yang
and Barron, 1999; Yu, 1996) then yields the lower bound.
2
σ 2 c−2
u δ
2
e
e
inf sup P kT − T kF ≥
≥ inf P(Te 6= Am
)
2
Te T ∈T
Te
where the infimum is taken over all estimators Te that are measurable functions of X and Y .
Let X = {X (i) : i = 1, . . . , n}, Y = {Y (i) : i = 1, . . . , n} and E = {(i) : i = 1, . . . , n}.
Using Fano’s inequality (see, e.g., Cover and Thomas, 1991), for any estimator Te, we have:
e
P[Te 6= Am
|X] ≥ 1 −
e
IX (Am
; Y ) + log 2
.
log m
Taking expectations over X on both sides, we have
e
P[Te 6= Am
]≥1−
e
EX [IX (Am
; Y )] + log 2
.
log m
For ` = 1, 2, . . . , m, let Q` denote the condition distribution of Y conditioned on X and
the event {T = A` }, and DKL (Q`1 ||Q`2 ) denote the Kullback-Leibler divergence between Q`1
48
and Q`2 . From the convexity of mutual information (see, e.g., Cover and Thomas, 1991), we
have the upper bound
IX (T ; Y ) ≤
1
m
X
m
2
DKL (Q`1 ||Q`2 ).
`1 ,`2 =1
Given our linear Gaussian observation model (1),
n
1 X
nkA`1 − A`2 k2n
(i) 2
(i)
`2
`1
,
X
i
=
,
X
i
−
hA
hA
.
2σ 2 i=1
2σ 2
DKL (Q`1 ||Q`2 ) =
Further if (18) holds, then
EX [IX (T ; Y )] ≤
n
2σ 2
X
m
2
EX [kA`1 − A`2 k2n ] ≤ c2u
`1 6=`2
n
2σ 2
X
m
2
kA`1 − A`2 k2F .
`1 6=`2
Based on our construction, there exists a set {A1 , A2 , . . . , Am } where each A` ∈ T such
that log m ≥ Cnδ 2 and
`1
`2
−1
c−1
u δ ≤ kA − A kF ≤ 8cu δ
for all `1 6= `2 ∈ {1, 2, . . . , m}. If (18) holds, then
EX kA`1 − A`2 k2n ≤ c2u kA`1 − A`2 k2F
and we can conclude that
EX [IX (T ; Y )] ≤ 32c2u σ −2 nδ 2 ,
and from the earlier bound due to Fano’s inequality, for and δ > 0 such that
1
32c2u σ 2 nδ 2 + log 2
≤ ,
log m
2
we are guaranteed that
n
o 1
e
P Te =
6 Am
≥ .
2
2
The proof is now completed because log m ≥ 128nδ and 32nδ 2 ≥ log 2.
Proof of Theorem 2. The proof for the upper bound follows directly from Lemma 4 with
d1 = d2 = m and d3 = p and noting that the overall covariance Σ ∈ R(nDM )×(nDM ) is
e since each of the samples is independent. Hence
block-structured with blocks Σ
c2` ≤ λmin (Σ) ≤ λmax (Σ) ≤ c2u .
49
To prove the lower bound, we use Theorem 6 and construct a suitable packing set for
T1 . The way we construct this packing is to construct two separate packing sets and select
the set with the higher packing number using a similar argument to that used in Raskutti
et al. (2012) which also uses two separate packing sets. The first packing set we consider
involves selecting the s-dimensional slice A..S where A ⊂ [j3 ] and S = {1, 2, ..., s}. Consider
2
vectorizing each slice so v = vec(A..S ) ∈ Rsm . Hence in order to apply Theorem 6, we define
2
the set T to be slices which is isomorphic to the vector space Rsm . Using Lemma 18 in
2
Appendix C, there exists a packing set {v 1 , v 2 , ..., v N } ∈ Rsm such that log N ≥ csm2 and
for all v `1 , v `2 where `1 6= `2 ,
δ2
≤ kv `1 − v `2 k2F ≤ δ 2
4
√
√
for any δ > 0. If we choose δ = c sm/ n, then Theorem 6 implies the lower bound
2
sm
,
min max kTe − T k2F ≥ cc−2
u
n
Te T ∈T1
with probability greater than 1/2.
The second packing set we construct is for the slice A11· ∈ Rp . Since in the third direction
only s of the p co-ordinates are non-zero, the packing number for any slice is analogous to
the packing number for s-sparse vectors with ambient dimension p. Letting v = A11· , we
need to construct a packing set for
{v ∈ Rp | kvk`0 ≤ s}.
Using Lemma 19 in Appendix C, there exists a discrete set {v 1 , v 2 , ..., v N } such that log N ≥
cs log(p/s) for some c > 0 and
δ2
≤ kv k − v ` k2`2 ≤ δ 2
8
for k 6= ` for any δ > 0. Setting δ 2 = sn−1 log(p/s),
s log(p/s)
min max kTe − T k2F ≥ cc−2
,
u
T
∈T
n
1
Te
with probability greater than 1/2.
Taking a maximum over lower bounds involving both packing sets completes the proof
of the lower bound in Theorem 2.
50
Proof of Theorem 3. The upper bound follows directly from Lemma 5 with d1 = d2 = m
and d3 = p and noting that the overall covariance Σ ∈ R(nDM )×(nDM ) is block-structured with
e since each of the samples is independent.
blocks Σ
To prove the lower bound, we use Theorem 6 and construct a suitable packing set for T2 .
Once again we construct two separate packings and choose the set that leads to the larger
minimax lower bound. For our first packing set, we construct a packing along one slice. Let
us assume A = (A··1 , ..., A··p ), where rank(A··1 ) ≤ r and
A··2 = · · · = A··p = 0.
If we let A··1 = M where M ∈ Rm×m then A = (M, 0, .., 0) ∈ Rm×m×p . Using Lemma 20 in
Appendix C, there exists a set {A1 , A2 , ..., AN } such that log N ≥ crm and
δ2
≤ kA`1 − A`2 k2F ≤ δ 2
4
p
for all `1 6= `2 and any δ > 0. Here we set δ = rm/n. Therefore using Theorem 6
min max kTe − T k2F ≥ cσ 2 c−2
u
Te
T ∈T2
rm
,
n
with probability greater than 1/2.
The second packing set for T2 involves a packing in the space of singular values since
p
X
rank(A··j ) ≤ r.
j=1
Let {σjk : k = 1, . . . , m} be the singular values of the matrix A··j . Under our rank constraint,
we have
p
m
X
X
I(σjk 6= 0) ≤ s.
j=1 k=1
mp
Let v ∈ R
where
v = vec((σjk )1≤j≤p,1≤k≤m ).
Note that
p
m
X
X
I(σjk 6= 0) ≤ r
j=1 k=1
51
implies kvk`0 ≤ r. Using Lemma 19, there exists a set {v 1 , v 2 , ..., v N }, such that log N ≥
cr log(mp/r) and for all `1 6= `2 ,
δ2
≤ kv `1 − v `2 k2`2 ≤ δ 2
4
for any δ > 0. If we set δ 2 = rn−1 log(mp/r). Therefore using Theorem 6,
r log(mp/r)
min max kTe − T k2F ≥ cσ 2 c−2
,
u
n
Te T ∈T2
with probability greater than 1/2. Hence taking a maximum over both bounds,
r max{m, log(p/r), log m}
r max{m, log(p/r)}
= cσ 2 c−2
,
min max kTe − T k2F ≥ cσ 2 c−2
u
u
n
n
Te T ∈T2
with probability greater than 1/2.
Proof of Theorem 4. The upper bound with
s
max{p, 2 log m}
λ≥3
µmin n
follows directly from Lemma 3 with d1 = p and d2 = d3 = m and (18) is satisfied with
c2u = 1/µmin and c2` = 1/µmax according to (36).
To prove the lower bound is similar to the proof for the lower bound in Theorem 2. Once
again we use Theorem 6 and construct a two suitable packing sets for T3 . The first packing
set we consider involves selecting an arbitrary subspace
Te := {A = (Aj1 ,j2 ,j3 )j1 ,j2 ,j3 | 1 ≤ j1 ≤
√
√
s, 1 ≤ j2 ≤ s, 1 ≤ j3 ≤ p}.
Now if we let v = vec(A), then v comes from an sp-dimensional vector space for any A ∈ Te .
Using Lemma 18 in Appendix C, there exists a packing set {v 1 , v 2 , ..., v N } ∈ Rsp such that
log N ≥ csp and for all v `1 , v `2 where `1 6= `2 ,
δ2
≤ kv `1 − v `2 k2F ≤ δ 2
4
for any δ > 0. If we choose δ =
p
sp/n, then Theorem 6 implies the lower bound
sp
,
min max kTe − T k2F ≥ cσ 2 c−2
u
n
Te T ∈T3
52
with probability greater than 1/2. Further c2u = 1/µmin .
The second packing set we construct is for the slice A1,j2 ,j3 for any 1 ≤ j2 , j3 ≤ m. Since
in the second and third direction only s of the co-ordinates are non-zero, we consider the
vector space
2
{v ∈ Rm | kvk`0 ≤ s}.
Once again using the standard standard hypercube construction in Lemma 19 in Appendix C,
there exists a discrete set {v 1 , v 2 , ..., v N } such that log N ≥ cs log(m2 /s) for some c > 0 and
δ2
≤ kv `1 − v `2 k2`2 ≤ δ 2
8
for `1 6= `2 for any δ > 0. Setting δ = sn−1 log(m2 /s) yields
√
2
2 −2 s log(m/ s)
e
min max kT − T kF ≥ cσ cu
,
n
Te T ∈T3
with probability greater than 1/2. Taking a maximum over lower bounds involving both
packing sets completes the proof of of our lower bound.
Proof of Theorem 5. The upper bound follows from a slight modification of the statement
in Lemma 5. In particular since R(A) = kA(12) k∗ + kA(13) k∗ + kA(23) k∗ , the dual norm is
R∗ (A) =
max
1≤k1 <k2 ≤3
kA(k1 k2 ) ks .
(38)
Hence, following the same technique as used in Lemma 5
r
r
max{d
,
d
max{d1 , d2 , d3 }
}
k1
k2
E[R∗ (G)] ≤ c max
=c
.
1≤k1 <k2 ≤3
n
n
(39)
It is also straightforward to see that s(T4 ) ≤ r.
To prove the lower bound, we construct three packing sets and select the one with the
largest packing number. Recall that
(12)
(13)
(23)
T4 = {A ∈ Rd1 ×d2 ×d3 : Aj1 j2 j3 = Aj1 j2 + Aj1 j3 + Aj2 j3 , A(k1 ,k2 ) ∈ Rdk1 ×dk2 ,
A(k1 ,k2 ) 1 = 0,
and
(A(k1 ,k2 ) )> 1 = 0
max rank(A(k1 ,k2 ) ) ≤ r}.
k1 ,k2
53
Therefore our three packings are for A(12) ∈ Rd1 ×d2 , A(13) ∈ Rd1 ×d3 , and A(23) ∈ Rd2 ×d3
assuming each has rank r. We focus on packing in A(12) ∈ Rd1 ×d2 since the approach is
similar in the other two cases. Using Lemma 17 from Appendix B in combination with
Theorem 6,
r min{d1 , d2 }
,
min max kTe − T k2F ≥ cσ 2 c−2
u
n
Te T ∈T4
with probability greater than 1/2. Repeating this process for packings in A(13) ∈ Rd1 ×d3 ,
and A(23) ∈ Rd2 ×d3 assuming each has rank r and taking a maximum over all three bounds
yields the overall minimax lower bound
r max{d1 , d2 , d3 }
,
min max kTe − T k2F ≥ cc−2
u
n
Te T ∈T4
with probability greater than 1/2.
References
A. Agarwal, S. Negahban, and M. J. Wainwright. Noisy matrix decomposition via convex
relaxation: Optimal rates in high dimensions. The Annals of Statistics, 27(2):1171–1197,
2012.
T. W. Anderson. The integral of a symmetric convex set and some probability inequalities.
Proc. of American Mathematical Society, 6:170–176, 1955.
T. W. Anderson. An Introduction to Multivariate Statistical Analysis. Wiley Series in
Probability and Mathematical Statistics. Wiley, New York, 1984.
P. Bartlett and S. Mendelson. Gaussian and Rademacher complexities: Risk bounds and
structural results. Journal of Machine Learning Research, 3:463–482, 2002.
S. Basu and G. Michailidis. Regularized estimation in sparse high-dimensional time series
models. Annals of Statistics, 43(4):1535–1567, 2015.
R. Bhatia. Matrix Analysis. Springer, New York, 1997.
P. Buhlmann and S. van de Geer. Statistical for High-Dimensional Data. Springer Series in
Statistics. Springer, New York, 2011.
54
S. Chen, M. R. Lyu, I. King, and Z. Xu. Exact and stable recovery of pairwise interaction
tensors. In Advances in Neural Information Processing Systems, 2013.
S. Cohen and M. Collins. Tensor decomposition for fast parsing with latent-variable pcfgs.
In Advances in Neural Information Processing Systems, 2012.
T.M. Cover and J.A. Thomas. Elements of Information Theory. John Wiley and Sons, New
York, 1991.
S. Dasgupta and A. Gupta. Empirical processes with a bounded ψ1 diameter. Geometric
and Functional Analysis, 20(4), 2010.
S. Gandy, B. Recht, and I. Yamada. Tensor completion and low-n rank tensor recovery via
convex optimization. Inverse Problems, 27, 2011.
Y. Gordon. On milmans inequality and random subspaces which escape through a mesh
in Rn . Geometric aspects of functional analysis, Israel Seminar 1986-87, Lecture Notes,
1317:84–106, 1988.
T. Hastie, R. Tibshirani, and M. Wainwright. Statistical Learning with Sparsity: The Lasso
and Generalizations. Monographs on Statistics and Applied Probability 143. CRC Press,
New York, 2015.
P. Hoff. Multilinear tensor regression for longitudinal relational data. Technical Report
TR-2003-08, Department of Statistics, University of Washington, 2003.
I. A. Ibragimov and R. Z. Has’minskii. Statistical Estimation: Asymptotic Theory. SpringerVerlag, New York, 1981.
T. G. Koldar and B. W. Bader. Tensor decompositions and applications. SIAM Review, 51:
455–500, 2009.
B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model selection.
The Annals of Statistics, 28(5):1303–1338, 1998.
M. Ledoux. The Concentration of Measure Phenomenon. Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2001.
55
M. Ledoux and M. Talagrand. Probability in Banach Spaces: Isoperimetry and Processes.
Springer-Verlag, New York, NY, 1991.
N. Li and B. Li. Tensor completion for on-board compression of hyperspectral images. In
17th IEEE International Conference on Image Processing (ICIP), pages 517–520, 2010.
H. Lütkepolhl. New introduction to multiple time series analysis. Springer, New York, 2006.
P. Massart. Concentration Inequalties and Model Selection. Ecole d’Eté de Probabilités,
Saint-Flour. Springer, New York, 2003.
Shahar Mendelson. Upper bounds on product and multiplier empirical processes. Technical
Report arXiv:1410.8003, Technion, I.I.T, 2014.
N. Mesgarani, M. Slaney, and S. Shamma. Content-based audio classification based on
multiscale spectro-temporal features. IEEE Transactions on Speech and Audio Processing,
14:920–930, 2006.
C. Mu, B. Huang, J. Wright, and D. Goldfarb. Square deal: Lower bounds and improved
relaxations for tensor recovery. In International Conference on Machine Learning, 2014.
S. Negahban and M. J. Wainwright. Estimation of (near) low-rank matrices with noise and
high-dimensional scaling. The Annals of Statistics, 39:1069–1097, 2011.
S. Negahban and M. J. Wainwright. Restricted strong convexity and weighted matrix completion. JMLR, 13:1665–1697, 2012.
S. Negahban, P. Ravikumar, M. J. Wainwright, and B. Yu. A unified framework for highdimensional analysis of m-estimators with decomposable regularizers. Statistical Science,
27(4):538–557, 2012.
G. Pisier. The Volume of Convex Bodies and Banach Space Geometry, volume 94 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, UK, 1989.
D. Pollard. Convergence of Stochastic Processes. Springer-Verlag, New York, 1984.
56
Z. Qin, K. Scheinberg, and D Goldfarb. Efficient block-coordinate descent algorithms for
the group lasso. Math. Program. Comput., 5:143–169, 2013.
G. Raskutti, M. J. Wainwright, and B. Yu. Restricted eigenvalue conditions for correlated
Gaussian designs. Journal of Machine Learning Research, 11:2241–2259, 2010.
G. Raskutti, M. J. Wainwright, and B. Yu. Minimax rates of estimation for high-dimensional
linear regression over `q -balls. IEEE Transactions of Information Theory, 57(10):6976–
6994, 2011.
G. Raskutti, M. J. Wainwright, and B. Yu. Minimax-optimal rates for sparse additive models
over kernel classes via convex programming. Journal of Machine Learning Research, 13:
398–427, 2012.
S. Rendle and L. Schmidt-Thieme. Pairwise interaction tensor factorization for personalized
tag recommendation. In ICDM, 2010.
S. Rendle, L. B. Marinho, A. Nanopoulos, and L. Schmidt-Thieme. Learning optimal ranking
with tensor factorization for tag recommendation. In SIGKDD, 2009.
G. Rockafellar. Convex Analysis. Princeton University Press, Princeton, 1970.
O. Semerci, N. Hao, M. Kilmer, and E. Miller. Tensor based formulation and nuclear norm
regularizatin for multienergy computed tomography. IEEE Transactions on Image Processing, 23:1678–1693, 2014.
N.D. Sidiropoulos and N. Nion. Tensor algebra and multi-dimensional harmonic retrieval in
signal processing for mimo radar. IEEE Transactions on Signal Processing, 58:5693–5705,
2010.
N. Simon, J. Friedman, and T. Hastie. A blockwise coordinate descent algorithm for penalized multiresponse and grouped multinomial regression. Technical report, Georgia Tech.,
November 2013.
D. Slepian. The one-sided barrier problem for gaussian noise. Bell System Tech. J, 41:
463–501, 1962.
57
M. Talagrand. New concentration inequalities in product spaces. Invent. Math., 126:503523,
1996.
R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society, Series B, 58(1):267–288, 1996.
B. Turlach, W.N. Venables, and S.J. Wright. Simultaneous variable selection. Technometrics,
27:349–363, 2005.
S. van de Geer. Empirical Processes in M-Estimation. Cambridge University Press, 2000.
S. van de Geer. Weakly decomposable regularization penalties and structured sparsity.
Scandivanian Journal of Statistics: Theory and Applications, 41:72–86, 2014.
Y. Yang and A. Barron. Information-theoretic determination of minimax rates of convergence. The Annals of Statistics, 27(5):1564–1599, 1999.
B. Yu. Assouad, Fano and Le Cam. Research Papers in Probability and Statistics: Festschrift
in Honor of Lucien Le Cam, pages 423–435, 1996.
M. Yuan and Y. Lin. Model selection and estimation in regression with grouped variables.
Journal of the Royal Statistical Society B, 68(1):49–67, 2006.
M. Yuan and C-H. Zhang. On tensor completion via nuclear norm minimization. Foundation
of Computational Mathematics, to appear, 2014.
Shuheng Zhou. Restricted eigenvalue conditions on subgaussian random matrices. Technical
Report arXiv:0912.4045, ETH Zurich, 2009.
A
Results for Gaussian random variables
In this section we provide some standard concentration bounds that we use throughout this
paper. First, we provide standard χ2 tail bounds. due to Laurent and Massart Laurent and
Massart (1998):
58
Lemma 13. Let Z be a centralized χ2 random variable with m degrees of freedom. Then for
all x ≥ 0,
√
P[Z − m ≥ 2 mx + 2x] ≤ exp(−x), and
√
P[Z − m ≤ −2 mx]
≤ exp(−x).
A.1
Gaussian comparison inequalities
The first result is a classical result from Anderson (1955).
Lemma 14 (Anderson’s comparison inequality). Let X and Y be zero-mean Gaussian random vectors with covariance ΣX and ΣY respectively. If ΣX − ΣY is positive semi-definite
then for any convex symmetric set C,
P(X ∈ C) ≤ P(Y ∈ C).
The following Lemma is Slepian’s inequality Slepian (1962) which allows to upper bound
the supremum of one Gaussian process by the supremum of another Gaussian process.
Lemma 15 (Slepian’s Lemma). Let {Gs , s ∈ S} and {Hs , s ∈ S} be two centered Gaussian
processes defined over the same index set S. Suppose that both processes are almost surely
bounded. For each s, t ∈ S, if E(Gs − Gt )2 ≤ E(Hs − Ht )2 , then E[sups∈S Gs ] ≤ E[sups∈S Hs ].
Further if E(G2s ) = E(Hs2 ) for all s ∈ S, then
P sup Gs > x ≤ P sup Hs > x ,
s∈S
s∈S
for all x > 0.
Finally, we require a standard result on the concentration of Lipschitz functions over
Gaussian random variables.
Theorem 7 (Theorem 3.8 from Massart (2003)). Let g ∼ N (0, Id×d ) be a d-dimensional
Gaussian random variable. Then for any function F : Rd → R such that |F (x) − F (y)| ≤
Lkx − yk`2 for all x, y ∈ Rd , we have
t2
P |F (g) − E[F (g)]| ≥ t ≤ 2 exp − 2 ,
2L
for all t > 0.
59
B
Suprema for i.i.d. Gaussian tensors
In this section we provide important results on suprema of i.i.d. Gaussian tensors over
different sets.
B.1
The group `2 -`∞ norm
Let G ∈ Rd1 ×d2 be an i.i.d. Gaussian matrix and define the set
V := {(u, v) ∈ Rd1 × Rd2 | kuk`2 ≤ 1 , kvk`1 ≤ 1}.
Using this notation, let us define the define the random quantity:
M (G, V ) := sup u> Gv.
(u,v)∈V
Then we have the following overall bound.
Lemma 16.
p
p
E[M (G, V )] ≤ 3( d1 + log d2 ).
Proof. Our proof user similar ideas to the proof of Theorem 1 in Raskutti et al. (2010). We
need to upper bound E[M (G, V )]. We are taking the supremum of the Gaussian process
u> Gv.
sup
kuk`2 ≤1, kvk`2 ≤1
eu,v over the set V and apply Slepian’s inWe now construct a second Gaussian process G
equality (see Lemma 15 in Appendix A.1) to upper bound
u> Gv
sup
kuk`2 ≤1, kvk`2 ≤1
eu,v . In particular, let us define the
by the supremum over our second Gaussian process. G
process as:
eu,v = g > u + h> v,
G
where the vectors (g, h) ∈ Rd1 × Rd2 are i.i.d. standard normals (also independent of each
other). It is straightforward to show that both u> Gv and g > u + h> v are zero-mean. Further
it is straightforward to show that
eu,v − G
eu0 ,v0 ) = ku − u0 k2 + kv − v 0 k2 .
Var(G
`2
`2
60
Now we show that
Var(u> Gv − u0> Gv 0 ) ≤ ku − u0 k2`2 + kv − v 0 k2`2 .
To this end, observe that
Var(u> Gv − u0> Gv 0 ) = kuv > − u0 v 0> k2F
= k(u − u0 )v > + u0 (v − v 0 )> k2F
= kvk2`2 ku − u0 k`22 + ku0 k2`2 kv − v 0 k2`2
+2(u> u0 − ku0 k`2 kuk`2 )(v > v 0 − kv 0 k`2 kvk`2 ).
First note that kvk2`2 ≤ kvk2`1 ≤ 1 for all v ∈ V and ku0 k2`2 ≤ 1. By the Cauchy-Schwarz
inequality, v > v 0 − kv 0 k`2 kvk`2 ≤ 0 and u> u0 − ku0 k`2 kuk`2 ≤ 0. Therefore
Var(u> Gv − u0> Gv 0 ) ≤ ku − u0 k`22 + kv − v 0 k2`2 .
Consequently using Lemma 15
E[M (G, V )] ≤ E[ sup g > u + sup h> v].
kuk`2 ≤1
kvk`1 ≤1
Therefore:
E[M (G, V )] ≤ E[ sup g > u + sup h> v]
kuk`2 ≤1
kvk`1 ≤1
= E[ sup g > u] + E[ sup h> v]
kuk`2 ≤1
kvk`1 ≤1
= E[kgk`2 ] + E[khk`∞ ].
By known results on Gaussian maxima (see e.g. Ledoux and Talagrand, 1991),
p
E[khk`∞ ] ≤ 3 log d2
and
E[kgk`2 ] ≤
p
p
3p
d1 + o( d1 ) ≤
Dj .
2
Therefore
E[M (G, V )] ≤
p
3p
d1 + 3 log d2 .
2
61
B.2
Spectral norm of tensors
Our proof is based on an extension of the proof techniques used for the proof of Proposition
1 in Negahban and Wainwright (2011).
Lemma 17. Let G ∈ Rd1 ×d2 ×···×dN be a random sample from an i.i.d. Gaussian tensor
ensemble. Then we have
E[kGks ] ≤ 4 log(4N )
N p
X
dk .
k=1
Proof. Recall the definition of kGks :
kGks =
hu1 ⊗ u2 ⊗ · · · ⊗ uN , Gi.
sup
(u1 ,u2 ,...,uN )∈S d1 −1 ×S d2 −1 ×···×S dN −1
Since each entry hu1 ⊗ u2 ⊗ · · · ⊗ uN , Gi is a zero-mean Gaussian random variable, kGks
is the supremum of a Gaussian process and therefore the concentration bound follows from
Theorem 7.1 in Ledoux (2001).
1
We use a standard covering argument to upper bound E[kGks ]. Let {u11 , u21 , . . . , uM
1 } be
a 1/2N covering number of the sphere S d1 −1 in terms of vector `2 -norm. Similarly for all
dk −1
k
. Therefore
2 ≤ k ≤ N , let {u1k , u2k , . . . , uM
k } be a 1/2N covering number of the sphere S
hu1 ⊗ u2 ⊗ · · · ⊗ uN −1 ⊗ uN , Gi
≤ hu1 ⊗ u2 ⊗ · · · ⊗ uN −1 ⊗ ujN , Gi + hu1 ⊗ u2 ⊗ · · · ⊗ uN −1 ⊗ (uN − ujN ), Gi.
Taking a supremum over both sides,
kGks ≤
max hu1 ⊗ u2 ⊗ · · · ⊗ uN −1 ⊗ ujN , Gi +
j=1,...,MN
1
kGks .
2N
Repeating this argument over all N directions,
kGks ≤ 2
max
j1 =1,2,...,M1 ,...,jN =1,2,...,MN
huj11 ⊗ uj22 ⊗ · · · ⊗ ujNN , Gi.
By construction, each variable huj11 ⊗uj22 ⊗· · ·⊗ujNN , Gi is a zero-mean Gaussian with variance
at most 1, so by standard bounds on Gaussian maxima,
p
p
p
E[kGks ] ≤ 4 log(M1 × M2 × · · · × MN ) ≤ 4[ log M1 + · · · + log MN ].
There exist a 1/2N -coverings of S dk −1 with log Mk ≤ dk log(4N ) which completes the proof.
62
C
Hypercube packing sets
In this section, we provide important results for the lower bound results. One key concept is
the so-called Hamming distance.The Hamming distance is between two vectors v ∈ Rd and
v 0 ∈ Rd is defined by:
0
dH (v, v ) =
d
X
I(vj 6= vj0 ).
j=1
Lemma 18. Let C = [−1, +1]d where d ≥ 6. Then there exists a discrete subset {v 1 , v 2 , ..., v m } ⊂
C, such that log m ≥ cd for some constant c > 0, and for all `1 6= `2 ,
δ2
≤ kv `1 − v `2 k2`2 ≤ δ 2 ,
4
for any δ > 0.
Proof. Let
d
δ
δ
v ∈ −√ , √
,
d d
`
i.e. a member of the d-dimensional hypercube re-scaled by
√
√
3δ/(2 d). Recall the definition
of Hamming distance provided above. In this case amounts to the places either vj or vj0 is
negative, but both or not negative. Then according to Lemma 4 in Yu (1996), there exists
a subset re-scaled of this hypercube v 1 , v 2 , ..., v m , such that
dH (v `1 , v `2 ) ≥
d
3
and log m ≥ cd. Clearly,
kv `1 − v `2 k2`2 =
3δ 2
δ2
dH (v `1 , v `2 ) ≥ .
4d
4
kv `1 − v `2 k2`2 ≤
3δ 2
3δ 2
×d≤
≤ δ2.
4d
4
Further,
This completes the proof.
Next we provide a hupercube packing set for the sparse subset of vectors. That is the set
V := {v ∈ Rd | kvk`0 ≤ s}.
This follows from Lemma 4 in Raskutti et al. (2011) which we state here for completeness.
63
Lemma 19. Let C = [−1, +1]d where d ≥ 6. Then there exists a discrete subset {v 1 , v 2 , ..., v m } ⊂
V ∩ C, such that log m ≥ cs log(d/s) for some c > 0, and for all `1 6= `2 ,
δ2
≤ kv `1 − v `2 k2`2 ≤ δ 2 ,
8
for any δ > 0.
Finally we present a packing set result from Lemma 6 in Agarwal et al. (2012) that packs
into the set of rank-r d1 × d2 matrices.
Lemma 20. Let min{d1 , d2 } ≥ 10, and let δ > 0. Then for each 1 ≤ r ≤ min{d1 , d2 },
there exists a set of d1 × d2 matrices {A1 , A2 , ..., Am } with rank-r with cardinality log m ≥
cr min{d1 , d2 } for some constant c > 0 such that
δ2
≤ kA`1 − A`2 k2F ≤ δ 2 ,
4
for all `1 6= `2 .
64
| 10math.ST
|
1
Stochastic Power System Simulation Using the
Adomian Decomposition Method
Nan Duan, Student Member, IEEE, and Kai Sun, Senior Member, IEEE
Abstract-- For dynamic security assessment considering
uncertainties in grid operations, this paper proposes an approach
for time-domain simulation of a power system having stochastic
loads. The proposed approach solves a stochastic differential
equation model of the power system in a semi-analytical way using
the Adomian decomposition method. The approach generates
semi-analytical solutions expressing both deterministic and
stochastic variables explicitly as symbolic variables so as to embed
stochastic processes directly into the solutions for efficient
simulation and analysis. The proposed approach is tested on the
New England 10-machine 39-bus system with different levels of
stochastic loads. The approach is also benchmarked with a
traditional stochastic simulation approach based on the EulerMaruyama method. The results show that the new approach has
better time performance and a comparable accuracy.
Index Terms—Adomian decomposition method, stochastic
differential equation, stochastic load, stochastic time-domain
simulation.
U
I. INTRODUCTION
NCERTAINTIES exist in operations of power grids [1]
Many factors such as random load consumptions and
unanticipated relay protection actions contribute to the
randomness of grid operations. It can be foreseen that a future
power grid will have more uncertainties and stochastic
behaviors in system operations due to the increasing
penetrations of responsive loads and intermittent renewable
generations. Thus, dynamic security assessment (DSA) of
power systems should be conducted in both deterministic and
stochastic manners. However, most of today’s power system
simulation software tools are still based on solvers of
deterministic differential-algebraic equations (DAEs) that do
not involve stochastic variables to model uncertainties in
system operating conditions.
In literature, there are three major approaches for the
modeling of a dynamic system having stochastic effects as
shown in Fig. 1: the master equation, the Fokker-Planck
equation [2][3] and Gillespie method [1][5]. The master
equation and the Fokker-Plank equation are widely applied in
the field of computational biology, which both focus on the
evolution of probability distribution; the Gillespie method
focuses on individual stochastic trajectories. The first two
approaches provide a more comprehensive understanding of
stochastic effects with a dynamic system but require solving
This work was supported by NSF grant ECCS-1610025.
Nan Duan and Kai Sun are with the department of EECS at the University
of Tennessee, Knoxville (e-mail: [email protected], [email protected]).
high dimensional partial differential equations, so they are
computationally difficult to be applied to simulations of
realistic power systems [6]. There have been works using the
Gillespie method for power system simulation [7]-[10].
Stochastic
Modeling
Master
Equation
Multiple runs
Gillespie
Algorithm
FokkerPlanck
Equation
Multiple runs
EulerMaruyama
method
Adomian
Decomposition
Method
Fig. 1. Stochastic modeling approaches.
In recent years, some researchers have contributed to power
system simulation in a less-deterministic manner. Reference [7]
proposed a systematic method to simulate the system behaviors
under the influence of stochastic perturbations on loads, bus
voltages and rotor speeds. This approach introduces stochastic
differential equations (SDEs) to represent stochastic
perturbations and solves the equations by Ito calculus, and then
a mean trajectory with the envelope on trajectory variations is
yielded by repeating simulations for many times. Papers [8][10] utilize a similar approach to study power system stability
under random effects. To analyze long term stability of a power
system with wind generation, a new SDE model is developed
in [11], which also applies the singular perturbation theory to
investigate the slow dynamics of the system with stochastic
wind generation. However, the time performance of such an
approach based on Euler-Maruyama method can hardly meet
the requirements for online power system simulation.
Especially, when the penetration of distributed energy
resources (DERs) reaches a high level, the distribution network
behaves in a more stochastic manner as seen from the
transmission network, and hence a large number of SDEs need
to be included in the power system model, which will
significantly influence the simulation speed. Also, the nature of
the Gillespie method requires a large number of simulations on
the same model to yield the mean trajectory as well as the
2
envelope on variations. Therefore, adding any extra SDE to the
existing set of SDEs will result in multiplying computing time
by a factor of hundreds or even thousands.
In our previous works [12]-[14], a new semi-analytical
approach for power system simulation has been proposed. That
approach applies the Adomain decomposition method (ADM)
to power system DAEs to derive a semi-analytical solution
(SAS) for each state variable as an explicit function of symbolic
variables including time, the initial system state and other
selected parameters on the system condition; then each function
is evaluated by plugging in values of its symbolic variables over
consecutive small time windows to make up a desired
simulation period so as to obtain the simulated trajectory of
each state variable. Since the form of every SAS is a summation
of finite terms for approximation, its evaluation can be fast and
parallelized among terms. Thus, compared to traditional
numerical integration based power system simulation, this
semi-analytical approach decomposes the computation into
offline derivation and online evaluation of an SAS and is better
fit for online power system simulation and a parallel computing
environment [14]. In fact, such a semi-analytical approach also
suggests a viable, alternative paradigm for fast stochastic
simulation. For example, early works by Adomian in the 1970s
utilized the ADM to solve nonlinear SDEs [15] by embedding
explicitly stochastic processes into the terms of an SAS.
For power system simulation in a stochastic manner, this
paper proposes an approach as an extension of the ADM based
approach proposed in [14]. Utilizing the semi-analytical nature
of an SAS yielded by the ADM, this new approach embeds a
stochastic model, e.g. a stochastic load model, into the SAS.
Evaluation of an SAS with the stochastic model whose
parameters are represented symbolically will not increase many
computational burdens compared to evaluation of an SAS for
deterministic simulation. Thus, an expected number of
simulation runs for one single case are achieved by evaluating
one SAS for the same number of times.
The rest of this paper is organized as follows: Section II
presents the SDE model of a power system that integrates
stochastic loads; Section III gives the ADM-based approach for
solving the power system SDEs for stochastic simulation;
Section IV uses a single-machine-infinite-bus (SMIB) system
to compare the fundamental difference between the ADMbased approach and the Euler-Maruyama approach in
mathematics; Section V introduces a criterion for defining the
stability of a general stochastic dynamical system, which is also
applied to power systems; Section VI validates the proposed
approach using the IEEE 10-machine 39-bus system with the
stochastic loads and compares the results and time performance
with those from the Euler-Maruyama approach; finally,
conclusions are drawn in Section VII.
II. POWER SYSTEM SDE MODEL WITH STOCHASTIC LOADS
A. Synchronous Generator Modeling
For a power system having K synchronous generators,
consider the 4th-order two-axis model (1) to model each
generator having saliency ignored [16]. All generators are
coupled through nonlinear algebraic equations (2) about the
network.
k k R
k R
R
k 2 H Pm k Pek D k
R
k
1
x dk x dk
i dk
E fdk e qk
e qk
T d0 k
1
iqk
e x qk x qk
e dk
dk
T
q0k
(1)
sin k e qk
cos k j ( e qk
sin k e dk
cos k )
E k e dk
def
I tk i R k ji Ik Y k* E
Pek e qk iqk e dk idk
(2)
iqk i Ik sin k i R k cos k , idk i R k sin k i Ik cos k
x dk
idk , e dk e dk
x qk
iqk
e qk e qk
In (1) and (2), ωR is the rated angular frequency; δk , ωk, Hk
and Dk are respectively the rotor angle, rotor speed, inertia and
damping coefficient of the machine k; Yk is the kth row of the
reduced admittance matrix Y; E is the column vector of all
generator’s electromotive forces (EMFs) and Ek is the kth
element; Pmk and Pek are the mechanical and electric powers;
Efdk is the internal field voltage;
,
, iqk, idk,
,
, xqk,
xdk, x’qk and x’dk are transient voltages, stator currents, opencircuit time constants, synchronous reactances and transient
reactances in q- and d-axes, respectively.
B. Stochastic Load Modeling
A stochastic model can be built based on analysis on real
data and assumptions on probabilistic characteristics of the
stochastic variables. Traditionally, uncertainties in loads of a
power system are ignored in time-domain simulation for the
sake of simplicity. However their stochastic behaviors are wellrecognized in [17]. Taking stochastic loads into consideration
will enable more realistic power system stability assessment.
This paper uses the Ornstein-Uhlenbeck process in [18] to
model the stochastic variations of a load in these SDEs:
y PL aP y PL bP W(t )
(3)
y QL aQ yQL bQ W(t )
(4)
where W(t) is the white noise vector whose dimension equals
the number of load buses, a and b parameters are drifting and
diffusion parameters of the SDEs, operator “ ” is the
Hadamard Product, i.e., element-wise multiplication, and yPL
and yQL are the stochastic variations in normal distributions.
The stochastic dynamic of the load is therefore modeled by
PL PL0 y PL
(5)
QL QL0 yQL
(6)
where PL0 and QL0 are the mean values of the active and
reactive loads, respectively.
Periodicities and autocorrelations have been observed in
historical data of loads on the daily basis. However, in the time
frame of seconds, loads at different substations have much
lower autocorrelations. Refer to [7], this paper sets the drifting
3
parameter on the autocorrelations of loads as 0.5 p.u./s.
T
III. PROPOSED ADM-BASED APPROACH TO SOLVING POWER
SYSTEM SDES
A. Modeling Stochastic Variables
Consider S stochastic variables y1 (t ), , yS (t ) , which could
be stochastic loads following S different distributions. Each yi(t)
can be transformed by function gi () in (7) from some i in a
normal distribution. For example, if yi(t) is a load represented
by a normal distribution with certain mean value, then i
specifies a zero-mean normal distribution as in (9) and gi ()
shifts it to around the desired mean value like in (5) and (6).
T
(7)
y (t ) g ( ) g ( ) g ( )
i
1
1
2
2
S
S
The Ornstein-Uhlenbeck process is utilized to generate each
where A n A1,n , A2, n , , AM , n
The next step is to apply inverse Laplace transform to both
sides of (15) and (16) to calculate the N-th order SAS of (10):
N
(17)
x SAS (t , y ) x n (t , y )
n 0
In the resulting SAS, stochastic variables in y appear
explicitly as symbolic variables.
IV. COMPARISON BETWEEN THE EULER-MARUYAMA
APPROACH AND THE PROPOSED APPROACH
This section applies both the Euler-Maruyama approach and
the proposed ADM-based approach to the SMIB system with a
stochastic load shown in Fig. 2 to illustrate the fundamental
difference between the two approaches.
E’∠δ
from (10).
ε (t ) a ε (t ) b W (t )
where
Ra
jXd’
R
jX
V∠0°
(8)
ε(t ) 1 (t ) 2 (t ) S (t )
T
RL
a(t ) a1 (t ) a2 (t ) aS (t )
T
jXL
b(t ) b1 (t ) b2 (t ) bS (t )
T
i (0, bi2 / 2 ai ) i 1, 2, , S
(9)
B. Solving SDEs Using the ADM
Consider a nonlinear system modeled by SDE (10) having
M deterministic state variables x1, …, xM, such as the state
variables of generators, exciters and speed governors, and S
stochastic variables y1, …, yS.
(10)
x (t ) f ( x (t ), y (t ))
x(t ) x1 (t ) x2 (t ) xM (t )
T
f () f1 ()
f 2 ()
f M ()
T
To solve x (t ) , the procedure in [14] can be used. First,
apply Laplace transformation to (10) to obtain
(11)
x (0) [f ( x, y )]
[x]
s
s
Then use (13) and (14) to calculate the Adomian polynomials
under the assumption of (12),
(12)
x(t ) xn (t )
n 0
f k (x, y ) Ak ,n (x0, x1, x n , y )
(13)
1 n
n
k
n fk xk , y
n !
k 0
0
(14)
Recursive formulas (15) and (16) can be derived by matching
terms of x (t ) and f () :
(15)
x x(0) s
0
xn1 An s n 0
The stochastic load is connected to the generator bus and has
its resistance RL and reactance XL modeled by stochastic
variables. Thus, the whole system is now modeled by DEs
(18a), (18b) and SDEs (18c), (18d).
R
R R
EV
Pm k3
k4 cos( ) k5 sin( )
D
2H
k1k2
1 (t )
RL a1RL bW
X a X b W(t)
2 L
2
L
(16)
(18a)
(18b)
(18c)
(18d)
where,
GL jBL
GS jBS
(19a)
1
RL jX L
1
Ra jX d
(19b)
1
R jX
(19c)
GR jBR
k1
n 0
Ak , n
Fig. 2. SMIB system with constant impedance load at generator bus.
(GL GR GS ) 2
( BL BR BS )
( BL BR BS )
(19d)
( BL BR BS ) 2
(GL GR GS )
(19e)
k2 (GL GR GS )
G ( B BR ) BS (GL GR ) BS ( BL BR ) GS (GL GR )
k3 E 2 S L
k1
k2
(19f)
k4 k2 ( BS GR GS BR ) k1 ( BS BR GS GR )
(19g)
k5 k2 ( BS BR GS GR ) k1 ( BS GR GS BR )
(19h)
Since RL and XL change stochastically, GL and BL cannot be
4
treated as constants in (19). The variances of RL and XL depend
on the values of drifting parameters a1 and a2 and diffusion
parameters b1 and b2, respectively.
To find the SAS of this system, the first step is to apply ADM
to DEs (18a) and (18b). Once the SAS of the system’s DEs is
derived, the SAS of the SDEs can be derived and incorporated
into it.
For instance, the 2nd order SAS for rotor speed ω is
2
(t ) n (t )
(20)
n 0
where,
0 (t ) (0)
1 (t )
(21)
tR D( (0) R )
Pm k3
2H
R
Therefore the infinite order SAS of (20c) is
s
t s
i1
1
(a1t )i
b1B(t ) b1 (a1 )i B(si )dsi ds2ds1 (33)
i!
i 0
i 1
0 0
0
RL (t ) RL (0)
Apply Maclaurin expansion of an exponential function and
lemma 2.3 in [19] to (33), the solution becomes
t
(22)
k E V
k4 E V
cos( (0)) 5
sin( (0))
k1k2
k1k2
E V
cos( (0))
k1k2
The 2nd order SAS of RL can be derived using ADM as
2
RL (t ) RL ,n (t )
(24)
n 0
where
(25)
RL ,0 (t ) RL (0) b1 B(t )
t
RL ,1 (t ) a1 RL (0)t a1b1 B ( s1 )ds1
RL (t ) RL (0)e a1t b1 B (t ) a1b1 ea1s a1t B( s )ds
(34)
0
Then apply the integration by parts formula
t 2 R D 2 ( (0) R )
2 (t )
8H 2
R
k E V
k E V
D ( Pm k3 4
cos( (0)) 5
sin( (0))) (23)
k1k2
k1k2
k E V
k E V
2 H R 5
cos( (0)) 2 H R 4
sin( (0))
k1k2
k1k2
2 H ( (0) R )
(24) and (28).
For some forms of SDEs, an analytical solution may exist,
which can be incorporated into the DEs’ SAS to directly derive
the SAS of the entire system. For example, the general
expression of the SAS terms of (18c) can be written as
sn1
t s1
tn
RL,n (t ) (1)n a1n RL (0) (1)n a1nb1 B( sn )dsn ds2 ds1 (32)
n!
0 0
0
(26)
t
t
0
0
as
at
as
e 1 dB(s) e 1 B(t ) a1e 1 B(s)ds
(35)
The close form solution can be found as
t
RL (t ) e a1t [ RL (0) b1 ea1s dB( s)]
(36)
0
In this case the symbolic variable RL in (20) can be replaced
by (36) instead of (24).
On the other hand, for the Euler-Maruyama approach
[20][21], since the deterministic model described by (18a) and
(18b) does not permit a close form solution, the sample
trajectories of (18) have to be numerically computed. The
numerical scheme for RL is shown in (37) and the same scheme
also applies to XL.
(37)
R L( , nt ) 1 R L( , nt ) a1 R L( , nt ) t b1 R L( , nt ) W
In practice the value of ΔW is dependent of the step size Δt
for integration.
(38)
W (0 , t 1 / 2 )
0
t s
RL ,2 (t ) a12 RL (0)
1
t2
a12b1 B( s2 )ds2 ds1
2!
0 0
(27)
Here, B(t) is the Brownian motion starting at origin and
dB(t)=W(t)dt.
Similarly, the 2nd order SAS of XL is,
2
X L (t ) X L , n (t )
(28)
n 0
where
X L ,0 (t ) X L (0) b2 B(t )
t
X L ,1 (t ) a2 RL (0)t a2b2 B( s1 )ds1
(29)
(30)
0
X L ,2 (t ) a22 RL (0)
2
t s1
t
a22b2 B( s2 )ds2 ds1
2!
0 0
(31)
To derive the SAS of the entire system considering both the
DEs and SDEs, replace the symbolic variables in the DEs’ SAS
representing the stochastic variables with the SDEs’ SAS, i.e.,
the 2nd order SAS of the system (18) can be derived by
replacing the symbolic variables RL and XL in (20) with SAS
V. STABILITY OF STOCHASTIC SYSTEMS
There are a variety of definitions on the stability of a
stochastic dynamical system in literature [23]-[26]. The
definition of “asymptotic stability in probability” in [26] can be
directly applied to a power system with stochastic variables.
That definition is a counterpart of the asymptotic Lyapunov
stability of a deterministic system.
Definition 1: Stability in probability
An equilibrium point is said to be stable in probability if for
given (0,1) and r, there exists σ(μ,r,t0) >0 such that,
P sup x(t ; x0 , t0 ) xeq r 1
t t0
(39)
whenever ||x0-xeq||<σ.
Definition 2: Asymptotic stability in probability
An equilibrium point is said to be asymptotic stable in
probability if it is stable in probability and for given (0,1)
, there exists σ0(μ)>0 such that,
5
P lim x (t ; x0 , t0 ) x eq 0 1
t
(40)
whenever ||x0-xeq||<σ0
To analyzed the stability of numerical simulation results, this
paper modifies (40) to (41) so that the stability can be accessed
using the results of finite time period simulations.
P x(t ; x0 , t0 ) xeq r0 1 t ts
(41)
where ts is a predefined time instant, ||r0|| is a small positive
number.
VI. CASE STUDIES
The proposed ADM-based approach is tested on the IEEE
10-machine 39-bus New England system as shown in Fig. 3.
Selected loads are assumed to change stochastically while all
generators are represented by deterministic models. In each
case study, the stochastic simulation result by the EulerMaruyama approach is used as the benchmark, and the 2nd order
SASs (i.e. N=2) are used and evaluated every 0.001 s. The value
of each stochastic variable is changed every 0.1 s. For each case,
100 sample trajectories are generated. The fault applied in all
cases is a 10-cycle 3-phase fault at bus 3 cleared by tripping line
3-4. All simulations are performed in MATLAB R2016a on a
desktop computer with an Intel Core i7-3770 3.40GHz CPU
and 8 GB RAM.
(a) Result from the ADM-based approach.
(b) Result from the Euler-Maruyama approach
Fig. 4. Simulation results of generator 1 rotor angle with loads connecting to
bus 3 and 4 represented by stochastic variable with 2% load variation.
From the simulation results, the deterministic system
response is indicated by the mean value and is asymptotically
stable. Use the stochastic system stability definition introduced
in Section V. When the loads at buses 3 and 4 have small
variances, the system behaves similar to a deterministic system,
which is asymptotically stable with a probability of 0.9 (ts=15
s, r0=0.05 rad/s).
Fig. 3. IEEE 10-machine 39-bus system.
A. Stochastic Loads at 5% with Low Variances
In the first case, model the loads at buses 3 and 4 (about 5%
of the system load) by Ornstein-Uhlenbeck process. The
variances of the loads are 2% of their mean values. The results
from the ADM-based approach and the Euler-Maruyama
approach are shown in Fig. 4. Among all the generators,
generator 1 has the shortest the electrical distance to bus 3 and
4, hence the rotor angle of it is presented in the following
results.
B. Stochastic Loads at 100% with Low Variances
In the second case, extend stochastic loads to all buses with
variances equal to 2% of their mean values. As shown in Fig.
5, the simulation results from two approaches agree with each
other, which reveal a less stable post-fault system response due
to increased uncertainties.
When all the system loads are stochastic, the system is
asymptotically stable with a probability of 0.6 (ts=15 s, r0=0.05
rad/s). Compared to the first case having only two stochastic
loads with the same r0 value, the probability of the system being
asymptotically stable reduces from 0.9 to 0.6.
Therefore, when the percentage of stochastic loads
increases, even though the load uncertainties are low and the
equilibrium point of the system is almost the same as its
deterministic model, the asymptotic stability of the system in
probability downgrades. That justifies the necessity of using
stochastic load models to study the stability of power systems
with a high penetration of stochastic loads.
6
(a) Result from the ADM-based approach.
(b) Result from the Euler-Maruyama approach.
Fig. 5. Simulation results of generator 1 rotor angle with all loads represented
by stochastic variable with 2% load variation.
In the third case, all the loads are represented by stochastic
loads and the variances of the loads are increased to 4% of the
mean values. This case may represent a scenario having DERs
widely deployed in distribution networks, which make the
aggregated bus load seen from each transmission or subtransmission substation behave more stochastically. The
simulation results from the ADM-based approach and EulerMaruyama approach are shown in Fig. 6.
The ADM-based approach agrees with the Euler-Maruyama
approach on the simulation results. Both of them show that the
system loses its stability when the variance of the loads
increases to 4% of their mean values. The instability is due to
the cumulative effect of stochastic load variations. The 90%
confidence envelope can be utilized as an indicator of the
system stability. Unlike Fig. 5, the 90% confidence envelope in
Fig. 6 is not bounded any more, indicating a 0.9 probability of
the system losing stability.
(a) ADM-based approach
C. Stochastic Loads at 100% with High Variances
(b) Euler-Maruyama approach
Fig. 7. Simulation results of bus voltage at bus 30 with all loads represented by
stochastic variable with 4% load variation.
(a) Result from the ADM-based approach.
(a) Result from the Euler-Maruyama approach.
Fig. 6. Simulation results of generator 1 rotor angle with all loads represented
by stochastic variable with 4% load variation.
Bus voltages also reflect the impact from high load
uncertainties as shown in Fig. 7 about the voltage magnitude of
bus 30, denoted by V30. With loads of high uncertainties, the
system has an increased risk of under- and over-voltage issues
because the imbalance between generation and load is
magnified by increased load uncertainties. That also indicates
the importance of stochastic power system simulation when
penetration of DERs becomes high.
From results of stochastic power system simulation, how the
probability distribution function (PDF) of a system variable
evolves in time during a post-contingency period can be
estimated and fit into an anticipated probability distribution for
analysis. As an example, if we assume V30 to follow a normal
distribution at each time instant with the mean value and
variance varying with time, Fig. 8 shows the evolutions of its
7
PDF using simulation results from both the ADM-based
approach and Euler-Maruyama approach for comparison. Fig.
8a basically matches Fig. 8b, indicating the accuracy of the
proposed ADM-based approach in reflecting the evaluation of
the PDF. From Fig.8, as time elapses, the PDF of the bus
voltage not only shifts the mean value but also increases the
variance indicated by the increasing width of the shape. Such
information is not available from deterministic power system
simulation. The longer the system is subjected to the effect of
stochastic variables the bigger variance and larger uncertainty
the system has in post-contingency dynamics.
Fig. 9. Mean value of generator 1’s rotor angle for case A.
Fig. 10. Standard deviation of generator 1’s rotor angle for case A.
As more loads are modeled as stochastic, the variance of
state variables grows accordingly. The mean value and standard
deviation of the rotor angle of generator 1 for case B are shown
in Fig. 11 and Fig. 12. In case B, the standard deviation reaches
its largest value 0.25 rad/s during the first swing, which is larger
than the largest standard deviation 0.1 rad/s in case A.
(a) ADM-based approach
Fig. 11. Mean value of generator 1’s rotor angle for case B.
(b) Euler-Maruyama approach
Fig. 8. Evolution of the pdf of the voltage magnitude at bus 30 from t=10 s to
20 s.
D. Variances of State Variables
To compare the accuracy of the numerical results from the
ADM-based approach and Euler-Maruyama approach, the
mean value and standard deviation of the 100 trajectories are
compared. For case A, as shown in Fig. 9 and Fig. 10, the ADMbased approach achieves comparable accuracy as the EulerMaruyama approach in terms of both mean value and standard
deviation value.
Fig. 12. Standard deviation of generator 1’s rotor angle for case B.
E. Comparison on Time Performances
The time performances for cases A, B and C of the ADMbased approach and Euler-Maruyama approach are compared in
TABLE I, from which the ADM-based approach takes less than
50% of the time cost of the Euler-Maruyama approach. The
advantage of the ADM-based approach in time performance is
more prominent when many simulation runs are required. As
discussed in [14], the ADM-based approach is inherently
8
suitable for parallel implementation, which could help further
improve the time performance if high-performance parallel
computers are available.
TABLE I
TIME PERFORMANCE COMPARISON OF STOCHASTIC LOAD CASES
Time costs (s)
Ito calculus single run
Ito calculus 100 runs
ADM single run
ADM 100 runs
Stochastic loads
at all buses
(Case B, C)
11.6
1165.1
5.1
511. 0
Stochastic loads
at buses 3 and 4
(Case A)
11.4
1142.4
5.1
503.6
[12]
[13]
[14]
[15]
[16]
[17]
VII. CONCLUSION
This paper proposes an alternative approach for stochastic
simulation of power systems. Using the SAS derived from the
ADM, the stochastic effects from load uncertainties can be
taken into considerations. The result from the proposed
approach is benchmarked with that from the Euler-Maruyama
approach. Since the evaluation of SASs is faster than the
integration with the Euler-Maruyama approach, the proposed
approach has an obviously advantage in time performance. This
is critical when a large number of simulation runs need to be
performed for simulating stochastic behaviors of a future power
grid having a high penetration of DERs. The simulation results
on different levels of stochastic loads show that when the level
of load uncertainty is low, the deterministic simulation is still
trustworthy compared to the mean-value trajectory from
stochastic simulation, but, once the level of load uncertainty
becomes high, the mean-value trajectory no longer represents
the true behavior of the system.
VIII. REFERENCES
[1]
I. Hiskens, J. Alseddiqui, "Sensitivity, approximation and uncertainty in
power system dynamic simulation," IEEE Trans. Power Systems, vol. 21,
no. 4, pp. 1808-1820, Nov. 2006.
[2] M. Tatari, M. Dehghan, M. Razzaghi, "Application of the Adomian
decomposition method for the Fokker-Planck equation," Math. and
Comput. Modelling, vol. 45, no. 5, pp.638-650, Mar. 2007.
[3] B. Spencer, Jr., L. Bergman, "On the numerical solution of the FokkerPlanck equation for nonlinear stochastic systems," Nonlinear Dynamics,
vol. 4, no. 4, pp. 357-372, Aug. 1993.
[4] Y. Saito, T. Mitsui, "Simulation of stochastic differential equations," Ann.
Inst. Stat. Math., vol. 45, no. 3, pp. 419-432, Sep. 1993.
[5] D. J. Higham, "An algorithmic introduction to numerical simulation of
stochastic differential equations," Soc. Ind. and Appl. Math. Review,
vol.43, no.3, pp.525-546, 2001.
[6] K. Wang, M. Crow, "Fokker Planck equation application to analysis of a
simplified wind turbine model," North American Power Symposium,
Champaign, IL, 2012.
[7] F. Milano, R. Zarate-Minano, "A systematic method to model power
systems as stochastic differential algebraic equations," IEEE Trans.
Power Systems, vol. 28, no. 4, pp. 4537-4544, Nov. 2013.
[8] W. Wu, K. Wang, G. Li, Y. Hu, "A stochastic model for power system
transient stability with wind power," IEEE PES General Meeting,
National Harbor, MD, 2014.
[9] K. Wang, M. Crow, "Numerical simulation of stochastic differential
algebraic equations for power system transient stability with random
loads," IEEE PES General Meeting, Detroit, MI, 2011.
[10] B. Yuan, M. Zhou, G. Li, X. Zhang, "Stochastic small signal stability of
power system with wind power generation," IEEE Trans. Power Systems,
vol. 30, no. 4, pp.1680-1689, Jul. 2015.
[11] X. Wang, H. D. Chiang, J. Wang, H. Liu, T. Wang, "Long-term stability
analysis of power systems with wind power based on stochastic
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
differential equations: model development and foundations," IEEE Trans.
Sustainable Energy, vol. 6, no. 4, pp. 1534-1542, Oct. 2015.
N. Duan, K. Sun, "Application of the Adomian decomposition method for
semi-analytic solutions of power system differential algebraic equations,"
IEEE PowerTech, Eindhoven, Netherlands, 2015.
N. Duan, K. Sun, "Finding semi-analytic solutions of power system
differential-algebraic equations for fast transient stability simulation,"
arXiv preprint arXiv:1412.0904, Dec. 2014.
N. Duan, K. Sun, "Power system simulation using the multistage
Adomian decomposition method," IEEE Trans. Power Systems, vol.32,
no. 1, pp. 430-441, Jan. 2017.
G. Adomian, "Nonlinear stochastic differential equations," Math. Anal.
And Applicat., vol. 55, no. 2, pp. 441-452, Aug. 1976.
J. Qi, K. Sun, W. Kang, "Optimal PMU placement for power system
dynamic state estimation by using empirical observability Gramian,"
IEEE Trans. Power Systems, vol. 30, no.4, pp. 2041-2054, Jul. 2015.
F. Galiana, E. Handschin, A. Fiechter, "Identification of stochastic
electric load models from physical data," IEEE Trans. Automat. Control,
vol. 19, no. 6, pp. 887-893, Dec. 1974.
T. Sauer, "Numerical solution of stochastic differential equations in
finance," in Handbook of Mathematical Functions, Springer Berlin
Heidelberg, 2012, pp. 529-550.
K. Nouri, "Study on stochastic differential equations via modified
Adomian decomposition method," U. P. B. Sci. Bull., Series A, vol. 78,
no. 1, pp. 81-90, 2016.
W. Cao, M. Liu, Z. Fan, "MS-stability of the Euler-Maruyama method for
stochastic differential delay equations," Appl. Math. and Comp., vol. 159,
no. 1, pp. 127-135, 2004.
M. Hutzenthaler, A. Jentzen, P. Kloeden, "Strong convergence of an
explicit numerical method for SDEs with non-globally Lipschitz
continuous coefficients", The Ann. of Appl. Probability, vol. 22, no.4, pp.
1611-1641, 2012.
G. Adomian, "A review of the decomposition method in applied
mathematics," J. of Math. Anal. and Applicat., vol. 135, no.2, pp. 501544, Nov. 1988.
U. H. Thygesen, "A survey of Lyapunov techniques for stochastic
differential equations," Dept. of Math. Modelling, Tech. Univ. of
Denmark, Lyngby, Denmark, Rep. IMM Technical Report nr. 18-1997.
X. Mao, Stochastic Differential Equations and Applications, 2nd edition,
Chichester, UK, Horwood, 2008.
K. Burrage, P. Burrage, T. Mitsui, "Numerical solutions of stochastic
differential equations – implementation and stability issues," J. of
Computational and Appl. Math., vol. 125, no. 1, pp. 171-182, 2000.
F. Kozin, "A survey of stability of stochastic systems," Automatica, vol.
5, no. 1, pp. 95-112, Jan. 1969.
| 3cs.SY
|
MagnifyMe: Aiding Cross Resolution Face Recognition via Identity Aware
Synthesis
Maneet Singh, Shruti Nagpal, Richa Singh, Mayank Vatsa, and Angshul Majumdar
IIIT-Delhi, India
arXiv:1802.08057v1 [cs.CV] 22 Feb 2018
{maneets, shrutin, rsingh, mayank, angshul}@iiitd.ac.in
Abstract
Enhancing low resolution images via super-resolution
or image synthesis for cross-resolution face recognition
has been well studied. Several image processing and machine learning paradigms have been explored for addressing the same. In this research, we propose Synthesis via
Deep Sparse Representation algorithm for synthesizing a
high resolution face image from a low resolution input image. The proposed algorithm learns multi-level sparse representation for both high and low resolution gallery images,
along with an identity aware dictionary and a transformation function between the two representations for face identification scenarios. With low resolution test data as input, the high resolution test image is synthesized using the
identity aware dictionary and transformation which is then
used for face recognition. The performance of the proposed
SDSR algorithm is evaluated on four databases, including
one real world dataset. Experimental results and comparison with existing seven algorithms demonstrate the efficacy
of the proposed algorithm in terms of both face identification and image quality measures.
Low Resolution
Face Image
Bicubic Interpolation
Low Resolution
Face Image
High Resolution
Gallery Image
Bicubic Interpolation
Figure 1: Images captured few minutes before Boston
Marathon Bombing, 2013, of suspect Dzhokhar Tsarnaev
(circled). The resolution of the circled image is less than
24 × 24, which is interpolated to (96 × 96).
covariate of face recognition with widespread applications.
Several researchers have shown that the performance of
state-of-the-art (SOTA) algorithms reduces while matching
cross-resolution face images [4, 30, 31]. In order to overcome this limitation, an intuitive approach is to generate a
high resolution image for the given low resolution input,
which can be provided as input to the face recognition engine. Figure 1 shows a sample real world image captured
before the Boston Bombing (2013). Since the person of
interest is at a distance, the face captured is thus of low
resolution. Upon performing bicubic interpolation to obtain a high resolution image, results in an image suffering from blur and poor quality. With the ultimate aim of
high recognition performance, the generated high resolution image should have good quality while preserving the
identity of the subject. As elaborated in the next subsection, while there exist multiple synthesis or super resolution techniques, we hypothesize that utilizing a (domain)
face-specific, recognition-oriented model for face synthesis
should result in improved recognition performance, especially for close-set recognition scenarios. To this effect, this
work presents a novel domain specific identity aware Synthesis via Deep Sparse Coding algorithm for synthesizing
a high resolution face image from a given low resolution
input image.
1. Introduction
Group images are often captured from a distance, in
order to capture multiple people in the image. In such
cases, the resolution of each face image is relatively smaller,
thereby resulting in errors during automated tagging. Similarly, in surveillance and monitoring applications, cameras
are often designed to cover the maximum field of view, this
often limits the size of face images captured, especially for
individuals at a distance. If we use these images to match
against high resolution images, e.g., profile images on social media or mugshot images captured by law enforcement,
then resolution gap between the two may lead to incorrect
results. This task of matching a low resolution input image
against a database of high resolution images is referred to
as cross resolution face recognition and it is a challenging
1
1.1. Literature Review
In literature, different techniques have been proposed
to address the problem of cross resolution face recognition. These can broadly be divided into transformation
based techniques and non-transformation based techniques.
Transformation based techniques address the resolution difference between images by explicitly introducing a transformation function either at the image or at the feature level.
Non-transformation techniques propose to learn/extract resolution invariant features or classifiers, in order to address
the resolution variations [4, 30]. In 2013, Wang et al. [31]
present an exhaustive review of the proposed techniques for
addressing cross resolution face recognition.
Peleg and Elad [21] propose a statistical model that uses
Minimum Mean Square Error estimator on high and low
resolution image pair patches for prediction. Lam [14] propose a Singular Value Decomposition based approach for
super resolving low resolution face images. Researchers
have also explored the domain of representation learning
to address the problem of cross resolution face recognition.
Yang et al. [32] propose learning dictionaries for low and
high resolution image patches jointly followed by learning a
mapping between the two. Yang et al. [33] propose a Sparse
Representation-based Classification approach in which the
face recognition and hallucination constraints are solved simultaneously. Gu et al. [13] propose convolutional sparse
coding where an image is divided into patches and filters
are learned to decompose a low resolution image into features. A mapping is learned to predict high resolution feature maps from the low resolution features. Mundunuri and
Biswas [19] propose a multi-dimensional scaling and stereo
cost technique to learn a common transformation matrix for
addressing the resolution variations.
A parallel area of research is that of super-resolution,
where research has focused on obtaining a high resolution
image from a given low resolution image, with the objective of maintaining/improving the visual quality of the input [3, 22, 28, 29]. There has been significant advancement in the field of super-resolution over the past several
years including recent representation learning architectures
[5, 6, 7, 16, 26] being proposed for the same. It is important to note that while such techniques can be utilized for
addressing cross resolution face recognition, however, they
are often not explicitly trained for face images, or for providing recognition-oriented results.
1.2. Research Contributions
This research focuses on cross resolution face recognition by proposing a recognition-oriented image synthesis algorithm, capable of handling large magnification factors. We propose a deep sparse representation based transfer learning approach termed as Synthesis via Deep Sparse
Representation (SDSR). The proposed identity aware syn-
thesis algorithm can be incorporated as a pre-processing
module prior to any existing face recognition engine to enhance the resolution of a given low resolution input. In order to ensure recognition-oriented synthesis, the proposed
model is trained using a gallery database having a single
image per subject. The results are demonstrated with four
databases and the effectiveness is evaluated in terms of both
no-reference image quality measure of the synthesized images and face identification accuracies with existing face
recognition models.
2. Synthesis via Deep Sparse Representation
Dictionary Learning algorithms have an inherent property of representing a given sample as a sparse combination of it’s basis functions [23]. This property is utilized in
the proposed SDSR algorithm to synthesize a high resolution image from a given low resolution input. The proposed
model learns a transformation between the representations
of low and high resolution images. That is, instead of interpolating the pixel values, this work focuses on interpolating a more abstract representation. Further, motivated by
the abstraction capabilities of deep learning, we propose to
learn the transformation from deeper levels of representation. Unlike traditional dictionary learning algorithms, we
propose to learn the transformation at deeper levels of representation. This leads to the key contribution of this work:
Synthesis via Deep Sparse Representation (SDSR), a transfer learning approach for synthesizing a high resolution image for a given low resolution input.
2.1. Preliminaries
Let X = [x1 |x2 |...|xn ] be the input training data with n
samples. Dictionary learning algorithms learn a dictionary
(D) and sparse representations (A) using data (X). The
objective function of dictionary learning is written as:
n
1 X 1 i
2
x − Dαi F + λ αi 1
(1)
min
D, A n
2
i=1
where, A = [α1 |α2 |....|αn ] are the sparse codes, k·k1 represents `1 -norm, and λ is the regularizing constant that controls how much weight is given to induce sparsity in the
representations. In Eq. 1, the first term minimizes the reconstruction error of the training samples, and the second
term is a regularization term on the sparse codes.
In literature, researchers have proposed extending a
single level dictionary to a multi-level dictionary to
learn multiple levels of representations of the given
data [20, 24, 25]. A k−level deep dictionary learns
k dictionaries D = {D1 , ..., Dk } and sparse coefficients
A = {A1 , ..., Ak } for a given input X.
n
1 X 1 i
2
(2)
x − D1 ...Dk αk,i F + λ αk,i 1
min
D, A n
2
i=1
The architecture of deep dictionary is inspired from the deep
learning techniques where deeper layers of feature learning enhance the level of abstraction learned by the network,
thereby learning meaningful latent variables.
In real world scenarios of surveillance or image tagging, the task is to match the low resolution test images
(probe) to the database of high resolution images known
as gallery images. Without loss of generality, we assume
that the target comprises of high resolution gallery images
while the source domain consists of low resolution images. In the proposed model, for low resolution face images
Xl , and high resolution face images Xh , k level deep dictionaries are learned in both source (GL = {G1l , ..., Gkl })
and target domain (GH = {G1h , ..., Gkh }). It is important
to note that the dictionaries are generated using the preacquired gallery images. Corresponding sparse representations, AL = {A1l , ..., Akl } and AH = {A1h , ..., Akh } are also
k,2
k,n
learned for all k levels, where Akh = [αk,1
h |αh |...|αh ] are
the representations learnt corresponding to the high resoluk,2
k,n
tion deep dictionary and Akl = [αk,1
l |αl |...|αl ] are the
th
representations learnt from the k level dictionary for the
low resolution images. The proposed algorithm learns a
transformation, M, between Akh and Akl . The optimization
formulation for Synthesis via Deep Sparse Representation
(SDSR), a k-level deep dictionary is written as:
n
1 X 1 i
xh − G1h ...Gkh αk,i
h
GH ,AH , n
2
G ,A ,M
i=1
2
+
min
F
L
+
1 i
x − G1l ...Gkl αk,i
l
2 l
2
2
k,i
+ λM αk,i
h − Mαl
+
F
k
X
j=1
k
X
j,i
(λj αh
j=1
j,i
(λj αl
)
1
)
1
F
where, λj are regularization parameters which control the
amount of sparsity in the learned representations of the j th
layer, while λM is the regularization constant for learning
the transformation function. GH and GL correspond to the
deep dictionaries learned for the high and low resolution
gallery images, respectively. The SDSR algorithm learns
multiple levels of dictionaries and corresponding representations for low and high resolution face images, along with
a transformation between the features learned at the deepest
layer.
Training SDSR Algorithm
Without loss of generality, training of the proposed SDSR
algorithm is explained with k = 2 (shown in Figure 2). For
1
L
+λ2
1
+
α2,i
l
1 i
x − G1l G2l α2,i
l
2 l
1
+ λM
α2,i
h
2
+ λ1 α1,i
l
2
− Mα2,i
l
F
F
1
(4)
Since the number of variables in Eq. 4 is large (even more
for deeper dictionaries), directly solving the optimization
problem may provide incorrect estimates and lead to overfitting. Therefore, greedy layer by layer training is applied.
It is important to note that since there is a l1 -norm regularizer on the coefficients of the first and the second layer, the
dictionaries G1h and G2h cannot be collapsed into one dictionary. In order to learn the estimates, Eq. 4 is split into
learning the first level representation, second level representation, and the transformation. From Eq. 2, the optimization
function for two level deep dictionary is as follows:
n
1 X 1 i
2
x − G1 G2 α2,i F +λ1 α1,i 1 +λ2 α2,i
min
2
G1 ,D2 , n
i=1
A1 ,A2
(5)
Assuming an intermediate variable, α1,i = G2 α2,i such
that A1 = [α1,1 |α1,2 |...|α1,n ], the above equation can be
modeled as a step-wise optimization of the following two
equations:
n
1 X 1 i
2
min
x − G1 α1,i F + λ1 α1,i 1
(6)
2
G1 , A 1 n
i=1
n
1 X 1
α1,i − G2 α2,i
min
2
G2 , A2 n
i=1
(3)
2.2.1
H
+λ2 α2,i
h
2.2. SDSR Algorithm
L
a two level deep dictionary, Eq. 3 can be written as:
n
1 X 1 i
2
1,i
min
xh − G1h G2h α2,i
h F + λ 1 αh
GH ,GL
n
2
A ,A ,M
i=1
2
F
+ λ2 α2,i
1
(7)
A deep dictionary of two levels (Eq. 5) requires two
steps for learning (Eq. 6, 7). Upon extending the formulation to k level deep dictionary, it would require exactly
k steps for optimization. The proposed SDSR algorithm
(Eq. 3) builds upon the above and utilizes k + 1 steps based
greedy layer-wise learning for a k level deep dictionary. k
steps are for learning representations using the deep dictionary architecture and the k + 1th step is for learning the
transformation between the final representations. Therefore, Eq. 4 is solved using an independent three step approach: (i) learn first level source (low resolution) and target (high resolution) domain dictionaries, (ii) learn second
level low and high resolution image dictionaries, and (iii)
learn a transformation between the final representations.
Using the concept in Eq. 5 - 7, in the first step, two separate level-1 (i.e. k = 1) dictionaries are learned from the
given input data for the low resolution (G1l ) and high resolution (G1h ) face images independently. Given the training
data consisting of low (Xl ) and high (Xh ) resolution face
1
Learn Sparse
Representations
Learn Sparse
Representations
Gh2
G1h
High Resolution Gallery Images
1st Level Dictionary
2nd Level Sparse Representations
2nd Level Dictionary
1st Level Sparse Representations
(a)
Learn Sparse
Representations
Learn Sparse
Representations
G1l
Low Resolution Gallery Images
Gl2
1st Level Dictionary
2nd Level Dictionary
1st Level Sparse Representations
Gl2
G1l
(b)
xltest
Learn
Transformation
‘M’
Using Transformation
M
2nd Level Sparse Representations
G1h
Gh2
Face Recognition
Engine
α1,test
h
α2,test
h
α2,test
l
α1,test
l
xhtest
Figure 2: Synthesis via Deep Sparse Representation algorithm for 2-level deep dictionary. (a) refers to the training of the
model, while (b) illustrates the high resolution synthesis of a low resolution input.
images, the following minimization is applied for the two
domains respectively:
n
1X1 i
2
min
xl − G1l α1,i
+ λ1 α1,i
(8)
l
l
2
1
1
1 n
2
Gl , A l
i=1
n
min
G1h , A1h
1X1 i
x − G1h α1,i
h
n i=1 2 h
2
2
+ λ1 α1,i
h
1
(9)
1,2
1,n
here,
A1l = [α1,1
and
l |αl |...|αl ]
1,1
1,2
1,n
1
Ah = [αh |αh |...|αh ] refer to the level-1 sparse
codes learned for the low and high resolution images
respectively. Each of the above two equations can be
optimized independently using an alternating minimization
dictionary learning technique over the dictionary G1 and
representation A1 [17]. After this step, G1h , G1l (dictionaries) and A1h , A1l (representations) are obtained for the two
varying resolution data.
In the second step, a deep dictionary is created by learning the second level dictionaries (G2l , G2h ) using the representations obtained from the first level (A1h and A1l ). That
is, two separate dictionaries, one for low resolution images
and one for high resolution images are learned using the
representations obtained at the first level as input features.
The equations for this step can be written as follows:
n
1X1
2 2,i 2
min
α1,i
+ λ2 α2,i
(10)
l − Gl αl
l
2
1
2
2 n
2
Gl , Al
i=1
n
1X1
2 2,i
min
α1,i
h − Gh αh
2
2 n
2
Gh , Ah
i=1
2
2
+ λ2
α2,i
h 1
(11)
2,2
2,n
here, A2l = [α2,1
l |αl |...|αl ] is the final represen-
tation obtained for the low resolution images and
2,2
2,n
A2h = [α2,1
h |αh |...|αh ] refers to the representation obtained for the high resolution images. Similar to the previous step, the equations can be solved independently using
alternating minimization over the dictionary and representations. After this step, G2l , G2h , A2l and A2h are obtained.
In order to synthesize from one resolution to another, the
third step of the algorithm involves learning a transformation between the deep representations of the two resolutions
(i.e. A2l and A2h ). The following minimization is solved to
obtain a transformation M.
1
2
min
A2h − MA2l F
(12)
M 2
The above equation is a least square problem having a
closed form solution. After training, the dictionaries
(G1l , G1h , G2l , G2h ) and the transformation function (M) are
obtained which are then used at test time.
2.2.2
Testing: Synthesizing High Resolution Face Image from Low Resolution Image
During testing, a low resolution test image, xtest
l , is input
to the algorithm. Using the trained gallery based dictionaries, G1l and G2l , first and second level representations
(α1,test
, α2,test
) are obtained for the given image:
l
l
xtest
= G1l α1,test
; α1,test
= G2l α2,test
l
l
l
l
(13)
The transformation function, M, learned in Eq. 12, is then
used to obtain the second level high resolution representation (α2,test
).
h
α2,test
= Mα2,test
h
l
(14)
Table 1: Summarizing the characteristics of the training and testing partitions of the databases used in experiments.
Dataset
CMU Multi-PIE [12]
CAS-PEAL [10]
Real World Scenarios [4]
SCface [11]
Training
Subjects
100
500
50
Training
Images
200
659
300
Testing
Subjects
237
540
1207
80
Testing
Images
474
705
1222
480
Gallery
Resolution
Probe Resolutions
8 × 8, 16 × 16, 24 × 24,
96 × 96
24 × 24, 32 × 32, 48 × 48
Probe
Resolution
Original
Image
Bicubic
Interp.
Dong et
al. [7]
Kim et
al. [15]
Gu et
al. [13]
Dong et
al. [6]
Peleg et
al. [21]
Yang et
al. [32]
Proposed
SDSR
CMU
MultiPIE
8×8
16×16
24×24
32×32
48×48
0.0±0.0
0.0±0.0
1.2±0.4
3.4±0.6
91.9±1.1
0.1±0.0
1.1±0.1
3.1±0.6
16.9±1.3
95.8±0.4
2.0±3.5
9.7±1.1
85.8±0.7
4.1±1.0
17.5±1.1
96.2±0.6
4.3±1.0
15.4±1.1
93.1±0.9
4.2±0.6
6.9±0.2
95.5±1.1
8.3±0.9
92.8±0.4
94.0±0.6
82.6±1.5
91.1±1.3
91.8±1.8
91.9±1.7
91.5±1.5
CAS-PEAL
8×8
16×16
24×24
32×32
48×48
0.0±0.0
0.0±0.0
0.4±0.6
3.7±0.7
63.4±1.7
0.0±0.0
0.2±0.3
14.9±1.7
76.5±1.8
90.8±1.5
0.4±0.2
5.4±1.2
46.5±2.5
2.3±0.8
11.8±1.1
75.8±2.3
1.9±0.7
8.1±2.3
77.7±2.1
2.5±0.7
2.1±0.7
72.0±0.7
3.1±1.6
74.0±2.6
73.3±3.3
92.8±0.7
94.4±1.1
95.3±1.4
95.6±1.1
95.4±1.5
SCface
Table 2: Rank-1 identification accuracies (%) obtained using Verilook (COTS-I) for cross resolution face recognition. The
target resolution is 96 × 96. The algorithms which do not support the required magnification factor are presented as 0 −0 .
24×24
32×32
48×48
1.1±0.2
1.8±0.5
6.5±0.6
0.8±0.1
2.5±0.3
9.5±1.9
0.4±0.2
2.2±0.4
6.9±0.6
0.4±0.2
2.0±0.0
6.7±1.2
1.5±0.3
2.3±0.3
7.7±0.6
1.3±0.3
0.7±0.3
7.5±1.3
1.8±0.5
7.3±0.9
6.8±0.7
14.7±3.3
15.6±1.3
18.5±2.6
Using Eq. 9 and Eq. 11 and the second level representation for the given image in the target domain, a synthesized
output of the given image is obtained. First α1,test
is calcuh
lated with the help of G2h and then xtest
h is obtained using
G1h , which is the synthesized image in the target domain.
1 1,test
α1,test
= G2h α2,test
; xtest
h = Gh αh
h
h
(15)
It is important to note that the synthesized high resolution
image is a sparse combination of the basis functions of the
learned high resolution dictionary. In order to obtain a good
quality, identity-preserving high resolution synthesis, the
dictionary is trained with the pre-acquired high resolution
database. This ensures that the basis functions of the trained
dictionaries span the latent space of the images. As will be
demonstrated via experiments as well, a key highlight of
this algorithm is to learn good quality, representative dictionaries with a single sample per subject as well. The high
resolution synthesized output image xtest
h can then be used
by any face identification engine for recognition.
3. Databases and Experimental Protocol
The effectiveness of the proposed SDSR algorithm
is demonstrated by evaluating the face recognition performance with original and synthesized images. Two
commercial-off-the-shelf face recognition systems (COTS),
Verilook (COTS-I) [2] and Luxand (COTS-II) [1] are used
on four different face databases. For Verilook, the face quality and confidence thresholds are set to minimum, in order
to reduce enrollment errors. The performance of the proposed algorithm is compared with six recently proposed
super-resolution and synthesis techniques by Kim et al.
[15]1 (kernel ridge regression), Peleg et al. [21]2 (sparse
representation based statistical prediction model), Gu et al.
[13]3 (convolutional sparse coding), Yang et al. [32]4 (dictionary learning), Dong et al. [6]5 (deep convolutional networks), and Dong et al. [7]6 (deep convolutional networks)
along with one of the most popular technique, bicubic interpolation. The results of the existing super-resolution algorithms are computed by using the models provided by
the authors at the links provided in the footnotes. It is to
be noted that not all the algorithms support all the levels of
magnification. For instance, the algorithm proposed by Kim
et al. [15] supports up to 4 levels of magnification whereas,
Yang et al.’s algorithm supports up to 2 levels of magnification.
Face Databases: Table 1 summarizes the statistics of
the databases in terms of training and testing partitions,
along with the resolutions. Details of the databases are
provided below:
1. CMU Multi-PIE Dataset [12]: Images pertaining
1 https://people.mpi-inf.mpg.de/
kkim/supres/supres.htm
2 http://www.cs.technion.ac.il/ẽlad/software/
3 http://www4.comp.polyu.edu.hk/
cslzhang/
4 http://www.ifp.illinois.edu/j̃yang29/
5 http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html
6 http://mmlab.ie.cuhk.edu.hk/projects/FSRCNN.html
to 337 subjects are selected with frontal pose, uniform
illumination, and neutral expression. 100 subjects are used
for training while the remaining 237 are in the test set.
2. CAS-PEAL Dataset [10] consists of face images of
1040 subjects. All subjects have a single, high-resolution
normal image and the dataset contains images of different
covariates such as lighting, expression, and distance. For
this research, normal images are used as the high resolution
gallery database while face images under the distance
covariate are downsampled and used as probe images.
3. SCface Dataset [11]: It consists of 130 subjects, each
having one high resolution frontal face image and multiple
low resolution images, captured from three distances using
surveillance cameras.
4. Real World Scenarios Dataset [4] contains images
of seven subjects associated with the London Bombing,
Boston Bombing, and Mumbai Attacks. Each subject
has one high resolution gallery image and multiple low
resolution test images. The test images are captured from
surveillance cameras and are collected from multiple
sources from the Internet. Since the number of subjects
are just seven, in order to mimic a real world scenario, the
gallery size is increased to create an extended gallery of
1200 subjects. Images from the Multi-PIE, ND Human
Identification Set-B [8], and MEDS[9] datasets are used for
the same.
Protocol: For all the datasets, a real world matching
protocol is followed. For each subject, multiple low resolution images are used as probe images which are matched
against the pre-acquired database of high resolution gallery
images. Only a single high resolution image per subject is
used as gallery. The proposed and comparative algorithms
are used to synthesize (or super-resolve) a high resolution
image from a given low resolution input. The magnification
factor varies from 2 (for probes of 48 × 48) to 12 (for
probes of 8 × 8) to match it against the gallery database
of size 96 × 96. For all the databases except the SCface,
test images are of sizes varying from 8 × 8 to 48 × 48. For
the SCface database, predefined protocol is followed and
probe resolutions are 24 × 24, 32 × 32, and 48 × 48. Face
detection is performed using face co-ordinates (if provided)
or using Viola Jones Face Detector [27] and synthetic
downsampling is performed to obtain lower resolutions.
All the experiments are performed with five times random
sub-sampling to ensure consistency.
Implementation Details: The SDSR algorithm is trained
using the pre-acquired gallery database for each dataset.
The regularization constant for sparsity is kept at 0.85.
Different dictionaries have different dimensions, based
on the input data. For instance, the two-level dictionaries
created for SCface dataset contain 100 and 80 atoms in the
first and second dictionary respectively. The source code
Input
Bicubic
Proposed
Input
Bicubic
Proposed
Figure 3: Sample images from SCface dataset incorrectly
synthesized by the SDSR algorithm for 32 × 32 input.
of the algorithm will be made publicly available in order to
ensure reproducibility of the proposed approach.
4. Results and Analysis
The proposed algorithm is evaluated with three sets of
experiments: (i) face recognition performance with resolution variations, (ii) image quality measure, and (iii) face
identification analysis with different dictionary levels. The
resolution of the gallery is set to 96 × 96. For the first experiment, the probe resolution varies from 8 × 8 to 48 × 48,
while it is fixed to 24 × 24 for the next two experiments.
4.1. Face Recognition across Resolutions
For all datasets and resolutions, results are tabulated in
Tables 2 to 4. The key observations pertaining to these set
of experiments are presented below:
8 × 8 and 16 × 16 probe resolutions: Except bicubic interpolation, none of the existing super resolution or synthesis algorithms used in this comparison support a magnification factor of 12 (for 8 × 8) or 6 (for 16 × 16); therefore,
the results on these two resolutions are compared with original resolution (when the probe is used as input to COTS as
it is, without any resolution enhancement) and bicubic interpolation only. As shown in the third and fourth columns
of the two tables, on the CMU Multi-PIE and CAS-PEAL
databases, matching with original and bicubic interpolated
images results in an accuracy of ≤ 1.1% whereas, the
images synthesized using the proposed algorithm provide
rank-1 accuracy of 82.6% and 92.8%, respectively.
24 × 24 and 32 × 32 probe resolutions: As shown in Table 2, on CMU Multi-PIE and CAS-PEAL databases with
test resolution of 24 × 24 and 32 × 32, the synthesized images obtained using the proposed SDSR algorithm, yield
a rank-1 accuracy of ≥ 91.8%. Other approaches yield a
rank-1 accuracy of less than 20%, except bicubic interpolation on 32 × 32 size which provides rank-1 accuracy of
76.5%. As shown in Table 3, similar performance trends are
observed using COTS-II on the two databases. For SCface,
the rank-1 accuracy with SDSR is significantly higher than
the existing approaches; however, due to the challenging
nature of the database, both commercial matchers provide
low rank-1 accuracies. Fig. 3 presents sample images from
the SCface dataset, incorrectly synthesized via the proposed
SDSR algorithm. Varying acquisition devices of the training and testing partitions, along with the covariates of pose
and illumination creates the problem further challenging.
Probe
Resolution
Original
Image
Bicubic
Interp.
Dong et
al. [7]
Kim et
al. [15]
Gu et
al. [13]
Dong et
al. [6]
Peleg et
al. [21]
Yang et
al. [32]
Proposed
SDSR
CMU
Multi-PIE
8×8
16×16
24×24
32×32
48×48
0.0±0.0
0.0±0.0
0.9±0.3
11.3±1.1
90.2±0.5
0.0±0.0
0.0±0.0
1.0±0.3
18.3±7.1
97.9±0.5
2.3±0.5
13.5±0.6
96.0±0.6
5.9±0.7
28.6±1.2
97.1±0.7
6.8±1.4
24.3±2.3
96.6±0.5
6.8±0.5
19.4±1.5
96.9±0.6
17.4±2.5
97.5±0.5
96.2±0.4
82.3±1.4
90.5±1.1
92.1±1.5
92.2±1.6
91.9±1.6
CAS-PEAL
8×8
16×16
24×24
32×32
48×48
0.0±0.0
0.0±0.0
0.5±0.4
11.1±2.8
88.1±0.6
0.0±0.0
0.6±0.6
49.3±1.3
92.5±2.0
95.4±1.4
2.3±0.8
27.9±1.1
85.7±2.8
10.2±1.1
34.6±2.3
93.3±1.4
7.3±1.5
31.7±2.3
93.3±1.4
5.8±0.4
15.2±3.3
91.4±1.4
28.0±1.9
93.6±1.9
90.8±1.7
91.7±0.9
93.3±0.7
93.7±1.4
93.9±1.3
93.9±1.5
SCface
Table 3: Rank-1 identification accuracies (%) obtained using Luxand (COTS-II) for cross resolution face recognition. The
target resolution is 96 × 96. The algorithms which do not support the required magnification factor are presented as 0 −0 .
24×24
32×32
48×48
1.1±0.2
2.2±0.4
9.7±1.7
1.5±1.0
3.2±0.4
12.6±1.7
1.2±0.5
3.5±0.5
9.6±1.1
0.8±0.1
2.6±0.7
11.6±1.3
2.2±0.5
4.0±0.7
10.1±1.8
1.9±0.6
2.6±1.0
11.7±2.0
2.8±0.7
11.4±1.2
11.9±1.0
14.7±2.4
15.7±1.3
19.1±3.4
CASPEAL
CMU
Multi-PIE
Real
World
SCface
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Figure 4: Probe images of 24 × 24 are superresolved/synthesized to 96×96. (a) corresponds to the original probe, (b)-(f) correspond to different techniques: bicubic interpolation, Kim et al [15], Gu et al. [13], Dong et al.
[6], Dong et al. [7], and the proposed SDSR algorithm.
48 × 48 probe resolution: Using COTS-I, the proposed
algorithm achieves improved performance than other techniques, except on the CMU Multi-PIE dataset, where it does
not perform as well. On all other databases, the proposed algorithm yields the best results. Upon analyzing both the Tables, it is clear that the proposed algorithm is robust to different recognition systems and performs well without any
bias for a specific kind of recognition algorithm.
Another observation is that with COTS-II, images superresolved using bicubic interpolation yield best results on the
first two databases. However, it should be noted that these
results are only observed for a magnification factor of 2 and
for images which were synthetically down-sampled. In real
world surveillance datasets, such as SCface, the proposed
approach performs best with both commercial systems.
Real World Scenarios Dataset: Table 4 summarizes the
results of COTS-I on Real World Scenarios dataset. Since
the gallery contains images from 1200 subjects, we sum-
marize the results in terms of the identification performance
with top 20% retrieved matches. It is interesting to observe
that for all test resolutions, the proposed algorithm significantly outperforms existing approaches. SDSR achieves a
53.3% identification accuracy on probe resolution of 8 × 8
and an accuracy of 60.0% for 48 × 48 test resolution.
Cross Dataset Experiments: The SDSR algorithm was
trained on the CMU Multi-PIE dataset, and tested on the
SCface dataset for a probe resolution of 24 × 24. A rank1 identification accuracy of 1.62% (1.92%) was obtained
using COTS-I (COTS-II), whereas a rank-5 identification
accuracy of 7.54% and 9.06% was obtained respectively.
The results showcase that the proposed model is still able
to achieve better recognition performance as compared to
other techniques. However, the drop in accuracy strengthens our hypothesis that using an identity-aware model for
performing synthesis is more beneficial for achieving higher
classification performance.
4.2. Quality Analysis
Fig. 4 shows examples of synthesized/super-resolved
images from multiple databases generated using the proposed and existing algorithms. In this figure, images of
96 × 96 are synthesized from low resolution images of
24 × 24. It can be observed that the output images obtained using existing algorithms (columns (b) - (f)) have artifacts in terms of blockiness and/or blurriness. However,
the quality of the images obtained using the proposed algorithm (column (g)) are significantly better than the other
algorithms. To compare the visual quality of the outputs, a
no reference image quality measure, BRISQUE [18] is utilized. Blind/Referenceless Image Spatial QUality Evaluator
(BRISQUE) computes the distortion in the image by using
the statistics of locally normalized luminance coefficients.
It is calculated in the spatial domain and is used to estimate
the losses of naturalness in an image. Lower the value, less
Table 4: Real World Scenarios: Recognition accuracy obtained in top 20% ranks, against a gallery of 1200 subjects using
COTS-I (Verilook) having resolution of 96x96.
Probe
Resolution
Original
Image
Bicubic Interpolation
Dong et
al. [7]
Kim et
al. [15]
Gu et al.
[13]
Dong et
al. [6]
Peleg et
al. [21]
Yang et
al. [32]
Proposed
SDSR
8×8
16×16
24×24
32×32
48×48
0.0
0.0
6.6
33.3
33.3
0.0
13.3
16.6
16.6
46.6
13.3
33.3
26.6
26.6
33.3
33.3
13.3
33.3
26.6
6.6
16.6
20.0
40.0
33.3
40.0
53.3
53.3
53.3
53.3
60.0
Table 5: Average no reference quality measure - BRISQUE [18] for probe resolution of 24 × 24 synthesized to 96 × 96,
obtained over five folds. A lower value for BRISQUE corresponds to lesser distortions in the image.
Database
Bicubic Interp.
Dong et al. [7]
Kim et al. [15]
Gu et al. [13]
Dong et al. [6]
Proposed SDSR
CMU Multi-PIE
CAS-PEAL
SCface
Real World
54.8 ± 0.1
60.0 ± 0.2
58.7 ± 0.1
57.5
28.94 ± 0.0
52.86 ± 0.0
52.86 ± 0.0
28.94
50.8 ± 0.1
54.3 ± 0.2
53.2 ± 0.2
54.5
52.8 ± 0.1
56.4 ± 0.1
54.9 ± 0.1
54.6
48.8 ± 0.1
53.4 ± 0.1
47.2 ± 0.1
49.54
26.2 ± 1.3
39.3 ± 0.3
34.2 ± 0.6
25.9
Table 6: Rank-1 accuracies (%) for varying levels of SDSR
algorithm with 24 × 24 probe and 96 × 96 gallery.
Database
COTS
CMU
Multi-PIE
Verilook
Luxand
Verilook
Luxand
Verilook
Luxand
CAS-PEAL
SCface
k=1
91.4
92.0
93.8
92.2
15.0
15.6
Dictionary Levels
k=2
k=3
91.8
91.8
92.1
92.5
95.3
93.7
93.7
93.6
14.7
15.2
14.7
15.3
distorted is an image. From Table 5, it can be seen that images obtained using the proposed SDSR algorithm have a
better (lower) BRISQUE score as compared to images generated with existing algorithms; a difference of at least 15
points is observed in the BRISQUE scores.
4.3. Effect of Dictionary Levels
As explained in the algorithm section, synthesis can be
performed at different levels of deep dictionary, i.e. with
varying values of k. This experiment is performed to analyze the effect of different dictionary levels on identification
performance. The proposed algorithm is used to synthesize
high resolution images (96 × 96, magnification factor of 4)
from input images of size 24 × 24 with varying dictionary
levels, i.e. k = 1, 2, 3. First level dictionary (k = 1) is
equivalent to shallow dictionary learning, whereas two and
three levels correspond to synthesis with deep dictionary
learning. Table 6 reports the rank-1 identification accuracies obtained with the two commercial matchers for four
databases. The results show that the proposed approach
with k = 2, generally, yields the best results. In some cases,
the proposed approach with k = 3 yields better results.
Generally, abstraction capability of deeper layers and overfitting are two effects in deep learning based approaches. In
Table 6, we observe the trade-off between these two. Most
of the datasets are moderately sized; therefore, we observe
good results in the second layer. In the third layer, overfitting offsets the abstraction, hence we see none to marginal
changes. Further, computational complexity with 3-level
deep dictionary features is higher and the improvements in
accuracy are not consistent across databases. On the other
hand, paired t-test on the results obtained by the shallow
dictionary and 2-level deep dictionary demonstrate statistical significance even with a confidence level of 95% (for
Verilook). Specifically, for a single image, synthesis with
level-1 dictionary requires 0.42 ms, level-2 requires 0.43
ms, and level-3 requires 0.45 ms.
5. Conclusion
The key contribution of this research is a recognitionoriented pre-processing module based on dictionary learning algorithm for synthesizing a high resolution face image
from low resolution input. The proposed SDSR algorithm
learns the representations of low and high resolution images
in a hierarchical manner along with a transformation between the representations of the two. The results are demonstrated on four databases with test image resolutions ranging from 8 × 8 to 48 × 48. Matching these requires generating synthesized high resolution images with a magnification
factor of 2 to 12. Results computed in terms of both image
quality measure and face recognition performance illustrate
that the proposed algorithm consistently yields good recognition results. Computationally, the proposed algorithm requires less than 1 millisecond for generating a synthesized
high resolution image which further showcases the efficacy
and usability of the algorithm for low resolution face recognition applications.
References
[1] Luxand. https://www.luxand.com. 5
[2] Verilook. http://www.neurotechnology.com/verilook.html. 5
[3] S. Baker and T. Kanade. Hallucinating faces. In IEEE International Conference on Automatic Face and Gesture Recognition, FG ’00, pages 83–, 2000. 2
[4] H. S. Bhatt, R. Singh, M. Vatsa, and N. K. Ratha. Improving cross-resolution face matching using ensemble-based cotransfer learning. IEEE Transactions on Image Processing,
23(12):5654–5669, December 2014. 1, 2, 5, 6
[5] R. Dahl, M. Norouzi, and J. Shlens. Pixel recursive super
resolution. In IEEE International Conference on Computer
Vision, 2017. 2
[6] C. Dong, C. C. Loy, K. He, and X. Tang.
Image
super-resolution using deep convolutional networks. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
38(2):295–307, 2016. 2, 5, 7, 8
[7] C. Dong, C. C. Loy, and X. Tang. Accelerating the superresolution convolutional neural network. In European Conference on Computer Vision, pages 391–407. Springer, 2016.
2, 5, 7, 8
[8] P. J. Flynn, K. W. Bowyer, and P. J. Phillips. Assessment of
time dependency in face recognition: An initial study. In International Conference on Audio-and Video-Based Biometric Person Authentication, pages 44–51, 2003. 6
[9] A. Founds, N. Orlans, G. Whiddon, and C. Watson. NIST
special database 32-multiple encounter dataset II (MEDSII).
National Institute of Standards and Technology, Tech. Rep,
2011. 6
[10] W. Gao, B. Cao, S. Shan, X. Chen, D. Zhou, X. Zhang,
and D. Zhao. The CAS-PEAL large-scale chinese face
database and baseline evaluations. IEEE Transactions on
Systems, Man, and Cybernetics - Part A: Systems and Humans, 38(1):149–161, January 2008. 5, 6
[11] M. Grgic, K. Delac, and S. Grgic. Scface - surveillance cameras face database. Multimedia Tools Application,
51(3):863–879, February 2011. 5, 6
[12] R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker.
Multi-PIE. Image Vision Computing, 28(5):807–813, May
2010. 5
[13] S. Gu, W. Zuo, Q. Xie, D. Meng, X. Feng, and L. Zhang.
Convolutional sparse coding for image super-resolution. In
IEEE International Conference on Computer Vision, December 2015. 2, 5, 7, 8
[14] M. Jian and K. M. Lam. Simultaneous hallucination and
recognition of low-resolution faces based on singular value
decomposition. IEEE Transactions on Circuits and Systems
for Video Technology, 25(11):1761–1772, November 2015.
2
[15] K. I. Kim and Y. Kwon. Single-image super-resolution using
sparse regression and natural image prior. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 32(6):1127–
1133, June 2010. 5, 7, 8
[16] C. Ledig, L. Theis, F. Huszar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, and
W. Shi. Photo-realistic single image super-resolution using
a generative adversarial network. In IEEE Conference on
Computer Vision and Pattern Recognition, 2017. 2
[17] H. Lee, A. Battle, R. Raina, and A. Y. Ng. Efficient sparse
coding algorithms. In Advances in Neural Information Processing Systems, pages 801–808. 2007. 4
[18] A. Mittal, A. K. Moorthy, and A. C. Bovik. No-reference
image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21(12):4695–4708, 2012. 7,
8
[19] S. P. Mudunuri and S. Biswas. Low resolution face recognition across variations in pose and illumination. IEEE
Transactions on Pattern Analysis and Machine Intelligence,
38(5):1034–1040, 2016. 2
[20] J. Ngiam, Z. Chen, S. A. Bhaskar, P. W. Koh, and A. Y. Ng.
Sparse filtering. In Advances in Neural Information Processing Systems, pages 1125–1133. 2011. 2
[21] T. Peleg and M. Elad. A statistical prediction model based
on sparse representations for single image super-resolution.
IEEE Transactions on Image Processing, 23(6):2569–2582,
June 2014. 2, 5, 7, 8
[22] G. Polatkan, M. Zhou, L. Carin, D. Blei, and I. Daubechies.
A bayesian nonparametric approach to image superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(2):346–358, 2015. 2
[23] R. Rubinstein, M. Zibulevsky, and M. Elad. Double sparsity:
Learning sparse dictionaries for sparse signal approximation.
IEEE Transactions on Signal Processing, 58(3):1553–1564,
March 2010. 2
[24] S. Tariyal, A. Majumdar, R. Singh, and M. Vatsa. Greedy
deep dictionary learning. CoRR, abs/1602.00203, 2016. 2
[25] J. J. Thiagarajan, K. N. Ramamurthy, and A. Spanias. Multilevel dictionary learning for sparse representation of images.
In 2011 Digital Signal Processing and Signal Processing Education Meeting (DSP/SPE), pages 271–276, 2011. 2
[26] T. Tong, G. Li, X. Liu, and Q. Gao. Image super-resolution
using dense skip connections. In IEEE International Conference on Computer Vision, 2017. 2
[27] P. Viola and M. J. Jones. Robust real-time face detection. International Journal Computer Vision, 57(2):137–154, 2004.
6
[28] N. Wang, D. Tao, X. Gao, X. Li, and J. Li. A comprehensive survey to face hallucination. International journal of
computer vision, 106(1):9–30, 2014. 2
[29] S. Wang, L. Zhang, Y. Liang, and Q. Pan. Semi-coupled dictionary learning with applications to image super-resolution
and photo-sketch synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 2216–2223,
2012. 2
[30] Z. Wang, S. Chang, Y. Yang, D. Liu, and T. S. Huang. Studying very low resolution recognition using deep networks.
In The IEEE Conference on Computer Vision and Pattern
Recognition, June 2016. 1, 2
[31] Z. Wang, Z. Miao, Q. M. Jonathan Wu, Y. Wan, and Z. Tang.
Low-resolution face recognition: a review. The Visual Computer, 30(4):359–386, 2013. 1, 2
[32] J. Yang, J. Wright, T. S. Huang, and Y. Ma. Image superresolution via sparse representation. IEEE Transactions on
Image Processing, 19(11):2861–2873, November 2010. 2, 5,
7, 8
[33] M. C. Yang, C. P. Wei, Y. R. Yeh, and Y. C. F. Wang. Recognition at a long distance: Very low resolution face recognition and hallucination. In International Conference on Biometrics, pages 237–242, May 2015. 2
| 1cs.CV
|
TTP: Tool for Tumor Progression
Johannes G. Reiter1 , Ivana Bozic2,3 , Krishnendu Chatterjee1 , Martin A.
Nowak2,3,4
arXiv:1303.5251v1 [q-bio.PE] 21 Mar 2013
1
IST Austria (Institute of Science and Technology Austria), Klosterneuburg, Austria
2
Program for Evolutionary Dynamics, Harvard University, Cambridge, USA
3
Department of Mathematics, Harvard University, Cambridge, USA
4
Department of Organismic and Evolutionary Biology, Harvard University,
Cambridge, USA
Abstract. In this work we present a flexible tool for tumor progression,
which simulates the evolutionary dynamics of cancer. Tumor progression
implements a multi-type branching process where the key parameters
are the fitness landscape, the mutation rate, and the average time of
cell division. The fitness of a cancer cell depends on the mutations it
has accumulated. The input to our tool could be any fitness landscape,
mutation rate, and cell division time, and the tool produces the growth
dynamics and all relevant statistics.
1
Introduction
Cancer is a genetic disease which is driven by the somatic evolution of cells [1],
where driver mutations for cancer increase the reproductive rate of cells through
different mechanisms, e.g. evading growth suppressors, sustaining proliferative
signaling, or resisting cell death [2]. Tumors are initiated by some genetic event
which increases the reproductive rate of previously normal cells. The evolution
of cancer (malignant tumor) is a multi-step process where cells need to receive
several mutations subsequently [3]. This phase of tumor progression is characterized by the uncontrolled growth of cells [2,4]. The requirement to accumulate
multiple mutations over time explains the increased risk of cancer with age.
There are several mathematical models to explain tumor progression and the
age incidence of cancer [5,6,7]. The models have also provided quantitative insights in the evolution of resistance to cancer therapy [8]. The models for tumor
progression are multi-type branching processes which represent an exponentially
growing heterogeneous population of cells, where the key parameters for the process are: (i) the fitness landscape of the cells (which determine the reproductive
rate), (ii) the mutation rate (which determines the accumulation of driver mutations), and (iii) the average cell division time (or the generation time for new
cells). The fitness landscapes allow the analysis of the effects of interdependent
(driver) mutations on the evolution of cancer [9].
In this work, we present a very flexible tool (namely, TTP, tool for tumor
progression) to study the dynamics of tumor progression. The input to our tool
could be any fitness landscape, mutation rate, and cell division time, and the tool
generates the growth dynamics and all relevant statistics (such as the expected
tumor detection time, or the expected appearance time of surviving mutants,
etc). Our stochastic computer simulation is an efficient simulation of a multitype branching process under all possible fitness landscapes, driver mutation
rates, and cell division times.
Our tool provides a quantitative framework to study the dynamics of tumor
progression in different stages of tumor growth. Currently, the data to understand the effects of complex fitness landscapes can only be obtained from patients
or animals suffering the disease. With our tool, playing with the parameters,
once the real-world data is reproduced, the computer simulations can provide
many simulation examples that would aid to understand these complex effects.
Moreover, once the correct mathematical models for specific types of cancer are
identified where the simulations match the real-world data, verification tools for
probabilistic systems can be used to further analyze and understand the tumor
progression process (such an approach has been followed in [10] for the verification of biological models). In this direction, results of specific fitness landscapes
of our tool have already been used in a biological application paper [11]. While
we present our tool for the discrete-time process (which provides a good approximation of the continuous-time process), results of our tool for the special case of
a uniform fitness landscape in the continuous-time process have also been shown
to have excellent agreement with the real-world data for the time to treatment
failure for colorectal cancer [8].
2
Model
Tumor progression is modeled as a discrete-time branching process (GaltonWatson process [12]). At each time step, a cell can either divide or die. The
phenotype i of a cancerous cell determines its division probability bi and is
encoded as a bit string of length four (i.e. {0, 1}4 ). The death probability di
follows from bi as di = 1 − bi . If a cell divides, one of the two daughter cells can
receive an additional mutation (i.e., a bit flips from wildtype 0 to the mutated
type 1) with probability u in one of the wildtype positions (e.g., cells of phenotype
1010 can receive an additional mutation only at positions two and four; cells
of phenotype 1111 can not receive any additional mutations). The branching
process is initiated by a single cell of phenotype i = 0000 (resident cell). The
resident cells are wildtype at all four positions and have a strictly positive growth
rate (i.e., b0000 − d0000 > 0).
Fitness landscapes. Our tool provides two predefined fitness landscapes for
driver mutations in tumor progression: (1) Multiplicative Fitness Landscape
(Mfl) and (2) Path Fitness Landscape (Pfl). Additionally, the user can also
define its own general fitness landscape (Gfl). A fitness landscape defines the
birth probability bi for all possible phenotypes i. Following the convention of the
standard modeling approaches, we let b0000 =1/2(1 + s0 ) be the birth probability
of the resident cells (i.e., cells of phenotype 0000) [9,11]. The growth coefficient
sj indicates the selective advantage provided by an additional mutation at position j in the phenotype.
Multiplicative fitness landscape. In the Mfl a mutation at position j of the
phenotype i of a cell results in a multiplication of its birth probability by (1+sj ).
Specifically, the birth probability bi of a cell with phenotype i is given by
bi =
4
Y
1
(1 + s0 )
(1 + sbj );
2
j=1
where sbj = 0 if the j-th position of i is 0; otherwise sbj = sj . Hence, each
additional mutation can be weighted differently and provides a predefined effect
(1+s1 ), (1+s2 ), (1+s3 ), or (1+s4 ) on the birth probability of a cell. Additional
mutations can also be costly (or neutral) which can be modeled by a negative sj
(or sj = 0). If s0 = s1 = s2 = s3 = s4 the fitness landscape reduces to the model
studied by Bozic et al. [9], which we call Emfl (Equal Multiplicative Fitness
Landscape) and is also predefined in our tool.
Path fitness landscape. The Pfl defines a certain path on which additional mutations need to occur to increase the birth probability of a cell. The predefined
path can be 0000 → 1000 → 1100 → 1110 → 1111, and again the growth coefficients sj determine the multiplicative effect of the new mutation on the birth
probability (see appendix for more details). Mutations not on this path are deleterious for the growth rate of a cell and its birth probability is set to 1/2(1 − v).
The parameter v (0 ≤ v ≤ 1) specifies the disadvantage for cells of all phenotypes
which do not belong to the given path.
General fitness landscapes. Our tool allows to input any fitness landscape as
follows: for bi for i ∈ {0, 1}4 , our tool can take as input the value of bi . In this
way, any fitness landscape can be a parameter to the tool.
Density limitation. In some situations, a tumor needs to overcome current
geometric or metabolic constraints (e.g. when the tumor needs to develop blood
vessels to provide enough oxygen and nutrients for further growth [13,11]). Such
growth limitations are modeled by a density limit (carrying capacity) for various phenotypes. Hence, the cells of a phenotype i grow first exponentially but
eventually reach a steady state around a given carrying capacity Ki . Only cells
with another phenotype (additional mutation) can overcome the density limit.
Logistic growth is modeled with variable growth coefficients sej = sj (1 − Xi /Ki )
where Xi is the current number of cells of phenotype i in the tumor. In this
model, initially sej ≈ sj (Xi Ki ), however, if Xi is on the order of Ki , sej
becomes approximately zero (details are given in the appendix).
3
Tool Implementation & Experimental Results
Our tool provides an efficient implementation of a very general tumor progression
model. Essentially, the tool implements the above defined branching processes
to simulate the dynamics of tumor growth and to obtain statistics about the
expected tumor detection time and the appearance of additional driver mutations
during different stages of disease progression. TTP can be downloaded from
here: http://pub.ist.ac.at/ttp.
For an efficient processing of the discrete-time branching process, the stochastic simulation samples from a multinomial distribution for each phenotype at
each time step [9,11]. The sample returns the number of cells which divided
with and without mutation and the number of cells which died in the current
generation (see the appendix for details). From the samples for each phenotype
the program calculates the phenotype distribution in the next generation. Hence,
the program needs to store only the number of cells of each phenotype during
the simulation. This efficient implementation of the branching process allows
the tool to simulate many patients within a second and to obtain very good
statistical results in a reasonable time frame.
a
b
0000
1000
1001
1011
108
106
0001
0011
1010
1011
0010
0101
1100
1110
1010
0100
0110
0111
1111
Number of cells
Number of cells
1010
104
102
1
0
10
20
106
104
102
1
30
tumor detection size
(109 cells ≈ 1cm3 )
108
0
10
Time (years)
20
30
Time (years)
c
d
1010
Probability density
Number of cells
Emfl
108
106
104
102
1
0
10
20
Time (years)
30
0.2
Mfl
Path
0.1
0
0
10
20
30
Time (years)
Fig. 1. Experimental results illustrating the variability of tumor progression. In panels a and b we show examples for two particular simulation runs where the cells grow
according to the Emfl and resident cells (blue) are constrained by a carrying capacity
of K0000 = 106 . In panel c the cells grow according to the Pfl. In panel d we show
statistical results for the probability density of tumor detection when cells grow according to different fitness landscapes. Parameter values: growth coefficients s0 = 0.004,
s1 = 0.006, s2 = 0.008 , s3 = 0.01, s4 = 0.012, and v = 0.01, mutation rate u = 10−6 ,
cell division time T = 3 days, tumor detection size 109 cells.
Modes. The tool can run in the following two modes: individual or statistics.
In the individual mode the tool produces the growth dynamics of one tumor in a
patient (see panels a, b, and c in Fig. 1). Furthermore, both the growth dynamics
and the phenotype distribution of the tumor are depicted graphically. In the
statistics mode the tool produces the probability distribution for the detection
time of the tumor (see panel d in Fig. 1) both graphically and quantitatively.
Additionally, the tool calculates for all phenotypes the appearance times of the
first surviving lineage, the existence probability, and the average number of cells
at detection time.
Features. TTP provides an intuitive graphical user interface to enter the parameters of the model and shows plots of the dynamics during tumor progression,
the phenotype distribution, or the probability density of tumor detection. These
plots can also be saved as files in various image formats. Furthermore, the tool
can create data files (tab-separated values) of the tumor growth history and
the probability distribution of tumor detection for any set of input parameters
(details on the format are given in the appendix).
Input parameters. In both modes, the tool takes the following input parameters: (i) growth coefficients s0 , s1 , s2 , s3 , and s4 (and v in the case of Pfl),
(ii) mutation rate u, (iii) cell generation time T , (iv) fitness landscape (Mfl,
Pfl, Emfl, or Gfl with the birth probability for each phenotype), and optional
(v) density limits for some phenotypes. In the individual mode, additionally, the
user needs to provide the number of generations which have to be simulated. In
the statistics mode, the additional parameters are: the tumor detection size and
the number of patients (tumors which survive the initial stochastic fluctuations)
which have to be simulated.
Experimental results. In panels a, b, and c of Fig. 1 we show examples of the
growth dynamics during tumor progression. Although we used exactly the same
parameters in panels a and b, we observe that the time from tumor initiation
until detection can be very different. In panel d we show the probability density
of tumor detection under various fitness landscapes. Further experimental results
are given in the appendix.
Case studies. Several results of these models have shown excellent agreement
with different aspects of real-world data. In [9], results for the expected tumor
size at detection time using a Emfl fit the reported polyp sizes of the patients
very well. Similarly, using a continuous-time branching process and a uniform
fitness landscape, results for the expected time to the relapse of a tumor after
start of treatment agree thoroughly with the observed times in patients [8].
Future work. In some ongoing work, we also investigate mathematical models
for tumor dynamics occurring during cancer treatment modeled by a continuoustime branching process. Thus an interesting extension of our tool would be to
model treatment as well. Another interesting direction is to model the seeding of
metastasis during tumor progression and hence simulate a “full” patient rather
than the primary tumor alone. Once faithful models of the evolution of cancer
have been identified, verification tools such as PRISM [14] and theoretical results
such as [15] might contribute to the understanding of these processes.
Acknowledgments. This work is supported by the ERC Start grant (279307:
Graph Games), the FWF NFN Grant (No S11407-N23, Rise), the FWF Grant
(No P 23499-N23), a Microsoft Faculty Fellow Award, the Foundational Questions in Evolutionary Biology initiative of the John Templeton Foundation, and
the NSF/NIH joint program in mathematical biology (NIH grant R01GM078986).
References
1. Vogelstein, B., Kinzler, K.W.: Cancer genes and the pathways they control. Nature
medicine 10(8) (2004) 789–799
2. Hanahan, D., Weinberg, R.a.: Hallmarks of cancer: the next generation. Cell
144(5) (2011) 646–674
3. Jones, S., Chen, W.D., Parmigiani, G., Diehl, F., Beerenwinkel, N., Antal, T.,
Traulsen, A., Nowak, M.A., Siegel, C., Velculescu, V.E., Kinzler, K.W., Vogelstein,
B., Willis, J., Markowitz, S.D.: Comparative lesion sequencing provides insights
into tumor evolution. PNAS 105(11) (2008) 4283–4288
4. Nowak, M.A.: Evolutionary Dynamics: Exploring the equations of life. The Belknap Press of Harvard University Press, Cambridge, MA (2006)
5. Komarova, N.L., Sengupta, A., Nowak, M.A.: Mutation-selection networks of cancer initiation: tumor suppressor genes and chromosomal instability. Journal of
Theoretical Biology 223(4) (2003) 433–450
6. Iwasa, Y., Michor, F., Nowak, M.A.: Stochastic tunnels in evolutionary dynamics.
Genetics 166(3) (2004) 1571–1579
7. Nowak, M.A., Michor, F., Komarova, N.L., Iwasa, Y.: Evolutionary dynamics of
tumor suppressor gene inactivation. PNAS 101(29) (2004) 10635–10638
8. Diaz, L.A., Williams, R.T., Wu, J., Kinde, I., Hecht, J.R., Berlin, J., Allen, B.,
Bozic, I., Reiter, J.G., Nowak, M.A., Kinzler, K.W., Oliner, K.S., Vogelstein, B.:
The molecular evolution of acquired resistance to targeted EGFR blockade in colorectal cancers. Nature 486(7404) (2012) 537–540
9. Bozic, I., Antal, T., Ohtsuki, H., Carter, H., Kim, D., Chen, S., Karchin, R.,
Kinzler, K.W., Vogelstein, B., Nowak, M.A.: Accumulation of driver and passenger
mutations during tumor progression. PNAS 107(43) (2010) 18545–18550
10. Sadot, A., Fisher, J., Barak, D., Admanit, Y., Stern, M., Hubbard, E., Harel, D.:
Toward verified biological models. Computational Biology and Bioinformatics,
IEEE/ACM Transactions on 5(2) (2008) 223–234
11. Reiter, J.G., Bozic, I., Allen, B., Chatterjee, K., Nowak, M.A.: The effect of one
additional driver mutation on tumor progression. Evolutionary Applications 6(1)
(2012) 34–45
12. Haccou, P., Jagers, P., Vatutin, V.A.: Branching Processes: Variation, Growth,
and Extinction of Populations. Cambridge University Press (2005)
13. Kerbel, R.S.: Tumor angiogenesis: past, present and the near future. Carcinogenesis
21(3) (March 2000) 505–15
14. Hinton, A., Kwiatkowska, M.Z., Norman, G., Parker, D.: Prism: A tool for automatic verification of probabilistic systems. In: TACAS. (2006) 441–444
15. Etessami, K., Stewart, A., Yannakakis, M.: Polynomial time algorithms for multitype branching processesand stochastic context-free grammars. In: STOC. (2012)
579–588
A
Appendix: Details of the Tool
TTP is available for download at http://pub.ist.ac.at/ttp. The tool is implemented in Java and runs on all operating systems which run a Java Virtual
Machine (JVM) of version 1.7 or above (www.java.com/getjava). All the necessary libraries are included in the jar-file.
A.1
Tool Features
Our tool supports various features in two running modes.
In the individual mode, TTP simulates the tumor growth dynamics for a
given number of generations. Plots of the growth dynamics over time and the
current phenotype distribution are produced simultaneously. (Both plots can be
saved in a PNG-file or SVG-file.) The full growth history for all cell types can
also be stored in a data-file (format is described in section A.2).
In the statistics mode, TTP simulates the given number of patients with
the same parameters and simultaneously shows the probability density of tumor
detection for a given detection size (109 cells correspond to a tumor volume of
approximately 1cm3 ). The average tumor detection time and the average fraction
of resident cells at detection are also shown during the simulations. After all
patients have been simulated, the existence probability at detection, the average
number of cells, and the average appearance year of the first surviving cell for
all phenotypes are calculated and shown in a new window. In addition the tool
shows the number of detected and died tumors per year in a separate window.
All these data is stored in a data-file (format is described in section A.2).
A.2
Installation and Implementation Details
TTP is written in Java and makes use of several other libraries. The tool requires
the Java Runtime Environment (JRE) of version 1.7 or above. To start TTP
double-click on ttp.jar or on the command line type: java -jar ttp.jar.
(Make sure that you have the permission to execute ttp.jar. On Mac OS invoking the tool from the command line can overcome the security restrictions.)
The tool is composed of the following components: the model implementation,
the statistics thread, the graphical user interface, and the plot generator.
Model implementation. The core component of the tool is the efficient implementation of the discrete-time branching process. Following Bozic et al. [9],
the number of cells Xi of phenotype i in the next generation (t + 1) is calculated
by sampling from the multinomial distribution
Prob[(Y1 , Y2 , Y3 ) = (y1 , y2 , y3 )] =
where y1 + y2 + y3 = Xi (t), y3 =
P
k
Xi (t)!
[bi (1 − u)y1 dyi 2 (bi u)y3 ]
y1 !y2 !y3 !
(1)
Mik , and
Xi (t + 1) = Xi (t) + Y1 − Y2 +
X
k
Mki .
(2)
The number of cells which give birth to an identical daughter cell is denoted
by Y1 , the number of cells which die is denoted by Y2 . The number of cells
which divide with an additional mutation is given by Y3 and the number of
cells mutated from phenotype k to i is given by Mki . In general, one can define
a mutation matrix to encode the probabilities mki that a cell of phenotype
k mutates to a cell of phenotype i. In our case this matrix is defined by the
sequential accumulation of mutations. A cell of some phenotype can receive an
additional mutation only on the positions in its bit-string encoding which are
wildtype (i.e. only bit flips from 0 to 1 are allowed). Mutations on all allowed
positions are equally likely. Back mutations are not considered.
Fitness landscapes. Our tool supports four fitness landscapes for additional driver
and passenger mutations: (i) Mfl, (ii) Emfl, (iii) Pfl, and (iv) Gfl. In principal, driver mutations increase the birth rate of a cell whereas passenger mutations
have no effect on the cell’s birth rate [9]. In the tables A1 and A2 we present
the complete definition of the Mfl and the Pfl, respectively. (The definition of
Emfl and Gfl have been given in Section 2.)
Table A1. Multiplicative Fitness Landscape.
Additional mutations
0
1
2
3
4
Phenotype
Birth probability
1/2(1 + s )
0000
0
1/2(1 + s )(1 + s )
1000
0
1
1/2(1 + s )(1 + s )
0100
0
2
1/2(1 + s )(1 + s )
0010
0
3
1/2(1 + s )(1 + s )
0001
0
4
1/2(1 + s )(1 + s )(1 + s )
1100
0
1
2
1/2(1 + s )(1 + s )(1 + s )
1010
0
1
3
1/2(1 + s )(1 + s )(1 + s )
1001
0
1
4
1/2(1 + s )(1 + s )(1 + s )
0110
0
2
3
1/2(1 + s )(1 + s )(1 + s )
0101
0
2
4
1/2(1 + s )(1 + s )(1 + s )
0011
0
3
4
1/2(1 + s )(1 + s )(1 + s )(1 + s )
1110
0
1
2
3
1/2(1 + s )(1 + s )(1 + s )(1 + s )
1101
0
1
2
4
1/2(1 + s )(1 + s )(1 + s )(1 + s )
1011
0
1
3
4
1/2(1 + s )(1 + s )(1 + s )(1 + s )
0111
0
1
3
4
1/2(1 + s)(1 + s )(1 + s )(1 + s )(1 + s )
1111
1
2
3
4
Density limit. Our tool allows a separate carrying capacity Ki for each phenotype i. When the Gfl is used, in the beginning of the simulation, the growth
coefficients si are calculated from the given bi for all phenotypes i ∈ {0, 1}4 since
the density limiting effects are based on the values for s˜i . As a technical detail,
for sizes Xi for which the birth probability bi would fall below 0 (or for which,
equivalently, s˜i would fall below -1), we set bi = 0.
Statistics thread. The statistics thread handles the simulation of many identical branching processes to obtain the statistical results. These simulations run
Table A2. Path Fitness Landscape.
Additional mutations
0
1
2
3
4
Phenotype
Birth probability
1/2(1 + s)
0000
1
1000
/2(1 + s0 )(1 + s1 )
1/2(1 − v)
0100
1/2(1 − v)
0010
1/2(1 − v)
0001
1
1100
/2(1 + s0 )(1 + s1 )(1 + s2 )
1/2(1 − v)
1010
1/2(1 − v)
1001
1/2(1 − v)
0110
1/2(1 − v)
0101
1/2(1 − v)
0011
1/2(1 + s )(1 + s )(1 + s )(1 + s )
1110
0
1
2
3
1/2(1 − v)
1101
1/2(1 − v)
1011
1/2(1 − v)
0111
1/2(1 + s )(1 + s )(1 + s )(1 + s )(1 + s )
1111
0
1
2
3
4
in a separate thread such that the GUI keeps responsive for user requests. After
completing all the necessary simulations a data-file with all relevant results is
automatically generated and stored to the execution directory of the tool.
Graphical user interface. The graphical user interface (GUI) component contains frames and forms required for the functionality of the tool. It also handles
all the user requests and distributes them to the other components. Within the
GUI the plots for the tumor progression dynamics (in individual mode) and for
the probability density of tumor detection (in statistics mode) are displayed.
Multiple screenshots of the GUI are shown in section A.4.
Plot generator. The plot generation is based on the free JFreeChart library
(www.jfree.org). For the generation of the Scalable Vector Graphics (SVG) the
Apache XML Graphics library (http://xmlgraphics.apache.org) is used. An
example for a plot generated by our tool is shown in Figure A2.
Data files. TTP produces various data files which can be used for further
analysis and processing. The data are given as tab-separated values where each
record is one line of the text files. In Listing A1 we show an example of a data file
generated in the statistics mode. Average results are given as comments which
start with a hash. In the individual mode the data file contains the number of
cells Xi for each phenotype i in all generations.
Fig. A2. Example for a generated plot of the tumor growth dynamics.
Listing A1. Generated data file in the statistics mode.
# Used f i t n e s s l a n d s c a p e : M u l t i p l i c a t i v e
# Growth c o e f f i c i e n t s : s0 =0.004 , s1 =0.006 , s2 =0.008 , s3 =0.01 , s4 =0.012
# Mutation r a t e : 1E−6
# Generation time ( days ) : 3 . 0
# S i m u l a t i o n f o r 10000 p a t i e n t s .
# Bin s i z e i s 20 g e n e r a t i o n s .
died
cumul died
generation
detected
cumul det
0
0
0
0.046003
0.920069
20
0
0
0.001938
0.958826
40
0
0
0.000705
0.972923
60
0
0
0.000358
0.98009
80
0
0
0.000227
0.984622
100
0
0
0.000149
0.987594
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5980
0
1
0
1
6000
0
1
0
1
# 0.7981% ( a b s : 10000) have been d e t e c t e d w i t h i n a t most
#
20000 g e n e r a t i o n s .
# 99.2% ( a b s : 1243006) tumors went e x t i n c t .
# Average y e a r o f d e t e c t i o n : 2 2 . 0 4
# Mutant appearance t i m e s :
# Mutant 0000 appeared i n a v e r a g e i n g e n e r a t i o n : 0 ( y e a r : 0 ) ,
#
e x i s t a n c e p r o b a b i l i t y : 100%,
#
a v e r a g e number o f c e l l s : 3267013.74
# Mutant 0001 appeared i n a v e r a g e i n g e n e r a t i o n : 1540 ( y e a r : 1 2 . 6 6 ) ,
#
e x i s t a n c e p r o b a b i l i t y : 100%,
#
a v e r a g e number o f c e l l s : 1925660.02
.
.
.
.
.
.
.
.
.
.
.
.
# 1253006 runs have been performed t o c r e a t e 10000 p a t i e n t s .
A.3
User Manual
TTP is invoked by a double-click on ttp.jar or by the command java -jar
ttp.jar. After the tool has started, the GUI can be used for all operations (see
Figure A3 for a screenshot of the GUI).
Input parameters. In the control panel the tool takes all the main parameters
for tumor progression: the fitness landscape, the mutation rate, and the cell
division time. If one of the prespecified fitness landscapes (Mfl, Pfl, or Emfl)
is used, the relevant growth coefficients have to be defined. If the general fitness
landscape is used, a pop-up window appears after the selection of the Gfl and
the specific birth probabilities for all the phenotypes can be defined. To add
density limits for specific phenotypes, a pop-up window appears after “density
limit” has been checked. For each phenotype a different density limit can be
given (−1 indicates that there is no limitation on this phenotype). To obtain
statistical results, the number of patients (i.e. the number of tumors with a
surviving lineage) and the tumor detection size (i.e. the number of cells when a
tumor can be detected) need to be provided.
Modes. After all the parameter values have been specified, the tool can either
run in the individual or the statistics mode. To simulate the growth dynamics
of a single tumor, click on “New simulation” and the tool runs in the individual
mode. Then any number of cell generations can be simulated until the tumor
consists of more than 1015 cells.
The statistics mode can be started by clicking on “Obtain statistics”. The
tool simulates the given number of tumors until they reach the detection size
and calculates all relevant statistics.
Output. In the individual mode, the tool generates plots for the growth dynamics and the phenotype distribution during the simulation. Furthermore, the
entire tumor growth dynamics for each phenotype can be stored as a data file
and the plots can be saved as PNG and SVG files (plots are stored to the folder
“charts” in the execution directory of the tool).
In the statistics mode, the tool generates the plot for the probability density
of tumor detection. Statistics about the appearance time of the mutants and the
detection and extinction year are shown in separate windows (see Figures A4
and A6 for screenshots). All the generated statistics are automatically saved to
a data file (see Listing A1).
A.4
Experimental Results & Screenshots
In this section we present some additional experimental results and multiple
screenshots of the tool. In Table A3 we compare the probability of tumor detection for the fitness landscapes Emfl, Mfl, and Pfl. In average our tool needs
approximately 90ms to simulate a tumor with 109 cells (on a dual core 2.67GHz
processor).
Table A3. Cumulative probability of tumor detection for different fitness landscapes.
Results are averages over 105 runs. Parameter values: growth coefficients s0 = 0.004,
s1 = 0.006, s2 = 0.008, s3 = 0.01, s4 = 0.012, v = 0.006, mutation rate u = 10−6 ,
detection size M = 109 .
generation
2000
2200
2400
2600
2800
3000
3200
3400
3600
3800
4000
4200
4400
4600
4800
5000
Emfl
0.0
0.0
0.0
0.004
0.068
0.289
0.567
0.774
0.892
0.949
0.977
0.989
0.995
0.998
0.999
1.0
Mfl
0.001
0.038
0.212
0.491
0.724
0.865
0.937
0.971
0.987
0.994
0.998
0.999
1.0
1.0
1.0
1.0
Pfl
0.0
0.0
0.006
0.075
0.288
0.559
0.766
0.887
0.947
0.976
0.989
0.995
0.998
0.999
1.0
1.0
Fig. A3. Graphical User Interface of TTP in the individual mode.
Fig. A4. Graphical User Interface of TTP in the statistics mode.
Fig. A5. Statistical results of the average detection and extinction year.
Fig. A6. Statistical results of the average appearance year, the existence probability
and the number of cells at detection time.
| 5cs.CE
|
COMPOSITION OF GRAY ISOMETRIES
SIERRA MARIE M. LAURESTA and Virgilio P. Sison
Institute of Mathematical Sciences and Physics
University of the Philippines Los Baños
College, Laguna 4031
E-mail: [email protected], [email protected]
Abstract
In coding theory, Gray isometries are usually defined as mappings
between finite Frobenius rings, which include the ring ℤ𝑚 of integers
modulo m and the finite fields. In this paper, we derive an isometric
mapping from ℤ8 to ℤ24 from the composition of the Gray isometries on
ℤ8 and on ℤ24 . The image under this composition of a ℤ8 -linear block
code of length n with homogeneous distance d is a (not necessarily
linear) quaternary block code of length 2n with Lee distance d.
Introduction
A (block) code of length n over a ring R is a non-empty set of n-tuples over R called
codewords. It is said to be linear of rate-k/n if it is a (not necessarily free) submodule, and is
completely determined by a 𝑘 × 𝑛 matrix G over R.
Codes over rings gained more attention when Hammons, Kumar, Calderbank, Sloane and Solé
[1] discovered in 1994 that certain very good but peculiar nonlinear codes over the binary field
𝔽2 can be viewed as images of linear codes over the integer ring ℤ4 under the so-called Gray
map 𝜙 from ℤ4 onto 𝔽22 defined by 0 ⟼ (0,0), 1 ⟼ (0,1), 2 ⟼ (1,1) and 3 ⟼ (1,0). The
map 𝜙 is an isometry or is weight-preserving, that is, the Lee weight of an element of ℤ4 is
equal to the Hamming weight of its image under 𝜙. The Hamming weight of a binary vector
is the number of nonzero components in the vector.
Carlet [2] introduced a generalization of 𝜙 to the ring ℤ2𝑘 of integers modulo 2𝑘 . Let k be a
positive integer, u an element of ℤ2𝑘 and 𝑢 = ∑𝑘𝑖=1 2𝑖−1 𝑢𝑖 its 2-adic expansion, where 𝑢𝑖 ∈ 𝔽2 .
The image of u by the generalized Gray map is the boolean function on 𝔽𝑘−1
given by
2
(𝑦1 , 𝑦2 , … , 𝑦𝑘−1 ) ⟶ 𝑢𝑘 + ∑𝑘−1
1=1 𝑢𝑖 𝑦𝑖 .
We identify this boolean function with a binary word of length 2𝑘−1 by simply listing its values.
𝑘−1
Thus the generalized Gray map is seen as a nonsurjective mapping from ℤ2𝑘 to 𝔽22 , and
its image is the Reed-Muller code of order 1, RM(1, k-1). The generalized Gray map is naturally
extended to the n-tuples.
When k = 2, RM(1,1) is the set of boolean functions on 𝔽2 , and we obtain the usual Gray map 𝜙
from ℤ4 to 𝔽22 . When k = 3, the generalized Gray map (which we denote by 𝜓) takes ℤ8 onto
RM(1,2) which is the set of boolean functions on 𝔽22 that give all the binary words in 𝔽42 with
even Hamming weight.
1
Methodology
We extend the usual Gray isometry 𝜙 as a bijective mapping from ℤ24 onto 𝔽42 . Table 1 shows
the binary image of an element of ℤ24 under 𝜙. Clearly the Lee weight 𝑤𝐿 of an element of ℤ24 is
equal to the Hamming weight 𝑤𝐻 of its binary image.
(𝑥, 𝑦) ∈ ℤ24
(0,0)
(0,1)
(0,2)
(0,3)
(1,0)
(1,1)
(1,2)
(1,3)
(2,0)
(2,1)
(2,2)
(2,3)
(3,0)
(3,1)
(3,2)
(3,3)
𝑤𝐿 (𝑥, 𝑦)
0
1
2
1
1
2
3
2
2
3
4
3
1
2
3
2
𝜙(𝑥, 𝑦) ∈ 𝔽42
(0,0,0,0)
(0,0,0,1)
(0,0,1,1)
(0,0,1,0)
(0,1,0,0)
(0,1,0,1)
(0,1,1,1)
(0,1,1,0)
(1,1,0,0)
(1,1,0,1)
(1,1,1,1)
(1,1,1,0)
(1,0,0,0)
(1,0,0,1)
(1,0,1,1)
(1,0,1,0)
𝑤𝐻 (𝜙(𝑥, 𝑦))
0
1
2
1
1
2
3
2
2
3
4
3
1
2
3
2
Table 1: The isometric Gray map 𝝓 on ℤ𝟐𝟒
We restrict 𝜙 −1 as a mapping from RM(1,2) to ℤ24 as follows:
𝜙 −1 (𝑥)
(0,0)
(1,1)
(0,2)
(1,3)
(2,2)
(3,3)
(2,0)
(3,1)
𝑥 ∈ 𝑅𝑀(1,2)
(0,0,0,0)
(0,1,0,1)
(0,0,1,1)
(0,1,1,0)
(1,1,1,1)
(1,0,1,0)
(1,1,0,0)
(1,0,0,1)
Table 2: The map 𝝓−𝟏 on 𝑹𝑴(𝟏, 𝟐)
For ℤ8 we apply the following homogeneous weight [3] and extend it coordinatewisely.
0
𝑤ℎ𝑜𝑚 (𝑥) = {4
2
𝑖𝑓 𝑥 = 0
𝑖𝑓 𝑥 = 4
𝑒𝑙𝑠𝑒
Table 3 shows the image of an element of ℤ8 in RM(1,2) under the generalized Gray map 𝜓. If
𝑢 ∈ ℤ8 has the 2-adic expansion 𝑢 = 𝑢1 + 2𝑢2 + 4𝑢3 , then 𝜓(𝑢) = 𝜓(𝑢1 + 2𝑢2 + 4𝑢3 ) =
(𝑢3 , 𝑢3 + 𝑢1 , 𝑢3 + 𝑢2 , 𝑢3 + 𝑢1 + 𝑢2 ). The mapping 𝜓 is weight preserving such that the
homogeneous weight of an element of ℤ8 is equal to the Hamming weight of its image in
RM(1,2).
2
𝑢 ∈ ℤ8
0
1
2
3
4
5
6
7
𝑤ℎ𝑜𝑚 (𝑢)
0
2
2
2
4
2
2
2
𝜓(𝑢) ∈ 𝑅𝑀(1,2)
(0,0,0,0)
(0,1,0,1)
(0,0,1,1)
(0,1,1,0)
(1,1,1,1)
(1,0,1,0)
(1,1,0,0)
(1,0,0,1)
𝑤𝐻 (𝜓(𝑢))
0
2
2
2
4
2
2
2
Table 3: The isometric Gray map 𝝍 on ℤ𝟖
Results and Discussion
We take the composition 𝜙 −1 𝜓: ℤ8 ⟶ ℤ24 . Table 4 shows the quaternary image of an element of
ℤ8 under 𝜙 −1 𝜓. If 𝑢 = 𝑢1 + 2𝑢2 + 4𝑢3 ∈ ℤ8 , then 𝜙 −1 𝜓(𝑢) = 𝜙 −1 𝜓(𝑢1 + 2𝑢2 + 4𝑢3 ) = (𝑢1 +
2𝑢3 , 𝑢1 + 2𝑢3 + 2𝑢2 ).
𝑢 ∈ ℤ8
0
1
2
3
4
5
6
7
𝜙 −1 𝜓(𝑢)
(0,0)
(1,1)
(0,2)
(1,3)
(2,2)
(3,3)
(2,0)
(3,1)
𝑤ℎ𝑜𝑚 (𝑢)
0
2
2
2
4
2
2
2
𝑤𝐿 (𝜙 −1 𝜓(𝑢))
0
2
2
2
4
2
2
2
Table 4: The isometric map 𝝓−𝟏 𝝍: ℤ𝟖 → ℤ𝟐𝟒
The mapping 𝜙 −1 𝜓 is weight preserving such that the homogeneous weight of an element of
ℤ8 is equal to the Lee weight of its image in ℤ24 . It is extended naturally to the n-tuples.
Let C be a linear block code of length n over ℤ8 with minimum homogeneous distance d. The
−1
image of C under 𝜙 −1 𝜓 is the set 𝜙 −1 𝜓(𝐶) = {𝑤 ∈ ℤ2𝑛
4 | 𝑤 = 𝜙 𝜓(𝑐) for 𝑐 ∈ 𝐶}.
Proposition. The set 𝜙 −1 𝜓(𝐶) has the following properties:
i.
ii.
iii.
𝜙 −1 𝜓(𝐶) is a (not necessarily linear) block code of length 2n over ℤ4 .
The Lee distance of 𝜙 −1 𝜓(𝐶) is equal to d.
Every codeword of 𝜙 −1 𝜓(𝐶) has even Lee weight.
To illustrate non-linearity, consider the rate-2/3 linear block code over ℤ8 generated by the
matrix
1 2 7
𝐺=(
)
0 2 4
This code has 32 codewords, minimum Hamming distance 1, and minimum homogeneous
distance 4. The codewords (6,6,6) and (7,6,1) generated by the information words (6,1) and
(7,4), respectively, have quaternary images (2,0,2,0,2,0) and (3,1,2,0,1,1) whose
superimposition is not in the code. This example also shows that 𝜙 −1 𝜓 is not an additive
homomorphism.
3
Conclusion and Recommedation
This paper offers a simple way to define isometric mappings from ℤ2𝑘 to ℤ2𝑟 , in general for
𝑟 < 𝑘, and to take the ℤ2𝑟 -image code of a ℤ2𝑘 -linear block code. Sufficient and necessary
conditions for the linearity of the ℤ2𝑟 -image can be determined. Extension of this construction to
galois rings is inevitable.
References
[1] R. Hammons, Jr., P. V. Kumar, A. R. Calderbank, N. J. A. Sloane, and P. Solé, “The ℤ4 linearity of Kerdock, Preparata, Goethals and related codes,” IEEE Trans. Inform. Theory,
vol. 40, no. 2, pp. 301 - 319, January 1994.
[2] C. Carlet, “ℤ2𝑘 -linear codes,” IEEE Trans. Inform. Theory, vol. 44, no. 4, pp. 1543-1546, July
1998.
[3] M. Greferath and S.E. Schmidt, “Gray isometries for finite chain rings and a nonlinear
ternary (36,312,15) code,” IEEE Trans. Inform.Theory, vol. 45, no. 7, pp. 2522--2524,
November 2001.
4
| 7cs.IT
|
Video Enhancement with Task-Oriented Flow
Tianfan Xue†
Google Research
Baian Chen
MIT CSAIL
Jiajun Wu
MIT CSAIL
Donglai Wei
Harvard University
arXiv:1711.09078v1 [cs.CV] 24 Nov 2017
William T. Freeman
MIT CSAIL, Google Research
Frame interpolation
t=3
t=2
t=1
(I-a) Input Low-frame-rate
Videos
(I-b) Epic Flow
(I-c) Interp. by Epic Flow
(I-d) Task-oriented Flow
(I-e) Interp. by Task-oriented Flow
Video denoising
t=7
t=4
t=1
(II-a) Input Noisy Videos
(II-b) Epic Flow
(II-c) Denoise by Epic Flow
(II-d) Task-oriented Flow
(II-e) Denoise by Task-oriented Flow
Figure 1: Many video processing tasks, e.g., temporal frame-interpolation (top) and video denoising (bottom), rely on flow
estimation. In many cases, however, precise optical flow estimation is intractable and could be suboptimal for a specific
task. For example, although EpicFlow [28] predicts precise movement of objects (I-b, the flow field aligns well with object
boundaries), small errors in estimated flow fields result in obvious artifacts in interpolated frames, like the obscure fingers in
(I-c). With the task-oriented flow proposed in this work (I-d), those interpolation artifacts disappear as in (I-e). Similarly, in
video denoising, our task-oriented flow (II-d) deviates from EpicFlow (II-b), but leads to a cleaner output frame (II-e). Flow
visualization is based on the color wheel shown on the corner of (I-b).
Abstract
Many video processing algorithms rely on optical flow to
register different frames within a sequence. However, a precise estimation of optical flow is often neither tractable nor
optimal for a particular task. In this paper, we propose taskoriented flow (TOFlow), a flow representation tailored for
specific video processing tasks. We design a neural network
with a motion estimation component and a video processing component. These two parts can be jointly trained in a
self-supervised manner to facilitate learning of the proposed
TOFlow. We demonstrate that TOFlow outperforms the traditional optical flow on three different video processing tasks:
frame interpolation, video denoising/deblocking, and video
super-resolution. We also introduce Vimeo-90K, a large-scale,
high-quality video dataset for video processing to better evaluate the proposed algorithm.
† This
work was done when Tianfan Xue was a student in MIT CSAIL.
1. Introduction
Motion estimation is a key component in video processing
tasks like temporal frame interpolation, video denoising, and
video super-resolution. Most motion-based video processing
algorithms use a two-step approach [16, 2, 15]. They first estimate motion between input frames and register them based on
estimated flow fields, and then process the registered frames
to generate the final output. Therefore, the accuracy of flow
estimation greatly affects the performance of these two-step
approaches.
However, precise flow estimation can be challenging and
slow. The brightness constancy assumption, which many motion estimation algorithms rely on, may fail due to variations
in lighting or pose and the presence of motion blur or occlusion. Also, many motion estimation algorithms involve
solving a large-scale optimization problem, making it inefficient for real-time applications. For example, the widely
used EpicFlow algorithm [28] takes about 16 seconds for each
frame of a million pixels.
Moreover, most motion estimation algorithms aim to solve
for a motion field that matches the actual objects in motion.
However, this may not be the best motion representation for
video processing. Figure 1 shows an example in frame interpolation. Even though EpicFlow calculates a precise motion
field (I-b) whose boundary is well-aligned with the fingers
in the image (I-c), the interpolated frame (I-c) based on it
contains obvious artifacts due to occlusion. In contrast, using
the task-oriented flow introduced in this work, the model generates better interpolation result (I-e), though the estimated
motion field (I-d) differs from optical flow in magnitude and
does not align with object boundaries. Similarly, in video
denoising, although EpicFlow shown in (II-b) matches the
boundary of the girl’s hair, the frame denoised with EpicFlow
is much noisier than the one with our task-oriented flow (II-e).
This suggests that for specific video processing tasks, there
exist flow representations that do not match the actual object
movement, but lead to better results.
In this paper, we propose to learn task-oriented flow
(TOFlow) by performing motion analysis and video processing jointly in an end-to-end trainable convolutional network.
Our network consists of three modules: the first one estimates
the motion fields between input frames; the second one registers all input frames based on estimated motion fields; and
the third one generates target output from registered frames.
These three modules are jointly trained to minimize the loss
between output frames and ground truth. Unlike other flow
estimation networks [27, 6], the flow estimation module in our
framework predicts a motion field tailored to a specific task,
e.g., frame interpolation or video denoising, as it is trained together with the corresponding video processing module. The
proposed TOFlow has several advantages: first, it significantly
outperforms optical flow algorithms on three video processing
tasks; second, it is highly efficient, taking only 15.8ms for an
input image with a resolution of 448×256; third, it can be
self-supervised, learning from unlabeled video frames.
To evaluate TOFlow, we build a large-scale, high-quality
video dataset for video processing. Most existing large video
datasets, like Youtube-8M [1], are designed for high-level
vision tasks like event classification. The videos are often of
low resolutions with significant motion blurs, making them
less useful for video processing. To evaluate video processing algorithms systematically, we introduce a new dataset,
Vimeo-90K, which consists of 89,800 high-quality video clips
(i.e., 720p or higher) downloaded from Vimeo∗ . We build
three benchmarks from these videos for interpolation, denoising/deblocking, and super-resolution, respectively. We hope
this dataset will contribute to future research in learning-based
video processing through its high-quality videos and diverse
examples.
The contributions of this paper are three-fold. First, we
propose TOFlow, a flow representation tailored to specific
∗ https://vimeo.com
video processing tasks, significantly outperforming standard
optical flow. Second, we propose a self-supervised and endto-end trainable video processing framework that can handle
various tasks, including frame interpolation, video denoising,
and video super-resolution. Third, we also build a large-scale,
high-quality video dataset, Vimeo-90K, for video processing.
2. Related Work
Optical Flow Estimation Dated back to Horn and
Schunck [9], most optical flow algorithms have sought to
minimize hand-crafted energy terms for image alignment and
flow smoothness [22, 3, 32]. Current state-of-the-art methods
like EpicFlow [28] and DC Flow [34] further exploit image
boundary and segment cues to improve the flow interpolation
among sparse matches. Recently, end-to-end deep learning
methods were proposed for faster inference, some trained
with supervision [6, 27], and some without [35]. In this work,
we used the same flow network as SpyNet [27]. But instead
of training it to minimize the flow estimation error, as SpyNet
does, we train it jointly with a video processing network to
learn a flow representation that is the best for a specific task.
Low-level Video Processing We focus on three video processing tasks: frame interpolation, video denoising, and video
super-resolution. Most existing algorithms in these areas
explicitly estimate the dense motion among input frames,
and then reconstruct the reference frame according to image
formation models for frame interpolation [2, 33, 36], video
super-resolution [17], and denoising [15, 30, 20]. We refer
readers to survey articles [23, 8] for comprehensive literature
reviews on these flourishing research topics.
Deep Learning for Video Enhancement Inspired by the
success of deep learning, researchers have directly modeled
video enhancement tasks as regression problems without representing motions, and have designed deep networks for frame
interpolation [21, 24] and super-resolution [14, 10, 12, 29].
Recently, with differentiable image sampling layers in
deep learning [11], motion information can be incorporated
into networks and trained jointly with the video enhancement
task. Such approaches have been applied to video interpolation [18], light-field video interpolation [31], object novel
view synthesis [37], eye gaze manipulation [7], and superresolution [5, 29]. Although many of these algorithms also
jointly train the flow estimation with the rest parts of network,
there is no systematical study on the advantage of joint training. In this paper, we illustrate the advantage of the trained
task-oriented flow through toy examples, and also demonstrate its superiority over general flow algorithm on various
real-world tasks. We also present a general framework that
can easily adapt to different video processing tasks.
3. Tasks
In the paper, we explore three video enhancement tasks:
frame interpolation, video denoising/deblocking, and video
super-resolution.
Temporal Frame Interpolation Given a low frame rate
video, a temporal frame interpolation algorithm generates
a high frame rate video by synthesizing additional frames
between two temporally neighboring frames. Specifically, let
I1 and I3 be two consecutive frames in an input video, the
task is to estimate the missing middle frame I2 . Temporal
frame interpolation doubles the video frame rate, and can be
recursively applied for even higher frame rates.
Video Denoising/Deblocking Given a degraded video with
artifacts from either the sensor or compression, video denoising/deblocking aims to remove the noise or compression artifacts to recover the original video. This is typically done by
aggregating information from neighboring frames. Specifically, Let I1 , I2 , ..., IN be N-consecutive frames in an input
video, the task of video denoising is to estimate the middle
∗
frame Iref
given degraded frames I1 , I2 , ..., IN as input. For
the ease of description, in the rest of paper, we simply call
both tasks as video denoising.
Video Super-Resolution (SR) Similar to video denoising,
given N consecutive low-resolution frames as input, the task
of video super-resolution is to recover the high-resolution
middle frame. In this work, we first upsample all the input
frames to the same resolution of the output using bicubic
interpolation, and our algorithm only needs to recover the
high-frequency component in the output image.
4. Task-Oriented Flow for Video Processing
Most motion-based video processing algorithms has two
steps: motion estimation and image processing. For example,
in temporal frame interpolation, most algorithms first estimate
how pixels move between input frames (frame 1 and 3), and
then move pixels to the estimated location in the output frame
(frame 2) [2]. Similarly, in video denoising, algorithms first
register different frames based on estimated motion fields
between them, and then remove noises by aggregating information from registered frames.
In this paper, we propose to use task-oriented flow
(TOFlow) to integrate the two steps. To learn the task-oriented
flow, we design an end-to-end trainable network with three
parts (Figure 3): a flow estimation module that estimates the
movement of pixels between input frames; an image transformation module that warps all the frames to a reference frame;
and a task-specific image processing module that performs
video interpolation, denoising, or super-resolution on registered frames. Because the flow estimation module is jointly
trained with the rest of the network, it learns to predict a flow
field that fits to a particular task.
4.1. Toy Example
Before discussing the details of network structure, we first
start with two synthetic sequences to demonstrate why our
TOFlow can outperform traditional optical flows. The left
Video Denoising
Frame Interpolation
?
Input frames
Input frames
Case I: With Ground Truth Flows
GT flow
Warped by Interpolated
GT flow
2nd frame
Case II: With Task-Oriented Flows
TOFlow
Warped by Interpolated
2nd frame
TOFlow
Case I: With Ground Truth Flows
GT flow
Warped by
TOFlow
Denoised
frame
Case II: With Task-Oriented Flows
TOFlow
Warped by
TOFlow
Denoised
frame
Figure 2: A toy example that demonstrates the effectiveness
of task oriented flow over the traditional optical flow. See
Section 4.1 for details.
of Figure 2 shows an example of frame interpolation, where
a green triangle is moving to the bottom in front of a black
background. If we warp both the first and third frames to the
second, even using the ground truth flow (Case I, left column),
there is an obvious doubling artifact in the warped frames due
to occlusion (Case I, middle column), which is a well-known
problem in the optical flow literature [2]. The final interpolation result based on these two warp frames still contains this
artifact (Case I, right column). In contrast, TOFlow does not
stick to object motion: the background should be static, but
it has non-zero motion (Case II, left column). With TOFlow,
however, there is barely any artifact in the warped frames
(Case II, middle column) and the interpolated frame looks
clean (Case II, right column). The hallucinated background
motion actually helps to reduce the doubling artifacts. This
shows that TOFlow can reduce errors and synthesize frames
better than the ground truth flow.
Similarly, on the right of Figure 2, we show an example of
video denoising. The random small boxes in the input frames
are synthetic noises. If we warp the first and the third frames
to the second using the ground truth flow, the noisy patterns
(random squares) remain, and the denoised frame still contains
some noise (Case I, right column. There are some shadows of
boxes on the bottom). But if we warp these two frames using
TOFlow (Case II, left column), those noisy patterns are also
reduced or eliminated (Case II, middle column), and the final
denoised frame base on them contains almost no noise. This
also shows that TOFlow learns to reduce the noise in input
frames by inpainting them with neighboring pixels, which
Flow Network
Input at
diff. scales
Frame 1
…
SPN
Flow net
…
…
Reference
…
Motion L1
Not used in interp.
Frame T
…
…
Motion L2
…
Motion L4
Image Processing Network for Interpolation (with Mask)
Improc net
Output
frame
Warped frame 𝐼21
Motion 𝑣21
Mask 𝑚21
′
Masked frame 𝐼21
Motion 𝑣23
Mask 𝑚23
′
Masked frame 𝐼23
SPN
Flow net
Input
frames
Motion
fields
Flow Estimation
Warped
input
Transformation
Image Processing
Interpolated
frame 𝐼2
Warped frame 𝐼23
Figure 3: Left: our model using task-oriented flow for video processing. Given an input video, we first calculate the motion
between frames through a task-oriented flo estimation network. We then warp input frames to the reference using spatial
transformer networks, and aggregate the warped frames to generate a high-quality output image. Right top: the detailed
structure of flow estimation network (the orange network on the left). Right bottom: the detailed structure of image processing
network for interpolation (the gray network on the left).
traditional flow cannot do.
Now we discuss the details of each module as follows.
the later modules of the network can transform the first and
the third frames to the second frame for synthesis.
4.2. Flow Estimation Module
4.3. Image Transformation Module
The flow estimation module calculates the motion fields between input frames. For a sequence with N frames (N = 3 for
interpolation and N = 7 for denoising and super-resolution),
we select the middle frame as the reference. The flow estimation module consists of N − 1 flow networks, all of which
have the same structure and share the same set of parameters.
Each flow network (the orange network in Figure 3) takes one
frame from the sequence and the reference frame as input,
and predicts the motion between them.
We use the multi-scale motion estimation framework proposed by [27] to handle the large displacement between
frames. The network structure is shown in the top right subfigure of Figure 3. The input to the network are Gaussian
pyramids of both the reference frame and another frame rather
than the reference. At each scale, a sub-network takes both
frames at that scale and upsampled motion fields from previous prediction as input, and calculates a more accurate motion
fields. We uses 4 sub-networks in a flow network, three of
which are shown Figure 3 (the yellow networks).
There is a small modification for frame interpolation,
where the reference frame (frame 2) is not an input to the
network, but what it should synthesize. To deal with that, the
motion estimation module for interpolation consists of two
flow networks, both taking both the first and third frames as
input, and predict the motion fields from the second frame to
the first and the third respectively. With these motion fields,
Using the predicted motion fields in the previous step, the
image transformation module registers all the input frames
to the reference frame. We use the spatial transformer networks [11] for registration, which synthesizes a new frame
after transformation using bilinear interpolation. One important property of this module is that it can back-propagate the
gradients from the image processing module to the flow estimation module, so we can learn a flow representation that
adapts to different video processing tasks.
4.4. Image Processing Module
We use another convolutional network as the image processing module to generate the final output. For each task,
we use a slightly different architecture. Please refer to our
appendix for details.
Occluded regions in warped frames As mentioned Section 4.1, occlusion often results in doubling artifacts in
the warped frames. To solve this, some interpolation algorithms [18, 24, 25] estimate occlusion masks and only use
pixels that are not occluded in interpolation. Inspired by this,
we also design an optional mask prediction network for frame
interpolation, in addition to the image processing module.
The mask prediction network takes the two estimated motion
fields as input, one from frame 2 to frame 1, and the other
from frame 2 to frame 3 (v21 and v23 in the bottom right of
Figure 3). It predicts two occlusion masks: m21 is the mask
Input 𝐼1
Input 𝐼3
Warp 𝐼1 by Epicflow
Warp 𝐼3 by Epicflow
Warp 𝐼1 by TOFlow
TOFlow interp (no mask)
Warp 𝐼3 by TOFlow TOFlow interp (use mask)
Figure 4: Comparison between Epicflow interpolation [28]
and TOFlow interpolation (both with and without mask).
of the warped frame 2 from frame 1 (I21 ), and m23 is the
mask of the warped frame 2 from frame 3 (I23 ). The invalid
regions in the warped frames (I21 and I23 ) are masked out
by multiplying them with their corresponding masks. The
middle frame is then calculated through another convolutional
neural network with both the warped frames (I21 and I23 ) and
0
0
the masked warped frames (I21
and I23
) as input. Please refer
to our appendix for details of the network structure.
Even without the mask prediction network, our flow estimation is mostly robust to occlusion. As shown in the third
column of Figure 4, the warped frames using TOFlow has little doubling artifacts. Therefore, just from two warped frames
without the learned masks, the network synthesizes a decent
middle frame (the top image of the right most column). The
mask network helps to remove some tiny artifacts, such as
the faint ghost of the bottom thumb circled by white (visible
when zoomed in).
4.5. Training
To accelerate the training procedure, we first pre-train
some modules of the network and then fine-tune all of them
together. Details are described below.
Pre-training the Flow Estimation Network Pre-training
the flow network consists of two steps. First, for all tasks,
we pre-train the motion estimation network on the Sintel
dataset [4], a realistically rendered video dataset with ground
truth optical flow, by minimizing the l1 difference between
estimated optical flow and the ground truth. In the second
step, for video denoising and super-resolution, we fine-tune
it with noisy or blurry input frames to improve its robustness
to these input. For video interpolation, we fine-tune it with
frames I1 and I3 from video triplets as input, minimizing
the l-1 difference between the estimated optical flow and the
ground truth flow V23 (or v21 ). This enables the flow network
to calculate the motion from the unknown frame I2 to frame
I3 given only frames I1 and I3 as input. Empirically we find
that this two-step pre-training can improve the convergence
speed.
Pre-training the Mask Network We also pre-train our occlusion mask estimation network for video interpolation as an
optional component of video processing network before joint
training. Two occlusion masks (m21 and m23 ) are estimated
together with the same network and only optical flow v21 , v23
as input. The network is trained by minimizing the l-1 loss
between the output masks and pre-computed occlusion masks.
Joint Training After pre-training, we train all the modules
jointly by minimizing the l1 loss between recovered frame and
the ground truth, without any supervision on estimated flow
fields. For optimization, we use ADAM [13] with a weight
decay of 10−4 . We run 15 epochs with batch size 1 for all
tasks. The learning rate for denoising/deblocking and superresolution is 10−4 , and the learning rate for interpolation is
3 × 10−4 .
5. The Vimeo-90K Dataset
To acquire high quality videos for video processing, previous works [17, 14] take videos by themselves, resulting
in video datasets that are small in size and limited in terms
of content. Alternatively, we resort to Vimeo where many
videos are taken with professional cameras on diverse topics.
In addition, we only search for videos without inter-frame
compression (e.g., H.264), so that each frame is compressed
independently, avoiding artificial signals introduced by video
codecs. As many videos are composed of multiple shots,
we use a simple threshold-based shot detection algorithm to
break each video into consistent shots and further use GIST
feature [26] to remove shots with similar scene background.
As a result, we collect a new video dataset from Vimeo,
consisting of 4,278 videos with 89,800 independent shots that
are different from each other in content. To standardize the input, we resize all frames to the fixed resolution 448×256. As
shown in Figure 5, frames sampled from the dataset contain
diverse content for both indoor and outdoor scenes. We keep
consecutive frames when the average motion magnitude is
between 1–8 pixels. The right column of Figure 5 shows the
histogram of flow magnitude over the whole dataset, where
the flow fields are calculated using SpyNet [27].
We further generate three benchmarks from the dataset for
the three video enhancement tasks studied in this paper.
Vimeo Interpolation Benchmark We select 73,171 frame
triplets from 14,777 video clips with the following three criteria for the interpolation task. First, more than 5% pixels
should have motion larger than 3 pixels between neighboring
frames. This criterion removes static videos. Second, l1 difference between the reference and the warped frame using
optical flow (calculated using SpyNet [27]) should be at most
15 pixels (the maximum intensity level of an image is 255).
This removes frames with large intensity change, which are
too hard for frame interpolation. Third, the average difference
between motion fields of neighboring frames (v21 and v23 )
should be less than 1 pixel. This removes non-linear motion,
15.1
14.1
12.1
13.1
11.1
9.1
10.1
8.1
7.1
6.1
5.1
4.1
2.1
3.1
1.1
Frequency
0.1
Flow
magnitude
(a) Sample frames
12.1
13.1
11.1
10.1
9.1
8.1
7.1
6.1
5.1
4.1
3.1
2.1
1.1
0.1
Frequency
(b) Pixel-wise flow frequency
Flow
magnitude
(c) Image mean flow frequency
Figure 5: The Vimeo-90K dataset. (a) Sampled frames from the dataset, demonstrating the high quality and wide coverage of
our dataset; (b) The histogram of flow magnitude of all pixels in Vimeo-90K; (c) The histogram of mean flow magnitude of all
images (the flow magnitude of an image is the average flow magnitude of all pixels in that image).
as most interpolation algorithms, including ours, are based on
linear motion assumption.
Vimeo Denoising/Deblocking Benchmark We select
91,701 frame septuplets from 38,990 video clips for the denoising task, using the first two criteria introduced for the
interpolation benchmark. For video denoising, we consider
two types of noises: a Gaussian noise with a standard deviation of 0.1, and mixed noises including a 10% salt-and-pepper
noise in addition to the Gaussian noise. For video deblocking,
we compress the original sequences using FFmpeg with codec
JPEG2000, format J2k, and quality value q = 20.
Vimeo Super-Resolution Benchmark We also use the
same set of septuplets for denoising to build the Vimeo
Super-Resolution Benchmark with down-sampling factor of
4: the resolution of input and output images are 112 × 64 and
448 × 256 respectively. To generate the low-resolution videos
from high-resolution input, we use the MATLAB imresize
function, which first blurs the input frames using cubic filters
and then downsamples videos using bicubic interpolation.
Methods
Vimeo interp.
DVF dataset [18]
PSNR
SSIM
PSNR
SSIM
SpyNet [27]
EpicFlow [28]
DVF [18]
AdaConv [24]
SepConv [25]
31.95
32.02
—
32.33
33.45
0.9601
0.9622
—
0.9568
0.9674
33.60
33.71
34.12
—
—
0.9633
0.9635
0.9631
—
—
Fixed flow
Fixed flow + Mask
29.09
30.10
0.9229
0.9322
31.61
32.23
0.9544
0.9575
TOFlow
TOFlow + Mask
33.53
33.73
0.9668
0.9682
34.54
34.58
0.9666
0.9667
Table 1: Quantitative comparison between different frame
interpolation algorithms on Vimeo interpolation test set and
DVF test set [18].
6.1. Frame Interpolation
Datasets We evaluate on two datasets: Vimeo Interpolation
Benchmark and the dataset used by [14].
6. Evaluation
Metrics We use two quantitative measure to evaluate the
performance of interpolation algorithms: peak signal-to-noise
ratio (PSNR) and structural similarity (SSIM) index.
In this section, we evaluate two variations of the proposed
network. The first one is to train each module separately:
we first pre-train motion estimation, and then train video
processing while fixing the flow module. This is similar to
the two-step video processing algorithms, and we refer to it
as Fixed Flow. The other one is to jointly train all modules as
described in Section 4.5, and we refer to it as TOFlow. Both
networks are trained on Vimeo benchmarks we collected.
We evaluate these two variations on three different tasks and
also compare with other state-of-the-art image processing
algorithms.
Baselines We first compare our framework with two-step
interpolation algorithms. For the motion estimation, we use
EpicFlow [28] and SpyNet [27]. To handle occluded regions
as mentioned in Section 4.4, we calculate the occlusion mask
for each frame using [38] and only use non-occluded regions
to interpolate the middle frame. Further, we compare with
state-of-the-art end-to-end models, Deep Voxel Flow (DVF)
[18], Adaptive Convolution (AdaConv) [24], and Separable
Convolution (SepConv) [25]. At last, we also compare with
Fixed Flow, which is another baseline two-step interpolation
algorithm.
Epicflow
AdaConv
SepConv
Fixed Flow
TOFlow
Ground truth
Figure 6: Comparison between different frame interpolation algorithms. Zoomed-in views are shown in lower right.
Vimeo-Mixed Dataset
Vimeo-BW Dataset
V-BM4D Dataset
Input noisy frame
Fixed Flow
TOFlow
Ground truth
VBM4D
TOFlow
Ground truth
Figure 7: Comparison between different algorithms on video denoising. The differences are clearer when zoomed-in.
Results Table 1 shows our quantitative results. On Vimeo
Interpolation Benchmark, TOFlow in general outperforms the
others interpolation algorithms, both the traditional two-step
interpolation algorithms (EpicFlow and SpyNet) and recent
deep-learning based algorithms (DVF, AdaConv, and SepConv), with a significant margin. Moreover, even our model
is trained on our Vimeo-90K dataset, it also out-performs
DVF on DVF dataset in both PSNR and SSIM. There is also
a significant boost over Fixed Flow, showing that the network
does learn a better flow representation for interpolation during
joint training.
Figure 6 also shows qualitative results. All the two-step
algorithms (EpicFlow and Fixed Flow) generate a doubling
artifacts, like the hand in the first row or the head in the second
row. AdaConv on the other sides does not have the doubling
artifacts, but it tends to generate blurry output, by directly
synthesizing interpolated frames without a motion module.
SepConv increases the sharpness of output frame compared
with AdaConv, but there are still artifacts (see the hat on
the bottom row). Compared with these methods, TOFlow
correctly recovers sharper boundaries and fine details even in
presence of large motion.
Baselines We compare our framework with the V-BM4D,
with the standard deviation of Gaussian noise as its additional
input on two grayscale datasets (Vimeo-BW and V-BM4D).
As before, we also compare with the fixed flow variant of
our framework on two RGB datasets (Vimeo-Gaussian and
Vimeo-Mixed).
Results On two RGB datasets, Vimeo-Gaussian and VimeoMixed, TOFlow beats Fixed Flow in both two measurements,
as shown in Table 2. The output of TOFlow also contains less
noise (the differences are clearer when zoomed-in), as shown
on the left side of Figure 7. This shows that TOFlow learns a
motion field for denoising.
6.2. Video Denoising/Deblocking
On two grayscale datasets, Vimeo-BW and V-BM4D,
TOFlow outperforms V-BM4D in SSIM, even we did not finetuned on V-BM4D dataset. Note that even though TOFlow
only achieves a comparable performance with V-BM4D in
PSNR, the output of TOFlow is much sharper than V-BM4D.
As shown in Figure 7, the words on the billboard are kept in
the denoised frame by TOFlow (the top right of Figure 7),
and leaves on the tree are also clearer (the bottom right of
Figure 7). Therefore, TOFlow beats V-BM4D in SSIM, which
better reflects human’s perception than PSNR.
Setup We first train and evaluate our framework on Vimeo
denoising benchmark, with either Gaussian noise or mixture
noise. To compare our network with V-BM4D [20], which is
a monocular video denoising algorithm, we transfer all videos
in Vimeo Denoising Benchmark to grayscale to create VimeoBW (Gaussian noise only), and retrain our network on it. We
also evaluate our framework on the V-BM4D dataset [20].
For video deblocking, Table 3 shows that TOFlow outperforms V-BM4D. Figure 8 also shows the qualitative comparison between TOFlow, Fixed Flow, and V-BM4D. Note
that the compression artifacts around the girl’s hair (top) and
the man’s nose (bottom) are completely removed by TOFlow.
The vertical line around the man’s eye (bottom) due to a
blocky compression is also removed by our algorithm.
Input compressed frames
VBM4D
Ground truth
TOFlow
Fixed Flow
Figure 8: Comparison between our algorithm and V-BM4D [20] on video deblocking. The difference are clearer when
zoomed-in.
Vimeo-Gaussian
Vimeo-Mixed
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
PSNR
SSIM
V-BM4D [20]
Fixed flow
—
28.38
—
0.9333
—
28.56
—
0.9200
27.38
—
0.8664
—
30.63
—
0.8759
—
TOFlow
29.10
0.9544
28.85
0.9407
29.41
0.9271
30.36
0.8855
Methods
Vimeo-BW
V-BM4D [20]
Table 2: Quantitative comparisons on video denoising
Input
Methods
Vimeo SR
BayesSR [16]
PSNR
SSIM
PSNR
SSIM
—
—
—
—
22.69
24.32
0.7746
0.8486
Full clip
DeepSR [14]
BayesSR [16, 19]
1 frame
Bicubic
29.79
0.9036
22.02
0.7203
DeepSR [14]
BayesSR [16, 19]
7 frames
Fixed flow
25.55
24.64
31.81
0.8498
0.8205
0.9288
21.85
21.95
22.85
0.7535
0.7369
0.7655
33.08
0.9417
23.54
0.8070
TOFlow
Table 4: Results of video super-resolution. Each clip in Vimeo
SR contains 7 frames, and each clip in BayesSR contains 30–
50 frames.
6.3. Video Super-Resolution
Datasets We evaluate our algorithm on two dataset: Vimeo
Super-Resolution Benchmark and the dataset provided by [16]
(BayesSR). The later one consists of 4 sequences, each of
which has 30 to 50 frames.
Baselines We compare our framework with bicubic upsampling, two video SR algorithms: BayesSR (we use the
version provided by Ma et al. [19]) and DeepSR [14], as
well as a baseline with a fixed flow estimation module. Both
BayesSR and DeepSR can take various number of frames as
input. Therefore, on BayesSR datset, we report two numbers.
One is using the whole sequence, the other is to only use
seven frames in the middle, as TOFlow and Fixed Flow only
take 7 frames as input.
Methods
Vimeo-Blocky
PSNR
SSIM
V-BM4D [20]
Fixed flow
35.75
36.52
0.9587
0.9636
TOFlow
36.92
0.9663
Table 3: Results on video deblocking
Results Table 4 shows our quantitative results. Our algorithm performs better than baseline algorithms when using 7
frames as input, and it also achieves comparison performance
to BayesSR when BayesSR uses all 30–50 frames as input
while our framework only uses 7 frames. We show qualitative results in Figure 9. Compared with either DeepSR [14]
or Fixed Flow, the jointly trained TOFlow generates sharper
output. Notice the words on the cloth (top) and the tip of
the knife (bottom) are clearer in the high-resolution frame
synthesized by TOFlow. This shows the effectiveness of joint
training.
In all the experiments, we train and evaluate our network
on a NVIDIA Titan X GPU. For an input clip with resolution
256 × 448, our network takes about 200 ms for interpolation
and 400ms for denoising or super-resolution (the input resolution to the super-resolution network is 64 × 112), where the
flow module takes 18 ms for each estimated motion field.
At last, Figure 10 also visualizes the motion fields learned
by different tasks. Even using the same network structure and
taking the same input frames, the estimated flows for different
tasks are very different: the flow field for interpolation is very
smooth, even on the occlusion boundary, while the flow field
for super-resolution has artificial movements along the texture
edges. This is indicates that the network may learn to encode
different information that is useful for different tasks in the
learned motion fields.
7. Conclusion
In this work, we propose a novel video processing model
that exploits task-oriented motion cues. Traditional video
Bicubic
DeepSR
Fixed Flow
TOFlow
Ground truth
Figure 9: Comparison between different super-resolution algorithms. A close-up is shown on the top left of each result. The
differences are clearer when zoomed-in.
Input
Flow for Interpolation
Flow for Denoising
Flow for SR
Flow for Deblocking
Figure 10: Visualization of motion fields for different tasks.
processing algorithms normally consist of two steps: motion
estimation and video processing based on estimated motion
fields. However, a genetic motion for all tasks might be suboptimal and the accurate motion estimation would be neither
necessary nor sufficient for these tasks. Our self-supervised
framework bypasses this difficulty by modeling motion signals in the loop. To evaluate our algorithm, we also create
a new dataset, Vimeo-90K, for video processing. Extensive
experiments on temporal frame interpolation, video denoising/deblocking, and video super-resolution demonstrate that
our algorithm achieves state-of-the-art performance.
Acknowledgements This work is supported by NSF RI1212849, NSF BIGDATA-1447476, Facebook, Shell Research, and Toyota Research Institute.
References
[1] S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici,
B. Varadarajan, and S. Vijayanarasimhan. Youtube-8m: A
large-scale video classification benchmark. arXiv:1609.08675,
2016. 2
[2] S. Baker, D. Scharstein, J. Lewis, S. Roth, M. J. Black, and
R. Szeliski. A database and evaluation methodology for optical
flow. IJCV, 92(1):1–31, 2011. 1, 2, 3
[3] T. Brox, A. Bruhn, N. Papenberg, and J. Weickert. High accuracy optical flow estimation based on a theory for warping. In
ECCV, 2004. 2
[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A
naturalistic open source movie for optical flow evaluation. In
ECCV, 2012. 5
[5] J. Caballero, C. Ledig, A. Aitken, A. Acosta, J. Totz, Z. Wang,
and W. Shi. Real-time video super-resolution with spatio-
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
temporal networks and motion compensation. In CVPR, 2017.
2
P. Fischer, A. Dosovitskiy, E. Ilg, P. Häusser, C. Hazırbaş,
V. Golkov, P. van der Smagt, D. Cremers, and T. Brox. Flownet:
Learning optical flow with convolutional networks. In ICCV,
2015. 2
Y. Ganin, D. Kononenko, D. Sungatullina, and V. Lempitsky.
Deepwarp: Photorealistic image resynthesis for gaze manipulation. In ECCV, 2016. 2
M. Ghoniem, Y. Chahir, and A. Elmoataz. Nonlocal video
denoising, simplification and inpainting using discrete regularization on graphs. Signal Process., 90(8):2445–2455, 2010.
2
B. K. Horn and B. G. Schunck. Determining optical flow. Artif.
Intell., 17(1-3):185–203, 1981. 2
Y. Huang, W. Wang, and L. Wang. Bidirectional recurrent
convolutional networks for multi-frame super-resolution. In
NIPS, 2015. 2
M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. In NIPS, 2015. 2, 4
A. Kappeler, S. Yoo, Q. Dai, and A. K. Katsaggelos. Video
super-resolution with convolutional neural networks. IEEE
TCI, 2(2):109–122, 2016. 2
D. P. Kingma and J. Ba. Adam: A method for stochastic
optimization. In ICLR, 2015. 5
R. Liao, X. Tao, R. Li, Z. Ma, and J. Jia. Video super-resolution
via deep draft-ensemble learning. In CVPR, 2015. 2, 5, 6, 8
C. Liu and W. Freeman. A high-quality video denoising algorithm based on reliable motion estimation. In ECCV, 2010. 1,
2
C. Liu and D. Sun. A bayesian approach to adaptive video
super resolution. In CVPR, 2011. 1, 8
C. Liu and D. Sun. On bayesian adaptive video super resolution.
IEEE TPAMI, 36(2):346–360, 2014. 2, 5
[18] Z. Liu, R. Yeh, X. Tang, Y. Liu, and A. Agarwala. Video frame
synthesis using deep voxel flow. In ICCV, 2017. 2, 4, 6
[19] Z. Ma, R. Liao, X. Tao, L. Xu, J. Jia, and E. Wu. Handling
motion blur in multi-frame super-resolution. In CVPR, 2015. 8
[20] M. Maggioni, G. Boracchi, A. Foi, and K. Egiazarian. Video
denoising, deblocking, and enhancement through separable 4-d
nonlocal spatiotemporal transforms. IEEE TIP, 21(9):3952–
3966, 2012. 2, 7, 8
[21] M. Mathieu, C. Couprie, and Y. LeCun. Deep multi-scale video
prediction beyond mean square error. In ICLR, 2016. 2
[22] E. Mémin and P. Pérez. Dense estimation and object-based
segmentation of the optical flow with robust techniques. IEEE
TIP, 7(5):703–719, 1998. 2
[23] K. Nasrollahi and T. B. Moeslund. Super-resolution: a comprehensive survey. MVA, 25(6):1423–1468, 2014. 2
[24] S. Niklaus, L. Mai, and F. Liu. Video frame interpolation via
adaptive convolution. In CVPR, 2017. 2, 4, 6
[25] S. Niklaus, L. Mai, and F. Liu. Video frame interpolation via
adaptive separable convolution. In ICCV, 2017. 4, 6
[26] A. Oliva and A. Torralba. Modeling the shape of the scene: A
holistic representation of the spatial envelope. IJCV, 42(3):145–
175, 2001. 5
[27] A. Ranjan and M. J. Black. Optical flow estimation using a
spatial pyramid network. In CVPR, 2017. 2, 4, 5, 6, 10
[28] J. Revaud, P. Weinzaepfel, Z. Harchaoui, and C. Schmid.
Epicflow: Edge-preserving interpolation of correspondences
for optical flow. In CVPR, 2015. 1, 2, 5, 6
[29] X. Tao, H. Gao, R. Liao, J. Wang, and J. Jia. Detail-revealing
deep video super-resolution. In ICCV, 2017. 2
[30] G. Varghese and Z. Wang. Video denoising based on a spatiotemporal gaussian scale mixture model. IEEE TCSVT,
20(7):1032–1040, 2010. 2
[31] T.-C. Wang, J.-Y. Zhu, N. K. Kalantari, A. A. Efros, and R. Ramamoorthi. Light field video capture using a learning-based
hybrid imaging system. In SIGGRAPH, 2017. 2
[32] A. Wedel, D. Cremers, T. Pock, and H. Bischof. Structure-and
motion-adaptive regularization for high accuracy optic flow. In
CVPR, 2009. 2
[33] M. Werlberger, T. Pock, M. Unger, and H. Bischof. Optical flow
guided tv-l1 video interpolation and restoration. In EMMCVPR,
2011. 2
[34] J. Xu, R. Ranftl, and V. Koltun. Accurate optical flow via direct
cost volume processing. In CVPR, 2017. 2
[35] J. J. Yu, A. W. Harley, and K. G. Derpanis. Back to basics:
Unsupervised learning of optical flow via brightness constancy
and motion smoothness. In ECCV Workshop, 2016. 2
[36] Z. Yu, H. Li, Z. Wang, Z. Hu, and C. W. Chen. Multi-level
video frame interpolation: Exploiting the interaction among
different levels. IEEE TCSVT, 23(7):1235–1248, 2013. 2
[37] T. Zhou, S. Tulsiani, W. Sun, J. Malik, and A. A. Efros. View
synthesis by appearance flow. In ECCV, 2016. 2
[38] C. L. Zitnick, S. B. Kang, M. Uyttendaele, S. Winder, and
R. Szeliski. High-quality video view interpolation using a
layered representation. ACM TOG, 23(3):600–608, 2004. 6
Appendices
Additional Qualitative Results In addition to the qualitative results shown in the main text (Figures 6–8), we show additional results on the following benchmarks: Vimeo Interpolation Benchmark (Figure 12), Vimeo Denoising Benchmark
(Figure 13 for grayscale videos), Vimeo Deblocking Benchmark (Figure 14), and Vimeo Super-resolution Benchmark
(Figure 15). To avoid cherry-picking, we randomly select
testing images from test datasets and do not show close-up as
Figures 6–9 in the main text. Differences between different
algorithms are more clearer when zoomed in.
Flow Estimation Module We used SpyNet [27] as our flow
estimation module. It consists of 4 sub-networks with the
same network structure, but each sub-network has an independent set of parameters. Each sub-network consists of 5 sets of
7x7 convolutional (with zero padding), batch normalization
and ReLU layers. The number of channels after each convolutional layer is 32, 64, 32, 16 and 2 respectively. The input
motion to the first network is a zero motion field.
Image Processing Module We use slight different structures in the image processing module for different tasks. For
temporal frame interpolation both with and without masks,
we build a residual network that consists of a averaging network and a residual network. The averaging network simply
averages the two transformed frames (from frame 1 and frame
3 respectively). The residual network also takes the two transformed frames as input, but calculates the difference between
the actual second frame and the average of two transformed
frames through a convolutional network consists of three convolutional layers, each of which is followed by a ReLU layer.
The kernel sizes of three layers are 9x9, 1x1, 1x1 respectively
(with zero padding), and the numbers of output channels are
64, 64, 3 respectively. The final output is the summation of
the outputs of these two networks (averaging network and
residual network).
For video denoising/deblocking, the image processing
module uses the same six-layer convolutional structure (3
convolutional layers and 3 ReLU layers) as interpolation, but
without the residual structure. We have also tried the residual
structure for denoising/deblocking, but there is no significant
improvement.
For video super-resolution, the image processing module
consists of 4 pairs of convolutional layers and ReLU layers.
The kernel sizes for these four layers are 9x9, 9x9, 1x1 and
1x1 respectively (with zero padding), and the numbers of
output channels are 64, 64, 64, and 3 respectively.
Mask Network Similar to our flow estimation module, our
mask estimation network is also a 4-level convolutional neural
network pyramid as in Figure 11. Each level consists of the
same sub-network structure with 5 sets of 7x7 convolutional
(with zero padding), batch normalization and ReLU layers,
but an independent set of parameters (output channels are 32,
64, 32, 16, and 2 respectively). For the first level, the input to
Mask Network
Estimated flow
at diff. scales
…
…
…
Masks L1
Masks L2
Masks L4
Figure 11: The structure of our mask network
the network is a concatenation of two estimated optical flow
fields (4 channels after concatenation), and the output is a
concatenation of two estimated masks (1 channel per mask).
From the second level, the inputs to the network switch to
a concatenation of 1) two estimated optical flow fields at
that resolution, and 2) bilinear-upsampled masks from the
previous level (the resolution is twice of the previous level).
In this way, the first level mask network estimates a rough
mask, and the rest refines high frequency details of the mask.
Epicflow
AdaConv
SepConv
Fixed flow
TOFlow
Ground Truth
Figure 12: Qualitative results for video interpolation. Samples are randomly selected from Vimeo Interpolation Benchmark.
The differences between different algorithms are clear only when zoomed in.
Input
Fixed Flow
TOFlow
Ground Truth
Input
Vbm4d
TOFlow
Ground Truth
Figure 13: Qualitative results for video denoising. The top five rows are results on color videos and bottom 4 rows are
grayscale videos. Samples are randomly selected from Vimeo Denoising Benchmark. The differences between different
algorithms are clear only when zoomed in.
Input
Vbm4d
TOFlow
Ground Truth
Figure 14: Qualitative results for video deblocking. Samples are randomly selected from Vimeo Deblocking Benchmark. The
differences between different algorithms are clear only when zoomed in.
Bicubic
DeepSR
Fixed flow
TOFlow
Ground Truth
Figure 15: Qualitative results for video super-resolution. Samples are randomly selected from Vimeo Super-resolution
Benchmark. The differences between different algorithms are clear only when zoomed in. DeepSR was originally trained on
30–50 images, but evaluated on 7 frames in this experiment, so there are some artifacts.
| 1cs.CV
|
World of Computer Science and Information Technology Journal (WCSIT)
ISSN: 2221-0741
Vol. 4, No. 7, 90-100, 2014
Intelligent Emergency Message Broadcasting in
VANET Using PSO
Ghassan Samara
Tareq Alhmiedat
Department of Computer Science, Zarqa University
Zarqa, Jordan
Department of Information Technology, Tabuk
University
Tabuk, Saudi Arabia
Abstract— The new type of Mobile Ad hoc Network which is called Vehicular Ad hoc Networks (VANET) created a fertile
environment for research.
In this research, a protocol Particle Swarm Optimization Contention Based Broadcast (PCBB) is proposed, for fast and effective
dissemination of emergency messages within a geographical area to distribute the emergency message and achieve the safety
system, this research will help the VANET system to achieve its safety goals in intelligent and efficient way.
Keywords- PSO; VANET; Message Broadcasting; Emergency System; Safety System.
The new techniques in this system should aim to make
the intelligent vehicle to think, communicate with other
vehicles and act to prevent hazards.
I. INTRODUCTION
Recent Year’s rapid development in wireless
communication networks has made Car to Car (C2C) and
Car to Infrastructure Communications (C2I) possible in
Mobile Ad hoc Networks (MANETs). This has given birth to
a new type of high mobile MANET called Vehicular Ad hoc
Networks (VANET) creating a fertile area of research aiming
for road safety, efficient driving experience, and
infotainment (Information and Entertainment) [1, and 2].
VANET safety applications depend on exchanging the
safety information among vehicles (C2C communication) or
between Vehicle to infrastructure (C2I Communication)
using the control channel, see figure 1.
Creating a safety system on the road is a very important
and critical concern for human today, each year nearly 1.3
million people die as a result of road traffic accidents – more
than 3000 deaths each day - and more than half of these
people are not travelling in a car, the injuries are about fifty
times of this number [3]. The number of cars in 2004 is
approximately estimated as 750 million cars around the
world [4], with an annually constant increase by 50 million
car around the world [5], with this constant raise, the
estimated number of cars nowadays exceeding one billion,
this raise the possibility to increase the number of crashes
and deaths on the roads, road traffic accidents are predicted
to become the fifth leading cause of death in the world,
resulting in an estimated 2.4 million death each year as stated
by the World Health Organization (WHO) [3], besides traffic
congestion makes a huge waste of time and fuel, this makes
developing an efficient safety system an urgent need on the
road.
Figure 1: VANET Structure
VANET safety communication can be made by two
means: Periodic Safety Message (called Beacon in this
paper) and Event Driven Message (called Emergency
Message in this paper), both sharing only one control
channel. The Beacon messages are status messages
containing status information about the sender vehicle like
position, speed, heading …etc. Beacons provide fresh
This research is funded by the Deanship of Research and
Graduate Studies in Zarqa University/ Jordan.
90
WCSIT 4 (7), 90 -100, 2014
information about the sender vehicle to the surrounding
vehicles in the network helping them to know the status of
the current network and predict the movement of vehicles.
Beacons are sent aggressively to neighboring vehicles 10
messages each second.
Depending on just one forwarder is not enough in a high
mobile network like VANET. Furthermore, authors didn’t
depend on beacons to gain the information. They proposed to
use hello message, which creates a chance to increase the
channel load.
Emergency Messages are messages sent by a vehicle
detect a potential dangerous situation on the road; this
information should be disseminated to alarm other vehicles
about a probable danger that could affect the incoming
vehicles. VANET is a high mobile network where the nodes
are moving in speeds that may exceed 120km/h, which
means that this vehicle move 33.33m/s, even if these vehicles
are very far from the danger, they will reach it very soon,
here milliseconds will be very important to avoid the danger
[6, and 7].
The contention period schemes (which is a waiting time
that the receiver waits before rebroadcasting the original
message received from the sender) are proposed by many
researchers[9], [10], [11], [12], [13], [14] and [15].
In [9], authors proposed the Link-based Distributed
Multi-hop Broadcast (LDMB), in which all the receivers of
the emergency message are potential forwarders. Each
forwarder computes and waits for contention time using
equation (3.2), if the contention time ends the forwarder will
start to rebroadcast the emergency message.
Emergency messages in VANET are sent in broadcast
fashion where all the vehicle inside the coverage area of the
sender should receive the message. The coverage area is not
enough as it is hardly reaches a 1000m (which is the DSRC
communication range) due to attenuation and fading effects.
Away vehicles from the danger should receive this critical
information to avoid the danger. Furthermore, the probability
of message reception can reach 99% in short distances and
can be as low as 20% at half of the communication range
(Moreno, 2004). Therefore, there should be a technique to
increase the emergency message reception with high
reliability and availability.
In [11] and [12], where authors proposed position-based
message forwarding strategy by sending the emergency
message in a broadcast fashion, and selecting the best
forwarder available. All vehicles receiving that message are
potential forwarders. In order to decide which node forwards
the message all receivers will be assigned a contention
window (waiting time); the contention window size will be
the smallest for the farthest node and the biggest size for the
nearest node, in other words, this protocol will give priority
for the farthest node to be the next forwarder.
The problem of the last two protocols that all the message
receivers will compute the waiting time and wait to make the
rebroadcast even the closest vehicles to the sender will do
and this will make the entire network vehicles busy for any
message received.
Duo to the high mobility of vehicles, the distribution of
nodes within the network changes rapidly, and unexpectedly
that wireless links initialize and break down frequently and
unpredictably. Therefore, broadcasting of messages in
VANETs plays a crucial rule in almost every application and
requires novel solutions that are different from any other
form of Ad-Hoc networks. Broadcasting of messages in
VANETs is still an open research challenge and needs some
efforts to reach an optimum solution.
Another protocol proposed by [13] called Emergency
Message Dissemination for Vehicular (EMDV) protocol, by
enabling the farthest vehicle within the transmission range to
make the rebroadcasting of the emergency message.
Choosing one forwarder vehicle is not appropriate in a
high mobile network like VANET as the position is always
changing, and the receiver vehicle may become out of range
when sending the message or simply the receiver can’t
receive the message because of the channel problems like
jam or denial of service, see figure 2.
Broadcasting requirements are: high reliability and high
dissemination speed with short latency in single-hop as well
as multi-hop communications. Problems associated with
regular broadcasting algorithms are: the high probability of
collision in the broadcasted messages, the lack of feedback
and the hidden node problem.
In this paper we concerned with proposing a new
intelligent broadcasting technique for the emergency
message in VANET aiming to increase the reception of the
emergency information.
II. RESEARCH BACKGROUND
Emergency message rebroadcast
In [8], authors proposed a street-based broadcast scheme
that utilizes neighbor’s information by exchanging hello
messages among vehicles, when any probable danger is
detected, a warning message is broadcasted to all neighbors.
The farthest vehicle is selected as a forwarder depending on
the information gained from the hello message, if the
preselected forwarder receives the message, it will
rebroadcast it.
Figure 2: Sender utilizing EMDV
In [10] authors proposed that the receivers of the message
will select random waiting times and make acknowledgment
91
WCSIT 4 (7), 90 -100, 2014
to avoid the re-transmissions from nodes closer to the
original sender.
Emergency message rebroadcast by network segments
Another way to rebroadcast the message is to divide the
network into segments proposed in [16, 17, and 18].
The acknowledgment scheme causes delay to the
rebroadcast.
In [16] authors proposed a protocol called Urban Multi–
hop broadcast (UMB) aiming to maximize the message
progress, and avoid broadcast storm, hidden node, and
reliability problems. The protocol assigns the duty of
forwarding and acknowledging the broadcast packets to only
one vehicle by dividing the road portion inside the
transmission range into segments and choosing the vehicle in
the furthest non-empty segment without prior topology
information. The source node transmits a broadcast control
packet, called Request to Broadcast (RTB), which contains
the position of the source and the segment size. On receiving
the RTB packet, nodes compute the distance between the
sender and the receiver. Then, nodes transmit a channel
jamming signal, called black–burst, that contains several
time–slots equal to their distance from the source (in number
of segments): the farther the distance, the longer the black–
burst. Each node transmits its black-burst and senses the
channel; if there is no other black-burst in the channel it
concludes that it is the farthest node from the source. Then
the node returns a Clear–to–Broadcast (CTB) control packet,
containing its identifier (ID), to the source.
In [14] authors proposed the Contention-Based
Forwarding (CBF) protocol where a vehicle sends a packet
as a broadcast message to all its neighbors. On receiving the
packet, neighboring vehicle will contend for forwarding the
packet. The node having the maximum progress to the
destination will have the shortest contention time and will
first rebroadcast the packet. If other nodes receive the
rebroadcast message, they will stop their contention and
delete the previously received message. This protocol mainly
proposed for forwarding the periodic safety message
(Beacons).
The problem of this protocol that there should be a
management technique to manage the contention for all the
neighboring vehicles, and there is a chance that the nearest
vehicle to the sender may not hear the rebroadcast of another
vehicle, here this vehicle will rebroadcast the message and
this called (hidden node problem (Tobagi and Kleinrock,
1975)) also it may lead to broadcast storm problem that
makes the protocol useless.
In [15] authors suggested that the emergency message
will be rebroadcasted by the receivers located at farther
distances from the sender by the selection of shorter waiting
times, see equation 1.
The Smart Broadcasting Protocol [17] addressed the
same objective as UMB using a different methodology.
Upon reception of a RTB message, each vehicle should
determine its segment and set a random back-off time. Each
segment has its own contention window size, i.e. if this
segment has contention window size (4) TS (time-slot);
vehicles in the furthest segment should randomly choose a
back-off time between (0) to (3) TS. Vehicles in the next
nearer segment choose a value between (4) to (7) TS, and so
on, as vehicles near the sender should wait for longer time.
In [7] authors proposed the Contention Based
Broadcasting (CBB) protocol for increasing the emergency
message reception and performance, the emergency message
will be broadcasted in multi-hop fashion, and the multi-hop
forwarders will be selected before the original message is
sent. CBB proven to achieve superiority over the EMDV
protocol as it choses more than one forwarder to rebroadcast
the emergency information and this gives the message a
chance to overcome the preselected forwarder failure.
Vehicles will decrement their backoff timers by one in
each time-slot while listening to the physical channel. While
waiting, if any vehicle receives a valid CTB message, it will
exit the contention time phase and listen to the incoming
broadcast. On the contrary, if any node finishes its backoff
timer, it will send the CTB containing its identity and
rebroadcast any incoming broadcast.
The criteria of choosing the forwarders depends on the
progress and on the segment localization, see figure 3, where
all the vehicles located in the final; none-empty segment are
a potential forwarder.
While in [18] authors proposed the Geographic random
forwarding (GeRaF) protocol, which divides the network
into equally adjacent sectors, the transmitter (source) elects
the sectors starting from the farthest one, by sending RTB
message, all the nodes in the elected sectors reply by CTB
message, if one node reply the CTB message, then this node
will become the next forwarder, if there are more than one
node sent the CTB message the source issue a collision
message and make a collision-resolution procedure to elect
the next forwarder depending on a probabilistic rule.
Many other approaches are discussed in details in our
previous paper (emergency message broadcasting)
III. THE PROPOSED PROTOCOL
Figure 3: Emergency message Sending and transmission range
This section presents a detailed design description for the
PCBB protocol, which aims to increase the percentage of
92
WCSIT 4 (7), 90 -100, 2014
reception for the emergency information by utilizing a
contention window, position based forwarding scheme, and
PSO intelligent technique.
Sending the message in single hop enables it to reach a
number of vehicles within a limited distance up to 1000 m
for the best cases [19]. However, this number should be
increased in order to warn more vehicles of possible dangers
before they reach the danger area.
Beacons and the emergency messages should be received
by all the neighboring vehicles with high probability and
reliability, because of the critical nature of the information
both provide. When a vehicle detects danger, it issues an
emergency message to warn other vehicles within the
network, and all the vehicles in the opposite direction of the
sender movement located in its transmission range must
receive such message. Covering the whole area does not
guarantee that all vehicles will receive the message because
of channel collisions and fading effects. The percentage of
emergency message reception for the network vehicles must
be as high as possible.
Sen ID
This paper proposes to categorize any emergency
message before sending it to make it easier for the receiver
vehicle to recognize the importance of the message being
received. Table 1 lists the codes for each category. For
example, when a vehicle receives two messages containing
categories 1 and 5, it processes the message that contains
category 1 first because it contains more critical information.
After assigning the message code, the sender should add
data to the message, such as the coordinates of the danger
zone or what the receiver should do; however, this aspect
would not be discussed in detail in the current study. The
proposed structure of the emergency message is shown in
Figure 4, where three inputs, namely Cid, MinB and MaxB,
are added to help the receiver vehicle determine what action
to take after receiving the emergency message.
002
Safety of Life
Cooperative Collision Warning
003
Safety
Intersection warning.
004
Safety
Transit Vehicle Signal Priority
005
Non-Safety
Toll Collection
006
Non-Safety
Service Announcement
007
Non-Safety
Movie Download(2 hours of MPEG 1)
MinB
MaxB
As mentioned earlier, the network is divided to several
segments to help the vehicle determine the next forwarder of
the emergency message. As proposed in [16], the
transmission range of the sender is divided into 10 segments
to make it easier for the sender to determine the last vehicle
in the last non-empty segment, which is eventually selected
as the next forwarder. For this paper, the distance is between
the sender and the forwarder. Authors in [16] established a
fixed distance of 1000 m. For the current study, however, if
the distance between the sender and the farthest vehicle is
Application
Emergency Break Warning/Avoidance
CId
As mentioned earlier assigning the forwarding job of the
emergency message to all the receiver vehicles of the
message may cause a broadcast storm problem, and
assigning the forwarding job just for one receiver vehicle
may not be appropriate sometimes as this specific forwarder
may not receive the emergency message. Hence vehicles in
the last segment, which should be the furthest one, will make
the forwarding of the emergency message if the forwarder
fails to receive and forward the message.
TABLE 1: EMERGENCY MESSAGE CLASSIFICATION
Safety of Life
Data
Choosing the next candidate forwarder is a process,
which begins by gathering the information obtained from
beacons received from neighbors. This information is
inserted and ordered into NT. The sender vehicle chooses the
farthest vehicle and assigns it to be the candidate forwarder.
The process of forwarding the emergency message to
increase the probability of reception so that the forwarded
signal can communicate with more vehicles on the road and
reach longer distances is the option used in this paper.
Choosing only one forwarder is inappropriate in high mobile
networks, such as VANET, because the forwarder might not
receive the emergency message. To solve this problem,
dividing the network to several segments is proposed.
Vehicles inside the last non-empty segment (i.e., the farthest
segment from the sender) wait for a period of time and
determine whether or not the candidate forwarder
rebroadcasted the emergency message. If none made the
rebroadcast, the vehicles located in the farthest segment
forwards the message. As mentioned earlier, assigning the
forwarding job of the emergency message to all the receiver
vehicles of the message may cause a broadcast storm
problem, and assigning the forwarding job to just one
receiver vehicle may be inappropriate because the specific
forwarder may not receive the emergency message. Hence,
vehicles in the last segment must forward the emergency
message if the forwarder fails to receive and forward the
message.
Every beacon received by a vehicle provides important
information about the sender status. This information is
utilized to form a rich and real time image about the current
network topology, which facilitates better network vehicle
communication. It also helps to be informed about the
potential dangers when they occur. When a vehicle has a
problem or detects a problem, it determines if the problem is
life critical or not. The life critical (Safety of Life) messages
will be given the highest priority and are then processed and
sent before any other kind of messages.
001
Msg ID
Where Sen ID: sender id, Code: Message Code, TS:
Time Stamp, Msg ID: Message ID, Data: Data sent, CId:
Forwarder Candidate ID, MinB: Minimum Boundary,
MaxB: Maximum Boundary.
Preparing to send
Priority
TS
Figure 4: Emergency message illustration.
In order to cover a wider area for message reception,
some neighboring vehicles can serve as potential forwarders,
and each forwarder has to wait for a certain period of time
(i.e., contention time) before forwarding the message.
Code
Code
93
WCSIT 4 (7), 90 -100, 2014
900 m, anything beyond 900 m is not considered. The
distance between the last vehicle and the sender is computed
using Equation (1).
segment should be expanded to include more vehicles, and
this could be determined using Equation (6).
(6)
(1)
where Dis is the distance between last vehicle and the
sender, SenPos is the position of the sender obtained from
GPS, and ForPos is the forwarder position or the last vehicle
in the last non-empty segment.
This calculation doubles the size of the last segment and
increases the number of the potential forwarders. If the
calculated number remains to be SucPer < Nmax%, the MinB
could be recalculated by multiplying Dif by 3 and so on. This
technique increases the number of the potential forwarders
and solves the preselected forwarder rebroadcast failure.
Determining the boundaries of the last non-empty
segment must be set dynamically depending on the channel
status and the network topology available, because it would
be pointless if this segment does not contain enough number
of vehicles for forwarding. At the same time, determining the
number of sufficient vehicles located in the last non-empty
segment must also depend on the channel status and network
topology. The sender vehicle has all the information required
to analyze the channel and draw the network topology. To
compute the boundaries, the CBB and PCBB protocols are
proposed. The CBB protocol depends on the selection of
boundaries of the last non-empty segment based on the
number of the vehicles located in this segment, and the
number of segments in the network. [16] suggested that the
number of the segments should be 10 segments. Computing
the segments and the boundaries could be done using
Equations (2), (3) and (4). Equation (2) assigns the distance
between the sender, and the farthest vehicle (forwarder) is
the boundary of the last non-empty segment. Equation (3)
computes the length of each segment, and Equation (4) finds
the location of the minimum boundary. MinB is the
minimum boundaries (Borders) where the last segment starts.
Nmax is the maximum number of the segments, and Dif is
the length of the segment.
The PCBB is an enhancement of the CBB and works
when the sender vehicle analyzes the dense locations of the
vehicles along its transmission range. In Figure 5, the vehicle
analyzes the location density in the network to form groups
for dense locations; the resulting network is then divided into
several groups (Figure 6).
Figure 5: vehicle analysis location density.
(2)
(3)
(4)
This means that the vehicles located in the area between
MinB and MaxB from the sender are considered as potential
forwarders of the emergency message. They are rebroadcast
if no vehicle makes the rebroadcast. Sometimes, the last
segment may have an insufficient number of vehicles. The
number of the potential forwarders must have a threshold to
determine if it is sufficient or not.
Figure 6: Analyzed network depending on network density and progress.
A dense or concentrated area has a number of vehicles
within just a small area. Thus, sending a message to vehicles
in concentrated areas increases the chance of receiving and
rebroadcasting the message. The probability of receiving the
rebroadcasted messages in this segment is also high, thus
eliminating the hidden node problem, which is considered
one of the most difficult problems encountered in
rebroadcasting emergency message in VANET. Equation (7)
is used to calculate how vehicles compute the dense
locations. The progress represents the upper bound of the last
segment and the length of the segment is also the distance
between the farthest vehicle in the segment and the first
vehicle located in the segment. For example, if the segment
that has progressed to 820 m, has 15 vehicles in 80 meters
between 820 m and 900 m from the location of the sender,
(5)
The number could be generated and tested using
Equation (5), where SucPer is the success percentage that the
last segment must fulfill before agreeing on the values of
MaxB and MinB, and NeiN is the total number of neighbor
vehicles in NT.
If SucPer > Nmax%, this means that the last segment
holds enough number of potential forwarder vehicles. The
result of
is subtracted by one vehicle, because the last
segment also holds the preselected potential forwarder. If the
SucPer < Nmax%, this means that the area of the last
94
WCSIT 4 (7), 90 -100, 2014
the segments with the higher progress and high number of
vehicles with smaller segments give higher fitness function.
Where MaxB is the highest boundary (border) for the last
segment, and MinB is the minimum boundary (Borders) from
which the segment starts.
(7)
After performing Equation (7), the vehicle inserts it in a
Progress List (PL), which then helps the vehicle in making
quicker analysis and decisions (Table 2).
This means that the vehicles located in the area between
MinB and MaxB from the sender are considered as potential
forwarders of the emergency message, and would have to
wait to rebroadcast in case no vehicle forwards the message.
When the sender decides to broadcast an emergency
message, it should examine the number of neighbors within
the back end of the coverage area. If the number is more than
one, then this protocol could be carried out; if the number of
the neighbors is zero, then the sender broadcasts the message
without specifying any forwarder. If the number of the
neighbors is equal to one, then the sender broadcasts the
message and specifies the forwarder without adding any
detail about the boundaries.
TABLE 2: PROGRESS LIST.
Progress (m)
Length of
Segment
No. of Vehicles
Fitness
Function
520
90
15
86
610
90
15
101
700
120
30
175
820
80
15
153
900
0
1
0
To compute for the contention time, Equation (12) is
performed, where the vehicles with the largest distance from
the sender have the shortest contention time to wait before
testing the channel to rebroadcast. Each vehicle tests its
progress from the sender by dividing its current position on
the maximum distance computed by the sender. The result of
this equation gives the waiting time for the contending
vehicles inside the last segment, giving the opportunity for
all the vehicles inside the last segment to recover from the
failure of the chosen forwarder. Thus, the protocol increases
the probability of resending the emergency message
consequently, increasing the percentage of sending the
emergency message and reaching longer distances at the
same time.
After performing Equation (7) on all the vehicles in the
segments, the sender vehicle takes the upper boundary of the
segment scores the higher fitness function then the PSO
optimization is applied
Fitv = lBestv w + C1
rand1 (pBestv - lBestv) + C2
rand2 (gBestv - lBestv), (8) [20].
lBestv = pBestv + Fitv (9), [20].
Where W is random number between W: 0.1 to 0.5, C1=
2, C2= 2, rand: random number 0.1 to 1, pBest is the last
lBest computed by the vehicle. w is the inertia weight of the
particles, random 1 and random 2 are two uniformly
distributed random numbers in the range [0, 1], and C1 and
C2 are specific parameters which control the relative effect
of the individual and global best particles.
Enabling just the last segment to contend eliminates the
hidden node problem, because all the potential forwarders
have high probabilities to sense the rebroadcasted message
when a forwarder resends the message. All potential
forwarders are located in a small and limited area. The
probability of reception can reach 99% at short distances, but
it could be as low as 20% at half the distance of the
communication range [13].
The lBest for the vehicle is obtained from the fitness
function computed using (7), which represents the best area
(segment dimension). The results indicate the sufficient
number of vehicles depending on the sender analysis. Pbest
is the previous fitness function computed by the vehicle,
while the gBest is best fitness function computed by the
vehicle obtained from the analysis of the information from
CRNT obtained from our previous paper [7]. The CRNT
gives extended information received from other vehicles
located in the neighborhood of the sender, which reduces the
error possibility that the vehicle might make during the
channel dense location analysis because PSO depends on
taking the neighbor’s information and history. From the
CRNT, the vehicle can conclude another global fitness
function from the neighboring vehicles analysis, which
influences the current analysis done by the current vehicle
Sending Steps: The sender dispatches an emergency
message warning other vehicles about any potential danger.
The sender analyzes the danger and selects the code. The
sender creates NT for its neighbors, and then selects the next
forwarder depending on the distance of the farthest vehicle.
The sender analyzes the dense location computing the fitness
function using Equation (7). The sender analyzes the
information gained from neighbors about dense locations
from CRNT and concludes the gBest. The PSO algorithm is
then applied to obtain the MinB, which represents the lower
bound of the segment. The sender creates the message and
inserts the values derived from steps 4 and 8 in the message,
after which it broadcasts the message to the network
To compute for the boundaries of the last segment, the
sender carries out the following equations:
The following illustrates the calculations using Equation
(7), which computes the fitness function for each segment.
This formula ensures that the vehicles having high progress
from the sender and having a large number of vehicles in a
small area can produce better fitness function. This is
because vehicles concentrated in a small area and are located
(10)
(11)
95
WCSIT 4 (7), 90 -100, 2014
far from the sender vehicle have better opportunities to
rebroadcast the emergency message with little chance of
failure.
Tc is the contention time that the segment vehicle has to
wait before checking the system to see if the emergency
message has been rebroadcasted by other vehicle or not,
Tslot: is the system time slot.
From Table 2 and after employing Equation (7):
Receiving emergency message steps: This section
represents the steps for receiving the emergency message,
which must be done efficiently. The receiver accepts a
message then checks the code if it is 1 or 2. The receiver also
checks if the message has been received before and if its ID
is the same as the forwarder ID, then, it rebroadcasts the
message immediately. If the ID is not the same, the receiver
calculates the distance between its current position and the
sender and tests if the current location falls within MinB and
MaxB. The receiver then prepares to forward the message.
The best value for the fitness function is 175; this value
means that the lBest is 700, which represents the lower
boundary for that segment.
The gBest is taken from the CRNT, because this protocol
provides the sender vehicle with information from other
vehicles. This information also enables the sender to analyze
the channel depending on the information from the other
vehicle giving more accurate data about the network.
Rebroadcasting steps:
The rebroadcasting job is only assigned to a limited
number of vehicles. It is not appropriate to assign this job to
all the receiver vehicles, because this can lead to a broadcast
storm problem [22] or the hidden node problem. The first
forwarder should be the first candidate selected by the
sender, and the forwarding steps are as follows.
pBest is the channel analysis history of the network made
from the sender vehicle, lBest is the boundary of the fitness
function, and gBest is the best analysis from the neighbors.
lBest = 700, pBest= 690, gBest= 720.
Apply PSO equation (8)
700 x 0.1 + 2 x 0.65(690 – 700) + 2 x 0.7 (720 –
700) = 70 – 13 + 28 =95
lBest = 690 + 95 eq (4.10)
The forwarder waits for a random back off time
depending on its contention window. The back off time is
used to avoid channel collision. The forwarder then senses
the channel and tests whether or not this message has been
transmitted from others. If no other vehicle rebroadcasts, the
message forwarder reserves the channel and the forwarder
broadcasts the message. The contention window for the first
candidate forwarder is 15 µs.
lBest = 785.
The MinB for this example is 785 m and the MaxB is 900
m, as obtained from Equation (10). The results imply that the
vehicles between 785 m and 900 m from the sender can be
considered as potential forwarders; they would have to
contend by means of the time to rebroadcast and be able to
overcome the preselected forwarder failure.
Each vehicle inside the last non-empty segment is
required to contend before trying to send the emergency
message and this could be done using the following steps.
The forwarder computes the contention window using
Equation (12). The vehicle contends for a random back off
time depending on the contention window. The vehicle
senses the channel and determines whether or not this
message has been transmitted from others. If no other vehicle
rebroadcasts the message, the vehicle reserves the channel
and broadcasts the message.
Receiving a message
Vehicles in VANET receive messages all the time. These
messages could be beacon, emergency, or service messages.
Each message should be analyzed to determine the level of
importance involved. If the message is holding safety critical
information, then the message code should be 1 or 2, and the
message is given higher priority for processing. The receiver
vehicle also has to ensure that this message has not been
received before to eliminate duplication. The receiver checks
the forwarder ID. If the current receiver ID is the same as
that of the forwarder, the receiver rebroadcasts the message
immediately. If the receiver is not a forwarder, then it must
compute the distance between its position and that of the
sender to determine if the receiver is located within the last
non-empty segment. If the receiver is located in the last nonempty segment, it starts contending and prepares itself to
rebroadcast. Each vehicle inside the last non-empty segment
must contend the time and compute the CW time (Waiting
time) using Equation (12). CW depends on the progress, and
the vehicles having the largest progress from the sender has
the shortest contention time before testing the channel to
make the rebroadcast.
PCBB and CBB both have the same goal of increasing
the percentage reception of the emergency information. The
main difference between them, however, is the potential
forwarder selection, where CBB depends on choosing the
last non-empty segment boundaries depending on the
number of the vehicles in the segment against a predefined
threshold, whereas the PCBB depends on selecting the
boundaries on vehicle saturation areas and utilizes the PSO
intelligent technique.
IV. SIMULATION
Simulation Setup
In order to test correctness of our protocol we made the
simulation using the commercial program Matlab®, the
distribution used is Nakagami distribution.
(12)
96
WCSIT 4 (7), 90 -100, 2014
Parameters used in our simulation are summarized in
table 1; all the simulations in this paper will adopt these
parameters.
PCBB can achieve better performance past the first 1000 m
distance representing the DSRC communication range for
the sender. The signal gets very weak in the last meters of the
sender communication range, but when the forwarder
rebroadcasts the message, the signal becomes stronger and
reaches greater distances.
We made our simulation for 10s including 200 vehicles
in 2 km road consisting of 3 lanes.
Simulation Parameters
Another difference between CBB/PCBB and EMDV is
that after several tries, CBB and PCBB never fail to
rebroadcast the emergency message, but EMDV sometimes
fails to do so, proving the effectiveness of the CBB protocol.
The PCBB protocol has also been shown to select forwarders
more carefully than CBB, which depends on a threshold,
while PCBB depends on traffic saturation and progress and
on the analysis made by neighboring vehicles. Furthermore,
PSO adopts the PSO intelligent algorithm, which takes other
vehicles' analysis into consideration, allowing for more
accurate selection of preselected forwarders.
Parameters used in the simulation experiment are
summarized in table 3; all the simulations in this paper will
adopt these parameters.
TABLE 3: SIMULATION CONFIGURATION PARAMETERS
Parameter
Radio
propagation
model
Value
Nakagami-m, m
=3
IEEE 802.11p data rate
PLCP header length
Symbol duration
Noise floor
SNR
6Mbps
8 μs
8 μs
-99dBm
10 - 40 dB
CW Min
CW Max
Slot time
SIFS time
DIFS time
Message size
Beacon Message Rate
Number of Vehicles
Road Length
Car Speed
Simulation Time
Road Type
Number of lanes
Neighbor entry size
15 μs
1023 μs
16 μs
32 μs
64 μs
512 bytes
10 Message / s
200
2 KM
20km – 120km
10 s
Highway
3 lanes
15 Bytes
Description
Model m=3 is
fixed
value,
recommended by
[13]
Fixed value
Fixed value
Fixed value
Fixed value
Adjustable to add
noise to the signal
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Fixed value
Figure 7: Probability of message reception of emergency message with
respect to the distance to the sender.
Figure 8 shows the message delay for CBB and EMDV
compared with PCBB emergency message delay. The
simulation computes the delay for broadcasting and
rebroadcasting of the original message, showing that the
EMDV during the time has a slightly higher delay than CBB
but not exceeding 50 µs. The delay shows a slight increase at
about 20 µs 1000 m away from the sender where the
rebroadcast starts to take effect. If the CBB has a shorter
delay starting from this point, it means that its rebroadcast
efficiency and decisions are made faster than those in
EMDV. PCBB has a slightly shorter delay (about 10 µs
shorter than CBB and 30 µs shorter than EMDV at the ninth
second), because the PSO is an intelligent technique that has
quick performance and response [23]. In safety systems with
a highly mobile network like VANET, a few microseconds
are critical in saving life or avoiding danger.
V. RESULTS
In order to enhance emergency message dissemination in
VANET, two contention- and position-based protocols have
been proposed and implemented, namely, CBB and PCBB
(an enhancement of the proposed CBB). In this section, the
EMDV protocol, which is the outcome of the NOW project
[13, and 22], is compared with the CBB and PCBB. The
results of the experiment are shown in Figures 7, 8, and 9.
The test performed concentrated on the probability of
emergency message reception, channel collision, and the
delay that the protocols may cause. The EMDV and DFPAV
protocols, both widely used in VANET today, are the results
of the NOW project, which is a collaboration between
Mercedes-Benz and Karlsruhe University [22]. Figure 7
shows the simulation results for the proposed CBB and
PCBB protocols. These have been simulated and tested in
terms of probability of emergency message reception;
afterwards, their performances are compared with that of the
EMDV protocol. The results show that all the protocols can
increase the performance and probability of emergency
message reception; more noteworthy is the fact that CBB and
97
WCSIT 4 (7), 90 -100, 2014
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
Figure 8: Delay measured after sending the emergency message with
respect to distance.
[7]
Figure 9 shows the collision produced by the three
protocols, all of which generated the same collision when
broadcasting emergency information. It is worth noting that
the collisions produced by CBB, PCBB and EMDV at the
beginning of the experiment do not increase. However, after
a period of time, sending a large number of emergency
messages resulted in an increase in the number of collision
for all the three protocols, with the difference between them
reaching 1% at the ninth second.
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
Figure 9: Collision measured after sending the emergency message.
[16]
VI. CONCLUSION
This Research has proposed the PCBB aiming to improve
road safety by achieving fast and efficient emergency
message transmission and delivery, utilizing the efficient and
newest intelligent technique (PSO), which helped to make
more accurate analysis and performance, and increased the
percentage of the emergency message reception without
affecting the channel collision.
[17]
[18]
[19]
98
Ghassan Samara, Wafaa AH Al-Salihy, R Sures, "Security Issues and
Challenges of Vehicular Ad Hoc Networks (VANET)", 4th
International Conference on New Trends in Information Science and
Service Science (NISS), 2010, 393-398.
Ghassan Samara, Wafaa AH Al-Salihy, R Sures, "Security Analysis
of Vehicular Ad Hoc Nerworks (VANET)", Second International
Conference on Network Applications Protocols and Services
(NETAPPS), 2010, 55-60.
Who,
W.
H.
(2011),
World
Health
Organization,
http://www.who.int/roadsafety/decade_of_action/plan/en/index.html,
visited on 30 April, 2014.
M Raya, D. J., P Papadimitratos, I Aad, JP Hubaux (2006), Certificate
Revocation in Vehicular Networks. Laboratory for computer
Communications and Applications (LCA) School of Computer and
Communication Sciences, EPFL, Switzerland.
Worldometers
(2011),
real
time
world
statistics,
www.worldometers.info, visited on 30 April, 2014.
Ghassan Samara, WAHA Alsalihy, S Ramadass, "Increase
Emergency Message Reception in VANET", Journal of Applied
Sciences, 2011, Volume 11, Pages 2606-2612.
Ghassan Samara, Wafaa Alsalihy, Sureswaran Ramadass,” Increasing
Network Visibility Using Coded Repitition Beacon Piggybacking”,
World Applied Sciences Journal (WASJ), Volume 13, Number 1, pp.
100 - 108, 2011.
Ching-Yi, Y. & Shou-Chih, L. (2010) Street Broadcast with Smart
Relay for Emergency Messages in VANET. 24th International
Conference on Advanced Information Networking and Applications
Workshops (WAINA), 323-328, IEEE.
Qiong, Y. & Lianfeng, S. (2010) A Multi-Hop Broadcast scheme for
propagation of emergency messages in VANET. 12th IEEE
International Conference on Communication Technology (ICCT),
1072-1075, IEEE.
Biswas, S., Tatchikou, R.,
Dion, F. (2006), Vehicle-to-vehicle
wireless communication protocols for enhancing highway traffic
safety, IEEE Communications Magazine, 44, 74-82, IEEE.
Torrent-Moreno, M. (2007), Inter-vehicle communications: assessing
information dissemination under safety constraints, 4th Annual
Conference on Wireless on Demand Network Systems and Services,
WONS '07, 59 – 64, IEEE.
Torrent-Moreno, M., Mittag, J., Santi, P. & Hartenstein, H. (2009),
Vehicle-to-Vehicle Communication: Fair transmit power control for
safety-critical information. Transactions on Vehicular Technology,
58, 3684-3703, IEEE
Torrent-Moreno, M. (2007), Inter-Vehicle Communications:
Achieving Safety in a Distributed Wireless Environment-Challenges,
Systems and Protocols, (Ph.D Paper), Universitatsverlag Karlsruhe,
ISBN: 978-3-86644-175-0.
Füßler, H., Widmer, J., Kasemann, M., Mauve, M. & Hartenstein, H.
(2003), Contention-based forwarding for mobile ad hoc networks, Ad
Hoc Networks, 1 (2003), 351-369.
Briesemeister, L., Schafers, L., Hommel, G. (2000), Disseminating
messages among highly mobile hosts based on inter-vehicle
communication, IEEE Intelligent Vehicles Symposium, IV 2000, 522
- 527, IEEE.
Korkmaz, G., Ekici, E., Özgüner, F. & Özgüner, Ü. (2004), Urban
multi-hop broadcast protocol for inter-vehicle communication
systems, 1st ACM international workshop on Vehicular ad hoc
networks, ACM.
Fasolo, E., Zanella, A., Zorzi, M. (2006), An effective broadcast
scheme for alert message propagation in vehicular ad hoc networks,
IEEE Int. Conf. on Communications ICC'06, 3960 – 3965, IEEE.
Zorzi, M. & Rao, R. R. (2003), Geographic random forwarding
(GeRaF) for ad hoc and sensor networks: energy and latency
performance, IEEE transactions on Mobile Computing, 349-365,
IEEE.
IEEE (2005), White Paper: DSRC technology and the DSRC Industry
consortium (DIC) prototype team.
WCSIT 4 (7), 90 -100, 2014
[20] Neo
Project,
(2014),
http://neo.lcc.uma.es/staff/jamal/portal/?q=content/particle-swarmoptimization-pso, Malaga University, visited on 10 March 2014.
[21] Ni, S. Y., Tseng, Y. C., Chen, Y. S. & Sheu, J. P. (1999a), The
broadcast storm problem in a mobile ad hoc network, 5th annual
ACM/IEEE international conference on Mobile computing and
networking, 151-162, ACM.
[22] Now, (2011). Network on Wheels project, http://www.network-onwheels.de/vision.html, accessed 5 May 2011.
[23] Mendes, R. (2004). Population Topologies and Their Influence in
Particle Swarm Performance (PhD thesis). Universidade do Minho.
Wait contention time
If channel = idle
{
If no rebroadcast
{
Rebroadcast ( )
} % end if
Else
{
Backoff = backoff -1
} % end if
} % end while
} % end if
} % end if
} % end if
}% end procedure
APPENDIX
Procedure DetectDanger ( )
{
Gather neighbor information
Select main forwarder
MaxB = forwarderlocation – senderlocation.
(MinB) = PSO ( )
Insert Emergency information in the message
Send Message
} % end procedure
Procedure Rebroadcast
Rebroadcast emergency message
End % procedure
Procedure Computevehicleconcentration ( )
{
Mindist = Computemindist ( )
Arrange NT descending
For i= 0 NT size -1
{
Take the average between two successive vehicles
If average < 2 * (last average) or number of vehicles compared
is 1 % if this vehicle is the first vehicle
{
Add to current segment % add the current vehicle to this
segment
Segment vehicles = segment vehicles + 1 % increase the number
of vehicles by 1
Vehicle location = location (i) %takes the location of the current
vehicle
Segment (segment count, dist) = Vehicle location - first element
location % compute the width of the
Procedure PSO ( )
{
Currentsegment = Computevehicleconcentration ( ) % to
compute the current concentration of vehicles
pBest = cfitness
cfitness = 0
For i=0 to Currentsegment size -1 % calculate the best result
{
% fitness function
If Fitness > cfitness
cfitness = Fitness
} % end if
} % end for
lBest = cfitness
segment
segment (segment count, Segment vehicles) = Segment vehicles
% to store the number of vehicles for the
neighborsegment = Computevehicleconcentration ( ) % to
compute the current concentration of vehicles
gfitness = 0
For x=0 to Currentsegment size -1 % calculate the best result
{
current segment
}
Else if the new average is more than the double of previous
value
{
segment count = segment count + 1
Vehicle location = location (i)
Segment vehicles = 1
Segment (segment count, progress) = Vehicle location
}
Return (segment)
}
% fitness function
If Fitness > cfitness
gfitness = Fitness
} % end if
} % end for
gBest = gfitness
}% end procedure
Procedure ReceiveEmerMessage ( )
{
If code = 1 or code = 2
{
If preselctedforwarder
{
Rebroadcast ( )
Else if candidate forwarder
{
Compute CW
Choose random backoff
While back <> 0
{
Figure A.1: Psedu-code Particle Swarm Optimization Contention Based
Broadcast Protocol (PCBB).
The procedure DetectDanger works when the vehicle
detects any danger, the first step for the sender is to order the
neighbors' information in NT, and select the first forwarder,
afterwards. It calls the PSO procedure that implements the
PSO algorithm to select the vehicles that overcome the
preselected forwarder’s failure.
99
WCSIT 4 (7), 90 -100, 2014
The procedure ReceiveEmerMessage works when the
vehicle receives an emergency message and checks if the
receiver is a forwarder or not.
The procedure Computevehicleconcentration analyzes
the neighbors to discover the location of the vehicles'
concentration.
100
| 9cs.NE
|
GENERATING THE IDEALS DEFINING UNIONS OF SCHUBERT VARIETIES
arXiv:1405.2945v1 [math.AG] 12 May 2014
ANNA BERTIGER
A BSTRACT. This note computes a Gröbner basis for the ideal defining a union of Schubert varieties.
More precisely, it computes a Gröbner basis for unions of schemes given by northwest rank conditions on the space of all matrices of a fixed size. Schemes given by northwest rank conditions include
classical determinantal varieties and matrix Schubert varieties–closures of Schubert varieties lifted
from the flag manifold to the space of matrices.
1. I NTRODUCTION
We compute a Gröbner basis, and hence ideal generating set, for the ideal defining a union of
schemes each given by northwest rank conditions with respect to an “antidiagonal term order.”
A scheme defined by northwest rank conditions is any scheme whose defining equations are of
the form “all k × k minors in the northwest i × j sub-matrix of a matrix of variables,” where i, j,
and k can take varying values. These schemes represent a generalization of classical determinantal
varieties–those varieties with defining equations all (r+1)×(r+1) minors of a matrix of variables.
One geometrically important collection of schemes defined by northwest rank conditions is the set
of matrix Schubert varieties. Matrix Schubert varieties are closures of the lift of Schubert varieties
from the complete flag manifold to matrix space [Ful92]. In general, a matrix Schubert variety
for a partial permutation π is the subvariety of matrix space given by the rank conditions that
the northwest i × j sub-matrix must have rank at most the number of 1s in the northwest i × j
sub-matrix of the partial permutation matrix for π. Notice that the set of matrix Schubert varieties
contains the set of classical determinantal varieties, which are the zero locus of all minors of a
fixed size on the space of all matrices of fixed size.
Matrix Schubert varieties associated to honest, that is non-partial, permutations are the closures
of the lifts of the corresponding Schubert varieties in the flag manifold, B− \GLn . If Xπ is the matrix
Schubert variety for an honest permutation π the projection
{full rank matrices} B− \ GLn C = F`Cn
sends Xπ ∩ GLn C onto the Schubert variety Xπ ⊆ F`Cn . Schubert varieties, orbits of B+ , stratify
F`Cn and give a basis for H∗ (F`Cn ). It is this application that led to the introduction of matrix
Schubert varieties in [Ful92]. Knutson and Miller showed that matrix Schubert varieties have a
rich algebro-geometric structure corresponding to beautiful combinatorics [KM05]. Fulton’s generators are a Gröbner basis with respect to any antidiagonal term order and their initial ideal is
the Stanley-Reisner ideal of the “pipe dream complex.” Further, Knutson and Miller show that
the pipe dream complex is shellable, hence the original ideal is Cohen-Macaulay. Pipe dreams,
the elements of the pipe dream complex, were originally called RC graphs and were developed
by Bergeron and Billey [BB93] to describe the monomials in polynomial representatives for the
classes corresponding to Schubert varieties in H∗ (F`Cn ).
The importance of Schubert varieties, and hence matrix Schubert varieties, to other areas of
geometry has become increasing evident. For example, Zelevinsky [Zel85] showed that certain
quiver varieties, sequences of vector space maps with fixed rank conditions, are isomorphic to
Date: May 14, 2014.
1
Schubert varieties. Knutson, Miller and Shimozono, [KMS06] produce combinatorial formulae for
quiver varieties using many combinatorial tools reminiscent of those for Schubert varieties.
1.1. Notation and Background. Much of the background surveyed here can be found in [MS05].
Let B− (respectively B+ ) denote the group of invertible lower triangular (respectively upper triangular) n × n matrices. Let M = (mi,j ) be a matrix of variables. In what follows π will be a possibly
partial permutation, written in one-line notation π(1) . . . π(n), with entries for π(i) undefined are
written ?. We shall write permutation even when we mean partial permutation in cases where
there is no confusion. A matrix Schubert variety Xπ is the closure B− πB+ in the affine space of all
matrices, where π is a permutation matrix and B− and B+ act by downward row and rightward
column operations respectively. Notice that for π an honest permutation Xπ is the closure of the
lift of Xπ = B− \B− πB+ ⊆ B− \GLn C to the space of n × n matrices.
The Rothe diagram of a permutation is found by looking at the permutation matrix and crossing
out all of the cells weakly below, and the cells weakly to the right of, each cell containing a 1. The
remaining empty boxes form the Rothe diagram. The essential boxes [Ful92] of a permutation are
those boxes in the Rothe diagram that do not have any boxes of the diagram immediately south
or east of them. The Rothe diagrams for 2143 and 15432 are given in Figure 1.1. In both cases the
essential boxes are marked with the letter e.
1
e
e
1
e
1
e
e
1
1
1
1
1
1
F IGURE 1.1. The Rothe diagrams and essential sets of 2143 (left) and 15432 (right).
The rank matrix of a permutation π, denoted r(π), gives in each cell r(π)ij the rank of the i × j
northwest-justified sub-matrix of the permutation matrix for π. For example, the rank matrix of
15432 is
1 1 1 1 1
1 1 1 1 2
1 1 1 2 3 .
1 1 2 3 4
1 2 3 4 5
Theorem 1.1 ( [Ful92]). Matrix Schubert varieties have radical ideal I(Xπ ) = Iπ given by determinants
representing conditions given in the rank matrix r(π), that is, the (r(π)ij + 1) × (r(π)ij + 1) determinants
of the northwest i × j sub-matrix of a matrix of variables. In fact, it is sufficient to impose only those rank
conditions r(π)ij such that (i, j) is an essential box for π.
Hereafter we call the determinants corresponding the to essential rank conditions, or the analogous determinants for any ideal generated by northwest rank conditions, the Fulton generators.
One special form of ideal generating set is a Gröbner basis. To define a Gröbner basis we set
a total ordering on the monomials in a polynomial ring such that 1 ≤ m and m < n implies
mp < np for all monomials m, n and p. Let init f denote the largest monomial that appears in the
polynomial f. A Gröbner basis for the ideal I is a set {f1 , . . . fr } ⊆ I such that init I := hinit f : f ∈
Ii = hinit f1 , . . . init fr i. Notice that a Gröbner basis for I is necessarily a generating set for I.
2
The antidiagonal of a matrix is the diagonal series of cells in the matrix running from the most
northeast to the most southwest cell. The antidiagonal term (or antidiagonal) of a determinant
is the product of the entries in the antidiagonal. For example, the antidiagonal of ac db is the
cells occupied by b and c, and correspondingly, in the determinant ad − bc the antidiagonal term
is bc. Term orders that select antidiagonal terms from a determinant, called antidiagonal term
orders have proven especially useful in understanding ideals of matrix Schubert varieties. There
are several possible implementations of an antidiagonal term order on an n×n matrix of variables,
any of which would suit the purposes of this paper. One example is weighting the top right entry
highest and decreasing along the top row before starting deceasing again at the right of the next
row; monomials are then ordered by their total weight.
Theorem 1.2 ( [KM05]). The Fulton generators for Iπ form a Gröbner basis under any antidiagonal term
order.
Typically we will denote the cells of a matrix that form antidiagonals by A or B. In what follows if A is the antidiagonal of a sub-matrix of M we will use the notation det(A) to denote the
determinant of this sub-matrix. We shall be fairly liberal in exchanging antidiagonal cells and the
corresponding antidiagonal terms, thus, for any antidiagonal term order, A = init det(A).
1.2. Statement of Result. Let I1 , . . . Ir be ideals defined by northwest rank conditions. We will
produce a Gröbner basis, and hence ideal generating set, for I1 ∩ · · · ∩ Ir . For each list of antidiagonals A1 , . . . , Ar , where Ai is the antidiagonal of a Fulton generator of Ii , we will produce a
Gröbner basis element gA1 ,...,Ar for ∩Ii . The generators gA1 ,...,Ar will be products of determinants,
though not simply the product of the r determinants corresponding to the Ai . For a fixed list of
antidiagonals A1 , . . . , Ar , build the generator gA1 ,...,Ar by:
(1) Begin with gA1 ,...,Ar = 1
(2) Draw a diagram with a dot of color i in each box of Ai and connect the consecutive dots of
color i with a line segment of color i.
(3) Break the diagram into connected components. Two dots are connected if they are either
connected by lines or are connected by lines to dots that occupy the same box.
(4) For each connected component, remove the longest series of boxes B such that there is
exactly one box in each row and column and the boxes are all in the same connected component. If there is a tie use the most northwest of the longest series of boxes. Note that B
need not be any of A1 , . . . , Ar . Multiply gA1 ,...,Ar by det(B). Remove this antidiagonal from
the diagram of the connected component, break the remaining diagram into components
and repeat.
Theorem 1.3. {gA1 ...Ar : Ai is an antidiagonal of a Fulton generator of Ii , 1 ≤ i ≤ r} form a Gröbner
basis, and hence a generating set, for ∩ri=1 Ii .
1.3. Acknowledgements. This work constitutes a portion of my PhD thesis completed at Cornell
University under the direction of Allen Knutson. I wish to thank Allen for his help, advice and
encouragement in completing this project. Thanks also go to Jenna Rajchgot for helpful discussions in the early stages of this work. I’d also like to thank the authors of computer algebra system
Macaulay2, [GS] which powered the computational experiments nessecary to do this work. I’m especially grateful to Mike Stillman who patiently answered many of my Macaulay2 questions over
the course of this work. Kevin Purbhoo gave very helpful comments on drafts of this manuscript
for which I cannot thank him enough.
2. E XAMPLES
We delay the proof of Theorem 1.3 to Section 3 and first give some examples of the generators
produced for given sets of antidiagonals. These examples are given by pictures of the antidiagonals on the left and corresponding determinantal equations on the right. Note that we only give
3
particular generators, rather than entire generating sets, which might be quite large. We then give
entire ideal generating sets for two smaller intersections.
If r = 1 then for each Fulton generator with antidiagonal A the algorithm produces the generator gA = det(A). Therefore, if we intersect only one ideal the algorithm returns the original set of
Fulton generators. The generator for the antidiagonal shown is exactly the determinant of the one
antidiagonal pictured:
m1,1 m1,2 m1,4
m3,1 m3,2 m3,4
m4,1 m4,2 m4,4
.
The generator for two disjoint antidiagonals is the product of the determinants corresponding
to the two disjoint antidiagonals:
m1,1 m1,2
m2,1 m3,2
m1,1 m1,2 m1,4
m3,1 m3,2 m3,4
m4,1 m4,2 m4,4
.
In general, if A1 , . . . Ar are disjoint antidiagonals then the then the algorithm looks at each Ai separately as they are part of separate components and the result is that gA1 ,...Ar = det(A1 ) · · · det(Ar ).
If A1 , . . . Ar overlap to form one antidiagonal X then the last step of the algorithm will occur
only once and will produce gA1 ,...Ar = det(X). For example,
m1,1
m2,1
m3,1
m4,1
m1,2
m2,2
m3,2
m4,2
m1,3
m2,3
m3,3
m4,3
m1,4
m2,4
m3,4
m4,4
.
In this example, there are two longest possible antidiagonals, the three cells occupied by the
green dots and the three cells occupied by the red dots. The ones occupied by the green dots are
more northwest, hence the generator for the three antidiagonals shown below is
4
m1,1 m1,2
m2,1 m2,2
m1,2 m1,3 m1,3
m2,2 m2,3 m2,3
m3,2 m3,3 m3,4
m4,2 m4,3
m5,2 m2,2
.
In the picture below, the longest possible anti diagonal uses all of the cells in the green anti
diagonal but only some of the cells in the red antidiagonal, however, there is only one possible
longest antidiagonal. Thus the generator is
m1,1 m1,2
m2,1 m2,2
m1,1
m2,1
m3,1
m4,1
m1,2
m2,2
m3,2
m4,2
m1,4
m2,4
m3,4
m4,4
m1,5
m2,5
m3,5
m4,5
m5,1
.
We now give two examples where the complete ideals are comparatively small. Firstly, we
calculate I(X231 ∪ X312 ) = I(X231 ) ∩ I(X312 ). I(X231 ) = hm1,1 , m2,1 i and I(X312 ) = hm1,1 , m1,2 i. The
antidiagonals and corresponding generators are shown below with antidiagonals from generators
of I(X231 ) shown in red and antidiagonals of generators of I(X312 ) shown in blue. Note that the
antidiagonals are only one cell each in this case.
m1,1
m1,1 m1,2
m1,1 m2,1
m1,2 m2,1
Theorem 1.3 results in
I(X231 ∪ X312 ) = I(X231 ) ∩ I(X312 ) = hm1,1 , m1,1 m1,2 , m1,1 m2,1 , m1,2 m2,1 i.
As a slightly larger example, consider I(X1423 ∪ X1342 ) = I(X1423 ) ∩ I(X1342 ). These generators
are given below in the order that the antidiagonals are displayed reading left to right and top to
bottom. The antidiagonals for I(X1423 are shown in red while the antidigaonals I(X1342 ) are shown
in blue. for Note that the full 4 × 4 grid is not displayed, but only the northwest 3 × 3 portion
where antidiagonals for these two ideals may lie.
5
Here Theorem 1.3 produces
m1,1
m1,1 m1,2
,
m2,1
m2,1 m2,2
m1,1 m1,2
m1,1
+
m3,1 m2,2
m1,1 m1,2
m1,2
+
m2,1 m3,2
m2,2
m1,2
m2,2
m1,1
m1,1 m1,3
,
m2,1 m2,3
m1,3
m2,2
m2,1 m2,2
m3,1 m2,2
m3,1 m1,2
,
m2,1 m2,2
m1,1 m1,3
m3,1 m3,2
m1,1 m1,2
,
m3,1 m2,3
m1,1 m1,3
,
m3,1 m2,3
m1,2 m1,3
m2,2 m2,2
m2,1 m2,2
m3,1 m2,3
m1,1 m1,2 m1,3
, m2,1 m2,2 m2,3
m3,1 m3,2 m3,4
3. P ROOF OF T HEOREM 1.3
We now prove the main result of this paper, Theorem 1.3, which states that the gA1 ,...,Ar generate
I1 ∩ · · · ∩ Ir .
We begin with a few fairly general statements:
Theorem 3.1 ( [Knu]). If {Ii : i ∈ S} are ideals generated by northwest rank conditions then init(∩i∈S Ii ) =
∩i∈S (init Ii ).
Lemma 3.2 ( [KM05]). If J ⊆ K are homogeneous ideals in a polynomial ring such that init J = init K
then J = K.
Lemma 3.3. Let IA and IB be ideals that define schemes of northwest rank conditions and let det(A) ∈ IA
and det(B) ∈ IB be determinants with antidiagonals A and B respectively such that A ∪ B = X and
A ∩ B 6= ∅. Then det(X) is in IA ∩ IB .
Proof. Let VX = V(det(X)), VA = V(IA ) and VB = V(IB ) be the varieties corresponding to the ideals
hdet(X)i, IA and IB . It is enough to show that VA ⊆ VX and VB ⊆ VX .
We will show that given a matrix with antidiagonal X with a sub-matrix with antidiagonal
A ⊆ X where the sub-matrix northwest of the cells occupied by A has rank at most length(A) − 1
then the full matrix has rank at most length(X) − 1. The corresponding statement for sub-matrix
with antidiagonal B can be proven by replacing A with B everywhere.
The basic idea of this proof is that we know the rank conditions on the rows and columns northwest of those occupied by A. The rank conditions given by A then imply other rank conditions as
adding either a row or a column to a sub-matrix can increase its rank by at most one.
6
column t + 1 column c
rank at most l
northwest of row k − t column c
rank at most l + k − c ≤ k − t − 2
northwest of column k row k − t
rank at most k − 2
row k − t
northwest of column k row k
row k
F IGURE 3.1. The proof of Lemma 3.3. The antidiagonal cells in A are marked in
black and the antidiagonal cells in X − A ⊆ B are marked in white.
Let k be the number of rows, also the number of columns in the antidiagonal X. Let the length
of A be l + 1, so the rank condition on all rows and columns northwest of those occupied by A
is at most l. Assume that the rightmost column of A is c and the leftmost column of A is t + 1.
Notice that this implies that the bottom row occupied by A is k − t, as the antidiagonal element in
column t + 1 is in row k − t. Thus, the northwest (k − t) × c of matrices in VA has rank at most l.
Notice c ≥ (t+1)+(l+1), with equality if A occupies a continuous set of columns, so matrices in
VA have rank at most l in the northwest (k−t)×(t+l+2). Adding k−c ≤ k−(n−t−l−2) columns to
this sub-matrix gives a (k−t)×k sub-matrix with rank at most l+k−c ≤ r+(k−t−l−2) = k−t−2.
Further, by the same principle, moving down t rows, the northwest k × k, i.e. the whole matrix
with antidiagonal X, has rank at most k − t − 2 + t = k − 2, hence has rank at most k − 1 and so is
in VX .
For a visual explanation of the proof of Lemma 3.3 see Figure 3.1.
Lemma 3.4. gA1 ,...,Ar ∈ Ii for 1 ≤ i ≤ r and hence
hgA1 ,...,Ar : Ai ranges over all antidiagonals for Fulton generators of Ii i ⊆ ∩ri=1 Ii .
Proof. Fix i. Let S be the first antidiagonal containing a box occupied by a box contained in Ai
added to gA1 ,...,Ar . We shall show that det(S) is in Ii and hence gA1 ,...,Ar ∈ Ii as it is a multiple of
det(S). If Ai ⊆ S then det(S) ∈ Ii either because S = Ai or S ( Ai in which case we apply Lemma
3.3. Otherwise, |S| ≥ |Ai | and S is weakly to the northwest of Ai . Therefore, there is a subset B
of S such that |B| = |Ai |, and B is weakly northwest of Ai . Hence, B is an antidiagonal for some
determinant in Ii , and again by Lemma 3.3 det(S) ∈ Ii .
Lemma 3.5. init gA1 ,...,Ar = A1 ∪ · · · ∪ Ar under any antidiagonal term order.
Proof. init gA1 ,...,Ar is a product of determinants, with collective antidiagonals A1 ∪ · · · ∪ Ar .
When we combine Lemma 3.5 and Theorem 3.1 we see that inithgA1 ,...Ar i = init(∩Ii ). Then,
Lemmas 3.2 and 3.4 combine to complete the proof of Theorem 1.3.
7
Note that Theorem 1.3 may produce an oversupply of generators. For example, if I1 = I2 , then
inputting the same set of p Fulton generators twice results in a Gröbner basis of p2 polynomials
for I1 ∩ I2 = I1 = I2 .
R EFERENCES
Nantel Bergeron and Sara Billey, RC-graphs and Schubert polynomials, Experiment. Math. 2 (1993), no. 4, 257–
269. MR 1281474 (95g:05107)
[Ful92] William Fulton, Flags, Schubert polynomials, degeneracy loci, and determinantal formulas, Duke Math. J. 65 (1992),
no. 3, 381–420. MR 1154177 (93e:14007)
[GS]
Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in algebraic geometry,
Available at http://www.math.uiuc.edu/Macaulay2/.
[KM05] Allen Knutson and Ezra Miller, Gröbner geometry of Schubert polynomials, Ann. of Math. (2) 161 (2005), no. 3,
1245–1318. MR 2180402 (2006i:05177)
[KMS06] Allen Knutson, Ezra Miller, and Mark Shimozono, Four positive formulae for type A quiver polynomials, Invent.
Math. 166 (2006), no. 2, 229–325. MR 2249801 (2007k:14098)
[Knu]
Allen Knutson, Frobenius splitting, point-counting and degeneration, Preprint, arXiv:0911.4941v1.
[MS05] Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, vol. 227,
Springer-Verlag, New York, 2005. MR 2110098 (2006d:13001)
[Zel85] A. V. Zelevinskiı̆, Two remarks on graded nilpotent classes, Uspekhi Mat. Nauk 40 (1985), no. 1(241), 199–200.
MR 783619 (86e:14027)
[BB93]
8
| 0math.AC
|
An O(log k log2 n)-competitive Randomized Algorithm for the
k-Sever Problem
arXiv:1510.07773v1 [cs.DS] 27 Oct 2015
Wenbin Chen∗†‡
Abstract
In this paper, we show that there is an O(log k log2 n)-competitive
randomized algorithm for the k-sever problem on any metric space
with n points, which improved the previous best competitive ratio
O(log2 k log3 n log log n) by Nikhil Bansal et al. (FOCS 2011, pages 267276).
Keywords: k-sever problem; Online algorithm; Primal-Dual method;
Randomized algorithm;
1
Introduction
The k-sever problem is to schedule k mobile servers to serve a sequence of requests in a metric
space with the minimum possible movement distance. In 1990, Manasse et al. introduced the ksever problem as a generalization of several important online problems such as paging and caching
problems [29] ( Its conference version is [28]), in which they proposed a 2-competitive algorithm
for the 2-sever problem and a n − 1-competitive algorithm for the n − 1 sever problem in a n-point
metric space. They still showed that any deterministic online algorithm for the k-sever problem is
of competitive ratio at least k. They proposed the well-known k-sever conjecture: for the k-sever
problem on any metric space with more than k different points, there exists a deterministic online
algorithm with competitive ratio k.
It was in [29] shown that the k-sever conjecture holds for two special cases: k = 2 and n = k + 1.
The k-sever conjecture also holds for the k-sever problem on a uniform metric. The special case
of the k-sever problem on a uniform metric is called the paging (also known as caching) problem.
Slator and Tarjan have proposed a k-competitive algorithm for the paging problem [31]. For some
other special metrics such as line, tree, there existed k-competitive online algorithms. Yair Bartal
∗
Email:[email protected]
Department of Computer Science, Guangzhou University, P.R. China
‡
State Key Laboratory for Novel Software Technology, Nanjing University, P.R. China
†
1
and Elias Koutsoupias show that the Work Function Algorithm for the k-sever problem is of kcompetitive ratio in the following special metric spaces: the line, the star, and any metric space
with k + 2 points [16]. Marek Chrobak and Lawrence L. Larmore proposed the k-competitive
Double-Coverage algorithm for the k-sever problem on trees [21].
For the k-sever problem on the general metric space, the k-sever conjecture remain open. Fiat
et al. were the first to show that there exists an online algorithm of competitive ratio that depends
only on k for any metric space: its competitive ratio is Θ((k!)3 ). The bound was improved later
by Grove who showed that the harmonic algorithm is of competitive ratio O(k2k ) [25]. The result
was improved to (2k log k) by Y.Bartal and E. Grove [14]. A significant progress was achieved by
Koutsoupias and Papadimitriou, who proved that the work function algorithm is of competitive
ratio 2k − 1 [27].
Generally, people believe that randomized online algorithms can produce better competitive
ratio than their deterministic counterparts. For example, there are several O(log k)-competitive
algorithms for the paging problem and a Ω(log k) lower bound on the competitive ratio in [24, 30,
1, 4]. Although there were much work [17, 13, 15], the Ω(log k) lower bound is still best lower bound
in the randomized case. Recently, N. Bansal et al. propose the first polylogaritmic-competitive
randomized algorithm for the k-sever problem on a general metrics spcace [3]. Their randomized
algorithm is of competitive ratio O(log2 k log3 n log log n) for any metric space with n points, which
improves on the deterministic 2k − 1 competitive ratio of Koutsoupias and Papadimitriou whenever
n is sub-exponential k.
For the k-server problem on the general metric space, it is widely conjectured that there is an
O(log k)-competitive randomized algorithm, which is called as the randomized k-server conjecture.
For the paging problem ( it corresponds to the k-sever problem on a uniform metric), there is
O(log k)-competitive algorithms [24, 30, 1]. For the weighted paging problem ( it corresponds to the
k-sever problem on a weighted star metric space), there were also O(log k)-competitive algorithms
[4, 9] via the online primal-dual method. More extensive literature on the k-server problem can be
found in [26, 18].
In this paper, we show that there exists a randomized k-sever algorithm of O(log k log2 n)competitive ratio for any metric space with n points, which improved the previous best competitive
ratio O(log2 k log3 n log log n) by Nikhil Bansal et al. [3].
In order to get our results, we use the online primal-dual method, which is developed by Buchbinder and Naor et al. in recent years. Buchbinder and Naor et al. have used the primal-dual
method to design online algorithms for many online problems such as covering and packing problems, the ad-auctions problem and so on [4, 5, 6, 7, 8, 10]. First, we propose a primal-dual
formulation for the fraction k-sever problem on a weighted hierarchical well-separated tree (HST).
Then, we design an O(ℓ log k)-competitive online algorithm for the fraction k-sever problem on a
weighted HST with depth ℓ. Since any HST with n leaves can be transformed into a weighted
HST with depth O(log n) with any leaf to leaf distance distorted by at most a constant [3], thus,
we get an O(log k log n)-competitive online algorithm for the fraction k-sever problem on an HST.
Based on the known relationship between the fraction k-sever problem and the randomized k-sever
problem, we get that there is an O(log k log n)-competitive randomized algorithm for the k-sever
problem on an HST with n points. By the metric embedding theory [22], we get that there is an
O(log k log2 n)-competitive randomized algorithm for the k-sever problem on any metric space with
n points.
2
2
Preliminaries
In this section, we give some basic definitions.
Definition 2.1. (Competitive ratio adapted from [32]) For a deterministic online algorithm DALG,
we call it r-competitive if there exists a constant c such that for any request sequence ρ, costDALG (ρ) ≤
r · costOP T (ρ) + c, where costDALG (ρ) and costOP T (ρ) are the costs of the online algorithm DALG
and the best offline algorithm OP T respectively.
For a randomized online algorithm, we have a similar definition of competitive ratio:
Definition 2.2. (Adapted from [32]) For a randomized online algorithm RALG, we call it rcompetitive if there exists a constant c such that for any request sequence ρ, E[costRALG (ρ)] ≤
r · costOP T (ρ) + c, where E[costRALG (ρ)] is the expected cost of the randomized online algorithm
RALG.
In order to analyze randomized algorithms for the k-sever problem, D. Türkouğlu introduce the
fractional k-sever problem [32]. On the fractional k-sever problem, severs are viewed as fractional
entities as opposed to units and an online algorithm can move fractions of servers to the requested
point.
Definition 2.3. (Fractional k-sever problem adapted from [32]) Suppose that there are a metric
space S and a total of k fractional severs located at the points of the metric space. Given a sequence
of requests, each request must be served by providing one unit server at requested point, through
moving fractional servers to the requested point. The cost of an algorithm for servicing a sequence
of requests is the cumulative sum of the distance incurred by each sever, where moving a w fraction
of a server for a distance of δ costs wδ.
In [11, 12], Bartal introduce the definition of a Hierarchical Well-Separated Tree (HST), into
which a general metric can be embedded with a probability distribution. For any internal node, the
distance from it to its parent node is σ times of the distance from it to its child node. The number
σ is called the stretch of the HST. An HST with stretch σ is called a σ-HST. In the following, we
give its formal definition.
Definition 2.4. (Hierarchically Well-Separated Trees (HSTs)[20]). For σ > 1, a σ-Hierarchically
Well-Separated Tree (σ-HST) is a rooted tree T = (V, E) whose edges length function d satisfies the
following properties:
(1). For any node v and any two children w1 , w2 of v, d(v, w1 ) = d(v, w2 ).
(2). For any node v, d(p(v), v) = σ · d(v, w), where p(v) is the parent of v and w is a child of v.
(3). For any two leaves v1 and v2 , d(p(v1 ), v1 ) = d(p(v2 ), v2 ).
Fakcharoenphol et al. showed the following result [22].
Lemma 2.5. If there is a γ-competitive randomized algorithm for the k-sever problem on an σ-HST
with all requests at the n leaves, then there exists an O(γσ log n)- competitive randomized online
algorithm for the k-server problem on any metric space with n points.
We still need the definition of a weighted hierarchically well-separated tree introduced in [3].
3
Definition 2.6. (Weighted Hierarchically Well-Separated Trees (Weighted HSTs)[3]). A weighted
σ-HST is a rooted tree satisfying the property (1), (3) of the definition 2.4 and the property:
d(p(v), v) ≥ σ · d(v, w) for any node v which is not any leaf or the root, where p(v) is the parent of v and w is a child of v.
In [3], Banasl et al. show that an arbitrary depth σ-HST with n leaves can be embedded into
an O(log n) depth weighted σ-HST with constant distortion, which is described as follows.
Lemma 2.7. Let T be a σ-HST T with n leaves which is of possibly arbitrary depth. Then, T can
be transformed into a weighted σ-HST T̃ with depth O(log n) such that: the leaves of T̃ and T are
2σ
factor.
the same, and leaf to leaf distance in T is distorted in T̃ by a at most σ−1
3
An O(log2 k) Randomized Algorithm for the k-Sever Problem on
an HST when n = k + 1
In this paper, we view the k-sever problem as the weighed caching problem such that the cost of
evicting a page out of the cache using another page satisfies the triangle inequality, i.e., a point is
viewed as a page; the set of k points that are served by k severs is viewed as the cache which holds
k pages; the distance of two points i and j is viewed the cost of evicting the corresponding page pi
out of the cache using the corresponding page pj .
Let [n] = {p1 , . . . , pn } denotes the set of n pages and d(pi , pj ) denotes the cost of evicting
the page pi out of the cache using the page pj for any pi , pj ∈ [n], which satisfies the triangle
inequality: for any pages i, j, s, d(pi , pi ) = 0; d(pi , pj ) = d(pj , pi ); d(pi , pj ) ≤ d(pi , ps ) + d(ps , pj ).
Let p1 , p2 , . . . , pM be the requested pages sequence until time M , where pt is the requested page at
time t. At each time step, if the requested page pt is already in the cache then no cost is produced.
Otherwise, P
the page pt must be fetched into the cache by evicting some other pages p in the cache
and a cost
d(p, pt ) is produced.
In this section, in order to clearly describe our algorithm design idea, we consider the case
n = k + 1.
First, we give some notations. Let σ-HST denote a hierarchically well-separated trees with
stretch factor σ. Let N be the number of nodes in a σ-HST and leaves be p1 , p2 , · · · , pn . Let ℓ(v)
denote the depth of a node v. Let r denote the root node. Thus, ℓ(r) = 0. For any leaf p, let ℓ
denote its depth, i.e. ℓ(p) = ℓ. Let p(v) denote the parent node of a node v. C(v) denote the set
of children of a node v. Let D denote the distance from the root to its a child. D(v) denote the
D
.
distance from a node v to its parent, i.e. D(v) = d(v, p(v)). It is easy to know that D(v) = σℓ(v)−1
Let Tv denote the subtree rooted at v and L(Tv ) denote the set of leaves in Tv . Let |Tv | denote the
number of the leaves in Tv . For a leaf pi , let A(pi , j) denote the ancestor node of pi at the depth
j. Thus, A(pi , ℓ) is pi , A(pi , 0) is the root r and so on. At time t, let variable xpi ,t denote the
fraction of pi that is in
Pthe cache and upi ,t denote the fraction
Pof pi that is out of cache. Obviously,
xpi ,t + upi ,t = 1 and
xp,t = k. For a node v, let uv,t =
up,t , i.e., it is the total fraction of
p∈[n]
p∈L(Tv )
P
pages in the subtree Tv which is out of the cache. It is easy to see that uv,t =
uw,t. Suppose
w∈C(v)
that at time 0, the set of initial k pages in the cache is I = {pi1 , . . . , pik }.
At time t, when the request pt arrives, if page pt is fetched mass ∆(pt , p) into the cache by
evicting out the page p in the cache, then the evicting cost is d(p, pt ) · ∆(pt , p). For a σ-HST metric,
4
suppose the path from pt to p in it is: pt , vj , . . . , v1 , v, v1′ , . . . , vj′ , p, where v is the first common
ancestor node of pt and p. By the definition of a σ-HST, we have D(pt ) = D(p) and D(vi ) = D(vi′ )
j
j
j
P
P
P
for any 1 ≤ i ≤ j. Thus, d(p, pt ) = D(pt ) +
D(vi ) + D(p) +
D(vi′ ) = 2D(pt ) +
2D(vi ). So,
the evicting cost is (2D(pt )+
cost incurred at time t is
N
P
j
P
i=1
i=1
i=1
2D(vi ))·∆(pt , p). Since p can be any page in [n]\{pt }, the evicting
i=1
2D(v) max{0, uv,t−1 − uv,t }. Thus, we give the LP formulation for the
v=1
fractional k-sever problem on a σ-HST as follows.
M P
N
M
P
P
(P)
Minimize
2D(v)zv,t +
∞ · upt ,t
t=1 v=1
t=1 P
up,t ≥ |S| − k;
Subject to ∀t > 0 and S ⊆ [n] with |S| > k,
p∈S
P
(up,t−1 − up,t );
∀t > 0 and a subtree Tv (v 6= r), zv,t ≥
(3.1)
(3.2)
p∈L(Tv )
∀t > 0 and node v, zv,t , uv,t ≥ 0;
(3.3)
For t = 0 and any leaf node p ∈ I, up,0 = 0;
(3.4)
For t = 0 and any leaf node p 6∈ I, up,0 = 1;
(3.5)
The first primal
constraintP
(3.1) states that
Pat any time t, if we take any set S of vertices with
P
xp,t = |S| − k, i.e., the total number of pages out
xp,t ≥ |S| −
up,t = |S| −
|S| > k, then
p∈S
p∈S
p∈[n]
of the cache is at lease |S| − k. The variables zv,t denote the total fraction mass of pages in Tv that
are moved out of the subtree Tv (Obviously, it is not needed to define a variable zr,t for the root
node). The fourth and fifth constraints ((3.4) and (3.5)) enforce the initial k pages in the cache
are pi1 , . . . , pik . The first term in the object function is the sum of the moved cost out of the cache
and the second term enforces the requirement that the page pt must be in the cache at time t (i.e.,
upt ,t = 0).
Its dual formulation is as follows.
M
P
P
P
γp
(|S| − k)aS,t +
(D)
Maximize
t=1 S⊆[n],|S|>k
Subject to ∀t and p ∈ [n] \ {pt },
p6∈I
P
aS,t −
S:p∈S
∀t = 0 and ∀p ∈ [n], γp −
ℓ
P
ℓ
P
(b(A(p,j),t+1 − b(A(p,j),t ) ≤ 0
(3.6)
j=1
bA(p,j),1 ≤ 0
(3.7)
j=1
∀t > 0 and any subtree Tv , bv,t ≤ 2D(v)
(3.8)
∀t > 0 and v and |S| > k, aS,t , bv,t ≥ 0
(3.9)
In the dual formulation, the variable aS,t corresponds to the constraint of the type (3.1); the
variable bv,t corresponds to the constraint of the type (3.2); The variable γp corresponds to the
constraint of the type (3.4) and (3.5).
Based on above primal-dual formulation, we extend the design idea of Bansal et al.’s primaldual algorithm for the metric task system problem on a σ-HST [10] to the k-sever problem on a
σ-HST. The design idea of our online algorithm is described as follows. During the execution of
our algorithm, it always maintains the following relation between the primal variable uv,t and dual
bv,t+1
ln(1 + k)) − 1). When the request pt arrives at
variable bv,t+1 : uv,t = f (bv,t+1 ) = |Tkv | (exp( 2D(v)
time t, the page pt is gradually fetched into the cache and other pages are gradually moved out of
5
the cache by some rates until pt is completely fetched into the cache(i.e, upt ,t is decreased at some
rate and other up,t is increased at some rate for any p ∈ [n] \ {pt } until upt ,t becomes 0). It can be
viewed that we move mass upt ,t out of leaf pt through its ancestor nodes and distribute it to other
leaves p ∈ [n] \ {pt }. In order to compute the exact distributed amount at each page p ∈ [n] \ {pt },
the online algorithm should maintain the following invariants:
1. ( Satisfying Dual Constraints:) It is tight for all dual constraints of type (3.6) on other
leaves [n] \ {pt }.
P
2. (Node Identity Property:) uv,t =
uw,t holds for each node v,.
w∈C(v)
We give more clearer description of the online algorithm process. At time t, when the request
pt arrives, we initially set upt ,t = upt ,t−1 . If upt ,t = 0, then we do nothing. Thus, the primal cost
and the dual profit are both zero. All the invariants continue to hold. If upt ,t 6= 0, then we start
to increase variable aS at rate 1. At each step, we would like to keep the dual constraints (3.6)
tight and maintain the node identity property. However, increasing variable aS violates the dual
constraints(3.6) on leaves in [n] \ {pt }. Hence, we increase other dual variables in order to keep
these dual constraints (3.6) tight. But, increasing these variables may also violate the node identity
property. So, it makes us to update other dual variables. This process results in moving initial upt ,t
mass from leaf pt to leaves [n] \ {pt }. We stop the updating process when upt ,t become 0.
In the following, we will compute the exact rate at which we should move mass upt ,t from pt
through its ancestor nodes at time t to other leaves in [n] \ {pt } in the σ-HST. Because of the space
limit, we put proofs of the following some claims in the Appendix. First, we show one property of
the function f .
Lemma 3.1.
duv,t
dbv,t+1
Proof. Since uv,t =
claim.
=
ln(1+k)
2D(v) (uv,t
bv,t+1
|Tv |
k (exp( 2D(v)
+
|Tv |
k ).
ln(1 + k)) − 1), we take the derivative over bv,t+1 and get the
In order to maintain the Node Identity Property: uv,t =
P
uw,t for each node v at any
w∈C(v)
time t, when uv,t is increased or decreased, it is also required to increase or decrease the children
of v at some rate. The connection between these rates is given.
Lemma 3.2. For a node v, if we increase variable bv,t+1 at rate h, then we have the following
equality:
P dbw,t+1
dbv,t+1
|Tv |
1
=
· (uw,t + |Tkv | )
σ (uv,t + k ) ·
dh
dh
w∈C(v)
We need one special case of lemma 3.2: when the variable bv,t+1 is increased (decreased) at rate
h, it is required that the increasing (decreasing) rate of all children of v is the same. By above
lemma, we get:
Lemma 3.3. For v a node, assume that we increase (or decrease ) the variable bv,t+1 at rate h. If
the increasing (or decreasing ) rate of each w ∈ C(v) is the same, then in order to keep the Node
Identity Property, we should set the increasing (or decreasing ) rate for each child w ∈ C(v) as
follows:
db
dbw,t+1
= σ1 · v,t+1
dh
dh
Repeatedly applying this lemma, we get the following corollary.
6
Corollary 3.4. For a node v with ℓ(v) = j and a path P from leaf pi ∈ Tv to v, if bv,t+1 is increased
( or decreased) at rate h and the increasing (decreasing) rate of all children of any v ′ ∈ P is the
P dbv′ ,t+1
db
1
) = (1 + Θ( σ1 )).
= v,t+1
· ψ(j), where ψ(j) = (1 + σ1 + σ12 + · · · + σℓ−j
same, then
dh
dh
v′ ∈P
We still require the following special case of lemma 3.2. Let w1 be the first child of the node
v. Assume that bw1 ,t+1 is increased (or decreased) at some rate and the rate of increasing (or
decreasing) bw′ ,t+1 is the same for every w′ ∈ C(v), w′ 6= w1 . If bv,t+1 is unchanged, then the
following claim should hold.
Lemma 3.5. Let w1 , . . . , wm be the children of a node v. Assume that we increase (or decrease)
db ′
db i ,t+1
w1 at rate h and also increase w2 to wm at the same rate h. For i ≥ 2, let wdh,t+1 be wdh
. If
we would like to maintain the amount uv,t unchanged, then we should have:
dbw′ ,t+1
dh
|T (w1 )|
k
|T (v)|
uv,t + k
uw1 ,t +
=
· (−
dbw1 ,t+1
dh
+
dbw′ ,t+1
)
dh
Theorem 3.6. When request pt arrives at time t, in order to keep the dual constraints tight
and node identity property, if aS,t is increased with rate 1, we should decrease every bA(pt ,j),t+1
(1 ≤ j ≤ ℓ) with rate:
dbA(pt ,j),t+1
daS,t
2+
1
|T
|T
|
|
n−1
= ψ(j)
[(uA(pt ,j),t + A(pk t ,j) )−1 − (uA(pt ,j−1),t + A(ptk,j−1) )−1 ].
For each sibling w of A(pt , j), increase bw,t+1 with the following rate:
dbw,t+1
daS,t
=
1
2+ n−1
ψ(j) (uA(pt ,j−1),t
+
|TA(pt ,j−1) | −1
)
k
Thus, we design an online algorithm for the fractional k-sever problem as follows (see Algorithm
3.1).
1:
2:
3:
4:
5:
6:
At time t = 0, we set bp,1 = γp = 0 for all p and set bA(p,j),1 = 0 for any 1 ≤ j ≤ ℓ.
At time t ≥ 1, when a request pt arrives:
Initially, we set up,t = up,t−1 for all p, and bp,t+1 is initialized to bp,t .
If upt ,t = 0, then do nothing.
Otherwise, do the following:
P
xp,t = k = n − 1. So, S = [n].
Let S = {p : up,t < 1}. Since k = n − 1, |S| >
x∈S
While upt ,t 6= 0:
8:
Increasing aS,t with rate 1;
9:
For each 1 ≤ j ≤ ℓ, decrease every bA(pt ,j),t+1 with rate:
7:
dbA(pt ,j),t+1
daS,t
2+
1
|T
|
|T
|
n−1
= ψ(j)
[(uA(pt ,j),t + A(pk t ,j) )−1 − (uA(pt ,j−1),t + A(ptk,j−1) )−1 ],
For each sibling w of A(pt , j), increase bw,t+1 with the following rate:
10:
11:
12:
dbw,t+1
daS,t
13:
For any
1
σ
·
1
2+ n−1
|TA(pt ,j−1) | −1
)
k
ψ(j) (uA(pt ,j−1),t +
node v ′ in the path from w to a leaf
=
dbv′ ,t+1
dh
in Tw , if w′ be the child of v ′ ,
dbw′ ,t+1
dh
=
Algorithm 3.1: The online primal-dual algorithm for the fractional k-sever problem on a σ-HST.
Theorem 3.7. The online algorithm for the fractional k-sever problem on a σ-HST is of competitive
ratio 15 ln2 (1 + k).
7
In [32], Duru Türkoğlu study the relationship between fractional version and randomized version
of the k-sever problem, which is given as follows.
Lemma 3.8. The fractional k-sever problem is equivalent to the randomized k-sever problem on
the line or circle, or if k = 2 or k = n − 1 for arbitrary metric spaces.
Thus, we get the following conclusion:
Theorem 3.9. There is a randomized algorithm with competitive ratio 15 ln2 (1 + k) for the k-sever
problem on a σ-HST when n = k + 1.
By lemma 2.5, we get the following conclusion:
Theorem 3.10. There is an O(log2 k log n) competitive randomized algorithm for the k-sever problem on any metric space when n = k + 1 .
4
An O(ℓ log k)-competitive Fractional Algorithm for the k-Sever
Problem on a Weighted HST with Depth ℓ
In this section, we first give an O(ℓ log k)-competitive fractional algorithm for the k-Sever problem
on a weighted σ-HST with depth ℓ.
We give another some notations for a weighted HST. Let T̃ be a weighted σ-HST. For a node
D(v)
= d(p(v),v)
whose depth is j, let σj = D(w)
d(v,w) where w is a child of v. By the definition of a weighted σHST, σj ≥ σ for all 1 ≤ j ≤ ℓ−1. For a node v ∈ T̃ , if any leaf p ∈ L(T (v)) such that up,t = 1, we call
it a full node. By this definition, for a full node, uv,t = |L(Tv )|. Otherwise, we call it non-full node.
Let N F C(v) is the set of non-full children node of v, i.e., N F C(v) = {w|w ∈ C(v) and w is a non-full
node }. For a node v, let N L(Tv ) denote the set of non-full leaf nodes in T̃v . Let S = {p|up,t < 1}.
Let P ′ denotes the path from pt to root r: {A(pt , ℓ) = Pt , A(pt , ℓ−1), . . . , A(pt , 1), . . . , A(pt , 0) = r}.
For a node v ∈ P ′ , if there exists a p ∈ S\{pt } such that v is the first common ancestor of pt and p, we
call it a common ancestor node in P ′ . Let CA(pt , S) denote the set of common ancestor nodes in P ′ .
Suppose thatP
CA(pt , S) = {A(pt , ℓh ), . . . , A(pt , ℓ2 ), A(pP
t , ℓ1 )}, where ℓ1 < ℓ2 < . . . < ℓh . For a node
uw,t . Thus, for a full node v, uv,t = 0.
up,t . It is easy to know that uv,t =
v, let uv,t =
p∈S
w∈N F C(v)
For any ℓh < j < ℓ, uA(pt ,j),t = uA(pt ,ℓ),t . For any 0 ≤ j < ℓ1 , uA(pt ,0),t = uA(pt ,j),t = uA(pt ,ℓ1 ),t . For
any ℓi−1 < j < ℓi , uA(pt ,j),t = uA(pt ,ℓi ),t .
The primal-dual formulation for the fractional k-sever problem on a weighted HST is the same
as that on a HST in section 3. Based on the primal-dual formulation, the design idea of our online
algorithm is similar to the design idea in section 3. During the execution of our algorithm, it keeps
the following relation between the primal variable uv,t and dual variable bv,t+1 : uv,t = f (bv,t+1 ) =
bv,t+1
|N L(Tv )|
(exp( 2D(v)
ln(1 + k)) − 1). This relation determines how much mass of upt ,t should be
k
gradually moved out of leaf pt and how it should be distributed among other leaves S \ {pt } until
pt is completely fetched into the cache, i.e. upt ,t = 0. Thus,
P at any time t, the algorithm maintains
up,t = n − k.
a distribution (up1 ,t , . . . , upn ,t ) on the leaves such that
p∈[n]
In order to compute the the exact rate at which we should move mass upt ,t from pt through its
ancestor nodes at time t to other leaves S \ {pt } in the weighted σ-HST, using similar argument to
that in section 3, we get following several claims. Because of the space limit, we put their proofs
in the Appendix.
8
Lemma 4.1.
duv,t
dbv,t+1
Proof. Since uv,t =
claim.
=
ln(1+k)
2D(v) (uv,t
|N L(Tv )|
).
k
+
bv,t+1
|N L(Tv )|
(exp( 2D(v)
k
ln(1 + k)) − 1), we take the derivative over bv,t+1 and get the
Lemma 4.2. For a node v with ℓ(v) = j, if we increase variable bv,t+1 at rate h, then we have the
following equality:
P
dbw,t+1
db
|N L(Tv )|
w )|
1
) · v,t+1
=
· (uw,t + |N L(T
)
σj (uv,t +
k
dh
dh
k
w∈N F C(v)
Lemma 4.3. For v a node with ℓ(v) = j, assume that we increase (or decrease ) the variable bv,t+1
at rate h. If the increasing (or decreasing ) rate of each w ∈ N F C(v) is the same, then in order
to keep the Node Identity Property, we should set the increasing (or decreasing ) rate for each
child w ∈ N F C(v) as follows:
dbw,t+1
db
= σ1j · v,t+1
dh
dh
Repeatedly applying this lemma, we get the following corollary.
Corollary 4.4. For a node v with ℓ(v) = j and a path P from leaf pi ∈ Tv to v, if bv,t+1 is increased
( or decreased) at rate h and the increasing (decreasing) rate of all children of any v ′ ∈ P is the same,
P dbv′ ,t+1
db
= v,t+1
· φ(j), where φ(j) = (1 + σ1j + σj ·σ1j+1 + · · · + σj ·σj+11...σℓ−1 ) ≤ (1 + Θ( σ1 )).
then
dh
dh
v′ ∈P
Lemma 4.5. Let w1 , . . . , wm be the non-full children node of a node v (i.e., any wi ∈ N F C(v)).
Assume that we increase (or decrease) w1 at rate h and also increase w2 to wm at the same rate
db ′
db i ,t+1
. If we would like to maintain the amount uv,t unchanged, then
h. For i ≥ 2, let wdh,t+1 be wdh
we should have:
dbw′ ,t+1
dh
|NL(Tw1 )|
k
v )|
uv,t + |NL(T
k
uw1 ,t +
=
· (−
dbw1 ,t+1
dh
+
dbw′ ,t+1
)
dh
Theorem 4.6. When request pt arrives at time t, in order to keep the dual constraints tight and
node identity property, if aS,t is increased with rate 1, we should decrease every bA(pt ,j),t+1 for each
j ∈ {ℓ1 + 1, ℓ2 + 1, . . . , ℓh + 1} with rate:
dbA(pt ,j),t+1
daS,t
ur,t + |S|
|N L(T
|N L(T
)|
)|
A(pt ,j−1)
A(pt ,j)
= φ(j) k [(uA(pt ,j),t +
)−1 − (uA(pt ,j−1),t +
)−1 ].
k
k
For each sibling w ∈ N F C(v) of A(pt , j), increase bw,t+1 with the following rate:
dbw,t+1
daS,t
|S|
=
ur,t + k
φ(j)
(uA(pt ,j−1),t +
|N L(TA(pt ,j−1) )| −1
)
k
Thus, we design an online algorithm for the fractional k-sever problem on a weighted σ-HST as
follows (see Algorithm 4.1).
Theorem 4.7. The online algorithm for the fractional k-sever problem on a weighted σ-HST with
depth ℓ is of competitive ratio 4ℓ ln(1 + k).
By lemma 2.7, we get:
Theorem 4.8. There exists an O(log k log n)-competitive fractional algorithm for the k-sever problem on any σ-HST.
In [3], Nikhil Bansal et al. show the following conclusion.
9
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
At time t = 0, we set bp,1 = γp = 0 for all p.
At time t ≥ 1, when a request pt arrives:
Initially, we set up,t = up,t−1 for all p, and bp,t+1 is initialized to bp,t .
If upt ,t = 0, then do nothing.
Otherwise, do the following:
Let S = {p : up,t < 1}. Suppose that CA(pt , S) = {A(pt , ℓh ), . . . , A(pt , ℓ2 ), A(pt , ℓ1 )}, where
0 ≤ ℓ1 < ℓ2 < . . . < ℓh < ℓ.
While upt ,t 6= 0:
Increasing aS,t with rate 1;
For each j ∈ {ℓ1 + 1, ℓ2 + 1, . . . , ℓh + 1}, decrease every bA(pt ,j),t+1 with rate:
|S|
dbA(pt ,j),t+1
ur,t + k
=
daS,t
φ(j) [(uA(pt ,j),t
|N L(TA(pt ,j−1) )| −1
) ].
k
+
|N L(TA(pt ,j) )| −1
)
k
− (uA(pt ,j−1),t +
For each sibling w ∈ N F C(v) of A(pt , j), increase bw,t+1 with the following rate:
dbw,t+1
daS,t
ur,t + |S|
|N L(T
)|
A(pt ,j−1)
= φ(j)k (uA(pt ,j−1),t +
)−1
k
14:
For any node v ′ in the path from w to a leaf in N L(Tw ), if w′ ∈ N F C(v ′ ) and
db ′
db ′
ℓ(v ′ ) = j, wdh,t+1 = σ1j · vdh,t+1
15:
For p ∈ S \ {pt }), if some up,t reaches the value of 1, then we update S ← S \ {p} and
16:
the set N F C(v) for each ancestor node v of p.
13:
Algorithm 4.1: The online primal-dual algorithm for the fractional k-sever problem on a weighted
σ-HST.
Lemma 4.9. Let T be a σ-HST with σ > 5. Then any online fractional k-sever algorithm on T can
be converted into a randomized k-sever algorithm on T with an O(1) factor loss in the competitive
ratio.
Thus, we get the following conclusion by Theorem 4.8:
Theorem 4.10. Let T be a σ-HST with σ > 5. There is a randomized algorithm for the k-sever
problem with a competitive ratio of O(log k log n) on T .
By lemma 2.5, we get the following conclusion:
Theorem 4.11. For any metric space, there is a randomized algorithm for the k-sever problem
with a competitive ratio of O(log k log2 n).
5
Conclusion
In this paper, for any metric space with n points, we show that there exist a randomized algorithm
with O(log k log2 n)-competitive ratio for the k-sever problem, which improved the previous best
competitive ratio O(log2 k log3 n log log n).
Acknowledgments
We would like to thank the anonymous referees for their careful readings of the manuscripts and
many useful suggestions.
10
Wenbin Chen’s research has been partly supported by the National Natural Science Foundation
of China (NSFC) under Grant No.11271097., the research projects of Guangzhou education bureau
under Grant No. 2012A074. and the project KFKT2012B01 from State Key Laboratory for Novel
Software Technology, Nanjing University.
References
[1] Dimitris Achlioptas, Marek Chrobak, and John Noga. Competitive analysis of randomized
paging algorithms. Theoretical Computer Science, 234(1-2):203-218, 2000.
[2] Avrim Blum, Carl Burch, and Adam Kalai. Finely-competitive paging. Proceedings of the 40th
Annual Symposium on Foundations of Computer Science, page 450-456, 1999.
[3] Nikhil Bansal, Niv Buchbinder, Aleksander Madry, Joseph Naor. A PolylogarithmicCompetitive Algorithm for the k-Server Problem. FOCS 2011, pages 267-276.
[4] Nikhil Bansal, Niv Buchbinder, and Joseph (Seffi) Naor. A primal-dual randomized algorithm
for weighted paging. Proceedings of the 48th Annual IEEE Symposium on Foundations of
Computer Science, pages 507-517, 2007.
[5] N. Buchbinder, K. Jain, and J. Naor. Online primal-dual algorithms for maximizing ad-auctions
revenue. Proc. 14th European Symp. on Algorithms (ESA), pp. 253-264, 2007.
[6] N. Buchbinder and J. Naor. Online primal-dual algorithms for covering and packing problems.
Proc. 12th European Symp. on Algorithms (ESA), volume 3669 of Lecture Notes in Comput.
Sci., pages 689–701. Springer, 2005.
[7] N. Buchbinder and J. Naor. Improved bounds for online routing and packing via a primal-dual
approach. Proc. 47th Symp. Foundations of Computer Science, pages 293–304, 2006.
[8] Niv Buchbinder, Joseph Naor. The Design of Competitive Online Algorithms via a Primal-Dual
Approach. Foundations and Trends in Theoretical Computer Science 3(2-3): 93–263 (2009).
[9] Nikhil Bansal, Niv Buchbinder, and Joseph (Seffi) Naor. Towards the randomized k-server
conjecture: A primal-dual approach. Proceedings of the 21st Annual ACM- SIAM Symposium
on Discrete Algorithms, pp. 40-55, 2010.
[10] Nikhil Bansal, Niv Buchbinder, and Joseph (Seffi) Naor. Metrical task systems and the ksever problem on HSTs. In ICALP’10: Proceedings of the 37th International Colloquium on
Automata, Languages and Programming, 2010, pp.287-298.
[11] Yair Bartal. Probabilistic approximations of metric spaces and its algorithmic applications.
Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science, pages
184-193, 1996.
[12] Yair Bartal. On approximating arbitrary metrices by tree metrics. Proceedings of the 30th
Annual ACM Symposium on Theory of Computing, pages 161-168, 1998.
11
[13] Yair Bartal, Bela Bollobas, and Manor Mendel. A ramsy-type theorem for metric spaces and its
applications for metrical task systems and related problems. Proceedings of the 42nd Annual
IEEE Symposium on Foundations of Computer Science, pages 396-405, 2001.
[14] Yair Bartal and Eddie Grove. The harmonic k-server algorithm is competitive. Journal of the
ACM, 47(1):1-15, 2000.
[15] Yair Bartal, Nathan Linial, Manor Mendel, and Assaf Naor. On metric ramsey-type phenomena. Proceedings of the 35th Annual ACM Symposium on Theory of Computing, pages
463C472, 2003.
[16] Yair Bartal, Elias Koutsoupias. On the competitive ratio of the work function algorithm for
the k-sever problem. Theoretical Computer Science, 324(2-3): 337-345.
[17] Avrim Blum, Howard J. Karloff, Yuval Rabani, and Michael E. Saks. A decomposition theorem
and bounds for randomized server problems. Proceedings of the 31st Annual IEEE Symposium
on Foundations of Computer Science, pages 197-207, 1992.
[18] Allan Borodin and Ran El-Yaniv. Online computation and competitive analysis. Cambridge
University Press, 1998.
[19] M. Chrobak and L. Larmore. An optimal on-line algorithm for k-servers on trees. SIAM Journal
on Computing, 20(1): 144-148, 1991.
[20] A. Coté, A. Meyerson, and L. Poplawski. Randomized k-server on hierarchical binary trees.
Proceedings of the 40th Annual ACM Symposium on Theory of Computing, pages 227-234,
2008.
[21] B. Csaba and S. Lodha. A randomized on-line algorithm for the k-server problem on a line.
Random Structures and Algorithms, 29(1): 82-104, 2006.
[22] Jittat Fakcharoenphol, Satish Rao, and Kunal Talwar. A tight bound on approximating arbitrary metrics by tree metrics. Proceedings of the 35th Annual ACM Symposium on Theory of
Computing, pages 448-455, 2003.
[23] A. Fiat, Y. Rabani, and Y. Ravid. Competitive k-server algorithms. Journal of Computer and
System Sciences, 48(3): 410-428, 1994.
[24] Amos Fiat, Richard M. Karp, Michael Luby, Lyle A. McGeoch, Daniel Dominic Sleator, and
Neal E. Young. Competitive paging algorithms. Journal of Algorithms, 12(4): 685-699, 1991.
[25] Edward F. Grove. The harmonic online k-server algorithm is competitive. Proceedings of the
23rd Annual ACM Symposium on Theory of Computing, pages 260-266, 1991.
[26] Elias Koutsoupias. The k-sever problem. Computer Science Review. Vol. 3. No. 2. Pages 105118, 2009.
[27] Elias Koutsoupias and Christos H. Papadimitriou. On the k-server conjecture. Journal of the
ACM, 42(5): 971-983, 1995.
12
[28] M.S. Manasse, L.A. McGeoch, and D.D. Sleator. Competitive algorithms for online problems.
Proceedings of the 20th Annual ACM Symposium on Theory of Computing, pages 322-333,
1988.
[29] M. Manasse, L.A. McGeoch, and D. Sleator. Competitive algorithms for server problems.
Journal of Algorithms, 11: 208-230, 1990.
[30] Lyle A. McGeoch and Daniel D. Sleator. A strongly competitive randomized paging algorithm.
Algorithmica, 6(6): 816-825, 1991.
[31] Daniel D. Sleator and Robert E. Tarjan. Amortized efficiency of list update and paging rules.
Communications of the ACM, 28(2): 202-208, 1985.
[32] Duru Türkoğlu. The k-sever problem and fractional analysis. Master’s Thesis, The University
of Chicago, 2005. http://people.cs.uchicago.edu/∼duru/papers/masters.pdf.
Appendix
Proofs for claims in section 3
The proof for Lemma 3.2 is as follows:
P
Proof. Since it is required to maintain uv,t =
uw,t, we take the derivative of both sides and
w∈C(v)
get that:
duv,t
dbv,t+1
·
dbv,t+1
dh
P
=
w∈C(v)
By lemma 3.1, we get:
duw,t
dbw,t+1
·
dbw,t+1
dh .
ln(1+k)
2D(v) (uv,t
+
D(v)
Since D(w)
= σ, we get :
P dbw,t+1
dbv,t+1
|Tv |
1
=
σ (uv,t + k ) ·
dh
dh
w∈C(v)
|Tv |
k )
P
=
w∈C(v)
· (uw,t +
dbw,t+1
dh
·
ln(1+k)
2D(w) (uw,t
+
|Tv |
k ).
|Tw |
k )
The proof for Lemma 3.3 is as follows:
Proof. By above Lemma 3.2, if the increasing (or decreasing ) rate of each w ∈ C(v) is the same,
we get that:
P
dbv,t+1
db
db
|Tv |
1
= w,t+1
·
· (uv,t + |Tkv | ).
(uw,t + |Tkw | ) = w,t+1
σ (uv,t + k ) ·
dh
dh
dh
So, we get that:
dbw,t+1
dh
=
1
σ
·
w∈C(v)
dbv,t+1
dh
The proof for Lemma 3.5 is as follows:
Proof. By lemma 3.2, in order to keep the amount uv,t unchanged, we get:
P
dbw1 ,t+1
dbw,t+1
1 )|
· (uw1 ,t + |T (w
)+
· (uw,t + |T (w)|
dh
k
dh
k ) = 0.
Thus,
db 1 ,t+1
( wdh
db
−
db
dbw′ ,t+1
)
dh
′
w∈C(v)\{w1 }
1 )|
· (uw1 ,t + |T (w
)
k
1 ,t+1
So, ( wdh
− wdh,t+1 ) · (uw1 ,t +
Hence, we get the claim.
|T (w1 )|
)
k
+
+
dbw′ ,t+1
dh
dbw′ ,t+1
dh
13
·
·
P
(uw,t +
w∈C(v)
(uv,t + |T (v)|
k )
|T (w)|
k )
=0
=0
The proof for Theorem 3.6 is as follows:
Proof. When request pt arrives at time t, we move mass upt ,t from pt through its ancestor nodes
to other leaves [n] \ {pt }, i.e. upt ,t is decreased and up,t is increased for any p ∈ [n] \ {pt }. Since
these mass moves out of each subtree TA(pt ,j) for each 1 ≤ j ≤ ℓ, uA(pt ,j),t is decreased. By
b
t ,j),t+1
uA(pt ,j),t = f (bA(pt ,j),t+1 ) = |Tkv | (exp( A(p
ln(1 + k)) − 1) (we need to keep this relation during
2D(v)
the algorithm), bA(pt ,j),t+1 also decreases for each 1 ≤ j ≤ ℓ. On the other hand, up,t is increased
for each p ∈ [n] \ {pt }. Thus, for each node v whose Tv doesn’t contain pt , its mass uv,t is also
increased. For each node v whose Tv doesn’t contain pt , it must be a sibling of some node A(pt , j).
For each 1 ≤ j ≤ ℓ, we assume that all siblings v ′ of node A(pt , j) increase bv′ ,t+1 at the same rate.
In the following, we will compute the increasing (or decreasing) rate of all dual variables in the
b
t ,j),t+1
be the decreasing rate of bA(pt ,j),t+1
σ-HST regarding aS . For 1 ≤ j ≤ ℓ, let ∇bj = − A(pda
S
b
regarding aS . For 1 ≤ j ≤ ℓ, let ∇b′j = w,t+1
daS be the increasing rate of bw,t+1 for any siblings w of
A(pt , j) regarding aS .
Using from top to down method, we can get a set of equations about the quantities ∇bj and
∇b′j . First, we consider the siblings of A(pt , 1) ( i.e. those nodes are children of root r, but they
are not A(pt , 1)). Let w be one of these siblings. If bw,t+1 is raised by ∇b′1 , by Corollary 3.4, the
sum of ∇b′ on any path from a leaf in Tw to w must be ψ(1) · ∇b′1 . Since aS is increasing with rate
1, it forces ψ(1) · ∇b′1 = 1 in order to maintain the dual constraint (3) tight for leaves in Tw .
This considers the dual constraints for these leaves. Now, this increasing mass must be canceled
out by decreasing the mass in TA(pt ,1) since the mass ur,t in Tr is not changed. Thus, in order to
maintain the “Node Identity Property” of root, by Lemma 3.5, we must set ∇b1 such that:
∇b′1 = (∇b1 + ∇b′1 ) ·
|T(A(p ,1)) |
t
)
k
(uA(pt ,1),t +
n
1+ n−1
For siblings of node A(pt , 2), we use the similar argument. Let w be a sibling of A(pt , 2).
Consider a path from a leaf in Tw to the w. Their dual constraint (3) already grows at rate
1 + ψ(1)∇b1 . This must be canceled out by increasing bw,t+1 , and if bw,t+1 is raised by ∇b′2 , by
corollary 3.4, the sum of ∇b′ on any path from a leaf in Tw to w must be ψ(2) · ∇b′2 . Thus, ∇b′2
must be set such that: ψ(2) · ∇b′2 = 1 + ψ(1) · ∇b1
Again, this increasing mass must be canceled out by decreasing the mass in TA(pt ,2) . In order
to keep the “Node Identity Property” of A(pt , 1), by Lemma 3.5. we must set ∇b2 such that:
∇b′2 = (∇b2 + ∇b′2 ) ·
|TA(p ,2) |
t
)
k
|T(A(p ,1) |
t
(uA(pt ,1),t +
)
k
(uA(pt ,2),t +
Continuing this method, we obtain a system of linear equations about all ∇bj and ∇b′j (1 ≤
j ≤ ℓ). For maintaining the dual constraints tight, we get the following equations:
ψ(1) · ∇b′1 = 1
ψ(2) · ∇b′2 = 1 + ψ(1) · ∇b1
..
.
ℓ−1
P
ψ(ℓ) · ∇b′ℓ = 1 +
ψ(i)∇bi
i=1
For keeping the node identity property, we get the following equations:
∇b′1 = (∇b1 + ∇b′1 ) ·
|T(A(p ,1)) |
t
)
k
(uA(pt ,1),t +
n
1+ n−1
14
∇b′2 = (∇b2 + ∇b′2 ) ·
..
.
∇b′ℓ
= (∇bℓ +
∇b′ℓ )
·
|T(A(p ,2)) |
t
)
k
|T(A(p ,1)) |
t
(uA(pt ,1),t +
)
k
(uA(pt ,2),t +
|T(A(p ,ℓ)) |
t
)
k
|T(A(p ,ℓ−1)) |
t
)
(uA(pt ,ℓ−1),t +
k
(uA(pt ,ℓ),t +
.
We continue to solve the system of linear equations.
For each 1 ≤ j ≤ ℓ,
j−1
P
ψ(j) · ∇b′j = 1 +
ψ(i)∇bi
=1+
j−2
P
i=1
ψ(i)∇bi + ψ(j − 1)∇bj−1
i=1
= ψ(j − 1) · ∇b′j−1 + ψ(j − 1)∇bj−1
= ψ(j − 1) · (∇b′j−1 + ∇bj−1 ).
|T(A(p ,j−1)) |
t
)
k
|T(A(p ,j−2)) |
t
(uA(pt ,j−2),t +
)
k
|T(A(p ,j−2)) |
t
)
(uA(pt ,j−2),t +
k
|T(A(p ,j−1)) |
t
(uA(pt ,j−1),t +
)
k
Since ∇b′j−1 = (∇bj−1 + ∇b′j−1 ) ·
ψ(j)∇b′j = ψ(j − 1)∇b′j−1 ·
(uA(pt ,j−1),t +
, we get:
Solving the recursion, we get:
n
1+ n−1
∇b′j =
|T(A(p ,j−1)) |
∇bj =
ψ(j)·(uA(pt ,j−1),t +
n
1+ n−1
ψ(j) · [(uA(pt ,j),t
t
k
+
)
|T(A(pt ,j)) | −1
)
k
− (uA(pt ,j−1),t +
|T(A(pt ,j−1)) | −1
) ]
k
The proof for Theorem 3.7 is as follows:
Proof. Let P denote the value of the objective function of the primal solution and D denote the
value of the objective function of the dual solution. Initially, let P = 0 and D = 0. In the following,
we prove three claims:
(1) The primal solution produced by the algorithm is feasible.
(2) The dual solution produced by the algorithm is feasible.
(3) P ≤ 15 ln2 (1 + k)D.
By three claims and weak duality of linear programs, the theorem follows immediately.
First,
we prove the claim (1) as follows. At any time t, since S = [n] and the algorithm keeps
P
= n − k = 1 = |S| − k. So, the primal constraints (3.1) are satisfied.
p∈[n]
Second, we prove the claim (2) as follows. By theorem 3.6, the dual constraints (3.6) are
satisfied. Obviously, dual constraints (3.7) are satisfied. For any node v, if bv,t+1 = 0, then
uv,t = 0; if bv,t+1 = 2D(v), then uv,t = |Tv |. Thus, the dual constraints (3.8) are satisfied.
Third, we prove claim (3) as follows. If the algorithm increases the variables aS,t at some time
∂D
= |S| − k = n − (n − 1) = 1. Let’s compute the primal cost. At depth j(1 ≤ j ≤ ℓ),
t, then: ∂a
S,t
we compute the movement cost of our algorithm by the change of ∇b′j as follows.
P
db′
du
2D(w) · dbw,t
· aSj
′
j
w∈C(A(pt ,j−1))\{A(pt ,j)}
P
= ∇b′j ·
2D(w)
w∈C(A(pt ,j−1))\{A(pt ,j)}
·
ln(1+k)
2D(w) (uw,t
15
+
|Tw |
k )
=
1
2+ n−1
ψ(j)
≤
5
2
P
(uw,t +
w∈C(A(pt ,j−1))\{A(pt ,j)}
|TA(p ,j−1) |
t
uA(p ,j−1),t +
k
P t
|T |
(uw,t + kw )
w∈C(A(pt ,j−1))\{A(pt ,j)}
|TA(p ,j−1) |
t
uA(pt ,j−1),t +
k
|Tw |
)
k
· ln(1 + k) ·
· ln(1 + k) ·
Let Bj denote uA(pt ,j),t +
|TA(pt ,j) |
.
k
P
Then
(uw,t +
w∈C(A(pt ,j−1))\{A(pt ,j)}
|Tw |
k )
= Bj−1 − Bj .
Hence, the total
cost over all ℓ levels is
Pℓ movement
Bj−1 −Bj
5
ln(1
+
k)
·
j=1 Bj−1
2
P
Bj
5
= 2 ln(1 + k) · ℓj=1 (1 − Bj−1
)
P
B
ℓ
j
≤ 52 ln(1 + k) · j=1 ln Bj−1
=
=
5
2
5
2
5
2
5
2
5
2
0
ln(1 + k) · ln B
Bℓ
ln(1 + k) · ln u B0+ 1
pt ,t
k
≤ ln(1 + k) · ln kB0
≤ ln(1 + k) · 2 ln k · B0
1
≤ ln(1 + k) · 2 ln k · (2 + n−1
)
2
= 15 ln (1 + k)
where the first inequality holds since 1 ≤ y − ln y for any 0 ≤ y ≤ 1.
Thus, we get P ≤ 15 ln2 (1 + k)D.
Let OP T be the cost of the best offline algorithm. Pmin be the optimal primal solution and
Dmax be the optimal dual solution. Then, Pmin ≤ OP T since OP T is a feasible solution for the
15 ln2 (1+k)D
P
P
≤
primal program. Based on the weak duality, Dmax ≤ Pmin . Hence, OP
T ≤ Pmin ≤
Pmin
15 ln2 (1+k)Dmax
Pmin
2
min
≤ 15 ln (1+k)P
= 15 ln2 (1 + k).
Pmin
So, the competitive ratio of this algorithm is 15 ln2 (1 + k).
Proofs for claims in section 4
The proof for Lemma 4.2 is as follows:
P
Proof. Since it is required to maintain uv,t =
uw,t, we take the derivative of both sides
w∈N F C(v)
and get that:
dbv,t+1
duv,t
=
dbv,t+1 ·
dh
P
duw,t
dbw,t+1
·
w∈N F C(v)
ln(1+k)
2D(v) (uv,t
By lemma 3.1, we get:
D(v)
Since D(w)
= σj , we get :
db
|N L(Tv )|
1
) · v,t+1
σj (uv,t +
k
dh
=
dbw,t+1
dh .
+
|N L(Tv )|
)
k
P
w∈N F C(v)
dbw,t+1
dh
=
P
w∈N F C(v)
· (uw,t +
dbw,t+1
dh
·
ln(1+k)
2D(w) (uw,t
+
|N L(Tw )|
).
k
|N L(Tw )|
)
k
The proof for Lemma 4.3 is as follows:
Proof. By Lemma 4.2, if the increasing (or decreasing ) rate of each w ∈ N F C(v) is the same, we
get that:
16
1
σj (uv,t
+
db
|N L(Tv )|
) · v,t+1
k
dh
So, we get that:
dbw,t+1
dh
=
=
1
σj
dbw,t+1
dh
·
P
·
(uw,t +
w∈N F C(v)
|N L(Tw )|
)
k
=
dbw,t+1
dh
· (uv,t +
|N L(Tv )|
).
k
dbv,t+1
dh
The proof for Lemma 4.5 is as follows:
Proof. By lemma 4.2, in order to keep the amount uv,t unchanged, we get:
P
dbw1 ,t+1
|N L(Tw1 )|
dbw,t+1
w )|
· (uw1 ,t +
)+
· (uw,t + |N L(T
) = 0.
dh
k
dh
k
Thus,
db 1 ,t+1
( wdh
db
−
db
dbw′ ,t+1
)
dh
w∈N F C(v)\{w1 }
|N L(Tw1 )|
· (uw1 ,t +
)
k
′
1 ,t+1
− wdh,t+1 ) · (uw1 ,t +
So, ( wdh
Hence, we get the claim.
|N L(Tw1 )|
)
k
+
+
dbw′ ,t+1
dh
dbw′ ,t+1
dh
·
·
P
(uw,t +
w∈N F C(v)
v )|
(uv,t + |N L(T
)
k
|N L(Tw )|
)
k
=0
=0
The proof for Theorem 4.6 is as follows:
Proof. When request pt arrives at time t, we move mass upt ,t from pt through its ancestor nodes to
other non-full leaves nodes S \ {pt }, i.e. upt ,t is decreased and up,t is increased for any p ∈ S \ {pt }.
Since these mass moves out of each subtree TA(pt ,j) for each 1 ≤ j ≤ ℓ, uA(pt ,j),t is decreased. By
b
v )|
t ,j),t+1
(exp( A(p
uA(pt ,j),t = f (bA(pt ,j),t+1 ) = |N L(T
ln(1 + k)) − 1) (we need to keep this relation
k
2D(v)
during the algorithm), bA(pt ,j),t+1 also decreases for each 1 ≤ j ≤ ℓ. On the other hand, up,t is
increased for each p ∈ S \ {pt }. Thus, for each non-full node v whose Tv doesn’t contain pt , its mass
uv,t is also increased. For each non-full node v whose Tv doesn’t contain pt , it must be a sibling of
some node A(pt , j) where j ∈ {ℓ1 , . . . , ℓh }. We assume that all siblings v ′ of any node v increase
bv′ ,t+1 at the same rate
In the following, we will compute the increasing (or decreasing) rate of all dual variables in
b
t ,j),t+1
be the decreasing rate of
the weighted σ-HST regarding aS . For 1 ≤ j ≤ ℓ, let ∇bj = − A(pda
S
b
bA(pt ,j),t+1 regarding aS . For each j ∈ {ℓ1 + 1, ℓ2 + 1, . . . , ℓh + 1}, let ∇b′j = w,t+1
daS be the increasing
rate of bw,t+1 for any siblings w ∈ N F C(A(pt , j)) of A(pt , j) regarding aS .
Using from top to down method, we can get a set of equations about the quantities ∇bj and
∇b′j . First, we consider the siblings of A(pt , ℓ1 + 1) ( i.e. those nodes are children of A(pt , ℓ1 ),
but they are not A(pt , ℓ1 + 1)). Let w be one of these siblings. If bw,t+1 is raised by ∇b′ℓ1 +1 , by
Corollary 4.4, the sum of ∇b′ on any path from a leaf in Tw to w must be φ(ℓ1 + 1) · ∇b′ℓ1 +1 . Since
aS is increasing with rate 1, it forces φ(ℓ1 + 1) · ∇b′ℓ1 +1 = 1 in order to maintain the dual constraint
(3) tight for non-full leaf nodes in Tw .
This considers the dual constraints for these non-full leaf nodes. Now, this increasing mass
must be canceled out by decreasing the mass in TA(pt ,ℓ1 +1) since the mass uA(pt ,ℓ1 ),t in TA(pt ,ℓ1 ) is
not changed. Thus, in order to maintain the “Node Identity Property” of A(pt , ℓ1 ), by Lemma 4.5,
we must set ∇b1 such that:
∇b′ℓ1 +1
= (∇bℓ1 +1 +
|NL(T(A(p ,ℓ +1)) )|
t 1
)
k
|S|
uA(pt ,ℓ1 ),t + k
|NL(T(A(p ,ℓ )) )|
t 2
(uA(pt ,ℓ2 ),t +
)
k
|S|
ur + k
∇b′ℓ1 +1 )
= (∇bℓ1 +1 + ∇b′ℓ1 +1 ) ·
·
(uA(pt ,ℓ1 +1),t +
For siblings of node A(pt , ℓ2 +1), we use the similar argument. Let w be a sibling of A(pt , ℓ2 +1).
Consider a path from a non-full leaf node in Tw to the w. Their dual constraint (3.6) already grows
17
at rate 1 + φ(ℓ1 + 1)∇bℓ1 +1 . This must be canceled out by increasing bw,t+1 , and if bw,t+1 is raised
by ∇b′ℓ2 +1 , by corollary 4.4, the sum of ∇b′ on any path from a leaf in Tw to w must be φ(2) · ∇b′2 .
Thus, ∇b′2 must be set such that: φ(ℓ2 + 1) · ∇b′ℓ2 +1 = 1 + φ(ℓ1 + 1) · ∇bℓ1 +1 .
Again, this increasing mass must be canceled out by decreasing the mass in TA(pt ,ℓ2 ) . In order
to keep the “Node Identity Property” of A(pt , ℓ2 ), by Lemma 4.5. we must set ∇bℓ2 +1 such that:
|NL(TA(p ,ℓ +1) )|
t 2
)
k
|NL(T(A(p ,ℓ ) )|
t
2
(uA(pt ,ℓ2 ),t +
)
k
|NL(TA(p ,ℓ ) )|
t
3
(uA(pt ,ℓ3 ),t +
)
k
|NL(T(A(p ,ℓ ) )|
t
2
(uA(pt ,ℓ2 ),t +
)
k
∇b′ℓ2 +1 = (∇bℓ2 +1 + ∇b′ℓ2 +1 ) ·
= (∇bℓ2 +1 + ∇b′ℓ2 +1 ) ·
(uA(pt ,ℓ2 +1),t +
Continuing this method, we obtain a system of linear equations about all ∇bj and ∇b′j (j ∈
{ℓ1 +1, ℓ2 +1, . . . , ℓh +1}). For maintaining the dual constraints tight, we get the following equations:
φ(ℓ1 + 1) · ∇b′ℓ1 +1 = 1
φ(ℓ2 + 1) · ∇b′ℓ2 +1 = 1 + φ(ℓ1 + 1) · ∇bℓ1 +1
..
.
h−1
P
ψ(ℓh + 1) · ∇b′ℓh +1 = 1 +
φ(ℓi + 1)∇bℓi +1
i=1
For keeping the node identity property, we get the following equations:
∇b′ℓ1 +1
= (∇bℓ1 +1 +
∇b′ℓ1 +1 )
·
∇b′ℓ2 +1 = (∇bℓ2 +1 + ∇b′ℓ2 +1 ) ·
..
.
∇b′ℓh +1
= (∇bℓh +1 +
∇b′ℓh +1 )
·
|NL(T(A(p ,1)) )|
t
)
k
|S|
ur + k
|NL(T(A(p ,ℓ )) )|
t 3
)
(uA(pt ,ℓ3 ),t +
k
|NL(T(A(p ,ℓ )) )|
t
2
(uA(pt ,ℓ2 ),t +
)
k
(uA(pt ,ℓ2 ),t +
|NL(T(A(p ,ℓ)) )|
t
)
k
|NL(T(A(p ,ℓ )) )|
t h
(uA(pt ,ℓh ),t +
)
k
(uA(pt ,ℓ),t +
.
We continue to solve the system of linear equations.
For each 1 ≤ j ≤ h,
j−1
P
φ(ℓj + 1) · ∇b′ℓj +1 = 1 +
φ(ℓi + 1)∇bℓi +1
=1+
j−2
P
i=1
i=1
φ(ℓi + 1)∇bℓi +1 + φ(ℓj−1 + 1)∇bℓj−1 +1
= φ(ℓj−1 + 1) · ∇b′ℓj−1 +1 + φ(j − 1)∇bℓj−1 +1
= φ(ℓj−1 + 1) · (∇b′ℓj−1 +1 + ∇bℓj−1 +1 ).
Since
∇b′ℓj−1 +1
= (∇bℓj−1 +1 +
∇b′ℓj−1 +1 ) ·
ψ(ℓj + 1)∇b′ℓj +1 = ψ(ℓj−1 + 1)∇b′ℓj−1 +1 ·
|NL(T(A(p ,ℓ )) )|
t j
)
k
|NL(T(A(p ,ℓ
)|
t j−1 ))
(uA(pt ,ℓj−1 ),t +
)
k
|NL(T(A(p ,ℓ
)|
))
t j−1
)
(uA(pt ,ℓj−1 ),t +
k
|NL(T(A(p ,ℓ )) )|
t j
(uA(pt ,ℓj ),t +
)
k
(uA(pt ,ℓj ),t +
, we get:
Solving the recursion, we get:
∇b′ℓj +1 =
ur + |S|
k
φ(ℓj +1)·(uA(pt ,ℓj ),t +
|NL(T(A(p ,ℓ )) )|
t j
)
k
|S|
∇bℓj +1 =
=
|N L(T(A(pt ,ℓj )) )|
|N L(T(A(pt ,ℓj+1 )) )|
)−1 − (uA(pt ,ℓj ),t +
)−1 ]
k
k
|S|
|N L(T(A(pt ,ℓj +1)) )|
|N L(T(A(pt ,ℓj )) )|
ur + k
)−1 − (uA(pt ,ℓj ),t +
)−1 ]
k
k
φ(ℓj +1) · [(uA(pt ,ℓj +1),t +
ur + k
φ(ℓj +1)
· [(uA(pt ,ℓj+1 ),t +
18
The proof for Theorem 4.7 is as follows:
Proof. Let P denote the value of the objective function of the primal solution and D denote the
value of the objective function of the dual solution. Initially, let P = 0 and D = 0. In the following,
we prove three claims:
(1) The primal solution produced by the algorithm is feasible.
(2) The dual solution produced by the algorithm is feasible.
(3) P ≤ 2 ln(1 + k)D.
By three claims and weak duality of linear programs, the theorem follows immediately.
The proof of claim (1) and (2) are similar to that of claim (1) and (2) in section 3.7.
Third, we prove claim (3) as follows. If the algorithm increases the variables aS,t at some time
∂D
= |S| − k. Let’s compute the primal cost. At depth j ∈ {ℓ1 + 1, ℓ2 + 1, . . . , ℓh + 1},
t, then: ∂a
S,t
we compute the movement cost of our algorithm by the change of ∇b′j as follows.
P
db′
du
2D(w) · dbw,t
· aSj
′
j
w∈N F C(A(pt ,j−1))\{A(pt ,j)}
= ∇b′j
P
2D(w) ·
w∈C(A(pt ,j−1))\{A(pt ,j)}
=
|S|
ur,t + k
φ(j)
P
=
p∈S
=
up,t + |S|
k
φ(j)
P
+
|N L(Tw )|
)
k
P
w )|
)
(uw,t + |NL(T
k
w∈C(A(pt ,j−1))\{A(pt ,j)}
|NL(TA(p ,j−1) )|
t
uA(pt ,j−1),t +
k
P
w )|
(uw,t + |NL(T
)
k
w∈C(A(pt ,j−1))\{A(pt ,j)}
|NL(TA(p ,j−1) )|
t
uA(pt ,j−1),t +
k
· ln(1 + k) ·
P
w )|
)
(uw,t + |NL(T
k
w∈C(A(pt ,j−1))\{A(pt ,j)}
|NL(TA(p ,j−1) )|
φ(j)
t
uA(pt ,j−1),t +
k
P
P
|S|−1
w )|
(uw,t + |NL(T
(
up,t + k +2(|S|−k))
)
k
w∈C(A(pt ,j−1))\{A(pt ,j)}
p∈S\{pt }
|NL(TA(p ,j−1) )|
φ(j)
t
uA(pt ,j−1),t +
k
P
|T (w)|
(uw,t + k )
w∈C(A(pt ,j−1))\{A(pt ,j)}
(|S|−1)
|NL(TA(p ,j−1) )|
k
t
uA(pt ,j−1),t +
k
P
w )|
)
(uw,t + |NL(T
k
w∈C(A(pt ,j−1))\{A(pt ,j)}
|NL(TA(p ,j−1) )|
t
uA(pt ,j−1),t +
k
(
<
· ln(1 + k) ·
ln(1+k)
2D(w) (uw,t
p∈S\{pt }
+2)
up,t + |S|−1
k
· ln(1 + k) ·
· ln(1 + k) ·
≤ (3(|S| − k) +
) · ln(1 + k) ·
≤ 4 · ln(1 + k) · (|S| − k) ·
≤ 4 · ln(1 + k) · (|S| − k), since
P
(uw,t +
w∈C(A(pt ,j−1))\{A(pt ,j)}
|N L(T
)|
A(pt ,j−1)
.
uA(pt ,j−1),t +
k
Where the first inequality holds since
|N L(Tw )|
)
k
≤
P
up,t < |S| − k, the reason is that the constraint
P
up,t =
at time t is not satisfied otherwise the algorithm stop increasing the variable up,t (Since
p∈S
P
up,t < |S| − k ⇔ upt ,t = 0, i.e the algorithm stop increasing the variables). In
|S| − k,
p∈S\{pt }
p∈S\{pt }
addition, when |S| ≥ k + 1, |S|−1
= |S| − k.
k
Thus, the total cost of all j depth is at most 4ℓ · ln(1 + k) · (|S| − k). Hence, we get P ≤
4ℓ̇ · ln(1 + k) · D.
19
So, the competitive ratio of this algorithm is 4ℓ ln(1 + k).
20
| 8cs.DS
|
1
On the Exact Solution to a
Smart Grid Cyber-Security Analysis Problem
I. I NTRODUCTION
A modern society relies critically on the proper operation of
the electric power distribution and transmission system, which
is supervised and controlled through Supervisory Control
And Data Acquisition (SCADA) systems. Through remote
terminal units (RTUs), SCADA systems measure data such
as transmission line power flows, bus power injections and
part of the bus voltages, and send them to the state estimator
to estimate the power network states (e.g., the bus voltage
phase angles and bus voltage magnitudes). The estimated
states are used for vital power network operations such as
optimal power flow (OPF) dispatch and contingency analysis
(CA) [1], [2]. See Fig. 1 for a block diagram of the above
functionalities. Any malfunctioning of these operations can
delay proper reactions in the control center, and lead to
significant social and economical consequences such as the
northeast US blackout of 2003.
The technology and the use of the SCADA systems have
evolved a lot since the 1970s when they were introduced. The
SCADA systems now are interconnected to office LANs, and
through them they are connected to the Internet. Hence, today
there are more access points to the SCADA systems, and also
more functionalities to tamper with. For example, the RTUs
can be subjected to denial-of-service attacks (A1 in Fig. 1).
The communicated data can be subjected to false data attacks
(A2). Furthermore, the SCADA master itself can be attacked
(A3). This paper focuses on the cyber security issue related
to false data attacks (A2), where the communicated metered
measurements are subjected to additive data attacks. A false
data attack can potentially lead to erroneous state estimates
The authors are with the ACCESS Linnaeus Center and the Automatic
Control Lab, the School of Electrical Engineering, KTH Royal Institute of
Technology, Sweden. {sou,hsan,kallej}@kth.se
This work is supported by the European Commission through the VIKING
project, the Swedish Research Council (VR) under Grant 2007-6350 and Grant
2009-4565 and the Knut and Alice Wallenberg Foundation.
Power Network
RTUs
RTUs
A2
AGC
Optimal Power Flow
.
.
.
.
.
.
EMS x
SSCADA Masster
Index Terms—Power network state estimation, security, operation research, optimization methods.
A1
State Estimatorr
Abstract—This paper considers a smart grid cyber-security
problem analyzing the vulnerabilities of electric power networks
to false data attacks. The analysis problem is related to a
constrained cardinality minimization problem. The main result
shows that an l1 relaxation technique provides an exact optimal
solution to this cardinality minimization problem. The proposed
result is based on a polyhedral combinatorics argument. It is
different from well-known results based on mutual coherence
and restricted isometry property. The results are illustrated on
benchmarks including the IEEE 118-bus and 300-bus systems.
SSCADA Masster
arXiv:1201.5019v2 [math.OC] 17 Sep 2012
Kin Cheong Sou, Henrik Sandberg and Karl Henrik Johansson
A3
Human
operator
Control center
Fig. 1: Block diagram of power network control center and
SCADA. RTUs connected to the substations transmit and
receive data from the control center using the SCADA system.
At the control center, a state estimate is computed and then
used by Energy Management Systems (EMS) to send out
commands to the power network. The human figures indicate
where a human is needed in the control loop. This paper
considers the false data attack scenario in A2.
by the state estimator, which can result in gross errors in OPF
dispatch and CA. In turn, these can lead to disasters of significant social and economical consequences. False data attack
on communicated metered measurements has been considered
in the literature (e.g., [3]–[10]). [3] was the first to point out
that a coordinated intentional data attack can be staged without
being detected by state estimation bad data detection (BDD)
algorithm, which is a standard part of today’s SCADA/EMS
system. [3]–[5], [7]–[10] investigate the construction problem
for such “unobservable” data attack, especially the sparse ones
involving relatively few meters to compromise, under various
assumptions of the network (e.g., DC power flow model [1],
[2]). In particular, [3] poses the attack construction problem as
a cardinality minimization problem to find the sparsest attack
including a given set of target measurements. [4], [5], [8] set
up similar optimization problems for the sparsest attack including a given measurement. [7], [10] seek the sparsest nonzero
attack and [9] finds the sparsest attack including exactly two
injection measurements. The solution information of the above
optimization problems can help network operators identify
the vulnerabilities in the network and strategically assign
2
protection resources (e.g., encryption of meter measurements)
to their best effect (e.g., [5], [6], [10]). On the other hand,
the unobservable data attack problem has its connection to
another vital EMS functionality, namely observability analysis
[1], [2]. In particular, solving the attack construction problem
can also solve an observability analysis problem (this is to be
explained in Section II-C). This connection was first reported
in [7], and was utilized in [11] to compute the sparsest critical
p-tuples for some integer p. This is a generalization of critical
measurements and critical sets [1].
To perform the cyber-security analysis in a timely manner,
it is important to solve the data attack construction problem
efficiently. This effort has been discussed, for instance, in [3]–
[5], [7]–[10]. The efficient solution to the attack construction
problem in [4] is the focus of this paper. The matching pursuit
method [12] employed in [3] and the basis pursuit method
[13] (l1 relaxation and its weighted variant) employed in [10]
are common efficient (i.e., polynomial-time) approaches to
suboptimally solve the attack construction problem. However,
these methods do not guarantee exact optimal solutions, and
in some cases they might not be sufficient (see for instance [8]
for a naive application of basis pursuit and its consequences).
While [7], [10] provide polynomial-time solution procedures
for their respective attack construction problems, the problems
therein are different from the one in this paper. Furthermore,
the considered problem in this paper cannot be solved as a
special case of [7], [10]. In particular, in [7] the attack vector
contains at least one nonzero entry. However, this nonzero
entry cannot be given a priori. [10] needs to restrict the number
of nonzero injection measurements attacked, while there is
no such requirement in the problem considered in this paper.
In [4], [5] a simple heuristics is provided to find suboptimal
solutions to the attack construction problem. This heuristics,
however, might not be sufficiently accurate. [8], [11], [23] is
most closely related to the current work. The distinctions will
be elaborated in Section IV-B.
The main conclusion of this paper is that basis pursuit
(i.e., l1 relaxation) can indeed solve the data attack construction problem exactly, under the assumption on the network metering system that no injection measurements are
metered. The limitations of this assumption will be discussed
in Section IV-A. In fact, the main result identifies a class
of cardinality minimization problems where basis pursuit can
provide exact optimal solutions. This class of problems include
as a special case the considered data attack construction
problem, under the assumption above.
Outline: Section II describes the state estimation model
and introduces the optimization problems considered in this
paper. Section III describes the main results of this paper – the
solution to the considered optimization problems. Section IV
compares the proposed result to related works. Section V
provides the proof the proposed main results. Section VI
numerically demonstrates the advantages of the proposed
results.
II. S TATE E STIMATION AND C YBER -S ECURITY A NALYSIS
O PTIMIZATION P ROBLEMS
A. Power network model and state estimation
A power network with n + 1 buses and ma transmission
lines can be described as a graph with n + 1 nodes and ma
edges. The graph topology can be specified by the (directed)
incidence matrix B0 ∈ R(n+1)×ma , in which the direction of
the edges can be assigned arbitrarily. The physical property
of the network is described by a nonsingular diagonal matrix
D ∈ Rma ×ma , whose nonzero entries are the reciprocals of
the reactance of the transmission lines.
The states of the network include bus voltage phase angles
and bus voltage magnitudes, the latter of which are typically
assumed to be constant (equal to one in the per unit system).
In addition, since one arbitrary bus is assigned as the reference
with zero voltage phase angle, the network states considered
n
can be captured in a vector θ ∈ [0, 2π) . The state estimator
estimates the states θ based on the measurements obtained
from the network. Under the DC power flow model [1], [2]
the measurement vector, denoted as z, is related to θ by
P DB T
z = Hθ + ∆z, where H ,
.
(1)
QBDB T
In (1), ∆z can be either a vector of random error or intentional
additive data attack (e.g., [3]). B ∈ Rn×ma is the truncated
incidence matrix (i.e., B0 with the row corresponding to
the reference node removed), and P consists of a subset
of rows of an identity matrices of appropriate dimension,
indicating which line power flow measurements are actually
taken. Together, P DB T θ is a vector of the power flows on the
transmission lines to be measured. Analogously, the matrix Q
selects the bus power injection measurements that are taken.
QBDB T θ is a vector of power injections at the buses to be
measured. Therefore, H is the measurement matrix, relating
the measured power quantities to the network states. The
number of rows of H is denoted m.
The measurements z and the network information H are
jointly used to find an estimate of the network states denoted
as θ̂. Assuming that the network is observable, it is wellestablished that the state estimate can be obtained using the
weighted least squares approach [1, Chapter 5] [2, Chapter 8]:
−1
θ̂ = (H T W H)
W H T z,
(2)
where W is a positive definite diagonal weighting matrix,
typically weighting more on the more accurate measurements.
The state estimate θ̂ is subsequently fed to other vital SCADA
functionalities such as OPF dispatch and CA. Therefore, the
accuracy and reliability of θ̂ is of paramount concern.
To detect possible faults in the measurements z, the BDD
test is commonly performed (see [1], [2]). In one typical
strategy, if the norm of the residual
−1
residual , z − H θ̂ = (I − H(H T W H)
W H T )∆z
is too big, then the BDD alarm will be triggered.
(3)
3
B. Unobservable data attack and security index
The BDD test is in general sufficient to detect the presence
of ∆z if it contains a single random error [1], [2]. However, in
face of a coordinated malicious data attack on multiple measurements the BDD test can fail. In particular, [3] considers
unobservable attack of the form
∆z = H∆θ
(4)
for an arbitrary ∆θ ∈ Rn . Since ∆z as defined in (4) would
result in a zero residual in (3), it is unobservable from the BDD
perspective. This was also experimentally verified in [14] in a
realistic SCADA system testbed. To quantify the vulnerability
of a network to unobservable attacks, [4] introduced the notion
of security index for an arbitrarily specified measurement. The
security index is the optimal objective value of the following
cardinality minimization problem:
minimize
n
kH∆θk0
subject to
H(k, :)∆θ = 1,
∆θ∈R
(5)
where k is given, indicating that the security index is computed
for measurement k. The symbol k · k0 denotes the cardinality
of a vector and H(k, :) denotes the k th row of H. The
security index is the minimum number of measurements an
attacker needs to compromise in order to attack measurement
k undetected. In particular, a small security index for a
particular measurement k means that in order to compromise k
undetected it is necessary to compromise only a small number
of additional measurements. This can imply that measurement
k is relatively easy to compromise in an unobservable attack.
As a result, the knowledge of the security indices allows the
network operator to pinpoint the security vulnerabilities of
the network, and to better protect the network with limited
resource. To model the case where certain measurements are
protected (hence cannot be attacked) [3], [5], [6], problem (5)
becomes:
minimize
kH∆θk0
n
∆θ∈R
subject to H(k, :)∆θ = 1
H(I, :)∆θ = 0,
(6)
where the protection index set I ⊂ {1, 2, . . . , m} is given,
H(I, :) denotes a submatrix of H with rows indexed by I.
By convention, the constraint H(I, :)∆θ = 0 is ignored when
I = ∅. Hence, (5) is a special case of (6).
C. Measurement set robustness analysis
J
subject to
kJ k0
rank(H(J¯, :)) < n
rank(H(J¯ ∪ {k}, :)) = n
k ∈ J.
I:
II:
H(k, :) 6= 0.
H has full column rank (= n).
The following three statements are true:
(a)
(b)
(c)
Problem (5) is feasible if and only if condition I is
satisfied.
Problem (7) is feasible if and only if conditions I and
II are satisfied.
If conditions I and II are satisfied, then (5) and (7)
are equivalent (see Definition 1 in Section V-A).
Note that if condition I is not satisfied, then the corresponding
measurement k should be removed from consideration. Also,
since measurement redundancy is a common practice in power
networks [1], [2], H can be assumed to have full column rank
(= n). Therefore, conditions I and II in Proposition 2.1 can be
justified in practice. Finally, note that Proposition 2.1 remains
true for arbitrary matrix H (not necessarily defined by (1)).
III. P ROBLEM S TATEMENT AND M AIN R ESULT
A. Problem statement
Problem (5) is also motivated from another important state
estimation analysis problem, namely observability analysis [1],
[2]. The measurement set, described by H in (1), is observable
if θ̂ can be uniquely determined by (2). An important question
of observability analysis is as follows:
minimize
In above, k is a given index and J¯ denotes the complement
of J (for index set I in the rest of the paper, Ī denotes its
complement). The meaning of (7) is as follows: J denotes
a subset of measurements from the measurement system
described by H. The condition that rank(H(J¯, :)) < n
means that the measurement system becomes unobservable
if the measurements associated with J are lost. That is, it
becomes impossible to uniquely determine θ̂ from H(J¯, :).
The problem in (7) seeks the minimum cardinality J which
must include a particular given measurement k. Therefore, if
there exist a measurement k which leads to an instance of
(7) with a very small objective value, then the measurement
system is not robust against meter failure. Special cases of (7)
have been extensively studied in the power system community.
For instance, the solution label sets of cardinalities one and
two are, respectively, referred to as critical measurements and
critical sets containing measurement k [1]. Their calculations
have been documented in, for example, [1], [15]–[18]. For the
more general cases where the minimum cardinality is p > 2,
the solution label set in (7) is a critical p-tuple which contains
the specified measurement k [11], [19]. Solving (5) solves (7)
as well. The justification is given by the following statement
inspired by [7], and proved in Appendix:
Proposition 2.1: Let H ∈ Rm×n and k ∈ {1, 2, . . . , m} be
given for problems (5) and (7). Denote the two conditions:
As discussed previously, this paper proposes an efficient
solution to the security index (i.e., attack construction) problem in (6). However, the proposed result focuses only on a
generalization of a special case of (6). In this special case, H
in (1) does not contain injection measurements:
H = P DB T .
(8)
(7)
The limitation of the assumption in (8) will be discussed in
Section IV-A, after the main result is presented.
4
In the Appendix, it is shown that the special case of (6)
with the assumption in (8) is equivalent to
P (Ī, :)B T ∆θ
minimize
n
∆θ∈R
subject to
0
P (k, :)B T ∆θ = 1
P (I, :)B T ∆θ = 0.
(9)
Instead of considering (9) directly, the proposed result pertains
to a more general optimization problem associated with a
totally unimodular matrix (i.e., the determinant of every square
submatrix is either −1, 0, or 1 [20]). In particular, the
following problem is the main focus of this paper:
minimize
n
x∈R
subject to
A(Ī, :)x
0
A(k, :)x = 1
A(I, :)x = 0,
(10)
where A ∈ Rm×n is a given totally unimodular matrix, and
k ∈ {1, 2, . . . , m} and I ⊂ {1, 2, . . . , m} are given. Since B
in (9) is an incidence matrix, P B T is a totally unimodular
matrix. Therefore, (10) is a generalization of (9). However,
neither (10) nor (6) includes each other as special cases.
C. Statement of main result
⋆
⋆
Theorem 3.1: Let (x⋆+ , x⋆− , y+
, y−
) be an optimal basic
feasible solution to (12), where A, k and I are defined in
(10). Then x⋆ , (x⋆+ − x⋆− ) is an optimal solution to (10).
Remark 3.1: Theorem 3.1 provides a complete procedure
for solving (10) via (12). If the standard form LP problem
in (12) is feasible, then it contains at least one basic feasible
solution (see the definition in Section V-A). Together with
the fact that the objective value is bounded from below (e.g.,
by zero), [22, Theorem 2.8, p. 66] implies that problem (12)
contains at least one optimal basic feasible solution, which can
be used to construct an optimal solution to (10) according to
Theorem 3.1. Conversely, if the feasible set of (12) is empty,
then the feasible set of (10) must also be empty because a
feasible solution to (10) can be used to construct a feasible
solution to (12).
Remark 3.2: To ensure that an optimal basic feasible solution to (12) is found if one exists, the simplex method (e.g.,
[22, Chapter 3]) can be used to solve (12).
The proof of Theorem 3.1 will be given in Section V. Before
that, the related work are reviewed, and the assumption in (8)
is discussed.
IV. R ELATED W ORK
B. l1 relaxation
Problem (10) is a cardinality minimization problem. In
general, no efficient algorithms have been found for solving
cardinality minimization problems [21], so heuristic or relaxation based algorithms are often considered. The l1 relaxation
(i.e., basis pursuit [13]) is a relaxation technique which has
received much attention. In l1 relaxation, instead of (10), the
following optimization problem is set up and solved:
minimize
n
x∈R
subject to
A(Ī, :)x
1
A(k, :)x = 1
A(I, :)x = 0,
(11)
where in the objective function in (11) the vector 1-norm
replaces the cardinality in (10). Problem (11) can be rewritten
as a linear programming (LP) problem in standard form [22,
pp. 4-6, p.17]:
minimize
x+ ,x− ,y+ ,y−
|Ī|
P
j=1
y+ (j) + y− (j)
A(Ī, :)(x+ − x− ) = y+ − y−
A(k, :)(x+ − x− ) = 1
A(I, :)(x+ − x− ) = 0
|Ī|
|Ī|
x+ ∈ Rn+ , x− ∈ Rn+ , y+ ∈ R+ , y− ∈ R+ ,
(12)
where |Ī| denotes the cardinality of the index set Ī.
If (x+ , x− , y+ , y− ) is a feasible solution to (12), then
x , x+ − x− is feasible to (10). Hence, an optimal solution
to (12), if it exists, corresponds to a suboptimal solution to
the original problem in (10). An important question is under
what conditions this suboptimal solution is actually optimal to
(10). An answer is provided by our main result, based on the
special structure in (10) and the fact that matrix A is totally
unimodular.
subject to
A. Rationale of the no injection assumption in (8)
Consider the case of (6) where I corresponds only to line
power flow measurements, then with the definition of H in
(1) it can be verified that (6) is equivalent to the following:
minimize
n
∆θ∈R
P (Ī, :)B T ∆θ
0
+ QB T DB∆θ
0
(13)
T
subject to P (k, :)B ∆θ = 1
P (I, :)B T ∆θ = 0.
This indicates that the considered problem in (9) is a relaxation [22] of the general case in (13). [8] utilizes this
observation and obtains satisfactory suboptimal solution to
(6). Alternatively, [11] considers indirectly accounting for
the term QB T DB∆θ 0 in the objective function of (13).
[11] demonstrates that solving the following problem provides
satisfactory suboptimal solution to (13)
minimize
n
∆θ∈R
subject to
¯ :)B T ∆θ
P̃ (Ĩ,
0
P̃ (k̃, :)B T ∆θ = 1
(14)
P̃ (Ĩ, :)B T ∆θ = 0,
with appropriately defined P̃ , I˜ and k̃. Notice that (14) has the
same form as (9). In conclusion, the “no injection assumption”
in (8) which leads to (9) introduces limitation, but it need
not be as restrictive as it might appear. The proposed result
in Theorem 3.1 still leads to a LP based approach to obtain
suboptimal solutions to (13) (and hence (6)).
B. Relationship with minimum cut based results
Nevertheless, the main strength of the current result lies
in the fact that it solves problem (10) where the A matrix
is totally unimodular. (10) includes (9) as a special case
5
where the corresponding constraint matrix is a transposed
graph incidence matrix. This distinguishes the current work
with other ones such as [8], [11], [23] which specialize in
solving (9) using graph-based minimum cut algorithms (e.g.,
[24]). One example of A which is totally unimodular but
not associated with a graph is the matrix with consecutive
ones property (i.e., if either for each row or for each column,
the 1’s appear consecutively) [25]. For a possible application,
consider a networked control system [26], [27] with one
controller and n sensor nodes. Each node contains a scalar
state value, constant over a period of m time slots. The nodes
need to transmit their state values through a shared channel
to the controller. Each node can keep transmitting over an
arbitrary period of consecutive time slots. At each time slot,
the measurement transmitted to the controller is the sum of
the state values of all transmitting nodes. Denote z ∈ Rm as
the vector of measurements transmitted over all time slots,
and θ ∈ Rn as the vector of node state values. Then the
measurements and the states are related by z = Aθ, where
A ∈ Rm×n is a (0, 1) matrix with consecutive ones in the
each column. Solving the observability problem in (7) with
H = A can identify the vulnerable measurement slots, which
should have higher priority in communication.
C. Relationship with compressed sensing type results
Problem (10) can be written in a form more common in
the literature. Consider only the case where the null space of
AT is not empty (otherwise rank(A) = m and (10) is trivial).
With a change of decision variable z = Ax, (10) can be posed
as:
minimize
kz(Ī)k0
m
z∈R
subject to
Lz = 0
z(I) = 0
z(k) = 1,
(15)
where L has full rank and LA = 0, and z(Ī) denotes a
sub-vector of z containing the entries corresponding to the
index set Ī. (15) can be written as the cardinality minimization
problem considered, for instance, in [21], [28]–[30]:
minimize
z̃
kz̃k0
subject to Φz̃ = b,
(16)
with appropriately defined matrix Φ and vector b. In this
subsection, we restrict the discussion to the standard case.
That is, (16) is feasible and Φ is a full rank matrix with
more columns than rows. As (16) is well-studied, certain
conditions regarding when its optimal solution can be obtained
by l1 relaxation are known. For example, [28], [31] report
a sufficient condition based on mutual coherence, which is
denoted as µ(Φ) and defined as
T
µ(Φ) = max
i6=j
|Φ(:, i) Φ(:, j)|
.
kΦ(:, i)k2 kΦ(:, j)k2
(17)
The sufficient condition [30] states that if there exists a feasible
solution z̃ in (16) which is sparse enough:
1
1
1+
,
(18)
kz̃k0 <
2
µ(Φ)
then z̃ is the unique optimal solution to (16) and its l1 relaxation (i.e., problem (16) with kz̃k1 replacing kz̃k0 ). Another
well-known sufficient condition is based on the restricted
isometry property (RIP) [21], [32]. For any integer s, the RIP
constant δs of matrix Φ is the smallest number satisfying
(1 − δs )kxk22 ≤ kΦxk22 ≤ (1 + δs )kxk22
(19)
for all vector x such that kxk0 ≤ s. The RIP-based sufficient
condition
√ [32] states that if for some s, Φ has a RIP constant
δ2s < 2 − 1, then any z̃ satisfying Φz̃ = b and kz̃k0 ≤ s
is necessarily the unique optimal solution to both (16) and
its l1 relaxation. It has been shown that certain type of
randomly generated matrices satisfy the above conditions with
overwhelming probabilities (e.g., [21] provides a RIP-related
result). However, the above conditions might not apply to (10),
which is the focus of this paper. For instance, consider A in
(10) being a submatrix of the transpose of the incidence matrix
of the 6-bus power network from [33]:
1 −1 0
0
0
0
0 1 −1 0
0
0
0 0
0
1
−1
0
.
A=
0
0
0
0
1
−1
1 0
0
−1
0
0
0 1
0
0 −1 0
0 0
1
0
0 −1
Let k = 6, and I = ∅.
b are
1 1 −1
Φ = 1 0 −1
0 0 0
Then the corresponding Φ in (16) and
−1 −1 0
0 −1 1
0
0 1
1
0
0
0
b = 0 .
1
(20)
For this Φ, (17) implies that µ(Φ) = 1. Therefore, the sparsity
bound in (18) becomes kz̃k0 < 1. This is too restrictive
to be practical. Similarly, for all s ≥ 1, the RIP constants
δ2s are at least one because Φ(:, 1) = −Φ(:, 3). Hence the
RIP-based sufficient condition would not be applicable either.
Nevertheless, the failure to apply these sufficient conditions
here does not mean that it is impossible to show that l1
relaxation can exactly solve (16). The mutual coherence and
RIP-based conditions characterize when a unique optimal
solution exists for both (16) and its l1 relaxation, while in
this paper uniqueness is not required. Indeed, for (16) with
T Φ
and b defined in (20), both −1 0 0 −1 0 1 0 and
T
−1 1 0 0 0 1 0 are optimal (this can be verified
by inspection). Using the CPLEX LP solver [34] in MATLAB
to solve the l1 relaxation leads to the first optimal solution.
It is the main contribution of this paper to show that this is
the case in general when (16) is defined by (10), even though
the optimal solution might not be unique. The reason why the
proposed result is applicable is that it is based on a polyhedral
combinatorics argument, which is different from those of the
mutual coherence and RIP based results.
V. P ROOF
OF THE
M AIN R ESULT
A. Definitions
The proof requires the following definitions:
6
Definition 1: Two optimization problems are equivalent
if there is an one-to-one correspondence of their instances.
The corresponding instances either are both infeasible, both
unbounded or both have optimal solutions. In the last case,
it is possible to construct an optimal solution to one problem
from an optimal solution to the other problem and vice versa.
In addition, the two problems have the same optimal objective
value.
Definition 2: A polyhedron in Rp is a subset of Rp
described by linear equality and inequality constraints. A
standard form polyhedron (as associated with a standard form
LP problem instance) is specified by {θ Cθ = d, θ ≥ 0} for
some given matrix C and vector d.
Definition 3: A basic solution [22, p. 50] of a polyhedron
in Rp is a vector satisfying all equality constraints. In addition,
out of all active constraints p of them are linearly independent.
For a standard form polyhedron with a constraint matrix of full
row rank, basic solutions can alternatively be defined by the
following statement [22, p. 53]:
Theorem 5.1: Consider a polyhedron {θ Cθ = d, θ ≥ 0},
and assume that C ∈ Rl×p and C has full row rank. A vector
θ is a basic solution if and only if Cθ = d and there exists an
index set J ⊂ {1, 2, . . . , p}, with |J | = l, such that det(C(:
, J )) 6= 0 and θ(i) = 0 if i ∈
/ J.
Definition 4: A basic feasible solution [22, p. 50] of a
polyhedron is a basic solution which is also feasible. By
convention, the terminology “a basic feasible solution to a LP
problem instance” should be understood as a basic feasible
solution of the polyhedron which defines the feasible set of
the instance.
B. Proof
Two lemmas, key to the proof, are presented first. The first
lemma states that problem (12), as set up by l1 relaxation, has
integer-valued optimal basic feasible solutions.
⋆
⋆
Lemma 5.1: Let (x⋆+ , x⋆− , y+
, y−
) be an optimal basic feasible solution to (12). Then it holds that x⋆ (i) , (x⋆+ (i) −
x⋆− (i)) ∈ {−1, 0, 1} for all 1 ≤ i ≤ n. In addition,
y ⋆ (j) , |A(Ī(j), :)x⋆ | ∈ {0, 1} for all 1 ≤ j ≤ |Ī|, where
Ī(j) denotes the j th element of Ī.
Proof: Assume that the feasible set of (12) is nonempty,
otherwise there is no basic feasible solution (cf. Definition 4).
The following two claims are made:
A(k, :) cannot be a linear combination of the rows
of A(I, :).
(b)
There exists I ′ ⊂ I such that either I ′ = ∅ or the
rows of A(I ′ , :) are linearly independent. In addition,
in both cases A(I ′ , :)θ = 0 and A(I, :)θ = 0 define
the same constraints.
Claims (a) and (b) together imply that problem (12) can be
written as a standard form LP problem with a constraint matrix
with full row rank (i.e., matrix C below):
(a)
minimize
θ
fT θ
subject to Cθ = d
θ ≥ 0,
(21)
with
A(Ī, :) −A(Ī, :) −I|Ī| I|Ī|
C , A(I ′ , :) −A(I ′ , :)
0
0
A(k, :) −A(k,
:) 0 0
0n×1
x+
0
x−
0n×1
d , 0 θ ,
y+ f , 1|Ī|×1 ,
1
y−
1|Ī|×1
(22)
where I|Ī| is an identity matrix of dimension |Ī|, and 1 is a
vector of all ones.
To see the claims, first note that (a) is implied by the
feasibility of (12). For (b), If I = ∅ or A(I, :) = 0, then
set I ′ = ∅. Otherwise, there exists I ′ ⊂ I with the properties
that |I ′ | = rank(A(I, :)), A(I ′ , :) has linearly independent
rows and A(I, :) = SA(I ′ , :) for some matrix S. On the
other hand, A(I ′ , :) = S ′ A(I, :) for some matrix S ′ , because
I ′ ⊂ I. Hence, A(I, :)θ = 0 and A(I ′ , :)θ = 0 define the
same constraints. This shows (b).
The next step of the proof is to show that every basic
solution of (21) has its entries being either −1, 0 or 1. Denote
the matrix B1 as the first 2n columns of C, and let B̃1
be any square submatrix of B1 . If B̃1 has two columns (or
rows) which are the same or negative of each other, then
det(B̃1 ) = 0. Otherwise, B̃1 is a (possibly row and/or column
permuted) square submatrix of A, and A is assumed to be
totally unimodular. Hence, det(B̃1 ) ∈ {−1, 0, 1} and B1 is
totally unimodular. Next consider the matrix B defined as
A(Ī, :) −A(Ī, :) −I|Ī| I|Ī| 0
B , C d = A(I ′ , :) −A(I ′ , :)
0
0 0
A(k, :) −A(k, :)
0
0 1
−I|Ī| I|Ī| 0
= B1
0
0 0 .
0
0 1
Denote the number of rows and the number of columns of B as
mB and nB respectively. Let J ⊂ {1, 2, . . . , nB } be any set
of column indices of B such that |J | = mB (so that B(:, J ) is
square). If B(:, J ) contains only columns of B1 , then det(B(:
, J )) ∈ {−1, 0, 1} since B1 is totally unimodular. Otherwise,
by repeatedly applying Laplace expansion on the columns of
B(:, J ) which are not columns of B1 , it can be shown that
det(B(:, J )) is equal to the determinant of a square submatrix
of B1 , which can only be −1, 0 or 1. Hence, by Cramer’s rule
the following holds: If v is the solution to the following system
of linear equations
B(:, J ) v = B(:, nB ), J ⊂ {1, . . . , nB − 1},
|J | = mB , and det(B(:, J )) 6= 0,
(23)
then
v(j) ∈ {−1, 0, 1},
∀ j.
(24)
Theorem 5.1 and (23) together imply that the nonzero entries
of all basic solutions to (21) are either −1, 0 or 1. Therefore,
the basic feasible solutions, which are also basic solutions, to
the polyhedron in (21) also satisfy this integrality property.
7
⋆
⋆
Finally, let (x⋆+ , x⋆− , y+
, y−
) be an optimal basic feasible
solution. Then feasibility (i.e., nonnegativity) implies that
x⋆+ (j) ∈ {0, 1},
⋆
y+
(j) ∈ {0, 1},
x⋆− (j) ∈ {0, 1},
⋆
y−
(j) ∈ {0, 1}, ∀ j.
(25)
minimize
The minimization excludes the possibility that, at optimality,
⋆
⋆
y+
(j) = y−
(j) = 1. Hence, it is possible to define x⋆ and y ⋆
such that
x⋆ (i) , (x⋆+ (i) − x⋆− (i)) ∈ {−1, 0, 1} ∀ i
⋆
⋆
y ⋆ (j) , (y+
(j) + y−
(j)) = |A(Ī(j), :) x⋆ | ∈ {0, 1} ∀ j.
(26)
The second lemma is concerned with a restricted version of
(10) with an infinity norm bound as follows:
minimize
x
A(Ī, :) x
(27)
Lemma 5.2: Optimization problems (10) and (27) are
equivalent.
Proof: Suppose (10) is feasible, then it has an optimal
solution denoted as x⋆ . Let Īx⋆ ⊂ Ī be the row index set such
that A(j, :)x⋆ 6= 0 if and only if j ∈ Īx⋆ . Then it is claimed
that there exists a common optimal solution to both (10) and
(27) with the same optimal objective value. The argument
is as follows. The property of x⋆ implies the feasibility of
′
(12) , which is denoted as a variant of (12) with Ī replaced
by Īx⋆ . By [22, Corollary 2.2, p. 65], problem (12)′ , as a
standard form LP problem, has at least one basic feasible
solution. Furthermore, since the optimal objective value of
(12)′ is bounded from below (e.g., by zero), [22, Theorem 2.8,
′
p. 66] implies that (12) has an optimal basic feasible solu⋆
⋆
⋆
⋆
tion (x̃+ , x̃− , ỹ+ , ỹ− ) which is integer-valued as specified by
Lemma 5.1. Denote x̃⋆ , (x̃⋆+ − x̃⋆− ). Then x̃⋆ is feasible to
both (10) and (27) since Īx⋆ ⊂ Ī, |A(Īx⋆ , :)x̃⋆ | ∈ {0, 1}|Īx⋆ |
and k ∈ Īx⋆ . Also, A(Ī, :)x̃⋆ 0 = A(Īx⋆ , :)x̃⋆ 0 ≤
A(Īx⋆ , :)x⋆ 0 = A(Ī, :)x⋆ 0 , as the inequality is true
′
because x̃⋆ is an optimal solution to (12) . Hence x̃⋆ is optimal
to both (10) and (27), with the same objective value.
Conversely, suppose (10) is infeasible, then (27) is also
infeasible. This concludes that (10) and (27) are equivalent.
⋆
⋆
Proof of Theorem 3.1: Let (x⋆+ , x⋆− , y+
, y−
) be an
optimal basic feasible solution to (12). Then there exist x⋆ and
y ⋆ as defined in Lemma 5.1. In particular, x⋆ = (x⋆+ − x⋆− ).
It can be verified that (x⋆ , y ⋆ ) is an optimal solution to the
following optimization problem:
minimize
x,y
subject to
|Ī|
P
y(i)
j=1
A(Ī, :)x ≤ y
−A(Ī, :)x ≤ y
A(k, :)x = 1
A(I, :)x = 0
0 ≤ y(j) ≤ 1 ∀ j = 1, 2, . . . , |Ī|,
x,y
|Ī|
P
y(i)
j=1
subject to
A(Ī, :)x ≤ y
−A(Ī, :)x ≤ y
A(k, :)x = 1
A(I, :)x = 0
y(j) ∈ {0, 1} ∀ j = 1, 2, . . . , |Ī|.
(28)
It can be verified that (28) is equivalent to (27). Then
Lemma 5.2 states that (28) is also equivalent to (10). Consequently, (x⋆ , y ⋆ ) being an optimal solution to (28) implies that
|Ī|
P
(10) is feasible with optimal objective value being
y ⋆ (j). A
j=1
0
subject to A(k, :) x = 1
A(I, :) x = 0
kAxk∞ ≤ 1.
where the inequalities above hold entry-wise. Because of the
property that y ⋆ (j) ∈ {0, 1} for all j, (x⋆ , y ⋆ ) is also an
optimal solution to
feasible solution to (10) is x⋆ . Since y ⋆ (j) = |A(Ī(j), :)x⋆ | ∈
|Ī|
P
{0, 1} ∀j, it holds that A(Ī, :)x⋆ 0 =
y ⋆ (j). Hence, x⋆
j=1
is an optimal solution to (10).
VI. N UMERICAL D EMONSTRATION
As a demonstration, instances of the restricted security index
problem in (9) are solved with P being an identity matrix
and I being empty. The incidence matrix B describes the
topology of one of the following benchmark systems: IEEE
14-bus, IEEE 57-bus, IEEE 118-bus, IEEE 300-bus and Polish
2383-bus and Polish 2736-bus [35]. For each benchmark, (9) is
solved for all possible values of k (e.g., 186 choices in the 118bus case and 411 choices in the 300-bus case). Two solution
approaches are tested. The first approach is the one proposed.
It is denoted the l1 approach, and includes the following steps:
1) Set up the LP problem in (12) with A being B T .
2) Solve (12) using a LP solver (e.g., CPLEX LP). Let
⋆
⋆
(x⋆+ , x⋆− , y+
, y−
) be its optimal solution.
⋆
3) Define ∆θ = x⋆+ − x⋆− . It is the optimal solution to (9),
according to Theorem 3.1.
The second solution approach to (9) is standard, and it was
applied also in [8], [11]. This second approach is referred to as
l0 approach, as (9) is formulated into the following problem:
P
minimize
y(j)
∆θ, y
subject to
j
B T ∆θ
−B T ∆θ
T
B(:, k) ∆θ
y(j)
≤
≤
=
∈
My
My
1
{0, 1} ∀ j,
(29)
where M is a constant required to be at least B T ∞ = kBk1
(i.e., the maximum column sum of the absolute values of the
entries of B) [8]. Because of the binary decision variables in
y, (29) is a mixed integer linear programming (MILP) problem
[22]. It can be solved by a standard solver such as CPLEX.
The correctness of the l0 approach is a direct consequence
that (29) is a reformulation of (9). As a result, both the l1
and l0 approaches are guaranteed to correctly solve (9) by
theory. Fig. 2 shows the sorted security indices (i.e., optimal
objective values of (9)) for the four larger benchmark systems.
8
8
118 bus
300 bus
2383 bus
2736 bus
7
3
10
2383 bus
l0 approach
l1 approach
2736 bus
2
10
solve time (sec)
The security indices are computed using the l1 approach. As a
comparison, the security indices are also computed using the
l0 approach, and they are shown in Fig. 3. The two figures
reaffirm the theory that the proposed l1 approach computes
the security indices exactly. Fig. 2 (or Fig. 3) indicates that
the measurement systems are relatively insecure, as there exist
many measurements with very low security indices (i.e., equal
to 1 or 2).
1
300 bus
10
118 bus
57 bus
0
10
14 bus
−1
10
6
−2
security index
10
1
2
3
5
4
5
6
case number
Fig. 4: Solve-time for computing all security indices for
different benchmark systems
4
3
2
VII. C ONCLUSION
1
0
0
500
1000
1500
2000
2500
3000
3500
ranked measurement index
Fig. 2: Security indices using the l1 approach
8
118 bus
300 bus
2383 bus
2736 bus
7
security index
6
5
4
3
2
The cardinality minimization problem is important but in
general difficult to solve. An example is shown in this paper
as the smart grid security index problem in (6). The l1
relaxation has demonstrated promise but to establish the cases
where it provides exact solutions is non-trivial. Well-known
results based on mutual coherence and RIP provide sufficient
conditions under which a unique optimal solution solves both
the cardinality minimization problem and its l1 relaxation.
However, this paper identifies a class of application motivated
problems (as in (10)) which can be shown to be solvable by l1
relaxation, even though results based on mutual coherence and
RIP cannot make the assertion. In fact, the optimal solution to
(10) might not be unique. The key property that leads to the
conclusion of this paper is total unimodularity of the constraint
matrix. The total unimodularity of matrix A in (10) leads to
two important consequences. (10) is equivalent to its ∞-norm
restricted version in (27). Furthermore, (27) can be solved
exactly by solving the LP problem in (12), thus establishing
the conclusion that l1 relaxation exactly solves (10).
A PPENDIX
1
A. Proof of the equivalence between (6) and (9)
0
0
500
1000
1500
2000
2500
3000
3500
ranked measurement index
Fig. 3: Security indices using the l0 approach
In terms of computation time performances, it is well-known
that the l0 approach is much more time-consuming than the
l1 approach since a MILP problem is much more difficult
to solve than a LP problem of the same size [22]. Fig. 4
shows the solve-time for computing all security indices for
each benchmark system, using the l1 and l0 approaches. It
verifies that the proposed l1 approach is more effective. In the
above illustration, all computations are performed on a dualcore Windows machine with 2.4GHz CPU and 2GB of RAM.
Note that the constraint H(I, :)∆θ = 0 implies that
kH∆θk0 = H(Ī, :)∆θ 0 . Since P consists of rows of an
identity matrix and D is diagonal and nonsingular, for all
J ⊂ {1, 2, . . . , m}, there exists a diagonal and nonsingular
matrix DJ such that P (J , :)D = DJ P (J , :). In particular,
let Dkk be a positive scalar such that P (k, :)D = Dkk P (k, :
) = P (k, :)Dkk . The above implies that for all ∆θ,
P (k, :)B T ∆θ = 1 if and only if
P (k, :)Dkk B T (Ddd −1 ∆θ) = P (k, :)DB T (Ddd −1 ∆θ) = 1.
In addition, for all ∆θ
P (I, :)B T ∆θ = 0 if and only if
Dkk −1 DI P (I, :)B T ∆θ = P (I, :)DB T (Ddd −1 ∆θ) = 0.
9
Finally, for all ∆θ
T
P (Ī, :)B ∆θ
0
−1
T
DĪ P (Ī, :)B ∆θ
=
Dkk
=
P (Ī, :)DB T (Ddd −1 ∆θ)
0
0
.
Applying the definition of H in (1) and a change of decision
variable to Dkk −1 ∆θ shows that (6) and (9) are equivalent.
B. Proof of Proposition 2.1
Part (a) is trivial.
For the necessary part of (b), condition I is necessary
because if H(k, :) = 0 then rank(H(J¯, :)) = rank(H(J¯ ∪
{k}, :)) for all J (meaning that (7) is infeasible). Condition
II is also necessary because if rank(H) < n. then there does
not exist any J such that rank(H(J¯, :)) = n.
For the sufficiency part of (b), assume that conditions I
and II are satisfied. Then by part (a) problem (5) is feasible.
Hence it has an optimal solution denoted as θ⋆ . Define Jθ⋆ ⊂
{1, 2, . . . , m} such that p ∈ Jθ⋆ if and only if H(p, :)θ⋆ 6= 0.
By definition of Jθ⋆ , rank(H(Jθ⋆ , :)) < n. Also, k ∈ Jθ⋆
because H(k, :)θ⋆ = 1. If rank(H(Jθ⋆ ∪ {k}, :)) = n, then
Jθ⋆ is feasible to (7), thus showing that (7) is feasible. To
show this, first consider the case when kHθ⋆ k0 = 1. Then
Jθ⋆ = {k} and rank(H(Jθ⋆ ∪ {k}, :)) = rank(H) = n
because of condition II (i.e., H has full column rank). Next
consider the case when kHθ⋆ k0 > 1 (i.e., |Jθ⋆ \ {k}| > 0). If
rank(H(Jθ⋆ ∪ {k}, :)) < n, then there exists θ̃ 6= 0 such
that H(Jθ⋆ ∪ {k}, :)θ̃ = 0. In particular, H(k, :)θ̃ = 0.
Also, condition II implies that H(Jθ⋆ \ {k}, :)θ̃ 6= 0 (since
otherwise H θ̃ = 0). Let q ∈ Jθ⋆ \ {k} such that H(q, :)θ̃ 6= 0.
Note also that by definition of Jθ⋆ , H(q, :)θ⋆ 6= 0. Construct
θ′ , (H(q, :)θ̃)θ⋆ − (H(q, :)θ⋆ )θ̃. Then H(k, :)θ′ = 1,
H(p, :)θ′ = 0 whenever H(p, :)θ⋆ = 0, but H(q, :)θ′ = 0
while H(q, :)θ⋆ 6= 0. This implies that θ′ is feasible to (5) with
a strictly less objective value than that of θ⋆ , contradicting the
optimality of θ⋆ . Therefore, the claim that rank(H(Jθ⋆ ∪{k}, :
)) = n is true. This implies that Jθ⋆ is feasible to (7),
establishing the sufficiency of part (b).
For part (c), under conditions I and II both (5) and (7)
are feasible. In addition, Jθ⋆ constructed in the proof of
the sufficiency part of (b) satisfies |Jθ⋆ | = kHθ⋆ k0 , for θ⋆
being an optimal solution to (5). This means that the optimal
objective function value of (7) is less than or equal to that of
(5). For the converse, suppose that J ⋆ is optimal to (7), then
the feasibility of J ⋆ implies that there exists θJ ⋆ 6= 0 such that
H(J¯⋆ , :)θJ ⋆ = 0. This also implies that kHθJ ⋆ k0 ≤ |J ⋆ |.
If H(k, :)θJ ⋆ = 0, then H(J¯⋆ ∪ {k}, :)θJ ⋆ = 0. This
implies that rank(H(J¯⋆ ∪ {k}, :)) < n, contradicting the
feasibility of J ⋆ . Therefore, there exists a scalar α such that
H(k, :)(αθJ ⋆ ) = 1. Consequently, αθJ ⋆ is feasible to (5) with
an objective function value less than or equal to the optimal
objective function value of (7).
R EFERENCES
[1] A. Abur and A. Expósito, Power System State Estimation. Marcel
Dekker, Inc., 2004.
[2] A. Monticelli, State Estimation in Electric Power Systems A Generalized
Approach. Kluwer Academic Publishers, 1999.
[3] Y. Liu, M. Reiter, and P. Ning, “False data injection attacks against
state estimation in electric power grids,” in 16th ACM Conference on
Computer and Communication Security, New York, NY, USA, 2009, pp.
21–32.
[4] H. Sandberg, A. Teixeira and K.H. Johansson, “On security indices for
state estimators in power networks,” in First Workshop on Secure Control
Systems, CPSWEEK 2010, 2010.
[5] G. Dán and H. Sandberg, “Stealth attacks and protection schemes for
state estimators in power systems,” in IEEE SmartGridComm, 2010.
[6] R. Bobba, K. Rogers, Q. Wang, H. Khurana, K. Nahrstedt, and T. Overbye, “Detecting false data injection attacks on dc state estimation,” in
the First Workshop on Secure Control Systems, CPSWEEK 2010, 2010.
[7] O. Kosut, L. Jia, R. Thomas, and L. Tong, “Malicious data attacks on
the smart grid,” IEEE Transactions on Smart Grid, vol. 2, pp. 645–658,
2011.
[8] K.C. Sou and H. Sandberg and K.H. Johansson, “Electric power network
security analysis via minimum cut relaxation,” in IEEE Conference on
Decision and Control, December 2011.
[9] A. Giani, E. Bitar, M. McQueen, P. Khargonekar, and K. Poolla, “Smart
grid data integrity attacks: Characterizations and countermeasures,” in
IEEE SmartGridComm, 2011.
[10] T.T. Kim and H.V. Poor, “Strategic protection against data injection
attacks on power grids,” IEEE Transactions on Smart Grid, vol. 2, pp.
326–333, June 2011.
[11] K.C. Sou, H. Sandberg and K.H. Johansson, “Computing Critical ktuples in Power Networks,” IEEE Transactions on Power Systems,
vol. 27, no. 3, pp. 1511–1520, 2012.
[12] S. Mallat and Z. Zhang, “Matching pursuit with time-frequency dictionaries,” IEEE Transactions on Signal Processing, vol. 41, pp. 3397–
3415, 1993.
[13] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition
by basis pursuit,” SIAM Journal on Scientific Computing, vol. 20, 1998.
[14] A. Teixeira, G. Dan, H. Sandberg, and K. Johansson, “Cyber security
study of a scada energy management system: stealthy deception attacks
on the state estimator,” in IFAC World Congress, Milan, Italy, 2011.
[15] G. Korres and G. Contaxis, “Identification and updating of minimally
dependent sets of measurements in state estimation,” IEEE Transactions
on Power Systems, vol. 6, no. 3, pp. 999 –1005, aug 1991.
[16] M. de Almeida, E. Asada, and A. Garcia, “Identifying critical sets
in state estimation using gram matrix,” in PowerTech, 2009 IEEE
Bucharest, 28 2009-july 2 2009, pp. 1 –5.
[17] M. Ayres and P. H. Haley, “Bad data groups in power system state
estimation,” IEEE Transactions on Power Systems, vol. 1, no. 3, pp. 1
–7, aug. 1986.
[18] K. Clements, G. Krumpholz, and P. Davis, “Power system state estimation residual analysis: An algorithm using network topology,” Power
Apparatus and Systems, IEEE Transactions on, vol. PAS-100, no. 4, pp.
1779 –1787, april 1981.
[19] J. London, J.B.A., L. Alberto, and N. Bretas, “Network observability:
identification of the measurements redundancy level,” in Power System
Technology, 2000. Proceedings. PowerCon 2000. International Conference on, vol. 2, 2000, pp. 577 –582 vol.2.
[20] A. Schrijver, “A course in combinatorial optimization,” CWI,
Amsterdam, Netherlands, 2010, online document available from
http://homepages.cwi.nl/˜lex/.
[21] E. Candès and T. Tao, “Decoding by linear programming,” Information
Theory, IEEE Transactions on, vol. 51, no. 12, pp. 4203 – 4215, Dec.
2005.
[22] J. Tsitsiklis and D. Bertsimas, Introduction to Linear Optimization.
Athena Scientific, 1997.
[23] J. Hendrickx, K.H. Johansson, R. Jungers, H. Sandberg, and K.C. Sou,
“An exact solution to the power networks security index problem and
its generalized min cut formulation,” in preparation, [Online]. Available:
http://arxiv.org/abs/1204.6174.
[24] M. Stoer and F. Wagner, “A simple min-cut algorithm,” J. ACM, vol. 44,
pp. 585–591, July 1997.
[25] A. Schrijver, Theory of Linear and Integer Programming. Wiley, 1998.
[26] J. Hespanha, P. Naghshtabrizi, and Y. Xu, “A survey of recent results
in networked control systems,” Proceedings of the IEEE, vol. 95, no. 1,
pp. 138 –162, jan. 2007.
[27] A. Bemporad, M. Heemels, and M. Johansson, Networked Control
Systems. Springer, 2010.
[28] D. L. Donoho and M. Elad, “Optimally sparse representation in general
(nonorthogonal) dictionaries via l1 minimization,” Proceedings of The
National Academy of Sciences, vol. 100, pp. 2197–2202, 2003.
10
[29] E. Candès, M. Wakin, and S. Boyd, “Enhancing sparsity by reweighted
l1 minimization,” Journal of Fourier Analysis and Applications, vol. 14,
pp. 877–905, 2008.
[30] A. M. Bruckstein, D. L. Donoho, and M. Elad, “From sparse solutions of
systems of equations to sparse modeling of signals and images,” SIAM
Review, vol. 51, 2009.
[31] R. Gribonval and M. Nielsen, “Sparse representations in unions of
bases,” Information Theory, IEEE Transactions on, vol. 49, no. 12, pp.
3320 – 3325, dec. 2003.
[32] E. J. Candés, “The restricted isometry property and its implications for
compressed sensing,” Comptes Rendus Mathematique, vol. 346, no. 910,
pp. 589 – 592, 2008.
[33] A. Wood and B. Wollenberg, Power generation, operation, and control.
J. Wiley & Sons, 1996.
[34] “Cplex,” http://www-01.ibm.com/software/integration/optimization/cplex-optimizer/.
[35] R. Zimmerman, C. Murillo-Sánchez, and R. Thomas, “MATPOWER
Steady-State Operations, Planning and Analysis Tools for Power Systems Research and Education,” IEEE Transacations on Power Systems,
vol. 26, no. 1, pp. 12–19, 2011.
| 5cs.CE
|
Accelerating Learning in Constructive Predictive
Frameworks with the Successor Representation
arXiv:1803.09001v1 [cs.LG] 23 Mar 2018
Craig Sherstan1 , Marlos C. Machado1 , Patrick M. Pilarski1
Abstract— Here we propose using the successor representation
(SR) to accelerate learning in a constructive knowledge system
based on general value functions (GVFs). In real-world settings
like robotics for unstructured and dynamic environments, it is
infeasible to model all meaningful aspects of a system and its
environment by hand due to both complexity and size. Instead,
robots must be capable of learning and adapting to changes in
their environment and task, incrementally constructing models
from their own experience. GVFs, taken from the field of
reinforcement learning (RL), are a way of modeling the world
as predictive questions. One approach to such models proposes
a massive network of interconnected and interdependent GVFs,
which are incrementally added over time. It is reasonable
to expect that new, incrementally added predictions can be
learned more swiftly if the learning process leverages knowledge
gained from past experience. The SR provides such a means
of separating the dynamics of the world from the prediction
targets and thus capturing regularities that can be reused across
multiple GVFs. As a primary contribution of this work, we show
that using SR-based predictions can improve sample efficiency
and learning speed in a continual learning setting where new
predictions are incrementally added and learned over time. We
analyze our approach in a grid-world and then demonstrate its
potential on data from a physical robot arm.
I. INTRODUCTION
A long standing goal in the pursuit of artificial general intelligence is that of knowledge construction—incrementally
modeling and explaining the world and the agent’s interaction with it directly from the agent’s own experience. This is
particularly important in fields such as continual learning [1]
and developmental robotics [2], where we expect agents to
be capable of learning, dynamically and incrementally, to
interact and succeed in complex environments.
One proposed approach for representing such world models is a collection of general value functions (GVFs) [3],
which models the world as a set of predictive questions each
defined by a policy of interest, a target signal of interest
under that policy, and a timescale (discounting schedule) for
accumulating the signal of interest. For example, a GVF on
a mobile robot could pose the question “How much current
will my wheels consume over the next second if I drive
straight forward?”
GVF questions are typically answered using temporaldifference (TD) methods [4] from the field of reinforcement
learning (RL) [5]. A learned GVF approximates the expected
future value of a signal of interest, directly representing the
relationship between the environment, policy, timescale, and
target signal as the output of a single predictive unit.
1
University of Alberta, Canada
{sherstan, machado, pilarski}@ualberta.ca
Nevertheless, despite the success RL algorithms have
achieved recently (e.g., [6], [7]), methods for answering multiple predictive questions from a single stream of experience
(critical in a robotic setting) are known to exhibit sample
inefficiency. In our setting of interest, where multitudes of
GVFs are learned in an incremental, sample by sample way,
this problem is multiplied. Ultimately, the faster an agent can
learn to approximate a new GVF, the better.
In this paper we show how one can accelerate learning in
a constructive knowledge system based on GVFs by sharing
the environment dynamics across the different predictors.
This is done with the successor representation (SR) [8],
which allows us to learn the world dynamics under a policy
independently of any signal being predicted.
We empirically demonstrate the effectiveness of our approach on both a tabular representation and on a robot
arm which uses function approximation. We evaluate our
algorithm in the continual learning setting where it is not
possible to specify all GVFs a priori, but rather GVFs are
added incrementally during the course of learning. As a key
result, we show that using a learned SR enables an agent to
learn newly added GVFs faster than when learning the same
GVFs in the standard fashion without the use of the SR.
II. BACKGROUND
We consider an agent interacting with the environment
sequentially. We use the standard notation in the reinforcement learning (RL) literature [5], modeling the problem as
a Markov Decision Process. Starting from state S0 ∈ S,
at each timestep the agent chooses an action, At ∈ A,
according to the policy distribution π : S × A → [0, 1], and
transitions to state St+1 ∈ S according to the probability
At
transition p(·|St , At ). For each transition St −−→
St+1
the agent receives a reward, Rt , from the reward function
R(St , At , St+1 ) ∈ R.
In this paper we focus on the prediction problem in RL, in
which the agent’s goal is to predict the value of a signal from
its current state (e.g., the cumulative sum of future rewards).
Note that throughout this paper we use upper case letters to
indicate random variables.
A. General Value Functions (GVFs)
The most common prediction made in RL is about the
expected return. The return is defined to be the sum of future
discounted rewardsPunder policy π starting from state s.
T
t
Formally, Gt =
t=0 γ Rt , with γ ∈ [0, 1] being the
discount factor and T being the final timestep, where T =
∞ denotes a continuing task. The function encoding the
prediction about the return is known as the value function
vπ (s) = Eπ [Gt |S0 = s].
GVFs [3] extend the notion of predictions to different
signals in the environment. This is done by replacing the
reward signal, Rt , by any other target signal, which we refer
to as the cumulant, Ct , and by allowing a state-dependent
discounting function, γt = γ(St ), instead of using a fixed
discounting factor. The general value of state s under policy
π is defined as:
X
vπ:γ (s) = c̄π + γ̄π
pπ (s0 |s)vπ (s0 ),
(1)
s0
where c̄π is the average cumulant from state s, γ̄π is the
average γ from state s, and pπ (s0 |s) is the probability of
transitioning from state s to s0 under policy π. This can also
be written in matrix form: vπ:γ = c̄π + γ̄ π Pπ vπ:γ , where
denotes element-wise multiplication.. Such an equation,
when solved, gives us
vπ:γ = (I − γ̄ π
Pπ )−1 c̄π ,
(2)
where I is the identity matrix, v, c̄, γ̄ ∈ R|S|×1 , and
Pπ ∈ [0, 1]|S|×|S| is a probability matrix such that [Pπ ]ij =
E[pπ (St+1 = sj |St = si )].
B. The Successor Representation (SR)
The successor representation [8] was initially proposed as
a representation capable of capturing state similarity in terms
of time. It is formally defined, for a fixed γ < 1, as:
∞
i
hX
ψπ (s, s0 ) = Eπ
γ t 1{St =s0 } |S0 = s .
t=0
In words, the SR encodes the expected number of times the
agent will visit a particular state when the sum is discounted
by γ over time1 . It can be re-written in matrix form as
Ψπ =
∞
X
(γPπ )t = (I − γPπ )−1 .
(3)
t=0
Importantly, the SR can be easily computed, incrementally,
by standard RL algorithms such as TD learning [4], since its
primary modification is to replace the reward signal by a
state visitation counter. Nevertheless, despite its simplicity,
the SR holds important properties we leverage in this paper.
The SR, in the limit, for a constant γ (see Eq. 3),
corresponds to the first factor of the solution in Eq. 2. Thus,
the SR can be seen as encoding the dynamics of the Markov
chain induced by the policy π and by the environment’s
transition probability p. If the agent has access to the SR, it
can accurately predict the (discounted) accumulated value of
any signal, from any state, by simply learning the expected
immediate value, c̄π , of that signal in each state. On the
other hand, if the agent does not use the SR, the agent must
also deal with the problem of credit assignment, having to
look at n-step returns to control for delayed consequences.
1 Note that Dayan describes the SR as predicting future state visitation
from time t onward. This is non-standard in RL as we typically describe
the return as predicting the signal from t + 1 onward.
Importantly, the dynamics encoded by the SR
for all signals learned under the same policy
function. This factorization of the solution
property we use in our work, described in the
are the same
and discount
is the main
next section.
III. METHODS
As aforementioned, we are interested in the problem of
knowledge acquisition in the continual learning setting [1],
where knowledge is encoded as predictive questions (GVFs).
In this setting it is not possible to specify all GVFs ahead of
time. Instead, GVFs must be added incrementally by some,
as yet unknown, mechanism. The standard approach would
be to learn each newly added prediction from scratch. In this
section we discuss how we can use the SR to accelerate learning by taking advantage of the factorization shown in Eq. 2.
Our method leverages the fact that the SR is independent of
the target signal being predicted, learning the SR separately
and re-using it when learning to predict new signals.
In the previous section, for clarity, we discussed the main
concepts in the tabular case. In real world applications,
where the state space is too large, assuming states can be
uniquely identified is not often feasible. Instead, we generally
represent states as a set of features φ(s) ∈ Rd where
d |S|. Because both ψ and c̄ are a function of feature
vector φ(S), they can easily be represented using function
approximation and learned using TD algorithms. In order
to present a more general version of our algorithm, we
introduce it here using the function approximation notation.
The first step in our algorithm is to compute the one-step
average cumulant, which we do with TD error:
δt = Ct+1 − c̄ φ(St ) .
(4)
If we use
linear function approximation to estimate c̄ then
c̄ φ(s) = φ(s)> w. The TD error for ψ is given as (note
that δ is a vector of length n)
δ t = φ(St ) + γt+1 ψ φ(St+1 ) − ψ φ(St ) .
(5)
This generalization of the SR to the function approximation
case is known as successor features [9].
If we use linear function approximation to estimate ψ
then ψ φ(s) = M > φ(s), where M ∈ Rd×d . Using the
usual semi-gradient method,2 often used with TD, we derive
a TD(0) stochastic gradient descent update as:
1
Mt+1 = Mt − α∇M (δ 2t )
2
= Mt + α∇M ψ(St ) ⊗ δ t
= Mt + αφ(St ) ⊗ δ t ,
where ∇M is the gradient with respect to M and ⊗ is
the outer product. Based on this derivation, as well as
Eq. 4 and 5, we obtain Algorithm 1. Note that the last two
lines of Algorithm 1, which update the SR for the state S 0 ,
are only required for the episodic case.
2 In semi-gradient methods, the effect of M on the prediction target
φ(St+1 ) + γψ φ(St+1 ) is ignored when computing the gradient.
Algorithm 1
TABLE I: Signal Primitives
GVF prediction with the SR
Input: Feature representation φ, policy π, discount function γ, and step-sizes αC , αSR
Output: Matrix M and vectors c̄i as predictors of Ci
Initialize w and M arbitrarily
while S 0 is not terminal do
Observe state S, take action A selected according to
π(S), and observe a next state S 0 and the cumulants Ci
δSR = φ(S) + γ(S 0 )M > φ(S 0 ) − M > φ(S)
M ← M + αSR φ(S) ⊗ δSR
for each cumulant Ci do
δCi = Ci − φ(S)> wi
wi ← wi + αCi φ(S)δCi
end for
end while
δSR = φ(S 0 ) − M > φ(S 0 )
M ← M + αSR φ(S 0 ) ⊗ δSR
This algorithm allows us to predict the cumulant Ci , in
state S, using the current estimate of the matrix M and the
weights wi . We can then obtain the final prediction by simply
>
>
>
computing ψ φ(S) w = M > φ(s) w = φ(s) M w.
This algorithm accelerates learning because, generally,
learning to estimate c̄ is faster than learning the GVF directly.
This is exactly what our algorithm does. When predicting a
new signal, it starts with its current estimate for the SR,
Ψ. At the end, the multiplication Ψc̄ is simply a weighted
average of the one-step predictions across all states, weighted
by the likelihood they will be visited. We provide empirical
evidence supporting this claim in the next sections.
IV. EVALUATION IN DAYAN’S GRID WORLD
We first evaluated our algorithm in a tabular grid world. Its
simplicity allowed us to analyze our method more thoroughly
since we were not bounded by the speed and complexity of
physical robots. The grid world we used was inspired by
Dayan’s [8] (see Figure 1a). Four actions are available in this
environment: up, down, left, right. Taking an action into a
wall (blue) results in no change in position. Transitions are
deterministic, i.e., a move in any direction moves the agent to
the next cell in the given direction, except when moving into
a wall. For each episode the agent spawns at location S and
the episode terminates when the agent reaches the goal G.
We generated fifty different signals for the agent to
predict. They were generated randomly from a collection
of primitives enumerated in Table I. They are composed
of two different primitives, one for each axis, like so:
(sigx (x + of f setx ) + biasx ) ∗ (sigy (y + of f sety ) + biasy ).
The bias and offset were drawn from [−2.0, 2.0) and [0, 10),
respectively. Offset and bias were not applied to either the
unit or shortest path primitives. Further, the shortest path
primitive was not combined with a second signal, but was
Primitive
Fixed value
Square wave
Sin wave
Random
binary
Random float
Unit
Shortest path
Parameters
value ∈ [−2.0, 2.0)
period ∈ [2, 40), invert ∈ [T rue, F alse]
period ∈ [2, 40)
Fixed random binary string generated over
the length of an axis.
Fixed random floats ∈ [0, 1.0) generated
over the length of an axis.
Fixed value of 1.0
transition cost ∈ [−10.0, −1.0),
goal reward ∈ [1.0, 10.0)
used on its own. Gaussian noise with a standard deviation
of 0.3 was applied on top of each signal. The shortest path
signal is inspired by a common reward function used in RL
where each transition has a cost (a negative reward) meant
to push the agent to completing a task in a timely manner
and reaching the goal produces a positive reward signal.
Our agent selects actions using -greedy action selection
where, at each timestep, with probability 1 − , it uses the
action specified by a hand-coded policy (see Figure 1a), and
otherwise chooses randomly from all four actions ( = 0.3
in our experiments). A tabular representation is used with
each grid cell uniquely represented by a one-hot encoding.
In this set of experiments we can compute the groundtruth predictors for the SR and the signal predictors. This is
done by taking the average return observed from each state.
The SR reference was averaged over 30,000 episodes and
the signal predictor references were averaged over 10,000
episodes. Each episode started at the start state and followed
the -greedy policy already described.
We first evaluated the predictive performance of our SR
learning algorithm with respect to the step-size for a variety
of γ values. We report the average over 30 trials of 10,000
episodes. We initialize the SR weights to 0.0. The squared
Euclidean distance was calculated between the predicted SR
and the reference SR for each timestep. These values were
summed over the run and the average was taken across the
runs. These averages are shown in Figure 1b.
Using the results in Figure 1b we evaluated the performance of the two signal prediction approaches by sweeping
across step-sizes for various γ. For each experimental run,
learning of a new signal was enabled incrementally every 50
episodes. This produced runs with a total length of 2,500
episodes where the first pair of GVFs added (the direct and
one-step predictors) were trained for 2,500 episodes and the
last added GVFs were trained for 50 episodes. Further, for
each run, the order in which the signals were added was
randomized. Thirty runs were performed. The weights of the
predictors and of the SR were initialized to 0.0. Notice that
the SR was being learned at the same time as the direct and
one-step predictions.
For each run a cumulative MSE for each signal i was
calculated according to Eq. 6. This equation computes the
total squared error between the predictor’s estimate, V and
the reference predictor’s estimate, V ∗ . For each episode, E,
the error of the current and previous episodes is averaged.
Then, for each signal, the maximum error for a given γ,
(a)
(b)
(c)
Fig. 1: a) Dayan’s grid world. Arrows indicate the hand-coded policy (from the start state the policy is to go up). Black
squares indicate the SR prediction given from the starting state, S, for γ = 1.0; the darker the square the higher the expected
visitation. Notice the graying around the central path caused by the -greedy action selection. b) MSE of the SR as a function
of step-size α for different values of γ. Lowest error is indicated by the markers. c) A comparison of the NMSE of the direct
(dashed lines) and SR-based (solid lines) predictions as a function of fixed step-size α and discount factor γ summed across
all signals. Lowest cumulative error is indicated by up arrows (direct predictions) and down arrows (SR-based predictions).
Note, that although difficult to see, confidence intervals of 95% are included in both b and c.
either in the direct or SR-based predictions, is found and used
to normalize the errors in the signal across that particular
value of γ (see Eq. 7). In this way we attempt to treat
the error of each signal equally. If this is not done the
errors of large magnitude signals dominate the results. These
normalized values are then summed across the signals and
the averages across all 30 runs are plotted in Figure 1c.
M SEi =
E T
1 XX
∗ 2
(Vi:t − Vi:t
)
E e t
(6)
M SEi:α:γ
maxα (M SEi:γ:direct , M SEi:γ:SR )
(7)
0
N M SEi:α:γ =
The advantage of the SR-based method is clear as γ
increases. This is to be expected since for γ = 0.0 both
methods are making one-step predictions. In the experiment
of Figure 1c, the SR performs better in the vast majority of
the signals, as shown in Table II. For all step-sizes not listed
the SR-based method was better on all signals. Analysis
of these cases where the direct method did better reveal
that some of the target signals have very small magnitudes,
suggesting the SR-based approach may be more susceptible
to signal-to-noise ratio. Further analysis remains to be done.
Finally, we analyzed how the prediction error of our
systems evolve with time. This is demonstrated in Figure 2
where we selected the best step-sizes for γ = 0.9 and
plotted the performance over time across 30 different runs.
In this case the order of the signals remained fixed so that
sensible averages could be plotted for each signal. Signal
performance was normalized as before and summed across
TABLE II: Signal performance for γ = 0.9 of Figure 1c.
α
0.25
0.5
0.75
1.0
Direct Better
3
5
4
1
SR-Based Better
47
45
46
49
Fig. 2: All predictors learn from scratch with new predictors
added in every 50 episodes. As the SR error (red, right axis)
goes low the SR-based predictors (green) are able to learn
faster than their direct (blue) counterparts. Shading indicates
a 95% confidence interval.
all active signals. As expected, we clearly see that the SRbased predictions (green) start with much higher error than
the direct (blue), but as the error of the SR (red) drops low the
newly added SR-based predictors are able to learn quicker,
with less peak and overall error than the direct predictor.
In the continual learning setting we never have the opportunity to tune for optimal step-sizes as we did in our
evaluation. Practically, fixed step-sizes are used for many
robotics settings in RL, but, in order to ensure stable learning,
small step-sizes are chosen. As we saw in Figure 1c, the
advantage of using the SR-based predictions is enhanced
with smaller step-sizes. Ideally, however, we would imagine
that a fully developed system would use some method of
adapting step-sizes, such as ADADELTA [10].
V. EVALUATION ON A ROBOT ARM
Tabular settings like Dayan’s grid world are useful for
enabling analysis and providing insight into the behavior
of our method. However, our goal is to accelerate learning
Fig. 3: The user controls the robot arm using a joystick to
trace the inside of the wire maze in a counter-clockwise
direction. Circuit path shown in blue.
on a real robot where states are not fully observed and
cannot be represented exactly; instead we must use function
approximation. Here we demonstrate our approach using a
robot arm and learning sensorimotor predictions with respect
to a human-generated policy. In our task a user controls a
robot arm via joystick to trace a counter-clockwise circuit
through the inside of the wire maze (see Figure 3) with a
rod held in the robot’s gripper. The user performed this task
for approximately 12 minutes completing around 50 circuits.
In this experiment we used six different prediction targets:
the current, position and speed of both the shoulder rotation
and elbow flexion joints. A new predictor was activated every
2,000 timesteps (Note that the robot reports sensor updates
at 30 timesteps/s). For this demonstration a discount factor
of γ = 0.95 was used. Four signals were used as input
to our function approximator: the current position and a
decaying trace of the position for the shoulder and elbow
joints. The decaying trace for joint j was calculated as
trj:t+1 = 0.8 ∗ trj:t + 0.2 ∗ posj:t+1 . These inputs were
normalized over the joint ranges observed in the experiment
and passed into a 4-dimensional tilecoding [5] with 100
tilings of width 1.0 and a total memory size of 2048.
Additionally a bias unit was added, resulting in a binary
feature vector of length 2049 with a maximum of 101 active
features on each timestep (hashing collisions can reduce this
number). We use a decaying step-size for all the predictors
where the step-size starts at 0.1 and decays linearly to zero
over the entire dataset. At each timestep this step-size is
further divided by the number of active features in φ(St ).
Finally, for each predictor, this step-size is offset such that
the step-size starts at 0.1 when it is first activated and decays
at the same rate as all the other predictors.
To compare the prediction error we compute a running
MSE for each signal according to Eq. 8, where at each
timestep t the sum is taken over all previous timesteps.
Unlike the previous tabular domain, we do not have the
ideal estimator to compare against and instead compare the
predictions, V , against the actual return, G. In order to
treat each signal equally we further normalize these errors
according to Eq. 9. Note that the NMSE allows us to compare
the predictions of a single signal between the two methods,
but does not tell us how accurate the predictions are nor does
Fig. 4: A 12 minute run tracing the maze circuit. A new
predictor is added every 2,000 timesteps. NMSE errors are
summed across all predictors.
it allow comparison between signals.
T
1 X
(Vi:t − Gi:t )2
T t=t
(8)
M SEi
max(M SEi:direct , M SEi:SR )
(9)
M SEi:t =
0
N M SEi =
Figures 4 and 5 show a single run, approximately 12
minutes in length. A single ordering of the predictors was
used. Figure 4 shows the error across all predictors while
Figure 5 separates out each predictor. Here we see a clear
advantage to using the SR-based predictions for most of the
signals. Unlike the previous tabular results, there is little
difference on the performance of the first predictor (shoulder
current), even while the SR is being learned. To investigate,
we ran experiments where each signal was learned from the
beginning of the run. We observed that performance was
rarely worse and sometimes even better when using the SRbased method. This suggests the SR-based approach is more
robust than expected, but further experimentation is needed.
VI. FURTHER ADVANTAGES WHEN SCALING
While this paper analyzed single policies and discount
functions, this is not the setting in which the GVF framework
is proposed to be used. Rather, it is imagined that massive
numbers of GVFS over many policies and timescales will
be used represent complex models of the world [3], [11]. In
this setting we note that using SR-based predictions can offer
additional benefits, allowing the robot to do more with less.
Consider, for a single policy π, a collection of SRs learned
for f discount functions and h one-step predictors. We can
then represent f h predictions using f + h predictors. A first
advantage is that far fewer GVFs need to be updated on each
timestep, saving computational costs. As a second benefit,
there is potential to reduce the number of weights used by the
system. For example, consider learning in a tabular setting,
with |S| states, using linear estimators. For f h predictions
the number of weights needed is : |w|direct = f h|S|,
|w|SR−based = f |S|2 + h|S|. It can be shown that for a
fixed f and S the total number of weights used by the direct
|S|
prediction approach is greater when h > ff −1
.
new predictor. We demonstrated this behaviour in both a
tabular grid world and on a robot arm. These results suggest
an effective method for improving the learning rate and
sample efficiency for robots learning in the real world.
There are several clear opportunities for further research
on this topic. The first is to provide greater understanding
into why, for a given fixed step-size, some (few) signals are
better predicted directly rather than through the SR. Further,
the work in using the SR with function approximation is
preliminary and more insight can yet be gained in this setting.
Another opportunity for research is to explore using SRbased predictions with state-dependent discount functions.
Finally, we suggest that SR-based predictions with deep
feature learning [12], [13] and an incrementally constructed
architecture would be a very powerful tool to support continual or developmental learning in robotic domains with
widespread real-world applications.
R EFERENCES
Fig. 5: The same results as Figure 4, but with the NMSE for
the individual predictors. Each is normalized from 0 to 1.
VII. RELATED WORK
The idea of the SR was originally introduced as a function
approximation method [8]. However, it has recently been
applied to other settings. It has been used, for instance,
in transfer learning problems allowing agents to generalize
better across similar but different tasks [9], [12], and to define
intrinsic rewards in option discovery algorithms [13].
GVFs were originally proposed as a method for building
an agent’s overall knowledge in a modular way [3]. To
date they have primarily been used with hand-coded fixed
policies [11]. The UNREAL agent [14] is a powerful demonstration of the usefulness of multiple predictions; pre-defined
auxiliary tasks, which can be viewed as GVFs, are shown to
accelerate and improve the robustness of learning.
Finally, the idea closest to this work is the concept of
Universal Value Functions (UVFAs) [15]. UVFAs are, as
GVFs, a generalization of value functions. However, instead
of generalizing them to multiple predictors (and discount
factors) they generalize value functions over goals in a
parametrized way. We believe our result and the idea of
UVFAs are complementary, and could in fact be eventually
combined in a future work.
VIII. CONCLUSIONS
In this paper we showed how the successor representation
(SR), although originally introduced for another purpose,
can be used to accelerate learning in a continual learning
setting in which a robot incrementally constructs models of
its world as a collection of predictions known as general
value functions (GVFs). The SR enables a given prediction
to be modularized into two components, one representing
the dynamics of the environment (the SR) and the other
representing the target signal (one-step signal prediction).
This allows a robot to reuse its existing knowledge when
adding a new prediction target, speeding up learning of the
[1] M. B. Ring, “Continual Learning in Reinforcement Environments,”
Ph.D. dissertation, The University of Texas at Austin, 1994.
[2] P. Oudeyer, F. Kaplan, and V. V. Hafner, “Intrinsic Motivation Systems for Autonomous Mental Development,” IEEE Transactions on
Evolutionary Computation, vol. 11, no. 2, pp. 265–286, 2007.
[3] R. S. Sutton, J. Modayil, M. Delp, T. Degris, P. M. Pilarski, A. White,
and D. Precup, “Horde: a Scalable Real-time Architecture for Learning
Knowledge from Unsupervised Sensorimotor Interaction,” in Proceedings of the International Joint Conference on Autonomous Agents and
Multiagent Systems (AAMAS), 2011, pp. 761–768.
[4] R. S. Sutton, “Learning to Predict by the Methods of Temporal
Differences,” Machine Learning, vol. 3, pp. 9–44, 1988.
[5] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction. MIT Press, 1998.
[6] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. A. Riedmiller, A. Fidjeland, G. Ostrovski,
S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran,
D. Wierstra, S. Legg, and D. Hassabis, “Human-level Control through
Deep Reinforcement Learning,” Nature, vol. 518, no. 7540, pp. 529–
533, 2015.
[7] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang,
A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap,
F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis,
“Mastering the Game of Go Without Human Knowledge,” Nature, vol.
550, pp. 354–359, 2017.
[8] P. Dayan, “Improving Generalization for Temporal Difference Learning: The Successor Representation,” Neural Computation, vol. 5, no. 4,
pp. 613–624, 1993.
[9] A. Barreto, W. Dabney, R. Munos, J. J. Hunt, T. Schaul, D. Silver, and
H. P. van Hasselt, “Successor Features for Transfer in Reinforcement
Learning,” in Advances in Neural Information Processing Systems
(NIPS), 2017, pp. 4058–4068.
[10] M. D. Zeiler, “ADADELTA: An Adaptive Learning Rate Method,”
CoRR, vol. abs/1212.5701, 2012.
[11] J. Modayil, A. White, and R. S. Sutton, “Multi-Timescale Nexting in
a Reinforcement Learning Robot,” Adaptive Behavior, vol. 22, no. 2,
pp. 146–160, 2014.
[12] T. D. Kulkarni, A. Saeedi, S. Gautam, and S. J. Gershman, “Deep Successor Reinforcement Learning,” CoRR, vol. abs/1606.02396, 2016.
[13] M. C. Machado, C. Rosenbaum, X. Guo, M. Liu, G. Tesauro, and
M. Campbell, “Eigenoption Discovery through the Deep Successor
Representation,” in Proceedings of the International Conference on
Learning Representations (ICLR), 2018.
[14] M. Jaderberg, V. Mnih, W. M. Czarnecki, T. Schaul, J. Z. Leibo, D. Silver, and K. Kavukcuoglu, “Reinforcement Learning with Unsupervised
Auxiliary Tasks,” in Proceedings of the International Conference on
Learning Representations (ICLR), 2017.
[15] T. Schaul, D. Horgan, K. Gregor, and D. Silver, “Universal Value Function Approximators,” in Proceedings of the International Conference
on Machine Learning (ICML), 2015, pp. 1312–1320.
| 2cs.AI
|
Foundations of Declarative Data Analysis Using Limit Datalog Programs
arXiv:1705.06927v2 [cs.AI] 12 Nov 2017
Mark Kaminski, Bernardo Cuenca Grau, Egor V. Kostylev, Boris Motik and Ian Horrocks
Department of Computer Science, University of Oxford, UK
{mark.kaminski, bernardo.cuenca.grau, egor.kostylev, boris.motik, ian.horrocks}@cs.ox.ac.uk
Abstract
Motivated by applications in declarative data analysis, we study Datalog Z —an extension of positive Datalog with arithmetic functions over integers. This language is known to be undecidable, so
we propose two fragments. In limit Datalog Z predicates are axiomatised to keep minimal/maximal
numeric values, allowing us to show that fact entailment is CO NE XP T IME-complete in combined,
and CO NP-complete in data complexity. Moreover,
an additional stability requirement causes the complexity to drop to E XP T IME and PT IME, respectively. Finally, we show that stable Datalog Z can
express many useful data analysis tasks, and so our
results provide a sound foundation for the development of advanced information systems.
1
Introduction
Analysing complex datasets is currently a hot topic in information systems. The term ‘data analysis’ covers a broad
range of techniques that often involve tasks such as data aggregation, property verification, or query answering. Such
tasks are currently often solved imperatively (e.g., using Java
or Scala) by specifying how to manipulate the data, and this
is undesirable because the objective of the analysis is often
obscured by evaluation concerns. It has recently been argued
that data analysis should be declarative [Alvaro et al., 2010;
Markl, 2014; Seo et al., 2015; Shkapsky et al., 2016]: users
should describe what the desired output is, rather than how to
compute it. For example, instead of computing shortest paths
in a graph by a concrete algorithm, one should (i) describe
what a path length is, and (ii) select only paths of minimum
length. Such a specification is independent of evaluation details, allowing analysts to focus on the task at hand. An evaluation strategy can be chosen later, and general parallel and/or
incremental evaluation algorithms can be reused ‘for free’.
An essential ingredient of declarative data analysis is an
efficient language that can capture the relevant tasks, and
Datalog is a prime candidate since it supports recursion.
Apart from recursion, however, data analysis usually also requires integer arithmetic to capture quantitative aspects of
data (e.g., the length of a shortest path). Research on combining the two dates back to the ’90s [Mumick et al., 1990;
Kemp and Stuckey, 1991; Beeri et al., 1991; Van Gelder,
1992; Consens and Mendelzon, 1993; Ganguly et al., 1995;
Ross and Sagiv, 1997], and is currently experiencing a revival
[Faber et al., 2011; Mazuran et al., 2013]. This extensive
body of work, however, focuses primarily on integrating recursion and arithmetic with aggregate functions in a coherent
semantic framework, where technical difficulties arise due to
nonmonotonicity of aggregates. Surprisingly little is known
about the computational properties of integrating recursion
with arithmetic, apart from that a straightforward combination is undecidable [Dantsin et al., 2001]. Undecidability also
carries over to the above formalisms and practical Datalogbased systems such as BOOM [Alvaro et al., 2010], DeALS
[Shkapsky et al., 2016], Myria [Wang et al., 2015], SociaLite
[Seo et al., 2015], Overlog [Loo et al., 2009], Dyna [Eisner
and Filardo, 2011], and Yedalog [Chin et al., 2015].
To develop a sound foundation for Datalog-based declarative data analysis, we study Datalog Z —negation-free Datalog with integer arithmetic and comparisons. Our main contribution is a new limit Datalog Z fragment that, like the existing data analysis languages, is powerful and flexible enough
to naturally capture many important analysis tasks. However,
unlike Datalog Z and the existing languages, reasoning with
limit programs is decidable, and it becomes tractable in data
complexity under an additional stability restriction.
In limit Datalog Z , all intensional predicates with a numeric
argument are limit predicates. Instead of keeping all numeric
values for a given tuple of objects, such predicates keep only
the minimal (min) or only the maximal (max) bounds of numeric values entailed for the tuple. For example, if we encode
a weighted directed graph using a ternary predicate edge, then
rules (1) and (2), where sp is a min limit predicate, compute
the cost of a shortest path from a given source node v0 to
every other node.
→ sp(v0 , 0)
sp(x, m) ∧ edge(x, y, n) → sp(y, m + n)
(1)
(2)
If these rules and a dataset entail a fact sp(v, k), then the cost
of a shortest path from v0 to v is at most k; hence, sp(v, k 0 )
holds for each k 0 ≥ k since the cost of a shortest path is also at
most k 0 . Rule (2) intuitively says that, if x is reachable from
v0 with cost at most m and hx, yi is an edge of cost n, then v 0
is reachable from v0 with cost at most m + n. This is different from Datalog Z , where there is no implicit semantic con-
nection between sp(v, k) and sp(v, k 0 ), and such semantic
connections allow us to prove decidability of limit Datalog Z .
We provide a direct semantics for limit predicates based on
Herbrand interpretations, but we also show that this semantics can be axiomatised in standard Datalog Z . Our formalism
can thus be seen as a fragment of Datalog Z , from which it
inherits well-understood properties such as monotonicity and
existence of a least fixpoint model [Dantsin et al., 2001].
Our contributions are as follows. First, we introduce limit
Datalog Z programs and argue that they can naturally capture
many relevant data analysis tasks. We prove that fact entailment in limit Datalog Z is undecidable, but, after restricting the use of multiplication, it becomes CO NE XP T IME- and
CO NP-complete in combined and data complexity, respectively. To achieve tractability in data complexity (which is
very important for robust behaviour on large datasets), we additionally introduce a stability restriction and show that this
does not prevent expressing the relevant analysis tasks.
The proofs of all results are given in the appendix of this
paper.
2
Preliminaries
In this section, we recapitulate the well-known definitions of
Datalog with integers, which we call Datalog Z .
Syntax A vocabulary consists of predicates, objects, object
variables, and numeric variables. Each predicate has an integer arity n, and each position 1 ≤ i ≤ n is of either object
or numeric sort. An object term is an object or an object variable. A numeric term is an integer, a numeric variable, or
of the form s1 + s2 , s1 − s2 , or s1 × s2 where s1 and s2 are
numeric terms, and +, −, and × are the standard arithmetic
functions. A constant is an object or an integer. The magnitude of an integer is its absolute value. A standard atom
is of the form B(t1 , . . . , tn ), where B is a predicate of arity
n and each ti is a term whose type matches the sort of position i of B. A comparison atom is of the form (s1 < s2 ) or
(s1 ≤ s2 ), where < and ≤ are the standard comparison predicates,Vand s1 Vand s2 are numeric terms. A rule r is of the
form i αi ∧ j βj → α, where α(i) are standard atoms, βj
are comparison atoms, and each variable in r occurs
V in some
αi . Atom h(r) = α is the head
V of r; sb(r) = i αi is the
standard body of r; cb(r) = j βj is the comparison body
of r; and b(r) = sb(r) ∧ cb(r) is the body of r. A ground
instance of r is obtained from r by substituting all variables
by constants. A (Datalog Z ) program P is a finite set of rules.
Predicate B is intensional (IDB) in P if B occurs in P in the
head of a rule whose body is not empty; otherwise, B is extensional (EDB) in P. A term, atom, rule, or program is ground
if it contains no variables. A fact is a ground, function-free,
standard atom. Program P is a dataset if h(r) is a fact and
b(r) = ∅ for each r ∈ P. We often say that P contains a
fact α and write α ∈ P, which actually means → α ∈ P. We
write a tuple of terms as t, and we often treat conjunctions
and tuples as sets and write, say, αi ∈ sb(r), |t|, and ti ∈ t.
Semantics A (Herbrand) interpretation I is a (not necessarily finite) set of facts. Such I satisfies a ground atom α,
written I |= α, if (i) α is a standard atom and evaluating the
arithmetic functions in α produces a fact in I, or (ii) α is a
comparison atom and evaluating the arithmetic functions and
comparisons produces true. The notion of satisfaction is extended to conjunctions of ground atoms, rules, and programs
as in first-order logic, where each rule is universally quantified. If I |= P, then I is a model of program P; and P entails
a fact α, written P |= α, if I |= α holds whenever I |= P.
Complexity In this paper we study the computational properties of checking P |= α. Combined complexity assumes that
both P and α are part of the input. In contrast, data complexity assumes that P is given as P 0 ∪ D for P 0 a program and
D a dataset, and that only D and α are part of the input while
P 0 is fixed. Unless otherwise stated, all numbers in the input
are coded in binary, and the size kPk of P is the size of its
representation. Checking P |= α is undecidable even if the
only arithmetic function in P is + [Dantsin et al., 2001].
Presburger arithmetic is first-order logic with constants 0
and 1, functions + and −, equality, and the comparison predicates < and ≤, interpreted over all integers Z. The complexity of checking sentence validity (i.e., whether the sentence
is true in all models of Presburger arithmetic) is known when
the number of quantifier alternations and/or the number of
variables in each quantifier block are fixed [Berman, 1980;
Grädel, 1988; Schöning, 1997; Haase, 2014].
3
Limit Programs
Towards introducing a decidable fragment of Datalog Z for
data analysis, we first note that the undecidability proof of
(plain) Datalog Z outlined by Dantsin et al. [2001] uses atoms
with at least two numeric terms. Thus, to motivate introducing our fragment, we first prove that undecidability holds
even if atoms contain at most one numeric term. The proof
uses a reduction from the halting problem for deterministic
Turing machines. To ensure that each standard atom in P has
at most one numeric term, combinations of a time point and
a tape position are encoded using a single integer.
Theorem 1. For P a Datalog Z program and α a fact, checking P |= α is undecidable even if P contains no × or − and
each standard atom in P has at most one numeric term.
We next introduce limit Datalog Z , where limit predicates
keep bounds on numeric values. This language can be seen
as either a semantic or a syntactic restriction of Datalog Z .
Definition 2. In limit Datalog Z , a predicate is either an object predicate with no numeric positions, or a numeric predicate where only the last position is numeric. A numeric predicate is either an ordinary numeric predicate or a limit predicate, and the latter is either a min or a max predicate. Atoms
with object predicates are object atoms, and analogously for
other types of atoms/predicates. A Datalog Z rule r is a limit
(Datalog Z ) rule if (i) b(r) = ∅, or (ii) each atom in sb(r) is
an object, ordinary numeric, or limit atom, and h(r) is an
object or a limit atom. A limit (Datalog Z ) program P is a
program containing only limit rules; and P is homogeneous
if it does not contain both min and max predicates.
In the rest of this paper we make three simplifying assumptions. First, numeric atoms occurring in a rule body are
function-free (but comparison atoms and the head can contain arithmetic functions). Second, each numeric variable in
a rule occurs in at most one standard body atom. Third, distinct rules in a program use different variables. The third
assumption is clearly w.l.o.g. because all variables are universally quantified so their names are immaterial. Moreover,
the first two assumptions are w.l.o.g. as well since, for each
rule, there exists a logically equivalent rule that satisfies these
assumptions. In particular, we can replace an atom such as
A(t, m1 + m2 ) with conjunction
A(t, m) ∧ I(m1 ) ∧ I(m2 ) ∧
(m ≤ m1 + m2 ) ∧ (m1 + m2 ≤ m)
where m is a fresh variable and I is a fresh predicate axiomatised to hold on all integers as follows:
→ I(0)
I(m) → I(m + 1)
I(m) → I(m − 1)
Also, we can replace atoms A1 (t1 , m) ∧ A2 (t2 , m) with conjunction A1 (t1 , m) ∧ A2 (t2 , m0 ) ∧ (m ≤ m0 ) ∧ (m0 ≤ m),
where m0 is a fresh variable.
Intuitively, a limit fact B(a, k) says that the value of B for
a tuple of objects a is at least k (if B is max) or at most k
(if B is min). For example, a fact sp(v, k) in our shortest
path example from Section 1 says that node v is reachable
from v0 via a path with cost at most k. To capture this intended meaning, we require interpretations I to be closed for
limit predicates—that is, whenever I contains a limit fact α, it
also contains all facts implied by α according to the predicate
type. In our example, this captures the observation that the
existence of a path from v0 to v of cost at most k implies the
existence of such a path of cost at most k 0 for each k 0 ≥ k.
Definition 3. An interpretation I is limit-closed if, for each
limit fact B(a, k) ∈ I where B is a min (resp. max) predicate, B(a, k 0 ) ∈ I holds for each integer k 0 with k ≤ k 0 (resp.
k 0 ≤ k). An interpretation I is a model of a limit program P
if I |= P and I is limit-closed. The notion of entailment is
modified to take into account only limit-closed models.
The semantics of limit predicates in a limit Datalog Z program P can be axiomatised explicitly by extending P with
the following rules, where Z is a fresh predicate. Thus, limit
Datalog Z can be seen as a syntactic fragment of Datalog Z .
→ Z(0)
Z(m) → Z(m + 1)
Z(m) → Z(m − 1)
B(x, m) ∧ Z(n) ∧ (m ≤ n) → B(x, n)
for each min predicate B in P
B(x, m) ∧ Z(n) ∧ (n ≤ m) → B(x, n)
for each max predicate B in P
Each limit program can be reduced to a homogeneous program; however, for the sake of generality, in our technical
results we do not require programs to be homogeneous.
Proposition 4. For each limit program P and fact α, a homogeneous program P 0 and fact α0 can be computed in linear
time such that P |= α if and only if P 0 |= α0 .
Intuitively, program P 0 in Proposition 4 is obtained by replacing all min (or all max) predicates in P by fresh max
(resp. min) predicates and negating their numeric arguments.
In Section 1 we have shown that limit Datalog Z can compute the cost of shortest paths in a graph. We next present
further examples of data analysis tasks that our formalism
can handle. In all examples, we assume that all objects in
the input are arranged in an arbitrary linear order using facts
first(a1 ), next(a1 , a2 ), . . . , next(an−1 , an ); we use this order to simulate aggregation by means of recursion.
Example 5. Consider a social network where agents are connected by the ‘follows’ relation. Agent as introduces (tweets)
a message, and each agent ai retweets the message if at least
kai agents that ai follows tweet the message, where kai is
a positive threshold uniquely associated with ai . Our goal
is to determine which agents tweet the message eventually.
To achieve this using limit Datalog Z , we encode the network
structure in a dataset Dtw containing facts follows(ai , aj ) if
ai follows aj , and ordinary numeric facts th(ai , kai ) if ai ’s
threshold is kai . Program Ptw , containing rules (3)–(8), encodes message propagation, where nt is a max predicate.
→ tw (as )
(3)
follows(x, y 0 ) ∧ first(y) → nt(x, y, 0)
follows(x, y) ∧ first(y) ∧ tw (y) → nt(x, y, 1)
nt(x, y 0 , m) ∧ next(y 0 , y) → nt(x, y, m)
0
(4)
(5)
(6)
0
nt(x, y , m) ∧ next(y , y) ∧
(7)
follows(x, y) ∧ tw (y) → nt(x, y, m + 1)
th(x, m) ∧ nt(x, y, n) ∧ (m ≤ n) → tw (x)
(8)
Specifically, Ptw ∪ Dtw |= tw (ai ) iff ai tweets the message.
Intuitively, nt(ai , aj , m) is true if, out of agents {a1 , . . . , aj }
(according to the order), at least m agents that ai follows
tweet the message. Rules (4) and (5) initialise nt for the first
agent in the order; nt is a max predicate, so if the first agent
tweets the message, rule (5) ‘overrides’ rule (4). Rules (6)
and (7) recurse over the order to compute nt as stated above.
Example 6. Limit Datalog Z can also solve the problem of
counting paths between pairs of nodes in a directed acyclic
graph. We encode the graph in the obvious way as a dataset
Dcp that uses object predicates node and edge. Program
Pcp , consisting of rules (9)–(14) where np and np 0 are max
predicates, then counts the paths.
node(x) → np(x, x, 1)
(9)
node(x) ∧ node(y) ∧ first(z) → np 0 (x, y, z, 0)
(10)
edge(x, z) ∧
np(z, y, m) ∧ first(z) → np 0 (x, y, z, m)
(11)
np 0 (x, y, z 0 , m) ∧ next(z 0 , z) → np 0 (x, y, z, m)
(12)
0
0
0
np (x, y, z , m) ∧ next(z , z) ∧
(13)
edge(x, z) ∧ np(z, y, n) → np 0 (x, y, z, m + n)
np 0 (x, y, z, m) → np(x, y, m)
(14)
Specifically, Pcp ∪ Dcp |= np(ai , aj , k) iff at least k paths
exist from node ai to node aj . Intuitively, np 0 (ai , aj , ak , m) is
true if m is at least the sum of the number of paths from each
a0 ∈ {a1 , . . . , ak } (according to the order) to aj for which
there exists an edge from ai to a0 . Rule (9) says that each
node has one path to itself. Rule (10) initialises aggregation
by saying that, for the first node z, there are zero paths from x
to y, and rule (11) overrides this if there exists an edge from
x to z. Finally, rule (12) propagates the sum for x to the next
z in the order, and rule (13) overrides this if there is an edge
from x to z by adding the number of paths from z 0 and z to y.
Example 7. Assume that, in the graph from Example 6, each
node ai is associated with a bandwidth bai limiting the number of paths going through ai to at most bai . To count the
paths compliant with the bandwidth requirements, we extend
Dcp to dataset Dbcp that additionally contains an ordinary
numeric fact bw (ai , bai ) for each node ai , and we define Pbcp
by replacing rule (14) in Pcp with the following rule.
np 0 (x, y, z, m) ∧ bw (z, n) ∧ (m ≤ n) → np(x, y, m)
Then, Pbcp ∪ Dbcp |= np(ai , aj , k) iff there exist at least k
paths from node ai to node aj , where the bandwidth requirement is satisfied for all nodes on each such path.
4
Fixpoint Characterisation of Entailment
Programs are often grounded to eliminate variables and thus
simplify the presentation. In limit Datalog Z , however, numeric variables range over integers, so a grounding can be
infinite. Thus, we first specialise the notion of a grounding.
Definition 8. A rule r is semi-ground if each variable in r is a
numeric variable that occurs in r in a limit body atom. A limit
program P is semi-ground if all of its rules are semi-ground.
The semi-grounding of P contains, for each r ∈ P, each rule
obtained from r by replacing each variable not occurring in
r in a numeric argument of a limit atom with a constant of P.
Obviously, P |= α if and only if P 0 |= α for P 0 the semigrounding of P. We next characterise entailment of limit
programs by pseudo-interpretations, which compactly represent limit-closed interpretations. If a limit-closed interpretation I contains B(b, k) where B is a min predicate, then
either the limit value ` ≤ k exists such that B(b, `) ∈ I and
B(b, k 0 ) 6∈ I for k 0 < `, or B(b, k 0 ) ∈ I holds for all k 0 ≤ k,
and dually for B a max predicate. Thus, to characterise the
value of B on a tuple of objects b in I, we just need the limit
value, or information that no such value exists.
Definition 9. A pseudo-interpretation J is a set of facts over
integers extended with a special symbol ∞ such that k = k 0
holds for all limit facts B(b, k) and B(b, k 0 ) in I.
Limit-closed interpretations correspond naturally and oneto-one to pseudo-interpretations, so we can recast the notions
of satisfaction and model using pseudo-interpretations. Unlike for interpretations, the number of facts in a pseudo-model
of a semi-ground limit program P can be bounded by |P|.
Definition 10. A limit-closed interpretation I corresponds to
a pseudo-interpretation J if I contains exactly all object and
ordinary numeric facts of J, and, for each limit predicate B,
each tuple of objects b, and each integer `, (i) B(b, k) ∈ I
for all k if and only if B(b, ∞) ∈ J, and (ii) B(b, `) ∈ I and
B(b, k) 6∈ I for all k < ` (resp. ` < k) and B is a min (resp.
max) predicate if and only if B(b, `) ∈ J.
Let J and J 0 be pseudo-interpretations corresponding to
interpretations I and I 0 . Then, J satisfies a ground atom α,
written J |= α, if I |= α; J is a pseudo-model of a program
P, written J |= P, if I |= P; finally, J v J 0 holds if I ⊆ I 0 .
Example 11. Let I be the interpretation consisting of A(1),
A(2), B(a, k) for k ≤ 5, and B(b, k) for k ∈ Z, where A is
an ordinary numeric predicate, B is a max predicate, and a
and b are objects. Then, {A(1), A(2), B(a, 5), B(b, ∞)} is
the pseudo-interpretation corresponding to I.
We next introduce the immediate consequence operator
TP of a limit program P on pseudo-interpretations. We assume for simplicity that P is semi-ground. To apply a rule
r ∈ P to a pseudo-interpretation J while correctly handling
limit atoms, operator TP converts r into a linear integer constraint C(r, J) that captures all ground instances of r applicable to the limit-closed interpretation I corresponding to J. If
C(r, J) has no solution, r is not applicable to J. Otherwise,
h(r) is added to J if it is not a limit atom; and if h(r) is a min
(max) atom B(b, m), then the minimal (maximal) solution
` for m in C(r, J) is computed, and J is updated such that
the limit value of B on b is at least (at most) `—that is, the
application of r to J keeps only the ‘best’ limit value.
Definition 12. For P a semi-ground limit program, r ∈ P,
and J a pseudo-interpretation, C(r, J) is the conjunction of
comparison atoms containing (i) cb(r); (ii) (0 < 0) if an object or ordinary numeric atom α ∈ sb(r) exists with α 6∈ J,
or a limit atom B(b, s) ∈ sb(r) exists with B(b, `) 6∈ J for
each `; and (iii) (` ≤ s) (resp. (s ≤ `)) for each min (resp.
max) atom B(b, s) ∈ sb(r) with B(b, `) ∈ J and ` 6= ∞.
Rule r is applicable to J if C(r, J) has an integer solution.
Assume r is applicable to J. If h(r) is an object or ordinary numeric atom, let hd(r, J) = h(r). If h(r) = B(b, s) is
a min (resp. max) atom, the optimum value opt(r, J) is the
smallest (resp. largest) value of s in all solutions to C(r, J),
or ∞ if no such bound on the value of s in the solutions to
C(r, J) exists; moreover, hd(r, J) = B(b, opt(r, J)).
Operator TP (J) maps J to the smallest (w.r.t. v) pseudointerpretation satisfying hd(r, J) for each r ∈ P applicable
to J. Finally, T0P = ∅, and TnP = TP (Tn−1
P ) for n > 0.
Example 13. Let r be A(x) ∧ (2 ≤ x) → B(x + 1) with A
and B max predicates. Then, C(r, ∅) = (2 ≤ x) ∧ (0 < 0)
does not have a solution, and therefore rule r is not applicable to the empty pseudo-interpretation. Moreover, for
J = {A(3)}, conjunction C(r, J) = (2 ≤ x) ∧ (x ≤ 3) has
two solutions—{x 7→ 2} and {x 7→ 3}—and therefore rule
r is applicable to J. Finally, B is a max predicate, and so
opt(r, J) = max{2 + 1, 3 + 1} = 4, and hd(r, J) = B(4).
Consequently, T{r} (J) = {B(4)}.
Lemma 14. For each semi-ground limit program P, operator
TP is monotonic w.r.t. v. Moreover, J |= P if and only if
TP (J) v J for each pseudo-interpretation J.
Monotonicity ensures existence of the closure T∞
P of P—
the least pseudo-interpretation such that TnP v T∞
P for each
n ≥ 0. The following theorem characterises entailment and
provides a bound on the number of facts in the closure.
Theorem 15. For P a semi-ground limit program and α a
∞
fact, P |= α if and only if T∞
P |= α; also, |TP | ≤ |P|; and
J |= P implies T∞
v
J
for
each
pseudo-interpretation
J.
P
The proofs for the first and the third claim of Theorem 15
use the monotonicity of TP analogously to plain Datalog.
The second claim holds since, for each n ≥ 0, each pair of
distinct facts in TnP must be derived by distinct rules in P.
5
Decidability of Entailment: Limit-Linearity
We now start our investigation of the computational properties of limit Datalog Z . Theorem 15 bounds the cardinality of
the closure of a semi-ground program, but it does not bound
the magnitude of the integers occurring in limit facts; in fact,
integers can be arbitrarily large. Moreover, due to multiplication, checking rule applicability requires solving nonlinear
inequalities over integers, which is undecidable.
Theorem 16. For P a semi-ground limit program and α a
fact, checking P |= α and checking applicability of a rule of
P to a pseudo-interpretation are both undecidable.
The proof of Theorem 16 uses a straightforward reduction
from Hilbert’s tenth problem.
Checking rule applicability is undecidable due to products
of variables in inequalities; however, for linear inequalities
that prohibit multiplying variables, the problem can be solved
in NP, and in polynomial time if we bound the number of
variables. Thus, to ensure decidability, we next restrict limit
programs so that their semi-groundings contain only linear
numeric terms. All our examples satisfy this restriction.
Definition 17. A limit rule r P
is limit-linear if each numeric
n
term in r is of the form s0 + i=1 si × mi , where (i) each
mi is a distinct numeric variable occurring in a limit body
atom of r, (ii) term s0 contains no variable occurring in a
limit body atom of r, and (iii) each si with i ≥ 1 is a term
constructed using multiplication ×, integers, and variables
not occurring in limit body atoms of r. A limit-linear program
contains only limit-linear rules.1
In the rest of this section, we show that entailment for limitlinear programs is decidable and provide tight complexity
bounds. Our upper bounds are obtained via a reduction to
the validity of Presburger formulas of a certain shape.
Lemma 18. For P a semi-ground limit-linear program
Wnand α
a fact, there exists a Presburger sentence ϕ = ∀x∃y. i=1 ψi
that is valid if and only if P |= α. Each ψi is a conjunction of possibly negated atoms. Moreover, |x| + |y| and
each kψi k are bounded polynomially by kPk + kαk. Number n is bounded polynomially by |P| and exponentially by
maxr∈P krk. Finally, the magnitude of each integer in ϕ is
bounded by the maximal magnitude of an integer in P and α.
The reduction in Lemma 18 is based on three main ideas.
First, for each limit atom B(b, s) in a semi-ground program
P, we use a Boolean variable def Bb to indicate that an atom
of the form B(b, `) exists in a pseudo-model of P, a Boolean
variable fin Bb to indicate whether the value of ` is finite, and
an integer variable val Bb to capture ` if it is finite. Second,
each rule of P is encoded as a universally quantified Presburger formula by replacing each standard atom with its encoding. Finally, entailment of α from P is encoded as a sentence stating that, in every pseudo-interpretation, either some
rule in P is not satisfied, or α holds; this requires universal
quantifiers to quantify over all models, and existential quantifiers to negate the (universally quantified) program.
Lemma 19 bounds the magnitude of integers in models of
Presburger formulas from Lemma 18. These bounds follow
from recent deep results on semi-linear sets and their connection to Presburger arithmetic [Chistikov and Haase, 2016].2
1
Note that each multiplication-free limit program can be normalised in polynomial time to a limit-linear program.
2
We thank Christoph Haase for providing a proof of this lemma.
Wn
Lemma 19. Let ϕ = ∀x∃y. i=1 ψi be a Presburger sentence where each ψi is a conjunction of possibly negated
atoms of size at most k mentioning at most ` variables, a
is the maximal magnitude of an integer in ϕ, and m = |x|.
Then, ϕ is valid if and only if ϕ is valid over models where
each integer variable assumes a value whose magnitude is
`
4
bounded by (2O(` log `) · ak` )n2 ·O(m ) .
Lemmas 18 and 19 provide us with bounds on the size of
counter-pseudo-models for entailment.
Theorem 20. For P a semi-ground limit-linear program, D
a dataset, and α a fact, P ∪ D 6|= α if and only if a pseudomodel J of P ∪ D exists where J 6|= α, |J| ≤ |P ∪ D|, and
the magnitude of each integer in J is bounded polynomially
by the largest magnitude of an integer in P ∪ D, exponentially by |P|, and double-exponentially by maxr∈P krk.
By Theorem 20, the following nondeterministic algorithm
decides P 6|= α:
1. compute the semi-grounding P 0 of P;
2. guess a pseudo-interpretation J that satisfies the bounds
given in Theorem 20;
3. if TP 0 (J) v J (so J |= P 0 ) and J 6|= α, return true.
Step 1 requires exponential (polynomial in data) time, and it
does not increase the maximal size of a rule. Hence, Step 2 is
nondeterministic exponential (polynomial in data), and Step 3
requires exponential (polynomial in data) time to solve a system of linear inequalities. Theorem 21 proves that these
bounds are both correct and tight.
Theorem 21. For P a limit-linear program and α a fact, deciding P |= α is CO NE XP T IME-complete in combined and
CO NP-complete in data complexity.
The upper bounds in Theorem 21 follow from Theorem 20,
in data complexity is shown by a reduction
from the square tiling problem, and CO NE XP T IME-hardness
in combined complexity is shown by a similar reduction from
the succinct version of square tiling.
CO NP-hardness
6
Tractability of Entailment: Stability
Tractability in data complexity is important on large datasets,
so we next present an additional stability condition that brings
the complexity of entailment down to E XP T IME in combined
and PT IME in data complexity, as in plain Datalog.
6.1
Cyclic Dependencies in Limit Programs
The fixpoint of a plain Datalog program can be computed in
PT IME in data complexity. However, for P a limit-linear program, a naı̈ve computation of T∞
P may not terminate since repeated application of TP can produce larger and larger numbers. Thus, we need a way to identify when the numeric argument ` of a limit fact A(a, `) diverges—that is, grows or
decreases without a bound; moreover, to obtain a procedure
tractable in data complexity, divergence should be detected
after polynomially many steps. Example 22 illustrates that
this can be achieved by analysing cyclic dependencies.
Example 22. Let Pc contain facts A(0) and B(0), and rules
A(m) → B(m) and B(m) → A(m + 1), where A and B
are max predicates. Applying the first rule copies the value
of A into B, and applying the second rule increases the value
of A; thus, both A and B diverge in T∞
Pc . The existence of
a cyclic dependency between A and B, however, does not
necessarily lead to divergence. Let program Pc0 be obtained
from Pc by adding a max fact C(5) and replacing the first
rule with A(m) ∧ C(n) ∧ (m ≤ n) → B(m). While a cyclic
dependency between A and B still exists, the increase in the
values of A and B is bounded by the value of C, which is independent of A or B; thus, neither A nor B diverge in T∞
Pc0 .
In the rest of this section, we extend <, +, and − to ∞
by defining k < ∞, ∞ + k = ∞, and ∞ − k = ∞ for each
integer k. We formalise cyclic dependencies as follows.
Definition 23. For each n-ary limit predicate B and each
tuple b of n − 1 objects, let vBb be a node unique for B
and b. The value propagation graph of a semi-ground limitlinear program P and a pseudo-interpretation J is the directed weighted graph GJP = (V, E, µ) defined as follows.
1. For each limit fact B(b, `) ∈ J, we have vBb ∈ V.
2. For each rule r ∈ P applicable to J with the head of
the form A(a, s) where vAa ∈ V and each body atom
B(b, m) of r where vBb ∈ V and variable m occurs
in term s, we have hvBb , vAa i ∈ E; such r is said to
produce the edge hvBb , vAa i in E.
3. For each r ∈ P and each edge e = hvBb , vAa i ∈ E produced by r, δre (J) = ∞ if opt(r, J) = ∞; otherwise,
opt(r, J) − `
if B and A are max, ` 6= ∞
−opt(r,
J)
−
`
if B is max, A is min, ` 6= ∞
δre (J) = −opt(r, J) + ` if B and A are min, ` 6= ∞
opt(r, J) + `
if B is min, A is max, ` 6= ∞
opt(r, J)
if ` = ∞
where ` is such that B(b, `) ∈ J. The weight of each
edge e ∈ E is then given by
µ(e) = max{δre (J) | r ∈ P produces e}.
A positive-weight cycle in GJP is a cycle for which the sum of
the weights of the contributing edges is greater than 0.
Intuitively, GJP describes how, for each limit predicate B
and objects b such that B(b, `) ∈ J, operator TP propagates ` to other facts. The presence of a node vBb in V indicates that B(b, `) ∈ J holds for some ` ∈ Z ∪ {∞}; this
` can be uniquely identified given vBb and J. An edge
e = hvBb , vAa i ∈ E indicates that at least one rule r ∈ P is
applicable to J where h(r) = A(a, s), B(b, m) ∈ sb(r), and
m occurs in s; moreover, applying r to J produces a fact
A(a, `0 ) where `0 satisfies ` + µ(e) ≤ `0 if both A and B are
max predicates, and analogously for the other types of A and
B. In other words, edge e indicates that the application of TP
to J will propagate the value of vBb to vAa while increasing it
by at least µ(e). Thus, presence of a positive-weight cycle in
GJP indicates that repeated rule applications might increment
the values of all nodes on the cycle.
6.2
Stable Programs
As Example 22 shows, the presence of a positive-weight cycle
in GJP does not imply the divergence of all atoms corresponding to the nodes in the cycle. This is because the weight of
such a cycle may decrease after certain rule applications and
so it is no longer positive.
This motivates the stability condition, where edge weights
in GJP may only grow but never decrease with rule application. Hence, once the weight of a cycle becomes positive,
it will remain positive and thus guarantee the divergence of
all atoms corresponding to its nodes. Intuitively, P is stable
if, whenever a rule r ∈ P is applicable to some J, rule r is
also applicable to each J 0 with larger limit values, and applying r to such J 0 further increases the value of the head.
Definition 24 defines stability as a condition on GJP . Please
note that, for all pseudo-interpretations J and J 0 with J v J 0
0
and GJP = (V, E, µ) and GJP = (V 0 , E 0 , µ0 ) the corresponding value propagation graphs, we have E ⊆ E 0 .
Definition 24. A semi-ground limit-linear program P is stable if, for all pseudo-interpretations J and J 0 with J v J 0 ,
0
GJP = (V, E, µ), GJP = (V 0 , E 0 , µ0 ), and each e ∈ E,
1. µ(e) ≤ µ0 (e), and
2. e = hvBb , vAa i and B(b, ∞) ∈ J imply µ(e) = ∞.
A limit-linear program is stable if its semi-grounding is stable.
Example 25. Program Pc from Example 22 is stable, while
Pc0 is not: for J = {A(0), C(0)} and J 0 = {A(1), C(0)}, we
have J v J 0 , but µ(hvA , vB i) = 0 and µ0 (hvA , vB i) = −1.
For each semi-ground program P and each integer n, we
have TnP v Tn+1
P , and stability ensures that edge weights
only grow after rule application. Thus, recursive application
of the rules producing edges involved in a positive-weight cycle leads to divergence, as shown by the following lemma.
Lemma 26. For each semi-ground stable program P, each
pseudo-interpretation J with J v T∞
P , and each node vAa
on a positive-weight cycle in GJP , we have A(a, ∞) ∈ T∞
P .
Algorithm 1 uses this observation to deterministically compute the fixpoint of P. The algorithm iteratively applies TP ;
however, after each step, it computes the corresponding value
propagation graph (line 4) and, for each A(a, `) where node
vAa occurs on a positive-weight cycle (line 5), it replaces
` with ∞ (line 6). By Lemma 26, this is sound. Moreover, since the algorithm repeatedly applies TP , it necessarily derives each fact from T∞
P eventually. Finally, Lemma 27
shows that the algorithm terminates in time polynomial in the
number of rules in a semi-ground program. Intuitively, the
proof of the lemma shows that, without introducing a new
edge or a new positive weight cycle in the value propagation graph, repeated application of TP necessarily converges
in O(|P|2 ) steps; moreover, the number of edges in GJP is
at most quadratic |P|, and so a new edge or a new positive
weight cycle can be introduced at most O(|P|2 ) many times.
Lemma 27. When applied to a semi-ground stable program
P, Algorithm 1 terminates after at most 8|P|6 iterations of
the loop in lines 2–8.
Lemmas 26 and 27 imply the following theorem.
Theorem 28. For P a semi-ground stable program, D a
dataset, and α a fact, Algorithm 1 decides P ∪ D |= α in time
polynomial in kP ∪ Dk and exponential in maxr∈P krk.
Since the running time is exponential in the maximal size
of a rule, and semi-grounding does not increase rule sizes,
Algorithm 1 Entailment for Semi-Ground Stable Programs
Input: semi-ground stable program P, fact α
Output: true if P |= α
1: J 0 ··= ∅
2: repeat
3:
J ··= J 0
4:
GJP ··= (V, E, µ)
5:
for each vAa ∈ V in a positive-weight cycle in GJP do
6:
replace A(a, `) in J with A(a, ∞)
7:
J 0 ··= TP (J)
8: until J = J 0
9: return true if J |= α and false otherwise
Algorithm 1 combined with a semi-grounding preprocessing
step provides an exponential time decision procedure for stable, limit-linear programs. This upper bound is tight since
entailment in plain Datalog is already E XP T IME-hard in combined and PT IME-hard in data complexity.
The first condition of Definition 31 ensures that each variable occurring in a numeric term contributes to the value of
the term. For example, it disallows terms such as 0 · x and
x − x, since a rule with such a term in the head may violate the second condition. Moreover, the second condition of
Definition 31 ensures that, if the value of a numeric variable
x occurring in the head ‘increases’ w.r.t. the type of the body
atom introducing x (i.e., x increases if it occurs in a max
body atom and decreases otherwise), then so does the value
of the numeric term in the head; this is essential for the first
condition of stability (cf. Definition 24). Finally, the third
condition of Definition 31 ensures that comparisons cannot
be invalidated by ‘increasing’ the values of the variables involved, which is required for both conditions of stability.
Type consistency is a purely syntactic condition that can
be checked by looking at one rule and one atom at a time.
Hence, checking type consistency is feasible in L OG S PACE.
Proposition 32. Each type-consistent limit-linear program is
stable.
Theorem 29. For P a stable program and α a fact, checking P |= α is E XP T IME-complete in combined and PT IMEcomplete in data complexity.
Proposition 33. Checking whether a limit-linear program is
type-consistent can be accomplished in L OG S PACE.
6.3
7
Type-Consistent Programs
Unfortunately, the class of stable programs is not recognisable, which can again be shown by a reduction from Hilbert’s
tenth problem.
Proposition 30. Checking stability of a limit-linear program
P is undecidable.
We next provide a sufficient condition for stability that captures programs such as those in Examples 5 and 6. Intuitively,
Definition 31 syntactically prevents certain harmful interactions. In the second rule of program Pc0 from Example 22,
numeric variable m occurs in a max atom and on the lefthand side of a comparison atom (m ≤ n); thus, if the rule is
applicable for some value of m, it is not necessarily applicable for each m0 ≥ m, which breaks stability.
Definition 31. A semi-ground limit-linear rule r is typeconsistent if
Pn
– each numeric term t in r is of the form k0 + i=1 ki × mi
where k0 is an integer and each ki , 1 ≤ i ≤ n, is a nonzero
integer, called the coefficient of variable mi in t;
– if h(r) = A(a, s) is a limit atom, then each variable occurring in s with a positive (resp. negative) coefficient also
occurs in a (unique) limit body atom or r that is of the same
(resp. different) type (i.e., min vs. max) as h(r); and
– for each comparison (s1 < s2 ) or (s1 ≤ s2 ) in r, each
variable occurring in s1 with a positive (resp. negative)
coefficient also occurs in a (unique) min (resp. max) body
atom, and each variable occurring in s2 with a positive
(resp. negative) coefficient also occurs in a (unique) max
(resp. min) body atom of r.
A semi-ground limit-linear program is type-consistent if all
of its rules are type-consistent. Moreover, a limit-linear program P is type-consistent if the program obtained by first
semi-grounding P and then simplifying all numeric terms as
much as possible is type-consistent.
Conclusion and Future Work
We have introduced several decidable/tractable fragments of
Datalog with integer arithmetic, thus obtaining a sound theoretical foundation for declarative data analysis. We see
many challenges for future work. First, our formalism should
be extended with aggregate functions. While certain forms
of aggregation can be simulated by iterating over the object domain, as in our examples in Section 3, such a solution may be too cumbersome for practical use, and it relies on the existence of a linear order over the object domain, which is a strong theoretical assumption. Explicit
support for aggregation would allow us to formulate tasks
such as the ones in Section 3 more intuitively and without relying on the ordering assumption. Second, it is unclear whether integer constraint solving is strictly needed
in Step 7 of Algorithm 1: it may be possible to exploit
stability of P to compute TP (J) more efficiently. Third,
we shall implement our algorithm and apply it to practical data analysis problems. Fourth, it would be interesting to establish connections between our results and existing
work on data-aware artefact systems [Damaggio et al., 2012;
Koutsos and Vianu, 2017], which faces similar undecidability
issues in a different formal setting.
Acknowledgments
We thank Christoph Haase for explaining to us his results
on Presburger arithmetic and semi-linear sets as well as for
providing a proof for Lemma 19. Our work has also benefited from discussions with Michael Benedikt. This research
was supported by the Royal Society and the EPSRC projects
DBOnto, MaSI3 , and ED3 .
References
[Alvaro et al., 2010] Peter Alvaro, Tyson Condie, Neil Conway, Khaled Elmeleegy, Joseph M. Hellerstein, and Rus-
sell Sears. BOOM analytics: exploring data-centric,
declarative programming for the cloud. In EuroSys. ACM,
2010.
[Beeri et al., 1991] Catriel Beeri, Shamim A. Naqvi, Oded
Shmueli, and Shalom Tsur. Set constructors in a logic
database language. J. Log. Program., 10(3&4), 1991.
[Berman, 1980] Leonard Berman. The complexitiy of logical theories. Theor. Comput. Sci., 11, 1980.
[Byrd et al., 1987] Richard H. Byrd, Alan J. Goldman, and
Miriam Heller. Recognizing unbounded integer programs.
Oper. Res., 35(1), 1987.
[Chin et al., 2015] Brian Chin, Daniel von Dincklage, Vuk
Ercegovac, Peter Hawkins, Mark S. Miller, Franz Josef
Och, Christopher Olston, and Fernando Pereira. Yedalog:
Exploring knowledge at scale. In SNAPL, 2015.
[Chistikov and Haase, 2016] Dmitry
Chistikov
and
Christoph Haase. The taming of the semi-linear set.
In ICALP, 2016.
[Consens and Mendelzon, 1993] Mariano P. Consens and
Alberto O. Mendelzon. Low complexity aggregation in
GraphLog and Datalog. Theor. Comput. Sci., 116(1), 1993.
[Damaggio et al., 2012] Elio Damaggio, Alin Deutsch, and
Victor Vianu. Artifact systems with data dependencies and
arithmetic. ACM Trans. Database Syst., 37(3):22:1–22:36,
2012.
[Dantsin et al., 2001] Evgeny Dantsin, Thomas Eiter, Georg
Gottlob, and Andrei Voronkov. Complexity and expressive
power of logic programming. ACM Comput. Surv., 33(3),
2001.
[Eisner and Filardo, 2011] Jason Eisner and Nathaniel Wesley Filardo. Dyna: Extending datalog for modern AI. In
Datalog, 2011.
[Faber et al., 2011] Wolfgang Faber, Gerald Pfeifer, and
Nicola Leone. Semantics and complexity of recursive aggregates in answer set programming. Artif. Intell., 175(1),
2011.
[Ganguly et al., 1995] Sumit Ganguly, Sergio Greco, and
Carlo Zaniolo. Extrema predicates in deductive databases.
J. Comput. Syst. Sci., 51(2), 1995.
[Grädel, 1988] Erich Grädel. Subclasses of presburger arithmetic and the polynomial-time hierarchy. Theor. Comput.
Sci., 56, 1988.
[Haase, 2014] Christoph Haase. Subclasses of Presburger
arithmetic and the weak EXP hierarchy. In CSL-LICS,
2014.
[Hougardy, 2010] Stefan Hougardy. The Floyd-Warshall algorithm on graphs with negative cycles. Inf. Process. Lett.,
110(8-9), 2010.
[Kannan, 1987] Ravi Kannan. Minkowski’s convex body
theorem and integer programming. Math. Oper. Res.,
12(3):415–440, 1987.
[Kemp and Stuckey, 1991] David B. Kemp and Peter J.
Stuckey. Semantics of logic programs with aggregates. In
ISLP, 1991.
[Koutsos and Vianu, 2017] Adrien Koutsos and Victor
Vianu. Process-centric views of data-driven business
artifacts. J. Comput. System Sci., 86:82–107, 2017.
[Loo et al., 2009] Boon Thau Loo, Tyson Condie, Minos N.
Garofalakis, David E. Gay, Joseph M. Hellerstein, Petros
Maniatis, Raghu Ramakrishnan, Timothy Roscoe, and Ion
Stoica. Declarative networking. Commun. ACM, 52(11),
2009.
[Markl, 2014] Volker Markl. Breaking the chains: On
declarative data analysis and data independence in the big
data era. PVLDB, 7(13), 2014.
[Mazuran et al., 2013] Mirjana Mazuran, Edoardo Serra,
and Carlo Zaniolo. Extending the power of datalog recursion. VLDB J., 22(4), 2013.
[Mumick et al., 1990] Inderpal Singh Mumick, Hamid Pirahesh, and Raghu Ramakrishnan. The magic of duplicates
and aggregates. In VLDB, pages 264–277, 1990.
[Papadimitriou, 1981] Christos H. Papadimitriou. On the
complexity of integer programming. J. ACM, 28(4), 1981.
[Ross and Sagiv, 1997] Kenneth A. Ross and Yehoshua Sagiv. Monotonic aggregation in deductive databases. J.
Comput. System Sci., 54(1), 1997.
[Schöning, 1997] Uwe Schöning. Complexity of presburger
arithmetic with fixed quantifier dimension. Theory Comput. Syst., 30(4), 1997.
[Seo et al., 2015] Jiwon Seo, Stephen Guo, and Monica S.
Lam. SociaLite: An efficient graph query language based
on datalog. IEEE Trans. Knowl. Data Eng., 27(7), 2015.
[Shkapsky et al., 2016] Alexander Shkapsky, Mohan Yang,
Matteo Interlandi, Hsuan Chiu, Tyson Condie, and Carlo
Zaniolo. Big data analytics with datalog queries on Spark.
In SIGMOD. ACM, 2016.
[Van Gelder, 1992] Allen Van Gelder. The well-founded semantics of aggregation. In PODS, 1992.
[von zur Gathen and Sieveking, 1978] Joachim von zur Gathen and Malte Sieveking. A bound on solutions of linear integer equalities and inequalities. Proc. AMS, 72(1),
1978.
[Wang et al., 2015] Jingjing Wang, Magdalena Balazinska,
and Daniel Halperin. Asynchronous and fault-tolerant
recursive datalog evaluation in shared-nothing engines.
PVLDB, 8(12), 2015.
A
Proofs for Section 3
Theorem 1. For P a Datalog Z program and α a fact, checking P |= α is undecidable even if P contains no × or − and each
standard atom in P has at most one numeric term.
Proof. We prove the claim by presenting a reduction of the halting problem for deterministic Turing machines on the empty
tape. Let M be an arbitrary deterministic Turing machine with finite alphabet Γ containing the blank symbol , the finite set of
states S containing the initial state s and the halting state h, and transition function δ : S × Γ → S × Γ × {L, R}. We assume
that M works on a tape that is infinite to the right, that it starts with the empty tape and the head positioned on the leftmost cell,
and that it never moves the head off the left edge of the tape.
We encode each time point i ≥ 0 using an integer 2i , and we index tape positions using zero-based integers; thus, at time i,
each position j ≥ 2i is necessarily empty, so we can encode a combination of a time point i and tape position j with 0 ≤ j < 2i
using a single integer 2i + j. We use this idea to encode the state of the execution of M using the following facts:
• Num(k) is true for each positive number k;
• Time(k) is true if k = 2i and so k encodes a time point i;
• Tape(a, 2i + j) says that symbol a occupies position j of the tape at time i, and it will be defined for each 0 ≤ j < 2i ;
• Pos(2i + j) says that the head points to position j of the tape at time i;
• State(q, 2i ) says that the machine is in state q at time i; and
• Halts is a propositional variable saying that the machine has halted.
We next give a Datalog Z program PM that simulates the behaviour of M on the empty tape. We represent each alphabet
symbol a ∈ Γ using an object constant a, and we represent each state q ∈ S using an object constant q. Furthermore, we
.
abbreviate (s ≤ t) ∧ (t < u) as (s ≤ t < u). Finally, we abbreviate conjunction (s ≤ t) ∧ (t ≤ s) as (s = t), and disjunction
.
(s < t) ∨ (t > s) as (s =
6 t). Strictly speaking, disjunctions are not allowed in rule bodies; however, each rule with a disjunction
.
in the body of the form ϕ ∧ (s =
6 t) → α corresponds to rules ϕ ∧ (s < t) → α and ϕ ∧ (t < s) → α, so we use the former
form for the sake of clarity. With these considerations in mind, program PM contains rules (15)–(24).
→ Num(1)
Num(x) → Num(x + 1)
→ Time(1)
Time(x) → Time(x + x)
→ Tape( , 1)
→ Pos(1)
→ State(s, 1)
State(h, x) → Halts
.
Time(x) ∧ Tape(v, y) ∧ Pos(z) ∧ (x ≤ y < x + x) ∧ (x ≤ z < x + x) ∧ (y =
6 z) → Tape(v, x + y)
Time(x) ∧ Num(u) ∧ (x + x + x ≤ u + u < x + x + x + x) → Tape( , u)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
Moreover, for each alphabet symbol a ∈ Γ and all states q, q 0 ∈ S such that δ(q, a) = (q 0 , a0 , D) where D ∈ {L, R} is a direction, PM contains rules (25)–(28).
Time(x) ∧ State(q, x) ∧ Tape(a, y) ∧ Pos(y) ∧ (x ≤ y < x + x) → Tape(a0 , x + y)
0
(25)
Time(x) ∧ State(q, x) ∧ Tape(a, y) ∧ Pos(y) ∧ (x ≤ y < x + x) → State(q , x + x)
(26)
Time(x) ∧ State(q, x) ∧ Tape(a, y) ∧ Pos(y) ∧ Num(u) ∧
.
(x ≤ y < x + x) ∧ (x + y = u + 1) → Pos(u)
if D = L
(27)
Time(x) ∧ State(q, x) ∧ Tape(a, y) ∧ Pos(y) ∧ Num(u) ∧
.
(x ≤ y < x + x) ∧ (x + y + 1 = u) → Pos(u)
if D = R
(28)
Rules (15)–(16) initialise Num so that it holds of all positive integers, and rules (17)–(18) initialise Time so that it holds for
each integer k = 2i . Rules (19)–(21) initialise the state of the M at time i = 0. Rule (22) derives Halts if at any point the
Turing machine enters the halting state h. The remaining rules encode the evolution of the state of M , and they are based on
the following idea: if variable x encodes a time point i using value 2i , then variable y encodes a position j for time point i if
x ≤ y < x + x holds; moreover, for such y, position j at time point i + 1 is encoded as 2i+1 + j = 2i + 2i + j and can be
obtained as x+y, and the encodings of positions j −1 and j +1 can be obtained as x + y − 1 and x + y + 1, respectively. Since
our goal is to prove undecidability by just using +, we simulate subtraction by looking for a value u such that x + y = u + 1.
With these observations in mind, one can see that rule (23) copies the unaffected part of the tape from time point i to time point
i + 1. Moreover, rule (24) pads the tape by filling each location j with 1.5 × 2i ≤ j < 2 × 2i with the blank symbol; since
division is not supported in our language, we express this condition as 3 × 2i ≤ 2 × j < 2 × 2 × 2i . Finally, rule (25) updates
the tape at the position of the head; rule (26) updates the state; and rules (27) and (28) move the head left and right, respectively.
Consequently, we have PM |= Halts if and only if M halts on the empty tape.
Proposition 4. For each limit program P and fact α, a homogeneous program P 0 and fact α0 can be computed in linear time
such that P |= α if and only if P 0 |= α0 .
Proof. Let P be an arbitrary limit program. Without loss of generality, we construct a program P 0 containing only max
predicates. For each min predicate A, let A0 be a fresh max predicate uniquely associated with A. We construct P 0 from P by
modifying each rule r ∈ P as follows:
1. if h(r) = A(t, s) where A is a min predicate, replace the head of r with A0 (t, −s);
2. for each body atom A(t, n) ∈ sb(r) where A is a min predicate and n is a variable, replace the atom with A0 (t, m) where
m is a fresh variable, and replace all other occurrences of n in the rule with −m;
3. for each body atom A(t, k) ∈ sb(r) where A is a min predicate and k is an integer, replace the atom with A0 (t, −k).
Finally, if α = A(a, k) is a min fact, let α0 = A(a, −k); otherwise, let α0 = α. Now consider an arbitrary interpretation I, and
let I 0 be the interpretation obtained from I by replacing each min fact A(a, k) with A0 (a, −k); it is straightforward to see that
I |= P if and only if I 0 |= P 0 , and that I |= α if and only if I 0 |= α0 . Thus, P |= α if and only if P 0 |= α0 .
B
Proofs for Section 4
We use the standard notion of substitutions—sort-compatible partial mappings of variables to constants. For ϕ a formula and σ
a substitution, ϕσ is the formula obtained by replacing each free variable x in ϕ on which σ is defined with σ(x).
Proposition B.1. For each semi-ground rule r = ϕ → α, each pseudo-interpretation J, and each mapping σ of the variables
of r to integers, σ is an integer solution to C(r, J) if and only if J |= ϕσ.
Proof. (⇒) Assume that σ is an integer solution to C(r, J). We consider each atom β ∈ ϕ and show that J |= βσ holds.
• If β is a comparison atom, the claim is straightforward due to β ∈ C(r, J).
• If β is an object atom or an ordinary numeric atom, β is ground and we have β ∈ J and J |= β; otherwise,
(0 < 0) ∈ C(r, J) would hold and so σ could not be a solution to C(r, J).
• If β is a max atom B(b, s), since (0 < 0) 6∈ C(r, J), either B(b, `) ∈ J for some integer ` ∈ Z and (s ≤ `) ∈ C(r, J),
or B(b, ∞) ∈ J. In the former case, since σ is a solution to C(r, J), we have sσ ≤ `, and, since B is a max predicate,
J |= B(b, sσ) holds. In the latter case, J |= B(b, sσ) holds due to {B(b, sσ)} v {B(b, ∞)}.
• If β is a min atom, the proof is analogous to the previous case.
The proof of the (⇐) direction is analogous and we omit it for the sake of brevity.
Definition B.2. Given a limit-closed interpretation I and a program P, let
V
V
IP (I) = {γ | i αi → β is a ground instance of a rule in P such that I |= i αi and γ is a fact such that {γ} v {β}}.
S
∞
n
Let I0P = ∅, let InP = IP (In−1
P ) for n > 0, and let IP =
n≥0 IP .
Lemma B.3. For each program P, operator IP is monotonic w.r.t. ⊆; moreover, for I an interpretation, I |= P if and only if
IP (I) = I, and I |= P implies I∞
P ⊆ I.
Proof. Operator IP is the standard immediate consequence operator of Datalog, but applied to the program P 0 obtained by
extending P with the rules from Section 3 encoding the semantics of limit predicates. Thus, all claims of this lemma hold in
the usual way [Dantsin et al., 2001].
Lemma B.4. For each limit-closed interpretation I and the corresponding pseudo-interpretation J, and for each semi-ground
limit program P, interpretation IP (I) corresponds to the pseudo-interpretation TP (J).
Proof. It suffices to show that, for each fact α, the following claims hold:
1. if α is an object fact or ordinary numeric fact, then α ∈ IP (I) if and only if α ∈ TP (J);
2. if α is a limit fact of the form A(a, k) where k is an integer, then α ∈ IP (I) if and only if {α} v TP (J); and
3. if α is a limit fact of the form A(a, ∞), then {A(a, k) | k ∈ Z} ⊆ IP (I) if and only if α ∈ TP (J).
(Claim 1) Consider an arbitrary object fact α of the form A(a); the proof for ordinary numeric facts is analogous.
• Assume α ∈ IP (I). Then, a rule r = ϕ → α ∈ P and a grounding σ of r exist such that I |= ϕσ; the head of r must be
α since P is semi-ground. But then, J |= ϕσ holds as well, so Proposition B.1 ensures that σ is a solution to C(r, J);
moreover, hd(r, J) = α, and thus we have α ∈ TP (J).
• Assume α ∈ TP (J). Then, there exist a rule r = ϕ → α ∈ P and an integer solution σ to C(r, J). Proposition B.1 then
ensures J |= ϕσ, and so I |= ϕσ holds as well. Thus, we have ασ = α ∈ IP (I).
(Claim 2) Consider an arbitrary max fact α of the form A(a, k); the proof for a min fact is analogous.
• Assume α ∈ IP (I). Then, a rule r = ϕ → A(a, s) ∈ P and a grounding σ of r exist such that I |= ϕσ and α = A(a, sσ).
But then, J |= ϕσ holds as well, so Proposition B.1 ensures that σ is a solution to C(r, J); moreover, sσ ≤ opt(r, J), and
therefore we have {α} v {hd(r, J)} = {A(a, opt(r, J))} v TP (J).
• Assume {α} v TP (J). Then, there exist a rule r = ϕ → A(a, s) ∈ P and an integer solution σ to C(r, J) such that
{α} v {hd(r, J)} = {A(a, sσ)} where sσ = opt(r, J). Proposition B.1 then ensures J |= ϕσ, and so I |= ϕσ holds as
well. Thus, γ ∈ IP (I) holds for each fact γ with {γ} v {A(a, sσ)}, so we have α ∈ IP (I).
(Claim 3) Consider an arbitrary max fact α of the form A(a, ∞); the proof for a min fact is analogous. In the following, let
S = {A(a, k) | k ∈ Z}.
• Assume S ⊆ IP (I). Program P contains only finitely many rules, so the infinitely many facts of S in IP (I) are produced
by a rule r = ϕ → A(a, s) ∈ P and an infinite sequence (σi )i≥0 of groundings of r such that, for each i, we have I |= ϕσi
and sσi < sσi+1 . But then, J |= ϕσi , so Proposition B.1 ensures that σi satisfies C(r, J) for each i ≥ 0; therefore,
opt(r, J) = ∞ and α ∈ TP (J) holds.
• Assume A(a, ∞) ∈ TP (J). Then, a rule r = ϕ → A(a, s) ∈ P exists such that opt(r, J) = ∞, so an infinite sequence
(σi )i≥0 of solutions to C(r, J) exists such that sσi < sσi+1 for each i ≥ 0. Proposition B.1 ensures J |= ϕσi for each
i ≥ 0, and so I |= ϕσi as well. Thus, for each k ∈ Z, some i ≥ 0 exists such that k ≤ sσi , and therefore we have
{A(a, k)} v {A(a, sσi )}; consequently, S ⊆ IP (I) holds.
Lemma 14. For each semi-ground limit program P, operator TP is monotonic w.r.t. v. Moreover, J |= P if and only if
TP (J) v J for each pseudo-interpretation J.
Proof. Immediate from Lemmas B.3 and B.4.
∞
Theorem 15. For P a semi-ground limit program and α a fact, P |= α if and only if T∞
P |= α; also, |TP | ≤ |P|; and J |= P
implies T∞
v
J
for
each
pseudo-interpretation
J.
P
Proof. By inductively applying Lemma B.4, for each n ≥ 0, the limit-closed interpretation InP clearly corresponds to the
∞
pseudo-interpretation TnP . Thus, I∞
P and TP also correspond on all object and ordinary numeric facts. Now consider an
arbitrary n-ary max predicate A and a tuple of n − 1 objects a, and for M = {k | A(a, k) ∈ I∞
P } consider the following cases.
n
• M = ∅. Then, for each n ≥ 0 and each k ∈ Z, we have A(a, k) 6∈ IP , which implies A(a, k) 6∈ TnP and A(a, ∞) 6∈ TnP .
∞
∞
Finally, T∞
P is the least (w.r.t. v) fixpoint of TP , so A(a, k) 6∈ TP and A(a, ∞) 6∈ TP holds as well.
0
• There exists ` = max M . Then, there exists n ≥ 0 such that A(a, `) ∈ InP , and A(a, `0 ) 6∈ Im
P for each ` > ` and m ≥ 0;
0
∞
is
the
least
(w.r.t.
v) fixpoint of
for
each
`
>
`
and
m
≥
0;
finally,
T
but then, A(a, `) ∈ TnP , and A(a, `0 ) 6∈ Tm
P
P
holds.
operator TP , so A(a, `) ∈ T∞
P
• M = Z. Then, for each k ∈ Z, there exists n ≥ 0 such that A(a, k) ∈ InP , and so TnP |= A(a, k) holds; but then,
A(a, ∞) ∈ TnP holds as well.
∞
Analogous reasoning holds for min predicates, so I∞
P corresponds to TP . But then, the first and the third claim of this theorem
follow straightforwardly from Lemma B.3. Moreover, pseudo-interpretations contain at most one fact per combination of a
limit predicate and a tuple of objects (of corresponding arity), and program P is semi-ground so each rule in P produces at
most one fact in T∞
P , which implies the second claim of this theorem.
C
Proofs for Section 5
Theorem 16. For P a semi-ground limit program and α a fact, checking P |= α and checking applicability of a rule of P to a
pseudo-interpretation are both undecidable.
Proof. We present a reduction from Hilbert’s tenth problem, which is to determine whether, given a polynomial P (x1 , . . . , xn )
over variables x1 , . . . , xn , equation P (x1 , . . . , xn ) = 0 has integer solutions. It is well known that the problem remains undecidable even if the solutions must be nonnegative integers, so we use that variant in this proof. For each such polynomial P , let
PP be the program containing rules (29)–(30) for A a unary min predicate and B a nullary object predicate; it is obvious that
PP |= B if and only if P (x1 , . . . , xn ) = 0 has a nonnegative integer solution.
→ A(0)
(29)
Vn
i=1
A(xi ) ∧ (P (x1 , . . . , xn ) ≤ 0) ∧ (P (x1 , . . . , xn ) ≥ 0) → B
(30)
Moreover, rule (30) is applicable to J = {A(0)} if and only if P (x1 , . . . , xn ) = 0 has a nonnegative integer solution.
Although Presburger arithmetic does not have propositional variables, these can clearly be axiomatised using numeric variables. Hence, in the rest of this section we use propositional variables in Presburger formulas for the sake of clarity.
Definition C.1. For each n-ary object predicate A, each (n + 1)-ary ordinary numeric predicate B, each (n + 1)-ary limit
predicate C, each n-tuple of objects a, and each integer k, let def Aa , def Bak , def Ca and fin Ca be distinct propositional
variables, and let val Ca a distinct integer variable.
Moreover, let C be ≤ (resp. ≥) if C is a max (resp. min) predicate.
V
For P a semi-ground program, Pres(P) = r∈P Pres(r) is the Presburger formula where Pres(r) = ∀y.r0 for y all numeric
variables in r and r0 is obtained by replacing each atom α in r with its encoding Pres(α) defined as follows:
• Pres(α) = α if α is a comparison atom;
• Pres(α) = def Aa if α is an object atom of the form A(a);
• Pres(α) = def Bak if α is an ordinary numeric atom of the form B(a, k); and
• Pres(α) = def Ca ∧ (¬fin Ca ∨ s C val Ca ) if α is a limit atom of the form C(a, s).
Let J be a pseudo-interpretation, and let µ be an assignment of Boolean and integer variables. Then, J corresponds to µ if
all of the following conditions hold for all A, B, C, and a as specified above, for each integer k ∈ Z:
• µ(def Aa ) = true if and only if A(a) ∈ J;
• µ(def Bak ) = true if and only if B(a, k) ∈ J;
• µ(def Ca ) = true if and only if C(a, ∞) ∈ J or there exists ` ∈ Z such that C(a, `) ∈ J;
• µ(fin Ca ) = true and µ(val Ca ) = k if and only if C(a, k) ∈ J.
Note that k in Definition C.1 ranges over all integers (which excludes ∞), µ(val Ca ) is an equal to some integer k, and J is
a pseudo-interpretation and thus cannot contain both C(a, ∞) and C(a, k); thus, C(a, ∞) ∈ J implies µ(fin Ca ) = false.
Also note that each assignment µ corresponds to precisely one J; however, each J corresponds to infinitely many assignments
µ since Definition C.1 does not restrict the value of variables other than def Aa , def Bak , def Ca , fin Ca , and val Ca . Moreover,
two assignments corresponding to the same pseudo-interpretation may differ on the value of val Ca if fin Ca is set to false in
both assignments, and they can differ on the values of fin Ca and val Ca if def Ca is set to false in both assignments.
Lemma C.2. Let J be a pseudo-interpretation and let µ be a variable assignment such that J corresponds to µ. Then,
1. J |= α if and only if µ |= Pres(α) for each ground atom α, and
2. J |= r if and only if µ |= Pres(r) for each semi-ground rule r.
Proof. (Claim 1) We consider all possible forms of α.
• α is a comparison atom. Then, the truth of α is independent from J so the claim is immediate.
• α = A(a) is an object fact. Then, Pres(α) = def Aa , and µ(def Aa ) = true if and only if def Aa ∈ J, so the claim holds.
• α = B(a, k) is an ordinary numeric fact. The proof is analogous to the case of object facts.
• α = C(a, k) is a limit fact. If J |= α, then either C(a, ∞) ∈ J or an integer ` exists such that C(a, `) ∈ J and k C `;
either way, µ(def Ca ) = true holds; moreover, µ(fin Ca ) = false holds in the former and µ(val Ca ) = ` holds in the latter
case; thus, µ |= Pres(r) clearly holds. The converse direction is analogous so we omit it for the sake of brevity.
(Claim 2) Let r be an arbitrary semi-ground rule, and let I be the limit-closed interpretation corresponding to J. By definition,
J |= r if and only if I |= r, and the latter is equivalent to I |= r0 for each ground instance r0 of r by the semantics of universal
quantification in first-order logic; but then, the latter claim is equivalent to J |= r0 for each ground instance r0 of r.
Now note that, by construction, we have Pres(βσ) = Pres(β)σ for each semi-ground atom β and each grounding σ, and
thus Pres(rσ) = Pres(r)σ. Finally, groundings of r can be equivalently seen as variable assignments to universally quantified
numeric variables in Pres(r), so Claim 2 follows immediately from Claim 1.
Wn
Lemma 18. For P a semi-ground limit-linear program and α a fact, there exists a Presburger sentence ϕ = ∀x∃y. i=1 ψi
that is valid if and only if P |= α. Each ψi is a conjunction of possibly negated atoms. Moreover, |x| + |y| and each kψi k are
bounded polynomially by kPk + kαk. Number n is bounded polynomially by |P| and exponentially by maxr∈P krk. Finally,
the magnitude of each integer in ϕ is bounded by the maximal magnitude of an integer in P and α.
Proof. Lemma C.2 immediately implies that P |= α if and only if the sentence ϕ0 = ∀x. Pres(α) ∨ ¬Pres(P) is valid, where
x contains all variables def Aa , def Bak , def Ca , fin Ca , and val Ca occurring in Pres(P) or Pres(α). Clearly, |x| is polynomially
bounded by kPk + kαk, and the magnitude of each integer in ϕ0 is bounded by the maximum magnitude of an integer in P
and α. Let ϕ1 be the sentence obtained from ϕ0 by converting each top-level conjunct of Pres(P) into form ∀yi .χi where χi
is in CNF. Formulae ϕ0 and ϕ1 are equivalent, and ϕ1 is of the form
Vn
V`i
χji ,
ϕ1 = ∀x. Pres(α) ∨ ¬ i=1 ∀yi . j=1
where n = |P| and, for each rule ri ∈ P, integer `i is exponentially bounded by kri k, and kχji k and |yi | are linearly bounded
by kri k. By moving all quantifiers to the front of the formula and pushing negations inwards, we finally obtain formula
Wn W`i
ψij ,
ϕ2 = ∀x∃y.Pres(α) ∨ i=1 j=1
Sn
where y = i=1 yi and each ψij is the negation-normal form of ¬χji . Formula ϕ2 is of the required form, |y| is bounded
Pn
polynomially by kPk, number i=1 `i is bounded polynomially by n = |P| and exponentially by maxr∈P krk, and kψij k is
bounded linearly by kPk.
Wn
Lemma 19. Let ϕ = ∀x∃y. i=1 ψi be a Presburger sentence where each ψi is a conjunction of possibly negated atoms
of size at most k mentioning at most ` variables, a is the maximal magnitude of an integer in ϕ, and m = |x|. Then, ϕ
is valid if and only if ϕ is valid over models where each integer variable assumes a value whose magnitude is bounded by
`
4
(2O(` log `) · ak` )n2 ·O(m ) .
Proof. Let m0 = |y|. Each ψi can be seen as a system of linear inequalities Si = (Ai x ≤ ci ) such that |x| ≤ `, |ci | ≤ k,
and where the maximal magnitude of all numbers in A and ci is bounded by ak . By Proposition 3 of Chistikov and Haase
[2016] (adapted from
S the work by von zur Gathen and Sieveking [1978]), the set of solutions to Si can be represented by
a semi-linear set i0 ∈Ni L(bi0 , Pi0 ) where bi0 ∈ Z` , Pi0 ⊆ Z` , |Ni | ≤ 2` , and the magnitude of all integers in bi0 and Pi0
Wn
S
is bounded by 2O(` log `) · ak` . Consequently, disjunction i=1 ψi corresponds to a semi-linear set j∈M L(bj , Pj ) where
0
0
O(` log `)
bj ∈ Zm+m W
, Pj ⊆ Zm+m , |M | ≤ n2` , and the magnitude
· ak` .
S of each integer in bj and Pj is still bounded by 2
n
Formula ∃y. i=1 ψi then corresponds to the projection of j∈M L(bj , Pj ) on the variables in x, which is a semi-linear set of
S
the form j∈M L(b0j , Pj0 ) where each b0j ∈ Zm is a projection of bj on x, and each Pj0 ⊆ Zm is a projection of Pj on x. Now,
Wn
0
Theorem 21 by Chistikov and Haase
S [2016] implies that the satisfying assignments to the formula ϕ = ¬∃y. i=1 ψi can be
represented as a semi-linear set j 0 ∈M 0 L(cj 0 , Qj 0 ) where the magnitude of each integer in each cj 0 and Qj 0 is bounded by
`
4
b = (2O(` log `) · ak` )n2 ·O(m ) . Since ϕ0 has a satisfying assignment if and only if it has a satisfying assignment involving only
numbers from some cj 0 , it follows that ϕ0 is satisfiable if and only if it is satisfiable over models where the absolute value of
every integer variable is bounded by b. This implies the claim of this lemma since ϕ is valid if and only if ϕ0 is unsatisfiable.
Theorem 20. For P a semi-ground limit-linear program, D a dataset, and α a fact, P ∪ D 6|= α if and only if a pseudo-model
J of P ∪ D exists where J 6|= α, |J| ≤ |P ∪ D|, and the magnitude of each integer in J is bounded polynomially by the largest
magnitude of an integer in P ∪ D, exponentially by |P|, and double-exponentially by maxr∈P krk.
Proof. The ⇐ direction is trivial. For the ⇒ direction assume that P ∪ D 6|= α holds, and let D0 be obtained
Wn from D by
removing each fact that does not unify with an atom in P or α; clearly, we have P ∪ D0 6|= α. Let ϕ = ∀x∃y. i=1 ψi be the
Presburger sentence from Lemma 18 for P ∪ D0 and α. Sentence ϕ is not valid and it satisfies the following conditions.
• Number m = |x| is polynomial in kP ∪ D0 k, which, in turn, is bounded by |P ∪ D0 | · maxr∈P∪D0 krk. Moreover, D0
contains only facts that unify with atoms in P and α, so m can be bounded further, namely linearly in the product cs, for
c = |P| and s = maxr∈P kPk.
• Number n is linear in the product of c and 2s .
• The size, and hence the number ` of variables in each ψi , are linear in s.
Let a be the maximal magnitude of an integer in P ∪ D0 (and thus in ϕ as well). By Lemma 19, an assignment µ exists such
2
s
O(s)
4
that µ 6|= ϕ and the magnitude of each integer variable is bounded by b = (2O(s log s) · aO(s ) )O(c·2 )·2 ·O((cs) ) . Clearly,
b is polynomial in a, exponential in c, and doubly-exponential in s, as required. Moreover, clearly µ |= Pres(P ∪ D0 ) and
µ 6|= Pres(α). Now let J 0 be the pseudo-model corresponding to µ; by Lemma C.2, we have J 0 |= P ∪ D0 and J 0 6|= α. By
construction, the magnitude of each integer in J 0 is bounded by b. Furthermore, let J be the restriction of J 0 to the facts that
unify with the head of at least one rule in P ∪ D0 ; clearly, we still have J |= P ∪ D0 and J 6|= α. Finally, |J| ≤ |P ∪ D0 | holds
by our construction, which implies our claim.
Lemma C.3. For each semi-ground, limit-linear program P, pseudo-interpretation J, and dataset D, there exists a polynomial p such that TP∪D (J) can be computed in nondeterministic polynomial time in kPk + kDk + kJk, and in deterministic
p(maxr∈P krk)
polynomial time in kPk + kDk + kJk
.
Proof. Let S = {hd(r, J) | rule r ∈ P ∪ D is applicable to J}. Program P is semi-ground, and therefore P ∪ D is semiground as well; thus, each rule of P ∪ D can contribute at most one fact to S, so we have |S| ≤ |P| + |D|. By Definition 12,
TP∪D (J) is the smallest (w.r.t. v) pseudo-interpretation such that TP∪D (J) |= S, so we can compute TP∪D (J) as the set
containing each object and ordinary numeric fact in S, each fact A(a, ∞) ∈ S for A a limit predicate, each fact A(a, `) ∈ S
such that A is a min (resp. max) predicate and A(a, k) ∈ S implies k 6= ∞ and k ≥ ` (resp. k ≤ `). To complete the proof of
this lemma, we next argue that set S can be computed within the required time bounds.
Consider an arbitrary rule r ∈ P ∪ D, and let J 0 be the subset of J containing all facts that unify with a body atom in r; note
that |J 0 | ≤ krk. Rule r is applicable to J if and only if conjunction C(r, J 0 ) has an integer solution. By construction, kC(r, J 0 )k
is linear in krk + kJ 0 k, the number of variables in C(r, J 0 ) and r is the same, |C(r, J 0 )| is linear in krk, and the magnitude of
each integer in C(r, J 0 ) is exponentially bounded in krk + kJk. But then, checking whether C(r, J) has an integer solution is
in NP w.r.t. krk + kJk, and in PT IME w.r.t. kJkp(krk) for some polynomial p, as we argue next.
• We first consider the former claim. Let a be the maximal magnitude of an integer in C(r, J 0 ). Conjunction C(r, J 0 ) contains
only the numbers from r and J 0 , whose magnitude is at most 2krk and 2kJk , respectively; thus, we have a ≤ 2krk+kJk .
Moreover, the results by Papadimitriou [1981] show that there exists a polynomial p1 such that the magnitude of an integer
in a solution to C(r, J 0 ) can be bounded by b = ap1 (krk) , and so there exists a polynomial p2 such that b ≤ 2p2 (krk+kJk) .
The binary representation of b thus requires at most krk + kJk bits, and so we can guess it in polynomial time.
• We next consider the latter claim. By Theorem 5.4 of Kannan [1987], checking satisfiability of C(r, J) over Z is fixedparameter tractable in the number n of variables in r–that is, there exists a polynomial p3 such that a solution to C(r, J)
can be computed in time O((krk + kJk) · 2p3 (n) ). Since n ≤ krk clearly holds, there exists a polynomial p4 such that the
satisfiability of C(r, J) can be checked in time that is thus kJkp4 (krk) .
Now assume that r is applicable to J. Then hd(r, J) = h(r) if h(r) is an object atom, so we assume that h(r) = A(a, s) is a
limit atom and argue that opt(r, J) can be computed within the required time bounds using the following two steps.
1. Depending on whether A is a min or a max predicate, we check whether there is a smallest/largest value for s in all solutions
to C(r, J)—that is, we check whether the integer linear program ‘minimise/maximise s subject to C(r, J)’ is bounded.
Byrd et al. [1987] showed that this amounts to checking boundedness of the corresponding linear relaxation, which in turn
can be reduced to checking linear feasibility and can be solved in deterministic polynomial time in krk + kJk.
2. If the above problem is bounded, we compute its optimal solution, which can be reduced to polynomially many (in
krk + kJk) feasibility checks, as shown by Papadimitriou [1981] (Corollary 2 with binary search). Each such feasibility
check is in NP w.r.t. krk + kJk, and in PT IME w.r.t. kJkp(krk) .
Thus, hd(r, J) can be computed in nondeterministic polynomial time in krk + kJk, and in deterministic polynomial time in
kJkp(krk) , which implies our claim.
Lemma C.4. Deciding P |= α is CO NP-hard in data complexity for P a limit-linear program and α a fact.
Proof. An instance T of the square tiling problem is given by an integer N coded in unary, a set T = {t0 , . . . , tM −1 } of
M tiles, and two compatibility relations H ⊆ T × T and V ⊆ T × T . The problem is to determine whether there exists a
tiling τ : {0, . . . , N − 1}2 → T of an N × N square such that hτ (i, j), τ (i + 1, j)i ∈ H holds for all 0 ≤ i < N − 1 and
0 ≤ j < N , and hτ (i, j), τ (i, j + 1)i ∈ V holds for all 0 ≤ i < N and 0 ≤ j < N − 1, which is known to be NP-complete.
Thus, to prove the claim of this lemma, we reduce the complement of the problem by presenting a fixed program Ptiling and a
dataset DT (that depends on T ), and showing that T has no solution if and only if Ptiling ∪ DT |= noSolution.
Our encoding uses object EDB predicates succ, incompatibleH , and incompatibleV ; ordinary numeric EDB predicates
shift, tileNo, numTiles, and maxTiling; nullary object IDB predicate noSolution; unary min IDB predicate I; and unary
.
max IDB predicate tiling. Program Ptiling contains rules (31)–(35), where (s = t) abbreviates (s ≤ t) ∧ (t ≤ s).
→ I(0)
→ tiling(0)
(31)
(32)
tiling(n) ∧ numTiles(nt) ∧
.
shift(x, y, s) ∧ tileNo(u, t) ∧ I(m1 ) ∧ I(m2 ) ∧ (n = m1 × nt × s + t × s + m2 ) ∧ (m2 < s) ∧
succ(x, x0 ) ∧
(33)
. 0
0
0
0 0
0
0
0
0
0
0
shift(x , y, s ) ∧ tileNo(u , t ) ∧ I(m1 ) ∧ I(m2 ) ∧ (n = m1 × nt × s + t × s + m2 ) ∧ (m02 < s0 ) ∧
incompatibleH (u, u0 ) → tiling(n + 1)
tiling(n) ∧ numTiles(nt) ∧
.
shift(x, y, s) ∧ tileNo(u, t) ∧ I(m1 ) ∧ I(m2 ) ∧ (n = m1 × nt × s + t × s + m2 ) ∧ (m2 < s) ∧
succ(y, y 0 ) ∧
(34)
. 0
0 0
0 0
0
0
0
0
0
0
shift(x, y , s ) ∧ tileNo(u , t ) ∧ I(m1 ) ∧ I(m2 ) ∧ (n = m1 × nt × s + t × s + m2 ) ∧ (m02 < s0 ) ∧
incompatibleV (u, u0 ) → tiling(n + 1)
tiling(n) ∧ maxTiling(m) ∧ (m < n) → noSolution
(35)
Dataset DT contains facts (36)–(42), where g0 , . . . , gN −1 are fresh objects, and ti for 0 ≤ i < M are distinct objects corre2
sponding to the tiles in T . Since N is coded in unary, although numbers M N − 1 and M i+N j are exponential in N , they can
be computed in polynomial time and represented using polynomially many bits.
→ numTiles(M )
→ tileNo(ti , i)
→ incompatibleH (ti , tj )
→ incompatibleV (ti , tj )
for each 0 ≤ i < M
for each 0 ≤ i, j < M such that (ti , tj ) ∈
/H
for each 0 ≤ i, j < M such that (ti , tj ) ∈
/V
(36)
(37)
(38)
(39)
for each 0 ≤ i < N − 1
(40)
(41)
for each 0 ≤ i, j < N
(42)
2
→ maxTiling(M N − 1)
→ succ(gi , gi+1 )
→ shift(gi , gj , M
i+N j
)
Our reduction uses the following idea. Facts (37) associate each tile ti with an integer i where 0 ≤ i < M ; hence, in
the rest P
of this discussion, we do not distinguish a tile from its number. This allows us to represent each tiling τ using a
number 0≤i,j<N τ (i, j) × M i+N j . Thus, given a number n that encodes a tiling, number t with 0 ≤ t < M corresponds
to the tile assigned to position (i, j) if n = m1 × M × M i+N j + t × M i+N j + m2 for some integers m1 and m2 where
0 ≤ m2 < M i+N j . Thus, if numeric variable n is assigned such an encoding of a tiling and numeric variable s is assigned the
factor M i+N j corresponding to a position (i, j), then conjunction
.
tileNo(u, t) ∧ I(m1 ) ∧ I(m2 ) ∧ (n = m1 × nt × s + t × s + m2 ) ∧ (m2 < s)
is true if and only if u is assigned the tile object corresponding to position (i, j) in the tiling encoded by n. To complete the
construction, we represent each position (i, j) by a pair of objects (gi , gj ), each of which is associated with the corresponding
factor M i+N j using facts (42). Facts (41) provide an ordering on gi , which allows us to identify adjacent positions. Finally,
fact (40) records the maximal number that encodes a tiling as we outlined earlier. Program Ptiling then simply checks through
all tilings: rule (32) ensures that the tiling encoded as 0 is checked; moreover, for each n such that tiling(n) holds, rules (33)
and (34) derive tiling(n + 1) if either the horizontal or the vertical compatibility requirement is violated for the tiling encoded
2
by n. Finally, rule (35) detects that no solution exists if tiling(M N ) is derived.
Lemma C.5. Deciding P |= α is CO NE XP T IME-hard for P a limit-linear program and α a fact.
Proof. We present a reduction from the succinct square tiling problem. An instance T of the problem is given by an integer
N coded in unary, a set T containing M tiles, and horizontal and vertical compatibility relations H and V , respectively, as in
the proof of Lemma C.4; however, the objective is to tile a square of 2N × 2N positions, which is known to be NE XP T IMEcomplete. Thus, to prove the claim of this lemma, we reduce the complement of the problem by presenting a program PT and
showing that T has no solution if and only if PT |= noSolution.
The main idea behind our reduction is similar to Lemma C.4. Program PT contains rules (43)–(45) that associate each tile
with a number using an ordinary numeric predicate tileNo, and encode the horizontal and vertical incompatibility relations
using the object predicates incompatibleH and incompatibleV .
→ tileNo(ti , i)
→ incompatibleH (ti , tj )
→ incompatibleV (ti , tj )
for each 0 ≤ i < M
for each 0 ≤ i, j < M such that (ti , tj ) ∈
/H
for each 0 ≤ i, j < M such that (ti , tj ) ∈
/V
(43)
(44)
(45)
The main difference to Lemma C.4 is that, in order to obtain a polynomial encoding, we cannot represent a position (i, j)
in the grid explicitly using a pair of objects. Instead, we encode each position using a pair (i, j) where i and j are N -tuples of
objects 0̄ and 1̄. If we read 0̄ and 1̄ as representing numbers 0 and 1, respectively, then each i and j can be seen as a binary
number in [0, 2N − 1]. By a slight abuse of notation, we often identify a tuple over 0̄ and 1̄ with the number it encodes and use
tuples in arithmetic expressions. While positions can be encoded using N bits, we will also need to ensure distance between
positions, which requires N + 1 bits. In the rest of this proof, 0̄ and 1̄ stand for tuples 0̄, . . . , 0̄ and 1̄, . . . , 1̄, respectively, whose
length is often implicit from the context where these tuples occur. Similarly, x, x0 , y, and y0 are tuples of distinct variables
whose length will also be clear from the context.
To axiomatise an ordering on numbers with N bits, program PT contains rules (46)–(49), where B is a unary object predicate,
succ is a 2N -ary object predicate, and succ 0 is a (2N + 2)-ary object predicate. Rules (46)–(48) ensure PT |= succ(i, j)
where i and j encode numbers with N bits such that j = 1 + i; in particular, rule (48) encodes binary incrementation, where
x1̄0̄ = 1 + x1̄0̄ holds for each position k and each k-tuple of zeros and ones x. Rules (46)–(47) and (49) ensure an analogous
property for succ 0 , but for numbers with N + 1 bits.
→ B (0̄)
(46)
→ B (1̄)
(47)
Vk
B (xi ) → succ(x, 0̄, 1̄, x, 1̄, 0̄)
for each 0 ≤ k < N where |x| = k and |1̄| = |0̄| = N − k − 1
(48)
Vk
0
for each 0 ≤ k < N + 1 where |x| = k and |1̄| = |0̄| = N − k
(49)
i=1
i=1
B (xi ) → succ (x, 0̄, 1̄, x, 1̄, 0̄)
2N
Analogously to the proof of Lemma C.4, we encoded tilings using numbers in [0, M 2 − 1]. To compute the maximum
number encoding a tiling, program PT contains rules (50)–(53), where maxTiling is a unary min predicate, and auxT is
a (2N + 1)-ary min predicate. Auxiliary rules (50)–(52) multiply M with itself as many times as there are grid positions,
N
so we have PT |= auxT (i, j, M 1+i+2 ·j ) for each position (i, j). Consequently, rule (53) ensures that, for all s, we have
2N
PT |= maxTiling(s) if and only if s ≥ M 2 − 1.
→ auxT (0̄, 0̄, M )
0
(50)
0
auxT (x, y, n) ∧ succ(x, x ) → auxT (x , y, M × n)
0
(51)
0
auxT (1̄, y, n) ∧ succ(y, y ) → auxT (0̄, y , M × n)
(52)
auxT (1̄, 1̄, n) → maxTiling(n − 1)
(53)
Unlike in the proof of Lemma C.4, we cannot include shift factors explicitly into PT since this would make the encoding
exponential; moreover, we could precompute shift factors using rules similar to (50)–(52), but then we would need to use values
from limit predicates in multiplication, which would not produce a limit-linear program. Therefore, we check tilings using a
different approach. As in the proof of Lemma C.4, our construction ensures that, for all s, we have PT |= tiling(s) if and only
if each tiling n with 0 ≤ n ≤ s does not satisfy the compatibility relations. Given a tiling encoded by n and a position (i, j), let
j
k
n
sn,i,j =
.
N
i+j·2
M
Program PT contains rules (54)–(57) where shiftedTiling is a max predicate of arity 2N + 1 and I is a unary min predicate.
These rules ensure that, for each i, j, and tiling n such that PT |= tiling(n), we have PT |= shiftedTiling(i, j, sn,i,j ). To
understand how this is achieved, we order the grid positions as follows:
(0, 0), (1, 0), . . . , (2N − 1, 0), (0, 1), (1, 1), . . . , (2N − 1, 1), . . . . . . , (0, 2N − 1), (1, 2N − 1), . . . , (2N − 1, 2N − 1)
Now consider an arbitrary position (i, j) and its successor (i0 , j0 ) in the ordering. The encoding of a tiling using an integer n
ensures sn,i,j = M × sn,i0 ,j0 + t holds, where 0 ≤ t < M is the number of the tile that n assigns to position (i, j). Thus, rule
(55) ensures that position (0̄, 0̄) satisfies the mentioned property, rule (56) handles adjacent positions of the form (i, j) and
(i + 1, j), and rule (57) handles adjacent positions of the form (1̄, j) and (0̄, j + 1).
→ I (0)
tiling(n) → shiftedTiling(0̄, 0̄, n)
.
shiftedTiling(x, y, n) ∧ succ(x, x ) ∧ I (`) ∧ I (m) ∧ (` < M ) ∧ (n = M × m + `) → shiftedTiling(x0 , y, m)
.
shiftedTiling(1̄, y, n) ∧ succ(y, y0 ) ∧ I (`) ∧ I (m) ∧ (` < M ) ∧ (n = M × m + `) → shiftedTiling(0̄, y0 , m)
0
0
(54)
(55)
(56)
(57)
0
Note that, for all n and n with n < n and each position (i, j), we have sn,i,j < sn0 ,i,j . Thus, since shiftedTiling is a max
predicate, the limit value for s in shiftedTiling(i, j, s) will always correspond to the limit value for n in tiling(n).
N
Checking horizontal compatibility is now easy, but checking vertical compatibility requires dividing sn,i,j by M 2 , which
would make the reduction exponential. Hence, PT checks compatibility using rules (58)–(61), where conflict is a max predicate
of arity 3N + 3. These rules ensure that, for each i, j, d, u, and tiling n such that PT |= tiling(n) and the position that precedes
(i, j) by distance d in the ordering cannot be labelled in n with tile u, we have PT |= conflict(i, j, d, u, sn,i,j ). To this end,
assume that (x, y) is labelled with tile u0 ; now if (u, u0 ) 6∈ H and x 6= 0̄ (i.e., the predecessor x0 of x exists), then rule (58)
says that the position preceding (x, y) by 0̄1̄ (i.e., the position to the left) cannot be labelled with u; moreover, if (u, u0 ) 6∈ V
and y 6= 0̄ (i.e., the predecessor y0 of y exists), then rule (59) says that the position preceding (x, y) by 1̄0̄ = 2N (i.e., the
position above) cannot be labelled with u. Moreover, rule (60) propagates such constraints from position (i, j) to (i − 1, j) while
reducing the distance by one, and rule (61) does so for positions (0̄, j) and (1̄, j − 1).
shiftedTiling(x, y, m) ∧ succ(x0 , x) ∧ incompatibleH (u, u0 ) ∧ tileNo(u0 , t0 ) ∧
.
I (`) ∧ (m = M × ` + t0 ) → conflict(x, y, 0̄1̄, u, m)
(58)
shiftedTiling(x, y, m) ∧ succ(y0 , y) ∧ incompatibleV (u, u0 ) ∧ tileNo(u0 , t0 ) ∧
.
I (`) ∧ (m = M × ` + t0 ) → conflict(x, y, 1̄0̄, u, m)
(59)
shiftedTiling(x, y, m) ∧ succ(x, x0 ) ∧ conflict(x0 , y, z0 , u, m0 ) ∧ succ 0 (z, z0 ) ∧
.
I (`) ∧ (` < M ) ∧ (m = M × m0 + `) → conflict(x, y, z, u, m)
(60)
shiftedTiling(1̄, y, m) ∧ succ(y, y0 ) ∧ conflict(0̄, y0 , z0 , u, m0 ) ∧ succ 0 (z, z0 ) ∧
.
I (`) ∧ (` < M ) ∧ (m = M × m0 + `) → conflict(1̄, y, z, u, m)
(61)
Program PT also contains rules (62)–(64) where invalid is a (2N + 1)-ary max predicate. These rules ensure that, for
each i, j, and each tiling n such that PT |= tiling(n) and there exists a position (i0 , j0 ) that comes after (i, j) in the position
order such that n does not satisfy the compatibility relations between (i0 , j0 ) and its horizontal or vertical successor, we have
PT |= invalid (i, j, sn,i,j ). Rule (62) determines invalidity at position (x, y) for conflicts with zero distance, and rules (63) and
(64) propagate this information to preceding positions analogously to rules (60) and (61).
.
conflict(x, y, 0̄, u, m) ∧ tileNo(u, t) ∧ I (`) ∧ (m = M × ` + t) → invalid (x, y, m)
(62)
shiftedTiling(x, y, m) ∧ succ(x, x0 ) ∧ invalid (x0 , y, m0 ) ∧
.
I (`) ∧ (` < M ) ∧ (m = M × m0 + `) → invalid (x, y, m)
(63)
shiftedTiling(1̄, y, m) ∧ succ(y, y0 ) ∧ invalid (0̄, y0 , m0 ) ∧
.
I (`) ∧ (` < M ) ∧ (m = M × m0 + `) → invalid (1̄, y, m)
(64)
Finally, program PT contains rules (65)–(67) where tiling is a unary max predicate and noSolution is a nullary predicate.
Rule (65) ensures that tiling encoded by 0 is checked. Based on our discussion from the previous paragraph, for each invalid
tiling n such that PT |= tiling(n), we have PT |= invalid (0̄, 0̄, s0̄,0̄,n ); moreover, s0̄,0̄,n = n so, if PT |= invalid (0̄, 0̄, n)
.
holds, then rule (66) ensures that tiling encoded by n + 1 is considered; atom (m = n) is needed in the rule since no numeric
variable is allowed to occur in more than one standard body atom. If we exhaust all available tilings, rule (67) determines that
no solution exists, just as in the proof of Lemma C.4.
→ tiling(0)
.
invalid (0̄, 0̄, m) ∧ tiling(n) ∧ (m = n) → tiling(n + 1)
tiling(n) ∧ maxTiling(m) ∧ (m < n) → noSolution
(65)
(66)
(67)
Based on our discussion of the consequences of PT , we conclude that instance T of the succinct tiling problem does not
have a solution if and only if PT |= noSolution.
Proposition C.6. For J a pseudo-interpretation and α a fact, J |= α if and only if {α} v J.
Proof. Consider an arbitrary pseudo-interpretation J and the corresponding limit-closed interpretation I. If J |= α, then α ∈ I,
so there exists a fact α0 ∈ J such that such that {α} v {α0 }, which implies {α} v J. Moreover, if {α} v J, then there exists
a fact α0 ∈ J such that {α} v {α0 } and, since I is limit-closed, we have α ∈ I, which implies J |= α.
Theorem 21. For P a limit-linear program and α a fact, deciding P |= α is CO NE XP T IME-complete in combined and CO NPcomplete in data complexity.
Proof. Lemmas C.4 and C.5 prove hardness. Moreover, the following nondeterministic algorithm decides P ∪ D 6|= α in time
polynomial in kDk + kαk, and exponential in kPk + kDk + kαk.
1. Compute the semi-grounding P 0 of P.
2. Guess a pseudo-interpretation J over the signature of P 0 ∪ D such that the number of facts in J and the absolute values
of all integers in J are bounded as in Theorem 20.
3. Check that J is a pseudo-model of P ∪ D; if not, return false.
4. Return false if J |= α and true otherwise.
Correctness of the algorithm follows from Theorem 20, so we next argue about its complexity. The mentioned data complexity
holds by the following observations.
• In step 1, kP 0 k, |P 0 |, and the time required to compute P 0 are all polynomial in kDk and constant in kαk.
• Since |P 0 | is polynomial in kDk and constant in kαk, and maxr∈P 0 krk is constant in kDk and kαk, the magnitude of the
integers in J is exponentially bounded in kαk + kDk by Theorem 20; thus, the number of bits needed to represent each
integer in J is polynomial in kαk + kDk. Furthermore, we have |J| ≤ |D| + |P 0 |, and |D| + |P 0 | is polynomial in kDk
and constant in kαk; thus, J can be guessed in step 2 in nondeterministic polynomial time in kαk + kDk.
• By Lemma 14, checking that J is a pseudo-model of P ∪ D amounts to checking TP 0 ∪D (J) = J. By Lemma C.3,
TP 0 ∪D (J) can be computed in deterministic polynomial time in kJk + kDk and hence in kDk (as kJk is polynomial in
kDk). Hence, step 3 requires deterministic polynomial time in kαk + kDk.
• By Proposition C.6, step 4 amounts to checking {α} v J, which can be done in time polynomial in kJk and kαk, and
hence polynomial in kDk + kαk as well.
Finally, the mentioned combined complexity holds by the following observations.
• In step 1, kP 0 k, |P 0 |, and time required to compute P 0 are all exponential in kPk + kDk and constant in kαk.
• Since |P 0 | is exponential in kPk + kDk and constant in kαk, and maxr∈P 0 krk is linear in kPk and constant in
kDk + kαk, the magnitude of the integers in J is doubly exponentially bounded in kPk + kDk + kαk by Theorem 20;
thus, the number of bits needed to represent each integer in J is exponential in kPk + kDk + kαk. Furthermore, we have
|J| < |D| + |P 0 |, and |D| + |P 0 | is exponential in kPk + kDk and constant in kαk; thus, J can be guessed in step 2 in
nondeterministic exponential time in kPk + kDk + kαk.
• By Lemma 14, checking that J is a pseudo-model of P ∪D amounts to checking TP 0 ∪D (J) = J. By Lemma C.3, polynomial p exists such that TP 0 ∪D (J) can be computed in deterministic polynomial time in kP 0 k + kDk + kJkp(maxr∈P 0 krk) ,
which, in turn, is bounded by O(2kPk+kDk ) + kDk + O(2(kPk+kDk)p(maxr∈P krk) ). Hence, step 3 requires deterministic
exponential time in kPk + kDk + kαk.
• By Proposition C.6, step 4 amounts to checking {α} v J, which can be done in time polynomial in kJk and kαk, and
hence in time exponential in kPk + kDk + kαk.
D
Proofs for Section 6
For arbitrary value propagation graph GJP = (V, E, µ), a path in GJP is a nonempty sequence π = v1 , . . . , vn of nodes from V
such that hvi , vi+1 i ∈ E holds for each 0 ≤ i < n; such π starts in v1 and ends in vn . We define |π| = n; moreover, by a slight
abuse of notation, we sometimes write π ∩ X = ∅ or vi ∈ π, where we identify π with the set of its nodes. A path π is simple
if all of its nodes are pair-wise distinct. A path π is a cycle if vn = v1 .
Definition D.1. Given a semi-ground, limit linear program P, a pseudo-interpretation J, value propagation graph
GJP = (V, E, µ), and a path π = v1 , . . . , vn in GJP , the weight µ(π) of π is defined as
µ(π) =
n−1
X
µ(hvi , vi+1 i).
i=1
Lemma D.2. Let P be a semi-ground and stable limit-linear program, let J be a pseudo-model of P, let GJP = (V, E, µ), and
let vAa , vBb ∈ V be nodes such that vAb is reachable from vBa by a path π. Then, for each k ∈ Z such that J |= B(b, k),
• J |= A(a, k + µ(π)) if A and B are both max predicates;
• J |= A(a, −k − µ(π)) if A is a min predicate and B is a max predicate;
• J |= A(a, k − µ(π)) if A and B are both min predicates; and
• J |= A(a, −k + µ(π)) if A is a max predicate and B is a min predicate.
Proof. We consider the case when A and B are both max predicates; the remaining cases are analogous. We proceed by
induction on the length of π. The base case (π is empty) is immediate. For the inductive step, assume that π = π 0 , vAa where
π 0 is a path starting at vBb and ending in node vCc . Then, there exists an edge e = hvCc , vAa i ∈ E, and e is produced by a rule
r = C(c, n) ∧ ϕ → A(a, s) ∈ P such that n is a variable occurring in s, and σ is a grounding of r such that
1. J |= (C(c, n) ∧ ϕ)σ, and
2. δre (J) = µ(e) = µ(hvCc , vAa i).
We next consider the case when C is a max predicate; the case when C is a min predicate is analogous. Let ` be such that
C(c, `) ∈ J. We have the following possibilities.
• If opt(r, J) and ` are both integers (i.e., they are not ∞), we have µ(e) = opt(r, J) − `.
• If opt(r, J) = ∞, then µ(e) = ∞ by Definition 23.
• If ` = ∞, then µ(e) = ∞ by Definition 24 and the fact that P is stable, and moreover opt(r, J) = ∞ by Definition 23.
Now for an arbitrary k ∈ Z such that J |= B(b, k), we consider the following two cases.
• µ(e) 6= ∞. The inductive hypothesis holds for π 0 , so J |= C(c, k + µ(π 0 )) and thus k + µ(π 0 ) ≤ `. Consequently,
we have µ(π) = µ(π 0 ) + µ(e) = µ(π 0 ) + opt(r, J) − ` ≤ opt(r, J) − k, and so k + µ(π) ≤ opt(r, J) holds. Moreover,
J |= P implies TP (J) = J by Lemma 14; thus, Proposition C.6 and the definition of TP imply J |= A(a, k + µ(π)).
• µ(e) = ∞. Clearly, µ(π) = ∞. Moreover, J |= P implies TP (J) = J by Lemma 14; thus, opt(r, J) = ∞, Proposition C.6, and the definition of TP imply J |= A(a, ∞).
Lemma 26. For each semi-ground stable program P, each pseudo-interpretation J with J v T∞
P , and each node vAa on a
positive-weight cycle in GJP , we have A(a, ∞) ∈ T∞
.
P
0
J
0
0
0
Proof. Let GJP = (V, E, µ), let J 0 = T∞
P , and let GP = (V , E , µ ). Now assume for the sake of a contradiction that there
J
exist a cycle π in GP and a node vAa ∈ π such that µ(π) > 0 and J 0 6|= A(a, ∞). Rule applicability is monotonic w.r.t. v, so
0
π is still a cycle in GJP , and, since P is stable, we have µ0 (π) ≥ µ(π) > 0. We consider the case when A is a max predicate;
the remaining case is analogous. Now vAa ∈ V ⊆ V 0 implies that A(a, k) ∈ J 0 for some k; moreover, J 0 6|= A(a, ∞) implies
k 6= ∞. But then, Lemma D.2 implies J 0 |= A(a, k + µ0 (π)); moreover, µ0 (π) > 0 implies that k + µ0 (π) is either ∞ or it is
an integer larger than k; either way, this contradicts our assumption that A(a, k) ∈ J 0 .
Lemma 27. When applied to a semi-ground stable program P, Algorithm 1 terminates after at most 8|P|6 iterations of the
loop in lines 2–8.
Proof. For J a pseudo-interpretation, A an (n + 1)-ary limit predicate, and a an n-tuple of objects such such that A(a, `) ∈ J,
let val(J, Aa) = ` if ` = ∞ or A is a max predicate and ` ∈ Z, and val(J, Aa) = −` if A is a min predicate and ` ∈ Z;
moreover, let R(J, Aa) be the set containing each rule r ∈ P that is applicable to J and where h(r) is of the form A(a, s).
By monotonicity of Datalog Z , we have R(J, Aa) ⊆ R(J 0 , Aa) for each J and J 0 such that J v J 0 . Moreover, for each edge
e = hvBb , vAa i ∈ E generated by a rule r ∈ R(J, Aa), Definition 23 ensures that the following property holds:
val(J, Bb) + µ(e) ≥ val(T{r} (J), Aa)
(∗)
To prove this lemma, we first show the following auxiliary claim.
Claim (♦). For each n ≥ 0 determining the pseudo-interpretation J = TnP and the value propagation graph GJP = (V, E, µ),
0
0
each n0 ≥ 1 determining the pseudo-interpretation J 0 = Tn+n
and the value propagation graph GJP = (V 0 , E 0 , µ0 ), each set
P
of nodes X ⊆ V, and each node vAa ∈ V of such that
1. E = E 0 ;
0
2. val(J 0 , Bb) = ∞ holds for each node vBb ∈ V 0 that occurs in GJP in a positive-weight cycle;
3. vAa 6∈ X;
4. val(J 0 , Aa) < val(TP (J 0 ), Aa);
5. |π| ≤ n0 holds for each simple path π in GJP that ends in vAa and satisfies π ∩ X = ∅;
6. for each node vBb ∈ X, there exists a path π in GJP that starts in vAa and ends in vBb ;
one of the following holds:
(i) val(J 0 , Cc) + µ0 (π) ≥ val(TP (J 0 ), Aa) for some node vCc ∈ X and path π in GJP starting in vCc and ending in vAa ;
(ii) R(J, Cc) ( R(J 0 , Cc) for some node vCc ∈ V.
Proof. For arbitrary n, we prove the claim by induction on n0 . For the base case n0 = 1, consider an arbitrary set X ⊆ V and
vertex vAa ∈ V that satisfy properties 1–6 of (♦). We distinguish two cases.
• There exists an edge e = hvBb , vAa i ∈ E such that val(J 0 , Bb) + µ0 (e) = val(TP (J 0 ), Aa). Now either vBb ∈ X or
vBb = vAa holds: if that were not the case, path π = vBa , vAa would be a simple path in GJP such that |π| = 2, which
would contradict property 5.
We next show that vBb = vAa is impossible. For the sake of a contradiction, assume that vBb = vAa holds, and thus we
have val(J 0 , Aa) + µ0 (e) = val(TP (J 0 ), Aa). By property 4, this implies val(J 0 , Aa) + µ0 (e) > val(J 0 , Aa), and hence
0
µ0 (e) > 0. Consequently, path π is a positive-weight cycle in GJP , and so, by property 2, we have val(J, Aa) = ∞, which,
in turn, contradicts property 4.
Consequently, we have vBb ∈ X. But then, since, by assumption, val(J 0 , Bb) + µ0 (e) = val(TP (J 0 ), Aa), part (i) of the
claim holds for vCc = vBb .
• For each edge hvBb , vAa i ∈ E we have val(J 0 , Bb) + µ0 (e) < val(TP (J 0 ), Aa). Then, for each rule r ∈ P that generates
an edge e = hvBb , vAa i, property (∗) ensures val(T{r} (J 0 ), Aa) < val(TP (J 0 ), Aa). Since
val(TP (J 0 ), Aa) =
max
r∈P,h(r)=A(a,s)
val(T{r} (J 0 ), Aa),
a rule r ∈ P exists that satisfies val(TP (J 0 ), Aa) = val(T{r} (J 0 ), Aa) but does not generate an edge in E ending
in vAa . Clearly, h(r) is of the form A(a, s) and r is applicable to J 0 , so r ∈ R(J 0 , Aa) holds. Moreover, r is
semi-ground; hence, if s were to contain a variable, this variable would occur in a limit body atom of r, and so r
would generate an edge in E; consequently, s is ground. Finally, if r were applicable to J, then {A(a, s)} v J 0
and so val(J 0 , Aa) ≥ val({A(a, s)}, Aa) = val(TP (J 0 ), Aa), which contradicts property 4. Consequently, we have
r∈
/ R(J, Aa), and so part (ii) of the claim holds for vCc = vAa .
For the inductive step, we assume that (♦) holds for n0 − 1 ≥ 1, each set X ⊆ V, and each node vAa ∈ V; and we consider
an arbitrary set X ⊆ V and vertex vAa ∈ V that satisfy properties 1–6 of (♦). By property 4, there exists a rule r ∈ P such
that val(J 0 , Aa) < val(TP (J 0 ), Aa) = val(T{r} (J 0 ), Aa). Now if r does not generate an edge in E, then in exactly the same
way as in the base case we conclude that part (ii) of claim (♦) holds for vCc = vAa ; consequently, in the rest of this proof
0
00
−1
we assume that r generates at least one edge in E. Let J 00 = Tn+n
and let GJP = (V 00 , E 00 , µ00 ). Then, E = E 00 = E 0
P
by property 1, and val(T{r} (J 00 ), Aa) ≤ val(J 0 , Aa) < val(T{r} (J 0 ), Aa), so there exists an edge e = hvBb , vAa i ∈ E such
that val(J 00 , Bb) < val(J 0 , Bb). Furthermore, since val(J 0 , Aa) < val(T{r} (J 0 ), Aa), if vBa were equal to vAa , then path
vAa , vAa would be a positive-weight cycle containing vAa , which contradicts property 2; hence, we have vBb 6= vAa and so
path vBb , vAa is simple. Now if vBb ∈ X holds, then, since r generates e and and val(T{r} (J 0 ), Aa) = val(TP (J 0 ), Aa), by
property (∗), we have val(J 0 , Bb) + µ0 (e) ≥ val(TP (J 0 ), Aa)—that is, part (i) of the claim holds for vCc = vBb . Therefore,
in the rest of this proof we assume vBb ∈
/ X. We now distinguish two cases.
J
• vBb is reachable from vAa in GP . We next show that the set X ∪ {vAa } and node vBb satisfy properties 5 and 6 of the
inductive hypothesis for n0 − 1.
For property 5, note that, since vBb is the direct predecessor of vAa in GJP , each simple path π in GJP that ends in vBb and
does not involve vAa can be extended to the simple path π, vAa that ends in vAa . Thus, we have
max{|π| | π is a simple path in GJP ending in vBb and π ∩ (X ∪ {vAa }) = ∅} <
max{|π| | π is a simple path in GJP ending in vAa and π ∩ X = ∅}.
Property 5 for X, vAa , and n0 ensures
max{|π| | π is a simple path in GJP ending in vAa and π ∩ X = ∅} ≤ n0 ,
which in turn implies
max{|π| | π is a simple path in GJP ending in vBb and π ∩ (X ∪ {vAa }) = ∅} ≤ n0 − 1.
Property 6 holds for X and vAa ; moreover, there exists a path from vBb to vAa via the edge e, so the property also holds
for the set X ∪ {vAa } and node vBb .
Property 3 (vBb ∈
/ X ∪ {vAa }) and property 4 (val(J 00 , Bb) < val(J 0 , Bb)) have already been established for X ∪ {vAa },
vBb , and n0 − 1; moreover, properties 1 and 2 do not depend on X, vAa , and n0 . Thus, we can apply the inductive
hypothesis and conclude that one of the following holds:
(i) val(J 00 , Cc) + µ00 (π) ≥ val(J 0 , Bb) holds for some node vCc ∈ X ∪ {vAa } and path π in GJP that starts in vCc and
ends in vBb ;
(ii) R(J, Cc) ( R(J 00 , Cc) holds for some node vCc ∈ V.
If (ii) is true, then case (ii) of claim (♦) holds since R(J 00 , Cc) ⊆ R(J 0 , Cc). Thus, we next assume that case (i) holds,
and we show that then part (i) of claim (♦) holds for vCc , X, and vAa .
We first show that vCc 6= vAa . For contradiction, assume vCc = vAa . Then val(J 00 , Aa) + µ00 (π) ≥ val(J 0 , Bb). Moreover, since r generates e, by property (∗) and property 4, we have
val(J 0 , Bb) + µ0 (e) ≥ val(T{r} (J 0 ), Aa) = val(TP (J 0 ), Aa) > val(J 0 , Aa).
Consequently, val(J 00 , Aa) + µ00 (π) + µ0 (e) > val(J 0 , Aa). Moreover, val(J 0 , Aa) ≥ val(J 00 , Aa) holds since TP is
monotonic, and µ0 (π) > µ00 (π) holds since P is stable. By these observations, we have
val(J 0 , Aa) + µ0 (π) + µ0 (e) > val(J 0 , Aa);
0
that is, µ0 (π) + µ0 (e) > 0. But then π, vAa is a positive-weight cycle in GJP , and so we have val(J 0 , Aa) = ∞, which
contradicts property 4.
Thus, we have vCc ∈ X. Then, from
val(J 00 , Cc) + µ00 (π) ≥ val(J 0 , Bb) and
val(J 0 , Bb) + µ0 (e) ≥ val(T{r} (J 0 ), Aa) = val(TP (J 0 ), Aa) we conclude
val(J 0 , Cc) + µ0 (π) + µ0 (e) ≥ val(TP (J 0 ), Aa)
as in the case for vCc = vAa . Since µ0 (π) + µ0 (e) = µ0 (π, vAa ), part (i) of claim (♦) holds for vCc , X, and vAa .
• vBb is not reachable from vAa in GJP . Then, by property 6, vBb is not reachable in GJP from any node in X; otherwise,
vBb would also be reachable in GJP from vAa via some node in X. Thus, no simple path in GJP ending in vBb involves
vAa or a node in X—that is, each such path can be extended to a simple path ending in vAa . Now property 5 ensures
max{|π| | π is a path in GJP ending in vAa and π ∩ X = ∅} ≤ n0 , which implies
max{|π| | π is a path in GJP ending in vBb } ≤ n0 − 1.
Thus, property 5 of the inductive hypothesis for n0 − 1 holds for the set ∅ and node vBb . Moreover, property 6 holds
vacuously for ∅, properties 3 and 4 have already been established for vBb , and properties 1 and 2 hold by assumption.
Thus, we can apply the inductive hypothesis for n0 − 1 to ∅ and vBb , and so one of the following holds:
(i) val(J 00 , Cc) + µ00 (π) ≥ val(J 0 , Bb) for some node vCc ∈ ∅ and path π in GJP that starts in vCc and ends in vBb ;
(ii) R(J, Cc) ( R(J 00 , Cc) for some node vCc in GJP .
Clearly, (i) is trivially false, so (ii) holds. But then, case (ii) of claim (♦) holds since R(J 00 , Cc) ⊆ R(J 0 , Cc).
Tn
Tn
Note that, for each n ≥ 0 and each simple path π in GP P , |π| is bounded by the number of nodes in GP P , which is in turn
n+m+1
bounded by m = |P|. Therefore, claim (♦) for n0 = m and X = ∅ ensures that, for each n ≥ 0 such that Tn+m
6= TP
,
P
one of the following holds:
Tn+m
1. val(Tn+m
, Cc) 6= ∞ for some node vCc that occurs in GP P in a positive weight cycle (so the value of the fact correP
sponding to vCc is set to ∞ in the next iteration of the main loop of the algorithm),
Tn+m
2. GP P
Tn
contains at least one edge that does not occur in GP P , or
Tn
3. R(TnP , Cc) ( R(Tn+m
, Cc) for some node vCc in GP P .
P
Tn
For each n ≥ 0, the size of the set R(TnP , Cc) for each node vCc and the number of nodes in GP P are both bounded by
Tn
m, and the number of edges in GP P is bounded by m2 . Thus, the number of iterations of the main loop is bounded by
m · (m + 1) · (m2 + 1) · (m2 + 1) ≤ 8m6 , where the first factor is given by Claim (♦), the second factor comes from the first
case above, the third factor comes from second case, and the fourth factor comes from the third case . Hence, Algorithm 1
reaches a fixpoint after at most 8m6 iterations of the main loop.
Theorem 28. For P a semi-ground stable program, D a dataset, and α a fact, Algorithm 1 decides P ∪ D |= α in time
polynomial in kP ∪ Dk and exponential in maxr∈P krk.
Proof. Partial correctness follows by Lemma 26, while termination follows by Lemma 27. Moreover, the number of iterations
of the main loop of Algorithm 1 is polynomially bounded in |P ∪ D|, and hence kJk in each such iteration is bounded by
kP ∪Dk. Consequently, lines 7 and 4 of Algorithm 1 require time that is worst-case exponential in maxr∈P krk and polynomial
in kP ∪ Dk by Lemma C.3. Moreover, lines 3, 6, and the check in line 8 require time polynomial in kJk and hence in kP ∪ Dk.
Finally, we argue that the check for positive-weight cycles in line 5 is feasible in time polynomial in kP ∪ Dk. Let G0 be
the graph obtained from GJP by negating all weights. Then, a maximal-weight path from vv1 to vv2 in GJP corresponds to the
least-weight path from vv1 to vv2 in G0 . Thus, detecting whether a node occurs in GJP in at least one positive-weight cycle
reduces to detecting whether the node occurs in G0 on a negative cycle (i.e., on a cycle with a negative sum of weights), which
can be solved in polynomial time using, for example, a variant of the Floyd-Warshall algorithm [Hougardy, 2010].
Theorem 29. For P a stable program and α a fact, checking P |= α is E XP T IME-complete in combined and PT IME-complete
in data complexity.
Proof. The E XP T IME lower bound in combined complexity and the PT IME lower bound in data complexity are inherited from
plain Datalog [Dantsin et al., 2001]. The PT IME upper bound in data is immediate by Theorem 28. For the E XP T IME upper
bound in combined complexity, note that, for P 0 = P00 ∪D the semi-grounding of P = P0 ∪D over constants in P, we have that
kP 0 k is exponentially bounded in kPk, whereas maxr∈P00 krk = maxr∈P0 krk. Hence, by Theorem 28, running Algorithm 1
on P 0 gives us an exponential-time decision procedure for P |= α.
Proposition 30. Checking stability of a limit-linear program P is undecidable.
Proof. We present a reduction from Hilbert’s tenth problem, which is to determine whether, given a polynomial P (x1 , . . . , xn )
over variables x1 , . . . , xn , equation P (x1 , . . . , xn ) = 0 has integer solutions. For each such polynomial P , we can assume
P
Qn
k
without loss of generality that P is of the form j (cj × i=1 xi j,i ) for cj ∈ Z and kj,i ≥ 0. Now let PP be the program
containing the following rule, where B is a unary max predicate, and A1 , . . . , An are distinct unary ordinary numeric predicates:
A1 (x1 ) ∧ · · · ∧ An (xn ) ∧ (P (x1 , . . . , xn ) ≤ 0) ∧ (P (x1 , . . . , xn ) ≥ 0) ∧ B(m) ∧ (m ≤ 0) → B(m + 1)
(68)
Note that rule (68) is limit-linear since variables x1 , . . . , xn do not occurs in a limit atom in the rule. We show that PP is stable
if and only if P (x1 , . . . , xn ) = 0 has no integer solutions.
Assume that P (x1 , . . . , xn ) = 0 has no integer solutions. Then, for each grounding σ of (68), at least one of the first two
comparison atoms in the rule is not satisfied, and so PP is trivially stable since, for each pseudo-interpretation J, the value
propagation graph GJP does not contain any edges.
Assume that substitution σ exists such that P (σ(x1 ), . . . , σ(xn )) = 0 holds, and let J1 and J2 be the following pseudointerpretations with the corresponding value propagation graphs:
J1 = {A1 (σ(x1 )), . . . , A1 (σ(xn ))} ∪ {B(0)}
GJP1P = (V1 , E1 , µ1 )
(69)
J2 = {A1 (σ(x1 )), . . . , A1 (σ(xn ))} ∪ {B(1)}
GJP2P
(70)
= (V2 , E2 , µ2 )
Then, we clearly have J1 v J2 and e = hvB , vB i ∈ E1 = E2 ; however, µ1 (e) = 1 and µ2 (e) = 0. Consequently, program PP
is not stable.
Lemma D.3. For each pseudo-interpretation J and each semi-ground type-consistent rule r, if r is applicable to J and it
contains a limit body atom B(b, n) ∈ sb(r) such that B(b, ∞) ∈ J and variable n occurs in h(r), then opt(r, J) = ∞.
Proof. Consider an arbitrary pseudo-interpretation J and rule r applicable to J that contains a limit body atom B(b, n) ∈ sb(r)
with B(b, ∞) ∈ J and n occurring in h(r). We consider the case when h(r) = A(a, s) for A a max predicate and variable
n occurs in s with a negative coefficient; the cases when A is a min predicate and/or n occurs in s with a positive coefficient
are analogous. Then, term s has the form ` × n + t for some negative integer ` and term t not containing n. Moreover, r is
type-consistent, so B is a min predicate. Since r is applicable to J, conjunction C(r, J) has a solution σ. We next show that
opt(r, J) = ∞ holds, for which it suffices to argue that, for each k ∈ Z, conjunction C(r, J) has a solution σ 0 such that sσ 0 ≥ k.
Let k0 = sσ and let σ 0 be the grounding of C(r, J) such that σ 0 (n) = σ(n) − |k − k0 | and σ 0 (m) = σ(m) for each variable
m 6= n. Since B(b, ∞) ∈ J, we have J |= B(b, σ 0 (n)). Moreover, r is type-consistent, σ satisfies all comparison atoms in
the body of r, B is a min predicate, and σ 0 (n) ≤ σ(n), so σ 0 also satisfies all comparison atoms in the body of r. Hence, σ 0 is
a solution to C(r, J). Then, the following calculation implies the claim of this lemma:
sσ 0 = ` × (σ(n) − |k − k0 |) + tσ = (` × σ(n) + tσ) − ` × |k − k0 | = k0 − ` × |k − k0 | ≥ k0 + |k − k0 | ≥ k.
Lemma D.4. For each type-consistent, limit-linear rule r with h(r) = A(t, s), each limit body atom B(t0 , n) ∈ sb(r) such that
n occurs in s, each semi-grounding σ of r such that n 6∈ dom(σ), and all pseudo-interpretations J1 and J2 such that J1 v J2
e
e
and rσ is applicable to J1 , we have δrσ
(J1 ) ≤ δrσ
(J2 ) where e = hvBt0 σ , vAtσ i.
Proof. Consider arbitrary r, B(t0 , n), σ, J1 , J2 , and e as stated in the lemma. We consider the case when A is a max and B
is a min predicate; the remaining cases are analogous. Let ϕ be the body of r and let σ be a semi-grounding of r. Moreover,
e
e
(J2 ) 6= ∞. Due to J1 v J2 , each solution to C(rσ, J1 ) is a solution
(J2 ) = ∞, so we next assume δrσ
the claim is trivial if δrσ
e
(J1 ) 6= ∞ holds. By Definition 23, we then have opt(rσ, J1 ) 6= ∞ and opt(rσ, J2 ) 6= ∞.
to C(rσ, J2 ) as well, so therefore δrσ
But then, by Lemma D.3, there exist k1 , k2 ∈ Z such that B(t0 σ, k1 ) ∈ J1 and B(t0 σ, k2 ) ∈ J2 ; since B is a min predicate,
e
e
(J2 ) = opt(r, J2 ) + k2 . Rule r is
(J1 ) = opt(r, J1 ) + k1 and δrσ
we have k2 ≤ k1 ; moreover, by Definition 23 we have δrσ
type-consistent, so variable n occurs negatively in s; thus, sσ is of the form s0 × n + s00 where s0 is a ground product evaluating
to a negative integer and s00 does not mention n. Moreover, opt(rσ, J1 ) 6= ∞, so there exists a grounding σ1 of rσ such that
e
J1 |= ϕσσ1 and opt(rσ, J1 ) = sσσ1 = s0 × k1 + s00 σ1 = δrσ
(J1 ) − k1 . Let σ2 be the substitution such that σ2 (n) = k2 and
σ2 (m) = σ1 (m) for m 6= n. Clearly, J2 satisfies all object and numeric atoms in ϕσσ2 . Then, we have the following:
e
δrσ
(J2 ) = opt(rσ, J2 ) + k2 ≥ sσσ2 + k2 = s0 × nσ2 + s00 σ2 + k2 = s0 × k2 + s00 σ1 + k2
e
Furthermore, we have already established s0 × k1 + s00 σ1 = δrσ
(J1 ) − k1 , which implies the following:
e
e
e
δrσ
(J2 ) ≥ s0 × k2 − s0 × k1 + δrσ
(J1 ) − k1 + k2 = (s0 + 1) × (k2 − k1 ) + δrσ
(J1 )
e
e
But then, s0 < 0 and k2 ≤ k1 clearly imply δrσ
(J2 ) ≥ δrσ
(J1 ), as required.
Proposition 32. Each type-consistent limit-linear program is stable.
Proof. For P a type-consistent program and σ a semi-grounding of P, condition 1 of Definition 24 follows by Lemma D.4, and
condition 2 of Definition 24 follows by Lemma D.3.
Proposition 33. Checking whether a limit-linear program is type-consistent can be accomplished in L OG S PACE.
Proof. Let P be a limit-linear program. We can check whether P is type-consistent by considering each rule r ∈ P independently. Note that the first type consistency condition is satisfied for every semi-ground limit-linear rule where all numeric terms
are simplified as much as possible; thus, no semi-grounding of r (with constants from P) where all numeric terms are simplified as much as possible can violate the first condition of Definition 31. Thus, it suffices to check whether a semi-grounding
of r (with constants from P) can violate the second or the third condition. In both cases, it suffices to consider at most one
atom α at a time (a limit head atom A(a, s) for the second condition or a comparison atom s1 < s2 or s1 ≤ s2 for the third
condition).PIn α, we consider at most one numeric term s at a time (s ∈ {s1 , s2 } for the third condition), where s is of the
n
form t0 + i=1 ti × mi and ti , for i ≥ 1, are terms constructed from integers, variables not occurring in limit atoms, and
multiplication. Moreover, for each such s, we consider each variable m occurring in s.
By assumption, m occurs in s, so we have mi = m for some i. For the second condition of Definition 31, we need to check
that, if the limit body atom B(s, mi ) introducing mi has the same (different) type as the head atom, then term ti can only be
grounded to positive (negative) integers or zero. For the third condition, we need to check that, if s = s1 and the limit body atom
B(s, mi ) introducing mi is min (max), then term ti can only be grounded to positive (negative) integers or 0, and dually for the
case s = s2 . Hence, in either case, it suffices to check whether term ti can be semi-grounded so that it evaluates to a positive
integer, a negative integer, or zero. We next discuss how this can be checked in logarithmic space. Let ti = t1i × · · · × tki ,
where each tji is an integer or a variable not occurring in a limit atom, and assume without loss of generality that we want to
check whether ti can be grounded to a positive integer; this is the case if and only if one of the following holds:
• all tji are integers whose product is positive;
• the product of all integers in ti is positive and P contains a positive integer;
• the product of all integers in ti is positive, P contains a negative integer, and the total number of variable occurrences in ti
is even;
• the product of all integers in ti is negative, P contains a negative integer, and the total number of variable occurrences in
ti is odd; or
• the product of all integers in ti is negative, P contains both positive and negative integers, and some variable tji has an odd
number of occurrences in ti .
Each of these conditions can be verified using a constant number of pointers into P and binary variables. This clearly requires
logarithmic space, and it implies our claim.
| 2cs.AI
|
arXiv:1702.02362v2 [cs.IT] 22 May 2017
The Necessity of Scheduling in
Compute-and-Forward
Ori Shmuel
Asaf Cohen
Omer Gurewitz
Department of Communication
System Engineering
Ben-Gurion University of the Negev
Email: [email protected]
Department of Communication
System Engineering
Ben-Gurion University of the Negev
Email: [email protected]
Department of Communication
System Engineering
Ben-Gurion University of the Negev
Email: [email protected]
Abstract—Compute and Forward (CF) is a promising relaying scheme which, instead of decoding single messages or
forwarding/amplifying information at the relay, decodes linear
combinations of the simultaneously transmitted messages. The
current literature includes several coding schemes and results
on the degrees of freedom in CF, yet for systems with a fixed
number of transmitters and receivers. It is unclear, however, how
CF behaves at the limit of a large number of transmitters.
In this paper, we investigate the performance of CF in that
regime. Specifically, we show that as the number of transmitters
grows, CF becomes degenerated, in the sense that a relay prefers
to decode only one (strongest) user instead of any other linear
combination of the transmitted codewords, treating the other
users as noise. Moreover, the sum-rate tends to zero as well. This
makes scheduling necessary in order to maintain the superior
abilities CF provides. Indeed, under scheduling, we show that
non-trivial linear combinations are chosen, and the sum-rate does
not decay, even without state information at the transmitters and
without interference alignment.
I. I NTRODUCTION
Compute and Forward (CF) [1] is a coding scheme which
enables receivers to decode linear combinations of transmitted
messages, exploiting the broadcast nature of wireless relay
networks. CF utilizes the shared medium and the fact that
a receiver, which received multiple transmissions simultaneously, can treat them as a superposition of signals, and decode
linear combinations of the received messages. Specifically,
together with the use of lattice coding, the obtained signal,
after decoding, can be considered as a linear combination
of the transmitted messages. This is due to an important
characteristic of lattice codes - every linear combination of
codewords is a codeword itself.
However, since the wireless channel suffers from fading,
the received signals are attenuated by real (and not integers)
attenuations factors, hence the received linear combination is
”noisy”. The receiver (e.g., a relay) then seeks a set of integer
coefficients, denoted by a vector a, to be as close as possible1
to the true channel coefficients.
This problem was elegantly associated with Diophantine
Approximation Theory in [2], and was compared to a similar
problem, that of finding a co-linear vector for the true channel
1 One can define different criteria for the goodness of the approximation,
for example, the minimum distance between the vectors elements.
coefficients vector (between the receiver and the transmitters).
In addition, the co-linear vector must be an integer valued
vector, due to the fact that it should represent the coefficients
of an integer linear combination of codewords. Based on this
theory, if one wishes to find an integer vector a that is close (in
terms of co-linearity) to a real vector h, then one must increase
||a|| in order to have a small approximation error between
them. The increase in the norm value leads to a significant
penalty in the achievable rate at the receiver and thus results
in a tradeoff between the goodness of the approximation and
the maximization of the rate.
The CF scheme was extended in many directions, such
as MIMO CF [3], linear receivers (Integer Forcing) [4] [5],
integration with interference alignment [2], scheduling [6] and
more [7], [8]. All the mentioned works considered a general
setting, where the number of transmitters is a parameter for
the system and all transmitters are active at all times. That
is, the receiver is able to decode a linear combination of
signals from a large number of transmitters as long as the
transmitters comply with the achievable rates at the receiver,
and still promise, to some extent, an acceptable performance.
However, in this work, we show that the number of simultaneous transmitters is of great importance when the number of
relays is fixed. In fact, this number cannot be considered solely
as a parameter but as a restriction since, when it grows, the
receiver will prefer to decode only the strongest user over all
possible linear combinations. This will make the CF scheme
degenerated, in the sense that the relay chooses a vector a
which is actually a unit vector (a line in the identity matrix),
thus treating all other signals as noise. In other words, the
linear combination chosen is trivial. Furthermore, we show
that as the number of transmitters grows, the scheme’s sumrate goes to zero as well. Thus, one is forced to use users
scheduling to maintain the superior abilities CF provide.
We conclude this paper with an optimistic view, that user
scheduling can improve the CF gain. We believe that this
can be done by suitable matching of linear combinations, i.e.
coding possibilities. Using simple Round Robin scheduling
and results for CF in fixed size systems, we lower bound the
sum-rate. We thus show that even for a simple scheduling
policy the system sum-rate does not decay to zero.
The paper is organized as follows. In Section II, the system
𝒘𝟏
𝒘𝟐
𝒘𝑳
𝜀"
𝜀#
.
.
.
𝜀$
𝒛𝟏 𝒚
𝟏
𝒙𝟏
𝒙𝟐
𝒛𝟐
𝐻
𝒛𝑴
𝒙𝑳
𝒚𝟐
𝒚𝑴
𝒟"
.𝟏
𝒖
𝒟#
.𝟐
𝒖
.
.
.
𝒟&
In CF each relay decodes a linear combination um of the
original messages, and forward it to the destination. With
enough linear combinations, the destination is able to recover
the desired (original) messages from all sources.
The main results in CF are the following.
.𝑴
𝒖
Fig. 1: Compute and Forward system model. L transmitters
communicate through a shared medium to M relays.
model is described. In Section III, we derive an analytical
expression for the probability of choosing a unit vector by the
relay, as the number of users grows. Section IV depicts the
behaviour of the sum-rate for this model, and in Section V
we present the advantage of using scheduling, along with a
simple scheduling algorithm.
II. S YSTEM MODEL AND K NOWN RESULTS
Consider a multi-user multi-relay network, where L transmitters are communicating to a single destination D via M
relays. The model is illustrated in figure 1. All relays form
a layer between the transmitters and the destination such that
each transmitter can communicate with all the relays. Each
transmitter draws a length-k message with equal probability
over a prime size finite field, wl ∈ Fkp , l = 1, 2, ..., L, where
Fp denotes the finite field with a set of p elements.
This message is then forwarded to the transmitter’s encoder,
El : Fkp → Rn , which maps length-k messages over the finite
field to length-n real-valued codewords, xl = El (wl ). Each
codeword is subject to a power constraint, kxl k2 ≤ nP . The
message rate of each transmitter is defined as the length of
the message measured in bits normalized by the number of
channel uses, that is, R = nk log p, which is equal2 for each
transmitter. Each transmitter then broadcasts it’s codeword to
the channel.
Hence, each relay m ∈ {1...M } observes a noisy linear
combination of the transmitted signals through the channel,
ym =
L
X
hml xl + zm
m = 1, 2, ..., M,
(1)
i=1
where hml ∼ N (0, 1) are the real channel coefficients and
z is an i.i.d., Gaussian noise, z ∼ N (0, In×n ). Let hm =
[hm1 , hm2 , ..., hmL ]T denote the vector of channel coefficients
at relay m. We assume that each relay knows its own channel
vector. After receiving the noisy linear combination, each relay
selects a scale coefficient αm ∈ R, an integer coefficients
vector am = (am1 , am2P
, ..., amL )T ∈ ZL , and attempts to
L
decode the lattice point l=1 aml xl from αm ym .
2 Note
that messages with different length can be allowed with zero padding
to attain a length-k message which will result in different rates for the
transmitters.
Theorem 1 ([1, Theorem 1]). For real-valued AWGN networks
with channel coefficient vectors hm ∈ RL and coefficients
vector am ∈ ZL , the following computation rate region is
achievable:
P
1
+
,
R(hm , am ) = max log
2 + P kα h − a k2
αm ∈R 2
αm
m m
m
(2)
where log+ (x) , max{log(x), 0}.
Theorem 2 ([1, Theorem 2]). The computation rate given in
Theorem 1 is uniquely maximized by choosing αm to be the
MMSE coefficient
αM M SE =
P hTm am
,
1 + P khm k2
which results in a computation rate region of
−1
1
P (hTm am )2
R(hm , am ) = log+ kam k2 −
.
2
1 + P khm k2
(3)
(4)
Note that the above theorems are for real channels and the
rate expressions for the complex channel are twice the above
([1, Theorems 3 and 4]).
Since the relay can decide which linear combination to
decode (i.e., the coefficients vector a), an optimal choice will
be one that maximizes the achievable rate. That is,
−1
1
P (hTm am )2
+
2
aopt
=
arg
max
log
ka
k
−
. (5)
m
m
1 + P khm k2
am ∈ZL \{0} 2
Remark 1 (The coefficients vector). The coefficients vector a
plays a significant role in the CF scheme. It dictates which
linear combination of the transmitted codewords the relay
wishes to decode. That is, each non-zero element signifies
the fact that the relay is interested in it’s corresponding
codeword. If, starting from a certain number of simultaneously
transmitting users, the coefficients vector the relay chooses is
always (or with high probability) a unit vector, this means
that essentially we treat all other users as noise and loose the
promised gain of CF.
The following Lemma bounds the search domain for the
maximization problem in (5).
Lemma 1 ([1, Lemma 1]). For a given channel vector h,
the computation rate R(hm , am ) in Theorem 2 is zero if the
coefficient vector a satisfies
kam k2 ≥ 1 + P khm k2 .
(6)
The problem of finding the optimal a can be done by
exhaustive search for small values of L. However, as L grows,
the problem becomes prohibitively complex quickly. In fact,
it becomes a special case of the lattice reduction problem,
which has been proved to be NP-complete. This can be seen
if we write the maximization problem of (5) as an equivalent
minimization problem [9]:
T
aopt
m = arg min f (am ) = am Gm am ,
(7)
am ∈ZL \{0}
where Gm = (1 + P khm k2 )I − P hm hTm . Gm can be
regarded as the Gram matrix of a certain lattice and am will be
the shortest basis vector and the one which minimize f . This
problem is also known as the shortest lattice vector problem
(SLV), which has known approximation algorithms due to its
hardness [10], [11]. The most notable of them is the LLL
algorithm [12], [13] which has an exponential approximation
factor which grows with the size of the dimension. However,
for special lattices, efficient algorithms exist [14]. In [9],
a polynomial complexity algorithm was introduced for the
special case of finding the best coefficient vector in CF.
III. P ROBABILITY OF A U NIT V ECTOR
In this section, we examine the coefficient vector at a single
relay, hence, we omit the index m in the expressions.
(a) L = 4
(b) L = 8
(c) L = 16
(d) L = 32
Fig. 2: Example for the magnitude of the elements of G for
different dimensions (i.e., different values of L), for P=10.
The graphs depict a single realization for each L, and were
interpolated for ease of visualization.
A. The Matrix G
Examining the matrix G, one can notice that as L, the
number of transmitters, grows, the diagonal elements grow
very fast relatively to the off-diagonal elements. Specifically,
each diagonal element is a random variable, which is a χ2L
r.v. minus a multiplication of two Gaussian r.vs., whereas
the off-diagonal elements are only a multiplication of two
Gaussian r.vs.. Of course, as L grows, the former has much
higher expectation value compared to later. Examples of G
are presented in Figure 2, for different dimensions. It is clear
that even for moderate number of transmitters, the differences
in values between the diagonal and off-diagonal elements are
significant.
Consider now the quadric form (7) we wish to minimize.
Any choice of a that is not a unit vector, will add more than
one element from the diagonal of G to it. When L is large, the
off-diagonal elements have little effect on the function value
compared to the diagonal elements. Therefore, intuitively, one
would prefer to have as little as possible elements from the
diagonal although the off-diagonal elements can reduce the
function value. This will happen if we choose a to be a unit
vector. In the reminder of this section, we make this argument
formal.
B. Minimization of the Quadratic form f
Note that the right term consists of all possible pairs (i, j)
such that i 6= j, a total of L(L−1)
elements.
2
We wish to understand when will a relay prefer a unit vector
over any other non-trivial vector a. Specifically, since a is
a function of the random channel h, we will compute the
probability of having a unit vector as the minimizer of f for
a given a. Or, alternatively, the probability that a certain nontrivial a will minimize f compared to a unit vector. We thus
wish to find the probability
P r(f (a) ≤ f (ei ))
= P r(kak2 + P (kak2 khk2 − (aT h)2 ) ≤ 1 + P (khk2 − h2i )),
(8)
where ei is a unit vector of size L with 1 at the i-th entry
and zero elsewhere, and a is any integer valued vector that is
not a unit vector. Note that (8) refers to any integer vector a,
including the vectors in the search domain such that kak ≤
p
1 + P khk2 ([1, Lemma 1]).
Note also that the right and left hand sides of the inequality
in equation (8) are dependent, hence direct computation of this
probability is not trivial. Still, this probability can be evaluated
exactly noting that the angle between a and h is what mainly
affects it. The details are in the theorem below.
The minimization function f (a) = aT Ga can be written as
C. The Optimality of ei VS. a Certain Vector a
aT Ga =
L
X
(1 − P (khk2 − h2i ))a2i − 2
i=1
= kak2 + P
L X
i−1
X
i=1 j=1
L X
i−1
X
(hi aj − hj ai )2
i=1 j=1
= kak2 + P (kak2 khk2 − (aT h)2 ).
P hi hj ai aj
Theorem 3. Under the CF scheme, the probability that a nontrivial vector a will be the coefficient vector aopt which maximize the achievable rate R(h, aopt ), i.e., minimize f (aopt ),
comparing with a unit vector ei , is upper bounded by
1 L−1
,
(9)
P r(f (a) ≤ f (ei )) ≤ 1 − IΦ(a)
2
2
where Ix (a, b) is the CDF of the Beta distribution with param1
1
eters a and b, and Φ(a) = 1 − kak
2 . Note that 2 ≤ Φ(a) ≤ 1
for any a which is not a unit vector.
Pr
0.4
In the context of this work, the main consequence of
Theorem 3, is the following.
0.3
Corollary 1. As the number of simultaneously transmitting
users grows, the probability that a non-trivial a will be the
maximizer for the achievable rate goes to zero. Specifically,
0.2
P r(f (a) ≤ f (ei )) ≤ e−LE1 (L) ,
The proofs will be given after the following discussion.
1) Discussion and simulation results: Corollary 1 clarifies
that for every P , as the number of users grows, the probability
of having a non-trivial vector a as the maximizer of the
achievable rate is going to 0. Note that the assumption of
L > 3, which arises naturally form this paper’s regime, along
with the fact that kak ≥ 2 grantees that E1 (L) is positive.
Figure 3 depicts the probability in (9), it’s upper bound given
in equation (10) and simulation results. From the analytic
results as well as the simulations on the rate of decay, one can
deduce that even for relatively small values of simultaneously
transmitting users (L > 20), the relay will prefer to choose a
unit vector. Also, one can observe from the results and from
the analytic bound that as the norm of a grows, the rate
of decay increases. This faster decay reflects the increased
penalty of approximating a real vector using an integer valued
vector.
2) Proofs: The proof of Theorem 3 is based on the lemma
below.
T
2
(a h)
Lemma 2. The distribution of kak
2 khk2 , which is the squared
cosine of the angle between an integer vector a and a standard
normal vector h, both of dimension L, is Beta( 12 , L−1
2 ).
Proof. Let Q be an orthogonal rotation matrix such that
Qa = a0 where a0 , is co-linear to the basis vector e1 . That
is, a0 = (kak, 0, ..., 0). Define h0 = Qh. Note that h0 is
a standard normal vector since E[h0 ] = E[Qh] = 0, and
QIQT = QQT = I. We have
(aT h)2
(aT h)2
(aT QT Qh)2
= T
= T T
2
2
T
kak khk
(a a)(h h)
(a Q Qa)(hT QT Qh)
((Qa)T Qh)2
((Qa)T Qh)2
=
=
T
T
((Qa) Qa)((Qh) Qh)
kQak2 kQhk2
2
=
ka0 k2 (e1 T h0 )2
(e1 T h0 )2
h0 1
=
=
.
0
2
2
0
2
2
0
2
ka k ke1 k kh k
ke1 k kh k
kh0 k2
(11)
Considering the above we have the equality of cos θ =
h0 21
. This expression can be represented as WW
+V ,
h021 +h0 22 +...+h02L
PL
2
0
2
02
2
where W = h 1 is a χ1 r.v. and V = i=2 h i is a χL−1 r.v.
independent in W . This ratio has a Beta(a, b) distribution,
with a = 12 and b = L−1
2 . Note that a and b correspond to the
degrees of freedom of W and V .
2
||a||2 =5
||a||=2
Simulation
Beta dist. bound
Min. dist. bound
0.1
(10)
where a is any integer vector that is not a unit vector and
E1 (L) = (1 − L3 ) log kak.
||a||2 =8
0
5
10
15
20
L
Fig. 3: The upper bounds given in (10)(solid lines), (9) (dashed
lines) and simulation results (dotted lines) for not having a unit
vector as the minimizer of f compared to various values of
kak2 as a function of simultaneously transmitting users.
Proof of Theorem 3. According to equation (8) we have,
P r(f (a) ≤ f (ei ))
= P r kak2 + P kak2 khk2 − (aT h)2 ≤ 1 + P (khk2 − h2i )
1 − kak2
= Pr
+ khk2 − h2i − kak2 khk2 + (aT h)2 ≥ 0
P
(a)
≤ P r khk2 − kak2 khk2 + (aT h)2 ≥ 0
(aT h)2
1
−
1
+
≥
0
= Pr
kak2
kak2 khk2
T
2
(a h)
1
= Pr
≥
1
−
kak2 khk2
kak2
(b)
1 L−1
,
,
= 1 − IΦ(a)
2
2
(12)
where (a) follows since we removed negative terms (kak2 >
1
1) and (b) follows from Lemma 2 with Φ(a) = 1 − kak
2.
The bound on the probability given in Theorem 3 consists
of a complicated analytic function IΦ(a) (·). Hence, corollary 1
includes a simplified bound which avoids the use of IΦ(a) (·),
yet keeps the nature of the result in Theorem 3. The proof of
Corollary 1 is based on the following lemma.
T
2
(a h)
Lemma 3. The CDF of kak
2 khk2 can be lower bounded by
the CDF of the minimum of L2 − 1 i.i.d. uniform random
variables in [0, 1].
Proof. We start by assuming that L is even where the case
of odd
L will be dealt with later. From 2Lemma 2, the r.v.
h1
(aT h)2
kak2 khk2 has the same distribution as khk2 , that is, for any
0 ≤ α ≤ 1,
(aT h)2
≤
α
kak2 khk2
2
h1
= Pr
≤
α
khk2
Pr
2
h1 + h22
≥ Pr
≤
α
khk2
h21 + h22
= Pr
≤
α
(h21 + h22 ) + ... + (h2L−1 + h2L )
(a)
L
(b)
= 1 − (1 − α) 2 −1 ,
(a) is true since a larger r.v. will yield lower probability. (b)
h21 +h22
is due to the observation that khk
can be represented as
2
PL
W
2
2
2
i=3 hi are independent
W +V , where W = h1 +h2 and V =
exponential r.vs. Note that V is essentially a sum of L2 − 1
independent
pairs. This ratio is distributed as the minimum
of L2 − 1 i.i.d. uniform [0, 1] random variables [15, Lemma
3.2]. This is since the ratio can be interpreted as the proportion
of the waiting time for the first arrival to the L2 arrival of a
Poisson process.
In case L is an odd number,2 we2 can increase the term in the
h +h
proof by replacing it with khk12 −h22 , resulting in a distribution
L
which is similar to the minimum of ( L−1
2 − 1) i.i.d. uniform
random variables in the same manner.
Proof of Corollary 1.
P r(f (a) ≤ f (ei ))
(a)
(aT h)2
1
≤ Pr
≥
1
−
kak2 khk2
kak2
L −1
b2c
(b)
1
≤ 1− 1−
2
kak
L−1 −1
2
1
≤
kak2
(13)
= e−LE1 (L) ,
where (a) and (b) follows from Lemmas 2 and 3 respectively, E1 (L) = (1 − L3 ) log kak and kak2 > 1.
The following lemma shows a simple property of the
optimal coefficients vector which shows that if the relay is
interested in only one transmitter, i.e. a unit vector as the
optimal coefficients vector, it will be the transmitter with the
strongest channel.
Lemma 4. For any channel vector h of size L with m =
arg maxi {|hi |}, the optimal coefficients vector aopt which
maximize the rate R(h, a) has to satisfy m = arg maxi {|ai |}
as well.
Proof. Suppose that there exist h for which (w.l.o.g.) |h1 | >
|hi | for all i 6= 1 and that the optimal coefficients vector aopt
satisfies (w.l.o.g.) |a2 | > |ai | for all i 6= 2. Considering the rate
expression R(h, a) we will show that by rearranging aopt a
b be a vector which is identical
higher rate can be attain. Let a
to aopt except the two first entries which are switched, i.e.
ab1 = aopt
and ab2 = aopt
1 . The values of the vectors are the
2
same thus we have kb
ak = kaopt k and the only term affecting
the rate is the scalar multiplication between aopt and h. We
first note that the signs of hi and its corresponding optimal
coefficient aopt
has to be equal or different for all i. I.e. the
i
the case which there exist i such that sign(hi ) = sign(aopt
i )
and j such that sign(hj ) 6= sign(aopt
j ) could not be possible.
This is due to the fact that the optimal coefficients vector has
to maximize the scalar multiplication. Therefore, considering
this property, h1 ab1 + h2 ab2 > h1 aopt
+ h2 aopt
1
2 . This means
b contradicting aopt
that the rate can be improved by choosing a
optimality. Specifically, we’ll get that as long as the maximal
value in any a is not in the same place as the maximal value
in h we can always improve the rate.
D. The Optimality of ei VS. All Possible Vectors a
Corollary 1 refers to the probability that a unit vector will
minimize f for a fixed a. Next we wish to explore this
probability for any possible a. For the purpose of clarity, (13)
gives an upper bound on the probability that a unit vector
will not minimize f compered to a certain possible integer
coefficients vectors with a certain kak2 . Where the probability
of having an optimal vector which is not a unit vector will be
the union of all probabilities for each a vector which satisfies
kak2 < 1 + P khk2 . Let us define P (ei ) as the probability
that a relay picked a unit vector as the coefficient vector, and
P (ei ) as the probability which any other vector was chosen.
In [9], a polynomial time algorithm for finding the optimal
coefficients vector a was given. The complexity result derives
from the fact that the cardinality of the set of all a vectors
3
(denoted
p as Φ) which are considered is upper bounded by
2L(d 1 + P khk2 e + 1). That is, any vector which does not
exist in this set has zero probability to be the one which
maximize the rate. We shell note this set here as A. Thus,
we wish to compute
[
P (ei ) =
P (f (a) ≤ f (ei )) ,
(14)
A
L
where A = {a ∈ Z : a ∈ Φ, a 6= ei ∀i}. Note that the
cardinality of A grows with the dimension of h, i.e., with L
and can be easily upper bounded as follows,
|A| ≤ 2L(d
p
1 + P khk2 e + 1)
≤ 2L(d1 + P khk2 e + 1) ≤ 2L(P khk2 + 3). (15)
Theorem 4. Under the CF scheme, the probability which any
other coefficients vector a will be chosen to maximize the
achievable rate R(h, aopt ) compared with a unit vector ei ,
as the number of simultaneously transmitting users grows, is
zero. That is,
lim P (ei ) = 0,
(16)
L→∞
Proof. We have,
lim P (ei ) = lim
L→∞
≤ lim
L→∞
L→∞
X
[
P (f (a) ≤ f (ei ))
A
P (f (a) ≤ f (ei ))
A
X 1
≤ lim
L→∞
kak2
L−1
2 −1
A
3 In [9], Φ is a set of points which the average of any consecutive points
is mapped to a different coefficients vector.
L−1
−1
1 2
≤ lim |A|
L→∞
2
Sum
Rate
7
(a)
L−1
−1
1 2
2
≤ lim 2L(P khk + 3)
L→∞
2
L−1 −1
L
X
1 2
= lim 2L(P
h2i + 3)
L→∞
2
i=1
L−1
L
−1
1 2
1X 2
(c)
h
= lim 2L2 P
L→∞
2
L i=1 i
L−1
−1
1 2
(d)
= lim 2L2 P
L→∞
2
6
(b)
= 4P lim L2 2−
5
4
3
2
1
ææææ
æ
æ
P=103
æ
æ
æ
æ
P=102
àààà
æ
æ à
à
æ
à
à
P=10
à
à ææ
æ
à
ììì
ì
à
ìì
à
ì
à æ
ì
ì
ì
à æ
ì
ì à æ
ì
ì à æ
ìì à
ììààæ
æ
àììì
ææ
ììì
àà
àà
ààììì
ììì
æææ
àààììì
æææ
àààìììì
ææææ
ààààà
ææææ
ìììììì
àààààààààà
ææææææ
ìììììì
æææææ
à
10
20
30
40
Trans.
50 Users
Fig. 4: The sum rate as give in (18) for the case of 4 relays as
a function of the number of simultaneously transmitting users,
for different values of P.
L−1
2
L→∞
= 4P lim L2 e−LE2 (L) = 0,
L→∞
where (a) is true since the term inside the sum is maximized
with kak2 = 2. (b) is due to (15), in (c) we multiplied and
divide with L and eliminate the limit term which is multiplied
by 3 since it goes to zero. (d) follows from the strong law
of large numbers were the normalized sum converge with
probability one to the expected value of χ21 r.v. which is one.
And lastly we define E2 (L) = 12 (1 − L1 ) log 2.
This result implies that the probability of having any non
unit vector as the rate maximizer is decreasing exponentially
to zero as the number of users grows.
IV. C OMPUTE AND F ORWARD S UM -R ATE
In order that relay m will be able to decode a linear
combination with coefficients vector am , all messages’ rates
which are involved in the linear combination must be within
the computation rate region [1]. I.e., all the messages for which
the corresponding entry in the coefficient vector is non zero.
That is,
Rl < min R(hm , am )
aml 6=0
(17)
Hence, the sum rate of the system is defined as the sum of
messages’ rates, i.e.,
L
X
l=1
min R(hm , am ).
m:aml 6=0
(18)
Following the results from previous subsections, we would
like to show that as the number of users grows, the system’s
sum-rate decreases to zero as well. That is, without scheduling
users, not only each individual rate is negligible, this is true
for the sum-rate as well. This will strengthen the necessity to
schedule users in CF.
Theorem 5. As L grows, the sum rate of CF is tends to zero,
that is,
L
X
lim
min R(hm , am ) = 0.
(19)
L→∞
l=1
m:aml 6=0
Proof. The proof outline is as follows. The sum rate expression is divided into two parts, which describe two scenarios.
The first is for the case where a relay chooses a unit vector
as the coefficients vector and the second is for the case where
any other vector is chosen. The probabilities for that are P (ei )
and P (ei ), respectively. Then, we show that each part goes to
zero by upper bounding the corresponding expressions. The
complete proof is given in appendix A.
Simulations for the sum-rate for different values of P can
be found in Figure 4. It is obvious that for large L, the sumrate decreases, hence, for a fixed number of relays there is no
use in scheduling a large number of users, as CF degenerates
to choosing unit vectors and treating other users as noise.
However, the simulations suggests a peak at a small number
of transmitters. We explore this in the next section.
V. S CHEDULING IN C OMPUTE AND F ORWARD
Theorem 4 and 5 suggest that a restriction on the number of
simultaneously transmitting users should be made. That is, in
order to apply the CF scheme for systems with a large number
of sources, scheduling a smaller number of users should take
place.
The most simple scheduling scheme is to schedule users
in a Round Robin (RR) manner, where in each transmission
only k users may transmit simultaneously. The value of k can
be optimized, yet as a thump rule, one can schedule M users
(similar to the number of relays) is each transmission to obtain
a sum-rate which is not going to zero. Figure 5 depicts such a
scenario. In fact, even higher sum rates can be obtained if the
number of scheduled users is higher than the number of relays,
i.e., the number for which the maximal sum-rate in Figure 4
is achieved. Still, it is clearly seen that it is not zero for M
scheduled sources, compared to the zero sum-rate when all L
users transmit and the relay use CF.
In fact, one can use existing results for the CF sum-rate for
the case of equal number of sources and relays, and describe
the sum-rate in each transmission under such a schedule.
Sum
Rate
12
Sum
Rate
5
4
æ
à
ì
æ
à
ì
æ
ò
à
ì
æ
à
ì
æ
à
ì
ò
ò
ò
ò
æ
à
æ
ì
à
ì
ò
ò
3
P=100
P=75
P=50
P=25
No Scheduling
Scheduling
2
1
5
10
æ
à
ì
æ
à
ì
æ
à
ì
æ
à
ì
æ
à
ì
æ
à
ì
10
ò
ò
ò
ò
ò
ò
8
æ
à
ì
Upper
æ
Best Eq.
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
æ
M=4
æ æ æ
æ æ
æ æ
M=8
à à
æ æ
à
à à
à à
æ æ
à à
à à
æ æ
à à
à à
à à
æ æ
à
à à à
à à
æ æ
à
à à à
æ æ
à à
à à
à
à
æ æ
à
à à à
æ
æ à à à à
à
æ
à
æ
à à à
æ
à à
à
à
æ
à à
æ
à à à à
à
à
ò
ò ò ò ò
ò
ò ò ò ò ò
ò ò
ò ò
ò ò
ò ò ò
ò
ò ò
ò ò ò
ò ò
ò ò ò ò ò
ò ò
ò ò
ò ò ò
ò ò ò ò
ò ò
ò ò
ò ò ò ò ò ò ò
Lower
6
4
ò
æ
à
ì
ò
15
æ
à
ì
ò
20
æ
à
ò
ì
25
2
æ
ò
ì
à
ò
æ
ì
à
Trans.
30 Users
5
10
15
20
25
Power
30 @dBD
Fig. 5: Simulation results for the average sum-rate per transmission. Here, the number of relays is 3 and scheduling was
performed in a Round Robin manner, where in every phase 3
sources were scheduled among the transmitting users.
Fig. 6: Simulation results for the average sum-rate compared
with the upper and lower bounds given in (20) and (21),
respectively, as a function of the transmission power for
M = 4 and M = 8.
According to [2], the sum rate for M sources and M relays
is upper bounded by,
any other linear combination of the transmitted signals. Thus,
CF becomes degenerate and it is more preferable to apply
scheduling for much smaller group size. We show that even
with a simple scheduling policy the sum-rate does not goes to
zero and would like, as future work, to proceed and explore
scheduling policies which exploit the decoding procedure of
CF.
M
X
l=1
min R(hm , am ) ≤
m:aml 6=0
1
log P + log log P , (20)
1 + 1/M
for M ≥ 2 and P ≥ 3. A very coarse lower bound can be
attained if the relays are forced to choose their coefficients
vectors such that each relay i chooses ai = ei . That is,
an interference channel where each relay i considers the
interferences from all other sources j, j 6= i as noise. Even
with this one has
M
X
l=1
≥
min
m:aml 6=0
M
X
l=1
Proof. The probabilities P (ei ) and P (ei ) define a partition on
the channel vectors a relays sees, specifically we define,
He = {h ∈ RL | arg min f (a) = ei }
R(hm , am )
min
m:aml 6=0
a∈ZL \{0}
1
log+
2
1 + P khm k2
1 + P (khm k2 − h2l )
L
He = {h ∈ R | arg min f (a) 6= ei }.
j6=l
which is not zero. Simulation results for the bounds and
the optimal CF coefficient vectors are presented in Figure 6.
From the aforesaid one can conclude that scheduling M users
for transmission is worthwhile with respect to the alternative
of permitting all users transmit simultaneously. Of course, the
scheduling policy has great impact on the performance, which
can be increased if, for example, one schedules groups whose
channel vectors are more suitable for CF.
That is, with probability P (ei ) a relay sees a channel vector
h ∈ He and with probability P (ei ) a relay sees a channel
vector h ∈ He . We note he and he as a channel vectors
which belongs to He and He respectively.
Under the above definitions, the sum rate can be written as
follows,
L
X
l=1
=
min R(hm , am )
m:aml 6=0
L
X
l=1
VI. C ONCLUSION AND FUTURE WORK
This work gives evidence for the necessity of user scheduling under the CF scheme for large number of simultaneously
transmitting users. We proved, with probability which goes
to one, that in this regime the optimal choice of decoding at
the relays is to decode the user with best channel instead of
(22)
a∈ZL \{0}
(21)
M
X
1
P h2l
P 2
≥
log 1 +
,
2
1+P
hj
l=1
A PPENDIX A
P ROOF FOR T HEOREM 5
P (ei ) min R(hem , ei )
(23)
m:eml 6=0
+ P (ei ) min R(hem , am ).
m:aml 6=0
We treat the two terms above separately where the second
term represents the sum rate for the case which the optimal
coefficients vectors may be any integer vector excluding the
unit vector ei . And the first term is for the case that the optimal
coefficients vector is ei . We will show that both terms goes to
zero while starting with the second term.
L
X
Returning to the expectation in (26) we have,
Z
e 2
E khm k =
= P (ei )
m:aml 6=0
L
X
l=1
(a)
Z
∞
≤
min
m:aml 6=0
α
R(hem , am )
e
≥ ) = 0,
lim P (RL
(25)
for all > 0.
Using The Markov and Jensen’s inequalities we have,
1 e
e
P (RL
≥ ) ≤ E RL
1
1
+
e 2
= E P (ei )L log 1 + P khm k
2
1
L
≤ P (ei ) log+ 1 + P E khem k2 ,
2
(27)
which means in words, all possible squared norm values which
belong to all vectors he . We define then, P (ξ) = P (khk2 ∈
Henorm ) as the probability to belong to Henorm . That is,
Z
fχ2L (x)dx = P (ξ).
Henorm
∞
x
fχ2L (x)
P (ξ)
α
fχ2L (x)
1
x
dx ≤
P (ei )
P (ei )
Z
dx
∞
0
xfχ2L (x)dx
L
.
P (ei )
(29)
R∞
Where, α satisfies α fχ2L (x)dx = P (ξ) and (a) is due to the
fact that P (ξ) ≥ P (ei ) since it may happen that two vectors
he and he would have the same squared norm value.
Applying the expectation’s upper bound in (26) we have,
1
L
P (ei ) log+ 1 + P E khem k2
2
L
PL
1
≤ P (ei ) log+ 1 +
2
P (ei )
s
(30)
(a) 1
L
PL
1 Lp
2
=
2P LP (ei )
≤ P (ei )
2
P (ei )
2
(b) 1 L p
√
1
≤
8P 2 L3 e−LE2 (L) = P L2 2Le−LE3 (L)
2
√
where (a) is due the bound log(1+x) ≤ 2x, (b) following
directly from Theorem 4 and E3 (L) = 41 (1 − L1 ) log 2.
Considering the above, as L grows, the second term of (23) is
going to zero, that is,
lim
(26)
therefore, we are interested in analyzing the expectation of
the squared norm values belonging to all channel vectors he .
Remember that, without any constraints, the channel vector
h is a Gaussian random vector which it’s squared norm follows
the χ2L distribution, we shell note as fχ2L (x). A single squared
norm value can belong to a several different Gaussian random
vectors. Hence, we define Henorm as the set of squared norm
values which belongs to he , formally,
Henorm = {khk2 ∈ R|h ∈ He },
P (ξ)
Z
dx ≤
=
≤ P (ei )L max R(hem , am )
m
= P (ei )L max
m
1 + P khem k2
1
+
log
2
kam k2 + P (kam k2 khem k2 − ((hem )T am )2 )
1
≤ P (ei )L max log+ 1 + P khem k2
m 2
1
= max P (ei )L log+ 1 + P khem k2 .
m
2
(24)
+
1
e
e 2
Define RL = P (ei )L 2 log 1 + P khm k , we would like
e p
to show that RL
→ 0, that is,
L→∞
fχ2L (x)
Henorm
P (ei ) min R(hem , am )
l=1
x
(28)
L
X
L→∞
P (ei ) min R(hem , am )
l=1
m:aml 6=0
√
1 P L2 L −LE3 (L)
√
≤ lim
e
=0
L→∞
0.5
(31)
for all > 0.
Thus, we are left with the first term in (23)
lim
L→∞
L
X
P (ei ) min R(hm , ei ) ≤
m:eil 6=0
l=1
L
X
l=1
min R(hm , ei )
m:eil 6=0
L
X
1
1 + P khm k2
+
= lim
min
log
L→∞
m:eil 6=0 2
1 + P (khm k2 − h2i )
l=1
M
X
(b)
1 + P khm k2
1
≤ lim
log+
L→∞
2
1 + P (khm k2 − h2i )
m=1
M
X
1 + P khm k2
1
+
log
lim
=
=0
L→∞ 1 + P (khm k2 − h2
2
i)
m=1
(a)
(32)
where in (a) we set the unit vector ei in the rate expression
R(hm , am ). The upper bound (b) is for the best case scenario
for which each relay has different unit vector ei . Finally, it is
clear that as L grows for each realization of hm the argument
of the log is going to 1.
R EFERENCES
[1] B. Nazer and M. Gastpar, “Compute-and-forward: Harnessing interference through structured codes,” IEEE Transactions on Information
Theory, vol. 57, no. 10, pp. 6463–6486, 2011.
[2] U. Niesen and P. Whiting, “The degrees of freedom of compute-andforward,” IEEE Transactions on Information Theory, vol. 58, no. 8, pp.
5214–5232, 2012.
[3] J. Zhan, B. Nazer, M. Gastpar, and U. Erez, “Mimo compute-andforward,” in 2009 IEEE International Symposium on Information Theory. IEEE, 2009, pp. 2848–2852.
[4] J. Zhan, B. Nazer, U. Erez, and M. Gastpar, “Integer-forcing linear
receivers,” IEEE Transactions on Information Theory, vol. 60, no. 12,
pp. 7661–7685, 2014.
[5] A. Sakzad, J. Harshan, and E. Viterbo, “Integer-forcing mimo linear
receivers based on lattice reduction,” Wireless Communications, IEEE
Transactions on, vol. 12, no. 10, pp. 4905–4915, 2013.
[6] W. He, C. Feng, C. I. Ionita, and B. Nazer, “Collision scheduling for cellular networks,” in Information Theory (ISIT), 2015 IEEE International
Symposium on. IEEE, 2015, pp. 1417–1421.
[7] L. Wei and W. Chen, “Compute-and-forward network coding design
over multi-source multi-relay channels,” IEEE Transactions on Wireless
Communications, vol. 11, no. 9, pp. 3348–3357, 2012.
[8] S.-N. Hong and G. Caire, “Compute-and-forward strategies for cooperative distributed antenna systems,” Information Theory, IEEE Transactions on, vol. 59, no. 9, pp. 5227–5243, 2013.
[9] S. Sahraei and M. Gastpar, “Compute-and-forward: Finding the best
equation,” in Communication, Control, and Computing (Allerton), 2014
52nd Annual Allerton Conference on. IEEE, 2014, pp. 227–233.
[10] D. Dadush, C. Peikert, and S. Vempala, “Enumerative lattice algorithms
in any norm via m-ellipsoid coverings,” in Foundations of Computer
Science (FOCS), 2011 IEEE 52nd Annual Symposium on. IEEE, 2011,
pp. 580–589.
[11] M. Alekhnovich, S. A. Khot, G. Kindler, and N. K. Vishnoi, “Hardness
of approximating the closest vector problem with pre-processing,” in
46th Annual IEEE Symposium on Foundations of Computer Science
(FOCS’05). IEEE, 2005, pp. 216–225.
[12] A. K. Lenstra, H. W. Lenstra, and L. Lovász, “Factoring polynomials
with rational coefficients,” Mathematische Annalen, vol. 261, no. 4, pp.
515–534, 1982.
[13] N. Gama and P. Q. Nguyen, “Finding short lattice vectors within
mordell’s inequality,” in Proceedings of the fortieth annual ACM symposium on Theory of computing. ACM, 2008, pp. 207–216.
[14] J. H. Conway and N. J. A. Sloane, Sphere packings, lattices and groups.
Springer Science & Business Media, 2013, vol. 290.
[15] K. P. Jagannathan, S. Borst, P. Whiting, and E. Modiano, “Efficient
scheduling of multi-user multi-antenna systems,” in Modeling and
Optimization in Mobile, Ad Hoc and Wireless Networks, 2006 4th
International Symposium on. IEEE, 2006, pp. 1–8.
| 7cs.IT
|
Multi-frame image super-resolution with fast
upscaling technique
Longguang Wang, Zaiping Lin, Xinpu Deng, Wei An
Abstract—Multi-frame image super-resolution (MISR) aims
to fuse information in low-resolution (LR) image sequence to
compose a high-resolution (HR) one, which is applied extensively in many areas recently. Different with single image super-resolution (SISR), sub-pixel transitions between multiple
frames introduce additional information, attaching more significance to fusion operator to alleviate the ill-posedness of
MISR. For reconstruction-based approaches, the inevitable
projection of reconstruction errors from LR space to HR
space is commonly tackled by an interpolation operator,
however crude interpolation may not fit the natural image
and generate annoying blurring artifacts, especially after fusion operator. In this paper, we propose an end-to-end fast
upscaling technique to replace the interpolation operator,
design upscaling filters in LR space for periodic sub-locations
respectively and shuffle the filter results to derive the final
reconstruction errors in HR space. The proposed fast upscaling technique not only reduce the computational complexity of
the upscaling operation by utilizing shuffling operation to
avoid complex operation in HR space, but also realize superior performance with fewer blurring artifacts. Extensive experimental results demonstrate the effectiveness and efficiency
of the proposed technique, whilst, combining the proposed
technique with bilateral total variation (BTV) regularization,
the MISR approach outperforms state-of-the-art methods.
Index Terms—multi-frame super-resolution, upscaling
technique, bilateral total variation, shuffling operation
I. INTRODUCTION
UE to the limited technical and manufacturing level,
the resolution of image may not be satisfied in video
surveillance [1][2], medical imaging [3][4][5], aerospace [6][7][8] and many other fields [9][10], where high
resolution (HR) images are commonly required and desired
for distinct image details. With charge-coupled device
(CCD) and CMOS image sensors developing rapidly in
recent decades, the increasing demand in image resolution
still cannot be satisfied, leading to attempts to steer clear of
the sensor issues but utilize computational imaging to improve the spatial resolution, namely super-resolution (SR).
Serving as a typical inverse problem, SR aims to recover
missing image details during image degradation process,
which is underdetermined, requiring additional information
to alleviate the ill-posedness. For single image super-resolution (SISR), lack of observation information
leads to attempts to exploit additional information to learn
how natural images are, and many example-based approaches [11][12][13][14][15][16] have been proposed. For
multi-frame image super-resolution (MISR), sub-pixel
transitions between multiple observations provide underly-
D
L. Wang is with the College of Electronic Science and Engineering, National University of Defense Technology, Changsha, China. (e-mail:
[email protected])
Z. Lin, X. Deng, and W. An are also with the College of Electronic Science and Engineering, National University of Defense Technology,
Changsha, China.
1
ing available information, therefore reconstruction-based
approach [17][20][21][25][26] is mainly concentrated on to
derive the high resolution (HR) image through maintaining
global consistency, which is intuitive and natural.
Concerning reconstruction-based MISR approaches, extensive works have been put forward focusing on the design of regularization to realize favorable results. Tikhonov
regularization SR method [17][18], as a representative
method, introduces smoothness constraints to suppress the
noise but results in the loss of detailed information in edge
regions. To realize edge preserving, total variation (TV)
operator is introduced as a regularization term [19][20][21],
however, leads to the deterioration of smoothness in local
flat region. Motivated by bilateral filter, Farsiu et al. [25]
proposed the bilateral total variation (BTV) operator measured by 𝐿1 norm, which integrates TV with bilateral filter
and realizes superior performance and robustness. Due to
the performance and simplicity of BTV, further improvement has attracted extensive investigation, Li et al. [31]
proposed the locally adaptive bilateral total variation
(LABTV) operator measured by fuzzy-entropy-based
neighborhood homogeneous measurement, realizing locally
adaptive regularization.
Among these reconstruction-based approaches, to maintain global consistency with multiple observations, reconstruction errors are commonly integrated in the cost function to penalize the discrepancy between reconstructed HR
image and LR observations. Within the iterative SR process,
inevitable projection of reconstruction error from LR space
to HR space is usually tackled by an interpolation operator
for simplicity, however this crude operation may introduce
additional errors and lead to deteriorated convergence and
performance, especially after fusion operation of MISR. In
this paper, we propose an end-to-end fast upscaling technique to replace the interpolation operation in the SR
framework. Firstly, we unfold the degradation model to
analyze underlying contributions of periodic sub-locations
to reconstruction error in LR space and design upscaling
filters correspondingly. Secondly, the filter results utilizing
designed upscaling filters are shuffled to derive the reconstruction errors in HR space. Finally, the reconstruction
errors are cooperated with regularization term to modify the
HR image iteratively until convergence. Extensive experiments are conducted to demonstrate the effectiveness and
efficiency of the proposed upscaling technique, besides,
combining the proposed technique with BTV regularization,
the MISR approach realizes state-of-the-art performance.
The rest of the paper is organized as follows: Section II
mainly formulates the problem of multi-frame image super-resolution. Section III presents the proposed upscaling
technique in detail. Section IV performs extensive experiments compared with other state-of-the-art approaches and
the conclusions are drawn in Section V.
𝐹𝑘
Original Image
𝐻𝑘
Geometric Wrap
Fig. 1.
𝐷𝑘
Blurring
𝑁𝑘
Downsample Add Noise
Sketch of degradation model.
II. MULTI-FRAME IMAGE SR PROBLEM FORMULATION
A. Degradation model
As the inverse process of image degradation, SR reconstruction is tightly dependent on the degradation model.
With many degrading factors existing like atmospheric
turbulence, optical blurring, relative motion and sampling
process, the degradation model of LR images can be formulated as:
𝒀𝑘 = 𝑫𝑘 𝑯𝑘 𝑭𝑘 𝑿 + 𝑵𝑘 , 𝑘 = 1,2, … 𝑀,
(1)
where 𝒀𝑘 , X represent 𝑘 𝑡ℎ LR image and HR image respectively, 𝑫𝑘 , 𝑯𝑘 , 𝑭𝑘 serve as decimation matrix, blurring operator and geometric warp matrix of 𝒀𝑘 respectively, and 𝑵𝑘 is the additional Gaussian noise in 𝒀𝑘 .
Note that although complex motions may be common in
real sequences which cannot be represented by a simple
parametric form and many works [28][29][30] tend to address this problem, global translational displacements between multiple frames serving as a fundamental issue, is
still the focus of this paper.
Generally assuming all LR images are generated under
the same condition, we can derive the following model:
𝒀𝑘 = 𝑫𝑯𝑭𝑘 𝑿 + 𝑵𝑘 , 𝑘 = 1,2, … 𝑀,
(2)
where 𝑫 and 𝑯 represent same decimation matrix and
blurring operator respectively in all LR images. The degradation model is further illustrated in Fig. 1.
B. Reconstruction-based SR process
In Bayesian framework, SR reconstruction is equivalent
to the estimation of HR image with given LR images,
where maximum a posteriori (MAP) estimator is extensively utilized as
*∏𝑀
𝑿=
𝑚
𝑘 1 𝑝(𝑿|𝒀𝑘 )+
*∏𝑀
=
𝑚
(3)
𝑘 1 𝑝(𝒀𝑘 |𝑿) 𝑝(𝑿)+
To solve the probabilistic maximization problem, equivalent minimization of reconstruction errors can be derived
as:
2
𝑿=
𝑚𝑖𝑛𝑿 *∑𝑀
(4)
𝑘 1‖𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 ‖ +
As insufficient information given in LR image sequence
𝒀𝑘 , reconstructing the original HR image 𝑿 is an underdetermined problem. To solve the ill-posed problem, regularization is commonly introduced as priori knowledge to
obtain a stable solution and (4) can be rewritten as:
2
𝑿=
𝑚𝑖𝑛 *∑𝑀
𝑘 1‖𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 ‖ + 𝜆𝑅(𝑿)+, (5)
where 𝑅(𝑿) is the regularization term of HR image 𝑿, 𝜆
serves as a regularization parameter weighting the reconstruction errors against the regularization cost.
Assuming decimation matrix 𝑫, blurring operator 𝑯
and geometric warp matrix 𝑭𝑘 are already known, the
minimization problem can be solved utilizing steepest descent approach as:
𝜕𝑅(𝑿)
𝑇 𝑇 𝑇
𝑿𝑙+1 = 𝑿𝑙 − 𝜂 .∑𝑀
/
𝑘 1 𝑭𝑘 𝑯 𝑫 (𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 ) + 𝜆
𝜕𝑿
(6)
where 𝑿𝑙 , 𝑿𝑙+1 are estimators of HR image 𝑿 in 𝑙 𝑡ℎ
and 𝑙 + 1𝑡ℎ iteration respectively, 𝜂 is the learning rate
representing the pace to approach the optimal during iterations. In this paper, the derivation of displacements and blur
kernel is not under consideration, we assume the blur kernel is already known and utilize optical flow method to
estimate the underlying displacements.
III. MULTI-FRAME IMAGE SR WITH FAST UPSCALING TECHNIQUE
In this section, we first present the proposed end-to-end
fast upscaling technique, introduce our motivation and its
formulation in detail before theoretical analysis on computational complexity and convergence. Then we integrate the
proposed upscaling technique with BTV regularization to
construct the overall MISR framework.
A. End-to-end upscaling technique
1) Motivation
As we can see from (6), projecting reconstruction errors
(𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 ) from LR space into HR space is required
in the inference, where interpolation operator commonly
plays a main role as the upscaling operator 𝑫𝑇 , and then
deblurring operator 𝑯𝑇 and inverse translation operator
𝑭𝑘 𝑇 are performed in HR space. Lacking in theoretical
basis, crude interpolation may introduce additional errors
leading to blurring artifacts, therefore it serves as a fundamental operator and requires small stepsize and enough
iterations to alleviate the deterioration, which is
sub-optimal and adds computational complexity [27].
In Shi et al. [27], an efficient sub-pixel convolutional
neural network is proposed, where an array of upscaling
filters are utilized to corporate with a shuffling operator to
upscale the final LR feature maps in to HR output, which is
located at the very end of the network. As [27] demonstrates increasing the resolution of LR image before image
enhancement increase the computational complexity, besides the commonly used interpolation methods do not
bring additional information to solve the ill-posed reconstruction problem, we are inspired to unfold the degradation model to analyze underlying contributions of periodic
sub-locations to reconstruction error in LR space and design similar array of upscaling filters. In this paper, we
propose an end-to-end upscaling technique to perform fast
and efficient upscaling operation, serving as a direct bridge
between reconstruction errors in LR space and HR space.
2) Formulation
Further analyze the degradation model shown in (2) from
the perspective of image, assuming the blurring operator 𝑯
is limited in a × region ( is odd for the symmetry of
blurring kernel) in HR space and the upscaling factor is
determined
sub-location
#4
sub-location
#3
sub-location
#2
sub-location
#1
original
d decimal
geometric wrap
after
translation
blurring
after
blurring
downsample
after
decimation
Fig. 2. Degradation process with respect to different sub-locations.
as 𝛾, namely the decimation operator 𝑫 is limited in a
𝛾 × 𝛾 region in HR space. Concerning translational displacements 𝒅 between LR images, only sub-pixel displacements 𝒅𝑠𝑢𝑏𝑝𝑖𝑥𝑒𝑙 are considered for pixel-level displacements do not bring additional information. Without
loss of generality, only positive sub-pixel displacements
𝒅𝑠𝑢𝑏𝑝𝑖𝑥𝑒𝑙 (𝒅𝑠𝑢𝑏𝑝𝑖𝑥𝑒𝑙 = 𝒅 − 𝑙𝑜𝑜 (𝒅)) are taken into consideration, so that the geometric wrap operator can be limited in a 2 × 2 region.
To illustrate the interaction between the concatenation of
operators, here we set
to be 3 with 𝛾 to be 2 and derive
the degradation model from the perspective of image as
shown in Fig. 2.
With the degradation process unfolded as shown in Fig.
2, it can been seen that different ranges of influence in LR
space correspond with different sub-locations, which inspires us to structure upscaling filters concerning periodic
sub-locations utilizing the differences between influence
ranges. As parameters including sub-pixel displacements
𝒅𝑠𝑢𝑏𝑝𝑖𝑥𝑒𝑙 , blurring kernel 𝑯 and upscaling factor 𝛾 are
all determined, the overall degradation process and the underlying contributions of periodic sub-locations to LR
space can be derived.
Remembering the projection of reconstruction errors
from LR space to HR space in (6), we structure upscaling
filters utilizing the underlying contributions of periodic
sub-locations to realize end-to-end upscaling of reconstruction errors in LR space. Within the probabilistic framework,
the upscaling operator can be equivalent to an optimal estimation problem as
𝑚 𝑝(ℎ𝑖 |𝑒𝑗,𝑗∈𝐶(𝑖) ) , 𝑖 = 1,2, … , 𝑁𝐻𝑅 ,
(7)
𝑖 =
ℎ𝑖
where 𝑖 , 𝑒𝑗 are reconstruction error of 𝑖 𝑡ℎ pixel in HR
space and 𝑗𝑡ℎ pixel in LR space respectively, 𝐶(𝑖) serves
as the influence range in LR space of pixel 𝑖 in HR space
and 𝑁𝐻𝑅 (𝑁𝐻𝑅 = 𝛾 2 𝑁𝐿𝑅 ) is the number of pixels in HR
space. Further assuming 𝑝𝑖𝑗 serves as the contribution of
𝑖 𝑡ℎ pixel in HR space to 𝑗𝑡ℎ pixel in LR space, minimization problem equivalent to (7) can be derived utilizing
greedy strategy
𝑖
=
𝑁𝐶(𝑖)
‖𝑝𝑖𝑗 ℎ𝑖
1
𝑚𝑖𝑛 *∑𝑗
ℎ𝑖
𝑝
− 𝑒𝑗 ‖𝑝 + , 𝑖 = 1,2, … , 𝑁𝐻𝑅 , (8)
where 𝑁𝐶(𝑖) is the number of pixels within 𝐶(𝑖). As the
influence ranges of different sub-locations can be limited in
a 3× 3 LR region in the case of Fig. 2, here we utilize 𝐿2
norm for simplicity and the closed-form solution can be
computed as
ℎ𝑖 =
∑9
𝑗=1 𝑝𝑖𝑗 𝑒𝑗
2
∑9
𝑗=1 𝑝𝑖𝑗
, 𝑖 = 1,2, … , 𝑁𝐻𝑅
(9)
Considering influence range and corresponding contributions are both sub-location dependent, namely HR pixels
with same sub-location share identical influence range and
contribution distribution, we intuitively separate the upscaling operator with respect to different sub-locations and
rewrite (9) for identical sub-location in a convolution form
due to the global consistent process.
𝑇 ∗𝑬
𝑮𝑖 = ‖𝑇𝑖 ‖2 = 𝑇′𝑖 ∗ 𝑬, 𝑖 = 1,2, … , 𝑁𝐻𝑅 ,
(10)
𝑖 2
where 𝑮𝑖 represents reconstruction error map for 𝑖 𝑡ℎ
sub-location in HR space, 𝑬 represents reconstruction
error map in LR space, namely 𝑬 = 𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 and 𝑇𝑖
serves as the contribution distribution concerning 𝑖 𝑡ℎ
sub-location. Regarding the 𝐿2 norm of 𝑇𝑖 as a normalization constant, we integrate it into 𝑇𝑖 to derive normalized contribution distribution 𝑇′𝑖 as filter masks. In this
way, the upscaling operator can be implemented by convolution operator, which realizes favorable efficiency.
As reconstruction errors with respect to ranged
sub-locations in HR space derived separately, a shuffling
operator is introduced to rearrange the elements in separate
error maps to a complete error map 𝑮 in HR space as
shown in Fig. 3. Utilizing the proposed end-to-end upscaling technique, we evade the interpolation operator which
may introduce additional errors, design filter masks according to contribution distribution concerning ranged
sub-locations and process the error map in LR space separately, finally a shuffling operator is implemented to derive
the final error map in HR space. As all processes in LR
P'1
P'2
Shuffling
P'3
P'4
reconstruction error map H
reconstruction error map E
reconstruction error map H1 ~ H 4
Fig. 3. End-to-end upscaling technique.
space are end-to-end, namely all processing results can be
mapped directly to corresponding HR space without intermediate operations, our upscaling technique can realize
superior efficiency and effectiveness, which is demonstrated in the following analysis and Section IV.
3) Theoretical analysis
In this section, theoretical analysis with respect to computational complexity and convergence are carried out respectively, we attempt to illustrate the superiority of the
proposed end-to-end upscaling technique theoretically.
Computational complexity
For conventional interpolation-based upscaling technique,
the reconstruction errors in LR space are commonly projected into HR space by an interpolation operator first, and
then processed by deblurring operator 𝑯𝑇 and inverse
translation operator 𝑭𝑘 𝑇 . In this way, deblurring operator
and inverse translation operator are both performed in HR
space, which adds computational complexity.
Although the complexities of interpolation-based upscaling technique and the proposed end-to-end upscaling
technique are both of order 𝛰(𝑁𝐻𝑅 ), where 𝑁𝐻𝑅 is the
number of pixels in HR space, the computation amounts
differ greatly. Assuming 𝑯, 𝑭𝑘 are limited in × and
2 × 2 region of HR space respectively, for interpolation-based upscaling technique, bicubic interpolation is
commonly utilized performing as weighted sum of neighboring 4 × 4 pixels in LR space, afterwards deblurring
operator 𝑯𝑇 and inverse translation operator 𝑭𝑘 𝑇 are
performed as weighted sum of neighboring × and
2 × 2 pixels respectively in HR space. For our proposed
end-to-end upscaling technique, the upscaling operator is
𝑎+1
performed as weighted sum of neighboring 𝑐𝑒𝑖𝑙(
+
1) × 𝑐𝑒𝑖𝑙(
𝑎+1
𝛾
𝛾
+ 1) pixels in LR space, which remarkably
scent and fixed stepsize are commonly utilized for interpolation-based upscaling technique, which requires small
stepsize and enough iterations to approach the optimal. As
interpolation-based upscaling technique introduces additional errors, the deviation of descent direction makes the
convergence process greatly time consuming.
While Newton-type methods typically tend to converge
in fewer iterations, the computation of Hessian matrix in
each iteration is required, leading to expensive computational cost. As we analyze our end-to-end upscaling technique theoretically, it can be regarded as a variation and
simplification of Newton-type method, which can realize
superior convergence.
Remembering the minimization problem in (5), as we
unfold the degradation model, it can be rewritten as
2
𝑿=
𝑚𝑖𝑛 *∑𝑁
(11)
𝑘 1‖𝑨𝑘 𝑿 − 𝒀𝑘 ‖ + 𝜆𝑅(𝑿)+,
where 𝑿, 𝒀𝑘 represent vectorized HR image and 𝑘 𝑡ℎ LR
image respectively, 𝑨𝑘 = ,𝒂𝑘1 , 𝒂𝑘2 , … , 𝒂𝑘𝐿 - serves as a
dictionary arranged in lexicographic order, which consists
of 𝑁𝐻𝑅 atoms. As analyzed before, different sub-locations
correspond to different influence ranges and contribution
distributions, we utilize this characteristic to construct
overcomplete dictionary as shown in Fig. 4.
For Newton method, the inference of 𝑿 can be written
as
𝑿𝑙+1 = 𝑿𝑙 − ∑𝑀
𝑘
𝜕𝑅(𝑿)
𝜕𝑿
2
𝜕 𝑅(𝑿)
2𝑨𝑘 𝑇 𝑨𝑘 +𝜆
2
𝜕𝑿
2𝑨𝑘 𝑇 (𝑨𝑘 𝑿−𝒀𝑘 )+𝜆
1
(12)
As dictionary 𝑨𝑘 is hard to manipulate, Newton method
cannot be directly utilized in general. Considering 𝑅(𝑿) is
commonly a non-linear operator and the computation of
second derivation is relatively difficult, besides the regularization parameter 𝜆 is usually small, we simplify (12)
and separate 𝑅(𝑿) term out as
𝑨 𝑇 (𝑨 𝑿−𝒀 )
𝜕𝑅(𝑿)
𝑘
𝑘
𝑘
reduces the computational complexity especially with larg𝑿𝑙+1 = 𝑿𝑙 − (∑𝑀
+ 𝜆 𝜕𝑿𝑇 ) (13)
𝑘 1
𝑨𝑘 𝑇 𝑨𝑘
2𝑨𝑘 𝑨𝑘
er upscaling factor 𝛾.
𝑇
Unfold the symmetric matrix 𝑨𝑘 𝑨𝑘 and we can derive
Convergence
For reconstruction-based MISR approaches, steepest de𝒂𝑘1 𝑇 𝒂𝑘−1 𝒂𝑘1 𝑇 𝒂𝑘𝑁𝐻𝑅
𝒂𝑘1 𝑇 𝒂𝑘1 𝒂𝑘1 𝑇 𝒂𝑘2
⋯
𝒂𝑘2 𝑇 𝒂𝑘1 𝒂𝑘2 𝑇 𝒂𝑘2
𝒂𝑘2 𝑇 𝒂𝐾−1 𝒂𝑘2 𝑇 𝒂𝑘𝑁𝐻𝑅
(14)
⋮
⋱
⋮
𝒂𝑘(𝑁𝐻𝑅 −1) 𝑇 𝒂𝑘(𝑁𝐻𝑅 −1) 𝒂𝑘(𝑁𝐻𝑅 −1) 𝑇 𝒂𝑘𝐿
𝒂𝑘(𝑁𝐻𝑅 −1) 𝑇 𝒂𝑘1 𝒂𝑘(𝑁𝐻𝑅 −1) 𝑇 𝒂𝑘2
⋯
𝑇
𝒂𝑘𝑁𝐻𝑅 𝑇 𝒂𝑘(𝑁𝐻𝑅 −1)
𝒂𝑘𝑁𝐻𝑅 𝑇 𝒂𝑘𝑁𝐻𝑅 ]
𝒂𝑘𝑁𝐻𝑅 𝑇 𝒂𝑘2
[ 𝒂𝑘𝑁𝐻𝑅 𝒂𝑘1
As we can see from Fig. 4, the atom is highly sparse and
As the inverse operation of 𝑨𝑘 𝑇 𝑨𝑘 is hard to manipulate,
𝑇
regional, namely 𝒂𝑘𝑖 𝒂𝑘𝑗 equals to zero except 𝑗 ∈ 𝜂(𝑖),
we only take diagonal elements into consideration, namely
where 𝜂(𝑖) represents the neighborhood of corresponding
regard 𝑨𝑘 𝑇 𝑨𝑘 as a diagonal matrix by ignoring other
HR pixel 𝑖 in HR space. Taking this into consideration,
non-diagonal non-zero entries, and rewrite (13) as
𝜕𝑅(𝑿)
𝑇
elements in 𝑨𝑇 𝑨 can be rearranged through placing cor𝑀 𝑨𝑘 (𝑨𝑘 𝑿−𝒀𝑘 )
𝑙+1
𝑙
𝜕𝑿
𝑿
=
𝑿
−
(∑
+
𝜆
) (15)
𝑇
𝑘 1 𝑑𝑖𝑎 (𝑨 𝑨 )
relative atoms closer, and then we can derive an approxi2𝑑𝑖𝑎 (𝑨𝑘 𝑇 𝑨𝑘 )
𝑘
𝑘
𝑇
mate diagonal matrix, namely most entries in 𝑨𝑘 𝑨𝑘
𝐴 𝑇 (𝐴 −𝑌 )
Rewrite 𝑘 𝑘 𝑇 𝑘 in an atom-wise way and we can
equal to zeros except diagonal ones and some other ones.
𝑑𝑖𝑎 (𝐴𝑘 𝐴𝑘 )
derive
sub-location
#1
sub-location
#3
X
Yk
Yk
sub-location
#2
X
sub-location
#4
dictionary AkT
X
Yk
Yk
X
Fig. 4. Procedure of dictionary 𝑨𝑘 .
.
𝐴𝑘 𝑇 (𝐴𝑘 −𝑌𝑘 )
/ =
𝑑𝑖𝑎 (𝐴𝑘 𝑇 𝐴𝑘 ) 𝑖
𝑎𝑘𝑖 𝑇 (𝐴𝑘 −𝑌𝑘 )
(16)
𝑎𝑘𝑖 𝑇 𝑎𝑘𝑖
If we push atom 𝒂𝑘𝑖 backwards into the corresponding LR
image, (16) can rewritten in a convolution form
𝑨𝑘 𝑇 (𝑨𝑘 𝑿−𝒀𝑘 )
(
) =
𝑑𝑖𝑎 (𝑨𝑘 𝑇 𝑨𝑘 ) 𝑖
𝑻𝑘𝑖 ∗(𝑨𝑘 𝑿−𝒀𝑘 )
,
‖𝑻𝑘𝑖 ‖2
2
(17)
where 𝑻𝑘𝑖 represents the contribution distribution map
corresponding to atom 𝒂𝑘𝑖 . Now we can see (15) performs
identical to (10) concerning reconstruction error in LR
space, illustrating the proposed end-to-end upscaling technique performs as a variation and simplification of Newton
method, which can realize superior convergence.
1
As 𝒂𝑘𝑖 is a one-sum atom, ‖𝑻 ‖2 performs as a magni𝑘𝑖 2
fication constant no less than 1. Considering 𝜆 is commonly small, the magnification effect on regularization
term can be ignored and we can derive
𝑿𝑙+1 = 𝑿𝑙 − (∑𝑀
𝑘
1
𝑨𝑘 𝑇 (𝑨𝑘 𝑿−𝒀𝑘 )
𝑨𝑘 𝑇 𝑨𝑘
+𝜆
𝜕𝑅(𝑿)
𝜕𝑿
)
𝑨𝑘 𝑇 (𝑨𝑘 𝑿−𝒀𝑘 )
1 𝑑𝑖𝑎 (𝑨 𝑇 𝑨 )
𝑘
𝑘
+𝜆
𝜕𝑅(𝑿)
𝜕𝑿
(19)
0.034
Interpolation-based technique
End-to-end technique
0.032
0.031
MSE
LR Y1
interpolation
R ( X l )
X l
HR X
strategy
However, the descent direction utilizing our end-to-end
upscaling technique is relatively more accurate, therefore
the convergence can be faster and more stable.
To further demonstrate the superior convergence, we
utilize Tikhonov regularization without loss of generality
and compare the convergence process of interpolation-based with the end-to-end upscaling technique, the
comparison of convergence process is shown in Fig. 5.
0.033
BTV
regularization
LR Y1 ~ YN
(18)
)
LR error map
HR error map
DHFk X Yk end-to-end
G
Considering our end-to-end upscaling technique performs as an approximate Newton method where the simplifications may introduce additional errors, therefore we also
applied a similar learning rate 𝜂 in (18) as (6) for stable
convergence
𝑿𝑙+1 = 𝑿𝑙 − 𝜂 (∑𝑀
𝑘
upscaling technique with BTV regularization to construct
the overall MISR framework. Due to the performance and
simplicity of BTV, it has become one of the most commonly applied regularization in SR process, therefore we
utilize BTV in our MISR framework to corporate with the
proposed end-to-end upscaling technique. The overall
framework is illustrated in Fig. 6 and further summarized
in Algorithm 1.
Fig. 6. Overall MISR framework
Algorithm 1: MISR utilizing end-to-end upscaling technique
Input: LR images 𝒀1 , 𝒀2 , … , 𝒀𝑀 , blurring kernel, upscaling factor 𝛾
Initialize: Select target image (𝒀1 for example), utilize
bicubic interpolation to derive initial HR image 𝑿 , estimate translational displacements 𝒅1𝑘 between target
image 𝒀1 and other LR images.
Loop until ‖𝑿𝑙 − 𝑿𝑙−1 ‖2 /‖𝑿𝑙−1 ‖2 < 𝜖 or 𝑙 𝑁𝑀𝑎𝑥
Compute error map 𝑫𝑯𝑭𝑘 𝑿 − 𝒀𝑘 in LR space
respectively
Perform end-to-end upscaling technique to derive
error map 𝑮 in HR space
Compute BTV regularization 𝑅(𝑿) and its deriva𝜕𝑅(𝑿)
tion
𝜕𝑿
Update 𝑿𝑙−1 to derive 𝑿𝑙 according to (19)
Output: Reconstructed HR image 𝑿
0.03
IV. EXPERIMENTAL RESULTS
0.029
0.028
0.027
0.026
0
50
100
150
200
250
300
Iteration
Fig. 5. Comparison of convergence process
As we can see form Fig. 5, our technique converges
within around 50 iterations while interpolation-based technique requires more than 150 iterations to converge,
demonstrating superior convergence of our end-to-end upscaling technique.
B. MISR framework
In this section, we integrate the proposed end-to-end
In this section, extensive experimental results are presented to demonstrate the effectiveness and efficiency of
the proposed end-to-end upscaling technique. We first perform experiments to demonstrate the effectiveness of our
upscaling technique through equipping it to various reconstruction-based MISR methods, and then the proposed
MISR framework is compared with other state-of-the-art
algorithms.
As described in the degradation model, the degraded LR
images are generated from an HR image through parallel
translations, blurring, downsampling and addition of noises.
In the experiments, the translational displacements 𝒅 are
randomly set with vertical and horizontal shifts randomly
sampled from a uniform distribution in [-5, 5]. The blurring
(a) Tikhonov
(b) TV
(c) BTV
(a) Tikhonov_ETE
(b) TV_ETE
(c) BTV_ETE
(d) LABTV
(d) LABTV_ETE
Fig. 7. Comparison of methods utilizing end-to-end technique with baselines on image Baby.
(a) Tikhonov
(b) TV
(c) BTV
(a) Tikhonov_ETE
(b) TV_ETE
(c) BTV_ETE
(d) LABTV
(d) LABTV_ETE
Fig. 8. Comparison of methods utilizing end-to-end technique with baselines on image Butterfly.
operator is realized utilizing a 5 × 5 Gaussian kernel with
standard deviation 𝜎 = 1 2. After geometric wrapping and
blurring operation, the images are then downsampled by a
factor 𝛾 = 3. Finally Gaussian noise with standard deviation 𝜎𝑛𝑜𝑖𝑠𝑒 = 1 is added.
In our scenario, we use 5 LR images to reconstruct an
HR image for MISR approaches, select the first one as target image without loss of generality and suppose the blurring kernel is already given. As human vision is more sensitive to brightness changes, all the SR methods are imple𝑁𝑅 = 1 𝑙𝑜
{
1
𝐼𝑀 =
.
2
𝑀
2
/,𝑀 𝐸 =
mented only in brightness channel (Y) with color channels
(UV) upscaled by bicubic interpolation for color images.
All the experiments are coded in Matlab R2011 and running on a workstation with Septuple Core i7 @ 3.60 GHz
CPUs and 16 GB memory.
For quantitative analysis and comparison of reconstruction performance, peak-signal-to-noise ratio (PSNR) and
mean structure similarity (SSIM) are utilized as metrics,
which are defined as
1
𝑀𝑁
𝑁
2
∑𝑀
𝑖 1 ∑𝑗 1(𝒀(𝑖, 𝑗) − 𝑿(𝑖, 𝑗))
,
(20)
𝑐 = (𝑘1 𝐿)2
,{ 1
𝑐2 = (𝑘2 𝐿)2
senting the dynamics of a pixel value, and 𝑘1 , 𝑘2 are generally set to be 0.01 and 0.03 respectively.
(2𝜇𝑥 𝜇𝑦+𝑐1 )(2𝜎𝑥 𝜎𝑦 +𝑐2 )
(𝜇𝑥 2 +𝜇𝑦2 +𝑐1 )(𝜎𝑥 2 +𝜎𝑦 2 +𝑐2 )
where 𝜇𝑥 , 𝜇𝑦 are mean value of image 𝑿 and 𝒀 respectively, 𝜎𝑥 , 𝜎𝑦 are standard variance of 𝑿 and 𝒀 respectively, 𝑐1 , 𝑐2 are two stabilizing constants with 𝐿 repre-
A. Evaluation of the proposed upscaling technique
(a) Original
(e) BTV
To validate the effectiveness and efficiency of the pro
(b) Bicubic
(c) Kim
(d) Yang
(f) LABTV
(g) MIScSR
(j) Proposed
Fig. 9. Reconstruction results for image Bridge by ranged methods.
posed end-to-end upscaling technique, we first select four
representative reconstruction-based MISR method consisting of Tikhonov, TV [19], BTV [22] and LABTV [24]
method as baseline methods, apply our technique to replace
interpolation-based technique in the SR pipelines (note as
Tikhonov_ETE for example) and conduct experiments on
Set5 dataset to compare the performance correspondingly.
Visual comparison is exhibited in Fig. 7 and 8 with quantitative results are presented in Table. I.
As we can see from Fig. 7 and 8, compared with corresponding baseline methods, methods equipped with our
end-to-end technique generate sharper edges and fine details, effectively alleviate the blurring effects with fewer
artifacts and realize superior visual quality, which demonstrates the effectiveness of our end-to-end technique.
From the quantitative results shown in Table. I we can
further see that the proposed technique remarkably improves the reconstruction performance with respect to
PSNR and SSIM in all the images of Set5, whilst accelerating the MISR process. The PSNR values have been improved by around 0.9 (dB) in average while SSIM values
also increased by around 0.015. Concerning computational
complexity, the running time of equipped methods is
shortened by around 0.3s in average with practicability
greatly enhanced.
TABLE I
COMPARISON OF PSNR, SSIM AND RUNNDING TIME. THE MEAN PERFROMANCE OF 50 EXPERIMENTS IS PRESENTED WITH THE PERFORMANCE IMPROVEMENT EQUIPPED WITH THE PROPOSED TECHNIQUE SHOWN IN BRACKETS (RED AND BOLD).
Set5
Metric Tikhonov Tikhonov_ETE
TV
TV_ETE
BTV
BTV_ETE
LABTV LABTV_ETE
PSNR
32.75
33.60(↑0.85)
32.76
33.61(↑0.85)
33.18
33.91(↑0.73)
33.38
33.99(↑0.61)
baby
SSIM
0.887
0.903(↑0.016)
0.886 0.903(↑0.017) 0.898 0.903(↑0.005)
0.901
0.904(↑0.003)
Time
1.67
1.19(↓0.48)
2.03
1.50(↓0.53)
2.87
2.34(↓0.53)
25.92
25.39(↓0.53)
PSNR
30.64
31.59(↑0.95)
30.66
31.61(↑0.95)
30.80
31.99(↑1.19)
31.00
31.99(↑0.99)
bird
SSIM
0.898
0.909(↑0.011)
0.898 0.909(↑0.011) 0.903 0.911(↑0.002)
0.904
0.906(↑0.002)
Time
0.75
0.47(↓0.28)
0.80
0.58(↓0.22)
1.06
0.82(↓0.24)
8.70
8.47(↓0.23)
PSNR
23.15
24.12(↑0.97)
23.19
24.13(↑0.94)
22.78
24.08(↑1.30)
23.01
24.23(↑1.22)
butterfly
SSIM
0.799
0.827(↑0.028)
0.801 0.827(↑0.026) 0.801 0.825(↑0.024)
0.806
0.832(↑0.026)
Time
0.61
0.39(↓0.22)
0.69
0.48(↓0.21)
0.90
0.70(↓0.20)
7.18
6.83(↓0.35)
PSNR
30.18
30.52(↑0.34)
30.18
30.53(↑0.35)
30.39
30.60(↑0.21)
30.46
30.64(↑0.18)
head
SSIM
0.726
0.743(↑0.017)
0.726 0.744(↑0.018) 0.741 0.748(↑0.007)
0.746
0.752(↑0.006)
Time
0.68
0.45(↓0.23)
0.78
0.54(↓0.24)
1.00
0.77(↓0.23)
8.50
8.32(↓0.18)
PSNR
27.71
28.90(↑1.19)
27.73
28.93(↑1.20)
27.43
28.97(↑1.54)
27.70
29.10(↑1.40)
woman
SSIM
0.882
0.901(↑0.019)
0.882 0.902(↑0.020) 0.885 0.904(↑0.019)
0.890
0.904(↑0.014)
Time
0.71
0.47(↓0.24)
0.79
0.55(↓0.24)
1.04
0.80(↓0.24)
8.53
7.98(↓0.55)
PSNR
28.89
29.75(↑0.86)
28.90
29.76(↑0.86)
28.92
29.91(↑0.99)
29.11
29.99(↑0.88)
average
SSIM
0.838
0.856(↑0.018)
0.838 0.857(↑0.019) 0.846 0.857(↑0.011)
0.849
0.859(↑0.010)
Time
0.88
0.59(↓0.29)
1.02
0.73(↓0.29)
1.37
1.09(↓0.28)
11.77
11.40(↓0.37)
B. Comparison with state-of-the-art methods
To further demonstrate the effectiveness and efficiency
of the proposed MISR framework, seven methods are selected to compare with our work. As bicubic interpolation
serves as the simplest SR approach, it is selected as a baseline method. Serving as most cited method in the field of
MISR for the performance and simplicity, Farsiu’s BTV
method [25] is selected besides with its variation Li’s
LABTV method [31]. As the popularity of learning-based
approaches increases, we also select Kato’s sparse coding
method for MISR (denote as MIScSR) [32] as one
state-of-the-art method in this field. In addition, Kim’s [34]
and Yang’s [33] methods considered as state-of-the-art
SISR methods are also introduced in the comparison.
For fair comparisons, the source codes of Kim’s and
Yang’s methods released in the authors’ homepages are
(a) Original
(e) BTV
directly implemented in our experiments. As no available
(b) Bicubic
(c) Kim
(d) Yang
(f) LABTV
(g) MIScSR
(j) Proposed
Fig. 10. Reconstruction results for image Commic by ranged methods.
(a) Original
(e) BTV
(b) Bicubic
(c) Kim
(f) LABTV
(g) MIScSR
(d) Yang
(j) Proposed
Fig. 11. Reconstruction results for image Foreman by ranged methods.
codes for other methods, we implement them according to
the instructions in [25][31][32], and the performance may
differ from the original. Note that Kato’s MIScSR method
is only utilized for comparison with upscaling factor 3 for
instructions in [32] only presented implementation details
and parameter settings under this condition. Extensive experiments are conducted on Set14 dataset, and the reconstruction results are exhibited in Figs. 9-13 with quantitative results presented in Table. III.
In our scenario, for bicubic method and SISR methods,
only the target LR image (first LR image) is utilized for
reconstruction, and for MISR methods, same registration
procedure is adopted. Note that the blurring kernel is sup-
posed to be given for all SR approaches for its derivation is
not the focus of this paper. Other detailed parameter settings for our MISR framework are summarized in Table.
II.
TABLE II
PARAMETER SETTINGS FOR THE PROPOSED UPSCALING
TECHNIQUE
Parameters
Values
𝜖: tolerance threshold
1 −
30
𝑁𝑀𝑎𝑥 : maximum iteration
0.1
𝜆: regularization parameter in (17)
0.1
𝜂: step size in (17)
From the perspective of visual quality, for SISR methods,
Kim’s method serving as the superior approach, has already
recovered the major structures of the scene, however it
(a) Original
(e) BTV
tends to oversmooth fine details. For MISR methods,
(b) Bicubic
(c) Kim
(d) Yang
(f) LABTV
(g) MIScSR
(j) Proposed
Fig. 12. Reconstruction results for image Girl by ranged methods.
blurring artifacts in BTV and LABTV methods are commonly noticeable, especially within edge and texture regions. Although the sparse representation alleviates the
blurring effect for MIScSR, some ragged edges are still
visible. By comparison, the proposed MISR approach produces sharper and clearer images with fine details and fewer artifacts.
From the quantitative results exhibited in Table. III we
can further see that our approach outperforms other
state-of-the-art methods in all the images of Set14 with
respect to reconstruction performance and efficiency.
Compared with Kim’s method, serving as the superior
SISR method, the PSNR value of our approach is improved
by 0.62 (dB) in average with processing efficiency 10 times
faster. Compared with MIScSR method, known as the
state-of-the-art MISR method, our approach improves the
PSNR value by 0.49 (dB) and runs nearly 8 times faster.
Comparing our approach with BTV method, the superiority
of the proposed end-to-end upscaling technique can be further extensively validated by the higher PSNR values with
shorter running time. Leave out the bicubic method, we can
see our approach performs as the most effective and efficient one among comparing methods with practical applications.
TABLE III
COMPARISON OF PSNR AND RUNNING TIME FOR RANGE METHODS. FOR MIMR METHODS, THE MEAN PERFROMANCE OF 50 EXPERIMENTS IS PRESENTED WITH STANDARD DERIVATION SHOWN IN BRACKETS. THE BEST RESULTS ARE SHOWN IN RED BOLD.
Set14
baboon
barbara
bridge
coastguard
comic
face
flowers
foreman
lenna
man
monarch
pepper
ppt
zebra
average
Bicubic
PSNR
Time
20.81
0.04
24.13
0.04
23.74
0.03
24.68
0.01
20.85
0.01
29.44
0.01
24.40
0.02
28.38
0.01
28.93
0.02
24.93
0.02
26.81
0.03
29.18
0.02
21.31
0.02
23.77
0.02
25.10
0.02
Kim
PSNR
Time
21.19
40.16
24.76
47.94
24.55
57.20
25.43
18.95
21.93
22.68
30.11
10.25
25.55
32.32
30.59
12.90
30.30
18.49
26.13
30.87
29.02
39.32
30.30
19.08
22.91
45.44
26.05
45.29
26.34
31.49
PSNR
21.16
24.75
24.53
25.44
21.77
29.99
25.41
30.06
29.99
25.92
28.39
30.10
22.52
25.58
26.11
Yang
Time
673.24
1103.15
980.27
319.60
288.58
275.17
424.51
253.73
412.01
414.53
624.27
569.97
579.91
506.18
530.37
BTV
PSNR
21.24(±0.01)
24.82(±0.02)
24.75(±0.02)
25.54(±0.02)
21.69(±0.01)
30.38(±0.03)
25.42(±0.01)
29.97(±0.03)
30.18(±0.03)
26.00(±0.02)
28.13(±0.02)
30.19(±0.03)
22.10(±0.01)
25.41(±0.02)
26.13(±0.02)
To further demonstrate the effectiveness of the proposed
MISR approach, experiments are conducted on Set5 dataset
concerning ranged upscaling factors and noise intensities
with results presented in Table. IV and Table. V.
As shown in Table. IV, with upscaling factor increased,
the effects of multiple observations are gradually erased
due to the growing ill-posedness, leading to the performances of MISR approaches deteriorating severely and
even inferior to SISR approaches in several conditions.
Although our approach also undergoes same deterioration,
Time
3.76
6.59
5.47
2.23
1.94
1.84
3.10
1.63
3.15
2.90
4.74
3.39
4.90
2.61
3.45
LABTV
PSNR
21.33(±0.02)
24.95(±0.02)
24.92(±0.04)
25.73(±0.04)
21.86(±0.02)
30.47(±0.04)
25.59(±0.02)
30.21(±0.05)
30.35(±0.04)
26.17(±0.03)
28.34(±0.03)
30.28(±0.04)
22.24(±0.02)
25.70(±0.04)
26.30(±0.03)
Time
33.77
61.41
57.74
22.66
19.50
17.23
30.09
15.69
30.35
27.00
41.31
27.05
36.04
23.98
31.70
MIScSR
PSNR
21.33(±0.00)
25.13(±0.00)
24.87(±0.00)
25.60(±0.00)
22.14(±0.00)
30.22(±0.00)
25.73(±0.00)
30.43(±0.00)
30.32(±0.00)
26.25(±0.00)
28.80(±0.00)
30.10(±0.00)
23.07(±0.00)
26.59(±0.00)
26.47(±0.00)
Time
27.95
48.36
41.48
15.69
14.31
12.13
24.05
11.89
21.99
18.70
28.23
18.74
24.78
16.43
23.20
Proposed
PSNR
21.64(±0.03)
25.37(±0.04)
25.61(±0.06)
26.40(±0.06)
22.59(±0.04)
30.64(±0.06)
26.28(±0.05)
31.04(±0.08)
30.86(±0.07)
26.88(±0.05)
29.30(±0.08)
30.54(±0.05)
23.11(±0.07)
27.21(±0.08)
26.96(±0.06)
Time
3.01
5.46
4.53
1.58
1.54
1.36
2.49
1.25
2.55
2.37
4.61
3.02
4.24
2.73
2.91
it still performs superiorly with highest PSNR values in
most conditions.
From Table. V we can further see that our approach performs strong robustness and tolerance to noises. As all the
SR methods are sensitive to noises and deteriorate with
noise intensity increased, the proposed approach still outperforms in average. Even under the condition of noise
intensity 3, the proposed approach performs +0.46 dB and
+0.42 dB compared with Kim’s method and LABTV
method respectively.
(a) Original
(e) BTV
(b) Bicubic
(c) Kim
(d) Yang
(f) LABTV
(g) MIScSR
(j) Proposed
Fig. 13. Reconstruction results for image Lenna by ranged methods.
TABLE IV
MAGNIFICATION × 2, × 3 AND × 4 PERFORMANCE IN TERMS OF PSNR ON SET5 DATASET. FOR MIMR METHODS, THE MEAN PERFROMANCE OF 20 EXPERIMENTS IS PRESENTED WITH STANDARD DERIVATION SHOWN IN BRACKETS. THE BEST RESULTS ARE
SHOWN IN RED BOLD.
Set5
baby
bird
butterfly
head
woman
average
baby
bird
butterfly
head
woman
average
baby
bird
butterfly
head
woman
average
Upscaling
factor
2
3
4
Bicubic
Kim
Yang
BTV
LABTV
MIScSR
Proposed
32.53
30.71
22.98
30.25
27.43
28.78
31.07
28.94
21.53
29.44
25.85
27.38
29.63
27.31
20.21
28.75
24.48
26.08
33.52
32.00
24.61
30.69
28.81
29.93
32.39
30.58
24.12
30.11
27.87
29.02
30.08
28.76
22.54
29.39
26.37
27.43
33.30
31.68
24.20
30.64
28.51
29.67
32.23
30.22
23.31
29.99
27.47
28.65
30.77
28.49
21.78
29.21
25.86
27.22
34.74(±0.03)
32.89(±0.02)
24.41(±0.01)
31.20(±0.02)
29.26(±0.02)
30.50(±0.02)
33.13(±0.05)
30.81(±0.03)
22.77(±0.01)
30.38(±0.03)
27.42(±0.03)
28.91(±0.03)
31.50(±0.09)
28.88(±0.08)
21.30(±0.03)
29.61(±0.06)
25.81(±0.04)
27.42(±0.06)
35.08(±0.05)
33.26(±0.03)
24.93(±0.02)
31.37(±0.03)
29.79(±0.04)
30.89(±0.03)
33.31(±0.07)
31.00(±0.05)
23.00(±0.02)
30.47(±0.04)
27.67(±0.05)
29.10(±0.05)
31.60(±0.11)
28.97(±0.10)
21.41(±0.04)
29.65(±0.07)
25.94(±0.05)
27.51(±0.07)
N.A.
N.A.
N.A.
N.A.
N.A.
N.A.
32.86(±0.00)
30.92(±0.00)
23.57(±0.00)
30.22(±0.00)
28.22(±0.00)
29.17(±0.00)
N.A.
N.A.
N.A.
N.A.
N.A.
N.A.
35.37(±0.05)
34.20(±0.06)
26.77(±0.04)
31.39(±0.03)
31.51(±0.07)
31.85(±0.05)
33.83(±0.08)
31.87(±0.07)
24.13(±0.07)
30.64(±0.06)
28.87(±0.13)
29.87(±0.08)
31.85(±0.12)
29.45(±0.15)
22.04(±0.11)
29.66(±0.09)
26.66(±0.11)
27.93(±0.12)
TABLE V
NOISE INTENSITY 1, 2 AND 3 PERFORMANCE IN TERMS OF PSNR ON SET5 DATASET. FOR MIMR METHODS, THE MEAN PERFROMANCE OF 50 EXPERIMENTS IS PRESENTED WITH STANDARD DERIVATION SHOWN IN BRACKETS. THE BEST RESULTS ARE SHOWN
IN RED BOLD.
Set5
baby
bird
butterfly
head
woman
average
baby
bird
butterfly
head
woman
average
baby
bird
butterfly
head
woman
average
Noise
intensity
1
2
3
Bicubic
Kim
Yang
BTV
LABTV
MIScSR
Proposed
31.07
28.94
21.53
29.49
25.85
27.38
30.89
28.84
21.50
29.37
25.80
27.28
30.63
28.69
21.47
29.18
25.71
27.14
32.39
30.58
24.12
30.15
27.87
29.02
32.09
30.37
24.07
29.99
27.74
28.85
31.61
30.04
23.96
29.64
27.52
28.55
32.23
30.22
23.31
30.04
27.47
28.65
31.87
30.00
23.27
29.80
27.33
28.45
31.38
29.68
23.18
29.49
27.12
28.17
33.13(±0.05)
30.81(±0.03)
22.77(±0.01)
30.41(±0.03)
27.42(±0.03)
28.91(±0.03)
32.74(±0.04)
30.56(±0.03)
22.72(±0.02)
30.17(±0.03)
27.29(±0.03)
28.70(±0.03)
32.24(±0.04)
30.25(±0.03)
22.66(±0.02)
29.90(±0.02)
27.14(±0.02)
28.44(±0.03)
33.31(±0.07)
31.00(±0.05)
23.00(±0.02)
30.51(±0.04)
27.67(±0.05)
29.10(±0.05)
32.89(±0.05)
30.73(±0.05)
22.94(±0.03)
30.25(±0.04)
27.53(±0.04)
28.87(±0.04)
32.33(±0.06)
30.41(±0.04)
22.87(±0.03)
29.96(±0.03)
27.36(±0.03)
28.59(±0.04)
32.86(±0.00)
30.92(±0.00)
23.57(±0.00)
30.27(±0.00)
28.22(±0.00)
29.17(±0.00)
32.17(±0.01)
30.48(±0.01)
23.48(±0.00)
29.85(±0.01)
27.97(±0.01)
28.79(±0.01)
31.24(±0.01)
29.84(±0.02)
23.34(±0.01)
29.27(±0.02)
27.58(±0.01)
28.25(±0.01)
33.83(±0.08)
31.87(±0.07)
24.13(±0.07)
30.66(±0.05)
28.87(±0.13)
29.87(±0.08)
33.07(±0.06)
31.33(±0.08)
23.93(±0.09)
30.27(±0.04)
28.56(±0.09)
29.43(±0.07)
32.36(±0.04)
30.82(±0.06)
23.78(±0.09)
29.89(±0.03)
28.24(±0.07)
29.01(±0.06)
V. CONCLUSIONS
In this paper, we propose an end-to-end fast upscaling
technique to replace the interpolation operator for recon-
struction-based MISR approaches. As we unfold the degradation model from the perspective of image, we find the
influence ranges and underlying contributions of periodic
sub-locations vary periodically, which inspires us to design
upscaling filters for periodic sub-locations respectively and
utilize a shuffling operator to realize effective fusion operation. Equipped with our upscaling technique, remarkable
improvements are realized with respect to reconstruction
performance and efficiency for reconstruction-based methods, besides the cooperation of our technique and BTV
regularization outperforms other state-of-the-art methods
demonstrated by extensive experiments.
REFERENCES
[1] V. Chandran, C. Fookes, F. Lin, S. Sridharan, “Investigation into optical flow super-resolution for surveillance applications,” Aprs Workshop on Digital Image Computing, vol. 26, no. 47, pp. 73–78, 2005.
[2] L. Zhang, H. Zhang, H. Shen, and P. Li, “A super-resolution reconstruction algorithm for surveillance images,” Signal Processing, vol.
90, no. 3, pp. 848–859, 2010.
[3] H. Greenspan, G. Oz, N. Kiryati, S. Peled, “Super-resolution in MRI,”
in Proc. IEEE ISBI, 2002, pp. 943-946.
[4] W. Shi, J. Caballero, C. Ledig, X. Zhuang, W. Bai, K. Bhatia, A. Marvao, T. Dawes, D. ORegan, and D. Rueckert, “Cardiac image super-resolution with global correspondence using multi-atlas patchmatch,” in Proc. Int. Conf. Medical Image Computing and Computer
Assisted Intervention (MICCAI), Jan. 2013, pp. 9–16.
[5] D. Trinh, M. Luong, F. Dibos, J. Rocchisani, C. Pham, and T. Q. Nguyen, “Novel Example-Based Method for Super-Resolution and Denoising of Medical Images,” IEEE Trans. Image Process., vol. 23, no.
4, pp. 1882-1895, Apr. 2014.
[6] A.J. Tatem, H.G. Lewis, P.M. Atkinson, M.S. Nixon, “Super-resolution target identification from remotely sensed images using
a hopfield neural network,” IEEE Trans. Geosci. Rem. Sens., vol. 39,
no. 4, pp. 781-796, Apr. 2001.
[7] M. W. Thornton, P. M. Atkinson, and D. a. Holland, “Sub-pixel mapping of rural land cover objects from fine spatial resolution satellite
sensor imagery using super-resolution pixel-swapping,” International
Journal of Remote Sensing, vol. 27, no. 3, pp. 473–491, Jan. 2006.
[8] K. Makantasis, K. Karantzalos, A. Doulamis, N. Doulamis, “Deep
supervised learning for hyperspectral data classification through convolutional neural networks,” in Proc. IEEE IGARSS, Jul. 2015, pp.
4959–4962.
[9] A.J. Tatem, H.G. Lewis, P.M. Atkinson, M.S. Nixon, “Super-resolution
land cover pattern prediction using a Hopfield neural network,” Remote Sens. Environ., vol. 79, no. 1, pp. 1–14, 2002.
[10] T. Goto, T. Fukuoka, F. Nagashima, S. Hirano and M. Sakurai, “Super-resolution System for 4K-HDTV, ” in Proc. Int Conf. Pattern
Recognition, Jan. 2014, pp. 4453–4458.
[11] K. Zhang, X. Gao, D. Tao and X. Li, “Multi-scale dictionary for
single image super-resolution,” in Proc. CVPR, Providence, RI, Jun.
2012, pp. 1114-1121.
[12] C. Yang, and M. H. Yang. “Fast direct super-resolution by simple
functions,” in Proc. ICCV, 2013, pp. 561-568.
[13] R. Timofte, V. De, and L. V. Gool. “Anchored neighborhood regression for fast example-based super-resolution,” in Proc. ICCV, 2014,
pp. 1920-1927.
[14] C. Dong, C. C. Loy, K. He and X. Tang, “Learning a deep convolutional network for image super-resolution,” In Proc. ECCV, 2014, pp.
184-199.
[15] S. Yang, Z. Wang, L. Zhang, and M. Wang, “Dual-geometric neighbor
embedding for image super resolution with sparse tensor,” IEEE
Trans. Image Process., vol. 23, no. 7, pp. 2793-2803, Jul. 2014.
[16] S. Gu, et al. “Convolutional sparse coding for image super-resolution,” in Proc. ICCV, 2016, pp. 1823-1831.
[17] N. Nguyen, P. Milanfar, G. Golub, “A computationally efficient superresolution image reconstruction algorithm,” IEEE Trans. Image
Process., vol. 10, no. 4, pp. 573–583, 2001.
[18] X. Zhang, E.Y. Lam, E.X. Wu, K.K. Wong, “Application of Tikhonov
regularization to super-resolution reconstruction of brain MRI image,”
Lecture Notes in Computer Science, vol. 4987, pp. 51–56, 2008.
[19] M. K. Ng, H. Shen, E. Y. Lam, L. Zhang, “A total variation regularization based super-resolution reconstruction algorithm for digital
video,” EURASIP J. Adv. Signal Process, vol. 2007, no. 74585, pp.
1-16, 2007.
[20] S. D. Babacan, R. Molina, A.K. Katsaggelos, “Parameter estimation
in TV image restoration using variational distribution approximation,”
IEEE Trans. Image Process., vol. 17, no. 3, pp. 326–339, Apr. 2008.
[21] Q. Yuan, L. Zhang, H. Shen, “Multiframe super-resolution employing
a spatially weighted total variation model,” IEEE Trans.Circuits Syst.
Video Technol., vol. 22, no. 3, pp. 379-392Mar. 2012.
[22] H. Shen, L. Zhang, B. Huang, P. Li, “A MAP approach for joint motion estimation segmentation and super resolution,” IEEE Trans. Image Process., vol. 16, no. 2, pp. 479–490, Jan. 2007.
[23] R. Molina, J. Mateos, A. K. Katsaggelos, and M. Vega, “Bayesian
multichannel image restoration using compound Gauss-Markov random fields,” IEEE Trans. Image Process., vol. 12, no. 12, pp. 1642–
1654, Feb. 2003.
[24] F. Humblot, A. Mohammad-Djafari, “Superresolution using hidden
Markov Model and Bayesian detection estimation framework,”
EURASIP J. Appl. Signal Process., vol. 2006, Article ID. 36971, pp.
1–16, Jan. 2006.
[25] S. Farsiu, M.D. Robinson, M. Elad, P. Milanfar, “Fast and robust
multi-frame superresolution,” IEEE Trans. Image Process., vol. 13, no.
10, pp. 1327–1344, Oct. 2004.
[26] P. Purkait and B. Chanda, “Super resolution image reconstruction
through bregman iteration using morphologic regularization,” IEEE
Trans. Image Process., vol. 21, no. 9, pp. 4029-4039, Sep. 2012.
[27] W. Shi, J. Caballero, F. Huszar, et al, “Real-time single image and
video super-resolution using an efficient sub-pixel convolutional neural network,” in Proc. IEEE CVPR, Sep. 2016, pp. 1874-1883.
[28] M. Protter, M. Elad, H. Takeda, and P. Milanfar, “Generalizing the
non-local-means to super-resolution reconstruction,” IEEE Trans.
Image Process., vol. 18, no. 1, pp. 36-51, Mar. 2008.
[29] H. Takeda, P. Milanfar, M. Protter, and M. Elad, “SuperResolution
without Explicit Subpixel Motion Estimation,” IEEE Trans. Image
Process., vol. 18, no. 9, pp. 1958-1975, Sep. 2009
[30] C. Liu, D. Sun, “On Bayesian adaptive video super resolution,” IEEE
Trans. Image Process., vol. 36, no. 2, pp. 346-360, Feb. 2014.
[31] X. Li, Y. Hu, X. Gao, D.Tao, B. Ning, “A multi-frame image super-resolution method,” Signal Processing, vol. 90, no. 2, pp.405-414,
2010.
[32] T. Kato, H. Hino, N. Murata, “Multi-frame image super resolution
based on sparse coding,” Neural Networks, vol. 66, pp. 64-78, 2015.
[33] J. Yang, J. Wright, T. Huang. “Image super-resolution as sparse representation of raw image patches,” IEEE Trans. Image Process., vol.
19, no. 11, pp. 2861,2010.
[34] K. I. Kim and Y. Kwon. “Single-image super-resolution using sparse
regression and natural image prior,” IEEE Trans. Pattern Anal. Mach.
Intell., vol. 32, no. 6, pp. 1127-1133, 2010.
| 1cs.CV
|
arXiv:1601.06367v1 [math.AC] 24 Jan 2016
THE ANNIHILATING-SUBMODULE GRAPH OF MODULES
OVER COMMUTATIVE RINGS II
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
Abstract. Let M be a module over a commutative ring R. In this paper,
we continue our study of annihilating-submodule graph AG(M ) which was
introduced in (The Zariski topology-graph of modules over commutative rings,
Comm. Algebra., 42 (2014), 3283–3296). AG(M ) is a (undirected) graph in
which a nonzero submodule N of M is a vertex if and only if there exists
a nonzero proper submodule K of M such that N K = (0), where N K, the
product of N and K, is defined by (N : M )(K : M )M and two distinct vertices
N and K are adjacent if and only if N K = (0). We prove that if AG(M ) is a
tree, then either AG(M ) is a star graph or a path of order 4 and in the latter
case M ∼
= F × S, where F is a simple module and S is a module with a unique
non-trivial submodule. Moreover, we prove that if M is a cyclic module with
at least three minimal prime submodules, then gr(AG(M )) = 3 and for every
cyclic module M , cl(AG(M )) ≥ |M in(M )|.
1. Introduction
Throughout this paper R is a commutative ring with a non-zero identity and M
is a unital R-module. By N ≤ M (resp. N < M ) we mean that N is a submodule
(resp. proper submodule) of M .
Define (N :R M ) or simply (N : M ) = {r ∈ R| rM ⊆ N } for any N ≤ M .
We denote ((0) : M ) by AnnR (M ) or simply Ann(M ). M is said to be faithful if
Ann(M ) = (0).
Let N, K ≤ M . Then the product of N and K, denoted by N K, is defined by
(N : M )(K : M )M (see [6]).
There are many papers on assigning graphs to rings or modules (see, for example,
[4, 7, 10, 11]). The annihilating-ideal graph AG(R) was introduced and studied in
[11]. AG(R) is a graph whose vertices are ideals of R with nonzero annihilators
and in which two vertices I and J are adjacent if and only if IJ = (0). Later, it
was modified and further studied by many authors (see, e.g., [1-3]).
In [7, 8], we generalized the above idea to submodules of M and defined the
(undirected) graph AG(M ), called the annihilating-submodule graph, with vertices
V (AG(M ))= {N ≤ M | there exists (0) 6= K < M with N K = (0)}. In this graph,
distinct vertices N, L ∈ V (AG(M )) are adjacent if and only if N L = (0). Let
AG(M )∗ be the subgraph of AG(M ) with vertices V (AG(M )∗ ) = {N < M with
(N : M ) 6= Ann(M )| there exists a submodule K < M with (K : M ) 6= Ann(M )
and N K = (0)}. Note that M is a vertex of AG(M ) if and only if there exists a
Date: April 8, 2018.
2010 Mathematics Subject Classification. primary 05C75, secondary 13C13.
Key words and phrases. annihilating-submodule graph, cyclic module, minimal prime submodule, chromatic and clique number.
1
2
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
nonzero proper submodule N of M with (N : M ) = Ann(M ) if and only if every
nonzero submodule of M is a vertex of AG(M ).
In this work, we continue our studying in [7, 8] and we generalize some results related to annihilating-ideal graph obtained in [1-3] for annihilating-submodule graph.
A prime submodule of M is a submodule P 6= M such that whenever re ∈ P for
some r ∈ R and e ∈ M , we have r ∈ (P : M ) or e ∈ P [13].
The prime radical radM (N ) or simply rad(N ) is defined to be the intersection
of all prime submodules of M containing N , and in case N is not contained in any
prime submodule, radM (N ) is defined to be M [13].
The notations Z(R), N il(R), and M in(M ) will denote the set of all zero-divisors,
the set of all nilpotent elements of R, and the set of all minimal prime submodules
of M , respectively. Also, ZR (M ) or simply Z(M ), the set of zero divisors on M , is
the set {r ∈ R| rm = 0 for some 0 6= m ∈ M }.
A clique of a graph is a complete subgraph and the supremum of the sizes of
cliques in G, denoted by cl(G), is called the clique number of G. Let χ(G) denote the
chromatic number of the graph G, that is, the minimal number of colors needed
to color the vertices of G so that no two adjacent vertices have the same color.
Obviously χ(G) ≥ cl(G).
In section 2 of this paper, we prove that if AG(M ) is a tree, then either AG(M )
is a star graph or is the path P4 and in this case M ∼
= F × S, where F is a simple
module and S is a module with a unique non-trivial submodule (see Theorem
2.7). Next, we study the bipartite annihilating-submodule graphs of modules over
Artinian rings (see Theorem 2.8). Moreover, we give some relations between the
existence of cycles in the annihilating-submodule graph of a cyclic module and the
number of its minimal prime submodules (see Theorem 2.18 and Corollary 2.19)).
Let us introduce some graphical notions and denotations that are used in what
follows: A graph G is an ordered triple (V (G), E(G), ψG ) consisting of a nonempty
set of vertices, V (G), a set E(G) of edges, and an incident function ψG that associates an unordered pair of distinct vertices with each edge. The edge e joins x
and y if ψG (e) = {x, y}, and we say x and y are adjacent. A path in graph G is a
finite sequence of vertices {x0 , x1 , . . . , xn }, where xi−1 and xi are adjacent for each
1 ≤ i ≤ n and we denote xi−1 − xi for existing an edge between xi−1 and xi .
A graph H is a subgraph of G, if V (H) ⊆ V (G), E(H) ⊆ E(G), and ψH is
the restriction of ψG to E(H). A bipartite graph is a graph whose vertices can be
divided into two disjoint sets U and V such that every edge connects a vertex in
U to one in V ; that is, U and V are each independent sets and complete bipartite
graph on n and m vertices, denoted by Kn,m , where V and U are of size n and
m, respectively, and E(G) connects every vertex in V with all vertices in U . Note
that a graph K1,m is called a star graph and the vertex in the singleton partition
is called the center of the graph. For some U ⊆ V (G), we denote by N (U ), the
set of all vertices of G \ U adjacent to at least one vertex of U . For every vertex
v ∈ V (G), the size of N (v) is denoted by d(v). If all the vertices of G have the
same degree k, then G is called k-regular, or simply regular. An independent set is
a subset of the vertices of a graph such that no vertices are adjacent. We denote
by Pn and Cn , a path and a cycle of order n, respectively. Let G and G′ be two
graphs. A graph homomorphism from G to G′ is a mapping φ : V (G) −→ V (G′ )
such that for every edge {u, v} of G, {φ(u), φ(v)} is an edge of G′ . A retract of G
is a subgraph H of G such that there exists a homomorphism φ : G −→ H such
THE ANNIHILATING-SUBMODULE GRAPH OF MODULES
3
that φ(x) = x, for every vertex x of H. The homomorphism φ is called the retract
(graph) homomorphism (see [19]).
2. The Annihilating-submodule graph II
An ideal I ≤ R is said to be nil if I consist of nilpotent elements; I is said to be
nilpotent if I n = (0) for some natural number n.
Proposition 2.1. Suppose that e is an idempotent element of R. We have the
following statements.
(a) R = R1 ⊕ R2 , where R1 = eR and R2 = (1 − e)R.
(b) M = M1 ⊕ M2 , where M1 = eM and M2 = (1 − e)M .
(c) For every submodule N of M , N = N1 × N2 such that N1 is an R1 submodule M1 , N2 is an R2 -submodule M2 , and (N :R M ) = (N1 :R1
M1 ) × (N2 :R2 M2 ).
(d) For submodules N and K of M , N K = N1 K1 × N2 K2 such that N =
N1 × N2 and K = K1 × K2 .
(e) Prime submodules of M are P × M2 and M1 × Q, where P and Q are prime
submodules of M1 and M2 , respectively.
Proof. This is clear.
We need the following lemmas.
Lemma 2.2. (See [5, Proposition 7.6].) Let R1 , R2 , . . . , Rn be non-zero ideals of
R. Then the following statements are equivalent:
(a) R R = R1 ⊕ . . . ⊕ Rn ;
(b) As an abelian group R is the direct sum of R1 , . . . , Rn ;
(c) There exist pairwise orthogonal central idempotents e1 , . . . , en with 1 =
e1 + . . . + en , and Ri = Rei , i = 1, . . . , n.
Lemma 2.3. (See [12, Theorem 21.28].) Let I be a nil ideal in R and u ∈ R be
such that u + I is an idempotent in R/I. Then there exists an idempotent e in uR
such that e − u ∈ I.
Lemma 2.4. (See [8, Lemma 2.4].) Let N be a minimal submodule of M and let
Ann(M ) be a nil ideal. Then we have N 2 = (0) or N = eM for some idempotent
e ∈ R.
Proposition 2.5. Let R/Ann(M ) be an Artinian ring and let M be a finitely
generated module. Then every nonzero proper submodule N of M is a vertex in
AG(M ).
Proof. Let N be a non-zero submodule of M . So there exists a maximal submodule
K of M such that N ⊆ K. Hence we have (0 :M (K : M )) ⊆ (0 :M (N : M )).
Since R/Ann(M ) is an Artinian ring, (K : M ) is a minimal prime ideal containing
Ann(M ). Thus (K : M ) ∈ Ass(M ). It follows that (K : M ) = (0 : m) for some
0 6= m ∈ M . Therefore N (Rm) = (0), as desired.
Lemma 2.6. Let M = M1 ×M2 , where M1 = eM , M2 = (1−e)M , and e (e 6= 0, 1)
is an idempotent element of R. If AG(M ) is a triangle-free graph, then one of the
following statements holds.
(a) Both M1 and M2 are prime R-modules.
4
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
(b) One Mi is a prime module for i = 1, 2 and the other one is a module with
a unique non-trivial submodule.
Moreover, AG(M ) has no cycle if and only if either M = F × S or M = F × D,
where F is a simple module, S is a module with a unique non-trivial submodule,
and D is a prime module.
Proof. If none of M1 and M2 is a prime module, then there exist r ∈ Ri (R1 = Re
and R2 = R(1 − e)), 0 6= mi ∈ Mi with ri mi = 0, and ri ∈
/ AnnRi (Mi ) for
i = 1, 2. So r1 M1 × (0), (0) × r2 M2 , and R1 m1 × R2 m2 form a triangle in AG(M ), a
contradiction. Thus without loss of generality, one can assume that M1 is a prime
module. We prove that AG(M2 ) has at most one vertex. On the contrary suppose
that {N, K} is an edge of AG(M2 ). Therefore, M1 × (0), (0) × N , and (0) × K form
a triangle, a contradiction. If AG(M2 ) has no vertex, then M2 is a prime module
and so part (a) occurs. If AG(M2 ) has exactly one vertex, then by [7, Theorem 3.6]
and Proposition 2.5, we obtain part (b). Now, suppose that AG(M ) has no cycle.
If none of M1 and M2 is a simple module, then choose non-trivial submodules Ni
in Mi for some i = 1, 2. So N1 × (0), (0) × N2, M1 × (0), and (0) × M2 form a cycle,
a contradiction. The converse is trivial.
Theorem 2.7. If AG(M ) is a tree, then either AG(M ) is a star graph or AG(M ) ∼
=
P4 . Moreover, AG(M ) ∼
= P4 if and only if M = F × S, where F is a simple module
and S is a module with a unique non-trivial submodule.
Proof. If M is a vertex of AG(M ), then there exists only one vertex N such that
Ann(M ) = (N : M ) and since AG(M )∗ is an empty subgraph, hence AG(M )
is a star graph. Therefore we may assume that M is not a vertex of AG(M ).
Suppose that AG(M ) is not a star graph. Then AG(M ) has at least four vertices. Obviously, there are two adjacent vertices N and K of AG(M ) such that
|V (N ) \ {K}| ≥ 1 and |V (K) \ {N }| ≥ 1. Let V (N ) \ {K} = {Ni }i∈Λ and
V (K) \ {N } = {Kj }j∈Γ . Since AG(M ) is a tree, we have V (N ) ∩ V (K) = ∅.
By [7, Theorem 3.4], diam(AG(M )) ≤ 3. So every edge of AG(M ) is of the form
{N, K}, {N, Ni } or {K, Kj }, for some i ∈ Λ and j ∈ Γ. Now, consider the following
claims:
Claim 1. Either N 2 = (0) or K 2 = (0). Pick p ∈ Λ and q ∈ Γ. Since AG(M )
is a tree, Np Kq is a vertex of AG(M ). If Np Kq = Nu , for some u ∈ Λ, then
KNu = (0), a contradiction. If Np Kq = Kv , for some v ∈ Γ, then N Kv = (0), a
contradiction. If Np Kq = N or Np Kq = K, then N 2 = (0) or K 2 = (0), respectively and the claim is proved.
Here, without loss of generality, we suppose that N 2 = (0). Clearly, (N : M )M * K
and (K : M )M * N .
Claim 2. Our claim is to show that N is a minimal submodule of M and K 2 6= (0).
To see that, first we show that for every 0 6= m ∈ N , Rm = N . Assume that
0 6= m ∈ N and Rm 6= N . If Rm = K, then K ⊆ N , a contradiction. Thus
Rm 6= K, and the induced subgraph of AG(M ) on N , K, and Rm is K3 , a
contradiction. So Rm = N . This implies that N is a minimal submodule of
M . Now if K 2 = (0), then we obtain the induced subgraph on N , K, and
(N : M )M + (K : M )M is K3 , a contradiction. Thus K 2 6= (0), as desired.
THE ANNIHILATING-SUBMODULE GRAPH OF MODULES
5
Claim 3. For every i ∈ Λ and every j ∈ Γ, Ni ∩ Kj = N . Let i ∈ Λ and
j ∈ Γ. Since Ni ∩ Kj is a vertex and N (Ni ∩ Kj ) = K(Ni ∩ Kj ) = (0), either
Ni ∩ Kj = N or Ni ∩ Kj = K. If Ni ∩ Kj = K, then K 2 = (0), a contradiction.
Hence Ni ∩ Kj = N and the claim is proved.
Claim 4. We complete the claim by showing that M has exactly two minimal
submodules N and K. Let L be a non-zero submodule properly contained in K.
Since N L ⊆ N K = (0), either L = N or L = Ni for some i ∈ Λ. So by the Claim
3, N ⊆ L ⊆ K, a contradiction. Hence K is a minimal submodule of M . Suppose
that L′ is another minimal submodule of M . Since N and K both are minimal submodules, we deduce that N L′ = KL′ = (0), a contradiction. So the claim is proved.
Now by the Claims 2 and 4, K 2 6= (0) and K is a minimal submodule of M . Then by
Lemma 2.4, K = eM for some idempotent e ∈ R. Now we have M ∼
= eM ×(1−e)M .
By Lemma 2.6, we deduce that either M = F × S and AG(M ) ∼
= P4 or R = F × D
and AG(M ) is a star graph. Conversely, we assume that M = F × S. Then AG(M )
has exactly four vertices (0) × S, F × (0), (0) × N , and F × N . Thus AG(M ) ∼
= P4
with the vertices (0) × S, F × (0), (0) × N , and F × N .
Theorem 2.8. Let R be an Artinian ring and AG(M ) is a bipartite graph. Then
either AG(M ) is a star graph or AG(M ) ∼
= P4 . Moreover, AG(M ) ∼
= P4 if and
only if M = F × S, where F is a simple module and S is a module with a unique
non-trivial submodule.
Proof. First suppose that R is not a local ring. Hence by [9, Theorem 8.9], R =
R1 × . . . × Rn , where Ri is an Artinian local ring for i = 1, . . . , n. By Lemma 2.2
and Proposition 2.1, since AG(M ) is a bipartite graph, we have n = 2 and hence
M∼
= M1 × M2 . If M1 is a prime module, then it is easy to see that M1 is a vector
space over R/Ann(M1 ) and so is a semisimple R-module. Hence by Lemma 2.6
and Theorem 2.7, we deduce that either AG(M ) is isomorphic to P2 or P4 . Now we
assume that R is an Artinian local ring. Let m be the unique maximal ideal of R
and k be a natural number such that mk M = 0 and mk−1 M 6= 0. Clearly, mk−1 M
is adjacent to every other vertex of AG(M ) and so AG(M ) is a star graph.
Proposition 2.9. Assume that Ann(M ) is a nil ideal of R.
(a) If AG(M ) is a finite bipartite graph, then either AG(M ) is a star graph or
AG(M ) ∼
= P4 .
(b) If AG(M ) is a regular graph of finite degree, then AG(M ) is a complete
graph.
Proof. (a). If M is a vertex of AG(M ), then AG(M ) has only one vertex N
such that Ann(M ) = (N : M ) and since AG(M )∗ is an empty subgraph, AG(M )
is a star graph. Thus we may assume that M is not a vertex of AG(M ) and
hence by [7, Theorem 3.3], M is not a prime module. Therefore [7, Theorem 3.6]
follows that R/Ann(M ) is an Artinian ring. If (R/Ann(M ), m/Ann(M )) is a local
ring, then there exists a natural number k such that mk M = 0 and mk−1 M 6= 0.
Clearly, mk−1 M is adjacent to every other vertex of AG(M ) and so AG(M ) is a
star graph. Otherwise, by [9, Theorem 8.9] and Lemma 2.2, there exist pairwise
orthogonal central idempotents modulo Ann(M ). By Lemma 2.3, it is easy to see
6
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
that M ∼
= eM × (1 − e)M , where e is an idempotent element of R and Lemma 2.6
implies that AG(M ) is a star graph or AG(M ) ∼
= P4 .
(b). If M is a vertex of AG(M ), since AG(M ) is a regular graph, then AG(M ) is a
complete graph. Hence we may assume that M is not a vertex of AG(M ). So M is
not a prime module, and hence rm = 0 such that 0 6= m ∈ M , r ∈
/ Ann(M ). It is
easy to see that (rM )(0 :M r) = (0). If the set of R-submodules of rM (resp., (0 :M
r))) is infinite, then (0 :M r) (resp., rM ) has infinite degree, a contradiction. Thus
rM and (0 :M r) have finite length. Since rM ∼
= M/(0 :M r), M has finite length so
that R/Ann(M ) is an Artinian ring. As in the proof of part (a), M ∼
= M1 × M2 . If
M1 has one non-trivial submodule N , then deg((0) × M2 ) > deg(N × M2 ) and this
contradicts the regularity of AG(M ). Hence, M1 is a simple module. Similarly, M2
is a simple module. So AG(M ) ∼
= K2 . Now suppose that (R/Ann(M ), m/Ann(M ))
is an Artinian local ring. Now as we have seen in part (a), there exists a natural
number k such that mk−1 M is adjacent to all other vertices and we deduce that
AG(M ) is a complete graph.
Let S be a multiplicatively closed subset of R. A non-empty subset S ∗ of M is
said to be S-closed if se ∈ S ∗ for every s ∈ S and e ∈ S ∗ . An S-closed subset S ∗
is said to be saturated if the following condition is satisfied: whenever ae ∈ S ∗ for
a ∈ R and e ∈ M , then a ∈ S and e ∈ S ∗ .
We need the following result due to Chin-Pi Lu.
Theorem 2.10. (See [16, Theorem 4.7].) Let M = Rm be a cyclic module. Let S ∗
be an S-closed subset of M relative to a multiplicatively closed subset S of R, and
N a submodule of M maximal in M \ S ∗ . If S ∗ is saturated, then ideal (N : M ) is
maximal in R \ S so that N is prime in M .
Theorem 2.11. If M is a cyclic module, Ann(M ) is a nil ideal, and |M in(M )| ≥
3, then AG(M ) contains a cycle.
Proof. If AG(M ) is a tree, then by Theorem 2.7, either AG(M ) is a star graph or
M∼
= F × S, where F is a simple module and S has a unique non-trivial submodule.
The latter case is impossible because |M in(F × S)| = 2. Suppose that AG(M ) is a
star graph and N is the center of star. Clearly, one can assume that N is a minimal
submodule of M . If N 2 6= (0), then by Lemma 2.4, there exists an idempotent
e ∈ R such that N = eM so that M ∼
= eM × (1 − e)M . Now by Proposition 2.1
and Lemma 2.6, we conclude that |M in(M )| = 2, a contradiction. Hence N 2 = 0.
Thus one may assume that N = Rm and (Rm)2 = (0). Suppose that P1 and P2
are two distinct minimal prime submodules of M . Since (Rm)2 = (0), we have
(Rm : M )2 ⊆ Ann(M ) ⊆ (Pi : M ), i = 1, 2. So (Rm : M )M = Rm ⊆ Pi , i = 1, 2.
Hence m ∈ Pi , i = 1, 2. Choose z ∈ (P1 : M ) \ (P2 : M ) and set S1 = {1, z, z 2, . . .},
S2 = M \ P1 , and S ∗ = S1 S2 . If 0 ∈
/ S ∗ , then Σ = {N < M | N ∩ S ∗ = ∅} is
not empty. Then Σ has a maximal element, say N . Hence by Theorem 2.10, N
is a prime submodule of M . Since N ⊆ P1 , we have N = P1 , a contradiction
because z ∈
/ (N : M ). So 0 ∈ S ∗ . Therefore, there exist positive integer k and
′
m ∈ S2 such that z k m′ = 0. Now consider the submodules (m), (m′ ), and z k M .
It is clear that (m) 6= (m′ ) and (m) 6= z k M . If (m) = z k M , then z ∈ (P2 : M ), a
contradiction. Thus (m), (m′ ), and z k M form a triangle in AG(M ), a contradiction.
Hence AG(M ) contains a cycle.
THE ANNIHILATING-SUBMODULE GRAPH OF MODULES
7
Theorem 2.12. Suppose that M is a cyclic module, radM (0) 6= (0), and Ann(M )
is a nil ideal. If |M in(M )| = 2, then either AG(M ) contains a cycle or AG(M ) ∼
=
P4 .
Proof. A similar argument to the proof of Theorem 2.11 shows that either AG(M )
contains a cycle or M ∼
= F × S, where F is a simple module and S is a module with
a unique non-trivial submodule. The latter case implies that AG(M ) ∼
= P4 (note
that radF ×D (0) = (0), where F is a simple module and D is a prime module).
We recall that N < M is said to be a semiprime submodule of M if for every
ideal I of R and every submodule K of M , I 2 K ⊆ N implies that IK ⊆ N .
Further M is called a semiprime module if (0) ⊆ M is a semiprime submodule.
Every intersection of prime submodules is a semiprime submodule (see [21]).
Theorem 2.13. Let S be a maltiplicatively closed subset of R containing no zerodivisors on finitely generated module M . Then cl(AG(MS )) ≤ cl(AG(M )). Moreover, AG(MS ) is a retract of AG(M ) if M is a semiprime module. In particular,
cl(AG(MS )) = cl(AG(M )), whenever M is a semiprime module.
Proof. Consider a vertex map φ : V (AG(M )) −→ V (AG(MS )), N −→ NS . Clearly,
NS 6= KS implies N 6= K and N K = (0) if and only if NS KS = (0). Thus φ
is surjective and hence cl(AG(MS )) ≤ cl(AG(M )). In what follows, we assume
that M a semiprime module. If N 6= K and N K = (0), then we show that
NS 6= KS . Without loss of generality we can assume that M is not a vertex of
AG(M ) and On the contrary suppose that NS = KS . Then NS2 = NS KS =
(N K)S = (0) and so N 2 = (0), a contradiction. This shows that the map φ is
a graph homomorphism. Now, for any vertex NS of AG(MS ), we can choice the
fixed vertex N of AG(M ). Then φ is a retract (graph) homomorphism which clearly
implies that cl(AG(MS )) = cl(AG(M )) under the assumption.
Corollary 2.14. If M is a finitely generated semiprime module, then cl(AG(T (M )) =
cl(AG(M )), where T = R \ Z(M ).
Since the chromatic number χ(G) of a graph G is the least positive integer r such
that there exists a retract homomorphism ψ : G −→ Kr , the following corollaries
follow directly from the proof of Theorem 2.13.
Corollary 2.15. Let S be a maltiplicatively closed subset of R containing no
zero-divisors on finitely generated module M . Then χ(AG(MS )) ≤ χ(AG(M )).
Moreover, if M is a semiprime module, then χ(AG(MS )) = χ(AG(M )).
Corollary 2.16. If M is a finitely generated semiprime module, then χ(AG(T (M )) =
χ(AG(M )), where T = R \ Z(M ).
Eben Matlis in [18, Proposition 1.5], proved that if {p1 , . . . , pn } is a finite set of
distinct minimal prime ideals of R and S = R \ ∪ni=1 pi , then Rp1 × . . . × Rpn ∼
= RS .
In [20], this result was generalized to finitely generated multiplication modules. In
Theorem 2.18, we use this generalization for a cyclic module.
Theorem 2.17. (See [20, Theorem 3.11].) Let {P1 , . . . , Pn } be a finite set of distinct minimal prime submodules of finitely generated multiplication module M and
S = R \ ∪ni=1 (Pi : M ). Then Mp1 × . . . × Mpn ∼
= MS , where pi = (Pi : M ) for
1 ≤ i ≤ n.
8
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
Theorem 2.18. Let M be a cyclic module and {P1 , . . . , Pn } be a finite set of
distinct minimal prime submodules of M . Then there exists a clique of size n.
Proof. Let M be a cyclic module and S = R \ ∪ni=1 pi , where pi = (Pi : M )
for 1 ≤ i ≤ n. Then since M is a multiplication module, by Theorem 2.17,
there exists an isomorphism φ : Mp1 × . . . × Mpn −→ MS . Let M = Rm, ei =
(0, . . . , 0, m/1, . . . , 0, . . . , 0) and φ(ei ) = ni /ti , where m ∈ M , 1 ≤ i ≤ n, and m/1 is
in the i-th position of ei . Consider the principal submodules Ni = (ni /ti ) = (ni /1)
in the module MS . By Lemma 2.2 and Proposition 2.1, the product of submodules
(0)×. . .×(0)×(m/1)Rpi ×(0)×. . .×(0) and (0)×. . .×(0)×(m/1)Rpj ×(0)×. . .×(0)
are zero, i 6= j. Since φ is an isomorphism, there exists tij ∈ S such that tij ri nj = 0,
for every i, j, 1 ≤ i < j ≤ n, where ni = ri m for some ri ∈ R. Let t = Π1≤i<j≤n tij .
We show that {(tn1 ), . . . , (tnn )} is a clique of size n in AG(M ). For every i, j, 1 ≤
i < j ≤ n, (Rtni )(Rtnj ) = (Rtnj : M )Rtni = (Rtnj : M )tri M = tri Rtnj = (0).
Since (tni )S = (ni /1) = Ni , we deduce that (tni ) are distinct non-trivial submodules of M .
Corollary 2.19. For every cyclic module M , cl(AG(M )) ≥ |M in(M )| and if
|M in(M )| ≥ 3, then gr(AG(M )) = 3.
Theorem 2.20. Let M be a cyclic module and radM (0) = (0). Then χ(AG(M )) =
cl(AG(M )) = |M in(M )|.
Proof. If |M in(M )| = ∞, then by Corollary 2.19, there is nothing to prove. Thus
suppose that |M in(M )| = {P1 , ..., Pn }, for some positive integer n. Let pi = (Pi :
M ) and S = R \ ∪ni=1 pi . By Theorem 2.17, we have Mp1 × . . . × Mpn ∼
= MS .
Clearly, cl(AG(MS )) ≥ n. Now we show that χ(AG(MS )) ≤ n. By [15, Corollary
3], Pi Rpi is the only prime submodule of M and since radM (0) = (0), every Mpi
is a simple Rpi -module. Define the map C : V (AG(MS )) −→ {1, 2, . . . , n} by
C(N1 × . . . × Nn ) = min{i| Ni 6= (0)}. Since each Mpi is a simple module, c is a
proper vertex coloring of AG(MS ). Thus χ(AG(MS )) ≤ n and so χ(AG(MS )) =
cl(AG(MS )) = n. Since radM (0) = (0), it is easy to see that S ∩ Z(M ) = ∅. Now
by theorem 2.13 and Corollary 2.15, we obtain the desired.
Theorem 2.21. For every module M , cl(AG(M )) = 2 if and only if χ(AG(M )) =
2.
Proof. For the first assertion, we use the same technique in [3, Theorem 13]. Let
cl(AG(M )) = 2. On the contrary assume that AG(M ) is not bipartite. So AG(M )
contains an odd cycle. Suppose that C := N1 − N2 − . . . − N2k+1 − N1 be a shortest
odd cycle in AG(M ) for some natural number k. Clearly, k ≥ 2. Since C is a
shortest odd cycle in AG(M ), N3 N2k+1 is a vertex. Now consider the vertices
N1 , N2 , and N3 N2k+1 . If N1 = N3 N2k+1 , then N4 N1 = (0). This implies that
N1 − N4 − . . . − N2k+1 − N1 is an odd cycle, a contradiction. Thus N1 6= N3 N2k+1 .
If N2 = N3 N2k+1 , then we have C3 = N2 − N3 - N4 − N2 , again a contradiction.
Hence N2 6= N3 N2k+1 . It is easy to check N1 , N2 , and N3 N2k+1 form a triangle in
AG(M ), a contradiction. The converse is clear.
The radical
√ of I, defined as the intersection of all prime ideals containing I,
the next theorem, we recall that if M is a finitely
denoted by I. Before stating
p
generated module, then (Q : M ) = (rad(Q) : M ), where Q < M (see [14] and
[17, Proposition 2.3]). Also, we know that if M is a finitely generated module, then
THE ANNIHILATING-SUBMODULE GRAPH OF MODULES
9
for every prime ideal p of R with p ⊇ Ann(M ), there exists a prime submodule P
of M such that (P : M ) = p (see [15, Theorem 2]).
Theorem 2.22. Assume that M is a finitely generated module, Ann(M ) is a nil
ideal, and |M in(M )| = 1. If AG(M ) is a triangle-free graph, then AG(M ) is a star
graph.
Proof. Suppose first that P is the unique minimal prime submodule of M . Since
M is not a vertex of AG(M ), hence Z(M ) 6= (0). So there exist non-zero elements
r ∈ R and m ∈ M such that rm = 0. It is easy to see that rM and Rm are
vertices of AG(M ) because (rM )(Rm) = 0. Since AG(M ) is triangle-free, Rm
or rM is a minimal submodule of M . Without loss of generality, we can assume
that Rm is a minimal submodule of M so that (Rm)2 = (0) (if rM is a minimal
submodule of M , then there exists 0 6= m′ ∈ M such that rM = Rm′ ). We
claim that Rm is the unique minimal submodule of M . On the contrary, suppose
that K is another minimal submodule of M . So either K 2 = K or K 2 = 0. If
K 2 = K, then by Lemma 2.4, K = eM for some idempotent element e ∈ R and
hence M ∼
= eM × (1 − e)M . This implies that |M in(M )| > 1, a contradiction.
If K 2 = 0, then we have C3 = K − (K : M )M + (Rm : M )M − Rm − K, a
contradiction. So Rm is the unique minimal submodule of M . Let V1 = N ((Rm)),
V2 = V (AG(M )) \ V1 , A = {K ∈ V1 |(Rm) ⊆ K}, B = V1 \ A, and C = V2 \ {Rm}.
We prove that AG(M ) is a bipartite graph with parts V1 and V2 . We may assume
that V1 is an independent set because AG(M ) is triangle-free. We claim that one
end of every edge of AG(M ) is adjacent to Rm and another end contains Rm. To
prove this, suppose that {N, K} is an edge of AG(M ) and Rm 6= N , Rm 6= K.
Since N (Rm) ⊆ Rm, by the minimality of Rm, either N (Rm) = (0) or Rm ⊆ N .
The latter case follows that K(Rm) = (0). If N (Rm) = (0), then K(Rm) 6= (0) and
hence Rm ⊆ K. So our plain is proved. This gives that V2 is an independent set
and N (C) ⊆ V1 . Since every vertex of A contains Rm and AG(M ) is triangle-free,
all vertices in A are just adjacent to Rm and so by [7, Theorem 3.4], N (C) ⊆ B.
Since one end of every edge is adjacent to Rm and another end contains Rm, we also
deduce that every vertex of C contains Rm and so every vertex of A ∪ V2 contains
Rm. Note that if Rm = P , then one end of each edge of AG(M ) is contained in Rm
and since Rm is a minimal submodule of M , AG(M ) is a star graph with center
Rm = P . Now, suppose that P 6= Rm. We claim that P ∈ A. Since Rm ⊆ P , it
suffices to show that (Rm)P = (0). To see this, let r ∈ (P : M ). We prove that
rm = 0. Clearly, (Rrm) ⊆ Rm. If rm = 0, then we are done. Thus Rrm = Rm
and so m = rsm for some s ∈ R. We have
p m(1 − rs) = 0. By [15, Theorem 2],
we have N il(R) = (P : M ) (note that Ann(M ) = (rad(0) : M ) = (P : M )).
Therefore 1 − rs is unit, a contradiction, as required. Since N (C) ⊆ B, if B = ∅,
then C = ∅ and so AG(M ) is a star graph with center Rm. It remains to show
that B = ∅. Suppose that K ∈ B and consider the vertex K ∩ P of AG(M ). Since
every vertex of A ∪ V2 contains Rm, yields K ∩ P ∈ B. Pick 0 6= m′ ∈ K ∩ P .
Since AG(M ) is triangle-free, one can find an element m′′ ∈ Rm′ such that Rm′′
is a minimal submodule of M and (Rm′′ )2 = (0). Since Rm is the unique minimal
submodule of M , we have Rm = Rm′′ ⊆ Rm′ . Thus Rm ⊆ K ∩ P , a contradiction.
So B = ∅ and we are done. Hence AG(M ) is a star graph whose center is Rm, as
desired.
10
HABIBOLLAH ANSARI-TOROGHY AND SHOKOUFEH HABIBI
Corollary 2.23. Assume that M is a finitely generated module, Ann(M ) is a nil
ideal, and |M in(M )| = 1. If AG(M ) is a bipartite graph, then AG(M ) is a star
graph.
References
[1] G. Aalipour, S. Akbari, R. Nikandish, M. J. Nikmehr, and F. Shaveisi, On the coloring
of the annihilating-ideal graph of a commutative ring. Discrete Mathematics. 312 (2012),
2620–2626.
, Minimal prime ideals and cycles in annihilating-ideal graphs. Rocky Mountain J.
[2]
Math. Vol. 43 (2013) No. 5, 1415–1425.
[3] G. Aalipour, S. Akbari, M. Behboodi, R. Nikandish, M. J. Nikmehr and F. Shaveisi, The
Classication of the Annihilating-Ideal Graphs of Commutative Rings. Algebra Colloquium.
21 (2) (2014), 249–256.
[4] D. F. Anderson and P. S. Livingston, The zero-divisor graph of a commutative ring. J.
Algebra, Springer, NJ. 217 (1999), 434–447.
[5] W. Anderson, K. R. Fuller, Rings and Categories of Modules, (New York-Heidelberg-Berlin,
Springer-Verlag, 1974).
[6] H. Ansari-Toroghy and F. Farshadifar, Product and dual product of submodules. Far East
J. Math. Sci. 25 (3) (2007), 447–455.
[7] H. Ansari-Toroghy and S. Habibi, The Zariski topology-graph of modules over commutative
rings, Comm. Algebra. 42 (2014), 3283–3296.
, The annihilating-submodule graph of modules over commutative rings, arXiv: 1601
[8]
.00916v1, submitted.
[9] M.F. Atiyah and I.G. Macdonald, Introduction to commutative algebra. Addison-Wesley.
(1969).
[10] I. Beck, Coloring of commutative rings. J. Algebra., 116 (1988), 208–226.
[11] M. Behboodi and Z. Rakeei, The annihilating-ideal graph of commutative rings I. J. Algebra
Appl. Vol. 10 (2011), No. 4, 727–739.
[12] T. Y. Lam, A First Course in Non-Commutative Rings. Springer - Verlag. (New York,
1991).
[13] Chin-Pi. Lu, Prime submodules of modules. Comment. Math. Univ. St. Pauli. 33 (1984),
no. 1, 61–69.
, M-radicals of submodules in modules II. Mathematica Japonica. 35 (5) (1990),
[14]
991–1001.
[15]
, Spectra of modules. Comm. Algebra. 23 (10) (1995), 3741–3752.
[16]
, Unions of prime submodules. Houston Journal of Mathematics., 23 (1997), no. 2,
203–213.
[17]
, Modules with Noetherian Spectrum. Comm. Algebra. 38 (2010), 807–828.
[18] E. Matlis, The minimal prime spectrum of a reduced ring. Illinois J. Math. 27 (3) (1983),
353–391.
[19] D. Reinard, Graph Theory. Grad, Texts in Math, Springer, NJ. (2005).
[20] K. Samei, Reduced multiplication modules. Math. Sci. 121 (2011), no.2, 121–132.
[21] H. A. Tavallaee and R. Varmazyar, Semi-radicals of submodules in modules, IUST International Journal of Engineering Science. 19 (2008), 21–27.
Department of pure Mathematics
Faculty of mathematical Sciences,
University of Guilan, P. O. Box 41335-19141, Rasht, Iran
E-mail: [email protected]
Department of pure Mathematics
Faculty of mathematical Sciences,
University of Guilan, P. O. Box 41335-19141, Rasht, Iran
E-mail: [email protected]
| 0math.AC
|
arXiv:1503.05436v4 [math.ST] 22 Feb 2018
Inference in Additively Separable Models with a
High-Dimensional Set of Conditioning Variables
Damian Kozbur
University of Zürich
Department of Economics
Schönberggasse 1, 8001 Zürich
email: [email protected]
Abstract. This paper studies nonparametric series estimation and inference
for the effect of a single variable of interest x on an outcome y in the presence of potentially high-dimensional conditioning variables z. The context is
an additively separable model E[y|x, z] = g0 (x) + h0 (z). The model is highdimensional in the sense that the series of approximating functions for h0 (z)
can have more terms than the sample size, thereby allowing z to have potentially very many measured characteristics. The model is required to be
approximately sparse: h0 (z) can be approximated using only a small subset
of series terms whose identities are unknown. This paper proposes an estimation and inference method for g0 (x) called Post-Nonparametric Double
Selection which is a generalization of Post-Double Selection. Standard rates
of convergence and asymptotic normality for the estimator are shown to hold
uniformly over a large class of sparse data generating processes. A simulation
study illustrates finite sample estimation properties of the proposed estimator
and coverage properties of the corresponding confidence intervals. Finally, an
empirical application estimating convergence in GDP in a country-level crosssection demonstrates the practical implementation of the proposed method.
Key Words: additive nonparametric models, high-dimensional sparse regression, inference under imperfect model selection. JEL Codes: C1.
1. Introduction
Nonparametric estimation in econometrics and statistics is useful in applications
where theory does not provide functional forms for relations between relevant observed variables. In many problems, primary quantities of interest can be computed
from the conditional expectation function of an outcome variable y given a regressor
of interest x,
E[y|x] = f0 (x).
Date: First version: September 2013. This version is of February 23, 2018.
Correspondence: Schönberggasse 1, 8001 Zürich, Department of Economics, University of Zürich,
[email protected].
I thank Christian Hansen, Tim Conley, Matt Taddy, Azeem Shaikh, Dan Nguyen, Dan Zhou,
Emily Oster, Martin Schonger, Eric Floyd, Kelly Reeve, and seminar participants at University
of Western Ontario, University of Pennsylvania, Rutgers University, Monash University, and the
Center for Law and Economics at ETH Zurich for helpful comments. I gratefully acknowledge
financial support from the ETH Postdoctoral Fellowship.
1
2
DAMIAN KOZBUR
In this case, nonparametric estimation is a flexible means for estimating unknown
f0 from data under minimal assumptions.
In most econometric models, however, it is also important to take into account
conditioning information, given through variables z. Failing to properly control for
such variables z will lead to incorrect estimates of the effects of x on y. When such
conditioning information is important to the problem, it is necessary to replace the
simple objective of learning the conditional expectation function f0 with the new
objective of learning a family of conditional expectation functions
E[y|x, z] = f0,z (x)
indexed by z.
This paper studies series estimation1 and inference of f0,z in a particular case
characterized by the following two main features.
1. f0,z is additively separable in x and z, meaning that
f0,z (x) = g0 (x) + h0 (z)
for some functions g0 and h0 .
2. The conditioning variables z are observable and high-dimensional.
Additively separable models are convenient in many economic problems because
any ceteris paribus effect of changing x to x0 is completely described by g0 . In
addition, a major statistical advantage in restricting to additively separable models
is that the individual components g0 , h0 can be estimated at faster rates than a
joint estimation of the family f0,z .2 Therefore, imposing additive separability in
contexts where such an assuption is justified is very helpful.
The motivation for studying a high-dimensional framework for z is to allow
researchers substantial flexibility in modeling conditioning information when the
primary object of interest is g0 . This framework allows analysis of particularly
rich or big datasets with a large number of conditioning variables.3 In this paper,
high-dimensionality of z is formally defined by the total number of terms in a series
expansion of h0 (z). This will allow many possibilities on the types of variables z
and functions h0 covered. For example, z can be high-dimensional itself, while h0
is approximately linear in the sense that
0
h0 (z) = z10 βh0 ,1 + ... + zL
βh0 ,L + o(1)
with L > n and βh0 ,j denoting the jth component of the vector βh0 and the
asymptotic o(1) valid for L → ∞. More generally, z itself can also have moderate
1Series estimation of nonparametric regression problems involves least squares estimation performed on a series expansion of the regressor variables. Series estimation is described more fully
in Section 2.
2Results on faster rates for separable models exist for both kernel methods (marginal integration and back-fitting methods) and series based estimators. For a general review of these issues,
see for example the textbook [41]. Additional discussion on the literature on additively separable
models is provided later in the introduction.
3In many cases, larger set of covariates can lend additional credibility to conditional exogeneity
assumptions. See the discussion in [15].
ADDITIVELY SEPARABLE
3
dimension, but any sufficiently expressive series expansion of h0 must have many
terms as a simple consequence of the curse of dimensionality.
A basic mechanical outline for the estimation and inference strategy presented
in this paper proceeds in the following steps.
1. Consider approximating dictionaries (equivalently series expansions) with
K terms, given by pK (x) = (p1K (x), ..., pKK (x)). Linear combinations of
pK (x) are used for approximating g0 (x). In addition, consider approximating dictionaries with L terms, q L (z) = (q1L (z), ..., qLL (z)), for approximating h0 (z). Possibly L > n.
2. Reduce the number of series terms for h0 in a way which continues to allow
robust inference. This requires multiple model selection steps.
3. Proceed with traditional series estimation and inference techniques on the
reduced dictionaries.
Strategies of this form are commonly referred to as post-model selection inference
strategies.
The primary targets of inference considered in this paper are real-valued functionals, g 7→ a(g) ∈ R. Specifically, let
θ0 = a(g0 ).
Leading examples of such functionals include the average derivative a(g) = E[g 0 (x)]
or the difference of a(g) = g(x20 ) − g(x10 ) for two distinct x10 , x20 of interest.
The main contribution of this paper is the construction of confidence sets that
cover θ0 to some pre-specified confidence level. Moreover, the construction is valid
uniformly over a large class of data generating processes which allow z to be highdimensional.
Current high-dimensional estimation techniques provide researchers with useful
tools for dimension reduction and dealing with datasets where the number of parameters exceeds the sample size.4 Most high-dimensional techniques require additional
structure to be imposed on the problem at hand in order to ensure good performance. One common structure for which reliable high-dimensional techniques exist
is sparsity. Sparsity means that the number of nonzero parameters is small relative
to the sample size. In this setting, common techniques include `1 -regularization
techniques like Lasso and Post-Lasso5. Other techniques include the Dantzig selector, Scad, and Forward Stepwise regression.
The literature on nonparametric estimation of additively separable models is well
developed. As mentioned above, additively separable models are useful since they
4Statistical models which are extremely flexible, and thus overparameterized, are likely to
overfit the data, leading to poor inference and out of sample performance. Therefore, when many
covariates are present, regularization is necessary.
5The Lasso is a shrinkage procedure which estimates regression coefficients by minimizing a
quadratic loss function plus an `1 penalty for the size of the coefficient. The nature of the penalty
gives Lasso favorable property that many parameter values are set identically to zero and thus
Lasso can also be used as a model selection technique. Post-Lasso fits an ordinary least squares
regression on variables with non-identically-zero estimated Lasso coefficients. For theoretical and
simulation results about the performance of these two methods, see [29] [52], [32] [23] [4], [5], [17],
[21], [20] [22], [23], [33], [38], [39], [42], [43], [44], [47], [52], [53], [55], [57], [9], [18], [9], among many
more.
4
DAMIAN KOZBUR
impose an intuitive restriction on the class of models considered, and as a result
provide higher quality estimates. Early study of additively separable models was
initiated in [19] and [31], who describe backfitting techniques. [25] propose marginal integration methods in the kernel context. [50] and [56] consider estimation
of derivatives in components of additive models. [26] develop local partitioned regression which can be applied more generally than the additive model. In terms of
series-based estimation, series estimators are particularly easy to use for estimating
additively separable models since series terms can be allocated to respective model
components. General large sample properties of series estimators have been derived
by [51], [27], [2], [28], [3] [45] [14] and many other references. Relative to kernel
estimation, series estimators are simpler to implement, but often require stronger
support conditions. Many additional references for both kernel and series based
estimation can be found in the reference text [41]. Finally, [34] consider estimation
of additively separable models in a setting where there are high-dimensional additive components. The authors propose and analyze a series estimation approach
with a Group-Lasso penalty to penalize different additive components. This paper therefore studies a very similar setting to the one in [34], but constructs a valid
procedure for forming confidence intervals rather than focusing on estimation error.
The main challenge in statistical inference or construction of confidence intervals
after model selection is in attaining robustness to model selection errors. When coefficients are small relative to the sample size (ie statistically indistinguishable from
zero), model selection mistakes are unavoidable.6 Such model selection mistakes can
lead to distorted statistical inference in much the same way that pretesting procedures lead to distorted inference. This intuition is formally developed in [46]
and [40]. Nevertheless, given the practical value of dimension reduction, and the
increasing prevalence of high-dimensional datasets, studying robust post-model selection inference techniques and post-regularization inference techniques is an active
area of current research. Offering solutions to this problem is the focus of a number
of recent papers; see, for example, [11], [8], [58], [12], [15], [54], [36], [10], and [16].7
This paper proposes a procedure called Post-Nonparametric Double Selection
for the additively separable model. The proposed procedure is a generalization
of the approach in [15] (named Post-Double-Selection). [15] gives robust statistical
inference for the slope parameter α0 of a treatment variable x with high-dimensional
control variables z in the context of a partially linear model E[y|x, z] = α0 x+h0 (z).8
The Post-Double Selection method selects elements of z in two steps. Step 1 selects
the terms in an expansion of z that are most useful for predicting x. Step 2 selects
terms in an expansion of z most useful for predicting y. A consequence of the
particular construction using two selection steps is that terms excluded by model
selection mistakes twice necessarily have a negligible effect on subsequent statistical
inference.9 Post-Nonparametric Double Selection replaces step 1 of Post-Double
6
Under some restrictive conditions, for example beta-min conditions which constrain nonzero
coefficients to have large magnitudes, perfect model selection can be attained.
7
Citations are ordered by date of first appearance on arXiv.
8Several authors have addressed the task of assessing uncertainties or estimation error of model
parameter estimates in a wide variety of models with high dimensional regressors (see, for example,
[11], [8], [58], [12], [15], [54], [36], and [10]).
9The use of two model selection steps is motivated partially by the intuition that two necessary
conditions for omitted variables bias to occur: an omitted variable exists which is (1) correlated
with the treatment x, and (2) correlated with the outcome y. Each selection step addresses one
ADDITIVELY SEPARABLE
5
Selection with selecting variables useful for predicting any test function ϕ(x) ∈ ΦK
for a sufficiently general class of functions ΦK .
This paper suggests a simple choice for ΦK which is based on the linear span
of pK (x). This choice is called ΦK,Span .10 Theoretical and simulation results show
that the suggested choice has favorable statistical properties uniformly under certain
sequences of data generating processes.
Working with a generalization of Post-Double Selection which dissociates the
first stage selection from the final estimation is useful for several reasons. One
reason is that the direct extension of Post-Double is not invariant to the choice of
dictionary pK (x) and leads natural to the consideration of more general ΦK . In
addition, applying the direct generalization of Post-Double selection may lead to
poorer statistical performance than using a larger, more robust ΦK . A simulation
study later in this paper explores these properties. Next, as a theoretical advantage,
in some cases a larger ΦK gives estimates and inference which are valid under weaker
rate conditions on K, n, etc. Finally, working dissociating the first stage helps in
terms of organizing the arguments in the proofs. In particular, various bounds
developed in the proof depend on a notion of density of ΦK within LinSpan(pK ).
This paper proves convergence rates and asymptotic normality for PostNonparametric Double Selection estimates of g0 (x) and θ0 respectively. The proofs
in the paper proceed by using the techniques in Newey’s analysis of series estimators (see [45]) and ideas in Belloni, Chernozhukov, and Hansen’s analysis of
Post-Double Selection (see [15]), along with careful tracking of a notion of density
of the set ΦK within the linear span of pK (x). The estimation rates for g0 obtained
in this paper match those of [45]. Next, a simulation study demonstrates finite
sample performance of the proposed procedure. Finally, an empirical example estimating the relationship of initial GDP to GDP growth in a cross-section of 90
countries illustrates the use of Post-Nonparametric Double Selection.
2. Series estimation with a reduced dictionary
This section establishes notation, reviews series estimation, and describes series
estimation on a reduced dictionary. The exposition begins with basic assumptions
on the observed data.
Assumption 1 (Data). The observed data, Dn , is given by n iid copies of random
variables (x, y, z) ∈ X × Y × Z indexed by 1 6 i 6 n, so that
Dn = (yi , xi , zi )ni=1 .
Here, yi are outcome variables, xi are explanatory variables of interest, and zi are
conditioning variables. In addition, Y ⊆ R and X ⊆ Rr for some integer r > 0 and
Z is a general measure space.
of the two concerns. In their paper, they prove that under the regularity right conditions, the two
described model selection steps can be used to obtain asymptotically normal estimates of α0 and
in turn to construct correctly sized confidence intervals.
10Alternative choices are possible and the analysis in the paper covers a general class of choices
for ΦK .
6
DAMIAN KOZBUR
Assumption 2 (Additive Separability). There is a random variable ε and functions g0 and h0 such that the following additive separability11 holds.
y = g0 (x) + h0 (z) + ε, E[ε|x, z] = 0.
Traditional series estimation of (g0 , h0 ) is carried out by performing least squares
regression on series expansions in x and z. Define a dictionary of approximating
functions by
(pK (x), q L (z))
where pK (x) = (p1K (x), ..., pKK (x)) and q L (z) = (q1L (z), ..., qLL (z)) are each
series of K and L functions such that their linear combinations can approximate g0 (x) and h0 (z). Construct the matrices P = [pK (x1 ), ..., pK (xn )]0 , Q =
[q L (z1 ), ..., q L (zn )]0 ,Y = (y1 , ..., yn )0 , and let βby,(K,L) = ([P Q]0 [P Q])−1 [P Q]0 Y be
the least squares estimate from Y on [P Q]. Let [βby,(K,L) ]g be the components of
βby,(K,L) corresponding to pK . Then gb(x) is defined by
0
gb(x) = pK (x) [βby,(K,L) ]g .
When L > n, quality statistical estimation is only feasible provided dimension
reduction or regularization is performed. A dictionary reduction selects new approximating terms,
(pK (x), q L (z))
reduction
−→
(p̃(x), q̃(z)),
comprised of a subset of the series terms in (pK (x), q L (x)). In this paper, because the primary objects of interest center around g0 (x), it will be the convention to always take p̃(x) = pK (x). The post-model-selection estimate of g0 (x)
is then defined analogously to the traditional series estimate. Let βby,(p̃,q̃) =
([P̃ Q̃]0 [P̃ Q̃])−1 [P̃ Q̃]0 Y where P̃ = [p̃(x1 ), ..., p̃(xn )]0 = P , Q̃ = [q̃(z1 ), ..., q̃(zn )]0
and as before, let [βby,(p̃,q̃) ]g be the components of βby,(p̃,q̃) corresponding to p̃. Then
gb is defined by
0
gb(x) = p̃(x) [βby,(p̃,q̃) ]g .
Finally, consider a functional12 a(g) ∈ R and as before, set θ0 = a(g0 ). One
sensible estimate for θ0 is given by
θb = a(b
g ).
In order to use θb for inference on θ0 , an approximate expression for the varib is necessary. As is standard, the expression for the variance will
ance var(θ)
K
0
b = ∂a(p (x) b) ([βby,(p̃,q̃) ]g ). Let
be approximated using the delta method. Let A
∂b
M = Idn − Q̃(Q̃0 Q̃)−1 Q̃0 be the projection matrix onto the space orthogonal to the
11This assumption simply rewrites the equation stated in the introduction in terms of a residual
ε. To ensure uniqueness of g0 , a further normalization is required. A common normalization in
the series context is g0 (0) = 0, which is sufficient for most common assumptions on h0 .
12Restriction to one dimensional functionals is for simplicity.
7
ADDITIVELY SEPARABLE
b = Y − [P̃ , Q̃]βby,(p̃,q̃) . Estimate Vb using the following
span of Q̃.13 Finally, let E
sandwich form:
bΩ
b −1 Σ
bΩ
b −1 A
b
Vb = A
b = n−1 P̃ 0 MP̃
Ω
b 2 MP̃ .
b = n−1 P̃ 0 Mdiag(E)
Σ
The following sections describe a dictionary reduction technique, along with regularity conditions, which imply that
b →d N (0, 1).
n1/2 Vb −1/2 (θ0 − θ)
The practical value of the results is that they formally justify approximate Gaussian inference for θ0 . An immediate corollary of the Gaussian limit is that for any
significance level γ ∈ (0, 1), with c1−γ/2 the (1 − γ/2)-quantile of the standard
Guassian distribution, it holds that
P(θ0 ∈ [θb − c1−γ/2 n1/2 Vb −1/2 , θb − c1−γ/2 n1/2 Vb −1/2 ]) → γ.
3. Dictionary Reduction by Post-Nonparametric Double Selection
The previous section described estimation using a generic dictionary reduction.
This section discusses one class of possibilities for constructing such reductions.
It is important to note that the coverage probabilities of the above confidence sets
depend critically on how the dictionary reduction is performed. In particular, naive
one-step methods will fail to produce correct inference. Formal results expanding
on this point can be found, for instance, in [46], [40]. Heuristically, the reason
resulting confidence intervals have poor coverage properties is due to model selection
mistakes.
To address this problem, this section proposes a procedure for selecting q̃(z). The
new procedure is a generalization of the methods in [15] who work in the context
of the partially linear model E[y|x, z] = α0 x + h0 (z). The methods described below
rely heavily on Lasso-based model selection. Therefore, a brief description of Lasso
is now provided. The following description of Lasso, which uses an overall penalty
level as well as term-specific penalty loadings follows [8] who are motivated by
allowing for heteroskedasticity.
For any random variable v with observations (v1 , ..., vn ), the Lasso estimate v
on q L (z) with penalty parameter λ and loadings lj is defined as a solution
βbv,L,Lasso ∈ arg min
n
X
(vi − q L (zi )0 b)2 + λ
i=1
L
X
|lj bj |.
j=1
The corresponding selected set Iv,L is defined as
Iv,L = {j : βbv,L,Lasso,j 6= 0}.
Finally, the corresponding Post-Lasso estimator is defined by
n
X
βbv,L,Post-Lasso ∈ arg
min
(vi − q L (zi )0 b)2 .
b:bj =0 for j ∈I
/ v,L
i=1
13When the required inverse does not exist, a pseudo-inverse may be used.
8
DAMIAN KOZBUR
Lasso is chosen over other model selection possibilities for several reasons. Foremost, Lasso is a simple, computationally efficient estimation procedure which produces sparse estimates because of its ability to set coefficients identically equal
to zero. In particular, |Iv,L | will generally be much smaller than n if a suitable
penalty level is chosen. The second reason is for the sake of continuity with the
previous literature; Lasso was used in [15]. The third reason is for concreteness.
There are indeed many alternative estimation or model selection procedures which
select a sparse set of terms which in principle can replace the Lasso. It is possible
to instead consider general model selection techniques in the course of developing
the subsequent theory. However, framing the discussion using Lasso allows explicit
calculation of bounds and explicit description of tuning parameters. This is also
helpful in terms of practical implementation of the procedures proposed below.
The quality of Lasso estimation is controlled by λ and lj . As the number of
different Lasso estimations increases (ie. with increasingly many different variables
v), the penalty parameter must be increased to ensure quality estimation uniformly
over all different v. The penalty parameter must also be increased with increasing
L. However, higher λ typically leads to more shrinkage bias in Lasso estimation.
Therefore, given lj , λ is usually chosen to be large enough to ensure quality performance, and no larger. See [8] for details.
For the sake of completeness, the Post-Double Selection procedure of [15] is now
reproduced for a partially linear model specified by E[y|x, z] = α0 x + h0 (z).
Algorithm 1. Post-Double Selection for the Partially Linear Model. (Reproduced
from [15]).
1. First Stage Model Selection Step. Perform Lasso regression x on q L (z) with
penalty λFS and loadings lFS,j . Let IFS be the set of selected terms.
2. Reduced Form Model Selection Step. Perform Lasso regression y on q L (z)
with penalty λRF and loadings lRF,j . Let IRF be the set of selected terms.
3. Post-Model Selection Estimation. Set IPD = IFS ∪ IRF and let q̃(z) =
[qjL (z)]j∈IPD . Estimate α0 with α
b based on least squares regression14 of y
on [x, q̃(z)].
Appendix A contains details about one possible method for choosing λFS , λRF
as well as lFS,j ,lRF,j . Arguments in [15] show that the choices of tuning parameters given in Appendix A are sufficient to guarantee a centered Gaussian sampling
distribution of α
b for α0 .
The simplest generalization of Post-Double Selection is to expand the first stage
selection step into K steps. More precisely, for k = 1, ..., K, perform Lasso regression of pkK (x) on q L (z), and set IFS,k as the selected terms. Then define
15
IFS = ∪K
This
k=1 IFS,k and continue to the reduced form and estimation steps.
approach has a few disadvantages. First, the selected variables can depend on
the particular dictionary pK (x). Ideally, the first stage model selection should be
approximately invariant to the choice of pK (x).
14In [15] heteroskedasticity-consistent standard errors are used for inference.
15A previous draft of this paper took this approach. Deriving theoretical results for this
approach requires stronger sparsity assumptions than required here.
9
ADDITIVELY SEPARABLE
Instead, consider a general class of test functions ΦK = {ϕ}. Concrete classes for
test functions are provided below. In the first stage in Post-Nonparametric Double
Selection, a Lasso step of ϕ(x) on q L (z) is performed for each ϕ ∈ ΦK .
Algorithm 2. Post-Nonparametric Double Selection
1. First Stage Model Selection Step. For each ϕ ∈ ΦK , perform Lasso regression ϕ(x) on q L (z) with penalty λϕ and loadings lϕ,j . Let Iϕ,L be the
selected terms. Let IΦK = ∪ϕ∈ΦK Iϕ,L be the union set of selected terms.
2. Reduced Form Model Selection Step. Perform Lasso regression y on q L (z)
with penalty λRF and loadings lRF,j . Let IRF be the set of selected terms.
3. Post-Model Selection Estimation. Set IΦK +RF = IΦK ∪ IRF . Estimate θ0
using θb based on the reduced dictionary
(p̃(x), q̃(z)) = (pK (x), [qjL (z)]j∈IΦK +RF ).
The following are several concrete, feasible options for ΦK . The first option is
named the Span option. This option is suggested for practical use and is the main
option in the simulation study as well as in the empirical example that follow.
ΦK, Span = {ϕ(x) ∈ LinSpan(pK (x)) : var(ϕ(x)) 6 1}.
The theory in the subsequent section is general enough to consider other options
for ΦK which might possibly be preferred in different contexts. Three additional
examples are as follows.
ΦK, Graded = {p11 (x)} ∪ {p12 (x), p22 (x)}... ∪ {p1K (x), ..., pKK (x)}
(1)
(1)
(m)
(m)
ΦK, Multiple = {p1K (x), ..., pKK (x)} ∪ ... ∪ {p1K (x), ..., pKK (x)}}
ΦK, Simple = {p1K (x), ..., pKK (x)}.
Appendix A again contains full implementation details for the Span option. This
includes one possible method for choosing λϕ , lϕ,j as well as lFS,j ,lRF,j which yield
favourable model selection properties. Discussion of the most important details is
given in the text below. The analysis in the next section gives conditions under
which θb attains a centered Gaussian limiting distribution.
Choosing ΦK optimally is an important problem, which is similar to the problem of dictionary selection.16 The Span option, ΦK, Span is used in the simulation
study as well as in the empirical example, since it performed well in initial simulations. Note that the definition of the set ΦK, Span depends on a population
quantity var(ϕ(x)) which may be unknown to the researcher. Note however, that
the identities of the covariates selected in the Lasso-based procedure described in
the appendix are invariant to rescaling of the left-hand side variable. The invariance
16The question of which option for Φ is optimal is likely application dependent. In order to
K
maintain focused, this question is not considered in detail in this paper but might be of interest
for future work.
10
DAMIAN KOZBUR
is a consequence of the method for choosing penalty loadings. Therefore, replacing the condition supx∈X kϕ(x)k2 6 1 with var(ϕ(x)) 6 1 is possible. The option
ΦK, Simple is the direct extension of Post Double Selection as given in [15]. The
set ΦK, Multiple corresponds to using multiple dictionaries, indexed (1), ..., (m) in
the notation above. For example ΦK, Multiple could include the union of B-splines,
orthogonal polynomials, and trigonometric polynomials, all in the first stage selection. The ΦK, Graded is appropriate when dictionaries are not nested with respect
to K. These include B-splines.
In order to set up a practical choice of penalty levels, the set proposed above,
ΦK is considered as a union17 :
ΦK,Span = ΦK 1 ∪ ΦK 2 ∪ ΦK 3
where
ΦK 1 = {x}
ΦK 2 = {p1K (x), ..., pKK (x)}
ΦK 3 = {ϕ(x) ∈ LinSpan(pK (x)) : var(ϕ(x)) 6 1}.
The reason then for decomposing ΦK,Span in this way is allow the use of different
penalty levels on each of the three sets ΦK 1 , ΦK 2 , ΦK 3 . In particular, λΦK 1 is the
penalty for a single heteroskedastic Lasso as described in [8]. λΦK 2 is a penalty
which adjusts for the presence of K different Lasso regressions with K → ∞. The
main proposed estimator sets λΦK3 = λΦK2 . This is less conservative than the
penalty level would be following [10] for a continuum of Lasso estimations.18 As
a result, any corresponding Lasso performance bounds do not hold uniformly over
ΦK3 . Rather the implied bounds hold only uniformly over any pre-specified K
element subsets of ΦK3 . The high-level model selection assumption below (see Assumption 10) indicates that these bounds are sufficient for the present purpose. In
the simulation study, a more conservative (higher) choice for λΦK3 is also considered. In terms of inferential quality, there is no noticeable difference between the
two choices of penalties in the data generating processes considered in the simulation study. As discussed above, penalty levels accounting for a set of different
Lassos estimated simultaneously must be higher to ensure quality estimation. This
leads to higher shrinkage bias. The above decomposition therefore addresses both
concerns about quality estimation and shrinkage bias by allowing smaller penalty
levels to be used on subsets of ΦK,Span . Because the decomposition is into a fixed,
finite number of terms (ie. into 3 terms), such an estimation strategy presents no
additional theoretical difficulties.
Another practical difficulty with this approach is computational. It is infeasible
to estimate a Lasso regression for every ϕ indexed by a continuum. Therefore, some
approximation must be made. The reference [10] gives suggestions for estimating a
continuum of Lasso regressions using a grid. This may be computationally expensive if K is even moderately large. An alternative heuristic approach is motivated
by the observation that qjL is selected into IΦK only when there is ϕ ∈ ΦK such
that j ∈ Iϕ,L . In the context of estimating θ0 , only the identity of selected terms is
17Some dictionaries pK (x) may not contain a term p
kK (x) = x. In this case, ϕ(x) = x can
be appended to ΦK . In addition, after rescaling, ΦK 1 ⊆ ΦK 2 ⊆ ΦK 3 is possible, and so the sets
have nonempty intersection. This causes no additional problems.
18Note, the normalization that kϕ(x)k 6 1 ensures that Φ
2
K3 is indexed by a compact set and
so λΦK 3 can chosen as described in [10] to account for a continuum of Lassos.
11
ADDITIVELY SEPARABLE
important (not their coefficients). For the implementation in this paper, a strategy
for approximating IΦK is adopted where for each j 6 L, a Lasso regression is run
using exactly one test function, ϕ̌j ∈ ΦK . The choice of ϕ̌j is made based on being
likely to select qjL relative to other ϕ ∈ ΦK . Specifically, for each j, ϕ̌j is set to
the linear combination of p1K , ..., pKK with highest marginal correlation to qjL .
Then the
S approximation to the first stage model selection step proceeds by using
IˇΦK = j6L Iϕ̌j (x) in place of IΦK . This is also detailed in the appendix.
The formal theory in the subsequent sections proceeds by working with a notion
of density of ΦK within a broader space of approximating functions for g(x). Aside
from added generality, working in this manner is helpful since it adds structure to
the proofs and it isolates exactly how the density of ΦK interacts with the final
estimation quality for θ0 .
4. Formal Theory
In this section, additional formal conditions are given which guarantee convergence and asymptotic normality of the Post-Nonparametric Double Selection.
There are undoubtedly many aspects of the estimation strategy that can be analyzed. These include important choices of tuning parameters and K.
The following definition helps characterize smoothness properties of target function g0 and approximating functions pK . Let g be a function on X. Define the
Sobolev norm |g|d = supx∈X max|c|6d |∂ |c| g/∂xc | where the inner maximum ranges
over multi-indeces c.
Assumption 3 (Regularity for pK ). For each K, there is a nonsingular matrix
BK such that the smallest eigenvalue of the matrix
ΩBK pK = E BK pK (x)(BK pK (x))0
is bounded uniformly away from zero in K. In addition, there is a sequence of
constants ζ0 (K) satisfying supx∈X kBK pK (x)k2 6 ζ0 (K) and n−1 ζ0 (K)2 K → 0 as
n → ∞.
Assumption 4 (Approximation of g0 ). There is an integer d > 0, a real number
αg0 > 0, and a sequence of vectors βg0 ,K which depend on K such that |g0 −
0
pK βg0 ,K |d = O(K −αg0 ) as K → ∞.
Assumptions 3 and 4 would be identical to Assumptions 2 and 3 from [45] if
there was no conditioning variable z present. These assumptions require that the
dictionary pK has certain regularity and can approximate g0 at a pre-specified rate.
The quantity ζ0 (K) is dictionary specific, and can be explicitly calculated in certain
cases. For instance, [45] gives that ζ0 = O(K 1/2 ) is possible for B-splines. Note
that values of αg0 can be derived for particular d, pK , and classes of functions
containing g0 . [45] also gives explicit calculation of αg0 for the leading cases when
pK (x) are power series and regression splines.
The next assumption quantifies the density of ΦK within LinSpan(pK ). In order
to do so, define the following. Let
ρ(g, ΦK ) =
inf
kg >1, η=(η1 ,...,ηkg )∈Rkg ,ϕ1 ,...,ϕkg ∈ΦK
LαZ sup |g(x) − ϕ(x)| + kηk1 .
x∈X
12
DAMIAN KOZBUR
Assumption 5 (Density of ΦK ). Each ϕ ∈ ΦK satisfies var(ϕ(x)) 6 1. There is
a constant αρ > 0 such that
sup
ρ(g, ΦK ) = O(K αρ ).
{g∈LinSpan(pK ): var(g(x))61}
There is nothing special about the constant 1 in var(ϕ(x)) 6 1. It is mainly a
tool for helping describe the density of ΦK . In addition, as mentioned above, the
set selected by Lasso as described in the appendix is invariant to rescaling of the
left-hand side variable. As a result, imposing restrictions on var(ϕ(x)) is without
loss of generality.
The density assumption is satisfied with αρ = 0 if the ΦK = ΦK,Span is used
since in that case, ρZ (g, ΦK ) is bounded uniformly in g. On the other hand, the
density assumption may only be satisfied with αρ = 1/2 or higher for the basic
ΦK Simple = {p1K (x), ..., pKK (x)} option.
The next assumptions concern sparse approximation properties of q L (z). Two
definitions are necessary before stating the assumption. First, a vector X is called
s-sparse if |{j : Xj 6= 0}| 6 s. Next, let πqL denote the linear projection operator. More precisely, for a square integrable random variable v, πqL v is defined
by πqL v(z) = q L (z)0 βv,L for βv,L such that E[(v − q L (z)0 βv,L )2 ] is minimized. For
functions ϕ of x such that ϕ(x) is square integrable, write βϕ,L = βϕ(x),L .
Assumption 6 (Sparsity). There is a sequence s0 > 1 and a constant αZ > 0 such
that the following hold.
1. There is a sequence of vectors βh0 ,L,s0 that are s0 -sparse with support S0
such that supz∈Z |h0 (z) − q L (z)0 βh0 ,L,s0 | = O(L−αZ ).
2. For all ϕ ∈ ΦK there are vectors βϕ,L,s0 that are s0 -sparse, all with common
support S0 , such that
sup sup |πqL ϕ(z) − q L (z)0 βϕ,L,s0 | = O(L−αZ ).
ϕ∈ΦK z∈Z
Assuming a uniform bound for the sparse approximation error for h0 is potentially stronger than necessary. At the moment of the writing of the manuscript, the
author sees
Pnno theoretical obstacle in terms of working under the weaker assumption n−1 i=1 |h0 (zi )−q L (zi )0 βh0 ,L,s0 |2 = Op (L−αZ ). In addition, the L−αZ rate is
imposed in order to maintain a parallel exposition relative to the O(K −αg0 ) term.
Other rates, for instance n−αZ , can also replace L−αZ , and this is done in [15], [8]
and other references.19 The same comment holds for the sparse approximation
conditions for ϕ ∈ ΦK .
Several references in the prior econometrics literature work with sparse approximation of the conditional expectation rather than the linear projection. In this
context, working with the conditional expectation places a higher burden on the
approximating dictionary q L . In particular, If the conditional expectation of ϕ(x)
given z can be approximated using s0 terms from q L , then the conditional expectation of ϕ(x)2 may potentially require O(s20 ) terms to approximate once interactions
are taken into account. This potentially requires the dictionary q L to contain a
prohibitively large amount of interaction terms. For this reason, the conditions in
this paper are cast in terms of linear projections.20
19using n−αZ is only more general if L grows faster than every polynomial of n.
20The author sees no theoretical obstructions in terms of applying the same arguments for
Lasso bounds in [8] without the conditional expectation assumption. The key ingredient in that
ADDITIVELY SEPARABLE
13
The next assumption imposes limitations on the dependence between x and z.
For example, in the case that ϕ(x) = x is an element of pK (x), this assumption
states that the residual variation after a linear regression of x on z is non-vanishing.
More generally, the assumption requires that population residual variation after
projecting pkK (x) away from z is non-vanishing uniformly K, L. One consequence
of Assumption 7 is that constants cannot be freely added to both g0 (x) and h0 (z).
This therefore requires the user to enforce a normalization condition like g0 (0) = 0
or E[g0 (x)] = 0. The simulation study and empirical illustration below both enforce
g0 (0) = 0.
Assumption 7 (Identifiability). For each K and for BK as in Assumption 3, the
matrix E[BK (pK (x)−πqL pK (z))(BK (pK (x)−πqL pK (z)))0 ] has eigenvalues bounded
uniformly away from zero in K, L. In addition, supz∈Z kBK πqL pK (x)k2 6 ζ0 (K).
The next condition restricts the sample Gram matrix of the second dictionary.
A standard condition for nonparametric estimation is that for a dictionary P , the
Gram matrix n−1 P 0 P eventually has eigenvalues bounded away from zero uniformly
in n with high probability. If K +L > n, then the matrix n−1 [P Q]0 [P Q] will be rank
deficient. However, in the high-dimensional setting, to assure good performance of
Lasso, it is sufficient to only control certain moduli of continuity of the empirical
Gram matrix. There are multiple formalizations of moduli of continuity that are
useful in different settings, see [17], [59] for explicit examples. This paper focuses
on a simple condition that seems appropriate for econometric applications. In particular, the assumption that only small submatrices of n−1 Q0 Q have well-behaved
eigenvalues will be sufficient for the results that follow. In the sparse setting, it
is convenient to define the following sparse eigenvalues of a positive semi-definite
matrix M :21
δ0 M δ
δ0 M δ
,
κ
(m)(M
)
:=
max
κmin (m)(M ) := min
max
2
2 .
16kδk0 6m kδk2
16kδk0 6m kδk2
In this paper, favorable behavior of sparse eigenvalues is taken as a high level
condition and the following is imposed.
Assumption 8 (Sparse Eigenvalues). There is a sequence sκ = sκ (n) such
that sκ → ∞ and such that the sparse eigenvalues obey κmin (sκ )(n−1 Q0 Q)−1 =
O(1) and κmax (sκ )(n−1 Q0 Q) = O(1) with probability 1 − o(1).
The assumption requires only that sufficiently small submatrices of the large p × p
empirical Gram matrix n−1 Q0 Q are well-behaved. This condition seems reasonable
and will be sufficient for the results that follow. Informally it states that no small
subset of covariates in q L suffer a multicollinearity problem. They could be shown
to hold under more primitive conditions by adapting arguments found in [9] which
build upon results in [57] and [49]; see also [48].
Pn
L
L
argument is that expression
i=1 q (zi )(πq L (zi ) − q (zi )βϕ,L ) stays suitably small. Note this
expression is a sum of mean zero independent random variables in the present context.
21In the sparse eigenvalue definition, k · k refers to the number of nonzero components of a
0
vector ( · ).
14
DAMIAN KOZBUR
Assumption 9 (High-Level Model Selection Performance). There are constants
αIΦ and αΦ and bounds
1. |IΦK | 6 K αIΦ O(s0 )
Pn
2. supϕ∈ΦK i=1 (q L (zi )0 (βϕ,L,s0 − βbϕ,L,Post-Lasso ))2 = O(K αΦ s0 log(L))
3. |IRF | = O(s0 )
Pn
L
0
2
b
4.
i=1 (q (zi ) (βy,L,s0 − βy,L,Post-Lasso )) = O(s0 log(L))
which hold with probability 1 − o(1).
The standard Lasso and Post-Lasso estimation rates when there is only one
outcome considered are s0 log(L) for the sum of the squared prediction errors, and
O(s0 ) for the number of selected covariates. Therefore, K αΦ is a uniform measure
of the loss of estimation quality stemming from the fact that Lasso estimation is
performed on all ϕ ∈ ΦK rather than just on a single outcome. Similarly, K αIΦ
measures the number of unique j selected in all first stage Lasso estimations. The
choice to present high-level assumptions is for generality - so that other model
selection techniques can also be applied. However, verification of the high level
bounds are available under additional regularity for Lasso estimation.
One reference on performance bounds for a continuum of Lasso estimation steps
is [10]. In that paper, the authors provide formal conditions (specifically Assumption 6.1) and prove that Statement 2 of Assumption 9 holds. The bounds in that
reference correspond to taking αΦ = 1/2. An important note is that the conditions
in [10] are slightly more stringent since the authors assume that βϕ,L,s0 and βy,L,s0
can be taken to approximate the conditional expectation of ϕ(x) and y given z
rather than just the linear projection. When |ΦK | finite, but grows only polynomially with n and L > n, αΦ = 0 is possible under further regularity conditions.
The main theoretical difficulty in verifying Assumption 9 using primitive conditions is in showing that the size of the set IΦK stays suitably small. [10] prove certain
performance bounds for a continuum of Lasso estimates under the assumption that
dim ΦK is fixed and state that their argument would hold for certain sequences
dim ΦK → ∞. [10] also proves that the size of the supports of the Lasso estimates,
|Iϕ,L | stay bounded uniformly by a constant multiple of s0 which does not depend
on n or ϕ. They do not, however, prove that the size of the union | ∪ϕ∈ΦK Iϕ,L |
remains similarly bounded. Therefore, their results do not imply a the existence
of a finite value of αIΦ . The later bound is required for the analysis of the above
proposed estimator. For a finite approximation to ΦK,Span (like ΦK, Simple ), there
is no difficulty calculating bounds on the total number of distinct selected terms.
This is because under regularity conditions standard in the literature, each Iϕ,L
satisfies |Iϕ,L | 6 O(s0 ) where the implied constants in the O(s0 ) terms can be
bounded uniformly over ϕ ∈ ΦK . In particular, when ΦK is finite, it is possible to
take K αIΦ = |ΦK |. This paper does not derive a bound for |IΦK, Span | as this would
likely lie outside the scope of this project. A valid alternative for which verifiable
bounds on the union of selected covariates is possible is to report estimates using
(
ΦK, Simple on the event that |IΦK, Span | > t(n)
b
ΦK =
ΦK, Span otherwise.
for some increasing threshold function t of n.
ADDITIVELY SEPARABLE
15
When {g ∈ LinSpan(pK ) : var(g(x)) 6 1} coincides with ΦK , so that ΦK is as
dense as possible, then Assumption 9 can be weakened in the following.
Assumption 10 (Alternative High-Level Model Selection Performance ). Suppose
that {g ∈ LinSpan(pK ) : var(g(x)) 6 1} ⊆ ΦK . Let ΦK 0 ⊂ ΦK be any nonrandom
fixed finite subset of at most K elements. There are constants αIΦ and αΦ and
bounds
1. |IΦK | 6 K αIΦ O(s0 )
Pn
(q L (zi )0 (βϕ,L,s − βbϕ,L,Post-Lasso ))2 = O(K αΦ s0 log(L))
2. sup
0
ϕ∈ΦK
i=1
0
3. |IRF | = O(s0 )
Pn
L
0
2
b
4.
i=1 (q (zi ) (βy,L,s0 − βy,L,Post-Lasso )) = O(s0 log(L))
which hold with probability 1 − o(1).
Assumption 10 is weaker than Assumption 9. However, Assumption 9 can be
more easily verified with primitive conditions by using finite sets ΦK .
Statements 2-4 can be attained under standard conditions with αΦ = 0 provided
a penalty adjusting for K different Lasso estimations is used. On the other hand,
using a conservative penalty as in [10] for the continuum of Lasso estimations like
in ΦK,Span would result in αΦ = 1/2. There is currently no proof that Statement
1 with αIΦ = 0 and Statement 2 with αΦ = 0 can hold simultaneously under
conditions standard in the econometrics literature.
It is interesting to note that the requirements to satisfy Assumption 10 are essentially pointwise bounds on the predictive performance of a set of Lasso estimations
along with a uniform bound on the identity of selected covariates. By contrast, [10]
prove uniform bounds on Lasso estimations along with pointwise bounds on the
identity of selected covariates. In practice, verification of the Condition 1 in Assumption 10 could be potentially very useful. This would allow the researcher to
use a penalty level which is smaller by a factor of K 1/2 , and would ultimately allow
more robustness without increasing variability of the final estimator.
For the choice of penalty parameters given in Appendix A for the Span option,
Conditions 2-4 of Assumption 10 can be verified under further regularity conditions like those given in [10] or [8] to yield αΦ = 0. Furthermore, Condition 1
b K mentioned on the previous
of Assumption 10 can be verified if an option like Φ
page is used. Most importantly, Assumption 10 serves a plausible high-level model
selection condition which is sufficient for proving the results that follow.
The next assumption describes moment conditions needed
Pn by applying certain
laws of large numbers, for instance for the quantities n−1 i=1 ε2i qjL (zi )2 .
Assumption 11 (Moment Conditions). The following moment conditions hold.
1. E[qjL (z)2 [BK (pK (x) − πqL pK (z))]2k ] is bounded away from zero uniformly
in K, L
2. E[|qjL (z)|3 ] is bounded uniformly in L
3. E[qjL (z)2 ε2 ] is bounded away from zero uniformly in L
4. E[|qjL (z)|3 |ε|3 ] is bounded uniformly in L.
The first statement of the assumption may also be seen as a stricter identifiability condition condition on the residual variation pK (x) − πqL pK (z). It rules
out situations where for instance x 6= 0 ⇔ qjL (z) = 0. Note that E[[BK (pK (x) −
πqL pK (z))]2k ] = 1 is given by the identifiability assumption. No direct assumption
16
DAMIAN KOZBUR
is needed about the corresponding third moment E[[BK (pK (x) − πqL pK (z))]3k ] = 1
since instead a reference to the bound ζ0 (K) is used.
The final assumption before the statement of Theorem 1 are rate conditions.
Assumption 12 (Rate Conditions). The following rate conditions hold.
1. s0 K αIΦ = o(sκ )
2. log(KL) = o(ζ0 (K)−1 n1/3 )
3. L−αZ n1/2 K −1/2 ζ0 (K) = O(1)
4. L−2αZ K 2αρ (K 1/2 n1/2 ζ0 (K)−1 + n1/2 + K log(L)1/2 ζ0 (K)−2 ) = O(1)
5. n−1/2 K 1/2 s0 log(L)ζ0 (K)−1 (K 2αρ +αΦ + K αρ +αΦ /2+αIΦ /2 ) = O(1)
1/2
1/2
6. n−1/2 s0 log(L)(K 2αρ +αΦ s0 + K αIΦ /2 ) = O(1).
The first statement ensures that the sparse eigenvalues remain well-behaved in
the with high probability over sets whose size is larger that the selected covariates. The second statement is used in conjunction with the above moment conditions to allow the use of moderate deviation bounds following [37]. The third
and fourth conditions are assumption on the sparse approximation error for q L (z).
The final two assumptions restrict the size of s0 and K and quantities depending on αρ , αΦ , and αIΦ relative to n. These assumptions can be unraveled for
certain choices of dictionaries. For example, as was noted above and by [45], for
B-splines, ζ0 (K) can be taken to be O(K 1/2 ). Using the simple option gives αρ = 1,
αΦ = 0 and αIΦ = 1. Then the conditions can be reduced to L−αZ n1/2 = O(1),
L−2αZ K 2 n1/2 = O(1), n−1/2 K 2 s0 log(L) = O(1).
The first result is a preliminary result which gives bounds on convergence rates
for the estimator gb. They are used in the course of the proof of Theorem 1 below, the
main inferential result of this paper. The proposition is a direct analogue of the rates
given in Theorem 1 of [45] which considers estimation of a conditional expectation g0
without model selection over a conditioning set. The rates obtained in Proposition
1 match the rates in [45]. To state it, let F0 be the distribution function of the
random variable x. In addition, let ζd (K) = max|c|6d supx∈X k∂ |c| BK pK (x)/∂xc k2 .
Theorem 1. Under Assumptions 1-8, 9 or 10, and 11-12, the Post-Nonparametric
Double Selection estimate gb for the function g0 satisfies the following bounds.
Z
(b
g (x) − g0 (x))2 dF0 (x) = Op (n−1 K + K −2αg0 ).
|b
g − g0 |d = Op (n−1/2 ζd (K)K 1/2 + K −αg0 ).
The next formal results concern inference for θ0 = a(g0 ). Recall that θ0 , is
estimated by θb = a(b
g ) and inference is conducted via the estimator Vb as described
in earlier sections.
Assumption 13 (Moments for Asymptotic Normality). E[ε4+δ |x, z] is bounded for
some δ > 0. var(ε|x, z) is bounded away from zero.
Note that the conditions in [45] require only that E[ε4 |x, z] is bounded. The
strengthened condition is needed for consistent variance estimation, in order to
construct a bound on the quantity maxi6n ε2i .
The following assumptions on the functional a are imposed. They are regularity
assumptions that imply that a attains a certain degree of smoothness. For example,
they imply that a is Fréchet differentiable.
ADDITIVELY SEPARABLE
17
Assumption 14 (Differentiability for a). The real valued functional a(g) ∈ R
is either linear or the following conditions hold. n−1 ζd (K)4 K 2 → 0. There is a
linear function D(g; ǧ) that is linear in g and such that for some constants C, ν >
0 and all ǧ, ǧˇ with |g − g0 |d < ν, |ǧ − g0 |d < ν, |ǧˇ − g0 |d < ν, it holds that
|a(g) − a(ǧ) − D(g − ǧ; ǧ)| 6 C(|g − ǧ|d )2 and |D(g; ǧ) − D(g; ǧˇ)| 6 C|g|d |ǧ − ǧˇ|d .
The function D is related to the functional derivative of a. The following assumption imposes further regularity on the continuity of the derivative. For shorthand,
let D(g) = D(g; g0 ).
The next rate conditions is used to ensure that estimates are undersmoothed.
The rate condition ensures that the estimation bias, which is heuristically captured
by K −αg0 , converges to zero faster than the estimation standard error.
Assumption 15 (Undersmoothing Rate Condition). n1/2 K −αg0 = o(1).
The next rate condition is used in order to bound quantities appearing in the
proof of Theorem 2. As was demonstrated in the case of Assumption 12, the rate
conditions can be unraveled for certain choices of K, pK , and ΦK .
Assumption 16 (Rate Conditions for Asymptotic Normality).
1. L−2αZ K 2αρ (ζ0 (K)K + ζ0 (K)4 K 1−2αρ + K log(L) + n1/2 ) = o(1)
2. n−1 s0 ζ0 (K) log(L)(K 1+2αρ +αΦ + K 1+αρ +αΦ /2+αIΦ /2 ) = o(1)
3. n−1 s20 log(L)2 (K 4αρ +2αΦ + K 2αρ +αΦ +αIΦ ) = o(1)
4. s0 K αIΦ n−1/2 ζ0 (K)K 1/2 + K −αg0 = o(1)
5. n2/(4+δ) ζ0 (K)n−1/2 K 1/2 = o(1).
The final two conditions divide the cases considered into two classes. The first
class (covered by Assumption 17) are functionals which fail to be mean-square differentiable and therefore cannot be estimated at the parametric n1/2 rate. The
second class (covered by Assumption 18) does attain the n1/2 rate. One example
with the functional of interest is evaluation of g at a point x0 : a(g) = g(x0 ). In this
case, a fails to be estimated that the parametric n1/2 rate in general
circumstances.
R
A second example is the weighted average derivative a(g) = w(x)∂g(x)/∂x for
a weight function w which satisfies regularity conditions. The Assumption 18
holds if w is differentiable, vanishes outside a compact set, and the density of x
is bounded away from zero wherever w is positive. In this case, a(g) = E[ψ(x)g(x)]
for ψ(x) = −φ(x)−1 ∂w(x)/∂x by a change of variables provided that x is continuously distributed with non vanishing density φ. These are one possible set
of
√ sufficient conditions under which the weighted average derivative does achieve
n-consistency.
Assumption 17 (Regularity for a in Absence of Mean-Square Differentiability).
There is a constant C̄ > 0 such that |D(g)| 6 C̄|g|d . There is β̄ dependent on K
0
such that for ḡ(x) = p(x)K β̄, it holds that E[ḡ(x)2 ] → 0 and D(ḡ) > C̄ > 0.
Assumption 18 (Conditions for n1/2 -Consistency). There is ψ(x) such that
E[ψ(x)2 ] finite and nonzero and such that D(g) = E[ψ(x)g(x)] and D(pkK ) =
0
E[ψ(x)pkK (x)] for every k. There is β̆ such that E[(ψ(x) − p(x)K β̆)2 ] → 0. Fi2
nally, the matrix V̄ = E[ψ(x) var(y|x, z)] is finite and nonzero.
Theorem 2 now establishes the validity of standard inference procedure after
model selection as well as validity of the plug in variance estimator.
18
DAMIAN KOZBUR
Theorem 2. Under Assumptions 1-8, 9 or 10, 11-17, the Post-Nonparametric
Double Selection estimate for the function θ0 satisfies
θb = θ0 + Op (n−1/2 ζd (K)).
In addition,
d
n1/2 V −1/2 (θb − θ) → N(0, 1)
and
d
n1/2 Vb −1/2 (θb − θ) → N(0, 1).
Under Assumptions 1-8, 9 or 10, 11-16, and 18,
d
n1/2 (θb − θ) → N(0, V̄ ) and
p
V − V̄ → 0.
5. Simulation study
The results stated in the previous section suggest that Post-Nonparametric Double Selection series estimation should exhibit good inferential properties for additively separable conditional expectation models when the sample size n is large. The
following simulation study is conducted in order to illustrate the implementation
and study the performance of the outlined procedure.
The simulation study is divided into two parts. The first part compares several
alternative estimators to Post-Nonparametric Double Selection. The second part
compares several Post-Nonparametric Double Selection estimates using different
choices for ΦK . This part demonstrates finite sample benefits from using the Span
option relative to the direct generalization of Post-Double Selection estimation (ie.
using the Simple option).
The following process generates the data in each simulation.
y = g0 (x) + h0 (z) + ε
2
g0 (x) = 10 sin(0.1x) − 0.5 sin(4πx4−x )
h0 (z) = z 0 βh0 ,L , βh0 ,L,j = −0.5 · (−0.65)j−1 1j6s0
zj ∼ N(0, 1), corr(zj1 , zj2 ) = 0.25|j1 −j2 |
ε ∼ N(0, 1)
x = 0.15v + 0.0375 − 3.75(stair(z 0 γ0 + v) + 0.375)FN(0,1) (10zs0 )...
+ 3.75(stair(0.5(z 0 γ0 + v)|z 0 γ0 + v|0.25 )(1 − FN(0,1) (10zs0 ))
γ0,L,j = −1.5(−0.75)j−1 1j6s0
v ∼ N(0, 1)
stair( · ) = 0.25
tanh(12( · )/2.5) − 12b( · )/2.5c − 6
.
2 tanh(6) + 0.5 + b( · )/2.5c
The study performs simulations for n ∈ {100, 150, ..., 500}. Two settings for the
parameter L are considered: L = n/2 and L = 2n. Finally, the sparsity level is set
to s0 = 6. Within each data generating process, 1000 simulation replications are
performed.
ADDITIVELY SEPARABLE
19
The data generating process is quite complicated. It is designed in order to create
correlations between the covariates z and various transformations of x. This allows
the data generating process to highlight many different statistical problems which
can arise using Nonparametric-Post Double Selection and alternative estimation
techniques all in one simulation study. Despite the complicated formulas for the
joint distribution of x and z, their realizations appear natural. Scatter plots of one
sample of n = 500 showing the respective bivariate distributions between z1 , ..., z6
and x are provided in Figure 6. Figure 5 provides a picture of the graph of g0 .
The simulations evaluate estimation of g0 and of θ0 defined by
θ0 = E[g00 (x)].
In order to avoid further complications, for each replication, the expectation and
thus true θ0 are calculated against the empirical distribution of x within that simulation replication.22
The first part of the simulation study considers the performances of five estimators23 for g0 and θ0 . Each estimator is a reduced series estimator based on initial
dictionaries consisting of a cubic spline expansion pK (x) for g0 (x) and a linear
expansion q L (z) = z for h0 (z).
1. Oracle. Estimator 1 is infeasible and sets q̃(z) = (z1 , .., zs0 ). This estimator serves as a benchmark for comparison to estimates in which the correct
support is known.
2. Span Post-Nonparametric Double. Estimator 2 selects q̃(z) using
Post-Nonparametric Double Selection with Φk given by the Span option,
as described in this paper.
3. Naive. Estimator 3 selects q̃(z) in one model selection step by performing
Lasso of y on q L (z).
4. OLS. Estimator 4 uses q̃(z) = z. In other words, this estimator does not
reduce the dictionary. This estimation strategy is only calculated provided
L < n.
5. Targeted Undersmoothing. Estimator 5 implements an alternative inferential procedure for dense functionals of high-dimensional parameters;
TU(1). This procedure was proposed in [30] and is described further below.
22Another possibility is to calculate against the population expectation of x. Under the assumption that the researcher knows the population distribution of x, this causes no further complication. If the distribution of x is unknown and estimated, this must however be taken into
account.
23
There are likely other sensible estimators beyond the 5 considered in the simulation section.
As pointed out by an anonymous reviewer, such estimators may include propensity score matching
on a continuous variable. Though such an approach may work well, the context here is not exactly
the same as usually seen in propensity score matching. In particular, the assumptions here do not
require unconfoundedness conditions. In addition, propensity score techniques are most commonly
applied to discrete treatement variables. There is some work on propensity score matching with a
continuous treatment; for example, see [35], who require the estimation of the conditional density
of treatment. In the high-dimensional setting, estimating the conditional density of x given z
would likely introduce complications beyond the scope of this paper.
20
DAMIAN KOZBUR
Detailed implementation descriptions are provided in Appendix A. For each
of the above estimators, the choice of pK (x) is made using a data-dependent
rule. First, an initial dictionary reduction q initial (z) is selected. For Oracle,
q initial (z) = (z1 , .., zs0 ). For the Span Post-Nonparametric Double and Naive estimators, q initial (z) is based on Lasso of y on q L (z) as implemented in Appendix A.
For OLS, q initial (z) = z. Next, BIC is used to choose a B-spline expansion pK (x).
Comparison of estimators 1-4 is standard in the post-model selection econometrics literature. The oracle estimator should be seen as a benchmark which is known
to provide good estimates if the true set, S0 , was known. The Naive estimator is
expected to perform poorly since it is not a uniformly valid estimator and susceptible to size-distortions arising from model selection mistakes. OLS is expected to
perform poorly due to potential problems related to overfitting.
Estimator 5 is a procedure called Targeted Undersmoothing which looks to correct distortions in inference from model selection mistakes. Targeted Undersmoothing appends covariates which significantly affect the value of the functional θb = a(b
g)
to an initially selected model (see [30]). It is appropriate for functionals of highdimensional models which depend on a growing number of parameters (dense functionals) and is therefore a potentially sensible procedure for inference for θ0 . This
estimator is detailed further in Appendix A.
The simulation results report several quantities which measure the performance
b bias
of each estimator. The results report standard deviation of the estimates θ,
of the estimates for θ0 , confidence interval length for estimates for θ0 , rejection
frequencies under the null for θ0 at the 5% level, mean number of series terms K
used, mean number of series terms selected from the original L, and integrated
squared error for g0 . The simulation results are reported in Figure 1 for L = n/2
and Figure 2 for L = 2n. The figures display the above mentioned simulation results
for each n = 100, ..., 500 with n changing over the horizontal axis.24 Note also that
across some of the estimators, some of the reported quantities will be identical. For
example, the point estimates for TU are identical to the Naive point estimates. The
selected K is identical for the Naive estimates as well as the Post-Nonparametric
Double Selection estimates.
In all of the simulations, the Post-Nonparametric Double Selection estimates
behave similarly to the Oracle estimates. The OLS estimates have wide confidence
intervals relative to the Post-Nonparametric Double Selection estimation, but have
similar coverage properties. The final estimator, Targeted Undersmoothing (TU),
is conservative in terms of coverage, with substantially larger intervals in every case.
On the other hand, the Naive estimator has poor coverage properties. For the
Naive estimator, after failing to control for the correct covariates, the increase in K
leads to an increasing bias. This highlights the fact that simply producing undersmoothed estimates of g0 by increasing K may not be adequate for reducing bias
and making quality statistical inference possible in the high-dimensional setting.
24Note that since s , the magnitude of coefficients β
0
h0 ,L and the joint distribution between
relevant covariates are all fixed in the simulations as n → ∞. Therefore, for sufficiently large n, all
relevant covariates would be identified with high probability, and all of the post-model selection
estimators would perform similarly. This simulation study therefore is identifying differences in
finite sample performance.
ADDITIVELY SEPARABLE
21
Figure 1. Simulation Results
This figure presents simulation results for the estimation of g0 and θ0 in the cases
n = 100, 150, ..., 500 with s0 = 6 and L = n/2 according to the data generating process described
in the text. Estimates are presented for the five estimators, Oracle, Post-Nonparametric Double
(PND Span), Naive, OLS and Targeted Undersmoothing (TU) as described in the text. The first
plot shows standard deviation of the respective estimates for θ0 . The second plot shows bias of
the estimates for θ0 . The third plot shows confidence interval length for estimates for θ0 . The
fourth plot shows rejection frequencies under the null for θ0 for a 5% level test. The fifth plot
shows the mean number of series terms K used. The sixth plot shows the mean number of series
terms from L selected. The seventh plot shows root mean integrated squared error for g0 .
Figures are based on 1000 simulation replications. n is always indexed by the horizontal axis.
22
DAMIAN KOZBUR
Figure 2. Simulation Results
This figure presents simulation results for the estimation of g0 and θ0 in the cases
n = 100, 150, ..., 500 with s0 = 6 and L = 2n according to the data generating process described
in the text. Estimates are presented for the four estimators, Oracle, Post-Nonparametric Double
(PND Span), Naive, and Targeted Undersmoothing (TU) as described in the text. The first plot
shows standard deviation of the respective estimates for θ0 . The second plot shows bias of the
estimates for θ0 . The third plot shows confidence interval length for estimates for θ0 . The fourth
plot shows rejection frequencies under the null for θ0 for a 5% level test. The fifth plot shows the
mean number of series terms K used. The sixth plot shows the mean number of series terms
from L selected. The seventh plot shows root mean integrated squared error for g0 . In each plot,
the horizontal axis denotes sample size n. Figures are based on 1000 simulation replications. n is
always indexed by the horizontal axis.
ADDITIVELY SEPARABLE
23
The second part of the simulation study compares four Post-Nonparametric Double Selection estimators which use different specifications for ΦK .
1. Span Post-Nonparametric Double. Estimator 1 is identical to the Span
Post-Nonparametric Double estimator in the first part of the simulation.
2. Conservative Span Post-Nonparametric Double. Estimator 2 uses
pK and ΦK as in the Span option, but in the decomposition ΦK,Span =
ΦK 1 ∪ ΦK 2 ∪ ΦK 3 , the penalty applied to ΦK 3 is more conservative, explicitly aimed at achieve Lasso performance bounds which hold uniformly over
all of ΦK .
3. Simple Post-Nonparametric Double. Estimator 3 uses pK as in the
Span, but uses ΦK = ΦK, Simple .
4. Alternative Spline Basis Simple Post-Nonparametric Double. Estimator 4 uses a different basis for selection. A QR decomposition is applied
to P in order to obtain orthonormal columns. Next, ΦK = ΦK, Simple is
used on the new orthogonalized data. Importantly, the new P spans the
same K-dimensional linear space in Rn as in the 3 previous estimators.
The estimates for the second part of the simulation are presented in Figures 3-4.
Note that all estimators are identical with regards to K, hence only one curve is
visible in the corresponding plots. In addition, the Conservative Span and Span
estimators have very similar performance in terms of standard deviation, bias, interval length, rejection frequency, and integrated squared error. The two estimators
are practically indistinguishable except in terms of the number of elements of q L
they select. They do not give numerically identical estimates or confidence intervals.
However, their differences are too small to be seen in Figures 3-4.
There are noticeable differences in the performance of the estimators. The Span
option is able to identify the highest number of relevant covariates, followed by
the Conservative Span option, the Simple option, and the Alternative Spline Basis
Simple option. The Span, Conservative Span, and Simple Post-Nonparametric
Double Selection procedures exhibit favorable finite sample properties for this data
generating process. In particular, for those estimators, the calculated rejection
frequencies move towards 5% as n increases.
By contrast, the Alternative Spline Basis Simple Post-Nonparametric Double
Selection procedure has very poor finite sample performance. It is unlikely that the
projection of the new orthogonalized basis onto q L has a good sparse representation.
This causes increased model selection mistakes in the first stage. Unlike in the
partially linear model, these mistakes can accumulate to cause more severe bias
since the number of first stage selection steps is growing with K. Note that the
Alternative Spline Basis estimator has similar performance to the Naive estimator
in the first part of the simulation study.
The Span and the Conservative Span options offer an opportunity to potentially
add additional robustness. These options select more variables than the Simple
option. There is no evidence from this simulation study that using the Span option
over-selects conditioning variables to the extent that rejection frequencies become
severely distorted or variability increases to an undesirable level.
24
DAMIAN KOZBUR
Figure 3. Simulation Results
This figure presents simulation results for the estimation of g0 and θ0 in the cases
n = 100, 150, ..., 500 with s0 = 6 and L = n/2 according to the data generating process described
in the text. Estimates are presented for four Post-Nonparametric Double Selection (PND)
estimators, Simple, Span, Conservative Span, and Alternative Spline Simple as described in the
text. The first plot shows standard deviation of the respective estimates for θ0 . The second plot
shows bias of the estimates for θ0 . The third plot shows confidence interval length for estimates
for θ0 . The fourth plot shows rejection frequencies under the null for θ0 for a 5% level test. The
fifth plot shows the mean number of series terms K used. The sixth plot shows the mean
number of series terms from L selected. The seventh plot shows root mean integrated squared
error for g0 . In each plot, the horizontal axis denotes sample size n. Figures are based on 1000
simulation replications. n is always indexed by the horizontal axis.
ADDITIVELY SEPARABLE
25
Figure 4. Simulation Results
This figure presents simulation results for the estimation of g0 and θ0 in the cases
n = 100, 150, ..., 500 with s0 = 6 and L = 2n according to the data generating process described
in the text. Estimates are presented for the four Post-Nonparametric Double Selection (PND)
estimators, Simple, Span, Conservative Span, and Alternative Spline Simple as described in the
text. The first plot shows standard deviation of the respective estimates for θ0 . The second plot
shows bias of the estimates for θ0 . The third plot shows confidence interval length for estimates
for θ0 . The fourth plot shows rejection frequencies under the null for θ0 for a 5% level test. The
fifth plot shows the mean number of series terms K used. The sixth plot shows the mean
number of series terms from L selected. The seventh plot shows root mean integrated squared
error for g0 . In each plot, the horizontal axis denotes sample size n. Figures are based on 1000
simulation replications. n is always indexed by the horizontal axis.
26
DAMIAN KOZBUR
Figure 5. Simulation study: g0
This figure depicts the function g0 used in the simulation study.
Figure 6. Simulation study: joint covariate distribution
This figure depicts the joint distribution between x and the first s0 = 6 covariates
as described in the above text. The plots are generated by one sample of size
n = 500.
ADDITIVELY SEPARABLE
27
Figure 7. GDP Growth Results
6. Empirical Example: GDP Growth
This section applies Post-Nonparametric Double Selection to an international
economic growth example. The data comes from the Barro and Lee [6] dataset
which contains a panel of 138 countries for the period of 1960 to 1985. This example
was also considered in [13], who apply Lasso techniques in the context of a highdimensional linear model for the purpose of locating important variables which
are predictive of GDP growth rates. This considers growth in GDP per capita as
a dependent variable y for the period 1965-85. The growth rate in GDP over a
period from t1 to t2 is commonly defined as log(GDPt2 /GDPt1 ).
Studying the factors that influence growth in GDP is a problem of central importance in economics. A difficulty with studying this problem empirically on a
cross-country level is that the number of observations is limited by the total number of countries. At the same time, the number of potential factors which influence
GDP growth can be large. This leads naturally to the need to regularize econometric estimation on any data on a cross-section of countries. This example specifically
studies the relation between initial GDP level and subsequent GDP growth in the
presence of a large number of other determinants of GDP growth. The interest in
studying this particular question is in testing the fundamental macroeconomic theory of convergence. Convergence predicts that countries with high initial GDP will
show lower levels of GDP growth, and conversely countries with low initial GDP
will show higher levels of GDP growth. There are many references for assumptions
which imply such convergence. See [1] and references therein.
This analysis considers a model with p = 62 covariates, which allows for a total of
n = 90 complete observations. Since p is comparably large relative to n, dimension
reduction in this setting is necessary. The goal here is to select a subset of these
covariates and briefly compare the resulting to predictions made in the growth
literature (see [7], [6] ). [24] and [6] contain complete definitions and discussion of
each of these variables. The estimated model is given by the specification
28
DAMIAN KOZBUR
yi = g0 ( log(GDPi ) − log(GDPi ) ) + h0 (zi ) + εi
where log(GDPi ) denotes the sample mean. The observed covariates enter linearly,
so that the expansion h0 (zi ) = zi1 β1 + .... + zi62 β62 is assumed.
The estimation is performed using cubic splines as detailed in Appendix A. g0
is normalized so that g0 (0) = 0. Estimates of several density-weighted average
derivatives of the effect of initial GDP on GDP growth are constructed using PostNonparametric Double Selection and are presented in Table 1. In addition, a scatter
plot of the primary variables of interest as well as an estimate of g0 are shown in
Figure 7.
A nonlinear specification for g0 allows testing of several hypotheses related to the
convergence of GDP. These include the hypothesis that conditional convergence can
depend on initial GDP. This is related to the idea of a poverty trap where countries
with smaller initial GDP exhibit less convergence (ie. the relationship between
initial GDP and GDP growth may be locally flat; see the reference text [1] for
additional background and details.) Conditional convergence could also imply that
at the high end of the initial GDP distribution, GDP growth is locally flat. The
existence of conditional convergence based on initial GDP can be tested by using a
nonlinear specification for g0 . In order to study the overall convergence, the data is
divided into quartiles. An average derivative is then estimated within each quartile.
In addition, an overall average derivative is estimated over the support of all initial
GDP observations. The respective average derivatives are then compared.
Estimates based on Post-Nonparametric Double Selection are presented in Table 1. The estimate for the overall weighted average derivative is -0.042 (std. err.
= 0.014, p = 0.003).25 The estimate is negative and statistically significant. This
result is consistent with convergence theory. In addition, the density-weighted average derivative is calculated for various smaller ranges of initial GDP. The empirical
distribution of initial GDP is divided into quartiles. Estimates for the weighted
average derivatives are calculated within each quartile. The estimated average
derivatives are 0.013 (std. err. = 0.022, p = 0.568) for Q1 , -0.042 (std. err. =
0.025, p = 0.087) for Q2 , -0.100 (std. err. = 0.041, p = 0.015) for Q3 , -0.062
(std. err. = 0.027, p = 0.022) for Q4 . The test of the hypothesis that the average
derivative in Q1 is equal to the average derivative over Q2 , Q3 , Q4 rejects the null
at the 5% level (p = 0.009, t stat. = -2.632). The test of the hypothesis that the
average derivative in Q4 is equal to the average derivative over Q1 , Q2 , Q3 fails to
reject the null at the 5% level (p = 0.908, t stat. = -0.115). The overall average
derivative estimate is negative and statistically significant. These estimates also
agree with and thus support the previous findings reported in [24], [6], [7], which
relied on ad-hoc reasoning for covariate selection. In addition, the analysis supports
the claim that conditional convergence is nonlinear in initial GDP, being flatter for
countries with lower initial GDP.
25p-values are calculated against a two-sided alternative for the null that the average derivative
is 0.
29
ADDITIVELY SEPARABLE
Table 1. Estimation Results for GDP Example.
Estimates
( S. E. )
[
95% C. I.
]
Average Derivative
-0.042
(0.014)
[ -0.070 -0.014 ]
Quartile-Specific Average Derivative
Q1
Q2
Q3
Q4
-0.013
-0.042
-0.100
-0.062
(0.022)
(0.025)
(0.041)
(0.027)
[
[
[
[
-0.031 0.056 ]
-0.090 0.006 ]
-0.182 -0.019 ]
-0.115 -0.009 ]
Additional Selected Variables
•
•
•
•
•
•
•
Life expectancy
Average schooling years in female population over age 25
Infant mortality rate
Female gross enrollment ratio for secondary education
Male gross enrollment ratio for secondary education
Total fertility rate
Population proportion under 15
Additional Hypothesis Tests
H0 : ave. deriv. (Q1 ) = ave. deriv. (Q2 -Q4 )
p-value = .009
t-statistic = -2.632
H0 : ave. deriv. (Q4 ) = ave. deriv. (Q1 -Q3 )
p-value = .9084
t-statistic = 0.115
Note. Post-Nonparametric Double Selection estimates with
b = 7.
B-spline basis. K
7. Conclusion
This paper considers the problem of selecting a conditioning set in the context of
nonparametric regression. Convergence rates and inference results are provided for
series estimators of a primary component of interest in additively separable models
with high-dimensional conditioning information. The finite sample performance of
several Post-Nonparametric Double Selection estimators are evaluated in a simulation study. Overall, the proposed Span option has good estimation and inferential
properties in the data generating processes considered.
30
DAMIAN KOZBUR
Appendix A. Implementation Details
A.1. Lasso Implementation Details.
A.1.1. Lasso implementation given penalty λ. In every case, penalty loadings `j are
chosen as described in [8] with one small modification. The procedure suggested in
[8] requires an initial penalty loadings which are constructed using initial estimates
= yi followed by an iterative
of regression residuals. Their suggestion is to use εbinitial
i
procedure. Here, instead, εbinitial
are
taken
as
the
linear
regression residuals after
i
regressing the outcome v on the 5 most marginally correlated qjL , ie, the 5 which
have the highest |corr(v,
d qjL (z)|. Such modification was also used in [30].
A.1.2. Penalty level choice for single outcome. In every case when a single outcome variable is considered in isolation (this includes the reduced form selection
step and the selection step corresponding to ΦK 1 ), Lasso is implemented with
penalty λ as described in [8]. For ease of reference, note that [8] suggest λ given by
−1
2cLasso FN(0,1)
(1 − αLasso /L) where cLasso > 1, αLasso → 0 are tuning parameters. In
every instance in this paper, cLasso = 1.01 and αLasso = .05 are used.
A.1.3. Penalty level choice for ΦK,Simple . In this case, K Lasso regressions are run
−1
simultaneously. In this case, for all ϕ ∈ ΦK , λ is given by 2cLasso FN(0,1)
(1 −
αLasso /L) where cLasso = 1.01 and αLasso = .05/K are used.
A.1.4. Penalty level choice and implementation for ΦK,Span . When the Span option
is used, ΦK,Span is decomposed ΦK, Span = ΦK 1 ∪ ΦK 2 ∪ ΦK 3 . Each component has
a corresponding penalty level applied to all ϕ within that component. On the first
−1
component, λΦK 1 = 2cLasso FN(0,1)
(1 − αLasso /L) where cLasso = 1.01 and αLasso =
−1
.05. On the second component, λΦK 2 = 2cLasso FN(0,1)
(1 − αLasso /L) where cLasso =
−1
(1 −
1.01 and αLasso = .05/K. On the third component, λΦK 3 = 2cLasso FN(0,1)
αLasso /L) where cLasso = 1.01 and αLasso = .05/K.
The following procedure is used for approximating IΦK in the case that a component of ΦK contains a continuum of test functions. For each j 6 L, a Lasso
regression ϕ̌j ∈ ΦK3 which is more likely to select qjL(z) than other ϕ ∈ ΦK .
Specifically, for each j, ϕ̌j is set to the linear combination of p1K , ..., pKK with
highest marginal correlation to qjL . Then the approximation to the first stage
S
model selection step proceeds by using IˇΦK 3 = j6L Iϕ̌j (x) in place of IΦK 3 .
A.1.5. Penalty level choice for ΦK,Span-Conservative . When the Conservative Span
option is used, ΦK, Span-Conservative is decomposed ΦK, Span-Conservative = ΦK 1 ∪
ΦK 2 ∪ ΦK 3 . Each component again has a corresponding penalty level applied to
−1
all ϕ within that component. On the first component, λΦK 1 = 2cLasso FN(0,1)
(1 −
αLasso /L) where cLasso = 1.01 and αLasso = .05. On the second component, λΦK 2 =
−1
2cLasso FN(0,1)
(1 − αLasso /L) where cLasso = 1.01 and αLasso = .05/K. On the third
−1
component, λΦK 3 = 2cLasso FN(0,1)
(1 − αLasso /L) where cLasso = 1.01K 1/2 and
αLasso = .05.
In order to approximate the variables selected on the continuum of Lasso estimates indexed by ΦK 3 , the identical procedure with the Span option above is used.
Note that the only difference between the Conservative Span option and the Span
option is in λΦK 3 .
ADDITIVELY SEPARABLE
31
A.2. pK Implementation Details.
In every simulation and in the empirical example, pK is constructed using a cubic
B-spline expansion. For fixed K, the approximating dictionary is chosen according
to the following procedure. Knots points t1 , ..., tK−3 are chosen according to the
following rule. Set
tmax = quantile0.95 (|x1 |, ..., |xn |) and tmin = −tmax .
Let ∆k = tk − tk−1 . For constants c1 , c2 > 0 set
∆k = c1 + c2 |(K − 2)/2 − k|
for k = 2, ..., K − 3.
The constants c1 , c2 serve to insert more knot points where the density of x
is higher. The choices for c1 , c2 are determined uniquely by the condition that
c1 = 2c2 and that the endpoints satisfy t1 = tmin and tK−3 = tmax . Next, the
B-spline formulation used here is given by the recursive formulation. Set
Bk,0 (x) = 1tk 6x<tk+1 .
Set Bk,0 = 0 for k outside of 1, ..., K − 3. In addition, for spline order o > 0,
tk+o+1 − x
x − tk
Bk,o−1 +
Bk+1,o−1 .
Bk,o (x) =
tk+o − tk
tk+o+1 − tk+1
Set (p1,K (x), ..., pK−3,K (x)) = (B1,3 (x), ...BK−3,3 (x)). The dictionary is completed
by adding the additional terms pK−2,K (x) = x, pK−1,K (x) = x2 , pK,K (x) = x3 .
b is chosen according to the following procedure. First, an initial set of terms
K
q initial (z) ⊆ q L (z) is selected. In each case, q initial (z) contains the terms IRF .
That is, the terms selected in a Lasso regression y on q L (z). Next, an initial value
b 0 6 2bn1/3 c is chosen to minimize BIC using (pK (x), q initial (z)). In the simulation
K
b 0 is constrained to be > 5. Finally, in order to ensure undersmoothing, K
b
study, K
b
b
is set to K = b(log10 (n))K0 c.
A.3. Targeted Undersmoothing Implementation Details. The following
procedure is used to estimate the Targeted Undersmoothing (TU; specifically
c K,I (θ0 ) be
TU(1); see [30]) confidence intervals for θ0 . For each I ⊆ {1, ..., p} let CI
the corresponding confidence interval for θ0 using K terms and the components of
q L corresponding to I. Then the full TU confidence interval is defined by the convex
cb
hull of ∪j6p CI
K,IRF ∪{j} (θ0 ). In this implementation, a truncated TU confidence incb
terval is calculated instead: ∪j62s CI
(θ0 ). This is done be the simulation
0
K,IRF ∪{j}
run time reduces to the order of a day (from the order of a month), and therefore
helps facilitate easier replicability. Changing the code to calculate the full TU confidence intervals is trivial. This also highlights that computing speed is another
advantage of the Post-Nonparametric Double procedure relative to TU in certain
settings. In terms of approximation error, the full TU estimator was implemented
for the case n = 100, p = 50 for 1000 replications. The full TU confidence intervals
as well as the truncated TU confidence intervals each made 9 false rejections. In
addition, the average interval length for the full TU intervals was 1.740 while the
average interval length for the truncated TU intervals was 1.722. Therefore, the
truncated and full TU confidence intervals show very similar performance in this
instance.
32
DAMIAN KOZBUR
Appendix B. Proofs
B.1. Preliminary Setup and Additional Notation. Throughout the course of
the proof, as much reference as possible is made to results in [45], [15]. This is done
in order to maximize clarity and to present a better picture of the overall argument.
In many cases, appealing directly to arguments in [45] is possible because many of
the bounds required for deriving asymptotic normality for series estimators depend
only on properties of gb, g0 , pK and D. Less direct appeal to bounds in the original
Post-Double Selection argument is possible, since those arguments do not track K,
and do not have notions of quantities stemming from ΦK like αρ , αΦ . However, the
main idea of decomposing pK into components in the span of, and orthogonal to
q L , remains as a theme throughout the proofs.
For any function ϕ, let ϕ(X) denote the vector [ϕ(x1 ), ϕ(x2 ), ..., ϕ(xn )]0 . Similarly, let φqL ϕ(Z) = [πqL ϕ(z1 ), πqL ϕ(z2 ), ..., πqL ϕ(zn )]0 . In addition, define the
following quantities.
1. Let m be the n × K matrix m = πqL pK (Z) = [πqL p1K (Z), ..., πqL pKK (Z)]
3. Let W = P − m
b = n−1 P 0 M̃ P
4. Let Ω
5. Let Ω = n−1 E[W 0 W ]
6. Let Ω̄ = n−1 W 0 W
7. Let m be partitioned m = [m1 , ..., mK ]
8. Let W be partitioned m = [W1 , ..., WK ]
9. For any ϕ, let Rϕ = Q(βϕ,L − βϕ,L,s0 )
10. Let Ry = Q(βy,L − βy,L,s0 )
11. For any ϕ, let Uϕ = ϕ(X) − Qβϕ,L
12. Let Uy = Y − Qβy,L
13. Let F = V −1/2
14. Let ϕa (x) be the function such that πqL ϕa (Z) = F A0 m
15. Let ma = F A0 m
16. Let Wa = ϕa (X) − ma .
Assume without loss of generality that BK = IdK , the identity matrix of order
K. The reason this is without loss of generality is that dictionary pK is used only
in the post-selection estimation, while ΦK is used for first stage model selection. In
addition, assume without loss of generality that Ω = IdK .
Throughout the exposition, there is a common naming convention for various
regression coefficients. Quantities of the form βbv,I always denotes the sample regression coefficients from regressing the variable v on the components specified
by I. This implies that the quantities βbϕ,Iϕ,L = βbϕ,L,Post-Lasso are equivalent,
since the specified components being regressed on are the same. In addition,
βbϕ,IΦK +RF = βbϕ,q̃ = βbϕ(X),IΦK +RF are equivalent. Next, quantities of the form
βv,L and βv,L,s0 without a hat accent are population quantities and are defined in
the text above.
ADDITIVELY SEPARABLE
33
B.2. Preliminary Lemmas.
Lemma 1. Under the assumptions of Theorem 1,
1. J1 := maxk6K n−1/2 kQ0 Wk k∞ = Op ( log(KL)1/2 )
2. J2 := n−1/2 kQ0 Ek∞ = Op ( log(L)1/2 )
0
3. J3 := n−1/2 kRm
Ek2 = Op ((KK αρ L−αZ ))
4. J4 := n−1/2 kRh0 0 W k2 = Op ( n1/2 ζ0 (K)L−αZ )
5. J5 := maxk6K n−1/2 kMmk k2
1/2
= Op ( n−1/2 K αρ K αΦ /2 s0 log(L)1/2 + L−αZ K αρ )
1/2
6. J6 := n−1/2 kMh0 (Z)k2 = Op ( n−1/2 K αρ K αΦ /2 s0 log(L)1/2 + L−αZ K αρ )
7. J7 := maxk6K kβbmk ,IΦK +RF − βpkK ,L,s0 k1
= Op ( n−1/2 K αρ K αΦ /2 s0 K αIΦ /2 log(L)1/2 + L−αZ K αρ )
8. J8 := kβbh0 ,IΦK +RF − βh0 ,L,s0 k1
= Op ( n−1/2 K αρ K αΦ /2 s0 K αIΦ /2 log(L)1/2 + L−αZ K αρ )
1/2
9. J9 := maxk6K kβbWk ,IΦK +RF k1 = Op ( n−1/2 s0 K αIΦ /2 log(KL)1/2 )
1/2
10. J10 := kβbE,IΦK +RF k1 = Op ( n−1/2 s0 K αIΦ /2 log(L)1/2 )
11. J11 := n−1/2 kQ0 Wa k∞ = Op ( log(KL)1/2 )
0
Ek2 = Op ((KK αρ L−αZ ))
12. J12 := n−1/2 kRm
a
1/2
13. J13 := n−1/2 kMma k2 = Op ( n−1/2 K αρ K αΦ /2 s0 log(L)1/2 + L−αZ K αρ )
14. J14 := kβbma ,IΦK +RF − βϕa ,L,s0 k1
= Op ( n−1/2 K αρ K αΦ /2 s0 K αIΦ /2 log(L)1/2 + L−αZ K αρ )
1/2
15. J15 := kβbWa ,IΦK +RF k1 = Op ( n−1/2 s0 K αIΦ /2 log(KL)1/2 )
0
15. J16 := n−1 kRm
W kF = Op (K 1/2 ζ0 (K)K αρ L−αZ ).
Proof.
Statement
1.
By Lemma 5 of [15],
|Q0 Wk |
are sufficient for maxk6Kj6L √Pn j
2
i=1
E[|qjL (z)|3 |Wik |3 ]1/3
2 ]1/2
E[qjL (z)2 Wik
−1 1/3
maxk6K,j6L
two conditions which together
2
qjL (zi ) Wki
= Op ((log KL)1/2 ) are that
= O(ζ0 (K)) and the rate condition log KL =
2 1/2
o(ζ0 (K) n ). Note that E[qjL (z)2 Wik
]
is bounded away from zero by assumption. In addition, by Hölder’s inequality, E[|qjL (z)|3 |Wik |3 ] 6 E[|qjL (z)|3 ]ζ0 (K)3 .
34
DAMIAN KOZBUR
This implies that the first condition holds. The second condition is given in the
assumptions.
Statement 2. Follows similarly as Statement 1.
Statement 3. This statement follows directly from the fact that E[ε|x, z] = 0,
0
E[ε2 |x, z] bounded, along with dim(Rm
E) = K and kRmk k∞ = O(L−αZ ), allowing
the use of the K-dimensional Chebyshev Inequality.
P
Statement 4. kRh0 0 W k2 = k i Rh0 ,i Wi k2 6 O(L−αZ )ζ0 (K) by the facts that
Rh0 = O(L−αZ ) and kWi k2 6 ζ0 (K).
Statement 5.
First note that the following two hold.
1. For any ϕ ∈ ΦK , MπqL ϕ(Z) = MRϕ + M(Qβϕ,L,s0 − Qβbϕ,Iϕ,L ).
2. For any g ∈ LinSpan(pK ), and any corresponding expansion g = η1 ϕ1 +
... + ηkg ϕkg + rg with η1 , ...ηkg ∈ R, ϕ1 , ...ϕkg ∈ ΦK ,
kMπqL g(Z)k2 6 kηk1
max
ϕ∈{ϕ1 ,...,ϕkg }
kQβϕ,L,s0 − Qβbϕ,Iϕ,L k2 + kηk1 kRg k2 + krg (Z)k2 .
To show the first of the above two statements, for each ϕ ∈ ΦK , note that
MπqL ϕ(Z) = MMπqL ϕ(Z)
= M(πqL ϕ(Z) − PπqL ϕ(Z))
= M(Qβϕ,L − P(ϕ(X) − Uϕ ))
= M(Qβϕ,L − Qβbϕ,IΦK +RF + PUϕ )
= MRϕ + M(Qβϕ,L,s0 − Qβbϕ,IΦK +RF ) + MPUϕ
= MRϕ + M(Qβϕ,L,s0 − Qβbϕ,IΦK +RF )
= MRϕ + M(Qβϕ,L,s0 − Qβbϕ,Iϕ,L ) + M(Qβbϕ,Iϕ,L − Qβbϕ,IΦK +RF )
{z
}
|
= M(PIϕ,L ϕ(X) − Pϕ(X))
= MP(PIϕ,L ϕ(X) − ϕ(X))
=0
⇒ MπqL ϕ(Z) = MRϕ + M(Qβϕ,L,s0 − Qβbϕ,Iϕ,L ).
This establishes the first claim. Now turn to the second claim. Note that using
the density assumption, there are ϕ1 , ..., ϕkg and a vector η = (η1 , ..., ηkg ) such that
g = η1 ϕ1 + ... + ηkg ϕkg + rg for some remainder rg , sufficiently small. Then
kMπqL g(Z)k2 = kη1 MπqL ϕ1 (X) + ... + ηkg MπqL ϕkg (Z) + MπqL rg (Z)k2
Next, looking at each ϕ in the above expansion (ie each ϕ ∈ {ϕ1 , ..., ϕkg }) and
combining the above expression gives
35
ADDITIVELY SEPARABLE
kMπqL g(Z)k2 = kη1 MRϕ1 + η1 M(Qβϕ1 ,L,s0 − Qβbϕ1 ,Iϕ1 ) + ...
... + ηkg MRϕkg + ηkg M(Qβϕkg ,L,s0 − Qβbϕkg ,Iϕk ) + MπqL rg (Z)k2 .
g
Applying Hölder’s inequality and the fact that M is a projection (and hence
non-expansive) gives the bound
6 kηk1
max
ϕ∈{ϕ1 ,...,ϕkg }
kQβϕ,L,s0 − Qβbϕ,Iϕ,L k2 + kηk1 kRg k2 + krg (Z)k2 .
These can then be applied directly to n−1/2 kMmk k2 . Under Assumption 9, note
that for mk , the corresponding η and Rmk satisfy kηk1 6 O(K αρ ) and kRmk k2 6
n1/2 O(L−αZ K αρ ). Then we have the bound
1/2
kMπqL g(Z)k2 = Op (K αρ K αΦ /2 s0
log(L)1/2 + n1/2 L−αZ K αρ ).
Under Assumption 10, note that for each mk , taking η = 1 and Rmk = 0 are
feasible by assumption. The result follows.
Statement 6.
n−1/2 kMh0 (Z)k2 = n−1/2 kM(Qβh0 ,L,s0 + Rh0 )k2 )
6 n−1/2 (kMQβh0 ,L,s0 k2 + kMRh0 k2 )
6 n−1/2 (kMQ(βg0 ,L,s0 + βh0 ,L,s0 − Qβg0 ,L,s0 )k2 + kMRh0 )k2
= n−1/2 (kM(Qβy,L,s0 − Qβg0 ,L,s0 )k2 + kMRh0 k2 )
6 n−1/2 (kMQβy,L,s0 k2 + kMQβg0 ,L,s0 k2 + kMRh0 k2 )
= n−1/2 (kMπqL y(Z)k2 + kMπqL g0 (Z)k2 + kMRh0 k2 )
The first two terms above, n−1/2 (kMQπqL y(Z)k2 + n−1/2 kMπqL g0 (Z)k2 ), are
1/2
Op (K αρ K αΦ /2 n−1/2 s0 log(L)1/2 +L−αZ K αρ ) by the same reasoning as Statement
−1/2
6. In addition n
kMRh0 k2 6 n−1/2 kRh0 k2 = O(L−αZ ) by assumption. This
gives
1/2
n−1/2 kMh0 (Z)k2 = Op (K αρ K αΦ /2 n−1/2 s0
log(L)1/2 + L−αZ K αρ ).
36
DAMIAN KOZBUR
Statement 7.
kβbmk ,IΦK +RF − βpkK ,L,s0 k1
6 |IΦK +RF |1/2 kβbmk ,IΦK +RF − βpkK ,L,s0 k2
1/2
= |IΦK +RF |1/2 (βbmk ,IΦK +RF − βpkK ,L,s0 )0 (βbmk ,IΦK +RF − βpkK ,L,s0 )
1/2
6 |IΦK +RF |1/2 Op (1) (βbmk ,IΦK +RF − βpkK ,L,s0 )0 (Q0IΦ +RF QIΦK +RF /n)(βbg,IΦK +RF − βpkK ,L,s0 )
K
= |IΦK +RF |1/2 Op (1)n−1/2 kPmk − QβpkK ,L,s0 k2
= |IΦK +RF |1/2 Op (1)n−1/2 kmk − Mmk − QβpkK ,L,s0 k2
= |IΦK +RF |1/2 Op (1)n−1/2 k − Mmk + Rmk k2
6 |IΦK +RF |1/2 Op (1)(J5 + O(L−αZ ))
1/2
= Op (s0 K αIΦ /2 )(J5 + O(L−αZ ))
αIΦ /2+1/2
= Op (s0
1/2
)Op (K αρ K αΦ /2 s0
log(L)1/2 + n1/2 L−αZ K αρ )
= Op (K αρ K αΦ /2 s10 K αIΦ /2 log(L)1/2 + n1/2 L−αZ K αρ ).
Statement 8. Proven analogously to Statement 7.
Statement 9.
max kβbWk ,IΦK +RF k1
k6K
= max k(Q̃0 Q̃)−1 Q̃0 Wk k1
k6K
6 ||IΦK +RF |1/2 max k(Q̃0 Q̃)−1 Q̃0 Wk k2
k6K
6
−1/2
|IΦK +RF |1/2 κmin (IΦK +RF |) max kn−1 Q0 Wk k∞
k6K
1/2
= Op (s0 K αIΦ /2 · 1 · n−1/2 log(KL)1/2 ).
Statement 10.
Proven analogously to Statement 9.
Statements 11-15.
Proven analogously to Statements 1,3,5,7,9.
Statement 16.
n−1
n
X
!1/2
Wi0 Rm,i
i=1
F
= n−1
X
kW 0 Rmk k22
k
!1/2
6 n−1
X
k
2
k
n2 ζ0 (K)2 kRm
k ∞
.
37
ADDITIVELY SEPARABLE
By the density assumption, kRmk k∞ 6 K αρ L−αZ . This then implies that
n
X
6 K 1/2 K αρ L−αZ .
n−1
Wi0 Rm,i
F
i=1
Lemma 2.
1. Ξ1 := n−1 kW 0 PW kF 6 n−1/2 KJ9 J1
2. Ξ2 := n−1 km0 MmkF 6 KJ52
3. Ξ3 := n−1 km0 MW kF 6 J16 + n−1/2 KJ7 J1
4. Ξ4 := n−1/2 km0 Mh0 (Z)k2 6 n1/2 K 1/2 J5 J6
5. Ξ5 := n−1/2 kW 0 Mh0 (Z)k2 6 J4 + K 1/2 J8 J1
6. Ξ6 := n−1/2 kW 0 PEk2 6 K 1/2 J9 J2
7. Ξ7 := n−1/2 km0 MEk2 6 J4 + K 1/2 J7 J2
8. Ξ8 := n−1/2 |m0a Mh0 (Z)| 6 n1/2 J5 J13
9. Ξ9 := n−1/2 |Wa0 Mh0 (Z)| 6 J12 + J14 J1
10. Ξ10 := n−1/2 |Wa0 PE| 6 J9 J11
11. Ξ11 := n−1/2 |m0a ME| 6 J12 + J7 J11 .
Proof.
Statement 1.
n−1 kW 0 PW kF
2
=
X
(n−1 Wk0 PWk̄ )2 =
k,k̄6K
=
X
0
(n−1 βbW
k ,IΦ
K +RF
Q0 Wk̄ )2
k,k̄6K
6
X
kn−1/2 βbWk ,IΦK +RF k21 kn−1/2 Q0 Wk̄ k2∞
k,k̄6K
=
X
kn−1/2 βbWk ,IΦK +RF k21
k6K
6K ·n
−1
X
k̄6K
J92
·K ·
J12
⇒ n−1 kW 0 PW kF 6 n−1/2 KJ1 J9 .
kn−1/2 Q0 Wk̄ k2∞
38
DAMIAN KOZBUR
Statement 2.
n−1 km0 MmkF
2
=
X
(n−1 m0k Mmk̄ )2 6
k,k̄6K
kn−1/2 Mmk k22 kn−1/2 Mmk̄ k22
k,k̄6K
2
=
X
X
kn−1/2 Mmk k22 6 K 2 J54
k6K
⇒ n−1 km0 MmkF 6 KJ52 .
Statement 3.
n−1 km0 MW kF = n−1 km0 W/n − m0 PW kF
0
= n−1 kRm
W + (QβpK ,L,s0 )0 W − m0 PW kF
0
= n−1 kRm
W + (QβpK ,L,s0 )0 W − (Qβbm,IΦK +RF )0 W kF
0
= n−1 kRm
W + (βpK ,L,s0 − βbm,IΦK +RF )0 Q0 W kF
0
6 n−1 kRm
W kF + n−1 k(βpK ,L,s0 − βbm,IΦK +RF )0 Q0 W kF .
0
W kF = J16 while
Then the first term in the last line is bounded above as n−1 kRm
the second term has
2
n−1 k(βpK ,L,s0 − βbm,IΦK +RF )0 Q0 W kF
X
= n−2
((βpk ,L,s0 − βbmk ,IΦK +RF )0 Q0 Wk̄ )2
k,k̄6K
6n
−2
X
kβpk ,L,s0 − βbmk ,IΦK +RF k21 kQ0 Wk̄ k2∞
k,k̄6K
= n−1
kβpk ,L,s0 − βbmk ,IΦK +RF k21
k6K
6n
−1
X
K·
J72
X
kn−1/2 Q0 Wk̄ k2∞
k̄6K
·K ·
J12 .
Therefore, n−1 km0 MW kF 6 J16 + n−1/2 KJ7 J1 .
Statement 4.
n−1/2 km0 Mh0 (Z)k2 6 n1/2 kn−1/2 Mh0 (Z)k2 K 1/2 max n−1/2 km0 Mk2
k6K
6n
1/2
K
1/2
J5 J6 .
39
ADDITIVELY SEPARABLE
Statement 5.
n−1/2 kW 0 Mh0 (Z)k2 = n−1/2 kW 0 h0 (Z) − W 0 Ph0 (Z)k2
= n−1/2 kW 0 h0 (Z) − W 0 Qβbh0 (Z),IΦK +RF k2
= n−1/2 kW 0 Rh0 + W 0 Qβh0 ,L,s0 − W 0 Qβbh0 (Z),IΦK +RF k2
= n−1/2 kW 0 h0 (Z)k2 + k(βbh0 (Z),IΦK +RF − β)0 Q0 W k2
6 J4 + K 1/2 max kβbh0 (Z),IΦK +RF − βh0 ,L,s0 k1 kn−1/2 Q0 Wk k∞
k6K
6 J4 + K
1/2
J8 J1 .
Statement 6.
2
X
(Wk0 PE)2
n−1/2 kW 0 PW k2 = n−1
k6K
=n
−1
X
0
(βbW
k ,IΦ
K +RF
Q0 E)2
k6K
6
X
kβbWk ,IΦK +RF k21 kn−1/2 Q0 Ek2∞
k6K
6 K · J92 · J22
⇒ n−1/2 kW 0 PEkF 6 K 1/2 J9 J2 .
Statement 7.
n−1/2 km0 MEk2 = n−1/2 km0 E/n − m0 PEkF
0
= n−1/2 kRm
E + (QβpK ,L,s0 )0 W − m0 PEkF
0
= n−1/2 kRm
E + (QβpK ,L,s0 )0 W − (Qβbm,IΦK +RF )0 Ek2
0
= n−1/2 kRm
E + (βpK ,L,s0 − βbm,IΦK +RF )0 Q0 Ek2
0
6 n−1/2 kRm
Ek2 + n−1/2 k(βpK ,L,s0 − βbm,IΦK +RF )0 Q0 Ek2 .
0
Then the first term in the last line is bounded above as n−1/2 kRm
Ek2 = J4 . Turning
to the second term,
2
n−1/2 k(βpK ,L,s0 − βbm,IΦK +RF )0 Q0 Ek2
X
= n−1
((βpk ,L,s0 − βbmk ,IΦK +RF )0 Q0 E)2
k6K
6
X
kβpk ,L,s0 − βbmk ,IΦK +RF k21 kn−1/2 Q0 Ek2∞
k6K
6 K · J72 · J22 .
Therefore, n−1/2 km0 MEk2 6 J4 + K 1/2 J7 J2 .
40
DAMIAN KOZBUR
Statements 8-11.
The argument is identical to the argument for Statements 4-7, adjusting appropriately for the fact that ma is 1-dimensional rather than K-dimensional.
The following corollaries follow directly from assumed rate conditions and the
above bounds. These are used in the proof of Theorems 1 and 2.
Corollary 1. Under the assumptions of Theorem 1,
1. Ξ1 + Ξ2 + Ξ3 = Op (n−1/2 ζ0 (K)K 1/2 )
2. n−1/2 (Ξ4 + Ξ5 + Ξ6 + Ξ7 ) = Op (n−1/2 K 1/2 + K −αg0 ).
Corollary 2. Under the assumptions of Theorem 2,
1. n−1/2 ζ0 (K)K 1/2 (Ξ4 + Ξ5 + Ξ6 + Ξ7 ) = op (1).
2. Ξ8 + Ξ9 + Ξ10 + Ξ11 = op (1).
B.3. Proof of Theorem 1.
Lemma 3.
b − ΩkF 6 Op (ζ0 (K)K 1/2 n−1/2 ) + Ξ1 + Ξ2 + Ξ3 = op (1)
1. kΩ
b
2. kΩ−1 − Ω−1 k2→2 = Op (ζ0 (K)K 1/2 n−1/2 ) + O(Ξ1 + Ξ2 + Ξ3 ).
Proof. The argument in Theorem 1 of [45] gives the bound kΩ̄ − ΩkF =
b =
OP (ζ0 (K)K 1/2 n−1/2 ). Next, using the decomposition, P = m + W , write Ω
0
0
0
0
0
(m + W ) M(m + W )/n = W W/n − W (Idn − M)W/n + m Mm/n + 2m MW/n.
b F 6 kW 0 PW/nkF + km0 Mm/nkF + k2m0 MW/nkF =
By triangle inquality, kΩ̄ − Ωk
Ξ1 + Ξ2 + Ξ3 . Bounds for each of the three above terms are established above along
b F = op (1). The last statement holds
with the assumed rate conditions give kΩ̄ − Ωk
by applying an expansion of the matrix inversion function around IdK .
b −1 = (IdK − (IdK − Ω))
b −1 = IdK + (IdK − Ω)
b + (IdK − Ω)
b 2 + ...
Ω
The sum given above is with probability → 1 absolutely convergent relative to the
b −1 −
Frobenius norm F. In addition, by the bound k · k2→2 6 k · kF , we have kΩ
2
1/2 −1/2
b
b
b
IdK k2→2 6 kΩ − IdK kF 6 kIdK − ΩkF + kIdK − ΩkF + ... = Op (ζ0 (K)K n
)+
O(Ξ1 + Ξ2 + Ξ3 ).
Note that since Ω has minimal eigenvalues bounded from below by assumption, it
b and Ω̄ are invertible with probability approaching 1. The reference
follows that Ω
b > 1/2} and later uses the fact that this
[45] works on the event 1n := {λmin (Ω)
event has probability → 1. This fact is used several times, however its use is only
implicitly in reference to arguments in [45].
b −1 n−1 P 0 MEk2 = Op (n−1/2 K 1/2 ).
Lemma 4. kΩ
Proof.
b −1 n−1 P 0 MEk2 6 kΩ
b −1 k2→2 n−1 kP 0 MEk2
kΩ
b −1 k2→2 (n−1 kW 0 Ek2 + n−1/2 Ξ6 + n−1/2 Ξ7 )
6 kΩ
kW 0 Ek2 = Op (n−1/2 K 1/2 = Op (n−1/2 ) by arguments in [45]. Bounds for n−1/2 Ξ6 +
n−1/2 Ξ7 follows from the previous Lemmas and from the assumed rate conditions.
ADDITIVELY SEPARABLE
41
b −1 P 0 M(g0 (X) − P βg ,K )/nk2 = Op (K −αg0 ).
Lemma 5. kΩ
0
Proof.
b −1 P 0 M(g0 (X) − P βg
kΩ
0 ,K
b −1 P 0 M(g0 (X) − P β)/n]1/2
)/nk2 = [(g0 (X) − P β)0 MP Ω
= Op (1)[(g0 (X) − P βg0 ,K )0 (g0 (Z) − P βg0 ,K )/n]1/2
= Op (K −αg0 )
b −1 P 0 M =
by assumption on (g0 (X) − P βg0 ,K ) and idempotency of MP Ω
0
−1
MP (P MMP ) M.
b −1 P 0 Mh0 (Z)/nk2 = op (n−1/2 ).
Lemma 6. kΩ
b has eigenvalues bounded below and above with probability approaching
Proof. Ω
1. Then,
b −1 P 0 Mh0 (Z)/nk2 6 Op (1)kP 0 Mh0 (Z)/nk2
kΩ
= Op (1)k(m + W )0 Mh0 (Z)/nk2
= n−1/2 Op (1)n−1/2 k(m + W )0 Mh0 (Z)/nk2
6 n−1/2 Op (1)(n−1/2 km0 Mh0 (Z)k2 + n−1/2 kW 0 Mh0 (Z)k2 )
= n−1/2 Op (1)(Ξ4 + Ξ5 )
= n−1/2 Op (1)op (1).
Lemma 7. kβbg − βg0 ,K k2 = Op (n−1/2 K 1/2 + K −αg0 ).
b −1 P 0 M b(g0 (X) −
b −1 P 0 ME + n−1 Ω
Proof. Note that ([βby,(p̃,q̃) ]g − βg0 ,K ) = n−1 Ω
I
−1 b −1 0
P βg0 ,K ) + n Ω P Mh0 (Z). Triangle inequality in conjuction with the bounds
described in the previous three lemmas give the result.
The final statement of Theorem 1 follows from the bound on kβbg − βg0 ,K k2 using
the arguments in [45].
B.4. Proof of Theorem 2. Recall that F = V −1/2 . Let ḡ = pK (x)0 βg0 ,K and
decompose the quantity n1/2 F [a(b
g ) − a(g0 )] by
n1/2 F [a(b
g ) − a(g0 )] = n1/2 F [a(b
g ) − a(g0 ) + D(b
g ) − D(g0 )
+ D(ḡ) − D(b
g)
+ D(g0 ) − D(ḡ)].
Lemma 8. n1/2 F [D(ḡ) − D(g0 )] = O(n1/2 K −αg0 ).
Proof. This follows from arguments given in the proof of Theorem 2 in [45]. Note
that the statement does not contain any reference to random quantities.
Lemma 9. |n1/2 F [a(b
g ) − a(g) − D(b
g ) + D(g)| = op (1).
Proof. Bounds on |b
g − g|d given by Theorem 1 imply that |n1/2 F [a(b
g ) − a(g0 ) −
D(b
g ) + D(g0 )| 6 Cn1/2 |b
g − g0 |2d = Op (n1/2 (n−1/2 ζd (K)K 1/2 + K −αg0 )2 ) = op (1).
This is again identical to the reasoning given in Theorem 2 in [45], since that
references uses only a bound on |b
g − g|d to prove the analogous result.
42
DAMIAN KOZBUR
The last step is to show that n1/2 F [D(b
g ) − D(ḡ)] →d N(0, 1).
Lemma 10. n1/2 F [D(b
g ) − D(ḡ)] →d N(0, 1).
Proof. Note that D(b
g ) can be expanded
K
b −1 n−1 P 0 MY )
D(b
g ) = D(p (x)0 [βby,(p̃,q̃) ]g ) = D(pK (x)0 Ω
b −1 n−1 P 0 M(g0 (X) + h0 (Z) + E))
= D(pK (x)0 Ω
b −1 n−1 (g0 (X) + h0 (Z) + E)
= D(pK (x))0 Ω
b −1 n−1 P 0 M(g0 (X) + h0 (Z) + E)
= A0 Ω
b −1 n−1 P 0 Mg0 (X) + A0 Ω
b −1 n−1 P 0 Mh0 (Z) + A0 Ω
b −1 n−1 P 0 ME.
= A0 Ω
In addition, D(ḡ) = D(pK (x)0 βg0 ,K ) = A0 βg0 ,K gives
b −1 n−1 P 0 Mg0 (X) − βg ,K ]
n1/2 F [D(b
g ) − D(ḡ)] = n1/2 F A0 [Ω
0
b −1 n−1 P 0 M(h0 (Z) + E)].
+ n1/2 F A0 [Ω
The above equation gives a decomposition of the right hand side into two terms,
which are next bounded separately. Before proceeding, note that the followb −1/2 k2 = Op (1),
b −1 k2 = Op (1), kF A0 Ω
ing bounds kF Ak2 = O(1), kF A0 Ω
0 −1
0 −1/2
kF A Ω k2 = Op (1), kF A Ω
k2 = O(1) all hold by arguments in [45]. Consider the first term.
b −1 P 0 Mg0 (X) − βg ,K ]|
|n1/2 F A0 [n−1 Ω
0
√
= | nF A0 [(P 0 MP/n)−1 P 0 M(G − P β)/n]|
√
b −1 P 0 M/ nk2 kg0 (X) − P βg ,K k2
6 kF A0 Ω
0
√
√
0 b −1 0
6 kF A Ω P M/ nk2 n max |g(xi ) − ḡ(xi )|
i6n
√
0 b −1/2
= kF A Ω
k2 n max |g(xi ) − ḡ(xi )|
i6n
√
0 b −1/2
6 kF A Ω
k2 n|g − ḡ|0
√
= Op (1)Op ( nK −α )
= op (1).
b −1 n−1 P 0 M(h0 (Z)+E). To handle this term, first bound
Next, consider n1/2 F A0 Ω
b −1 − Ω−1 )P 0 M(h0 (Z) + E)|
|n−1/2 F A0 (Ω
b −1 − Ω−1 )k2 kP 0 M(h0 (Z) + E)k2
6 n−1/2 kF A0 (Ω
b −1 − Ω−1 )k2 (n−1/2 kP 0 M(h0 (Z) + E)k2 )
= kF A0 (Ω
b −1 − Ω−1 )k2 (n−1/2 k(m + W )0 M(h0 (Z) + E)k2 )
= kF A0 (Ω
b −1 − Ω−1 )k2 (Ξ4 + Ξ5 + Ξ6 + Ξ7 )
6 kF A0 (Ω
b −1 − Ω−1 k2→2 kF A0 k2 (Ξ4 + Ξ5 + Ξ6 + Ξ7 )
6 kΩ
= op (1).
ADDITIVELY SEPARABLE
43
Next consider the last remaining term for which a central limit result will be shown.
√
nF A0 Ω−1 P 0 M(h0 (Z) + E)/n
√
= nF A0 Ω−1 (W + m)0 M(h0 (Z) + E)/n
√
= n(F A0 Ω−1 W + ma )0 M(h0 (Z) + E)/n
√
√
√
= nF A0 Ω−1 W ME + nm0a M(h0 (Z) + E)/n + nWa0 Mh0 (Z)/n
√
√
√
√
= nF A0 Ω−1 W E − nF A0 Ω−1 W PE + nm0a M(h0 (Z) + E)/n + nWa0 Mh0 (Z)/n
√
= nF A0 Ω−1 W 0 E/n + op (1).
bound in the equation
array above holds by the fact that
√
√ Note that the last√op (1)
0
0
W
PE+
nm
M(h
(Z)+E)/n+
nW
Mh0 (Z)/n| 6 Ξ8 +Ξ9 +Ξ10 +Ξ11 .
| nF A0 Ω−1
0
a
a
√
0 −1
0
The term nF A Ω W E/n satisfies the conditions Lindbergh-Feller Central Limit
Theorem, by arguments given in [45].
The previous three lemmas prove that n1/2 F [a(b
g ) − a(g0 )] → N(0, 1).
The next set of arguments bound Vb −V . For ν as in the statement of Assumption
b and u =
b −1 AF
14, Define the event Ag = {|b
g − g0 |d < ν/2}. Define u
b = 1A g Ω
P
−1
0 2
1Ag Ω AF . In addition, define Σ̄ = i Wi Wi εi /n, an infeasible sample analogue
of Σ.
Lemma 11.
b − Ak2 = op (1)
1. kA
2. kb
u − uk2 = op (1)
3. kΣ̄ − ΣkF = op (1)
4. |b
uΣ̄b
u−u
b0 Σb
u| = op (1).
Proof. Statement 1. In the case that a(g) is linear in g, then a(p0 β) = A0 β =⇒
b = A. Therefore, consider the case that a(g) is not linear in g. Using arguments
A
identical to those in [45], 1Ag = 1 with probability → 1, and
b − Ak2 6 C · ζd (K)|b
1Ag kA
g − g|d .
Statement 2. This follows from arguments in [45].
Statement 3. This follows from arguments in [45].
Statement 4. An immediate implication of Statement 3 is that 1Ag |b
uΣ̄b
u−u
b0 Σb
u| =
2
0
2
|b
u (Σ̄ − Σ)b
u| 6 kb
uk2 kΣ̄ − Σk2→2 = Op (1)op (1).
Lemma 12. maxi6n |h0 (zi ) − b
h(zi )| = op (1).
Proof. First note that
max |h0 (zi ) − b
h(zi )| 6 max |h0 (zi ) − q L (zi )0 βh0 ,L,s0 |
i
i
+ max |q L (zi )0 [βby,(p̃,q̃) ]h − q L (zi )0 βh0 ,L,s0 |.
i
44
DAMIAN KOZBUR
The first term has the bound maxi |h0 (zi ) − q(xi )0 η| = Op (L−αZ ) by assumption.
Next,
max |q L (zi )0 [βby,(p̃,q̃) ]h − q L (zi )0 βh ,L,s | = max |q L (zi )0 ([βby,(p̃,q̃) ]h − βh ,L,s )|
0
i
0
0
i
0
L
6 max kq (zi )k∞ k[βby,(p̃,q̃) ]h − βh0 ,L,s0 k1
i
Then,
k[βby,(p̃,q̃) ]h − βh
0 ,L,s0
k1 = kβby−bg,IΦK +RF − βh0 ,L,s0 k1
= kβbg0 ,IΦK +RF + βbh0 ,IΦK +RF + βbε,IΦK +RF − βbgb,IΦK +RF − βh0 ,L,s0 k1 .
6 kβbh0 ,IΦK +RF − βh0 ,L,s0 k1 + kβbε,IΦK +RF k1 + kβbg0 −bg,IΦK +RF k1
= J8 + J10 + kβbg0 −bg,IΦK +RF k1 .
Next,
kβbg −bg,I
0
ΦK +RF
k1
6 |IΦK +RF |1/2 kβbg0 −bg,IΦK +RF k2
= |IΦK +RF |1/2 k(Q0IΦ
K +RF
QIΦK +RF /n)−1 Q0IΦ
6 |IΦK +RF |1/2 κmin (|IΦK +RF |)−1 kQ0IΦ
1/2
6 |IΦK +RF |
κmin (|IΦK +RF |)
−1
|IΦK +RF |
6 |IΦK +RF |κmin (|IΦK +RF |)−1 kQ0IΦ
= |IΦK +RF |κmin (|IΦK +RF |)−1
=
K +RF
K +RF
(g0 (X) − gb(X))/nk2
(g0 (X) − gb(X))/nk2
1/2
kQ0IΦ
K +RF
(g0 (X) − gb(X))/nk∞
(g0 (X) − gb(X))/nk∞
!
n
X
−1
max n
|qjL (zi )| kg0 (X) − gb(X)k∞
K +RF
j
i=1
Op (s10 K αIΦ )Op (1)op (n−1/2 ζ0 (K)K 1/2
+ K −αg0 ).
Putting these together, it follows from the assumed rate conditions that
max |h0 (zi ) − b
h(zi )| = op (1).
i
Next, let ∆g0 i = g0 (xi ) − gb(xi ) and ∆h0 i = h0 (zi ) − b
h(zi ). Then above lemma
states maxi6n ∆h0 i = op (1). In addition maxi6n |∆g0 i | 6 |b
g − g|0 = op (1). Let
ωi2 = u0 Wi Wi0 u and ω
bi2 = u
b0 Wi Wi0 u
b.
Lemma 13. |F Vb F − u
bΣ̄b
u| = op (1).
Proof.
b − Σ̄)b
1Ag |F Vb F − u
b0 Σ̄b
u| = |b
u0 (Σ
u| =
n
X
ci W
ci0 εb2i u
u
b0 W
b/n −
i=1
6
n
X
i=1
ωi2 (b
ε2i − ε2i )/n +
n
X
u
b0 Wi Wi0 ε2i u
b/n
i=1
n
X
i=1
(b
ωi2 − ωi2 )b
ε2i /n .
45
ADDITIVELY SEPARABLE
Both terms on the right hand side will be bounded. Consider the first term. Expanding (b
ε2i − ε2i ) gives
n
X
n
X
ωi2 (b
ε2i − ε2i )/n 6
ωi2 ∆21i /n +
i=1
i=1
+2
n
X
n
X
ωi2 ∆22i /n +
i=1
ωi2 ∆1i εi /n + 2
n
X
n
X
ωi2 ∆1i ∆2i/n
i=1
ωi2 ∆2i εi /n .
i=1
Pn
Pni=1
Note that i=1 ωi2 /n, i=1 ωi2 |εi | = Op (1) by arguments is in [45]. The five terms
above are then bounded in order of their appearence by
n
n
X
X
ωi2 ∆21i /n 6 max |∆1i |
ωi2 /n = op (1)Op (1)
i=1
n
X
i6n
ωi2 ∆22i /n 6 max |∆2i |
i6n
i=1
n
X
ωi2 ∆1i ∆2i /n
i=1
n
X
6 max |∆1i | max |∆2i |
i6n
i6n
i6n
n
X
n
X
ωi2 /n = op (1)Op (1)
i=1
ωi2 |εi |/n = op (1)Op (1)
i=1
ωi2 ∆2i εi /n 6 max |∆2i |
i6n
i=1
ωi2 |εi |/n = op (1)Op (1)
i=1
ωi2 ∆1i εi /n 6 max |∆1i |
i=1
n
X
i=1
n
X
n
X
ωi2 |εi |/n = op (1)Op (1).
i=1
The second term is bounded by
n
X
ci W
ci0 − Wi Wi0 )b
u
b(W
ε2i u
b/n 6 max |ε2i |
i6n
i=1
n
X
ci W
ci0 − Wi Wi0 )b
u
b(W
u/n
i=1
n
X
2
2
ci W
ci0 − Wi Wi0 )/nk2→2 = max |b
b − Ω̄k2→2
6 max |b
εi |kb
uk2 k
(W
ε2i |kb
uk22 kkΩ
i6n
i6n
i=1
b − Ω̄k2→2
max |ε2i | + max |b
ε2i − ε2i | kb
uk22 kkΩ
i6n
i6n
= Op (n2/δ ) + op (1) Op (1)(Op (ζ0 (K)n−1/2 K 1/2 ) + Ξ1 + Ξ2 + Ξ3 )
6
= op (1).
where the last bounds come from the rate condition in Assumption 9 and
maxi6n |b
ε2i − ε2i | = op (1) by maxi6n |∆1i | + |∆2i | = op (1).
These results give the conclusion that
d
n1/2 Vb −1/2 (θb − θ) = n1/2 (F Vb F )−1/2 (θb − θ) → N (0, 1).
Calculations which give the rates of convergence in each of the cases of Assumption
17 or of Assumption 18, as well as the proof of the second statement of Theorem
2, use the same arguments as in [45]. This concludes the proof.
46
DAMIAN KOZBUR
References
[1] P. Aghion and P.W. Howitt. The Economics of Growth. MIT Press, 2008.
[2] Donald D.W. Andrews and Yoon-Jae Whang. Additive interactive regression models: Circumvention of the curse of dimensionality. Econometric Theory, 6(4):466479, 12 1990.
[3] Donald W. K. Andrews. Asymptotic normality of series estimators for nonparametric and
semiparametric regression models. Econometrica, 59(2):307–345, 1991.
[4] J. Bai and S. Ng. Forecasting economic time series using targeted predictors. Journal of
Econometrics, 146:304–317, 2008.
[5] J. Bai and S. Ng. Boosting diffusion indices. Journal of Applied Econometrics, 24, 2009.
[6] R.J. Barro and J.W. Lee. Data set for a panel of 139 countries.
NBER,http://www.nber.org/pub/barro.lee.html (1994).
[7] Robert J. Barro and Jong-Wha Lee. Losers and winners in economic growth. Working Paper
4341, National Bureau of Economic Research, April 1993.
[8] A. Belloni, D. Chen, V. Chernozhukov, and C. Hansen. Sparse models and methods for
optimal instruments with an application to eminent domain. Econometrica, 80:2369–2429,
2012. Arxiv, 2010.
[9] A. Belloni and V. Chernozhukov. Least squares after model selection in high-dimensional
sparse models. Bernoulli, 19(2):521–547, 2013. ArXiv, 2009.
[10] A. Belloni, V. Chernozhukov, I. Fernndez-Val, and C. Hansen. Program evaluation and causal
inference with high-dimensional data. Econometrica, 85(1):233–298, 2017.
[11] A. Belloni, V. Chernozhukov, and C. Hansen. Lasso methods for gaussian instrumental variables models. 2010 arXiv:[math.ST], http://arxiv.org/abs/1012.1297, 2010.
[12] A. Belloni, V. Chernozhukov, and C. Hansen. Inference for high-dimensional sparse econometric models. Advances in Economics and Econometrics. 10th World Congress of Econometric
Society. August 2010, III:245–295, 2013.
[13] Alexandre Belloni and Victor Chernozhukov. High Dimensional Sparse Econometric Models:
An Introduction, pages 121–156. Springer Berlin Heidelberg, Berlin, Heidelberg, 2011.
[14] Alexandre Belloni, Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Some new
asymptotic theory for least squares series: Pointwise and uniform results. Journal of Econometrics, 186(2):345 – 366, 2015. High Dimensional Problems in Econometrics.
[15] Alexandre Belloni, Victor Chernozhukov, and Christian Hansen. Inference on treatment effects after selection amongst high-dimensional controls with an application to abortion on
crime. Review of Economic Studies, 81(2):608–650, 2014.
[16] Alexandre Belloni, Victor Chernozhukov, Christian Hansen, and Damian Kozbur. Inference
in high-dimensional panel models with an application to gun control. Journal of Business &
Economic Statistics, 34(4):590–605, 2016.
[17] P. J. Bickel, Y. Ritov, and A. B. Tsybakov. Simultaneous analysis of Lasso and Dantzig
selector. Annals of Statistics, 37(4):1705–1732, 2009.
[18] P. Bühlmann and S. van de Geer. Statistics for High-Dimensional Data: Methods, Theory
and Applications. Springer, 2011.
[19] Andreas Buja, Trevor Hastie, and Robert Tibshirani. Linear smoothers and additive models.
Ann. Statist., 17(2):453–510, 06 1989.
[20] F. Bunea, A. Tsybakov, and M. H. Wegkamp. Sparsity oracle inequalities for the lasso.
Electronic Journal of Statistics, 1:169–194, 2007.
[21] F. Bunea, A. B. Tsybakov, , and M. H. Wegkamp. Aggregation and sparsity via `1 penalized
least squares. In Proceedings of 19th Annual Conference on Learning Theory (COLT 2006)
(G. Lugosi and H. U. Simon, eds.), pages 379–391, 2006.
[22] F. Bunea, A. B. Tsybakov, and M. H. Wegkamp. Aggregation for Gaussian regression. The
Annals of Statistics, 35(4):1674–1697, 2007.
[23] E. Candès and T. Tao. The Dantzig selector: statistical estimation when p is much larger
than n. Ann. Statist., 35(6):2313–2351, 2007.
[24] Been-Lon Chen. Economic growth : Robert J. Barro and Xavier Sala-i-Martin, (McGraw-Hill,
1995), 539 pp. Journal of Economic Dynamics and Control, 21(4-5):895–898, May 1997.
[25] R. Chen, W. Härdle, O. B. Linton, and E. Severance-Lossin. Nonparametric estimation of additive separable regression models. In Wolfgang Härdle and Michael G. Schimek, editors, Statistical Theory and Computational Aspects of Smoothing, pages 247–265, Heidelberg, 1996.
Physica-Verlag HD.
ADDITIVELY SEPARABLE
47
[26] Norbert Christopeit and Stefan G. N. Hoderlein. Local partitioned regression. Econometrica,
74(3):787–817, 2006.
[27] Dennis D. Cox. Approximation of least squares regression on nested subspaces. Ann. Statist.,
16(2):713–732, 06 1988.
[28] Brian J. Eastwood and A. Ronald Gallant. Adaptive rules for seminonparametric estimators
that achieve asymptotic normality. Econometric Theory, 7(3):307–340, 1991.
[29] Ildiko E. Frank and Jerome H. Friedman. A statistical view of some chemometrics regression
tools. Technometrics, 35(2):109–135, 1993.
[30] C. Hansen, D. Kozbur, and S. Misra. Targeted Undersmoothing. ArXiv e-prints, June 2017.
[31] Trevor Hastie and Robert Tibshirani. [generalized additive models]: Rejoinder. Statist. Sci.,
1(3):314–318, 08 1986.
[32] Trevor Hastie, Robert Tibshirani, and Jerome Friedman. Elements of Statistical Learning:
Data Mining, Inference, and Prediction. Springer, New York, NY, 2009.
[33] Jian Huang, Joel L. Horowitz, and Fengrong Wei. Variable selection in nonparametric additive
models. Ann. Statist., 38(4):2282–2313, 2010.
[34] Jian Huang, Joel L. Horowitz, and Fengrong Wei. Variable selection in nonparametric additive
models. Ann. Statist., 38(4):2282–2313, 08 2010.
[35] Guido Imbens and Keisuke Hirano. The propensity score with continuous treatments. 2004.
[36] Adel Javanmard and Andrea Montanari. Confidence intervals and hypothesis testing for highdimensional regression. Journal of Machine Learning Research, 15:2869–2909, 2014.
[37] Bing-Yi Jing, Qi-Man Shao, and Qiying Wang. Self-normalized cramr-type large deviations
for independent random variables. Ann. Probab., 31(4):2167–2215, 2003.
[38] Keith Knight. Shrinkage estimation for nearly singular designs. Econometric Theory, 24:323–
337, 2008.
[39] V. Koltchinskii. Sparsity in penalized empirical risk minimization. Ann. Inst. H. Poincar
Probab. Statist., 45(1):7–57, 2009.
[40] Hannes Leeb and Benedikt M. Pötscher. Can one estimate the unconditional distribution of
post-model-selection estimators? Econometric Theory, 24(2):338–376, 2008.
[41] Qi Li and Jeffrey Scott Racine. Nonparametric Econometrics: Theory and Practice. Princeton
University Press: Princeton, NJ, 2006.
[42] K. Lounici. Sup-norm convergence rate and sign concentration property of lasso and dantzig
estimators. Electron. J. Statist., 2:90–102, 2008.
[43] K. Lounici, M. Pontil, A. B. Tsybakov, and S. van de Geer. Taking advantage of sparsity in
multi-task learning. arXiv:0903.1468v1 [stat.ML], 2010.
[44] N. Meinshausen and B. Yu. Lasso-type recovery of sparse representations for high-dimensional
data. Annals of Statistics, 37(1):2246–2270, 2009.
[45] Whitney K. Newey. Convergence rates and asymptotic normality for series estimators. Journal
of Econometrics, 79:147–168, 1997.
[46] Benedikt M. Pötscher. Confidence sets based on sparse estimators are necessarily large.
Sankhyā, 71(1, Ser. A):1–18, 2009.
[47] Mathieu Rosenbaum and Alexandre B. Tsybakov. Sparse recovery under matrix uncertainty.
The Annals of Statistics, 38(5):2620–2651, 2010.
[48] M. Rudelson and S. Zhou. Reconstruction from anisotropic random measurements. IEEE
Transactions on Information Theory, 59(6):3434–3447, June 2013.
[49] Mark Rudelson and Roman Vershynin. On sparse reconstruction from fourier and gaussian
measurements. Communications on Pure and Applied Mathematics, 61:10251045, 2008.
[50] Eric Severance-Lossin and Stefan Sperlich. Estimation of derivatives for additive separable
models. Statistics, 33(3):241–265, 1999.
[51] Charles J. Stone. Additive regression and other nonparametric models. The Annals of Statistics, 13(2):689–705, 1985.
[52] R. Tibshirani. Regression shrinkage and selection via the lasso. J. Roy. Statist. Soc. Ser. B,
58:267–288, 1996.
[53] S. A. van de Geer. High-dimensional generalized linear models and the lasso. Annals of
Statistics, 36(2):614–645, 2008.
[54] Sara van de Geer, Peter Bhlmann, Yaacov Ritov, and Ruben Dezeure. On asymptotically
optimal confidence regions and tests for high-dimensional models. Ann. Statist., 42(3):1166–
1202, 06 2014.
48
DAMIAN KOZBUR
[55] M. Wainwright. Sharp thresholds for noisy and high-dimensional recovery of sparsity using
`1 -constrained quadratic programming (lasso). IEEE Transactions on Information Theory,
55:2183–2202, May 2009.
[56] Lijian Yang, Stefan Sperlich, and Wolfgang Hrdle. Derivative estimation and testing in generalized additive models. Journal of Statistical Planning and Inference, 115(2):521 – 542,
2003.
[57] C.-H. Zhang and J. Huang. The sparsity and bias of the lasso selection in high-dimensional
linear regression. Ann. Statist., 36(4):1567–1594, 2008.
[58] Cun-Hui Zhang and Stephanie S. Zhang. Confidence intervals for low dimensional parameters in high dimensional linear models. Journal of the Royal Statistical Society: Series B
(Statistical Methodology), 76(1):217–242, 2014.
[59] S. Zhou. Restricted eigenvalue conditions on subgaussian matrices. ArXiv:0904.4723v2, 2009.
| 10math.ST
|
Asymptotics For High Dimensional Regression
M -Estimates: Fixed Design Results
Lihua Lei∗, Peter J. Bickel†, and Noureddine El Karoui‡
arXiv:1612.06358v1 [math.ST] 19 Dec 2016
Department of Statistics, University of California, Berkeley
December 20, 2016
Abstract
We investigate the asymptotic distributions of coordinates of regression M-estimates
in the moderate p/n regime, where the number of covariates p grows proportionally with the sample size n. Under appropriate regularity conditions, we establish the coordinate-wise asymptotic normality of regression M-estimates assuming
a fixed-design matrix. Our proof is based on the second-order Poincaré inequality
(Chatterjee, 2009) and leave-one-out analysis (El Karoui et al., 2011). Some relevant
examples are indicated to show that our regularity conditions are satisfied by a broad
class of design matrices. We also show a counterexample, namely the ANOVA-type
design, to emphasize that the technical assumptions are not just artifacts of the
proof. Finally, the numerical experiments confirm and complement our theoretical
results.
1
Introduction
High-dimensional statistics has a long history (Huber, 1973; Wachter, 1976, 1978) with
considerable renewed interest over the last two decades. In many applications, the researcher collects data which can be represented as a matrix, called a design matrix and
denoted by X ∈ Rn×p , as well as a response vector y ∈ Rn and aims to study the connection between X and y. The linear model is among the most popular models as a starting
point of data analysis in various fields. A linear model assumes that
y = Xβ ∗ + ,
(1)
where β ∗ ∈ Rp is the coefficient vector which measures the marginal contribution of each
predictor and is a random vector which captures the unobserved errors.
The aim of this article is to provide valid inferential results for features of β ∗ . For
example, a researcher might be interested in testing whether a given predictor has a
negligible effect on the response, or equivalently whether βj∗ = 0 for some j. Similarly,
linear contrasts of β ∗ such as β1∗ − β2∗ might be of interest in the case of the group
comparison problem in which the first two predictors represent the same feature but are
collected from two different groups.
An M-estimator, defined as
n
β̂(ρ) = arg min
β∈Rp
1X
ρ(yi − xTi β)
n i=1
∗ contact:
(2)
[email protected]. Support from Grant FRG DMS-1160319 is gratefully acknowledged.
from Grant FRG DMS-1160319 is gratefully acknowledged.
‡ Support from Grant NSF DMS-1510172 is gratefully acknowledged.
AMS 2010 MSC: Primary: 62J99, Secondary: 62E20;
Keywords: M-estimation, robust regression, high-dimensional statistics, second order Poincaré inequality, leave-one-out analysis.
† Support
1
where ρ denotes a loss function, is among the most popular estimators used in practice
(Relles, 1967; Huber, 1973). In particular, if ρ(x) = 21 x2 , β̂(ρ) is the famous Least Square
Estimator (LSE). We intend to explore the distribution of β̂(ρ), based on which we can
achieve the inferential goals mentioned above.
The most well-studied approach is the asymptotic analysis, which assumes that the
scale of the problem grows to infinity and use the limiting result as an approximation.
In regression problems, the scale parameter of a problem is the sample size n and the
number of predictors p. The classical approach is to fix p and let n grow to infinity. It
has been shown (Relles, 1967; Yohai, 1972; Huber, 1972, 1973) that β̂(ρ) is consistent in
terms of L2 norm and asymptotically normal in this regime. The asymptotic variance
can be then approximated by the bootstrap (Bickel & Freedman, 1981). Later on, the
studies are extended to the regime in which both n and p grow to infinity but p/n
converges to 0 (Yohai & Maronna, 1979; Portnoy, 1984, 1985, 1986, 1987; Mammen,
1989). The consistency, in terms of the L2 norm, the asymptotic normality and the
validity of the bootstrap still hold in this regime. Based on
q these results, we can construct
d β̂j (ρ)) where Var(
d β̂j (ρ)) is
a 95% confidence interval for β0j simply as β̂j (ρ) ± 1.96 Var(
calculated by bootstrap. Similarly we can calculate p-values for the hypothesis testing
procedure.
We ask whether the inferential results developed under the low-dimensional assumptions and the software built on top of them can be relied on for moderate and highdimensional analysis? Concretely, if in a study n = 50 and p = 40, can the software built
upon the assumption that p/n ' 0 be relied on when p/n = .8? Results in random matrix
theory (Marčenko & Pastur, 1967) already offer an answer in the negative side for many
PCA-related questions in multivariate statistics. The case of regression is more subtle:
For instance for least-squares, standard degrees of freedom adjustments effectively take
care of many dimensionality-related problems. But this nice property does not extend to
more general regression M-estimates.
Once these questions are raised, it becomes very natural to analyze the behavior
and performance of statistical methods in the regime where p/n is fixed. Indeed, it will
help us to keep track of the inherent statistical difficulty of the problem when assessing
the variability of our estimates. In other words, we assume in the current paper that
p/n → κ > 0 while let n grows to infinity. Due to identifiability issues, it is impossible to
make inference on β ∗ if p > n without further structural or distributional assumptions.
We discuss this point in details in Section 2.3. Thus we consider the regime where
p/n → κ ∈ (0, 1). We call it the moderate p/n regime. This regime is also the natural
regime in random matrix theory (Marčenko & Pastur, 1967; Wachter, 1978; Johnstone,
2001; Bai & Silverstein, 2010). It has been shown that the asymptotic results derived
in this regime sometimes provide an extremely accurate approximations to finite sample
distributions of estimators at least in certain cases (Johnstone, 2001) where n and p are
both small.
1.1
Qualitatively Different Behavior of Moderate p/n Regime
First, β̂(ρ) is no longer consistent in terms of L2 norm and the risk Ekβ̂(ρ)−β ∗ k2 tends to
a non-vanishing quantity determined by κ, the loss function ρ and the error distribution
through a complicated system of non-linear equations (El Karoui et al., 2011; El Karoui,
2013, 2015; Bean et al., 2012). This L2 -inconsistency prohibits the use of standard
perturbation-analytic techniques to assess the behavior of the estimator. It also leads to
qualitatively different behaviors for the residuals in moderate dimensions; in contrast to
the low-dimensional case, they cannot be relied on to give accurate information about the
distribution of the errors. However, this seemingly negative result does not exclude the
possibility of inference since β̂(ρ) is still consistent in terms of L2+ν norms for any ν > 0
and in particular in L∞ norm. Thus, we can at least hope to perform inference on each
coordinate.
Second, classical optimality results do not hold in this regime. In the regime p/n → 0,
the maximum likelihood estimator is shown to be optimal (Huber, 1964, 1972; Bickel &
2
Doksum, 2015). In other words, if the error distribution is known then the M-estimator
associated with the loss ρ(·) = − log f (·) is asymptotically efficient, provided the design
is of appropriate type, where f (·) is the density of entries of . However, in the moderate
p/n regime, it has been shown that the optimal loss is no longer the log-likehood but
an other function with a complicated but explicit form (Bean et al., 2013), at least
for certain designs. The suboptimality of maximum likelihood estimators suggests that
classical techniques fail to provide valid intuition in the moderate p/n regime.
Third, the joint asymptotic normality of β̂(ρ), as a p-dimensional random vector, may
be violated for a fixed design matrix X. This has been proved for least-squares by Huber
(1973) in his pioneering work. For general M-estimators, this negative result is a simple
consequence of the results of El Karoui et al. (2011): They exhibit an ANOVA design
(see below) where even marginal fluctuations are not Gaussian. By contrast, for random
design, they show that β̂(ρ) is jointly asymptotically normal when the design matrix is
elliptical with general covariance by using the non-asymptotic stochastic representation
for β̂(ρ) as well as elementary properties of vectors uniformly distributed on the uniform
sphere in Rp ; See section 2.2.3 of El Karoui et al. (2011) or the supplementary material
of Bean et al. (2013) for details. This does not contradict Huber (1973)’s negative result
in that it takes the randomness from both X and into account while Huber (1973)’s
result only takes the randomness from into account. Later, El Karoui (2015) shows that
each coordinate of β̂(ρ) is asymptotically normal for a broader class of random designs.
This is also an elementary consequence of the analysis in El Karoui (2013). However, to
the best of our knowledge, beyond the ANOVA situation mentioned above, there are no
distributional results for fixed design matrices. This is the topic of this article.
Last but not least, bootstrap inference fails in this moderate-dimensional regime. This
has been shown by Bickel and Freedman (1983) for least-squares and residual bootstrap
in their influential work. Recently, El Karoui and Purdom (2015) studied the results
to general M-estimators and showed that all commonly used bootstrapping schemes,
including pairs-bootstrap, residual bootstrap and jackknife, fail to provide a consistent
variance estimator and hence valid inferential statements. These latter results even apply
to the marginal distributions of the coordinates of β̂(ρ). Moreover, there is no simple,
design independent, modification to achieve consistency (El Karoui & Purdom, 2015).
1.2
Our Contributions
In summary, the behavior of the estimators we consider in this paper is completely different in the moderate p/n regime from its counterpart in the low-dimensional regime.
As discussed in the next section, moving one step further in the moderate p/n regime is
interesting from both the practical and theoretical perspectives. The main contribution
of this article is to establish coordinate-wise asymptotic normality of β̂(ρ) for certain fixed
design matrices X in this regime under technical assumptions. The following theorem
informally states our main result.
Theorem (Informal Version of Theorem 3.1 in Section 3). Under appropriate conditions
on the design matrix X, the distribution of and the loss function ρ, as p/n → κ ∈ (0, 1),
while n → ∞,
β̂j (ρ) − Eβ̂j (ρ)
, N (0, 1) = o(1)
max dTV L q
Var(β̂j (ρ))
1≤j≤p
where dTV (·, ·) is the total variation distance and L(·) denotes the law.
It is worth mentioning that the above result can be extended to finite dimensional
linear contrasts of β̂. For instance, one might be interested in making inference on β1∗ −β2∗
in the problems involving the group comparison. The above result can be extended to
give the asymptotic normality of β̂1 − β̂2 .
Besides the main result, we have several other contributions. First, we use a new
approach to establish asymptotic normality. Our main technique is based on the secondorder Poincaré inequality (SOPI), developed by Chatterjee (2009) to derive, among many
3
other results, the fluctuation behavior of linear spectral statistics of random matrices.
In contrast to classical approaches such as the Lindeberg-Feller central limit theorem,
the second-order Poincaré inequality is capable of dealing with nonlinear and potentially
implicit functions of independent random variables. Moreover, we use different expansions
for β̂(ρ) and residuals based on double leave-one-out ideas introduced in El Karoui et al.
(2011), in contrast to the classical perturbation-analytic expansions. See aforementioned
paper and follow-ups. An informal interpretation of the results of Chatterjee (2009) is
that if the Hessian of the nonlinear function of random variables under consideration is
sufficiently small, this function acts almost linearly and hence a standard central limit
theorem holds.
Second, to the best of our knowledge this is the first inferential result for fixed (non
ANOVA-like) design in the moderate p/n regime. Fixed designs arise naturally from an
experimental design or a conditional inference perspective. That is, inference is ideally
carried out without assuming randomness in predictors; see Section 2.2 for more details.
We clarify the regularity conditions for coordinate-wise asymptotic normality of β̂(ρ)
explicitly, which are checkable for LSE and also checkable for general M-estimators if the
error distribution is known. We also prove that these conditions are satisfied with by a
broad class of designs.
The ANOVA-like design described in Section 3.3.4 exhibits a situation where the
distribution of β̂j (ρ) is not going to be asymptotically normal. As such the results of
Theorem 3.1 below are somewhat surprising.
For complete inference, we need both the asymptotic normality and the asymptotic
bias and variance. Under suitable symmetry conditions on the loss function and the error
distribution, it can be shown that β̂(ρ) is unbiased (see Section 3.2.1 for details) and
thus it is left to derive the asymptotic variance. As discussed at the end of Section 1.1,
classical approaches, e.g. bootstrap, fail in this regime. For least-squares, classical results
continue to hold and we discuss it in section 5 for the sake of completeness. However,
for M-estimators, there is no closed-form result. We briefly touch upon the variance
estimation in Section 3.4.2. The derivation for general situations is beyond the scope of
this paper and left to the future research.
1.3
Outline of Paper
The rest of the paper is organized as follows: In Section 2, we clarify details which
are mentioned in the current section. In Section 3, we state the main result (Theorem
3.1) formally and explain the technical assumptions. Then we show several examples of
random designs which satisfy the assumptions with high probability. In Section 4, we
introduce our main technical tool, second-order Poincaré inequality (Chatterjee, 2009),
and apply it on M-estimators as the first step to prove Theorem 3.1. Since the rest of the
proof of Theorem 3.1 is complicated and lengthy, we illustrate the main ideas in Appendix
A. The rigorous proof is left to Appendix B. In Section 5, we provide reminders about the
theory of least-squares estimation for the sake of completeness, by taking advantage of its
explicit form. In Section 6, we display the numerical results. The proof of other results
are stated in Appendix C and more numerical experiments are presented in Appendix D.
2
2.1
More Details on Background
Moderate p/n Regime: a more informative type of asymptotics?
In Section 1, we mentioned that the ratio p/n measures the difficulty of statistical inference. The moderate p/n regime provides an approximation of finite sample properties
with the difficulties fixed at the same level as the original problem. Intuitively, this regime
should capture more variation in finite sample problems and provide a more accurate approximation. We will illustrate this via simulation.
4
Consider a study involving 50 participants and 40 variables; we can either use the
asymptotics in which p is fixed to be 40, n grows to infinity or p/n is fixed to be 0.8,
and n grows to infinity to perform approximate inference. Current software rely on lowdimensional asymptotics for inferential tasks, but there is no evidence that they yield more
accurate inferential statements than the ones we would have obtained using moderate
dimensional asymptotics. In fact, numerical evidence (Johnstone, 2001; El Karoui et al.,
2013; Bean et al., 2013) show that the reverse is true.
We exhibit a further numerical simulation showing that. Consider a case that n = 50,
has i.i.d. entries and X is one realization of a matrix generated with i.i.d. gaussian
(mean 0, variance 1) entries. For κ ∈ {0.1, 0.2, . . . , 0.9} and different error distributions,
we use the Kolmogorov-Smirnov (KS) statistics to quantify the distance between the
finite sample distribution and two types of asymptotic approximation of the distribution
of β̂1 (ρ).
Specifically, we use the Huber loss function ρHuber,k with default parameter k = 1.345
(Huber, 2011), i.e.
1 2
|x| ≤ k
2x
ρHuber,k (x) =
k(|x| − 12 k) |x| > k
Specifically, we generate three design matrices X (0) , X (1) and X (2) : X (0) for small sample
case with a sample size n = 50 and a dimension p = nκ; X (1) for low-dimensional
asymptotics (p fixed) with a sample size n = 1000 and a dimension p = 50κ; and X (2)
for moderate-dimensional asymptotics (p/n fixed) with a sample size n = 1000 and a
dimension p = nκ. Each of them is generated as one realization of an i.i.d. standard
gaussian design and then treated as fixed across K = 100 repetitions. For each design
matrix, vectors of appropriate length are generated with i.i.d. entries. The entry
has either a standard normal distribution, or a t3 -distribution, or a standard Cauchy
distribution, i.e. t1 . Then we use as the response, or equivalently assume β ∗ = 0,
and obtain the M-estimators β̂ (0) , β̂ (1) , β̂ (2) . Repeating this procedure for K = 100 times
results in K replications in three cases. Then we extract the first coordinate of each
(0)
(1) K
(2) K
estimator, denoted by {β̂k,1 }K
k=1 , {β̂k,1 }k=1 , {β̂k,1 }k=1 . Then the two-sample KolmogorovSmirnov statistics can be obtained by
r
r
n
n
(0)
(1)
KS1 =
max |F̂n (x) − F̂n (x)|, KS2 =
max |F̂n(0) (x) − F̂n(2) (x)|,
2 x
2 x
(r)
(r)
where F̂n is the empirical distribution of {β̂k,1 }K
k=1 . We can then compare the accuracy
of two asymptotic regimes by comparing KS1 and KS2 . The smaller the value of KSi , the
better the approximation.
Figure 1 displays the results for these error distributions. We see that for gaussian errors and even t3 errors, the p/n-fixed/moderate-dimensional approximation is uniformly more accurate than the widely used p-fixed/low-dimensional approximation. For
Cauchy errors, the low-dimensional approximation performs better than the moderatedimensional one when p/n is small but worsens when the ratio is large especially when
p/n is close to 1. Moreover, when p/n grows, the two approximations have qualitatively different behaviors: the p-fixed approximation becomes less and less accurate while
the p/n-fixed approximation does not suffer much deterioration when p/n grows. The
qualitative and quantitative differences of these two approximations reveal the practical
importance of exploring the p/n-fixed asymptotic regime. (See also Johnstone (2001).)
2.2
Random vs fixed design?
As discussed in Section 1.1, assuming a fixed design or a random design could lead to
qualitatively different inferential results.
In the random design setting, X is considered as being generated from a super population. For example, the rows of X can be regarded as an i.i.d. sample from a distribution known, or partially known, to the researcher. In situations where one uses
techniques such as cross-validation (Stone, 1974), pairs bootstrap in regression (Efron &
5
Kolmogorov−Smirnov Statistics
Distance between the small sample and large sample distribution
normal
t(3)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.45
0.40
cauchy
●
0.50
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
0.35
●
●
●
●
●
●
●
●
●
●
0.30
●
●
●
0.25
0.25
0.50
0.75
0.25
0.50
0.75
0.25
0.50
0.75
kappa
Asym. Regime
●
p fixed
●
p/n fixed
Figure 1: Axpproximation accuracy of p-fixed asymptotics and p/n-fixed asymptotics:
each column represents an error distribution; the x-axis represents the ratio κ of the dimension and the sample size and the y-axis represents the Kolmogorov-Smirnov statistic;
the red solid line corresponds to p-fixed approximation and the blue dashed line corresponds to p/n-fixed approximation.
Efron, 1982) or sample splitting (Wasserman & Roeder, 2009), the researcher effectively
assumes exchangeability of the data (xTi , yi )ni=1 . Naturally, this is only compatible with
an assumption of random design. Given the extremely widespread use of these techniques
in contemporary machine learning and statistics, one could argue that the random design setting is the one under which most of modern statistics is carried out, especially
for prediction problems. Furthermore, working under a random design assumption forces
the researcher to take into account two sources of randomness as opposed to only one
in the fixed design case. Hence working under a random design assumption should yield
conservative confidence intervals for βj∗ .
In other words, in settings where the researcher collects data without control over the
values of the predictors, the random design assumption is arguably the more natural one
of the two.
However, it has now been understood for almost a decade that common random design
assumptions in high-dimension (e.g. xi = Σ1/2 zi where zi,j ’s are i.i.d with mean 0 and
variance 1 and a few moments and Σ “well behaved”) suffer from considerable geometric
limitations, which have substantial impacts on the performance of the estimators considered in this paper (El Karoui et al., 2011). As such, confidence statements derived from
that kind of analysis can be relied on only after performing a few graphical tests on the
data (see El Karoui (2010)). These geometric limitations are simple consequences of the
concentration of measure phenomenon (Ledoux, 2001).
On the other hand, in the fixed design setting, X is considered a fixed matrix. In this
case, the inference only takes the randomness of into consideration. This perspective
is popular in several situations. The first one is the experimental design. The goal is
to study the effect of a set of factors, which can be controlled by the experimenter, on
the response. In contrast to the observational study, the experimenter can design the
experimental condition ahead of time based on the inference target. For instance, a oneway ANOVA design encodes the covariates into binary variables (see Section 3.3.4 for
details) and it is fixed prior to the experiment. Other examples include two-way ANOVA
designs, factorial designs, Latin-square designs, etc. (Scheffe, 1999).
Another situation which is concerned with fixed design is the survey sampling where
the inference is carried out conditioning on the data (Cochran, 1977). Generally, in order
to avoid unrealistic assumptions, making inference conditioning on the design matrix
X is necessary. Suppose the linear model (1) is true and identifiable (see Section 2.3 for
details), then all information of β ∗ is contained in the conditional distribution L(y|X) and
hence the information in the marginal distribution L(X) is redundant. The conditional
6
inference framework is more robust to the data generating procedure due to the irrelevance
of L(X).
Also, results based on fixed design assumptions may be preferable from a theoretical
point of view in the sense that they could potentially be used to establish corresponding
results for certain classes of random designs. Specifically, given a marginal distribution
L(X), one only has to prove that X satisfies the assumptions for fixed design with high
probability.
In conclusion, fixed and random design assumptions play complementary roles in
moderate-dimensional settings. We focus on the least understood of the two, the fixed
design case, in this paper.
2.3
Modeling and Identification of Parameters
The problem of identifiability is especially important in the fixed design case. Define
β ∗ (ρ) in the population as
n
β ∗ (ρ) = arg min
β∈Rp
1X
Eρ(yi − xTi β).
n i=1
(3)
One may ask whether β ∗ (ρ) = β ∗ regardless of ρ in the fixed design case. We provide
an affirmative answer in the following proposition by assuming that i has a symmetric
distribution around 0 and ρ is even.
d
Proposition 2.1. Suppose X has a full column rank and i = −i for all i. Further
assume ρ is an even convex function such that for any i = 1, 2, . . . and α 6= 0,
1
(Eρ(i − α) + Eρ(i + α)) > Eρ(i ).
(4)
2
Then β ∗ (ρ) = β ∗ regardless of the choice of ρ.
The proof is left to Appendix C. It is worth mentioning that Proposition 2.1 only
requires the marginals of to be symmetric but does not impose any constraint on the
dependence structure of . Further, if ρ is strongly convex, then for all α 6= 0,
1
(ρ(x − α) + ρ(x + α)) > ρ(x).
2
As a consequence, the condition (4) is satisfied provided that i is non-zero with positive
probability.
If is asymmetric, we may still be able to identify β ∗ if i are i.i.d. random variables.
In contrast to the last case, we should incorporate an intercept term as a shift towards
the centroid of ρ. More precisely, we define α∗ (ρ) and β ∗ (ρ) as
n
(α∗ (ρ), β ∗ (ρ)) = arg min
α∈R,β∈Rp
1X
Eρ(yi − α − xTi β).
n i=1
Proposition 2.2. Suppose (1, X) is of full column rank and i are i.i.d. such that
Eρ(1 − α) as a function of α has a unique minimizer α(ρ). Then β ∗ (ρ) is uniquely
defined with β ∗ (ρ) = β ∗ and α∗ (ρ) = α(ρ).
The proof is left to Appendix C. For example, let ρ(z) = |z|. Then the minimizer of
Eρ(1 − a) is a median of 1 , and is unique if 1 has a positive density. It is worth pointing
out that incorporating an intercept term is essential for identifying β ∗ . For instance, in
the least-square case, β ∗ (ρ) no longer equals to β ∗ if Ei 6= 0. Proposition 2.2 entails that
the intercept term guarantees β ∗ (ρ) = β ∗ , although the intercept term itself depends on
the choice of ρ unless more conditions are imposed.
If i ’s are neither symmetric nor i.i.d., then β ∗ cannot be identified by the previous
criteria because β ∗ (ρ) depends on ρ. Nonetheless, from a modeling perspective, it is
popular and reasonable to assume that i ’s are symmetric or i.i.d. in many situations.
Therefore, Proposition 2.1 and Proposition 2.2 justify the use of M-estimators in those
cases and M-estimators derived from different loss functions can be compared because
they are estimating the same parameter.
7
3
3.1
Main Results
Notation and Assumptions
Let xTi ∈ R1×p denote the i-th row of X and Xj ∈ Rn×1 denote the j-th column of X.
Throughout the paper we will denote by Xij ∈ R the (i, j)-th entry of X, by X[j] ∈
Rn×(p−1) the design matrix X after removing the j-th column, and by xTi,[j] ∈ R1×(p−1)
the vector xTi after removing j-th entry. The M-estimator β̂(ρ) associated with the loss
function ρ is defined as
β̂(ρ) = arg min
β∈Rp
n
n
k=1
k=1
1X
1X
ρ(yk − xTk β) = arg min
ρ(k − xTk (β − β ∗ ))
n
β∈Rp n
(5)
We define ψ = ρ0 to be the first derivative of ρ. We will write β̂(ρ) simply β̂ when no
confusion can arise.
When the original design matrix X does not contain an intercept term, we can simply
replace X by (1, X) and augment β into a (p + 1)-dimensional vector (α, β T )T . Although
being a special case, we will discuss the question of intercept in Section 3.2.2 due to its
important role in practice.
Equivariance and reduction to the null case
β̂ −Eβ̂j
is invariant to the choice of β ∗ , provided that β ∗
Notice that our target quantity √j
Var(β̂j )
is identifiable as discussed in Section 2.3, we can assume β ∗ = 0 without loss of generality.
In this case, we assume in particular that the design matrix X has full column rank. Then
yk = k and
n
1X
ρ(k − xTk β).
β̂ = arg min
β∈Rp n
k=1
Similarly we define the leave-j-th-predictor-out version as
n
β̂[j] = arg min
β∈Rp−1
1X
ρ(k − xTk,[j] β).
n
k=1
Based on these notations we define the full residuals Rk as
Rk = k − xTk β̂,
k = 1, 2, . . . , n
and the leave-j-th-predictor-out residual as
rk,[j] = k − xTk,[j] β̂[j] ,
k = 1, 2, . . . , n,
j = 1, . . . , p.
Three n × n diagonal matrices are defined as
D = diag(ψ 0 (Rk ))nk=1 ,
D̃ = diag(ψ 00 (Rk ))nk=1 ,
D[j] = diag(ψ 0 (rk,[j] ))nk=1 .
(6)
We say a random variable Z is σ 2 -sub-gaussian if for any λ ∈ R,
EeλZ ≤ e
λ2 σ 2
2
.
In addition, we use Jn ⊂ {1, . . . , p} to represent the indices of parameters which are
of interest. Intuitively, more entries in Jn would require more stringent conditions for the
asymptotic normality.
Finally, we adopt Landau’s notation (O(·), o(·), Op (·), op (·)). In addition, we say an =
Ω(bn ) if bn = O(an ) and similarly, we say an = Ωp (bn ) if bn = Op (an ). To simplify the
logarithm factors, we use the symbol polyLog(n) to denote any factor that can be upper
1
bounded by (log n)γ for some γ > 0. Similarly, we use polyLog(n)
to denote any factor
1
0
that can be lower bounded by (log n)γ 0 for some γ > 0.
8
3.2
Technical Assumptions and main result
Before stating the assumptions, we need to define several quantities of interest. Let
T
T
X X
X X
λ+ = λmax
, λ− = λmin
n
n
be the largest (resp. smallest) eigenvalue of the matrix
canonical basis vector and
XT X
n .
Let ei ∈ Rn be the i-th
T
T
D[j] X[j] )−1 X[j]
)ei .
hj,1,i , (I − D[j] X[j] (X[j]
hj,0 , (ψ(r1,[j] ), . . . , ψ(rn,[j] ))T ,
Finally, let
(
|hTj,1,i Xj |
|hTj,0 Xj |
∆C = max max
, max
j∈Jn ||hj,0 ||2 i≤n,j∈Jn ||hj,1,i ||2
)
,
Qj = Cov(hj,0 )
Based on the quantities defined above, we state our technical assumptions on the design
matrix X followed by the main result. A detailed explanation of the assumptions follows.
1
, K1 , K2 =
A1 ρ(0) = ψ(0) = 0 and there exists positive numbers K0 = Ω polyLog(n)
O (polyLog(n)), such that for any x ∈ R,
K0 ≤ ψ 0 (x) ≤ K1 ,
|ψ 00 (x)|
d p 0
( ψ (x)) = p
≤ K2 ;
dx
ψ 0 (x)
A2 i = ui (Wi ) where (W1 , . . . , Wn ) ∼ N (0, In×n ) and ui are smooth functions with
ku0i k∞ ≤ c1 and ku00i k∞ ≤
c2 for some c1 , c2 = O(polyLog(n)). Moreover, assume
mini Var(i ) = Ω
1
polyLog(n)
.
A3 λ+ = O(polyLog(n)) and λ− = Ω
A4 minj∈Jn
XjT Qj Xj
tr(Qj )
=Ω
1
polyLog(n)
1
polyLog(n)
;
;
A5 E∆8C = O (polyLog(n)).
Theorem 3.1. Under assumptions A1 − A5, as p/n → κ for some κ ∈ (0, 1), while
n → ∞,
β̂
−
E
β̂
j
j
, N (0, 1) = o(1),
max dTV L q
j∈Jn
Var(β̂j )
where dTV (P, Q) = supA |P (A) − Q(A)| is the total variation distance.
We provide several examples where our assumptions hold in Section 3.3. We also
provide an example where the asymptotic normality does not hold in Section 3.3.4. This
shows that our assumptions are not just artifacts of the proof technique we developed,
but that there are (probably many) situations where asymptotic normality will not hold,
even coordinate-wise.
3.2.1
Discussion of Assumptions
Now we discuss assumptions A1 - A5. Assumption A1 implies the boundedness of the
first-order and the second-order derivatives of ψ. The upper bounds are satisfied by most
loss functions including the L2 loss, the smoothed L1 loss, the smoothed Huber loss,
etc. The non-zero lower bound K0 implies the strong convexity of ρ and is required for
technical reasons. It can be removed by considering first a ridge-penalized M-estimator
9
and taking appropriate limits as in El Karoui (2013, 2015). In addition, in this paper we
consider the smooth loss functions and the results can be extended to non-smooth case
via approximation.
Assumption A2 was proposed in Chatterjee (2009) when deriving the second-order
Poincaré inequality discussed in Section 4.1. It means that the results apply to nonGaussian distributions, such as the uniform distribution on [0, 1] by taking ui = Φ, the
cumulative distribution function of standard normal distribution. Through the gaussian
concentration (Ledoux, 2001), we see that A2 implies that i are c21 -sub-gaussian. Thus
A2 controls the tail behavior of i . The boundedness of u0i and u00i are required only for the
direct application of Chatterjee’s results. In fact, a look at his proof suggests that one can
obtain a similar result to his Second-Order Poincaré inequality involving moment bounds
on u0i (Wi ) and u00i (Wi ). This would be a way to weaken our assumptions to permit to have
the heavy-tailed distributions expected in robustness studies. Since we are considering
strongly convex loss-functions, it is not completely unnatural to restrict our attention to
light-tailed errors. Furthermore, efficiency - and not only robustness - questions are one
of the main reasons to consider these estimators in the moderate-dimensional context.
The potential gains in efficiency obtained by considering regression M-estimates (Bean
et al., 2013) apply in the light-tailed context, which further justify our interest in this
theoretical setup.
Assumption A3 is completely checkable since it only depends on X. It controls the
singularity of the design matrix. Under A1 and A3, it can be shown that the objective
function is strongly convex
withcurvature (the smallest eigenvalue of the Hessian matrix)
1
everywhere.
lower bounded by Ω polyLog(n)
Assumption A4 is controlling the left tail of quadratic forms. It is fundamentally
connected to aspects of the concentration of measure phenomenon (Ledoux, 2001). This
condition is proposed and emphasized under the random design setting by El Karoui et
al. (2013). Essentially, it means that for a matrix Qj ,which does not depend on Xj , the
quadratic form XjT Qj Xj should have the same order as tr(Qj ).
Assumption A5 is proposed by El Karoui (2013) under the random design settings. It
is motivated by leave-one-predictor-out analysis. Note that ∆C is the maximum of linear
contrasts of Xj , whose coefficients do not depend on Xj . It is easily checked for design
matrix X which is a realization of a random matrix with i.i.d sub-gaussian entries for
instance.
Remark 3.2. In certain applications, it is reasonable to make the following additional
assumption:
A6 ρ is an even function and i ’s have symmetric distributions.
Although assumption A6 is not necessary to Theorem 3.1, it can simplify the result. Under
d
assumption A6, when X is full rank, we have, if = denotes equality in distribution,
n
β̂ − β ∗ = arg min
η∈Rp
n
1X
1X
ρ(i − xTi η) = arg min
ρ(−i + xTi η)
n i=1
η∈Rp n i=1
n
d
= arg min
η∈Rp
1X
ρ(i + xTi η) = β ∗ − β̂.
n i=1
This implies that β̂ is an unbiased estimator, provided it has a mean, which is the case
here. Unbiasedness is useful in practice, since then Theorem 3.1 reads
β̂j − βj∗
, N (0, 1) = o(1) .
max dTV L q
j∈Jn
Var(β̂j )
For inference, we only need to estimate the asymptotic variance.
10
3.2.2
An important remark concerning Theorem 3.1
When Jn is a subset of {1, . . . , p}, the coefficients in Jnc become nuisance parameters.
Heuristically, in order for identifying βJ∗n , one only needs the subspaces span(XJn ) and
span(XJnc ) to be distinguished and XJn has a full column rank. Here XJn denotes the
sub-matrix of X with columns in Jn . Formally, let
Σ̂Jn =
1 T
X (I − XJnc (XJTnc XJnc )− XJTnc )XJn
n Jn
where A− denotes the generalized inverse of A, and
λ̃+ = λmax Σ̂Jn , λ̃− = λmin Σ̂Jn .
Then Σ̂Jn characterizes the behavior of XJn after removing the effect of XJnc . In particular, we can modify the assumption A3 by
1
.
A3* λ̃+ = O(polyLog(n)) and λ̃− = Ω polyLog(n)
Then we are able to derive a stronger result in the case where |Jn | < p than Theorem 3.1
as follows.
Corollary 3.3. Under assumptions A1-2, A4-5 and A3*, as p/n → κ for some κ ∈
(0, 1),
β̂
−
E
β̂
j
j
, N (0, 1) = o(1).
max dTV L q
j∈Jn
Var(β̂j )
It can be shown that λ̃+ ≤ λ+ and λ̃− ≥ λ− and hence the assumption A3* is weaker
than A3. It is worth pointing out that the assumption A3* even holds when XJcn does not
have full column rank, in which case βJ∗n is still identifiable and β̂Jn is still well-defined,
although βJ∗nc and β̂Jnc are not; see Appendix C-2 for details.
3.3
Examples
Throughout this subsection (except subsubsection 3.3.4), we consider the case where
X is a realization of a random matrix, denoted by Z (to be distinguished from X).
We will verify that the assumptions A3-A5 are satisfied with high probability under
different regularity conditions on the distribution of Z. This is a standard way to justify
the conditions for fixed design (Portnoy, 1984, 1985) in the literature on regression Mestimates.
3.3.1
Random Design with Independent Entries
First we consider a random matrix Z with i.i.d. sub-gaussian entries.
Proposition 3.4. Suppose Z has i.i.d. mean-zero σ 2-sub-gaussian
entries with Var(Zij ) =
1
2
τ > 0 for some σ = O(polyLog(n)) and τ = Ω polyLog(n) , then, when X is a realization of Z, assumptions A3-A5 for X are satisfied with high probability over Z for
Jn = {1, . . . , p}.
In practice, the assumption of identical distribution might be invalid. In fact the
assumptions A4, A5 and the first part of A3 (λ+ = O (polyLog(n))) are still satisfied with
high probability if we only assume the independence between entries and boundedness of
certain moments. To control λ− , we rely on Litvak et al. (2005) which assumes symmetry
of each entry. We obtain the following result based on it.
11
Proposition 3.5. Suppose Z has independent σ 2 -sub-gaussian entries with
d
Zij = −Zij , Var(Zij ) > τ 2
1
for some σ = O (polyLog(n)) and τ = Ω polyLog(n)
, then, when X is a realization of Z,
assumptions A3-A5 for X are satisfied with high probability over Z for Jn = {1, . . . , p}.
Under the conditions of Proposition 3.5, we can add an intercept term into the design
matrix. Adding an intercept allows us to remove the mean-zero assumption for Zij ’s. In
fact, suppose Zij is symmetric with respect to µj , which is potentially non-zero, for all i,
then according to section 3.2.2, we can replace Zij by Zij − µj and Proposition 3.6 can
be then applied.
Proposition 3.6. Suppose Z = (1, Z̃) and Z̃ ∈ Rn×(p−1) has independent σ 2 -subgaussian entries with
d
Z̃ij − µj = µj − Z̃ij , Var(Z̃ij ) > τ 2
1
and arbitrary µj . Then, when X is a
for some σ = O (polyLog(n)), τ = Ω polyLog(n)
realization of Z, assumptions A3*, A4 and A5 for X are satisfied with high probability
over Z for Jn = {2, . . . , p}.
3.3.2
Dependent Gaussian Design
To show that our assumptions handle a variety of situations, we now assume that the
observations, namely the rows of Z, are i.i.d. random vectors with a covariance matrix
i.i.d.
Σ. In particular we show that the Gaussian design, i.e. zi ∼ N (0, Σ), satisfies the
assumptions with high probability.
i.i.d.
Proposition
3.7. Suppose zi ∼ N (0, Σ) with λmax (Σ) = O (polyLog(n)) and λmin (Σ) =
1
Ω polyLog(n) , then, when X is a realization of Z, assumptions A3-A5 for X are satisfied
with high probability over Z for Jn = {1, . . . , p}.
This result extends to the matrix-normal design (Muirhead, 1982)[Chapter 3], i.e.
(Zij )i≤n,j≤p is one realization of a np-dimensional random variable Z with multivariate
gaussian distribution
vec(Z) , (z1T , z2T , . . . , znT ) ∼ N (0, Λ ⊗ Σ),
and ⊗ is the Kronecker product. It turns out that assumptions A3 − A5 are satisfied if
both Λ and Σ are well-behaved.
Proposition 3.8. Suppose Z is matrix-normal with vec(Z) ∼ N (0, Λ ⊗ Σ) and
1
λmax (Λ), λmax (Σ) = O (polyLog(n)) , λmin (Λ), λmin (Σ) = Ω
.
polyLog(n)
Then, when X is a realization of Z,assumptions A3-A5 for X are satisfied with high
probability over Z for Jn = {1, . . . , p}.
In order to incorporate an intercept term, we need slightly more stringent condition
on Λ. Instead of assumption A3, we prove that assumption A3* - see subsubsection 3.2.2
- holds with high probability.
Proposition 3.9. Suppose Z contains an intercept term, i.e. Z = (1, Z̃) and Z̃ satisfies
the conditions of Proposition 3.8. Further assume that
1
maxi |(Λ− 2 1)i |
1
mini |(Λ− 2 1)i |
= O (polyLog(n)) .
(7)
Then, when X is a realization of Z, assumptions A3*, A4 and A5 for X are satisfied
with high probability over Z for Jn = {2, . . . , p}.
12
When Λ = I, the condition (7) is satisfied. Another non-trivial example is the exchangeable case where Λij are all equal for i 6= j. In this case, 1 is an eigenvector of Λ and
1
1
hence it is also an eigenvector of Λ− 2 . Thus Λ− 2 1 is a multiple of 1 and the condition
(7) is satisfied.
3.3.3
Elliptical Design
Furthermore, we can move from Gaussian-like structure to generalized elliptical models
where zi = ζi Σ1/2 Zi where {ζi , Zij : i = 1, . . . , n; j = 1, . . . , p} are independent random
variables, Zij having for instance mean 0 and variance 1. The elliptical family is quite
flexible in modeling data. It represents a type of data formed by a common driven factor
and independent individual effects. It is widely used in multivariate statistics (Anderson
(1962); Tyler (1987)) and various fields, including finance (Cizek et al., 2005) and biology
(Posekany et al., 2011). In the context of high-dimensional statistics, this class of model
was used to refute universality claims in random matrix theory (El Karoui, 2009). In
robust regression, El Karoui et al. (2011) used elliptical models to show that the limit
of kβ̂k22 depends on the distribution of ζi and hence the geometry of the predictors. As
such, studies limited to Gaussian-like design were shown to be of very limited statistical
interest. See also the deep classical inadmissibility results (Baranchik, 1973; Jurečkovà &
Klebanov, 1997). However, as we will show in the next proposition, the common factors ζi
do not distort the shape of the asymptotic distribution. A similar phenomenon happens
in the random design case - see El Karoui et al. (2013); Bean et al. (2013).
Proposition 3.10. Suppose Z is generated from an elliptical model, i.e.
Zij = ζi Zij ,
where ζi are independent random variables taking values in [a, b] for some 0 < a < b < ∞
and Zij are independent random variables satisfying the conditions of Proposition 3.4 or
Proposition 3.5. Further assume that {ζi : i = 1, . . . , n} and {Zij : i = 1, . . . , n; j =
1, . . . , p} are independent. Then, when X is a realization of Z, assumptions A3-A5 for
X are satisfied with high probability over Z for Jn = {1, . . . , p}.
Thanks to the fact that ζi is bounded away from 0 and ∞, the proof of Proposition
3.10 is straightforward, as shown in Appendix C. However, by a more refined argument
and assuming identical distributions ζi , we can relax this condition.
Proposition 3.11. Under the conditions of Proposition 3.10 (except the boundedness of
ζi ) and assume ζi are i.i.d. samples generated from some distribution F , independent of
n, with
α
P (ζ1 ≥ t) ≤ c1 e−c2 t ,
for some fixed c1 , c2 , α > 0 and F −1 (q) > 0 for any q ∈ (0, 1) where F −1 is the quantile
function of F and is continuous. Then, when X is a realization of Z, assumptions A3-A5
for X are satisfied with high probability over Z for Jn = {1, . . . , p}.
3.3.4
A counterexample
Consider a one-way ANOVA situation. In other words, let the design matrix have exactly
1 non-zero entry per row, whose value is 1. Let {ki }ni=1 be integers in {1, . . . , p}. And
let Xi,j = 1(j = ki ). Furthermore, let us constrain nj = |{i : ki = j}| to be such that
1 ≤ nj ≤ 2bp/nc. Taking for instance ki = (i mod p) is an easy way to produce such a
matrix. The associated statistical model is just yi = i + βk∗i .
It is easy to see that
X
X
β̂j = arg min
ρ(yi − βj ) = arg min
ρ(i − (βj − βj∗ )) .
β∈R
β∈R
i:ki =j
i:ki =j
This is of course a standard location problem. In the moderate-dimensional setting we
consider, nj remains finite as n → ∞. So β̂j is a non-linear function of finitely many
random variables and will in general not be normally distributed.
13
For concreteness, one can take ρ(x) = |x|, in which case β̂j is a median of {yi }{i:ki =j} .
The cdf of β̂j is known exactly by elementary order statistics computations (see David
and Nagaraja (1981)) and is not that of a Gaussian random variable in general. In fact,
the ANOVA design considered here violates the assumption A3 since λ− = minj nj /n =
O (1/n). Further, we can show that the assumption A5 is also violated, at least in the
least-square case; see Section 5.1 for details.
3.4
3.4.1
Comments and discussions
Asymptotic Normality in High Dimensions
In the p-fixed regime, the asymptotic distribution is easily defined as the limit of L(β̂) in
terms of weak topology (Van der Vaart, 1998). However, in regimes where the dimension
p grows, the notion of asymptotic distribution is more delicate. a conceptual question
arises from the fact that the dimension of the estimator β̂ changes with n and thus there
is no well-defined distribution which can serve as the limit of L(β̂), where L(·) denotes
the law. One remedy is proposed by Mallows (1972). Under this framework, a triangular
2
array {Wn,j , j = 1, 2, . . . , pn }, with EWn,j = 0, EWn,j
= 1, is called jointly asymptotically
pn
normal if for any deterministic sequence an ∈ R with kan k2 = 1,
pn
X
L
an,j Wn,j → N (0, 1).
j=1
When the zero mean and unit variance are not satisfied, it is easy to modify the definition
by normalizing random variables.
Definition 3.12 (joint asymptotic normality).
{Wn : Wn ∈ Rpn } is jointly asymptotically normal if and only if for any sequence
{an : an ∈ Rpn },
!
aTn (Wn − EWn )
→ N (0, 1).
L p
aTn Cov(Wn )an
The above definition of asymptotic normality is strong and appealing but was shown
not to hold for least-squares in the moderate p/n regime (Huber, 1973). In fact, Huber
(1973) shows that β̂ LS is jointly asymtotically normal only if
max(X(X T X)−1 X T )i,i → 0.
i
When p/n → κ ∈ (0, 1), provided X is full rank,
max(X(X T X)−1 X T )i,i ≥
i
p
1
tr(X(X T X)−1 X T ) = → κ > 0.
n
n
In other words, in moderate p/n regime, the asymptotic normality cannot hold for all
linear contrasts, even in the case of least-squares.
In applications, however, it is usually not necessary to consider all linear contrasts but
instead a small subset of them, e.g. all coordinates or low dimensional linear contrasts
such as β1∗ − β2∗ . We can naturally modify Definition 3.12 and adapt to our needs by
imposing constraints on an . A popular concept, which we use in Section 1 informally,
is called coordinate-wise asymptotic normality and defined by restricting an to be the
canonical basis vectors, which have only one non-zero element. An equivalent definition
is stated as follows.
Definition 3.13 (coordinate-wise asymptotic normal).
14
{Wn : Wn ∈ Rpn } is coordinate-wise asymptotically normal if and only if for any
sequence {jn : jn ∈ {1, . . . , pn }},
!
Wn,jn − EWn,jn
p
L
→ N (0, 1).
Var(Wn,jn )
A more convenient way to define the coordinate-wise asymptotic normality is to introduce a metric d(·, ·), e.g. Kolmogorov distance and total variation distance, which induces
the weak convergence topology. Then Wn is coordinate-wise asymptotically normal if and
only if
!
!
Wn,j − EWn,j
, N (0, 1) = o(1).
max d L p
j
Var(Wn,j )
3.4.2
Discussion about inference and technical assumptions
Variance and bias estimation
To complete the inference, we need to compute the bias and variance. As discussed in
Remark 3.2, the M-estimator is unbiased if the loss function and the error distribution are
symmetric. For the variance, it is easy to get a conservative estimate via resampling methods such as Jackknife as a consequence of Efron-Stein’s inequality; see El Karoui (2013)
and El Karoui and Purdom (2015) for details. Moreover, by the variance decomposition
formula,
h
i
h
i
h
i
Var(β̂j ) = E Var(β̂j |X) + Var E(β̂j |X) ≥ E Var(β̂j |X) ,
the unconditional variance, when X is a random design matrix, is a conservative estimate.
The unconditional variance can be calculated by solving a non-linear system; see El Karoui
(2013) and Donoho and Montanari (2016).
However, estimating the exact variance is known to be hard. El Karoui and Purdom
(2015) show that the existing resampling schemes, including jacknife, pairs-bootstrap,
residual bootstrap, etc., are either too conservative or too anti-conservative when p/n is
large. The challenge, as mentioned in El Karoui (2013); El Karoui and Purdom (2015),
is due to the fact that the residuals {Ri } do not mimic the behavior of {i } and that
the resampling methods effectively modifies the geometry of the dataset from the point
of view of the statistics of interest. We believe that variance estimation in moderate
p/n regime should rely on different methodologies from the ones used in low-dimensional
estimation.
Technical assumptions
On the other hand, we assume that ρ is strongly convex. One remedy would be adding
a ridge regularized term as in El Karoui (2013) and the new problem is amenable to analysis with the method we used in this article. However, the regularization term introduces
a non-vanishing bias, which is as hard to be derived as the variance. For unregularized Mestimators, the strong convexity is also assumed by other works (El Karoui, 2013; Donoho
& Montanari, 2016). However, we believe that this assumption is unnecessary and can
be removed at least for well-behaved design matrices. Another possibility, for errors that
have more than 2 moments is to just add a small quadratic term to the loss function, e.g.
λx2 /2 with a small λ. Finally, we recall that in many situations, least-squares is actually
more efficient than `1 -regression (see numerical work in Bean et al. (2013)) in moderate
dimensions. This is for instance the case for double-exponential errors if p/n is greater
than .3 or so. As such working with strongly convex loss functions is as problematic
for moderate-dimensional regression M-estimates as it would be in the low-dimensional
setting.
To explore traditional robustness questions, we will need to weaken the requirements
of Assumption A2. This requires substantial work and an extension of the main results
15
of Chatterjee (2009). Because the technical part of the paper is already long, we leave
this interesting statistical question to future works.
4
Proof Sketch
Since the proof of Theorem 3.1 is somewhat technical, we illustrate the main idea in this
section.
First notice that the M-estimator β̂ is an implicit function of independent random
variables 1 , . . . , n , which is determined by
n
1X
xi ψ(i − xi β̂) = 0.
n i=1
(8)
The Hessian matrix of the loss function in (5) is n1 X T DX D0 λ− Ip under the notation
introduced in section 3.1. The assumption A3 then implies that the loss function is
strongly convex, in which case β̂ is unique. Then β̂ can be seen as a non-linear function
of i ’s. A powerful central limit theorem for this type of statistics is the second-order
Poincaré inequality (SOPI), developed in Chatterjee (2009) and used there to re-prove
central limit theorems for linear spectral statistics of large random matrices. We recall
one of the main results for the convenience of the reader.
Proposition 4.1 (SOPI; Chatterjee, 2009). Let W = (W1 , . . . , Wn ) = (u1 (W1 ), . . . , un (Wn ))
i.i.d.
where Wi ∼ N (0, 1) and ku0i k∞ ≤ c1 , ku00i k∞ ≤ c2 . Take any g ∈ C 2 (Rn ) and let ∇i g,
∇g and ∇2 g denote the i-th partial derivative, gradient and Hessian of g. Let
κ0 =
E
n
X
! 21
∇i g(W )
4
,
1
1
κ1 = (Ek∇g(W )k42 ) 4 ,
κ2 = (Ek∇2 g(W )k4op ) 4 ,
i=1
and U = g(W ). If U has finite fourth moment, then
!
!
√
2 5(c1 c2 κ0 + c31 κ1 κ2 )
U − EU
, N (0, 1) ≤
dTV L p
.
Var(U )
Var(U )
From (8), it is not hard to compute the gradient and Hessian of β̂j with respect to .
Recalling the definitions in Equation (6) on p. 8, we have
Lemma 4.2. Suppose ψ ∈ C 2 (Rn ), then
∂ β̂j
= eTj (X T DX)−1 X T D
∂T
(9)
∂ β̂j
= GT diag(eTj (X T DX)−1 X T D̃)G
∂∂T
where ej is the j-th cononical basis vectors in Rp and
(10)
G = I − X(X T DX)−1 X T D.
Recalling the definitions of Ki ’s in Assumption A1 on p. 9, we can bound κ0 , κ1 and
κ2 as follows.
Lemma 4.3. Let κ0j , κ1j , κ2j defined as in Proposition 4.1 by setting W = and g(W ) =
β̂j . Let
1
(11)
Mj = EkeTj (X T DX)−1 X T D 2 k∞ ,
then
κ20j
≤
K12
3
(nK0 λ− ) 2
· Mj ,
κ41j
K12
≤
,
(nK0 λ− )2
16
κ42j
≤
K24
3
(nK0 λ− ) 2
·
K1
K0
4
· Mj .
As a consequence of the second-order Poincaré inequality , we can bound the total variation distance between β̂j and a normal distribution by Mj and Var(β̂j ). More
precisely, we prove the following Lemma.
Lemma 4.4. Under assumptions A1-A3,
β̂j − Eβ̂j
, N (0, 1) = Op
max dTV L q
j
Var(β̂j )
1
maxj (nMj2 ) 8
n · minj Var(β̂j )
!
· polyLog(n) .
Lemma 4.4 is the key to prove Theorem 3.1. To obtain the coordinate-wise asymptotic
normality, it is left to establish an upper bound for Mj and a lower bound for Var(β̂j ).
In fact, we can prove that
Lemma 4.5. Under assumptions A1 - A5,
1
polyLog(n)
, min Var(β̂j ) = Ω
.
max Mj = O
j
j
n
n · polyLog(n)
Then Lemma 4.4 and Lemma 4.5 together imply that
β̂
−
E
β̂
polyLog(n)
j
j
max dTV L q
, N (0, 1) = O
= o(1).
1
j
n8
Var(β̂j )
Appendix A, provides a roadmap of the proof of Lemma 4.5 under a special case where
the design matrix X is one realization of a random matrix with i.i.d. sub-gaussian entries.
It also serves as an outline of the rigorous proof in Appendix B.
4.1
Comment on the Second-Order Poincaré inequality
Pn
Notice that when g is a linear function such that g(z) = i=1 ai zi , then the Berry-Esseen
inequality (Esseen, 1945) implies that
!
!
Pn
|ai |3
W − EW
dK L p
, N (0, 1) Pi=1
3 ,
n
Var(W )
( i=1 a2i ) 2
where
dK (F, G) = sup |F (x) − G(x)|.
x
On the other hand, the second-order Poincaré inequality implies that
!
!
!
!
1
Pn
4 2
W − EW
W − EW
i=1 ai
, N (0, 1) ≤ dTV L p
, N (0, 1) Pn
dK L p
.
2
Var(W )
Var(W )
i=1 ai
This is slightly worse than the Berry-Esseen bound and requires stronger conditions on
the distributions of variates but provides bounds for TV metric instead of Kolmogorov
metric. This comparison shows that second-order Poincaré inequality can be regarded as
a generalization of the Berry-Esseen bound for non-linear transformations of independent
random variables.
5
Least-Squares Estimator
The Least-Squares Estimator is a special case of an M-estimator with ρ(x) = 12 x2 . Because
the estimator can then be written explicitly, the analysis of its properties is extremely
simple and it has been understood for several decades (see arguments in e.g. Huber
(1973)[Lemma 2.1] and Huber (1981)[Proposition 2.2]). In this case, the hat matrix
H = X(X T X)−1 X T captures all the problems associated with dimensionality in the
problem. In particular, proving the asymptotic normality simply requires an application
of the Lindeberg-Feller theorem.
It is however somewhat helpful to compare the conditions required for asymptotic
normality in this simple case and the ones we required in the more general setup of
Theorem 3.1. We do so briefly in this section.
17
5.1
Coordinate-Wise Asymptotic Normality of LSE
Under the linear model (1), when X is full rank,
β̂ LS = β ∗ + (X T X)−1 X T ,
thus each coordinate of β̂ LS is a linear contrast of with zero mean. Instead of assumption
A2, which requires i to be sub-gaussian, we only need to assume maxi E|i |3 < ∞, under
which the Berry-Essen bound for non-i.i.d. data (Esseen, 1945) implies that
T
−1 T
T
−1 T 3
β̂j − βj∗
, N (0, 1) kej (X X) X k3 ≤ kej (X X) X k∞ .
dK L q
kej (X T X)−1 X T k2
keTj (X T X)−1 X T k32
Var(β̂ )
j
This motivates us to define a matrix specific quantity Sj (X) such that
Sj (X) =
keTj (X T X)−1 X T k∞
keTj (X T X)−1 X T k2
(12)
then the Berry-Esseen bound implies that maxj∈Jn Sj (X) determines the coordinate-wise
asymptotic normality of β̂ LS .
Theorem 5.1. If E maxi |i |3 < ∞, then
β̂
−
β
E|i |3
LS,j
0,j
max dK q
, N (0, 1) ≤ A ·
3 · max Sj (X),
j∈Jn
(E2i ) 2 j∈Jn
Var(β̂LS,j )
where A is an absolute constant and dK (·, ·) is the Kolmogorov distance, defined as
dK (F, G) = sup |F (x) − G(x)|.
x
It turns out that maxj∈Jn Sj (X) plays in the least-squares setting the role of ∆C in
assumption A5. Since it has been known that a condition like Sj (X) → 0 is necessary
for asymptotic normality of least-square estimators (Huber (1973)[Proposition 2.2]), this
shows in particular that our Assumption A5, or a variant, is also needed in the general
case. See Appendix C-4.1 for details.
5.2
Discussion
Naturally, checking the conditions for asymptotic normality is much easier in the leastsquares case than in the general case under consideration in this paper. In particular:
1. Asymptotic normality conditions can be checked for a broader class of random
design matrices. See Appendix C-4.2 for details.
kX k
.
2. For orthogonal design matrices, i.e X T X = cId for some c > 0, Sj (X) = kXjj k∞
2
Hence, the condition Sj (X) = o(1) is true if and only if no entry dominates the
j − th row of X.
3. The ANOVA-type counterexample we gave in Section 3.3.4 still provides a counterexample. The reason now is different: namely the sum of finitely many independent
random variables is evidently in general non-Gaussian. In fact, in this case, Sj (X) =
√1 is bounded away from 0.
nj
Inferential questions are also extremely simple in this context and essentially again dimensionindependent for the reasons highlighted above. Theorem 5.1 naturally reads,
β̂j − βj∗
d
q
→ N (0, 1).
T
T
−1
σ ej (X X) ej
18
(13)
Estimating σ is still simple under minimal conditions provided n − p → ∞: see Bickel
and Freedman (1983)[Theorem 1.3] or standard computations concerning the normalized
residual sum-of-squares (using variance computations for the latter may require up to 4
moments for i ’s). Then we can replace σ in (13) by σ̂ with
n
σ̂ 2 =
1 X 2
Rk
n−p
k=1
where Rk = yk − xTk β̂ and construct confidence intervals for βj∗ based on σ̂. If n − p does
not tend to ∞, the normalized residual sum of squares is evidently not consistent even in
the case of Gaussian errors, so this requirement may not be dispensed of.
6
Numerical Results
As seen in the previous sections and related papers, there are five important factors that
affect the distribution of β̂: the design matrix X, the error distribution L(), the sample
size n, the ratio κ, and the loss function ρ. The aim of this section is to assess the
quality of the agreement between the asymptotic theoretical results of Theorem 3.1 and
the empirical, finite-dimensional properties of β̂(ρ). We also perform a few simulations
where some of the assumptions of Theorem 3.1 are violated to get an intuitive sense of
whether those assumptions appear necessary or whether they are simply technical artifacts
associated with the method of proof we developed. As such, the numerical experiments
we report on in this section can be seen as a complement to Theorem 3.1 rather than
only a simple check of its practical relevance.
The design matrices we consider are one realization of random design matrices of the
following three types:
i.i.d.
(i.i.d. design) : Xij ∼ F ;
i.i.d.
i.i.d.
(elliptical design) : Xij = ζi X̃ij , where X̃ij ∼ N (0, 1) and ζi ∼ F . In addition,
{ζi } is independent of {X̃ij };
(partial Hadamard design) : a matrix formed by a random set of p columns of a n×n
Hadamard matrix, i.e. a n × n matrix whose columns are orthogonal with entries
restricted to ±1.
Here we consider two candidates for F in i.i.d. design and elliptical design: standard
normal distribution N (0, 1) and t-distribution with two degrees of freedom (denoted t2 ).
For the error distribution, we assume that has i.i.d. entries with one of the above two
distributions, namely N (0, 1) and t2 . The t-distribution violates our assumption A2.
To evaluate the finite sample performance, we consider the sample sizes n ∈ {100, 200, 400, 800}
and κ ∈ {0.5, 0.8}. In this section we will consider a Huber loss with k = 1.345 (Huber,
1981), i.e.
1 2
|x| ≤ k
2x
ρ(x) =
k2
kx − 2 |x| > k
k = 1.345 is the default in R and yields 95% relative efficiency for Gaussian errors in
low-dimensional problems. We also carried out the numerical work for L1 -regression, i.e.
ρ(x) = |x|. See Appendix D for details.
6.1
Asymptotic Normality of A Single Coordinate
First we simulate the finite sample distribution of β̂1 , the first coordinate of β̂. For each
combination of sample size n (100, 200, 400 and 800), type of design (i.i.d, elliptical and
Hadamard), entry distribution F (normal and t2 ) and error distribution L() (normal
and t2 ), we run 50 simulations with each consisting of the following steps:
(Step 1) Generate one design matrix X;
19
(Step 2) Generate the 300 error vectors ;
(Step 3) Regress each Y = on the design matrix X and end up with 300 random samples
(1)
(300)
of β̂1 , denoted by β̂1 , . . . , β̂1
;
b
(Step 4) Estimate the standard deviation of β̂1 by the sample standard error sd;
h
i
(k)
b β̂ (k) + 1.96 · sd
b for each
(Step 5) Construct a confidence interval I (k) = β̂1 − 1.96 · sd,
1
k = 1, . . . , 300;
(Step 6) Calculate the empirical 95% coverage by the proportion of confidence intervals
which cover the true β1 = 0.
Finally, we display the boxplots of the empirical 95% coverages of β̂1 for each case in
Figure 2. It is worth mentioning that our theories cover two cases: 1) i.i.d design with
normal entries and normal errors (orange bars in the first row and the first column), see
Proposition 3.4; 2) elliptical design with normal factors ζi and normal errors (orange bars
in the second row and the first column), see Proposition 3.10.
We first discuss the case κ = 0.5. In this case, there are only two samples per
parameter. Nonetheless, we observe that the coverage is quite close to 0.95, even with a
sample size as small as 100, in both cases that are covered by our theories. For other cases,
it is interesting to see that the coverage is valid and most stable in the partial hadamard
design case and is not sensitive to the distribution of multiplicative factor in elliptical
design case even when the error has a t2 distribution. For i.i.d. designs, the coverage
is still valid and stable when the entry is normal. By contrast, when the entry has a t2
distribution, the coverage has a large variation in small samples. The average coverage
is still close to 0.95 in the i.i.d. normal design case but is slightly lower than 0.95 in the
i.i.d. t2 design case. In summary, the finite sample distribution of β̂1 is more sensitive
to the entry distribution than the error distribution. This indicates that the assumptions
on the design matrix are not just artifacts of the proof but are quite essential.
The same conclusion can be drawn from the case where κ = 0.8 except that the
variation becomes larger in most cases when the sample size is small. However, it is
worth pointing out that even in this case where there is 1.25 samples per parameter, the
sample distribution of β̂1 is well approximated by a normal distribution with a moderate
sample size (n ≥ 400). This is in contrast to the classical rule of thumb which suggests
that 5-10 samples are needed per parameter.
6.2
Asymptotic Normality for Multiple Marginals
Since our theory holds for general Jn , it is worth checking the approximation for multiple
coordinates in finite samples. For illustration, we consider 10 coordinates, namely β̂1 ∼
β̂10 , simultaneously and calculate the minimum empirical 95% coverage. To avoid the
finite sample dependence between coordinates involved in the simulation, we estimate the
empirical coverage independently for each coordinate. Specifically, we run 50 simulations
with each consisting of the following steps:
(Step 1) Generate one design matrix X;
(Step 2) Generate the 3000 error vectors ;
(Step 3) Regress each Y = on the design matrix X and end up with 300 random samples
of β̂j for each j = 1, . . . , 10 by using the (300(j − 1) + 1)-th to 300j-th response
vector Y ;
b j for
(Step 4) Estimate the standard deviation of β̂j by the sample standard error sd
j = 1, . . . , 10;
h
i
(k)
(k)
b j , β̂ (k) + 1.96 · sd
b j for
(Step 5) Construct a confidence interval Ij = β̂j − 1.96 · sd
j
each j = 1, . . . , 10 and k = 1, . . . , 300;
20
^
Coverage of β1 (κ = 0.5)
normal
^
Coverage of β1 (κ = 0.8)
t(2)
normal
1.00
●
●
●
●
●
0.95
●
1.00
1.00
●
●
●
●
●
●
●
0.90
●
●
●
●
●
●
1.00
0.95
●
●
●
●
●
●
●
●
100
200
400
800
ellip
●
Coverage
0.90
0.95
●
●
0.90
ellip
Coverage
●
iid
●
iid
0.95
t(2)
1.00
0.90
1.00
0.90
100
200
400
800
100
200
400
hadamard
hadamard
0.95
0.95
0.90
800
100
200
400
Sample Size
Entry Dist.
●
normal
t(2)
800
Sample Size
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 2: Empirical 95% coverage of β̂1 with κ = 0.5 (left) and κ = 0.8 (right) using
Huber1.345 loss. The x-axis corresponds to the sample size, ranging from 100 to 800;
the y-axis corresponds to the empirical 95% coverage. Each column represents an error
distribution and each row represents a type of design. The orange solid bar corresponds
to the case F = Normal; the blue dotted bar corresponds to the case F = t2 ; the red
dashed bar represents the Hadamard design.
(Step 6) Calculate the empirical 95% coverage by the proportion of confidence intervals
which cover the true βj = 0, denoted by Cj , for each j = 1, . . . , 10,
(Step 7) Report the minimum coverage min1≤j≤10 Cj .
If the assumptions A1 - A5 are satisfied, min1≤j≤10 Cj should also be close to 0.95 as
a result of Theorem 3.1. Thus, min1≤j≤10 Cj is a measure for the approximation accuracy
for multiple marginals. Figure 3 displays the boxplots of this quantity under the same
scenarios as the last subsection. In two cases that our theories cover, the minimum
coverage is increasingly closer to the true level 0.95. Similar to the last subsection,
the approximation is accurate in the partial hadamard design case and is insensitive
to the distribution of multiplicative factors in the elliptical design case. However, the
approximation is very inaccurate in the i.i.d. t2 design case. Again, this shows the
evidence that our technical assumptions are not artifacts of the proof.
On the other hand, the figure 3 suggests using a conservative variance estimator, e.g.
the Jackknife estimator, or corrections on the confidence level in order to make simultaneous inference on multiple coordinates. Here we investigate the validity of Bonferroni
correction by modifying the step 5 and step 6. The confidence interval after Bonferroni
correction is obtained by
h
i
(k)
(k)
b j , β̂ (k) + z1−α/20 · sd
bj
Ij = β̂j − z1−α/20 · sd
(14)
j
where α = 0.05 and zγ is the γ-th quantile of a standard normal distribution. The
(k)
proportion of k such that 0 ∈ Ij for all j ≤ 10 should be at least 0.95 if the marginals
are all close to a normal distribution. We modify the confidence intervals in step 5 by
(k)
(14) and calculate the proportion of k such that 0 ∈ Ij for all j in step 6. Figure 4
displays the boxplots of this coverage. It is clear that the Bonferroni correction gives the
valid coverage except when n = 100, κ = 0.8 and the error has a t2 distribution.
21
^
^
Min. coverage of β1 ~ β10 (κ = 0.5)
normal
^
^
Min. coverage of β1 ~ β10 (κ = 0.8)
t(2)
normal
0.95
0.90
t(2)
0.95
●
●
●
●
●
●
●
●
0.90
●
●
●
●
●
●
●
●
0.80
0.95
●
0.90
●
●
●
●
●
ellip
●
●
Coverage
0.95
●
ellip
Coverage
iid
0.80
0.95
0.90
●
●
iid
0.80
●
●
●
●
●
0.80
0.95
0.80
100
200
400
800
100
200
400
hadamard
hadamard
0.90
0.90
0.80
800
100
200
400
Sample Size
Entry Dist.
●
normal
t(2)
800
100
200
400
800
Sample Size
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 3: Mininum empirical 95% coverage of β̂1 ∼ β̂10 with κ = 0.5 (left) and κ = 0.8
(right) using Huber1.345 loss. The x-axis corresponds to the sample size, ranging from
100 to 800; the y-axis corresponds to the minimum empirical 95% coverage. Each column
represents an error distribution and each row represents a type of design. The orange
solid bar corresponds to the case F = Normal; the blue dotted bar corresponds to the
case F = t2 ; the red dashed bar represents the Hadamard design.
7
Conclusion
We have proved coordinate-wise asymptotic normality for regression M-estimates in the
moderate-dimensional asymptotic regime p/n → κ ∈ (0, 1), for fixed design matrices
under appropriate technical assumptions. Our design assumptions are satisfied with high
probability for a broad class of random designs. The main novel ingredient of the proof
is the use of the second-order Poincaré inequality. Numerical experiments confirm and
complement our theoretical results.
22
^
^
Bonf. coverage of β1 ~ β10 (κ = 0.5)
normal
^
^
Bonf. coverage of β1 ~ β10 (κ = 0.8)
t(2)
normal
1.00
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
iid
0.95
iid
0.95
●
0.90
0.90
1.00
●
●
●
●
●
●
●
●
●
●
●
●
●
0.95
●
●
0.90
0.90
1.00
1.00
hadamard
hadamard
0.95
●
ellip
ellip
0.95
Coverage
1.00
Coverage
t(2)
1.00
0.95
0.90
0.90
100
200
400
800
100
200
400
800
100
200
400
Sample Size
Entry Dist.
●
normal
t(2)
800
100
200
400
800
Sample Size
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 4: Empirical 95% coverage of β̂1 ∼ β̂10 after Bonferroni correction with κ = 0.5
(left) and κ = 0.8 (right) using Huber1.345 loss. The x-axis corresponds to the sample size,
ranging from 100 to 800; the y-axis corresponds to the empirical uniform 95% coverage
after Bonferroni correction. Each column represents an error distribution and each row
represents a type of design. The orange solid bar corresponds to the case F = Normal;
the blue dotted bar corresponds to the case F = t2 ; the red dashed bar represents the
Hadamard design.
References
Anderson, T. W. (1962). An introduction to multivariate statistical analysis. Wiley New
York.
Bai, Z., & Silverstein, J. W. (2010). Spectral analysis of large dimensional random
matrices (Vol. 20). Springer.
Bai, Z., & Yin, Y. (1993). Limit of the smallest eigenvalue of a large dimensional sample
covariance matrix. The annals of Probability, 1275–1294.
Baranchik, A. (1973). Inadmissibility of maximum likelihood estimators in some multiple regression problems with three or more independent variables. The Annals of
Statistics, 312–321.
Bean, D., Bickel, P., El Karoui, N., Lim, C., & Yu, B. (2012). Penalized robust regression
in high-dimension. Technical Report 813, Department of Statistics, UC Berkeley.
Bean, D., Bickel, P. J., El Karoui, N., & Yu, B. (2013). Optimal M-estimation in highdimensional regression. Proceedings of the National Academy of Sciences, 110 (36),
14563–14568.
Bickel, P. J., & Doksum, K. A. (2015). Mathematical statistics: Basic ideas and selected
topics, volume i (Vol. 117). CRC Press.
Bickel, P. J., & Freedman, D. A. (1981). Some asymptotic theory for the bootstrap. The
Annals of Statistics, 1196–1217.
Bickel, P. J., & Freedman, D. A. (1983). Bootstrapping regression models with many
parameters. Festschrift for Erich L. Lehmann, 28–48.
Chatterjee, S. (2009). Fluctuations of eigenvalues and second order poincaré inequalities.
Probability Theory and Related Fields, 143 (1-2), 1–40.
Chernoff, H. (1981). A note on an inequality involving the normal distribution. The
Annals of Probability, 533–535.
23
Cizek, P., Härdle, W. K., & Weron, R. (2005). Statistical tools for finance and insurance.
Springer Science & Business Media.
Cochran, W. G. (1977). Sampling techniques. John Wiley & Sons.
David, H. A., & Nagaraja, H. N. (1981). Order statistics. Wiley Online Library.
Donoho, D., & Montanari, A. (2016). High dimensional robust m-estimation: Asymptotic
variance via approximate message passing. Probability Theory and Related Fields,
166 , 935-969.
Durrett, R. (2010). Probability: theory and examples. Cambridge university press.
Efron, B., & Efron, B. (1982). The jackknife, the bootstrap and other resampling plans
(Vol. 38). SIAM.
El Karoui, N. (2009). Concentration of measure and spectra of random matrices: applications to correlation matrices, elliptical distributions and beyond. The Annals of
Applied Probability, 19 (6), 2362–2405.
El Karoui, N. (2010). High-dimensionality effects in the markowitz problem and other
quadratic programs with linear constraints: Risk underestimation. The Annals of
Statistics, 38 (6), 3487–3566.
El Karoui, N. (2013). Asymptotic behavior of unregularized and ridge-regularized
high-dimensional robust regression estimators: rigorous results. arXiv preprint
arXiv:1311.2445 .
El Karoui, N. (2015). On the impact of predictor geometry on the performance on highdimensional ridge-regularized generalized robust regression estimators. Technical
Report 826, Department of Statistics, UC Berkeley.
El Karoui, N., Bean, D., Bickel, P., Lim, C., & Yu, B. (2011). On robust regression with
high-dimensional predictors. Technical Report 811, Department of Statistics, UC
Berkeley.
El Karoui, N., Bean, D., Bickel, P. J., Lim, C., & Yu, B. (2013). On robust regression
with high-dimensional predictors. Proceedings of the National Academy of Sciences,
110 (36), 14557–14562.
El Karoui, N., & Purdom, E. (2015). Can we trust the bootstrap in high-dimension?
Technical Report 824, Department of Statistics, UC Berkeley.
Esseen, C.-G. (1945). Fourier analysis of distribution functions. a mathematical study of
the laplace-gaussian law. Acta Mathematica, 77 (1), 1–125.
Geman, S. (1980). A limit theorem for the norm of random matrices. The Annals of
Probability, 252–261.
Hanson, D. L., & Wright, F. T. (1971). A bound on tail probabilities for quadratic forms
in independent random variables. The Annals of Mathematical Statistics, 42 (3),
1079–1083.
Horn, R. A., & Johnson, C. R. (2012). Matrix analysis. Cambridge university press.
Huber, P. J. (1964). Robust estimation of a location parameter. The Annals of Mathematical Statistics, 35 (1), 73–101.
Huber, P. J. (1972). The 1972 wald lecture robust statistics: A review. The Annals of
Mathematical Statistics, 1041–1067.
Huber, P. J. (1973). Robust regression: asymptotics, conjectures and monte carlo. The
Annals of Statistics, 799–821.
Huber, P. J. (1981). Robust statistics. John Wiley & Sons, Inc., New York.
Huber, P. J. (2011). Robust statistics. Springer.
Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in principal components analysis. Annals of statistics, 295–327.
Jurečkovà, J., & Klebanov, L. (1997). Inadmissibility of robust estimators with respect
to l1 norm. Lecture Notes-Monograph Series, 71–78.
Latala, R. (2005). Some estimates of norms of random matrices. Proceedings of the
American Mathematical Society, 133 (5), 1273–1282.
Ledoux, M. (2001). The concentration of measure phenomenon (No. 89). American
Mathematical Soc.
Litvak, A. E., Pajor, A., Rudelson, M., & Tomczak-Jaegermann, N. (2005). Smallest
singular value of random matrices and geometry of random polytopes. Advances in
24
Mathematics, 195 (2), 491–523.
Mallows, C. (1972). A note on asymptotic joint normality. The Annals of Mathematical
Statistics, 508–515.
Mammen, E. (1989). Asymptotics with increasing dimension for robust regression with
applications to the bootstrap. The Annals of Statistics, 382–400.
Marčenko, V. A., & Pastur, L. A. (1967). Distribution of eigenvalues for some sets of
random matrices. Mathematics of the USSR-Sbornik , 1 (4), 457.
Muirhead, R. J. (1982). Aspects of multivariate statistical theory (Vol. 197). John Wiley
& Sons.
Portnoy, S. (1984). Asymptotic behavior of M-estimators of p regression parameters when
p2 /n is large. i. consistency. The Annals of Statistics, 1298–1309.
Portnoy, S. (1985). Asymptotic behavior of M estimators of p regression parameters when
p2 /n is large; ii. normal approximation. The Annals of Statistics, 1403–1417.
Portnoy, S. (1986). On the central limit theorem in Rp when p → ∞. Probability theory
and related fields, 73 (4), 571–583.
Portnoy, S. (1987). A central limit theorem applicable to robust regression estimators.
Journal of multivariate analysis, 22 (1), 24–50.
Posekany, A., Felsenstein, K., & Sykacek, P. (2011). Biological assessment of robust noise
models in microarray data analysis. Bioinformatics, 27 (6), 807–814.
Relles, D. A. (1967). Robust regression by modified least-squares. (Tech. Rep.). DTIC
Document.
Rosenthal, H. P. (1970). On the subspaces ofl p (p¿ 2) spanned by sequences of independent random variables. Israel Journal of Mathematics, 8 (3), 273–303.
Rudelson, M., & Vershynin, R. (2009). Smallest singular value of a random rectangular
matrix. Communications on Pure and Applied Mathematics, 62 (12), 1707–1739.
Rudelson, M., & Vershynin, R. (2010). Non-asymptotic theory of random matrices:
extreme singular values. arXiv preprint arXiv:1003.2990 .
Rudelson, M., & Vershynin, R. (2013). Hanson-wright inequality and sub-gaussian concentration. Electron. Commun. Probab, 18 (82), 1–9.
Scheffe, H. (1999). The analysis of variance (Vol. 72). John Wiley & Sons.
Silverstein, J. W. (1985). The smallest eigenvalue of a large dimensional wishart matrix.
The Annals of Probability, 1364–1368.
Stone, M. (1974). Cross-validatory choice and assessment of statistical predictions. Journal of the Royal Statistical Society. Series B (Methodological), 111–147.
Tyler, D. E. (1987). A distribution-free M-estimator of multivariate scatter. The Annals
of Statistics, 234–251.
Van der Vaart, A. W. (1998). Asymptotic statistics. Cambridge university press.
Vershynin, R. (2010). Introduction to the non-asymptotic analysis of random matrices.
arXiv preprint arXiv:1011.3027 .
Wachter, K. W. (1976). Probability plotting points for principal components. In Ninth
interface symposium computer science and statistics (pp. 299–308).
Wachter, K. W. (1978). The strong limits of random matrix spectra for sample matrices
of independent elements. The Annals of Probability, 1–18.
Wasserman, L., & Roeder, K. (2009). High dimensional variable selection. Annals of
statistics, 37 (5A), 2178.
Yohai, V. J. (1972). Robust M estimates for the general linear model. Universidad
Nacional de la Plata. Departamento de Matematica.
Yohai, V. J., & Maronna, R. A. (1979). Asymptotic behavior of M-estimators for the
linear model. The Annals of Statistics, 258–268.
25
APPENDIX
A
Proof Sketch of Lemma 4.5
In this Appendix, we provide a roadmap for proving Lemma 4.5 by considering a special
case where X is one realization of a random matrix Z with i.i.d. mean-zero σ 2 -subgaussian entries. Random matrix
theory (Geman, 1980; Silverstein,
√ 1985; Bai & Yin,
√
1993) implies that λ+ = (1 + κ)2 + op (1) = Op (1) and λ− = (1 − κ)2 + op (1) = Ωp (1).
Thus, the assumption A3 is satisfied with high probability. Thus, the Lemma 4.4 in p. 16
holds with high probability. It remains to prove the following lemma to obtain Theorem
3.1.
Lemma A.1. Let Z be a random matrix with i.i.d. mean-zero σ 2 -sub-gaussian entries
and X be one realization of Z. Then under assumptions A1 and A2,
1
polyLog(n)
,
min Var(β̂j ) = Ωp
,
max Mj = Op
1≤j≤p
1≤j≤p
n
n · polyLog(n)
where Mj is defined in (11) in p.16 and the randomness in op (·) and Op (·) comes from
Z.
A-1
Upper Bound of Mj
First by Proposition E.3,
λ+ = Op (1),
λ− = Ωp (1).
In the rest of the proof, the symbol E and Var denotes the expectation and the variance
1
conditional on Z. Let Z̃ = D 2 Z, then Mj = EkeTj (Z̃ T Z̃)−1 Z̃ T k∞ . Let H̃j = I −
T
T
Z̃[j] (Z̃[j]
Z̃[j] )−1 Z̃[j]
, then by block matrix inversion formula (see Proposition E.1), which
we state as Proposition E.1 in Appendix E.
!−1
T
Z̃
Z̃
Z̃1
[1]
1
T
−1 T
(Z̃ Z̃) Z̃ =
T
Z̃[1]
Z̃[1]
Z̃[1]
T
1
1 −Z̃1T Z̃[1] (Z̃[1]
Z̃[1] )−1
Z̃1
= T
Z̃[1]
∗
∗
Z̃1 (I − H̃1 )Z̃1
T
1
Z̃1 (I − H̃1 )
= T
.
∗
Z̃1 (I − H̃1 )Z̃1
Z̃1T Z̃1
T
Z̃[1]
Z̃1
This implies that
M1 = E
kZ̃1T (I − H̃1 )k∞
.
Z̃1T (I − H̃1 )Z̃1
(A-1)
Since Z T DZ/n K0 λ− I, we have
1
1
= eT1 (Z̃ T Z̃)−1 e1 = eT1 (Z T DZ)−1 e1 = eT1
T
n
Z̃1 (I − H̃1 )Z̃1
Z T DZ
n
−1
e1 ≤
1
nK0 λ−
and we obtain a bound for M1 as
1
M1 ≤
EkZ̃1T (I − H̃1 )k∞
EkZ1T D 2 (I − H̃1 )k∞
=
.
nK0 λ−
nK0 λ−
Similarly,
1
1
1
T
T
EkZjT D 2 (I − D 2 Z[j]
(Z[j]
DZ[j] )−1 Z[j] D 2 )k∞
EkZjT D 2 (I − H̃j )k∞
Mj ≤
=
.
nK0 λ−
nK0 λ−
1
26
(A-2)
The vector in the numerator is a linear contrast of Zj and Zj has mean-zero i.i.d. subgaussian entries. For any fixed matrix A ∈ Rn×n , denote Ak by its k-th column, then
ATk Zj is σ 2 kAk k22 -sub-gaussian (see Section 5.2.3 of Vershynin (2010) for a detailed discussion) and hence by definition of sub-Gaussianity,
t2
P (|ATk Zj | ≥ σkAk k2 t) ≤ 2e− 2 .
Therefore, by a simple union bound, we conclude that
t2
P (kAT Zj k∞ ≥ σ max kAk k2 t) ≤ 2ne− 2 .
k
√
Let t = 2 log n,
P (kAT Zj k∞ ≥ 2σ max kAk k2
k
p
2
log n) ≤ = o(1).
n
This entails that
T
kA Zj k∞ = Op
max kAk k2 · polyLog(n) = Op (kAkop · polyLog(n)) .
k
(A-3)
1
with high probability. In Mj , the coefficient matrix (I − Hj )D 2 depends on Zj through
D and hence we cannot use (A-3) directly. However, the dependence can be removed by
replacing D by D[j] since ri,[j] does not depend on Zj .
Since Z has i.i.d. sub-gaussian entries, no column is highly influential. In other words,
the estimator will not change drastically after removing j-th column. This would suggest
Ri ≈ ri,[j] . It is proved by El Karoui (2013) that
polyLog(n)
√
sup |Ri − ri,[j] | = Op
.
n
i,j
It can be rigorously proved that
kZjT D(I − H̃j )k∞ − kZjT D[j] (I − Hj )k∞ = Op
1
polyLog(n)
n
,
1
T
T
2
2
where Hj = I − D[j]
Z[j] (Z[j]
D[j] Z[j] )−1 Z[j]
D[j]
; see Appendix A-1 for details. Since
D[j] (I − Hj ) is independent of Zj and
kD[j] (I − Hj )kop ≤ kD[j] kop ≤ K1 = O (polyLog(n)) ,
it follows from (A-2) and (A-3) that
kZjT D[j] (I
− Hj )k∞ = Op
polyLog(n)
n
.
In summary,
Mj = Op
A-2
A-2.1
polyLog(n)
n
.
(A-4)
Lower Bound of Var(β̂j )
Approximating Var(β̂j ) by Var(bj )
It is shown by El Karoui (2013)1 that
1 Nj
β̂j ≈ bj , √
n ξj
(A-5)
1 El Karoui (2013) considers a ridge regularized M estimator, which is different from our setting.
However, this argument still holds in our case and proved in Appendix B.
27
where
n
1 X
Nj = √
Zij ψ(ri,[j] ),
n i=1
1 T
T
T
Z (D[j] − D[j] Z[j] (X[j]
D[j] X[j] )−1 Z[j]
D[j] )Zj .
n j
ξj =
It has been shown by El Karoui (2013) that
max |β̂j − bj | = Op
j
polyLog(n)
n
.
Thus, Var(β̂j ) ≈ Var(bj ) and a more refined calculation in Appendix A-2.1 shows that
polyLog(n)
.
| Var(β̂j ) − Var(bj )| = Op
3
n2
It is left to show that
Var(bj ) = Ωp
A-2.2
1
n · polyLog(n)
.
(A-6)
Bounding Var(bj ) via Var(Nj )
By definition of bj ,
Var(bj ) = Ωp
polyLog(n)
n
⇐⇒ Var
Nj
ξj
= Ωp (polyLog(n)) .
As will be shown in Appendix B-6.4,
Var(ξj ) = Op
polyLog(n)
n
.
As a result, ξj ≈ Eξj and
Var
Nj
ξj
Nj
Eξj
≈ Var
=
Var(Nj )
.
(Eξj )2
As in the previous paper (El Karoui, 2013), we rewrite ξj as
ξj =
1
1
1
1 T 21
T
T
2
2
2
Zj D[j] (I − D[j]
Z[j] (X[j]
D[j] X[j] )−1 Z[j]
D[j]
)D[j]
Zj .
n
The middle matrix is idempotent and hence positive semi-definite. Thus,
ξj ≤
Then we obtain that
1 T
Z D[j] Zj ≤ K1 λ+ = Op (polyLog(n)) .
n j
Var(Nj )
= Ωp
(Eξj )2
Var(Nj )
polyLog(n)
,
and it is left to show that
Var(Nj ) = Ωp
A-2.3
1
polyLog(n)
.
(A-7)
Bounding Var(Nj ) via tr(Qj )
Recall the definition of Nj (A-5), and that of Qj (see Section 3.1 in p.8), we have
Var(Nj ) =
1 T
Z Qj Zj
n j
Notice that Zj is independent of ri,[j] and hence the conditional distribution of Zj given
Qj remains the same as the marginal distribution of Zj . Since Zj has i.i.d. sub-gaussian
28
entries, the Hanson-Wright inequality (Hanson & Wright, 1971; Rudelson & Vershynin,
2013; see Proposition E.2), shown in Proposition E.2, implies that any quadratic form of
Zj , denoted by ZjT Qj Zj is concentrated on its mean, i.e.
2
ZjT Qj Zj ≈ EZj , ZjT Qj Zj = (EZ1j
) · tr(Qj ).
As a consequence, it is left to show that
tr(Qj ) = Ωp
A-2.4
n
polyLog(n)
.
(A-8)
Lower Bound of tr(Qj )
By definition of Qj ,
tr(Qj ) =
n
X
Var(ψ(ri,[j] )).
i=1
To lower bounded the variance of ψ(ri,[j] ), recall that for any random variable W ,
Var(W ) =
1
E(W − W 0 )2 .
2
(A-9)
where W 0 is an independent copy of W . Suppose g : R → R is a function such that
|g 0 (x)| ≥ c for all x, then (A-9) implies that
Var(g(W )) =
c2
1
E(g(W ) − g(W 0 ))2 ≥ E(W − W 0 )2 = c2 Var(W ).
2
2
(A-10)
In other words, (A-10) entails that Var(W ) is a lower bound for Var(g(W )) provided
that the derivative of g is bounded away from 0. As an application, we see that
Var(ψ(ri,[j] )) ≥ K02 Var(ri,[j] )
and hence
tr(Qj ) ≥ K02
n
X
Var(ri,[j] ).
i=1
By the variance decomposition formula,
Var(ri,[j] ) = E Var ri,[j] (i) + Var E ri,[j] (i) ≥ E Var ri,[j] [i] ,
where (i) includes all but i-th entry of . Given (i) , ri,[j] is a function of i . Using (A-10),
we have
Var(ri,[j] |(i) ) ≥ inf
i
∂ri,[j]
∂i
2
· Var(i |(i) ) ≥ inf
i
∂ri,[j]
∂i
2
· Var(i ).
This implies that
Var(ri,[j] ) ≥ E Var ri,[j] [i]
≥ E inf
∂ri,[j]
∂i
2
2
!
· min Var(i ).
i
Summing Var(ri,[j] ) over i = 1, . . . , n, we obtain that
tr(Qj ) =
n
X
X
Var(ri,[j] ) ≥ E
i=1
i
∂ri,[j]
inf
∂i
· min Var(i ).
i
It will be shown in Appendix B-6.3 that under assumptions A1-A3,
X
∂ri,[j] 2
n
E
inf
= Ωp
.
∂i
polyLog(n)
i
This proves (A-8) and as a result,
min Var(β̂j ) = Ωp
j
29
1
n · polyLog(n)
.
(A-11)
B
B-1
Proof of Theorem 3.1
Notation
To be self-contained, we summarize our notations in this subsection. The model we
considered here is
y = Xβ ∗ +
where X ∈ Rn×p be the design matrix and is a random vector with independent entries.
β̂ −Eβ̂j
Notice that the target quantity √j
is shift invariant, we can assume β ∗ = 0 without
Var(β̂j )
loss of generality provided that X has full column rank; see Section 3.1 for details.
Let xTi ∈ R1×p denote the i-th row of X and Xj ∈ Rn×1 denote the j-th column
of X. Throughout the paper we will denote by Xij ∈ R the (i, j)-th entry of X, by
X(i) ∈ R(n−1)×p the design matrix X after removing the i-th row, by X[j] ∈ Rn×(p−1) the
design matrix X after removing the j-th column, by X(i),[j] ∈ R(n−1)×(p−1) the design
matrix after removing both i-th row and j-th column, and by xi,[j] ∈ R1×(p−1) the vector
xi after removing j-th entry. The M-estimator β̂ associated with the loss function ρ is
defined as
n
1X
ρ(k − xTk β).
(B-12)
β̂ = arg min
p
n
β∈R
k=1
Similarly we define the leave-j-th-predictor-out version as
n
β̂[j]
1X
ρ(k − xTk,[j] β).
= arg min
p
n
β∈R
(B-13)
k=1
Based on these notation we define the full residual Rk as
Rk = k − xTk β̂,
k = 1, 2, . . . , n
(B-14)
the leave-j-th-predictor-out residual as
rk,[j] = k − xTk,[j] β̂[j] ,
k = 1, 2, . . . , n, j ∈ Jn .
(B-15)
D̃ = diag(ψ 00 (Rk )),
(B-16)
D̃[j] = diag(ψ 00 (rk,[j] )).
(B-17)
T
T
G[j] = I − X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] .
(B-18)
Four diagonal matrices are defined as
D = diag(ψ 0 (Rk )),
D[j] = diag(ψ 0 (rk,[j] )),
Further we define G and G[j] as
G = I − X(X T DX)−1 X T D,
Let Jn denote the indices of coefficients of interest. We say a ∈]a1 , a2 [ if and only if a ∈
[min{a1 , a2 }, max{a1 , a2 }]. Regarding the technical assumptions, we need the following
quantities
T
T
X X
X X
λ+ = λmax
, λ− = λmin
(B-19)
n
n
be the largest (resp. smallest) eigenvalue of the matrix
canonical basis vector and
hj,0 = (ψ(r1,[j] ), . . . , ψ(rn,[j] ))T ,
XT X
n .
Let ei ∈ Rn be the i-th
hj,1,i = GT[j] ei .
(B-20)
Finally, let
(
|hTj,0 Xj |
|hTj,1,i Xj |
, max
∆C = max max
j∈Jn ||hj,0 ||
i≤n,j∈Jn ||hj,1,i ||
30
)
,
(B-21)
Qj = Cov(hj,0 ).
(B-22)
We adopt Landau’s notation (O(·), o(·), Op (·), op (·)). In addition, we say an = Ω(bn ) if
bn = O(an ) and similarly, we say an = Ωp (bn ) if bn = Op (an ). To simplify the logarithm
factors, we use the symbol polyLog(n) to denote any factor that can be upper bounded
1
to denote any factor that can be
by (log n)γ for some γ > 0. Similarly, we use polyLog(n)
1
0
lower bounded by (log n)γ 0 for some γ > 0.
Finally we restate all the technical assumptions:
1
A1 ρ(0) = ψ(0) = 0 and there exists K0 = Ω polyLog(n)
, K1 , K2 = O (polyLog(n)),
such that for any x ∈ R,
d p 0
|ψ 00 (x)|
≤ K2 ;
( ψ (x)) = p
dx
ψ 0 (x)
K0 ≤ ψ 0 (x) ≤ K1 ,
A2 i = ui (Wi ) where (W1 , . . . , Wn ) ∼ N (0, In×n ) and ui are smooth functions with
ku0i k∞ ≤ c1 and ku00i k∞ ≤
c2 for some c1 , c2 = O(polyLog(n)). Moreover, assume
mini Var(i ) = Ω
1
polyLog(n)
.
A3 λ+ = O(polyLog(n)) and λ− = Ω
A4 minj∈Jn
XjT Qj Xj
tr(Qj )
=Ω
1
polyLog(n)
1
polyLog(n)
;
;
A5 E∆8C = O (polyLog(n)).
B-2
Deterministic Approximation Results
In Appendix A, we use several approximations under random designs, e.g. Ri ≈ ri,[j] . To
prove them, we follow the strategy of El Karoui (2013) which establishes the deterministic
results and then apply the concentration inequalities to obtain high probability bounds.
Note that β̂ is the solution of
n
0 = f (β) ,
1X
xi ψ(i − xTi β),
n i=1
we need the following key lemma to bound kβ1 − β2 k2 by kf (β1 ) − f (β2 )k2 , which can be
calculated explicily.
Lemma B.1. [El Karoui (2013), Proposition 2.1] For any β1 and β2 ,
kβ1 − β2 k2 ≤
1
kf (β1 ) − f (β2 )k2 .
K0 λ−
Proof. By the mean value theorem, there exists νi ∈]i − xTi β1 , i − xTi β2 [ such that
ψ(i − xTi β1 ) − ψ(i − xTi β2 ) = ψ 0 (νi ) · xTi (β2 − β1 ).
Then
n
1X 0
ψ (νi )xi xTi (β1 − β2 )
n i=1
2
!
n
X
1
≥ λmin
ψ 0 (νi )xi xTi · kβ1 − β2 k2
n i=1
kf (β1 ) − f (β2 )k2 =
≥ K0 λ− kβ1 − β2 k2 .
31
Based on Lemma B.1, we can derive the deterministic results informally stated in
Appendix A. Such results are shown by El Karoui (2013) for ridge-penalized M-estimates
and here we derive a refined version for unpenalized M-estimates. Throughout this subsection, we only assume assumption A1. This implies the following lemma,
Lemma B.2. Under assumption A1, for any x and y,
p
p
p
|ψ(x)| ≤ K1 |x|, | ψ 0 (x)− ψ 0 (y)| ≤ K2 |x−y|, |ψ 0 (x)−ψ 0 (y)| ≤ 2 K1 K2 |x−y| , K3 |x−y|.
To state the result, we define the following quantities.
n
1
1X
T = √ max max kxi k2 , max kXj k2 , E =
ρ(i ),
i
j∈Jn
n i=1
n
n
U=
1X
xi (ψ(i ) − Eψ(i ))
n i=1
(B-23)
n
,
U0 =
2
1X
xi Eψ(i )
n i=1
.
(B-24)
2
The following proposition summarizes all deterministic results which we need in the proof.
Proposition B.3. Under Assumption A1,
(i) The norm of M estimator is bounded by
kβ̂k2 ≤
1
(U + U0 );
K0 λ−
(ii) Define bj as
1 Nj
bj = √
n ξj
where
n
1 X
Nj = √
Xij ψ(ri,[j] ),
n i=1
Then
ξj =
1 T
T
T
X (D[j] − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] )Xj ,
n j
√
√
1
2K1
max |bj | ≤ √ ·
· ∆C · E,
j∈Jn
n K0 λ−
(iii) The difference between β̂j and bj is bounded by
max |β̂j − bj | ≤
j∈Jn
1 2K12 K3 λ+ T
· ∆3C · E.
·
7
n
2
K04 λ−
(iv) The difference between the full and the leave-one-predictor-out residual is bounded
by
!
√
√
1
2K12 K3 λ+ T 2
2K1
3
2
max max |Ri − ri,[j] | ≤ √
· ∆C · E + 3
· ∆C · E .
7
j∈Jn i
n
2
K04 λ−
K02 λ−
Proof.
(i) By Lemma B.1,
kβ̂k2 ≤
1
kf (0)k2
kf (β̂) − f (0)k2 =
,
K0 λ−
K0 λ−
since β̂ is a zero of f (β). By definition,
n
n
n
1X
1X
1X
f (0) =
xi ψ(i ) =
xi (ψ(i ) − Eψ(i )) +
xi Eψ(i ).
n i=1
n i=1
n i=1
This implies that
kf (0)k2 ≤ U + U0 .
32
(ii) First we prove that
ξj ≥ K0 λ− .
(B-25)
Since all diagonal entries of D[j] is lower bounded by K0 , we conclude that
!
X T D[j] X
≥ K0 λ− .
λmin
n
Note that ξj is the Schur’s complement (Horn & Johnson, 2012, chapter 0.8) of
X T D[j] X
,
n
we have
ξj−1
=
eTj
X T D[j] X
n
!−1
ej ≤
1
,
K0 λ−
which implies (B-25). As for Nj , we have
Nj =
XjT hj,0
khj,0 k XjT hj,0
√
= √ 2·
.
khj,0 k2
n
n
(B-26)
The the second term is bounded by ∆C by definition, see (B-21). For the first term,
the assumption A1 that ψ 0 (x) ≤ K1 implies that
Z x
Z x 0
1 2
ψ (y)
· ψ(y)dy =
ψ (x).
ρ(x) = ρ(x) − ρ(0) =
ψ(y)dy ≥
K
2K
1
1
0
0
Here we use the fact that sign(ψ(y)) = sign(y). Recall the definition of hj,0 , we
obtain that
sP
sP
n
n
2
p
ψ(r
)
khj,0 k2
i,[j]
i=1
i=1 ρ(ri,[j] )
√
=
≤ 2K1 ·
.
n
n
n
Since β̂[j] is the minimizer of the loss function
n
Pn
i=1
ρ(i − xTi,[j] β[j] ), it holds that
n
1X
1X
ρ(ri,[j] ) ≤
ρ(i ) = E.
n i=1
n i=1
Putting together the pieces, we conclude that
p
√
|Nj | ≤ 2K1 · ∆C E.
By definition of bj ,
(B-27)
√
√
1
2K1
|bj | ≤ √ ·
∆C E.
n K0 λ−
(iii) The proof of this result is almost the same as El Karoui (2013). We state it here for
the sake of completeness. Let b̃j ∈ Rp with
(b̃j )j = bj ,
T
T
(b̃j )[j] = β̂[j] − bj (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj
(B-28)
where the subscript j denotes the j-th entry and the subscript [j] denotes the subvector formed by all but j-th entry. Furthermore, define γj with
(γj )j = −1,
T
T
(γj )[j] = (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj .
Then we can rewrite b̃j as
(b̃j )j = −bj (γj )j ,
(b̃j )[j] = β̂[j] − bj (γj )[j] .
33
(B-29)
By definition of β̂[j] , we have [f (β̂[j] )][j] = 0 and hence
[f (b̃j )][j] = [f (b̃j )][j] − [f (β̂[j] )][j] =
n
i
h
1X
xi,[j] ψ(i − xTi b̃j ) − ψ(i − xTi,[j] β̂[j] ) .
n i=1
(B-30)
By mean value theorem, there exists νi,j ∈]i − xTi b̃j , i − xTi,[j] β̂[j] [ such that
ψ(i − xTi b̃j ) − ψ(i − xTi,[j] β̂[j] ) = ψ 0 (νi,j )(xTi,[j] β̂[j] − xTi b̃j )
=ψ 0 (νi,j )(xTi,[j] β̂[j] − xTi,[j] (b̃j )[j] − Xij bj )
i
h
T
T
=ψ 0 (νi,j ) · bj · xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
Let
di,j = ψ 0 (νi,j ) − ψ 0 (ri,[j] )
(B-31)
and plug the above result into (B-30), we obtain that
n
i
h
1X
T
T
xi,[j] · ψ 0 (ri,[j] ) + di,j · bj · xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
n i=1
[f (b̃j )][j] =
n
h
i
1X 0
T
T
= bj ·
ψ (ri,[j] )xi,[j] xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
n i=1
n
1X
T
T
di,j xi,[j] (xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij )
n i=1
i
1h T
T
T
T
X[j] D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − X[j]
D[j] Xj
= bj ·
n
n
1X
di,j xi,[j] · xTi γj
+ bj ·
n i=1
!
n
1 X
di,j xi,[j] xTi γj .
= bj ·
n i=1
+ bj ·
Now we calculate [f (b̃j )]j , the j-th entry of f (b̃j ). Note that
n
[f (b̃j )]j =
=
1X
Xij ψ(i − xTi b̃j )
n i=1
n
n
n
n
h
i
1X
1X
T
T
Xij ψ(ri,[j] ) + bj ·
Xij (ψ 0 (ri,[j] ) + di,j ) · xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
n i=1
n i=1
h
i
1X
1X 0
T
T
Xij ψ(ri,[j] ) + bj ·
ψ (ri,[j] )Xij xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
n i=1
n i=1
!
n
1X
T
+ bj ·
di,j Xij xi γj
n i=1
!
n
X
1
1 T
1
T
T
2
= √ Nj + bj ·
D[j] X[j] )−1 X[j]
D[j] Xj −
X D[j] X[j] (X[j]
ψ 0 (ri,[j] )Xij
n j
n i=1
n
!
n
1X
T
+ bj ·
di,j Xij xi γj
n i=1
!
n
1
1X
T
= √ Nj − bj · ξj + bj ·
di,j Xij xi γj
n i=1
n
!
n
1X
T
= bj ·
di,j Xij xi γj
n i=1
=
34
where the second last line uses the definition of bj . Putting the results together, we
obtain that
!
n
1X
T
f (b̃j ) = bj ·
di,j xi xi · γj .
n i=1
This entails that
kf (b̃j )k2 ≤ |bj | · max |di,j | · λ+ · kγj k2 .
(B-32)
i
Now we derive a bound for maxi |di,j |, where di,j is defined in (B-36). By Lemma
B.2,
|di,j | = |ψ 0 (νi,j ) − ψ 0 (ri,[j] )| ≤ K3 |νi,j − ri,[j] | = K3 |xTi,[j] β̂[j] − xTi b̃j |.
By definition of b̃j and hj,1,i ,
T
T
|xTi,[j] β̂[j] − xTi b̃j | = |bj | · xTi,[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj − Xij
T
T
=|bj | · |eTi (I − X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] )Xj |
=|bj | · |hTj,1,i Xj | ≤ |bj | · ∆C khj,1,i k2 ,
(B-33)
where the last inequality is derived by definition of ∆C , see (B-21). Since hj,1,i
T
T
is the i-th column of matrix I − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
, its L2 norm is upper
bounded by the operator norm of this matrix. Notice that
1
1
1
−1
T
T
T
T
2
2
2
I − D[j]
X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j]
D[j]2 .
I − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
= D[j]
The middle matrix in RHS of the displayed atom is an orthogonal projection matrix
and hence
1
1
K1 2
−1
T
T
2
kop · kD[j]2 kop ≤
kI − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
kop ≤ kD[j]
. (B-34)
K0
Therefore,
max khj,1,i k2 ≤ max kI −
i,j
j∈Jn
T
T
D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
kop
and thus
r
max |di,j | ≤ K3
i
≤
K1
K0
12
,
K1
· |bj | · ∆C .
K0
(B-35)
(B-36)
As for γj , we have
K0 λ− kγj k22
≤
γjT
X T D[j] X
n
XjT Dj Xj
=(γj )2j ·
+ (γj )T[j]
n
!
γj
T
X[j]
D[j] X[j]
n
!
(γj )[j] + 2γj
XjT D[j] X[j]
(γj )[j]
n
Recall the definition of γj in (B-37), we have
!
T
X[j]
D[j] X[j]
1
T
T
T
(γj )[j]
(γj )[j] = XjT D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj
n
n
and
γj
XjT D[j] X[j]
1
T
T
(γj )[j] = − XjT D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] Xj .
n
n
35
As a result,
K0 λ− kγj k22
1
1
1
1
1
T
T
2
2
2
2
D[j] X[j] )−1 X[j]
D[j]
≤ XjT D[j]
(I − D[j]
X[j] (X[j]
)D[j]
Xj
n
1
≤
2
Xj k22
kD[j]
n
1
1
2
≤
1
T
T
2
2
· I − D[j]
X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j]
kD[j] Xj k22
n
≤
op
K1 kXj k22
≤ T 2 K1 ,
n
where T is defined in (B-23). Therefore we have
s
K1
T.
kγj k2 ≤
K0 λ−
(B-37)
Putting (B-32), (B-36), (B-37) and part (ii) together, we obtain that
s
r
K1
K1
∆C |bj | ·
T
kf (b̃j )k2 ≤ λ+ · |bj | · K3
K0
K0 λ−
s
r
K1
K1
1 2K1
2
≤ λ+ ·
∆ E · K3
∆C ·
T
n (K0 λ− )2 C
K0
K0 λ−
=
1 2K12 K3 λ+ T
· ∆3C · E.
·
5
n
3
2
K0 λ−
By Lemma B.1,
kβ̂ − b̃j k2 ≤
kf (β̂) − f (b̃j )k2
kf (b̃j )k2
1 2K12 K3 λ+ T
· ∆3C · E.
=
≤ ·
7
K0 λ−
K0 λ−
n
4
2
K0 λ−
Since β̂j − bj is the j-th entry of β̂ − b̃j , we have
|β̂j − bj | ≤ kβ̂ − b̃j k2 ≤
1 2K12 K3 λ+ T
·
· ∆3C · E.
7
n
2
K04 λ−
(iv) Similar to part (iii), this result has been shown by El Karoui (2013). Here we state
a refined version for the sake of completeness. Let b̃j be defined as in (B-28), then
|Ri − ri,[j] | = |xTi β̂ − xTi,[j] β̂[j] | = |xTi (β̂ − b̃j ) + xTi b̃j − xTi,[j] β̂[j] |
≤ kxi k2 · kβ̂ − b̃j k2 + |xTi b̃j − xTi,[j] β̂[j] |.
Note that kxi k2 ≤
√
nT , by part (iii), we have
1 2K12 K3 λ+ T 2
· ∆3C · E.
kxi k2 · kβ̂ − b̃j k2 ≤ √
7
n
K 4λ 2
0
(B-38)
−
On the other hand, similar to (B-36), by (B-33),
r
√
√
K1
1
2K1
T
T
|xi b̃j − xi,[j] β̂[j] | ≤
· |bj | · ∆C ≤ √ · 3
· ∆2C · E.
K0
n K2λ
0 −
(B-39)
Therefore,
1
|Ri − ri,[j] | ≤ √
n
2K12 K3 λ+ T 2
7
2
K04 λ−
36
√
·
∆3C
·E +
2K1
3
K02 λ−
·
∆2C
·
√
!
E
.
B-3
Summary of Approximation Results
Under our technical assumptions, we can derive the rate for approximations via Proposition B.3. This justifies all approximations in Appendix A.
Theorem B.4. Under the assumptions A1 - A5,
(i)
T ≤ λ+ = O (polyLog(n)) ;
(ii)
max |β̂j | ≤ kβ̂k2 = OL4 (polyLog(n)) ;
j∈Jn
(iii)
max |bj | = OL2
j∈Jn
polyLog(n)
√
n
;
(iv)
max |β̂j − bj | = OL2
j∈Jn
polyLog(n)
n
;
(v)
max max |Ri − ri,[j] | = OL2
j∈Jn
i
polyLog(n)
√
n
.
Proof. (i) Notice that Xj = Xej , where ej is the j-th canonical basis vector in Rp , we
have
kXj k2
XT X
= eTj
ej ≤ λ+ .
n
n
Similarly, consider the X T instead of X, we conclude that
kxi k2
XX T
≤ λmax
= λ+ .
n
n
Recall the definition of T in (B-23), we conclude that
p
T ≤ λ+ = O (polyLog(n)) .
(ii) Since i = ui (Wi ) with ku0i k∞ ≤ c1 , the gaussian concentration property (Ledoux,
2001, chapter 1.3) implies that i is c21 -sub-gaussian and hence E|i |k = O(ck1 ) for
any finite k > 0. By Lemma B.2, |ψ(i )| ≤ K1 |i | and hence for any finite k,
E|ψ(i )|k ≤ K1k E|i |k = O(ck1 ).
By part (i) of Proposition B.3, using the convexity of x4 and hence
a+b 4
2
≤
a4 +b4
2 ,
1
8
E(U + U0 )4 ≤
(EU 4 + U04 ).
4
(K0 λ− )
(K0 λ− )4
Pn
Recall (B-24) that U = n1 i=1 xi (ψ(i ) − Eψ(i )) 2 ,
Ekβ̂k42 ≤
U 4 = (U 2 )2 =
1
n4
n
X
2
xTi xi0 (ψ(i ) − Eψ(i ))(ψ(i0 ) − Eψ(i0 ))
i,i0 =1
2
n
X
1 X
= 4
kxi k22 (ψ(i ) − Eψ(i ))2 +
|xTi xi0 |(ψ(i ) − Eψ(i ))(ψ(i0 ) − Eψ(i0 ))
n
0
i=1
i6=i
37
1
= 4
n
+
X
n
kxi k42 (ψ(i ) − Eψ(i ))4 +
(2|xTi xi0 |2 + kxi k22 kxi0 k22 )(ψ(i ) − Eψ(i ))2 (ψ(i0 ) − Eψ(i0 ))2
i6=i0
i=1
X
X
|xTi xi0 |
·
|xTk xk0 |
· (ψ(i ) − Eψ(i ))(ψ(i0 ) − Eψ(i0 ))(ψ(k ) − Eψ(k ))(ψ(k0 ) − Eψ(k0 ))
others
Since ψ(i ) − Eψ(i ) has a zero mean, we have
E(ψ(i ) − Eψ(i ))(ψ(i0 ) − Eψ(i0 ))(ψ(k ) − Eψ(k ))(ψ(k0 ) − Eψ(k0 )) = 0
for any (i, i0 ) 6= (k, k 0 ) or (k 0 , k) and i 6= i0 . As a consequence,
n
1 X
kxi k42 E(ψ(i ) − Eψ(i ))4
EU = 4
n
i=1
X
T
2
2
2
2
2
0
0
0
0
(2|xi xi |2 + kxi k2 kxi k2 )E(ψ(i ) − Eψ(i )) E(ψ(i ) − Eψ(i ))
+
4
i6=i0
n
X
X
1
kxi k42 E(ψ(i ) − Eψ(i ))4 + 3
kxi k22 kxi0 k22 E(ψ(i ) − Eψ(i ))2 E(ψ(i0 ) − Eψ(i0 ))2 .
≤ 4
n
0
i=1
i6=i
4
4
a +b
4
For any i, using the convexity of x4 , hence ( a+b
2 ) ≤
2 , we have
E(ψ(i ) − Eψ(i ))4 ≤ 8E ψ(i )4 + (Eψ(i ))4 ≤ 16Eψ(i )4 ≤ 16 max Eψ(i )4 .
i
By Cauchy-Schwartz inequality,
E(ψ(i ) − Eψ(i ))2 ≤ Eψ(i )2 ≤
q
p
Eψ(i )4 ≤ max Eψ(i )4 .
i
Recall (B-23) that kxi k22 ≤ nT 2 and thus,
1
16n · n2 T 4 + 3n2 · n2 T 4 · max Eψ(i )4
i
n4
1
≤ 4 · (16n3 + 3n4 )T 4 max Eψ(i )4 = O (polyLog(n)) .
i
n
EU 4 ≤
On the other hand, let µT = (Eψ(1 ), . . . , Eψ(n )), then kµk22 = O(n · polyLog(n))
and hence by definition of U0 in (B-24),
r
kµk22
kµT Xk2
1p T
T
=
µ XX µ ≤
· λ+ = O (polyLog(n)) .
U0 =
n
n
n
In summary,
Ekβ̂k42 = O (polyLog(n)) .
(iii) By mean-value theorem, there exists ax ∈ (0, x) such that
ρ(x) = ρ(0) + xψ(0) +
x2 0
ψ (ax ).
2
By assumption A1 and Lemma B.2, we have
ρ(x) =
x2 0
x2 0
K3 x2
ψ (ax ) ≤
kψ k∞ ≤
,
2
2
2
where K3 is defined in Lemma B.2. As a result,
Eρ(i )8 ≤
K3
2
38
8
16
E16
i = O(c1 ).
Recall the definition of E in (B-23) and the convexity of x8 , we have
n
EE 8 ≤
1X
Eρ(i )8 = O(c16
1 ) = O (polyLog(n)) .
n i=1
(B-40)
Under assumption A5, by Cauchy-Schwartz inequality,
q
√
√
E(∆C E)2 = E∆2C E ≤ E∆4C · EE 2 = O (polyLog(n)) .
Under assumptions A1 and A3,
√
2K1
= O (polyLog(n)) .
K0 λ−
Putting all the pieces together, we obtain that
polyLog(n)
√
max |bj | = OL2
.
j∈Jn
n
(iv) Similarly, by Holder’s inequality,
E(∆3C E)2 = E∆6C E 2 ≤ E∆8C
34
· EE 8
41
= O (polyLog(n)) ,
and under assumptions A1 and A3,
2K12 K3 λ+ T
7
= O (polyLog(n)) .
2
K04 λ−
Therefore,
max |β̂j − bj | = OL2
j∈Jn
polyLog(n)
n
.
(v) It follows from the previous part that
√
E(∆2C · E)2 = O (polyLog(n)) .
Under assumptions A1 and A3, the multiplicative factors are also O (polyLog(n)),
i.e.
√
2K1
2K12 K3 λ+ T 2
= O (polyLog(n)) ,
= O (polyLog(n)) .
3
7
4
2
K0 λ−
K02 λ−
Therefore,
max max |Ri − ri,[j] | = OL2
j∈Jn
B-4
i
polyLog(n)
√
n
.
Controlling Gradient and Hessian
Proof of Lemma 4.2. Recall that β̂ is the solution of the following equation
n
1X
xi ψ(i − xTi β̂) = 0.
n i=1
Taking derivative of (B-41), we have
!
∂ β̂
∂ β̂
T
X D I − X T = 0 =⇒ T = (X T DX)−1 X T D.
∂
∂
39
(B-41)
This establishes (9). To establishes (10), note that (9) can be rewritten as
(X T DX)
∂ β̂
= X T D.
∂T
(B-42)
Fix k ∈ {1, · · · , n}. Note that
∂Ri
∂i
∂ β̂
=
− xTi
= I(i = k) − xTi (X T DX)−1 X T D.
∂k
∂k
∂k
Recall that G = I − X(X T DX)−1 X T D, we have
∂Ri
= eTi Gek ,
∂k
(B-43)
where ei is the i-th canonical basis of Rn . As a result,
∂D
= D̃ diag(Gek ).
∂k
(B-44)
Taking derivative of (B-42), we have
XT
∂ β̂
∂D
∂D ∂ β̂
X T + (X T DX)
= XT
T
∂k ∂
∂k ∂
∂k
=⇒
∂ β̂
∂D
= (X T DX)−1 X T
I − X(X T DX)−1 X T D
T
∂k ∂
∂k
=⇒
∂ β̂
= (X T DX)−1 X T D̃ diag(Gek )G,
∂k ∂T
where G = I − X(X T DX)−1 X T D is defined in (B-18) in p.30. Then for each j ∈
{1, · · · , p} and k ∈ {1, . . . , n},
∂ β̂j
= eTj (X T DX)−1 X T D̃ diag(Gek )G = eTk GT diag(eTj (X T DX)−1 X T D̃)G
∂k ∂T
where we use the fact that aT diag(b) = bT diag(a) for any vectors a, b. This implies that
∂ β̂j
= GT diag(eTj (X T DX)−1 X T D̃)G
∂∂T
Proof of Lemma 4.3. Throughout the proof we are using the simple fact that kak∞ ≤
kak2 . Based on it, we found that
1
1
eTj (X T DX)−1 X T D 2
≤ eTj (X T DX)−1 X T D 2
∞
q
= eTj (X T DX)−1 X T DX(X T DX)−1 ej
q
1
= eTj (X T DX)−1 ej ≤
1 .
(nK0 λ− ) 2
1
Thus for any m > 1, recall that Mj = E eTj (X T DX)−1 X T D 2
1
E eTj (X T DX)−1 X T D 2
≤E eTj (X T DX)−1 X T D
≤
Mj
(nK0 λ− )
m−1
2
∞
2
(B-45)
,
m
∞
1
2
1
∞
· eTj (X T DX)−1 X T D 2
.
m−1
2
(B-46)
40
We should emphasize that we cannot use the naive bound that
1
E eTj (X T DX)−1 X T D 2
m
∞
1
=⇒ eTj (X T DX)−1 X T D 2
m
1
≤ E eTj (X T DX)−1 X T D 2
polyLog(n)
m
√
= OL
n
∞
1
m ,
(nK0 λ− ) 2
≤
2
(B-47)
since it fails to guarantee the convergence of TV distance. We will address this issue after
deriving Lemma 4.4.
By contrast, as proved below,
1
polyLog(n)
1
eTj (X T DX)−1 X T D 2
= Op (Mj ) = Op
<< p
.
(B-48)
n
∞
nK0 λ−
Thus (B-46) produces a slightly tighter bound
1
eTj (X T DX)−1 X T D 2
∞
= OLm
polyLog(n)
n
.
m+1
2m
It turns out that the above bound suffices to prove the convergence. Although (B-48)
m+1
implies the possibility to sharpen the bound from n− 2m to n−1 using refined analysis,
we do not explore this to avoid extra conditions and notation.
• Bound for κ0j
First we derive a bound for κ0j . By definition,
4
∂ β̂j
∂
β̂
j
≤ E
κ20j = E
T
∂
∂T
2
∞
4
∂ β̂j
·
∂T
2
.
2
By Lemma 4.2 and (B-46) with m = 2,
E
∂ β̂j
∂T
2
1
≤ E eTj (X T DX)−1 X T D 2
∞
2
∞
· K1 =
K1 Mj
1
(nK0 λ− ) 2
.
On the other hand, it follows from (B-45) that
∂ β̂j
∂T
2
= eTj (X T DX)−1 X T D
2
2
1
≤ K1 · eTj (X T DX)−1 X T D 2
2
2
≤
2
K1
. (B-49)
nK0 λ−
Putting the above two bounds together we have
κ20j ≤
K12
3
(nK0 λ− ) 2
· Mj .
(B-50)
• Bound for κ1j
As a by-product of (B-49), we obtain that
κ41j
∂ β̂j
=E
∂T
4
≤
2
K12
.
(nK0 λ− )2
(B-51)
• Bound for κ2j
Finally, we derive a bound for κ2j . By Lemma 4.2, κ2j involves the operator norm
of a symmetric matrix with form GT M G where M is a diagonal matrix. Then by the
triangle inequality,
GT M G
op
≤ kM kop · GT G
41
2
op
= kM kop · kGkop .
Note that
1
1
1
1
D 2 GD− 2 = I − D 2 X(X T DX)−1 X T D 2
is a projection matrix, which is idempotent. This implies that
1
1
1
1
D 2 GD− 2
= λmax D 2 GD− 2 ≤ 1.
op
Write G as D
− 12
1
2
(D GD
− 12
1
2
)D , then we have
kGkop ≤ D
1
2
− 12
· D GD
− 21
· D
op
r
1
2
≤
op
op
K1
.
K0
Returning to κ2j , we obtain that
4
κ42j = E GT diag(eTj (X T DX)−1 X T D̃)G
op
4
8
· kGkop
≤ E eTj (X T DX)−1 X T D̃
∞
≤E
=E
4
eTj (X T DX)−1 X T D̃
∞
1
1
K1
K0
eTj (X T DX)−1 X T D 2 D− 2 D̃
4
4
K1
·
K0
∞
4
Assumption A1 implies that
1
|ψ 00 (Ri )|
∀i, p
≤ K2 & hence kD− 2 D̃kop ≤ K2 .
0
ψ (Ri )
Therefore,
1
1
eTj (X T DX)−1 X T D 2 D− 2 D̃
4
1
≤ K24 · eTj (X T DX)−1 X T D 2
∞
4
∞
.
By (B-46) with m = 4,
κ42j
≤
K24
3
(nλ− ) 2
·
K1
K0
4
· Mj .
(B-52)
Proof of Lemma 4.4. By Theorem B.4, for any j,
Eβ̂j4 ≤ Ekβ̂k42 < ∞.
Then using the second-order Poincaré inequality (Proposition 4.1),
!
3
β̂
−
E
β̂
c
c
κ
+
c
κ
κ
j
j
1
2
0j
1j
2j
1
, N (0, 1) = O
max dT V L q
j∈Jn
Var(
β̂
j)
Var(β̂j )
1
1
M2
M4
j
j
n 34 + n 87
=O
· polyLog(n)
=O
Var(β̂j )
1
1
(nMj2 ) 4 + (nMj2 ) 8
n Var(β̂j )
!
· polyLog(n) .
It follows from (B-45) that nMj2 = O (polyLog(n)) and the above bound can be simplified
as
!
1
(nMj2 ) 8
β̂
−
E
β̂
j
j
max dT V L q
, N (0, 1) = O
· polyLog(n) .
j∈Jn
n Var(β̂j )
Var(β̂j )
42
Remark B.5. If we use the naive bound (B-47), by repeating the above derivation, we
√
), in which case,
obtain a worse bound for κ0,j = O( polyLog(n)
) and κ2 = O( polyLog(n)
n
n
max dT V
j∈Jn
β̂
−
E
β̂
j
j
L q
, N (0, 1) = O
Var(β̂j )
polyLog(n)
n Var(β̂j )
!
.
1
However, we can only prove that Var(β̂j ) = Ω( n1 ). Without the numerator (nMj2 ) 8 , which
1
will be shown to be O(n− 8 polyLog(n)) in the next subsection, the convergence cannot be
proved.
B-5
Upper Bound of Mj
As mentioned in Appendix A, we should approximate D by D[j] to remove the functional
(1)
dependence on Xj . To achieve this, we introduce two terms, Mj
(1)
Mj
1
(2)
2
k∞ ),
= E(keTj (X T DX)−1 X T D[j]
Mj
(1)
(2)
and Mj , defined as
1
2
= E(keTj (X T D[j] X)−1 X T D[j]
k∞ ).
(1)
(2)
We will first prove that both |Mj − Mj | and |Mj − Mj | are negligible and then derive
(2)
an upper bound for Mj .
B-5.1
(1)
Controlling |Mj − Mj |
By Lemma B.2,
1
1
2
kD 2 − D[j]
k∞ ≤ K2 max |Ri − ri,[j] | , K2 Rj ,
i
and by Theorem B.4,
q
polyLog(n)
2
√
ERj = O
.
n
(1)
Then we can bound |Mj − Mj | via the fact that kak∞ ≤ kak2 and algebra as follows.
(1)
1
1
1
1
2
|Mj − Mj | ≤ E(keTj (X T DX)−1 X T (D 2 − D[j]
)k∞ )
2
)k2 )
≤ E(keTj (X T DX)−1 X T (D 2 − D[j]
r
r
=
1
1
2
E(keTj (X T DX)−1 X T (D 2 − D[j]
)k22 )
≤
1
1
2
E(eTj (X T DX)−1 X T (D 2 − D[j]
)2 X(X T DX)−1 ej ).
By Lemma B.2,
q
p
| ψ 0 (Ri ) − ψ 0 (ri,[j] )| ≤ K2 |Ri − ri,[j] | ≤ K2 Rj ,
thus
1
1
2
(D 2 − D[j]
)2 K22 R2j I
K22 2
R D.
K0 j
This entails that
(1)
− 21
|Mj − Mj | ≤ K2 K0
q
E(R2j · eTj (X T DX)−1 X T DX(X T DX)−1 ej )
q
1
−2
E(R2j · eTj (X T DX)−1 ej )
q
K2
polyLog(n)
p
.
≤√
E(R2j ) = O
n
nK0 λ−
= K2 K0
43
(1)
Bound of |Mj
B-5.2
(2)
− Mj |
First we prove a useful lemma.
Lemma B.6. For any symmetric matrix N with kN kop < 1,
(I − (I + N )−1 )2
N2
.
(1 − kN kop )2
Proof. First, notice that
I − (I + N )−1 = (I + N − I)(I + N )−1 = N (I + N )−1 ,
and therefore
(I − (I + N )−1 )2 = N (I + N )−2 N.
Since kN kop < 1, I + N is positive semi-definite and
1
I.
(1 − kN kop )2
(I + N )−2
Therefore,
N (I + N )−2 N
(1)
N2
.
(1 − kN kop )2
(2)
We now back to bounding |Mj − Mj |. Let Aj = X T D[j] X, Bj = X T (D − D[j] )X.
By Lemma B.2,
kD − D[j] k∞ ≤ K3 max |Ri − ri,[j] | = K3 Rj
i
and hence
kBj kop ≤ K3 Rj · nλ+ I , nηj .
where ηj = K3 λ+ · Rj . Then by Theorem B.4.(v),
polyLog(n)
E(ηj2 ) = O
.
n
Using the fact that kak∞ ≤ kak2 , we obtain that
(1)
|Mj
1
(2)
1
T
T
−1 T
2
2
− Mj | ≤ E(keTj A−1
X D[j]
k∞ )
j X D[j] − ej (Aj + Bj )
r
1
≤
1
2
T
T
−1 X T D 2 k2 )
E(keTj A−1
j X D[j] − ej (Aj + Bj )
[j] 2
q
−1 )X T D X(A−1 − (A + B )−1 )e
E eTj (A−1
j
j
j
[j]
j − (Aj + Bj )
j
q
−1 )A (A−1 − (A + B )−1 )e
= E eTj (A−1
j
j
j
j
j − (Aj + Bj )
j
=
The inner matrix can be rewritten as
−1
−1
(A−1
)Aj (A−1
)
j − (Aj + Bj )
j − (Aj + Bj )
−1
−1
−1
−1
−1
−1
−1
−1
−1
−1
− 12
=Aj 2 (I − (I + Aj 2 Bj Aj 2 )−1 )Aj 2 Aj Aj 2 (I − (I + Aj 2 Bj Aj 2 )−1 )Aj
−1
=Aj 2 (I − (I + Aj 2 Bj Aj 2 )−1 )2 Aj 2 .
−1
(B-53)
−1
Let Nj = Aj 2 Bj Aj 2 , then
−1
−1
1
1
kNj kop ≤ kAj 2 kop · kBj kop · kAj 2 kop ≤ (nK0 λ− )− 2 · nηj · (nK0 λ− )− 2 =
44
ηj
.
K0 λ−
On the event {ηj ≤ 21 K0 λ− }, kNj kop ≤ 12 . By Lemma B.6,
(I − (I + Nj )−1 )2 4Nj2 .
This together with (B-53) entails that
−1
−1
−1
eTj (A−1
)Aj (Aj−1 − (Aj + Bj )−1 )ej = eTj Aj 2 (I − (I + Nj )−1 )2 Aj 2 ej
j − (Aj + Bj )
−1
−1
−1
−1
−1
−1
−1
≤4eTj Aj 2 Nj2 Aj 2 ej = eTj A−1
j Bj Aj Bj Aj ej ≤ kAj Bj Aj Bj Aj kop .
Since Aj nK0 λ− I, and kBj kop ≤ nηj , we have
−1
−1
−1 3
2
kA−1
j Bj Aj Bj Aj kop ≤ kAj kop · kBj kop ≤
1
1
· η2 .
·
n (K0 λ− )3 j
Thus,
K0 λ−
−1
−1
−1
E eTj (A−1
−
(A
+
B
)
)A
(A
−
(A
+
B
)
)e
·
I
η
≤
j
j
j
j
j
j
j
j
j
2
T −1
polyLog(n)
1
1
−1
2
·
Eη
=
O
.
≤E ej Aj Bj A−1
B
A
e
·
≤
j j
j
j
j
n (K0 λ− )3
n2
On the event {ηj > 21 K0 λ− }, since nK0 λ− I Aj nK1 λ+ I and Aj + Bj nK0 λ− I,
−1
−1
|eTj (A−1
)Aj (A−1
)ej |
j − (Aj + Bj )
j − (Aj + Bj )
−1 2
≤nK1 λ+ · |eTj (A−1
) ej |
j − (Aj + Bj )
T
−2
≤nK1 λ+ · 2|eTj A−2
ej |
j ej | + 2|ej (Aj + Bj )
≤
4nK1 λ+
1 4K1 λ+
= ·
.
2
(nK0 λ− )
n (K0 λ− )2
This together with Markov inequality implies htat
K0 λ−
−1
−1
−1
E eTj (A−1
−
(A
+
B
)
)A
(A
−
(A
+
B
)
)e
·
I
η
>
j
j
j
j
j
j
j
j
j
2
K0 λ−
1 4K1 λ+
· P ηj >
≤ ·
n (K0 λ− )2
2
1 4K1 λ+
4
≤ ·
·
· Eηj2
n (K0 λ− )2 (K0 λ− )2
polyLog(n)
=O
.
n2
Putting pieces together, we conclude that
q
(1)
(2)
−1 )A (A−1 − (A + B )−1 )e
|Mj − Mj | ≤ E eTj (A−1
j
j
j
j
j − (Aj + Bj )
j
s
K0 λ−
−1
−1
T
−1
−1
≤ E ej (Aj − (Aj + Bj ) )Aj (Aj − (Aj + Bj ) )ej · I ηj >
2
s
K0 λ−
−1 )A (A−1 − (A + B )−1 )e · I η ≤
+ E eTj (A−1
−
(A
+
B
)
j
j
j
j
j
j
j
j
j
2
polyLog(n)
.
=O
n
B-5.3
(2)
Bound of Mj
Similar to (A-1), by block matrix inversion formula (See Proposition E.1),
1
1
2
eTj (X T D[j] X)−1 X T D[j] =
45
2
XjT D[j]
(I − Hj )
1
1
2
2
XjT D[j]
(I − Hj )D[j]
Xj
,
1
1
T
T
2
2
where Hj = D[j]
X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j]
. Recall that ξj ≥ K0 λ− by (B-25), so we
have
1
1
2
2
(I − Hj )D[j]
Xj = nξj ≥ nλ− .
XjT D[j]
As for the numerator, recalling the definition of hj,1,i , we obtain that
1
1 T
T
2
Xj (I − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j] ) · D[j]
n
∞
p
1 T
−1
T
≤ K1 ·
X (I − D[j] X[j] (X[j] D[j] X[j] ) X[j] )
n j
∞
p
p
= K1 max hTj,1,i Xj ≤ K1 ∆C max khj,1,i k2 .
1
2
kXjT D[j]
(I − Hj )k∞ =
i
i
As proved in (B-35),
max khj,1,i k2 ≤
i
K1
K0
21
.
This entails that
1
K1
2
kXjT D[j]
(I − Hj )k∞ ≤ √
· ∆C = OL1 (polyLog(n)) .
K0
Putting the pieces together we conclude that
1
(2)
Mj
B-5.4
≤
2
EkXjT D[j]
(I − Hj )k∞
nλ−
=O
polyLog(n)
n
.
Summary
Based on results from Section B.5.1 - Section B.5.3, we have
polyLog(n)
Mj = O
.
n
Note that the bounds we obtained do not depend on j, so we conclude that
polyLog(n)
max Mj = O
.
j∈Jn
n
B-6
B-6.1
Lower Bound of Var(β̂j )
Approximating Var(β̂j ) by Var(bj )
By Theorem B.4,
max E(β̂j − bj )2 = O
j
polyLog(n)
n2
,
max Eb2j = O
j
polyLog(n)
n
.
Using the fact that
β̂j2 − b2j = (β̂j − bj + bj )2 − b2j = (β̂j − bj )2 + 2(β̂j − bj )bj ,
we can bound the difference between Eβ̂j2 and Eb2j by
Eβ̂j2 − Eb2j = E(β̂j − bj )2 + 2|E(β̂j − bj )bj | ≤ E(β̂j − bj )2 + 2
q
q
polyLog(n)
E(β̂j − bj )2 Eb2j = O
.
3
n2
Similarly, since |a2 − b2 | = |a − b| · |a + b| ≤ |a − b|(|a − b| + 2|b|),
polyLog(n)
|(Eβ̂j )2 − (Ebj )2 | ≤ E|β̂j − bj | · E|β̂j − bj | + 2E|bj | = O
.
3
n2
46
Putting the above two results together, we conclude that
polyLog(n)
Var(β̂j ) − Var(bj ) = O
.
3
n2
(B-54)
Then it is left to show that
Var(bj ) = Ω
B-6.2
1
n · polyLog(n)
.
Controlling Var(bj ) by Var(Nj )
Recall that
1 Nj
bj = √
n ξj
where
n
1 X
Nj = √
Xij ψ(ri,[j] ),
n i=1
ξj =
1 T
T
T
X (D[j] − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] )Xj .
n j
Then
n Var(bj ) = E
Nj
Nj
−E
ξj
ξj
2
=E
Nj − ENj
ENj
Nj
+
−E
ξj
ξj
ξj
2
.
Using the fact that (a + b)2 − ( 12 a2 − b2 ) = 12 (a + 2b)2 ≥ 0, we have
1
n Var(bj ) ≥ E
2
B-6.3
Nj − ENj
ξj
2
−E
ENj
Nj
−E
ξj
ξj
2
,
1
I1 − I2 .
2
(B-55)
Controlling I1
The Assumption A4 implies that
tr(Cov(hj,0 ))
1 T
Var(Nj ) = Xj Qj Xj = Ω
.
n
npolyLog(n)
1
It is left to show that tr(Cov(hj,0 ))/n = Ω polyLog(n)
. Since this result will also be used
later in Appendix C, we state it in the following the lemma.
Lemma B.7. Under assumptions A1 - A3,
2
n−p+1
tr(Cov(ψ(hj,0 )))
K4
1
≥ 02 ·
· min Var(i ) = Ω
.
i
n
K1
n
polyLog(n)
Proof. The (A-10) implies that
Var(ψ(ri,[j] )) ≥ K02 Var(ri,[j] ).
(B-56)
Note that ri,[j] is a function of , we can apply (A-10) again to obtain a lower bound for
Var(ri,[j] ). In fact, by variance decomposition formula, using the independence of 0i s,
Var(ri,[j] ) = E Var ri,[j] (i) + Var E ri,[j] (i) ≥ E Var ri,[j] (i) ,
where (i) includes all but the i-th entry of . Apply A-10 again,
∂ri,[j]
Var ri,[j] (i) ≥ inf
i
∂i
2
· Var(i ),
and hence
∂ri,[j]
Var(ri,[j] ) ≥ E Var ri,[j] (i) ≥ E inf
∂i
47
2
· Var(i ).
(B-57)
∂ri,[j]
∂i .
Now we compute
Similar to (B-43) in p.40, we have
∂rk,[j]
= eTi G[j] ek ,
∂i
(B-58)
where G[j] is defined in (B-18) in p.30. When k = i,
1
1
1
∂ri,[j]
−1
−1
−1
2
2
2
= eTi G[j] ei = eTi D[j]2 D[j]
G[j] D[j]2 D[j]
ei = eTi D[j]
G[j] D[j]2 ei .
∂i
(B-59)
By definition of G[j] ,
−1
1
1
1
T
T
2
2
2
D[j] X[j] )−1 X[j]
D[j]
G[j] D[j]2 = I − D[j]
X[j] (X[j]
.
D[j]
1
T
T
2
Let X̃[j] = D[j]
X[j] and Hj = X̃[j] (X̃[j]
X̃[j] )−1 X̃[j]
. Denote by X̃(i),[j] the matrix X̃[j]
after removing i-th row, then by block matrix inversion formula (See Proposition E.1),
T
X̃(i),[j] + x̃i,[j] x̃Ti,[j] )−1 x̃i,[j]
eTi Hj ei = x̃Ti,[j] (X̃(i),[j]
=
=
T
(X̃(i),[j]
X̃(i),[j] )−1
x̃Ti,[j]
−
T
T
X̃(i),[j] )−1 x̃i,[j] x̃Ti,[j] (X̃(i),[j]
X̃(i),[j] )−1
(X̃(i),[j]
T
1 + x̃Ti,[j] (X̃(i),[j]
X̃(i),[j] )−1 x̃i,[j]
T
x̃Ti,[j] (X̃(i),[j]
X̃(i),[j] )−1 x̃i,[j]
T
1 + x̃Ti,[j] (X̃(i),[j]
X̃(i),[j] )−1 x̃i,[j]
!
x̃i,[j]
.
This implies that
1
−1
2
eTi D[j]
G[j] D[j]2 ei = eTi (I − Hj )ei =
1
T
1 + x̃Ti,[j] (X̃(i),[j]
X̃(i),[j] )−1 x̃i,[j]
1
=
1
2
1
T
T D2 e
1 + eTi D[j] X[j] (X(i),[j]
D(i),[j] X(i),[j] )−1 X[j]
[j] i
≥
=
≥
≥
1
1
2
1
T
T D2 e
1 + K0−1 eTi D[j] X[j] (X(i),[j]
X(i),[j] )−1 X[j]
[j] i
1
T
T e
1 + K0−1 (D[j] )i,i · eTi X[j] (X(i),[j]
X(i),[j] )−1 X[j]
i
1
1+
T
T e
K0−1 K1 eTi X[j] (X(i),[j]
X(i),[j] )−1 X[j]
i
1
K0
·
.
T
T e
K1 1 + eTi X[j] (X(i),[j]
X(i),[j] )−1 X[j]
i
T
T
Apply the above argument to Hj = X[j] (X[j]
X[j] )−1 X[j]
, we have
1
T
T
= eTi (I − X[j] (X[j]
X[j] )−1 X[j]
)ei .
T (X T
−1 X e
1 + eTi X[j]
[j] i
(i),[j] X(i),[j] )
Thus, by (B-56) and (B-57),
Var(ψ(ri,[j] )) ≥
K04
T
T
· [eTi (I − X[j] (X[j]
X[j] )−1 X[j]
)ei ]2 .
K12
Summing i over 1, . . . , n, we obtain that
n
tr(Cov(hj,0 ))
K4 1 X T
T
T
≥ 02 ·
[e (I − X[j] (X[j]
X[j] )−1 X[j]
)ei ]2 · min Var(i )
i
n
K1 n i=1 i
48
(B-60)
2
1
T
−1 T
tr(I − X[j] (X[j] X[j] ) X[j] ) · min Var(i )
i
n
2
K4
n−p+1
= 02 ·
· min Var(i )
i
K1
n
1
by assumption A2, we conclude that
Since mini Var(i ) = Ω polyLog(n)
K4
≥ 02 ·
K1
tr(Cov(hj,0 ))
=Ω
n
1
polyLog(n)
.
In summary,
Var(Nj ) = Ω
1
polyLog(n)
.
Recall that
ξj =
1 T
1
T
T
Xj (D[j] − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] )Xj ≤ XjT D[j] Xj ≤ K1 T 2 ,
n
n
we conclude that
I1 ≥
B-6.4
Var(Nj )
=Ω
(K1 T 2 )2
1
polyLog(n)
.
(B-61)
Controlling I2
By definition,
2
1
1
1
Nj
I2 = E ENj
−E
+ ENj E − E
ξj
ξj
ξj
ξj
2
ENj
Nj
1
= Var
+ ENj E − E
ξj
ξj
ξj
2
1
1
2
= (ENj ) · Var
+ Cov Nj ,
ξj
ξj
1
1
2
≤ (ENj ) · Var
+ Var(Nj ) Var
ξj
ξj
1
= ENj2 · Var
.
ξj
(B-62)
By (B-27) in the proof of Theorem B.4,
ENj2 ≤ 2K1 E(E · ∆2C ) ≤ 2K1
q
EE 2 · E∆4C = O (polyLog(n)) ,
where the last equality uses the fact that E = OL2 (polyLog(n)) as proved in (B-40). On
the other hand, let ξ˜j be an independent copy of ξj , then
Var
1
ξj
1
= E
2
1
1
−
˜
ξj
ξj
!2
=
1 (ξj − ξ˜j )2
E
.
2
ξj2 ξ˜j2
Since ξj ≥ K0 λ− as shown in (B-25), we have
1
1
1
Var
≤
E(ξj − ξ˜j )2 =
· Var(ξj ).
ξj
2(K0 λ− )4
(K0 λ− )4
(B-63)
To bound Var(ξj ), we propose to using the standard Poincaré inequality (Chernoff, 1981),
which is stated as follows.
49
Proposition B.8. Let W = (W1 , . . . , Wn ) ∼ N (0, In×n ) and f be a twice differentiable
function, then
2
∂f (W )
.
Var(f (W )) ≤ E
∂W
2
In our case, i = ui (Wi ), and hence for any twice differentiable function g,
Var(g()) ≤ E
∂g()
∂W
2
∂g()
∂
·
∂
∂W T
=E
2
2
2
≤ max ku0i k∞ · E
i
2
∂g()
∂
2
.
2
Applying it to ξj , we have
Var(ξj ) ≤ c21 · E
∂ξj
∂
2
.
(B-64)
2
For given k ∈ {1, . . . , n}, using the chain rule and the fact that dB −1 = −B −1 dBB −1 for
any square matrix B, we obtain that
∂
T
T
D[j] − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j]
∂k
∂D[j]
∂D[j]
T
T
T
T ∂D[j]
=
−
X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
∂k
∂k
∂k
∂D
[j]
T
T
T
T
+ D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j]
∂k
∂D[j]
G[j]
=GT[j]
∂k
T
T
where G[j] = I − X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] as defined in last subsection. This implies
that
∂D[j]
∂ξj
1
= XjT GT[j]
G[j] Xj .
∂k
n
∂k
Then (B-64) entails that
2
n
1 X
T T ∂D[j]
E Xj G[j]
Var(ξj ) ≤ 2
G[j] Xj
n
∂k
(B-65)
k=1
∂D
. Similar to (B-44) in p.40 and recalling the definition of D[j] in
First we compute ∂[j]
k
(B-17) and that of G[j] in (B-18) in p.30, we have
∂D[j]
= D̃[j] diag(G[j] ek ) diag(D̃[j] G[j] ek ),
∂k
Let Xj = G[j] Xj and X̃j = Xj ◦ Xj where ◦ denotes Hadamard product. Then
XjT GT[j]
∂D[j]
∂D[j]
G[j] Xj = XjT
Xj = XjT diag(D̃[j] G[j] ek )Xj = X̃jT D̃[j] G[j] ek .
∂k
∂k
Here we use the fact that for any vectors x, a ∈ Rn ,
xT diag(a)x =
n
X
ai x2i = (x ◦ x)T a.
i=1
This together with (B-65) imply that
Var(ξj ) ≤
n
1 X
1
E(X̃jT D̃[j] G[j] ek )2 = 2 E X̃jT D̃[j] G[j]
n2
n
k=1
50
2
=
2
1
E(X̃jT D̃[j] G[j] GT[j] D̃[j] X̃j )
n2
Note that G[j] GT[j] kG[j] k2op I, and D̃[j] K3 I by Lemma B.2 in p.(B.2). Therefore we
obtain that
K2
1
2
2
2
X̃j ≤ 23 · E G[j] op · kX̃j k22
Var(ξj ) ≤ 2 E G[j] op · X̃jT D̃[j]
n
n
K2
K32
2
2
4
4
= 2 E G[j] op · kXj k4 ≤ 3 E G[j] op · kXj k∞
n
n
As shown in (B-34),
kG[j] kop ≤
K1
K0
12
.
On the other hand, notice that the i-th row of G[j] is hj,1,i (see (B-20) for definition), by
definition of ∆C we have
kXj k∞ = kG[j] Xj k∞ = max |hTj,1,i Xj | ≤ ∆C · max khj,1,i k2 .
i
By (B-35) and assumption A5,
kXj k∞ ≤ ∆C ·
K1
K0
21
This entails that
= OL4 (polyLog(n)) .
Var(ξj ) = O
polyLog(n)
n
.
Combining with (B-62) and (B-63), we obtain that
polyLog(n)
I2 = O
.
n
B-6.5
Summary
Putting (B-55), (B-61) and (B-62) together, we conclude that
1
1
polyLog(n)
1
−O
=Ω
=⇒ Var(bj ) = Ω
.
n Var(bj ) = Ω
polyLog(n)
n · polyLog(n)
polyLog(n)
n
Combining with (B-54),
Var(β̂j ) = Ω
C
polyLog(n)
n
.
Proof of Other Results
C-1
Proofs of Propositions in Section 2.3
Proof of Proposition 2.1. Let Hi (α) = Eρ(i − α). First we prove that the conditions
d
imply that 0 is the unique minimizer of Hi (α) for all i. In fact, since i = −i ,
Hi (α) = Eρ(i − α) =
1
(Eρ(i − α) + ρ(−i − α)) .
2
Using the fact that ρ is even, we have
Hi (α) = Eρ(i − α) =
1
(Eρ(i − α) + ρ(i + α)) .
2
By (4), for any α 6= 0, Hi (α) > Hi (0). As a result, 0 is the unique minimizer of Hi . Then
for any β ∈ Rp
n
n
n
n
1X
1X
1X
1X
Eρ(yi − xTi β) =
Eρ(i − xTi (β − β ∗ )) =
Hi (xTi (β − β ∗ )) ≥
Hi (0).
n i=1
n i=1
n i=1
n i=1
51
The equality holds iff xTi (β − β ∗ ) = 0 for all i since 0 is the unique minimizer of Hi . This
implies that
X(β ∗ (ρ) − β ∗ ) = 0.
Since X has full column rank, we conclude that
β ∗ (ρ) = β ∗ .
Proof of Proposition 2.2. For any α ∈ R and β ∈ Rp , let
n
1X
Eρ(yi − α − xTi β).
n i=1
G(α; β) =
Since αρ minimizes Eρ(i − α), it holds that
n
n
1X
1X
G(α; β) =
Eρ(i − α − xTi (β − β ∗ )) ≥
Eρ(i − αρ ) = G(αρ , β ∗ ).
n i=1
n i=1
Note that αρ is the unique minimizer of Eρ(i − α), the above equality holds if and only
if
α − αρ
T
∗
α + xi (β − β ) ≡ αρ =⇒ (1 X)
= 0.
β − β∗
Since (1 X) has full column rank, it must hold that α = αρ and β = β ∗ .
C-2
Proofs of Corollary 3.3
Proposition C.1. Suppose that i are i.i.d. such that Eρ(1 − α) as a function of α has
a unique minimizer αρ . Further assume that XJnc contains an intercept term, XJn has
full column rank and
span({Xj : j ∈ Jn }) ∩ span({Xj : j ∈ Jnc }) = {0}
Let
(C-66)
(
)
n
1X
T
Eρ(yi − xi β) .
βJn (ρ) = arg min min
c n
βJn
βJn
i=1
Then βJn (ρ) = βJ∗n .
Proof. let
n
G(β) =
1X
Eρ(yi − xTi β).
n i=1
For any minimizer β(ρ) of G, which might not be unique, we prove that βJn (ρ) = βJ∗n . It
follows by the same argument as in Proposition 2.2 that
xTi (β(ρ)−β ∗ ) ≡ α0 =⇒ X(β(ρ)−β ∗ ) = α0 1 =⇒ XJn (βJn (ρ)) = −XJnc (β(ρ)Jnc −βJ∗nc )+α0 1.
Since XJnc contains the intercept term, we have
XJn (βJn (ρ) − βJ∗n ) ∈ span({Xj : j ∈ Jnc }).
It then follows from (C-68) that
XJn (βJn (ρ) − βJ∗n ) = 0.
Since XJn has full column rank, we conclude that
βJn (ρ) = βJ∗n .
52
The Proposition C.1 implies that βJ∗n is identifiable even when X is not of full column
rank. A similar conclusion holds for the estimator β̂Jn and the residuals Ri . The following
two propositions show that under certain assumptions, β̂Jn and Ri are invariant to the
choice of β̂ in the presense of multiple minimizers.
Proposition C.2. Suppose that ρ is convex and twice differentiable with ρ00 (x) > c > 0
for all x ∈ R. Let β̂ be any minimizer, which might not be unique, of
n
F (β) ,
1X
ρ(yi − xTi β)
n i=1
Then Ri = yi − xi β̂ is independent of the choice of β̂ for any i.
Proof. The conclusion is obvious if F (β) has a unique minimizer. Otherwise, let β̂ (1) and
β̂ (2) be two different minimizers of F denote by η their difference, i.e. η = β̂ (2) − β̂ (1) .
Since F is convex, β̂ (1) + vη is a minimizer of F for all v ∈ [0, 1]. By Taylor expansion,
F (β̂ (1) + vη) = F (β̂ (1) ) + v∇F (β̂ (1) )η +
v2 T 2
η ∇ F (β̂ (1) )η + o(v 2 ).
2
Since both β̂ (1) + vη and β̂ (1) are minimizers of F , we have F (β̂ (1) + vη) = F (β̂ (1) ) and
∇F (β̂ (1) ) = 0. By letting v tend to 0, we conclude that
η T ∇2 F (β̂ (1) )η = 0.
The hessian of F can be written as
∇2 F (β̂ (1) ) =
1 T
cX T X
X diag(ρ00 (yi − xTi β̂ (1) ))X
.
n
n
Thus, η satisfies that
ηT
cX T X
η = 0 =⇒ Xη = 0.
n
(C-67)
This implies that
y − X β̂ (1) = y − X β̂ (2)
and hence Ri is the same for all i in both cases.
Proposition C.3. Suppose that ρ is convex and twice differentiable with ρ00 (x) > c > 0
for all x ∈ R. Further assume that XJn has full column rank and
span({Xj : j ∈ Jn }) ∩ span({Xj : j ∈ Jnc }) = {0}
(C-68)
Let β̂ be any minimizer, which might not be unique, of
n
F (β) ,
1X
ρ(yi − xTi β)
n i=1
Then β̂Jn is independent of the choice of β̂.
Proof. As in the proof of Proposition C.2, we conclude that for any minimizers β̂ (1) and
β̂ (2) , Xη = 0 where η = β̂ (2) − β̂ (1) . Decompose the term into two parts, we have
XJn ηJn = −XJcn ηJnc ∈ span({Xj : j ∈ Jnc }).
It then follows from (C-68) that XJn ηJn = 0. Since XJn has full column rank, we conclude
(1)
(2)
that ηJn = 0 and hence β̂Jn = β̂Jn .
53
Proof of Corollary 3.3. Under assumption A3*, XJn must have full column rank.
Otherwise there exists α ∈ R|Jn | such that XJn α, in which case αT XJTn (I −HJnc )XJn α = 0.
This violates the assumption that λ̃− > 0. On the other hand, it also guarantees that
span({Xj : j ∈ Jn }) ∩ span({Xj : j ∈ Jnc }) = {0}.
This together with assumption A1 and Proposition C.3 implies that β̂Jn is independent
of the choice of β̂.
c
c
c
Let B1 ∈ R|Jn |×|Jn | , B2 ∈ R|Jn |×|Jn | and assume that B2 is invertible. Let X̃ ∈ Rn×p
such that
X̃Jn = XJn − XJnc B1 , X̃Jnc = XJnc B2 .
Then rank(X) = rank(X̃) and model (1) can be rewritten as
y = X̃ β˜∗ +
where
β̃J∗n = βJ∗n ,
β̃J∗nc = B2−1 βJ∗nc + B1 βJ∗n .
˜
Let β̂ be an M-estimator, which might not be unique, based on X̃. Then Proposition C.3
˜
˜
shows that β̂Jn is independent of the choice of β̂, and an invariance argument shows that
˜
β̂Jn = β̂Jn .
In the rest of proof, we use ˜· to denote the quantity obtained based on X̃. First we show
that the assumption A4 is not affected by this transformation. In fact, for any j ∈ Jn ,
by definition we have
span(X̃[j] ) = span(X[j] )
and hence the leave-j-th-predictor-out residuals are not changed by Proposition C.2. This
implies that h˜j,0 = hj,0 and Q̃j = Qj . Recall the definition of hj,0 , the first-order condition
of β̂ entails that X T hj,0 = 0. In particular, XJTnc hj,0 = 0 and this implies that for any
α ∈ Rn ,
0 = Cov(XJTnc hj,0 , αT hj,0 ) = XJnc Qj α.
Thus,
X̃jT Q̃j X̃j
tr(Q̃j )
=
XjT Qj Xj
(Xj − XJcn (B1 )j )T Qj (Xj − XJnc (B1 )j )
=
.
tr(Qj )
tr(Qj )
Then we prove that the assumption A5 is also not affected by the transformation. The
above argument has shown that
h̃Tj,0 X̃j
kh̃j,0 k2
=
hTj,0 Xj
.
khj,0 k2
I|Jn | 0
On the other hand, let B =
, then B is non-singular and X̃ = XB. Let
−B1 B2
B(j),[j] denote the matrix B after removing j-th row and j-th column. Then B(j),[j] is
also non-singular and X̃[j] = X[j] B(j),[j] . Recall the definition of hj,1,i , we have
T
T
h̃j,1,i = (I − D̃[j] X̃[j] (X̃[j]
D̃[j] X̃j )−1 X̃[j]
)ei
T
T
T
= (I − D[j] X[j] B(j),[j] (B(j),[j]
X[j]
D[j] Xj B(j),[j] )−1 B(j),[j]
X[j] )ei
T
= (I − D[j] X[j] (X[j]
D[j] Xj )−1 X[j] )ei
= hj,1,i .
On the other hand, by definition,
T
T
T
T
X[j]
hj,1,i = X[j]
(I − D[j] X[j] (X[j]
D[j] X[j] )−1 X[j]
)ei = 0.
54
Thus,
hTj,1,i X̃j = hTj,1,i (Xj − XJcn (B1 )j ) = hTj,1,i Xj .
In summary, for any j ∈ Jn and i ≤ n,
h̃Tj,1,i X̃j
kh̃j,1,i k2
=
hTj,1,i Xj
.
khj,1,i k2
Putting the pieces together we have
˜ C = ∆C .
∆
By Theorem 3.1,
β̂
−
E
β̂
j
j
, N (0, 1) = o(1).
max dTV L q
j∈Jn
Var(β̂j )
provided that X̃ satisfies the assumption A3.
Now let U ΛV be the singular value decomposition of XJnc , where U ∈ Rn×p , Λ ∈ Rp×p , V ∈
Rp×p with U T U = V T V = Ip and Λ = diag(ν1 , . . . , νp ) being the diagonal matrix formed
by singular values of XJnc . First we consider the case where XJnc has full column rank,
p
then νj > 0 for all j ≤ p. Let B1 = (XJTn XJn )− XJTn XJn and B2 = n/|Jnc |V T Λ−1 . Then
X̃ T X̃
1
=
n
n
XJTn (I − XJnc (XJTnc XJnc )−1 XJnc )XJn
0
0
nI
.
This implies that
λmax
X̃ T X̃
n
!
n
o
= max λ̃max , 1 ,
The assumption A3* implies that
!
X̃ T X̃
λmax
= O(polyLog(n)),
n
λmin
λmin
X̃ T X̃
n
X̃ T X̃
n
!
n
o
= min λ̃min , 1 .
!
=Ω
1
polyLog(n)
.
By Theorem 3.1, we conclude that
Next we consider the case where XJcn does not have full column rank. We first remove
the redundant columns from XJcn , i.e. replace XJnc by the matrix formed by its maximum
linear independent subset. Denote by X this matrix. Then span(X) = span(X) and
span({Xj : j 6∈ Jn }) = span({Xj : j 6∈ Jn }). As a consequence of Proposition C.1 and
C.3, neither βJ∗n nor β̂Jn is affected. Thus, the same reasoning as above applies to this
case.
C-3
Proofs of Results in Section 3.3
First we prove two lemmas regarding the behavior of Qj . These lemmas are needed for
justifying Assumption A4 in the examples.
Lemma C.4. Under assumptions A1 and A2,
kQj kop ≤ c21
K32 K1
,
K0
kQj kF ≤
where Qj = Cov(hj,0 ) as defined in section B-1.
55
√
nc21
K32 K1
K0
Proof of Lemma C.4. By definition,
||Qj ||op = sup αT Qj α
α∈Sn−1
where Sn−1 is the n-dimensional unit sphere. For given α ∈ Sn−1 ,
αT Qj α = αT Cov(hj,0 )α = Var(αT hj,0 )
It has been shown in (B-59) in Appendix B-6.3 that
∂ri,[j]
= eTi G[j] ek ,
∂k
T
T
where G[j] = I − X[j] (X[j]
D[j] X[j] )−1 X[j]
D[j] . This yields that
!
n
n
n
X
X
∂ri,[j]
∂ X
αi ψ(ri,[j] ) =
=
αi ψ 0 (ri,[j] ) ·
αi ψ 0 (ri,[j] ) · eTi G[j] = αT D̃[j] G[j] .
∂ i=1
∂
i=1
i=1
By standard Poincaré inequality (see Proposition B.8), since i = ui (Wi ),
!
!
n
n
2
X
∂ X
0 2
αi ψ(ri,[j] )
αi ψ(ri,[j] ) ≤ max ||uk ||∞ · E
Var
k
∂ i=1
i=1
≤c21 · E αT D̃[j] G[j] GT[j] D̃[j] α ≤ c21 EkD̃[j] G[j] GT[j] D̃[j] k22 ≤ c21 EkD̃j k2op kG[j] k2op .
We conclude from Lemma B.2 and (B-34) in Appendix B-2 that
kG[j] k2op ≤
kD̃[j] kop ≤ K3 ,
K1
.
K0
Therefore,
||Qj ||op = sup Var
α∈Sn−1
and hence
||Qj ||F ≤
√
n
X
!
αi ψ(Ri )
≤ c21
i=1
n||Qj ||op ≤
√
n · c21
K32 K1
K0
K32 K1
.
K0
Lemma C.5. Under assumptions A1 - A3,
tr(Qj ) ≥ K ∗ n = Ω(n · polyLog(n)),
where K ∗ =
K04
K12
·
n−p+1 2
n
· mini Var(i ).
Proof. This is a direct consequence of Lemma B.7 in p.47.
Throughout the following proofs, we will use several results from the random matrix
theory to bound the largest and smallest singular values of Z. The results are shown in
Appendix E. Furthermore, in contrast to other sections, the notation P (·), E(·), Var(·)
denotes the probability, the expectation and the variance with respect to both and Z in
this section.
Proof of Proposition 3.4. By Proposition E.3,
√
√
λ+ = (1 + κ)2 + op (1) = Op (1), λ− = (1 − κ)2 − op (1) = Ωp (1)
and thus the assumption A3 holds with high probability. By Hanson-Wright inequality
(Hanson & Wright, 1971; Rudelson & Vershynin, 2013; see Proposition E.2), for any given
deterministic matrix A,
t2
t
P (|ZjT AZj − EZjT AZj | ≥ t) ≤ 2 exp −c min
,
σ 4 kAk2F σ 2 kAkop
56
for some universal constant c. Let A = Qj and conditioning on Z[j] , then by Lemma C.4,
we know that
√
K 2 K1
K 2 K1
, kQj kF ≤ nc21 3
kQj kop ≤ c21 3
K0
K0
and hence
t2
t
T
T
P Zj Qj Zj − E(Zj Qj Zj Z[j] ) ≤ −t Z[j] ≤ 2 exp −c min
,
.
σ 4 · nc41 K34 K12 /K02 σ 2 c21 K32 K1 /K0
(C-69)
Note that
2
E(ZjT Qj Zj Z[j] ) = tr(E[Zj ZjT |Z[j] ]Qj ) = EZ1j
tr(Qj ) = τ 2 tr(Qj ).
By Lemma C.5, we conclude that
!
!
T
Q
Z
Z
ZjT Qj Zj
t
t
j
j
j
Z[j] ≤ P
P
Z[j]
≤ τ2 −
≤ τ2 −
tr(Qj )
nK ∗
tr(Qj )
tr(Qj )
t2
t
≤2 exp −c min
,
.
(C-70)
σ 4 · nc41 K34 K12 /K02 2σ 2 c21 K32 K1 /K0
Let t = 21 τ 2 nK ∗ and take expectation of both sides over Z[j] , we obtain that
!
ZjT Qj Zj
τ2
K ∗2 τ 4
K ∗τ 2
P
≤
≤ 2 exp −cn min
,
tr(Qj )
2
4σ 4 c41 K34 K12 /K02 2σ 2 c21 K32 K1 /K0
and hence
P
ZjT Qj Zj
τ2
≤
min
j∈Jn tr(Qj )
2
!
≤ 2n exp −cn min
K ∗2 τ 4
K ∗τ 2
,
4σ 4 c41 K34 K12 /K02 2σ 2 c21 K32 K1 /K0
= o(1).
(C-71)
This entails that
min
j∈Jn
ZjT Qj Zj
= Ωp (polyLog(n)).
tr(Qj )
Thus, assumption A4 is also satisfied with high probability. On the other hand, since Zj
has i.i.d. mean-zero σ 2 -sub-gaussian entries, for any deterministic unit vector α ∈ Rn ,
αT Zj is σ 2 -sub-gaussian and mean-zero, and hence
t2
P (|αT Zj | ≥ t) ≤ 2e− 2σ2 .
Let αj,i = hj,1,i /khj,1,i k2 and αj,0 = hj,0 /khj,0 k2 . Since hj,1,i and hj,0 are independent
of Zj , a union bound then gives
p
t2 +4σ 2 log n
t2
2σ 2
P ∆C ≥ t + 2σ log n ≤ 2n2 e−
= 2e− 2σ2 .
By Fubini’s formula (Durrett, 2010, Lemma 2.2.8.),
E∆8C
Z
∞
Z
7
√
2σ log n
7
8t P (∆C ≥ t)dt ≤
=
0
8t dt +
0
Z
∞
√
2σ log n
8t7 P (∆C ≥ t)dt
Z ∞
p
p
p
log n)8 +
8(t + 2σ log n)7 P (∆C ≥ t + 2σ log n)dt
Z0 ∞
p
t2
7
≤ (2σ log n)8 +
64(8t7 + 128σ 7 (log n) 2 ) · 2e− 2σ2 dt
= (2σ
0
= O(σ 8 · polyLog(n)) = O (polyLog(n)) .
(C-72)
This, together with Markov inequality, guarantees that assumption A5 is also satisfied
with high probability.
57
Proof of Proposition 3.5. It is left to prove that assumption A3 holds with high probability. The proof of assumption A4 and A5 is exactly the same as the proof of Proposition
3.5. By Proposition E.4,
λ+ = Op (1).
On the other hand, by Proposition E.7 (Litvak et al., 2005),
T
Z Z
< c1 ≤ e−c2 n .
P λmin
n
and thus
λ− = Ωp (1).
Proof of Proposition 3.6. Since Jn excludes the intercept term, the proof of assumption A4 and A5 is still the same as Proposition 3.5. It is left to prove assumption A3. Let
R1 , . . . , Rn be i.i.d. Rademacher random variables, i.e. P (Ri = 1) = P (Ri = −1) = 12 ,
and
Z ∗ = diag(B1 , . . . , Bn )Z.
Then (Z ∗ )T Z ∗ = Z T Z. It is left to show that the assumption A3 holds for Z ∗ with high
probability. Note that
(Zi∗ )T = (Bi , Bi x̃Ti ).
For any r ∈ {1, −1} and borel sets B1 , . . . , Bp ⊂ R,
P (Bi = r, Bi Z̃i1 ∈ B1 , . . . , Bi Z̃i(p−1) ∈ Bp−1 )
= P (Bi = r, Z̃i1 ∈ rB1 , . . . , Z̃i(p−1) ∈ rBp−1 )
= P (Bi = r)P (Z̃i1 ∈ rB1 ) . . . P (Z̃i(p−1) ∈ rBp−1 )
= P (Bi = r)P (Z̃i1 ∈ B1 ) . . . P (Z̃i(p−1) ∈ Bp−1 )
= P (Bi = r)P (Bi Z̃i1 ∈ B1 ) . . . P (Bi Z̃i(p−1) ∈ Bp−1 )
where the last two lines uses the symmetry of Z̃ij . Then we conclude that Zi∗ has independent entries. Since the rows of Z ∗ are independent, Z ∗ has independent entries. Since Bi
d
are symmetric and sub-gaussian with unit variance and Bi Z̃ij = Z̃ij , which is also symmetric and sub-gaussian with variance bounded from below, Z ∗ satisfies the conditions
of Propsition 3.5 and hence the assumption A3 is satisfied with high probability.
Proof of Proposition 3.8 (with Proposition 3.7 being a special case). Let Z∗ =
1
1
Λ− 2 ZΣ− 2 , then Z∗ has i.i.d. standard gaussian entries. By Proposition 3.6, Z∗ satisfies
assumption A3 with high probability. Thus,
!
T
1
1
Σ 2 Z∗T ΛZ∗ Σ 2
Z∗ Z∗
λ+ = λmax
≤ λmax (Σ) · λmax (Λ) · λmax
= Op (polyLog(n)),
n
n
and
1
λ− = λmin
1
Σ 2 Z∗T ΛZ∗ Σ 2
n
!
≥ λmin (Σ) · λmin (Λ) · λmin
Z∗T Z∗
n
= Ωp
1
polyLog(n)
1
.
As for assumption A4, the first step is to calculate E(ZjT Qj Zj |Z[j] ). Let Z̃ = Λ− 2 Z,
then vec(Z̃) ∼ N (0, I ⊗ Σ). As a consequence,
Z̃j |Z̃[j] ∼ N (µ̃j , σj2 I)
where
1
−2
µ̃j = Z̃[j] Σ−1
Z[j] Σ−1
[j],[j] Σ[j],j = Λ
[j],[j] Σ[j],j .
58
Thus,
Zj |Z[j] ∼ N (µj , σj2 Λ)
where µj = Z[j] Σ−1
[j],[j] Σ[j],j . It is easy to see that
λ− ≤ min σj2 ≤ max σj2 ≤ λ+ .
j
(C-73)
j
It has been shown that Qj µj = 0 and hence
ZjT Qj Zj = (Zj − µj )T Qj (Zj − µj ).
1
1
1
Let Zj = Λ− 2 (Zj − µj ) and Q̃j = Λ 2 Qj Λ 2 , then Zj ∼ N (0, σj2 I) and
ZjT Qj Zj = ZjT Q̃j Zj .
By Lemma C.4,
kQ̃j kop ≤ kΛkop · kQj kop ≤ λmax (Λ) · c21
and hence
kQ̃j kF ≤
√
nλmax (Λ) · c21
K32 K1
,
K0
K32 K1
.
K0
By Hanson-Wright inequality (Hanson & Wright, 1971; Rudelson & Vershynin, 2013; see
Proposition E.2), we obtain a similar inequality to (C-69) as follows:
P |ZjT Qj Zj − E(ZjT Qj Zj Z[j] )| ≥ t Z[j]
"
(
)#
t2
t
≤2 exp −c min
.
,
σj4 · nλmax (Λ)2 c41 K34 K12 /K02 σj2 λmax (Λ)c21 K32 K1 /K0
On the other hand,
E(ZjT Qj Zj |Z[j] ) = E(ZjT Q̃j Zj |Z[j] ) = σj2 tr(Q̃j ).
By definition,
1
1
1
1
tr(Q̃j ) = tr(Λ 2 Qj Λ 2 ) = tr(ΣQj ) = tr(Qj2 ΛQj2 ) ≥ λmin (Λ) tr(Qj ).
By Lemma C.5,
tr(Q̃j ) ≥ λmin (Λ) · nK ∗ .
Similar to (C-70), we obtain that
!
ZjT Qj Zj
t
2
≥ σj −
Z[j]
P
tr(Qj )
nK ∗
"
(
)#
t2
t
≤2 exp −c min
,
.
σj4 · nλmax (Λ)2 c41 K34 K12 /K02 σj2 λmax (Λ)c21 K32 K1 /K0
Let t = 12 σj2 nK ∗ , we have
P
ZjT Qj Zj
σj2
≥
tr(Qj )
2
!
≤ 2 exp −cn min
K ∗2
K∗
,
4λmax (Λ)2 c41 K34 K12 /K02 2λmax (Λ)c21 K32 K1 /K0
and a union bound together with (C-73) yields that
min
j∈Jn
ZjT Qj Zj
1
1
= Ωp min σj2 ·
= Ωp
.
j
tr(Qj )
polyLog(n)
polyLog(n)
59
=o
1
n
As for assumption A5, let
1
αj,0 =
1
Λ 2 hj,0
,
khj,0 k2
Λ 2 hj,1,i
khj,1,i k2
αj,i =
then for i = 0, 1, . . . , p,
p
λmax (Λ).
kαj,i k2 ≤
Note that
hTj,0 Zj
T
= αj,0
Zj ,
khj,0 k2
hTj,1,i Zj
T
= αj,i
Zj
khj,1,i k2
using the same argument as in (C-72), we obtain that
E∆8C = O λmax (Λ)4 · max σj8 · polyLog(n) = O (polyLog(n)) ,
j
and by Markov inequality and (C-73),
E(∆8C |Z) = Op E∆8C = Op (polyLog(n)).
Proof of Proposition 3.9. The proof that assumptions A4 and A5 hold with high
probability is exactly the same as the proof of Proposition 3.8. It is left to prove assump1
tion A3*; see Corollary 3.3. Let c = (mini |(Λ− 2 1)i |)−1 and Z = (c1 Z̃). Recall the the
definition of λ̃+ and λ̃− , we have
λ̃+ = λmax (Σ{1} ),
λ̃− = λmin (Σ{1} ),
where
Σ{1} =
1 T
Z̃
n
I−
11T
n
Z̃.
Rewrite Σ{1} as
Σ{1}
1
=
n
11T
I−
n
It is obvious that
span
T
11T
Z̃
I−
Z̃ .
n
11T
I−
n
Z̃ ⊂ span(Z).
As a consequence
λ̃+ ≤ λmax
ZT Z
n
λ̃− ≥ λmin
,
It remains to prove that
T
Z Z
λmax
= Op (polyLog(n)) ,
n
λmin
ZT Z
n
ZT Z
n
.
= Ωp
1
polyLog(n)
.
To prove this, we let
1
Z∗ = Λ− 2 Z
1
1
1
0
0
Σ
1
− 21
, (ν Z̃∗ ),
where ν = cΛ− 2 1 and Z̃∗ = Λ− 2 Z̃Σ− 2 . Then
!
T
T
1
1
Z Z
Σ 2 Z∗T ΛZ∗ Σ 2
Z∗ Z∗
λmax
= λmax
≤ λmax (Σ) · λmax (Λ) · λmax
,
n
n
n
60
and
λmin
ZT Z
n
1
1
= λmin
Σ 2 Z∗T ΛZ∗ Σ 2
n
!
≥ λmin (Σ) · λmin (Λ) · λmin
It is left to show that
T
Z∗ Z∗
= Op (polyLog(n)),
λmax
n
λmin
Z∗T Z∗
n
= Ωp
Z∗T Z∗
n
1
polyLog(n)
.
.
By definition, mini |νi | = 1 and maxi |νi | = O (polyLog(n)), then
!
!
T
Z∗ Z∗
kνk22
νν T
Z̃∗T Z̃∗
Z̃∗T Z̃∗
λmax
≤ λmax
+
= λmax
+
.
n
n
n
n
n
Since Z̃∗ has i.i.d. standard gaussian entries, by Proposition E.3,
!
Z̃∗T Z̃∗
= Op (1).
λmax
n
Moreover, kνk22 ≤ n maxi |νi |2 = O(n · polyLog(n)) and thus,
T
Z∗ Z∗
= Op (polyLog(n)).
λmax
n
On the other hand, similar to Proposition 3.6,
Z∗ = diag(B1 , . . . , Bn )Z∗
where B1 , . . . , Bn are i.i.d. Rademacher random variables. The same argument in the
proof of Proposition 3.6 implies that Z∗ has independent entries with sub-gaussian norm
bounded by kνk2∞ ∨ 1 and variance lower bounded by 1. By Proposition E.7, Z∗ satisfies
assumption A3 with high probability. Therefore, A3* holds with high probability.
Proof of Proposition 3.10. Let Λ = (λ1 , . . . , λn ) and Z be the matrix with entries
Zij , then by Proposition 3.4 or Proposition 3.5, Zij satisfies assumption A3 with high
probability. Notice that
T 2
T
Z Λ Z
Z Z
2
λ+ = λmax
≤ λmax (Λ) · λmax
= Op (polyLog(n)),
n
n
and
λ− = λmin
Z T Λ2 Z
n
2
≥ λmin (Λ) · λmin
ZT Z
n
= Ωp
1
polyLog(n)
.
Thus Z satisfies assumption A3 with high probability.
Conditioning on any realization of Λ, the law of Zij does not change due to the independence between Λ and Z. Repeating the arguments in the proof of Proposition 3.4 and
Proposition 3.5, we can show that
ZjT Q̃j Zj
tr(Q̃j )
= Ωp
1
polyLog(n)
,
and E
max
i=0,...,n;j=1,...,p
T
|α̃j,i
Zj |8 = Op (polyLog(n)),
(C-74)
where
Q̃j = ΛQj Λ,
α̃j,0 =
Λhj,0
,
kΛhj,0 k2
61
α̃j,1,i =
Λhj,1,i
.
kΛhj,1,i k2
Then
ZjT Qj Zj
ZjT Q̃j Zj tr(ΛQj Λ)
ZjT Q̃j Zj
·
= Ωp
=
≥ a2 ·
tr(Qj )
tr(Qj )
tr(Q̃j )
tr(Q̃j )
1
polyLog(n)
,
(C-75)
and
"
E∆8C
=E
max
i=0,...,n;j=1,...,p
≤ b8 E
max
kΛhj,1,i k2
kΛhj,0 k2
, max
· max max
j
khj,0 k2 i,j khj,1,i k2
T
|α̃j,i
Zj |8
T
|α̃j,i
Zj |8
i=0,...,n;j=1,...,p
8 #
(C-76)
= Op (polyLog(n)).
By Markov inequality, the assumption A5 is satisfied with high probability.
Proof of Proposition 3.11. The concentration inequality of ζi plus a union bound
imply that
2
2
P max ζi > (log n) α ≤ nc1 e−c2 (log n) = o(1).
i
Thus, with high probability,
T 2
T
4
Z Λ Z
Z Z
λmax = λmax
≤ (log n) α · λmax
= Op (polyLog(n)).
n
n
Let n0 = b(1 − δ)nc for some δ ∈ (0, 1 − κ). Then for any subset I of {1, . . . , n} with
size n0 , by Proposition E.6 (Proposition E.7), under the conditions of Proposition 3.4
(Proposition 3.5), there exists constants c3 and c4 , which only depend on κ, such that
T
ZI ZI
< c3 ≤ e−c4 n
P λmin
n
where ZI represents the sub-matrix of Z formed by {Zi : i ∈ I}, where Zi is the i-th row
of Z. Then by a union bound,
T
ZI ZI
n
P min0 λmin
< c3 ≤
e−c4 n .
n0
n
|I|=n
By Stirling’s formula, there exists a constant c5 > 0 such that
n
o
n!
n
≤
c
exp
(−
δ̃
log
δ̃
−
(1
−
δ̃)
log(1
−
δ̃))n
= 0
5
0
n
n !(n − n0 )!
where δ̃ = n0 /n. For sufficiently small δ and sufficiently large n,
−δ̃ log δ̃ − (1 − δ̃) log(1 − δ̃) < c4
and hence
P
min0 λmin
|I|=n
ZIT ZI
n
< c3
< c5 e−c6 n
for some c6 > 0. By Borel-Cantelli Lemma,
lim inf
min
n→∞ |I|=b(1−δ)nc
λmin
ZIT ZI
n
≥ c3
On the other hand, since F −1 is continuous at δ, then
a.s.
ζ(b(1−δ)nc) → F −1 (δ) > 0.
62
a.s..
(C-77)
where ζ(k) is the k-th largest of {ζi : i = 1, . . . , n}. Let I ∗ be the set of indices corresponding to the largest b(1 − δ)nc ζi0 s. Then with probability 1,
T 2
T 2
T
Z Λ Z
ZI ∗ Λ I ∗ ZI ∗
Z Z
= lim inf λmin
≥ lim inf ζ(b(1−δ)nc) · lim inf λmin
lim inf λmin
n→∞
n→∞
n→∞
n→∞
n
n
n
T
ZI ZI
≥ lim inf ζ(b(1−δ)nc) · lim inf
min
λmin
n→∞
n→∞ |I|=b(1−δ)nc
n
≥ c3 F −1 (δ)2 > 0.
To prove assumption A4, similar to (C-75) in the proof of Proposition 3.10, it is left to
show that
tr(ΛQj Λ)
1
= Ωp
.
min
j
tr(Qj )
polyLog(n)
Furthermore, by Lemma C.5, it remains to prove that
n
min tr(ΛQj Λ) = Ωp
.
j
polyLog(n)
Recalling the equation (B-60) in the proof of Lemma B.7, we have
eTi Qj ei ≥
1
K0
.
·
T
T
T
K1 1 + ei Z[j] (Z(i),[j] Z(i),[j] )−1 Z[j] ei
(C-78)
By Proposition E.5,
v
u
u
P tλmax
ZjT Zj
n
!
> 3C1 ≤ 2e−C2 n .
On the other hand, apply (C-77) to Z(i),[j] , we have
P
min
|I|=b(1−δ)nc
(Z(i),[j] )TI (Z(i),[j] )I
n
λmin
!
!
< c3
< c5 e−c6 n .
A union bound indicates that with probability (c5 np + 2p)e− min{C2 ,c6 }n = o(1),
!
!
T
Z[j]
Z[j]
(Z(i),[j] )TI (Z(i),[j] )I
2
≤ 9C1 , min
min
λmin
≥ c3 .
max λmax
j
i,j |I|=b(1−δ)nc
n
n
This implies that for any j,
λmax
T
Z[j]
Z[j]
!
n
T 2
Z[j]
Λ Z[j]
= λmax
!
n
2
≤ ζ(1)
· 9C12
and for any i and j,
λmin
T
Z(i),[j]
Z(i),[j]
n
!
= λmin
T
2
ζ(i)
Z(i),[j]
Z(i),[j]
!
n
2
≥ min{ζ(b(1−δ)nc) , ζ(b(1−δ)nc) + 1} ·
min
|I|=b(1−δ)nc
λmin
2
(Z(i),[j] )TI ζ(i)
(Z(i),[j] )I
n
2
≥c3 min{ζ(b(1−δ)nc) , ζ(b(1−δ)nc) + 1} > 0.
Moreover, as discussed above,
2
ζ(1) ≤ (log n) α , min{ζ(b(1−δ)nc) , ζ(b(1−δ)nc) + 1} → F −1 (δ)
63
!
almost surely. Thus, it follows from (C-78) that with high probability,
eTi Qj ei ≥
K0
1
·
T (Z T
−1 Z e
K1 1 + eTi Z[j]
[j] i
(i),[j] Z(i),[j] )
K0
1
·
T Z
[j]
K1 1 + eT Z[j]
ei · c3 (F −1 (δ))2
i
n
K0
1
.
≥
·
K1 1 + (log n) α4 · 9C12 · c3 (F −1 (δ))2
≥
The above bound holds for all diagonal elements of Qj uniformly with high probability.
Therefore,
K0
n
1
2
tr(ΛQj Λ) ≥ ζ(b(1−δ)nc)
·b(1−δ)nc·
=
Ω
·
.
p
K1 1 + (log n) α4 · 9C12 · c3 (F −1 (δ))2
polyLog(n)
As a result, the assumption A4 is satisfied with high probability. Finally, by (C-76), we
obtain that
T
E∆8C ≤ E
max
|α̃j,i
Zj |8 · kΛk8op .
i=0,...,n;j=1,...,p
By Cauchy’s inequality,
E∆8C ≤
r
E
max
i=0,...,n;j=1,...,p
T Z |16 ·
|α̃j,i
j
q
E max ζi16 .
i
Similar to (C-72), we conclude that
E∆8C = O (polyLog(n))
and by Markov inequality, the assumption A5 is satisfied with high probability.
C-4
C-4.1
More Results of Least-Squares (Section 5)
The Relation Between Sj (X) and ∆C
In Section 5, we give a sufficient and almost necessary condition for the coordinatewise asymptotic normality of the least-square estimator β̂ LS ; see Theorem 5.1. In this
subsubsection, we show that ∆C is a generalization of maxj∈Jn Sj (X) for general Mestimators.
Consider the matrix (X T DX)−1 X T , where D is obtain by using general loss functions,
then by block matrix inversion formula (see Proposition E.1),
−1
X1T DX1 X1T DX[1]
X1T
eT1 (X T DX)−1 X T = eT1
T
T
T
X[1] DX1 X[1] DX[1]
X[1]
=
≈
T
T
X1T (I − DX[1] (X[1]
DX[1] )−1 X[1]
)
T DX )−1 X T D)X
X1T (D − DX[1] (X[1]
1
[1]
[1]
T
T
X1T (I − D[1] X[1] (X[1]
D[1] X[1] )−1 X[1]
)
T DX )−1 X T D)X
X1T (D − DX[1] (X[1]
1
[1]
[1]
where we use the approximation D ≈ D[1] . The same result holds for all j ∈ Jn , then
T
T
kX1T (I − D[1] X[1] (X[1]
D[1] X[1] )−1 X[1]
)k∞
keTj (X T DX)−1 X T k∞
≈
.
T
T
T
T
T
−1
T
−1
kej (X DX) X k2
kX1 (I − D[1] X[1] (X[1] D[1] X[1] ) X[1] )k2
T
T
Recall that hTj,1,i is i-th row of I − D[1] X[1] (X[1]
D[1] X[1] )−1 X[1]
, we have
max
i
|hTj,1,i X1 |
keTj (X T DX)−1 X T k∞
≈
.
khj,1,i k2
keTj (X T DX)−1 X T k2
The right-handed side equals to Sj (X) in the least-square case. Therefore, although of
complicated form, assumption A5 is not an artifact of the proof but is essential for the
asymptotic normality.
64
C-4.2
Additional Examples
Benefit from the analytical form of the least-square estimator, we can depart from subgaussinity of the entries. The following proposition shows that a random design matrix
Z with i.i.d. entries under appropriate moment conditions satisfies maxj∈Jn Sj (Z) = o(1)
with high probability. This implies that, when X is one realization of Z, the conditions
Theorem 5.1 are satisfied for X with high probability over Z.
Proposition C.6. If {Zij : i ≤ n, j ∈ Jn } are independent random variables with
1
1. maxi≤n,j∈Jn (E|Zij |8+δ ) 8+δ ≤ M for some δ, M > 0;
2. mini≤n,j∈Jn Var(Zij ) > τ 2 for some τ > 0
3. P (Z has full column rank) = 1 − o(1);
4. EZj ∈ span{Zj : j ∈ Jnc } almost surely for all j ∈ Jn ;
where Zj is the j-th column of Z. Then
max Sj (Z) = Op
j∈Jn
1
1
n4
= op (1).
A typical practically interesting example is that Z contains an intercept term, which is
not in Jn , and Zj has i.i.d. entries for j ∈ Jn with continuous distribution and sufficiently
many moments, in which case the first three conditions are easily checked and EZj is a
multiple of (1, . . . , 1), which belongs to span{Zj : j ∈ Jnc }.
In fact, the condition 4 allows Proposition C.6 to cover more general cases than the
above one. For example, in a census study, a state-specific fix effect might be added into
the model, i.e.
yi = αsi + ziT β ∗ + i
where si represents the state of subject i. In this case, Z contains a sub-block formed by
zi and a sub-block with ANOVA forms as mentioned in Example 1. The latter is usually
incorporated only for adjusting group bias and not the target of inference. Then condition
4 is satisfied if only Zij has same mean in each group for each j, i.e. EZij = µsi ,j .
Proof of Proposition C.6. By Sherman-Morison-Woodbury formula,
eTj (Z T Z)−1 Z T =
ZjT (I − Hj )
ZjT (I − Hj )Zj
T
T
where Hj = Z[j] (Z[j]
Z[j] )−1 Z[j]
is the projection matrix generated by Z[j] . Then
Sj (Z) =
keTj (Z T Z)−1 Z T k∞
kZjT (I − Hj )k∞
q
=
.
keTj (Z T Z)−1 Z T k2
ZjT (I − Hj )Zj
(C-79)
Similar to the proofs of other examples, the strategy is to show that the numerator, as a
linear contrast of Zj , and the denominator, as a quadratic form of Zj , are both concentrated around their means. Specifically, we will show that there exists some constants C1
and C2 such that
n
o
1
1
. (C-80)
max
sup
P kAZj k∞ > C1 n 4 + P ZjT AZj < C2 n = o
j∈Jn
n
n×n
2
A∈R
,A =A,
tr(A)=n−p+1
If (C-80) holds, since Hj is independent of Zj by assumptions, we have
T
kZ
(I
−
H
)k
1
1
C1
C1
j ∞
j
P Sj (Z) ≥ √
· n− 4 = P q
≥√
· n− 4
T
C2
C
2
Z (I − Hj )Zj
j
65
1
≤P k(I − Hj )Zj k∞ > C1 n 4 + P ZjT (I − Hj )Zj < C2 n
1
T
4
=E P k(I − Hj )Zj k∞ > C1 n
Z[j] + E P Zj (I − Hj )Zj < C2 n Z[j]
1
≤
sup
P kAZj k∞ > C1 n 4 + P ZjT AZj < C2 n
(C-81)
A∈Rn×n ,A2 =A,tr(A)=n−p+1
(
≤ max
j∈Jn
1
4
P kAZj k∞ > C1 n
sup
)
+P
A∈Rn×n ,A2 =A,tr(A)=n−p+1
ZjT AZj
< C2 n
=o
1
.
n
(C-82)
Thus with probability 1 − o(|Jn |/n) = 1 − o(1),
1
C1
· n− 4
max Sj (Z) ≤ √
j∈Jn
C2
and hence
max Sj (Z) = Op
j∈Jn
1
.
1
n4
Now we prove (C-80). The proof, although looks messy, is essentially the same as the
proof for other examples. Instead of relying on the exponential concentration given by
the sub-gaussianity, we show the concentration in terms of higher-order moments.
In fact, for any idempotent A, the sum square of each row is bounded by 1 since
X
A2ij = (A2 )j,j ≤ λmax (A2 ) = 1.
i
By Jensen’s inequality,
2
2
EZij
≤ (E|Zij |8+δ ) 8+δ .
For any j, by Rosenthal’s inequality (Rosenthal, 1970), there exists some universal constant C such that
!4+δ/2
8+δ
n
n
n
X
X
X
2
Aij Zij
≤C
|Aij |8+δ E|Zij |8+δ +
A2ij EZij
E
i=1
i=1
i=1
!4+δ/2
n
n
X
X
2
8+δ
2
2
≤C
|Aij | E|Zij |
+
Aij EZij
i=1
i=1
!4+δ/2
n
n
X
X
8+δ
2
2
≤ CM
Aij +
Aij
≤ 2CM 8+δ .
i=1
i=1
1
Let C1 = (2CM 8+δ ) 8+δ , then for given i, by Markov inequality,
!
n
X
1
1
P
Aij Zij > C1 n 4 ≤ 2+δ/4
n
i=1
and a union bound implies that
1
P kAZj k∞ > C1 n 4 ≤
1
n1+δ/4
1
=o
.
n
(C-83)
Now we derive a bound for ZjT AZj . Since p/n → κ ∈ (0, 1), there exists κ̃ ∈ (0, 1 − κ)
such that n − p > κ̃n. Then
EZjT AZj
=
n
X
2
Aii EZij
> τ 2 tr(A) = τ 2 (n − p + 1) > κ̃τ 2 n.
i=1
To bound the tail probability, we need the following result:
66
(C-84)
Lemma C.7 (Bai and Silverstein (2010), Lemma 6.2). Let B be an n × n nonrandom
matrix and W = (W1 , . . . , Wn )T be a random vector of independent entries. Assume that
EWi = 0, EWi2 = 1 and E|Wi |k ≤ νk . Then, for any q ≥ 1,
q
q
E|W T BW − tr(B)|q ≤ Cq (ν4 tr(BB T )) 2 + ν2q tr(BB T ) 2 ,
where Cq is a constant depending on q only.
It is easy to extend Lemma C.7 to non-isotropic case by rescaling. In fact, denote σi2
by the variance of Wi , and let Σ = diag(σ1 , . . . , σn ), Y = (W1 /σ1 , . . . , Wn /σn ). Then
1
1
W T BW = Y T Σ 2 BΣ 2 Y,
1
1
with Cov(Y ) = I. Let B̃ = Σ 2 BΣ 2 , then
1
1
1
1
B̃ B̃ T = Σ 2 BΣB T Σ 2 ν2 Σ 2 BB T Σ 2 .
This entails that
1
1
tr(B̃ B̃ T ) ≤ nu2 tr(Σ 2 BB T Σ 2 ) = ν2 tr(ΣBB T ) ≤ ν22 tr(BB T ).
On the other hand,
q2
1
q
q
q
q
1
tr(B̃ B̃ T ) 2 ≤ nλmax (B̃ B̃ T ) 2 = nν22 λmax Σ 2 BB T Σ 2
≤ nν2q λmax (BB T ) 2 .
Thus we obtain the following result
Lemma C.8. Let B be an n×n nonrandom matrix and W = (W1 , . . . , Wn )T be a random
vector of independent mean-zero entries. Suppose E|Wi |k ≤ νk , then for any q ≥ 1,
q
q
E|W T BW − EW T BW |q ≤ Cq ν2q (ν4 tr(BB T )) 2 + ν2q tr(BB T ) 2 ,
where Cq is a constant depending on q only.
Apply Lemma C.8 with W = Zj , B = A and q = 4 + δ/2, we obtain that
E|ZjT AZj − EZjT AZj |4+δ/2 ≤ CM 16+2δ (tr(AAT ))2+δ/4 + tr(AAT )2+δ/4
for some constant C. Since A is idempotent, all eigenvalues of A is either 1 or 0 and thus
AAT I. This implies that
tr(AAT ) ≤ n,
tr(AAT )2+δ/4 ≤ n
and hence
E|ZjT AZj − EZjT AZj |4+δ/2 ≤ 2CM 16+2δ n2+δ/4
for some constant C1 , which only depends on M . By Markov inequality,
P
|ZjT AZj − EZjT AZj | ≥
κ̃τ 2 n
2
≤ 2CM 16+2δ
2
κ̃τ 2
4+δ/2
·
1
n2+δ/4
.
Combining with (C-84), we conclude that
P ZjT AZj < C2 n = O
2
1
n2+δ/4
=o
1
n
(C-85)
where C2 = κ̃τ2 . Notice that both (C-83) and (C-85) do not depend on j and A. Therefore, (C-80) is proved and hence the Proposition.
67
D
Additional Numerical Experiments
In this section, we repeat the experiments in section 6 by using L1 loss, i.e. ρ(x) = |x|.
L1 -loss is not smooth and does not satisfy our technical conditions. The results are
displayed below. It is seen that the performance is quite similar to that with the huber
loss.
^
Coverage of β1 (κ = 0.5)
normal
^
Coverage of β1 (κ = 0.8)
t(2)
normal
1.00
●
●
●
●
●
●
0.95
0.90
1.00
1.00
●
●
●
●
●
●
●
●
0.90
1.00
0.95
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
100
200
ellip
0.95
Coverage
0.90
ellip
Coverage
●
iid
●
iid
0.95
t(2)
1.00
0.90
1.00
0.90
100
200
400
800
100
200
400
hadamard
hadamard
0.95
0.95
0.90
800
100
200
400
Sample Size
Entry Dist.
●
normal
t(2)
800
400
800
Sample Size
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 5: Empirical 95% coverage of β̂1 with κ = 0.5 (left) and κ = 0.8 (right) using
L1 loss. The x-axis corresponds to the sample size, ranging from 100 to 800; the y-axis
corresponds to the empirical 95% coverage. Each column represents an error distribution
and each row represents a type of design. The orange solid bar corresponds to the case
F = Normal; the blue dotted bar corresponds to the case F = t2 ; the red dashed bar
represents the Hadamard design.
68
^
^
Min. coverage of β1 ~ β10 (κ = 0.5)
normal
^
^
Min. coverage of β1 ~ β10 (κ = 0.8)
t(2)
normal
0.95
0.90
t(2)
0.95
●
●
●
●
●
●
●
●
0.90
●
●
●
●
●
●
●
●
0.80
0.95
●
0.90
●
●
●
●
●
ellip
●
●
Coverage
0.95
●
ellip
Coverage
iid
0.80
0.95
0.90
●
●
iid
0.80
●
●
●
●
●
0.80
0.95
0.80
100
200
400
800
100
200
400
hadamard
hadamard
0.90
0.90
0.80
800
100
200
400
Sample Size
Entry Dist.
●
normal
800
100
200
400
800
Sample Size
t(2)
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 6: Mininum empirical 95% coverage of β̂1 ∼ β̂10 with κ = 0.5 (left) and κ = 0.8
(right) using L1 loss. The x-axis corresponds to the sample size, ranging from 100 to 800;
the y-axis corresponds to the minimum empirical 95% coverage. Each column represents
an error distribution and each row represents a type of design. The orange solid bar
corresponds to the case F = Normal; the blue dotted bar corresponds to the case F = t2 ;
the red dashed bar represents the Hadamard design.
^
^
Bonf. coverage of β1 ~ β10 (κ = 0.5)
normal
^
^
Bonf. coverage of β1 ~ β10 (κ = 0.8)
t(2)
normal
1.00
●
●
●
●
●
●
●
●
●
●
●
●
●
●
iid
0.95
iid
0.95
●
0.90
0.90
1.00
●
●
●
●
●
●
●
●
●
●
●
●
●
0.95
●
●
0.90
0.90
1.00
1.00
hadamard
hadamard
0.95
●
ellip
ellip
0.95
Coverage
1.00
Coverage
t(2)
1.00
●
0.95
0.90
0.90
100
200
400
800
100
200
400
800
100
200
400
Sample Size
Entry Dist.
●
normal
t(2)
800
100
200
400
800
Sample Size
hadamard
Entry Dist.
●
normal
t(2)
hadamard
Figure 7: Empirical 95% coverage of β̂1 ∼ β̂10 after Bonferroni correction with κ = 0.5
(left) and κ = 0.8 (right) using L1 loss. The x-axis corresponds to the sample size,
ranging from 100 to 800; the y-axis corresponds to the empirical uniform 95% coverage
after Bonferroni correction. Each column represents an error distribution and each row
represents a type of design. The orange solid bar corresponds to the case F = Normal;
the blue dotted bar corresponds to the case F = t2 ; the red dashed bar represents the
Hadamard design.
69
E
Miscellaneous
In this appendix we state several technical results for the sake of completeness.
Proposition E.1 (Horn & Johnson, 2012, formula (0.8.5.6)). Let A ∈ Rp×p be an invertible matrix and write A as a block matrix
A11 A12
A=
A21 A22
with A11 ∈ Rp1 ×p1 , A22 ∈ R(p−p1 )×(p−p1 ) being invertible matrices. Then
−1
−1
A11 + A−1
A21 A−1
−A−1
11 A12 S
11
11 A12 S
A−1 =
−S −1 A21 A−1
S −1
11
where S = A22 − A21 A−1
11 A12 is the Schur’s complement.
Proposition E.2 (Rudelson & Vershynin, 2013; improved version of the original form
by Hanson & Wright, 1971). Let X = (X1 , . . . , Xn ) ∈ Rn be a random vector with
independent mean-zero σ 2 -sub-gaussian components Xi . Then, for every t,
t
t2
,
P |X T AX − EX T AX| > t ≤ 2 exp −c min
σ 4 kAk2F σ 2 kAkop
Proposition E.3 (Bai & Yin, 1993). If {Zij : i = 1, . . . , n, j = 1, . . . , p} are i.i.d. random
variables with zero mean, unit variance and finite fourth moment and p/n → κ, then
T
T
√ 2
√
Z Z a.s.
Z Z a.s.
→ (1 + κ) , λmin
→ (1 − κ)2 .
λmax
n
n
Proposition E.4 (Latala, 2005). Suppose {Zij : i = 1, . . . , n, j = 1, . . . , p} are independent mean-zero random variables with finite fourth moment, then
sX
sX
sX
q
2 + max
2 +
4
E λmax (Z T Z) ≤ C max
EZij
EZij
EZij
4
i
j
j
i
i,j
4
for some universal constant C. In particular, if EZij
are uniformly bounded, then
λmax
ZT Z
n
r
p
= Op 1 +
.
n
Proposition E.5 (Rudelson & Vershynin, 2010). Suppose {Zij : i = 1, . . . , n, j =
1, . . . , p} are independent mean-zero σ 2 -sub-gaussian random variables. Then there exists
a universal constant C1 , C2 > 0 such that
s
r
T
!
2
Z Z
p
P
λmax
> Cσ 1 +
+t
≤ 2e−C2 nt .
n
n
Proposition E.6 (Rudelson & Vershynin, 2009). Suppose {Zij : i = 1, . . . , n, j =
1, . . . , p} are i.i.d. σ 2 -sub-gaussian random variables with zero mean and unit variance,
then for ≥ 0
s
!
r
T
Z Z
p−1
≤ (1 −
) ≤ (C)n−p+1 + e−cn
P
λmin
n
n
for some universal constants C and c.
70
Proposition E.7 (Litvak et al., 2005). Suppose {Zij : i = 1, . . . , n, j = 1, . . . , p} are
independent σ 2 -sub-gaussian random variables such that
d
Zij = −Zij ,
Var(Zij ) > τ 2
for some σ, τ > 0, and p/n → κ ∈ (0, 1), then there exists constants c1 , c2 > 0, which
only depends on σ and τ , such that
T
Z Z
< c1 ≤ e−c2 n .
P λmin
n
71
| 10math.ST
|
arXiv:1610.03046v4 [math.GT] 16 Dec 2017
The classification of Kleinian groups of
Hausdorff dimensions at most one
Yong Hou∗
Institute for Advanced Study
Princeton University
Abstract
In this paper we provide the complete classification of Kleinian
group of Hausdorff dimensions less than 1. In particular, we prove
that every purely loxodromic Kleinian groups of Hausdorff dimension
< 1 is a classical Schottky group. This upper bound is sharp. As an
application, the result of [4] then implies that every closed Riemann
surface is uniformizable by a classical Schottky group. The proof relies
on the result of Hou [6], and space of rectifiable Γ-invariant closed
curves.
1
Introduction and Main Theorem
We take Kleinian groups to be finitely generated, torsion-free, discrete subgroups of PSL(2, C). The main theorem is:
Theorem 1.1 (Classification). Any purely loxodromic Kleinian group Γ with
limit set of Hausdorff dimension < 1 is a classical Schottky group. This bound
is sharp.
∗
Supported by Ambrose Monell Fundation
1
We note that by Selberg lemma, torsion-free is not really a restriction,
since any finitely generated discrete subgroup of PSL(2, C) has a finite index
torsion-free subgroup.
As an application, we have the following Corollary 1.3, which is a resolution of a folklore problem of Bers on classical Schottky group uniformization
of closed Riemann surface. Corollary 1.3 follows from the work of Hou [4]:
Theorem 1.2 (Hou [4]). Every closed Riemann surface is uniformizable by a
Schottky group of Hausdorff dimension < 1, i.e. every point in moduli space
has a Hausdorff dimension < 1 fiber in the Schottky space.
Corollary 1.3. (Uniformization) Every closed Riemann surface can be uniformized by a classical Schottky group.
1.1
Strategy of proof
First let us recall the result of Hou [6]:
Theorem 1.4 (Hou). There exists λ > 0 such that any Kleinian group with
limit set of Hausdorff dimension < λ is a classical Schottky group
Define Hc = sup{λ|satisfies Theorem 1.4}. Hc is the maximal parameter
such that if Γ is a Schottky group of Hausdorff dimension < Hc then Γ is
classical Schottky group. Hence Theorem 1.1 can be rephrased as: Hc ≥ 1.
We prove by contradiction, so from now on and throughout the paper we
assume that Hc < 1, then we will show that Hc is not maximal.
Recall that the Hausdorff dimension function on the Schottky space of
c
rank g is real analytic. It is a consequence of Theorem 1.4 that, JH
g (rankg Schottky groups of Hausdorff dimension < Hc , see Section 2) is a 3g −
3 dimensional open and connected submanifold of Jg , the rank-g Schottky
space.
c
The proof is done as follows. First we note that, ∂JH
must contain a
g
non-classical Schottky group, otherwise Hc is not maximal by definition, see
Proposition 2.12. Second we show that, if Hc < 1 then every element of the
c
boundary ∂JH
g is either a classical schottky group or, it is not a Schottky
group, Lemma 3.11. This contradicts the first fact, hence we must have
Hc ≥ 1. The bulk of the paper is devoted to proof the second fact, which we
now summarize the idea in the following.
It is a result of Bowen [2] that, a Schottky group Γ has Hausdorff dimension < 1, if and only if there exist a rectifiable Γ-invariant closed curve. Let
2
R(S 1 , W ) be the space of bounded length closed curves which intersects the
compact set W ⊂ C and equipped with Fréchet metric. It is complete space,
see Section 2. We show that if Hc < 1 then, every quasi-circle with bounded
length of Γ is the limit of a sequence of quasi-circles of Γn in R(S 1 , W ).
We also show that if Γ is a Schottky group, then every quasi-circle of Γ has
an open neighborhood in the relative topology of ΨΓ (see section 3) such
that, every element of the open neighborhood is a quasi-circle of Γ. We also
define linearity and transversality invariant for quasi-circles, and show that
quasi-circles of classical Schottky groups preserve these invariants, and nonclassical Schottky groups do not have transverse linear quasi-circles.
Given a quasi-circle of a Schottky group Γ, we show that there exists
an open neighborhood (in the relative topology of space of rectifiable curves
with respect to Frechet metric) about the quasi-circle such that, every point
in the open neighborhood is a quasi-circle of Γ, see Lemma 3.9. Next assume that we have a sequence of classical Schottky groups Γn → Γ to a
Schottky group, and are all of Hasudorff dimensions less than one. We
then study singularity formations of classical fundamental domains of Γn
when Γn → Γ. These singularities are of three types: tangent, degenerate,
and collapsing. We show that all these singularities will imply that there
exists a quasi-circle such that, every open neighborhood about this quasicircle will contain some points which is not a quasi-circle. Essentially, the
existence of a singularity will be obstruction to the existence of any open
neighborhood that are of quasi-circles, see Lemma 3.11. Hence it follows
from these results that, if Γn → Γ with Γn classical and, all Hausdorff dimensions are of less than one then Γ must be a classical Schottky group.
Acknowledgement
This work is made possible by unwavering supports and insightful conversations from Peter Sarnak, whom I’m greatly indebted to. It is the groundbreaking works of Peter Sarnak that has guided the author to study this
problem at first place. I wish to express my deepest gratitude and sincere
appreciation to Dave Gabai for the continuous of amazing supports and encouragements which allowed me to complete this work.
I want to express my sincere appreciation to the referee for detailed reading and, helpful comments and suggestions. I also want express sincere appreciation to Ian Agol, Matthew de Courcy-Irland, for reading of the previous
draft.
This paper is dedicated to my father: ShuYing Hou.
3
2
Quasi-circles and generating Jordan curves
Schottky group Γ of rank g is defined as convex-cocompact discrete faithful
representation of the free group Fg in PSL(2, C). It follows that Γ is freely
generated by purely loxodromic elements {γi }g1 . This implies we can find
collection of open topological disks Di , Di+g , 1 ≤ i ≤ g of disjoint closure
D̄i ∩ D̄i+g = ∅ in the Riemann sphere ∂H3 = C with boundary curves ∂ D̄i =
ci , ∂ D̄i+g = ci+g . By definition ci , ci+g are closed Jordan curves in Riemann
sphere ∂H3 , such that γi (ci ) = ci+g and γi (Di ) ∩ Di+g = ∅. Whenever there
exists a set {γ1 , ..., γg } of generators with all {ci , ci+g }g1 as circles, then it is
called a classical Schottky group with {γ1 , ..., γg } classical generators.
Schottky space Jg is defined as space of all rank g Schottky groups up
to conjugacy by PSL(2, C). By normalization, we can chart Jg by 3g − 3
complex parameters. Hence Jg is 3g − 3 dimensional complex manifold. The
bihomolomorphic Auto(Jg ) group is Out(Fg ), which is isomorphic to quotient
of the handle-body group. Denote by Jg,o the set of all elements of Jg that
are classical Schottky groups. Note that Jg,o is open in Jg . On the other
hand it is nontrivial result due to Marden that Jg,o is non-dense subset of
Jg . However, it follows from Theorem 1.4, Jλg ⊂ Jg,o is 3g − 3 dimensional
open connected submanifold. Here Jλg denotes space of Schottky groups of
Hausdorff dimension < λ.
Some notations:
• Given Γ a Kleinian group, we denote by ΛΓ and ΩΓ and DΓ its limit set,
region of discontinuity, and Hausdorff dimension respectively throughout this paper.
• Given a fundamental domain F of Γ, we denote the orbit of F under
actions of Γ by FΓ . We also say FΓ is a classical fundamental domain
of classical Schottky group if ∂F are disjoint circles.
Definition 2.1 (Quasi-circles). Given a geometrically finite Kleinian group
Γ, a closed Γ-invariant Jordan curve that contains the limit set ΛΓ is called
quasi-circle of Γ.
Remark: From now on, we make the global assumption throughout this paper
that Hc < 1, and all Schottky groups are of Hausdorff dimension DΓ ≤ Hc ,
if not stated otherwise.
Next we give a construction of quasi-circles of Γ which is a generalization
of the construction by Bowen [2].
4
Let F be a fundamental domain of Γ, and {ci }2g
1 be the collection of
2g disjoint Jordan curves comprising ∂F. Let ζ denote collection of arcs
ζ = {ζi } connecting points pi ∈ ci , p0i ∈ ci+g for 1 ≤ i ≤ g, and arcs on
ci+g that connects pi+g to γi (pi ) and p0i+g ∈ ci+g to γi (p0i ). So ζ is a set of
g disjoint curves connecting disjoint points on collection of Jordan curves of
∂F (Figure 1 ).
Figure 1: Quasi-circle
ηΓ = ΛΓ ∪ ∪γ∈Γ γ(ζ) defines a Γ-invariant closed curve containing ΛΓ . ηΓ
defines a quasi-circle of Γ. Obviously there are infinitely many quasi-circles
and different ζ gives a different quasi-circles. Note that, the simply connected
regions C \ ηΓ gives the Bers simultaneous uniformization of Riemann surface
ΩΓ /Γ.
Definition 2.2 (Generating curve). Given a quasi-circle ηΓ of Γ. We say
a collection of disjoint curves ζ is a generating curve of ηΓ , if ηΓ can be
generated by ζ.
Note that the quasi-circles constructed in [2], which requires that pi+g is
a imagine of pi under element of γi , is a subset of the collection that we have
defined here. In fact, this generalization is also used for the construction of
quasi-circles of non-classical Schottky groups.
Proposition 2.3. Every quasi-circle ηΓ of Γ is generated by some generating
curves ζ.
Proof. Let ηΓ be a quasi-circle of Γ. Let F be a fundamental domain of Γ.
Set ξ = F ∩ ηΓ and ξ¯ = ∂F ∩ ηΓ . Then ξ, ξ¯ consists of collection of disjoint
5
curves which only intersects along ∂F. Hence we have, γ(ξ) ∩ γ 0 (ξ) = ∅ for
¯
γ, γ 0 ∈ Γ with γ 6= γ 0 . Since Γ(ξ)−Γ(ξ) = ΛΓ we have have ηΓ = ΛΓ ∪Γ(ξ ∪ ξ),
¯
hence ξ ∪ ξ is a generating curve of ηΓ .
Definition 2.4 (Linear quasi-circle). We call a quasi-circle ηΓ linear if,
ηΓ \ΛΓ consists of points, circular arcs or lines.
Note that, if ηΓ is linear then there exists FΓ such that, ηΓ ∩ FΓ and
ηΓ ∩ ∂FΓ are piece-wise circular arcs or lines.
We say an arc ζ ⊂ ηΓ ∩FΓ is orthogonal if the tangents at intersections on
∂FΓ are orthogonal with ∂FΓ , and an arc ξ ⊂ ηΓ ∩ ∂FΓ is parallel if ξ ⊂ ∂FΓ .
Definition 2.5 (Right-angled quasi-circle). Given a linear quasi-circle ηΓ of
Γ, if all linear arcs intersect at right-angle then we say ηΓ is right-angled
quasi-circle.
Definition 2.6 (Transverse quasi-circle). Given a quasi-circle ηΓ of Γ, we
say ηΓ is transverse quasi-circle if ηΓ intersects ∂FΓ orthogonally for some
FΓ and, ηΓ have no parallel arc. Otherwise, we say ηΓ is non-transverse.
Definition 2.7 (Parallel quasi-circle). Given a quasi-circle ηΓ of Γ, we say
ηΓ is parallel quasi-circle if there exists some arc η of ηΓ such that η ⊂ ∂FΓ
Proposition 2.8. Transverse quasi-circles always exists for a given Schottky
group Γ.
Proof. Let F be bounded by 2g distinct Jordan closed curves and take any
curve connecting pi ∈ ci , pi+g ∈ ci+g such that pi+g = γi (pi ) and pi+g ∈ ci+g =
γi (pi+g ) for 1 ≤ i ≤ g, that intersects ci , ci+g orthogonally.
It should be noted that a quasi-circle of Γ in general is not necessarily
rectifiable. For instance, if we take ζ to be some non-rectifiable generating
curves then, ηΓ will be non-rectifiable. Recall a curve is said to be rectifiable
if and only if the 1-dimensional Hausdorff measure of the curve is finite. This
is not the only obstruction to rectifiability, in fact we have the following result
of Bowen:
Theorem 2.9 ([2]). For a given Schottky group Γ, the Hausdorff dimension
of limit set is DΓ < 1, if and only if there exists a rectifiable quasi-circle for
Γ.
6
The proof of Theorem 2.9 relies on the fact that the Poincare series of Γ
converges if and only if DΓ < 1 [2].
Proposition 2.10. Let Γ be a Schottky group of DΓ < 1. Suppose a given
generating curve ζ is a rectifiable curve. Then ηΓ is a rectifiable quasi-circle
of Γ.
Proof. Let µ1 be the 1-dimensional Hausdorff measure. Since DΓ < 1, we
0
have µ1 (ΛΓ ) = 0. Let Γ = {γk }∞
k=1 , let γk denotes the derivative of γk . Then
we have,
1
µ (ηΓ ) =
∞
X
µ1 (γk (ζ))
k=1
1
µ (ζ)
∞
X
|γk0 (z)|,
z ∈ ζ.
k=1
Also since DΓ < 1 if and only if Poincare series satisfies
implies that ηΓ is rectifiable if and only if ζ rectifiable.
P∞
k=1
|γk0 | < ∞. This
Let W ⊂ C be a compact set. Denote the space of closed curves with
bounded length in C that intersect with W by:
R(S 1 , W ) ⊂ {h : S 1 → C|h(S 1 ) ∩ W 6= ∅,
h continuous rectifiable map}.
For h1 , h2 ∈ R(S 1 , W ) let `(h1 ), `(h2 ) be it’s respective arclength. The
Fréchet distance is defined as,
dF (h1 , h2 ) = inf{sup |h1 (σ1 ) − h2 (σ2 )|; σ1 , σ2 ∈ Homeo(S 1 )} + |`(h1 ) − `(h2 )|.
For a given compact W ⊂ C, the space of closed curves with bounded
length R(S 1 , W ) is a metric space with respect to dF . Two curves in φ, ψ ∈
R(S 1 , W ) are same if there exists parametrization σ such that ψ(σ) = φ and
`(φ) = `(ψ). The topology on R(S 1 , W ) is defined with respect to the metric
dF , see [1] p388. Let ξ be a generating curve for ηΓ . Fix a indexing of Γ, set
ξi = ∪i1 γj (ξ). Let σi be a parametrization of ξi such that σ = ∪i σi define a
parametrization of ηΓ . Then dF (ηΓ1 , ηΓ2 ) ≤ M (inf σ1 ,σ2 |ξ(σk1 ) − ζ(σk2 )| + |`(ξ) −
`(ζ)|) for some M, k. This implies continuity of ηΓ with respect to generating
curve ξ.
Proposition 2.11. For a given compact W ⊂ C, the space R(S 1 , W ) is
complete metric space with respect to dF .
7
Proof. Let {φi } ⊂ R(S 1 , W ) be a Cauchy sequence. φi are rectifiable curves
of bounded length and so there exists {σi } Lipschitz parameterizations with
bounded Lipschitz constants, such that {φi (σi (t))} are uniformly Lipschitz.
Then completeness follows from the fact that all curves of R(S 1 , W ) are
contained within some large compact subset of C.
Hc
c
Proposition 2.12. There exists a Γ ∈ JH
g in the closure of Jg , such that
Γ is not a classical Schottky group.
c
Proof. Suppose false, then every element Γ ∈ JH
g with Hausdorff dimension
≤ Hc is a classical Schottky group. Since the classical Schottky space Jg,o is
c
open in the Schottky space Jg , we have a open neighborhood U of JH
g in Jg
c
such that U ⊂ JH
g,o . By definition of Hc , it is maximal. Hence there are nonclassical Schottky groups of Hausdorff dimension arbitrarily close to Hc and
> Hc . Then there exists a sequence of non-classical Schottky groups {Γn }
of Hausdorff dimensions DΓn & Hc . Let Γ = limn Γn . Since by assumption,
all Schottky groups of Hausdorff dimension Hc is classical, we must have Γ
either it is not a Schottky group, or it is a classical Schottky group. This
implies that for large n we must have Γn ∈ U , which is a contradiction
to the sequence Γn been all non-classical Schottky groups, hence Hc is not
maximal.
Proposition 2.13. There exists a non-classical Schottky group Γ of DΓ =
Hc , and sequence of {Γn } classical Schottky groups such that DΓn < DΓ with
sup DΓn = DΓ .
Proof. It follows from that the Hausdorff dimension map D : Jg → (0, 2) is
real analytic map, and by Proposition 2.12, there exists a non-classical Γ with
c
DΓ = Hc . Theorem 1.4 implies JH
g is open submanifold of Jg,o , hence there
c
exists a sequence of classical Schottkys {Γn } ⊂ JH
g with DΓn → DΓ .
3
Schottky Space and rectifiable curves
Take {Γn } and Γ as given by Proposition 2.13. In particular, {Γn } is a sequence of classical Schottky groups such that Γn → Γ. Denote QC(C) space
of quasiconformal maps on C. It follows from quasiconformal deformation
theory of Schottky space, for Γc classical Schottky group of Hausdorff dimension < Hc we can write ,
Jg = {f ◦ γ ◦ f −1 ∈ PSL(2, C)|f ∈ QC(C), γ ∈ Γc }/PSL(2, C).
8
Remark: From now on throughout rest of the paper, we fix Γc to be a
classical Schottky group with Hausdorff dimension DΓc < Hc .
Notations: Set H to be the collection of all Schottky groups of Hausdorff
dimension < Hc .
Note that there exists a sequence of quasiconformal maps fn and f of C
such that, we can write Γn = fn (Γ) and Γ = f (Γc ). Here we write f (Γ) :=
{f ◦ g ◦ f −1 |g ∈ Γ} for a given Kleinian group Γ and quasiconformal map f.
Schottky space Jg can also be considered as subspace of C3g−3 . This provides Jg analytic structure as 3g − 3-dimensional complex analytic manifold.
Proposition 3.1. Let {Γn } be a sequence of Schottky groups with Γn → Γ
to a Schottky group Γ. Let F be a fundamental domain of Γ. There exists a
sequence of fundamental domain {Fn } of Γn such that Fn → F.
Proof. Let ∪2g
i=1 Ci = ∂F be the Jordan curves which is the boundary of F.
Set Cn = fn−1 (∪2g
i=1 Ci ). Then Cn is the boundary of a fundamental domain of
−1
fn (Γ). Hence we have a fundamental domain Fn of Γn defined by Cn with
Fn → F.
Lemma 3.2. Suppose fn → f and Γ = f (Γc ). Every quasi-circle with
bounded length of Γn = fn (Γc ) is in R(S 1 , W ) for some compact W.
Proof. Let ηn be a quasi-circle of Γn . Note ΛΓn ⊂ ηn . Since limit set ΛΓn →
ΛΓ , and limit set Λ(Γ) is compact, and ηn is rectifiable, we can find some
compact set W ⊃ ∪n ΛΓn ∪ ΛΓ .
Given {Γn } ⊂ H with Γn → Γ, let EΓn denote the collection of all bounded
dF
length quasi-circles of Γn . We define ΨΓ = ∪EΓn to be the closure of the
set of bounded length quasi-circles of Γn in R(S 1 , W ).
Proposition 3.3. The subspace ΨΓ is compact.
Proof. Curves in ΨΓ are bounded length and we have parametrization with
bounded Lipschitz constants. It follows from Ascoli-Arzela theorem we have
uniform convergence topology on ΨΓ . Since curves in R(S 1 , W ) are all contained in some large compact set of C, we have ΨΓ is closed and bounded,
hence compact.
9
We define O(η), open sets about η ∈ ΨΓ in relative topology given by
O(η) ∩ ΨΓ for some open set O(η) ⊂ R(S 1 , W ).
Let Γ be a Schottky group. Let F be a fundamental domain of Γ. For
O(ηΓ ) of ηΓ ∈ ΨΓ , and suppose every element is quasi-circle, and let ζξ
denote a generating curve of ξ ∈ O(ηΓ ) with respect to F. Then we have,
O(ζ) = ∪ξ∈O(ηΓ ) ζξ the collection of all generating curves of the open set
O(ηΓ ) gives a open set of generating curves of ηΓ . On set of collection of all
generating curves of elements of ΨΓ , we define the topology as ξηn → ξη , if
and only if ηn → η for ηn , η ∈ ΨΓ .
We will sometime denote by η∞ ∈ ∂ΨΓ a curve which is the limit of
rectifiable quasi-circles of {Γn }.
Proposition 3.4. Let {Γn } ⊂ H with Γn → Γ. Then every bounded length
quasi-circle ηΓ of Γ is in ΨΓ . In addition, if ηn are linear then ηΓ is linear
quasi-circles of Γ.
Proof. Let Γn = fn−1 (Γ). Note that since DΓn < Hc , we have DΓ ≤ Hc < 1.
Define ηn = fn−1 (ηΓ ), for all n. Then {ηn } is a sequence of Jordan closed
curves. It follows from Proposition 2.3, we have a generating curve ζ of ηΓ . So
ηΓ = ΛΓ ∪∪γ∈Γ γ(ζ), and we have ηn = fn−1 (ΛΓ )∪fn−1 (∪γ∈Γ γ(ζ)). Since ΛΓn =
fn−1 (ΛΓ ) and fn−1 (∪γ∈Γ γ(ζ)) = ∪γ∈Γ fn−1 γfn (fn−1 (ζ)) so its ∪γn ∈Γn γn (ζn ), where
ζn = fn−1 (ζ). Hence ζn is a generating curve of ηn which are quasi-circles of
Γn . Denote by ζΓ a generating curve for ηΓ . Since ζ is rectifiable curve,
modify ζn if necessary, we can assume {ζn } are rectifiable curves.
Let z ∈ ζn , since DΓn < Hc , we have the 1-dimension Hausdorff measure
1
µ (ηn ):
X
µ1 (ηn ) µ1 (ζn )
|γ 0 (z)| < ∞,
γ∈Γn
0
where γ is the derivative of γ. Hence {ηn } are rectifiable quasi-circles. It
follows that there exists c > 0 such that µ1 (ηn ) < cµ1 (ηΓ ) for large n, hence
{ηn } are bounded quasi-circles of R(S 1 , W ), and by Proposition 3.3, we have
ηn → ηΓ and ηΓ ∈ ΨΓ . Finally, if ηn are linear then ζn are linear and since
Mobius maps preserves linearity, we have ηΓ is linear.
Definition 3.5 (Good-sequences). For a given sequence {ηn } of quasi circles
of {Γn } ⊂ H with Γn → Γ, we say {ηn } is a good-sequence of quasi-circles
if, it is convergent sequence and η∞ is a quasi-circle of Γ. We also call {ηn }
(non)transverse good-sequence if all ηn are also (non)transverse.
10
Lemma 3.6 (Existence). Let {Γn } ⊂ H be a sequence of Schottky groups
with Γn → Γ. There exists a good-sequence {ηn } of quasi-circles of {Γn }. In
addition, if {ηn } is also (non)transverse then η∞ is (non)transverse.
Proof. Let fn be quasi-conformal maps such that Γn = fn (Γ). Let ηΓ be a
quasi-circle of Γ. Then ηn = fn (ηΓ ) is a good-sequence of quasi-circles. The
(non)transverse property obviously is preserved.
Corollary 3.7 (Linear-invariant). Let {ηn } be a good-sequence of linear
quasi-circles of {Γn } ⊂ H. η∞ is a linear quasi-circle of Γ.
Proof. Linearity is obviously preserved at η∞ .
Corollary 3.8 (Tranverse-invariant). Let {ηn } be a good-sequence of quasicircles of {Γn } ⊂ H. Then η∞ is a (non)transverse linear quasi-circle of Γ if
and only if {ηn } is a (non)transverse.
Proof.
Lemma 3.9 (Open). Let {Γn } ⊂ H. Let Γn → Γ be a Schottky group. Let
ηΓ be a quasi-circle of Γ. Then there exists a open neighboredood O(ηΓ ) of ηΓ
in ΨΓ such that every elements of O(ηΓ ) is a quasi-circle of Γ.
Proof. Let F be a fundamental domain of Γ. Let ξ be the generating curve of
ηΓ with respect to F. Let Fn be the fundamental domain of Γn with Fn → F.
By Proposition 3.4, we have ηn → ηΓ . Let U(ξn ) to be the open set about the
generating curve ξn of ηn , and set O(ξn ) = U(ξn ) ∩ ΨΓ . Then O(ξn ) is open
sets of generating curves in ΨΓ (Figure 2). Let O(ξn ) = fn (O(ξc )) for O(ξc ) a
open set of generating curves for ηΓc = fn−1 (ηn ). Since ηΓ is a quasi-circle, and
Fn → F, for large n we can choose sufficiently small neighborhood O0 (ξc )
such that f (O0 (ξc )) is a open neighborhood of generating curves of ηΓ . Let
ξn0 ∈ fn (O0 (ξc )) and ηn0 be generated by ξn0 . Assuming O0 (ξc ) is sufficiently
small neighborhood, we have length `(ηn0 ) < c`(ηΓ ) for some small c ≥ 1 for
0
large n and ηn0 generated by all ξn0 ∈ fn (O0 (ξc )). Since ηn → ηΓ , we have η∞
is quasi-circle of Γ with bounded length. This defines a open set which all
elements are quasi-circles in ΨΓ .
Corollary 3.10. Let {Γn } ⊂ H with Γn → Γ a Schottky group Γ. Let {ηn }
be a good-sequence of quasi-circles of {Γn }. Then there exists an open neighborhood O(η∞ ) of η∞ such that every element of O(η∞ ) is a quasi-circle of
Γ.
11
Figure 2: Open set of generating curves about ζ
Proof. Follows from Lemma 3.6 and Lemma 3.9.
Next we analyze the formations of singularities for a given sequence of
classical Schottky groups converging to a Schottky group. These types of
singularities has been studied in [5, 6].
Lemma 3.11 (Singularity). Let {Γn } ⊂ H with Γn → Γ a Schottky group.
Assume that Γ is non-classical Schottky group. Then there exists ηΓ such
that, every O(ηΓ ) contains a non-quasi-circle.
Here we say a closed curve is a non-quasi-circle, if it’s non-Jordan(contains
a singularity point), or it’s not Γ-invariant.
Proof. For each n, let Fn be a classical fundamental domain of Γn . Given a
sequence of classical fundamental domains {Fn } the convergence is consider
as follows: {∂Fn } is collection of 2g circles {ci,n }i=1,...,2g in the Riemann
sphere C, pass to a subsequence if necessary, then limn ci,n , is either a point
or a circle. We say {Fn } convergents to G, if G is a region that have boundary
consists of limn ci,n for each i, which necessarily is either a point or a circle.
Note that G is not necessarily a fundamental domain, nor it’s necessarily
connected.
Let lim Fn = G. By assumption that Γ is not a classical Schottky group,
we have G is not a classical fundamental domain of Γ. We have ∂G consists of
circles or points. However these circles may not be disjoint. More precisely,
we have the following possible degeneration of circles of ∂Fn which gives ∂G
of at least one of following singularities types:
• Tangency: Contains tangent circles.
12
• Degeneration: Contains a circles degenerates into a point.
• Collapsing: Contains two circles collapses into one circle. Here we have
two concentric circles centered at origin and rest circles squeezed in
between these two and these two collapse into a single circle in G.
Consider ∂G contains a tangency:
Let p be a tangency point. Let ηΓ be a quasi-circle of Γ which pass
through point p. Assume the Lemma is false, then all sufficiently small open
neighborhood O(ηΓ ) contains only quasi-circles of Γ. It follows from Proposition 3.2, we have a sequence with ηn → ηΓ . Let ξn be the generating curves
of ηn with respect to Fn . Define φn as follows. Note that ξn ∩ ∂Fn consists
of points or linear arcs. We define ζn be a generating curve with ζn ∩ ∂Fn
to consists of linear arcs, with one of its end point to be the point that converges into a tangency in ∂G. In addition we also require the arcs on ∂Fn
that with a end points which converges to tangency point be be connect by a
arc between the other end points. It is clear that every O(ηn ) ∩ Fn contains a
generating curve of this property. Let φn be the quasi-circle generated by ζn ,
and set φ = lim φn , then φ contains a loop singularity at a point of tangency
(Figure 3). Hence φ is non-quasi-circle of Γ. Since every O(ηΓ ) contains such
a φ we must have the lemma to be true for this type of singularity.
Figure 3: Tangency singularity
Consider ∂G contains a degeneration:
In this case, we must have any quasi-circle ηn pass through a degeneration
point. Note that we can’t have all circles degenerates into a single point.
There are two possibilities to have degenerate points. Case (A): two circles
merge into a single degenerate point p. Case (B): A circle degenerates into a
point on a circle.
13
Consider (A). In this case, any quasi-circles η∞ will have two possible
properties, either there is a point q on some circle of ∂G such that every
quasi-circle must pass through q, or two separate arcs of η∞ meet at p. The
second possibility implies there exists a loop singularity at p, hence η∞ can
not be a Jordan curve (Figure 4). Therefore we only have the first possibility
for η∞ . But it follows from Proposition 3.4, all rectifiable quasi-circles of
Γ is the limit of some sequence of quasi-circles of Γn , and hence all must
pass through q. However q is not a limit point, hence we can have some
quasi-circle not passing through q, a contradiction.
Figure 4: Degenerate singularities
Now consider (B). Here we can assume that there exists at least two circles
that do not degenerates into points. Otherwise, we will have the third type
(collapsing) singularity, which we will consider next. Having some circles
degenerates into a point on to a circle at p, we have a sequence of quasicircles ηn which passes through p. This is given by the generating curves that
have curves connecting the degenerating circle to point pn → p converging
to linear arc intersecting orthogonally at boundary. But any neighboring
quasi-circle of ηn will converges to a η∞ with a loop singularity at p, which
is not a Jordan curve (Figure 5).
Finally, we note that if we have a degeneration point p which is not a
limit point, then there exists a neighboring quasi-circle of ηΓ that misses the
point p. Hence any O(ηΓ ) will contains some curve which is the limit of a
sequence of quasi-circles that do not pass through the point. This gives a
non-quasi-circle in O(ηΓ ).
Consider ∂G contain collapsing:
Let C denote the collapsed circle. First suppose that there exists γ ∈ Γ
such that it has a fixed point not on C. Then there are infinitely many
elements with fixed points not on C. Let γn ∈ Γn with γn → γ. Since
14
Figure 5: Degenerate singularities
γn (Fn ) → γ(G) = γ(C), we must have γ(C) either identical or disjoint. Suppose that not all fixed points of elements of Γ is contained in C. Then we
have infinitely many fixed points not in C. Take three points a, b, c fixed
points of elements of Γ with a ∈ C and b, c 6∈ C. For sufficiently large n, we
must have some βn ∈ Γn such that a, b, c is contained in three distinct disk
of the complement of βn (Fn ). This follows from the fact that the orbit FΓn
of Fn will have images with disks converging to fixed points, and since they
are distinct fixed points, we must have some disks that only contain one of
the points only. Since βn (Fn ) → β(C), so converges to a circle. But a, b, c
are contained in distinct disks bounded by circles of ∂βn (Fn ) for all large n,
which implies a, b, c must lies on β(C). Hence β(C) ∩ C 6= ∅, but they are not
identical circles, which is a contradiction. Hence all fixed points of elements
of Γ are ⊂ C, which implies ΛΓ ⊂ C. Since Γ is Schottky group, we must have
Γ is Fuchsian group of second kind. By [3], Γ is classical Schottky group, a
contradiction.
4
Proof of Theorem 1.1
Proof. (Theorem 1.1)
We proof by contradiction. Suppose that Hc < 1.
First note that by Selberg Lemma we can just assume Kleinian group to
be torsion-free.
Now note that if a Kleinian group Γ of DΓ < 1 then it must be free.
To show this, assume otherwise. Since DΓ < 1 and Γ is purely loxodromic,
it is convex-cocompact of second-kind. There exists an imbedded surface
R = ΩΓ /Γ in H3 /Γ. If R is incompressible then, subgroup π1 (R) ⊂ Γ have
Dπ1 (R) = 1 which is contradiction. If R is compressible then, we can cut along
15
compression disks. We either end with incompressible surface as before or
after finitely many steps of cutting we obtain topological ball, which implies
H3 /Γ is handle-body, hence free.
Let {Γn } ⊂ H, and Γn → Γ, with Γ a Schottky group. It follows
from Lemma 3.9, that there exists an open set O(ηΓ ) such that every element is quasi-circle of Γ. Now if Γ is non-classical Schottky group then by
Lemma 3.11, we must have every open set contains some non-quasi-circle,
in-particular we must have O(ηΓ ) contain a non-quasi-circle. But this gives
a contradiction, hence we must have Γ is a classical Schottky group.
Finally, sharpness comes from the fact that, there exists Kleinian groups
which is not free of Hausdorff dimension equal to one. Hence we have our
result.
E-mail: [email protected]
References
[1] Choquet-Bruhat, Y. (2009). General relativity and the Einstein equations. Oxford university press.
[2] Bowen, R. (1979). Hausdorff dimension of quasi-circles. I.H.E.S. 50. p.
11-25.
[3] Button, J. All Fuchsian Schottky groups are classical Schottky groups.
Geometry&Topology Mono. Vol 1. p. 117-125.
[4] Hou, Y. (2016). On smooth moduli space of Riemann surfaces.
[5] Hou, Y. (2010). Kleinian groups of small Hausdorff dimension are classical Schottky groups. I. Geometry & Topology. 14. p 473-519
[6] Hou, Y. (2014). All finitely generated Kleinian groups of
small Hausdorff dimension are classical Schottky groups.
http://arxiv.org/abs/1307.2677
[7] Phillips, R., Sarnak, P. The Laplacian for domains in hyperbolic space
and limit sets of Kleinian groups. Acta Math., 155:173-241
16
| 4math.GR
|
arXiv:1508.06610v1 [cs.DS] 26 Aug 2015
Technische Universität München
Full-text and Keyword Indexes for String
Searching
by
Aleksander Cisłak
Master’s Thesis in Informatics
Department of Informatics
Technische Universität München
Full-text and Keyword Indexes for String Searching
Indizierung von Volltexten und Keywords für Textsuche
Author: Aleksander Cisłak ([email protected])
Supervisor: prof. Burkhard Rost
Advisors: prof. Szymon Grabowski; Tatyana Goldberg, MSc
Master’s Thesis in Informatics
Department of Informatics
August 2015
Declaration of Authorship
I, Aleksander Cisłak, confirm that this master’s thesis is my own work and I have documented all sources and material used.
Signed:
Date:
ii
Abstract
String searching consists in locating a substring in a longer text, and two strings can be
approximately equal (various similarity measures such as the Hamming distance exist).
Strings can be defined very broadly, and they usually contain natural language and
biological data (DNA, proteins), but they can also represent other kinds of data such as
music or images.
One solution to string searching is to use online algorithms which do not preprocess
the input text, however, this is often infeasible due to the massive sizes of modern data
sets. Alternatively, one can build an index, i.e. a data structure which aims to speed up
string matching queries. The indexes are divided into full-text ones which operate on
the whole input text and can answer arbitrary queries and keyword indexes which store
a dictionary of individual words. In this work, we present a literature review for both
index categories as well as our contributions (which are mostly practice-oriented).
The first contribution is the FM-bloated index, which is a modification of the well-known
FM-index (a compressed, full-text index) that trades space for speed. In our approach,
the count table and the occurrence lists store information about selected q-grams in
addition to the individual characters. Two variants are described, namely one using
O(n log2 n) bits of space with O(m + log m log log n) average query time, and one with
linear space and O(m log log n) average query time, where n is the input text length
and m is the pattern length. We experimentally show that a significant speedup can be
achieved by operating on q-grams (albeit at the cost of very high space requirements,
hence the name “bloated”).
In the category of keyword indexes we present the so-called split index, which can efficiently solve the k-mismatches problem, especially for 1 error. Our implementation in the
C++ language is focused mostly on data compaction, which is beneficial for the search
speed (by being cache friendly). We compare our solution with other algorithms and
we show that it is faster when the Hamming distance is used. Query times in the order
of 1 microsecond were reported for one mismatch for a few-megabyte natural language
dictionary on a medium-end PC.
A minor contribution includes string sketches which aim to speed up approximate string
comparison at the cost of additional space (O(1) per string). They can be used in
the context of keyword indexes in order to deduce that two strings differ by at least k
mismatches with the use of fast bitwise operations rather than an explicit verification.
Acknowledgements
I would like to thank prof. Szymon Grabowski for his constant support, advice, and
mentorship. He introduced me to the academia, and I would probably not pursue the
scientific path if it were not for him. His vast knowledge and ability to explain things
are simply unmatched.
I would like to thank prof. Burkhard Rost and Tatyana Goldberg for their helpful remarks
and guidance in the field of bioinformatics.
I would like to thank the whole “Gank Incoming” Dota 2 team; the whole Radogoszcz
football and airsoft pack; Ester Cłapińska, who keeps my ego in check; Łukasz Duda,
whose decorum is still spoiled; Edwin Fung, for not letting the corporate giant consume
him; Łukasz Fuchs, who still promises the movie; Florentyna Gust, who always supported
me in difficult times, for being a cat from Korea; Jacek Krasiukianis, for introducing me
to the realm of street fighter games; Rafał Madaj, for biking tours; Jakub Przybylski,
who recently switched focus from whisky to beer (while remaining faithful to the interdisciplinary field of malt engineering), for solving world problems together; and Wojciech
Terepeta, who put up with my face being his first sight in the morning.
I am indebted to Ozzy Osbourne, Frank Sinatra, and Roger Waters for making the world
slightly more interesting. I would like to thank the developers of the free yUML software
(http://yuml.me/) for making my life somewhat easier.
Many thanks also goes to my family as well as to all the giants who lent me their shoulders
for a while.
iv
Contents
Declaration of Authorship
ii
Abstract
iii
Acknowledgements
iv
Contents
v
1 Introduction
1.1 Applications . . . . . . . . . . . . .
1.1.1 Natural language . . . . . .
1.1.2 Bioinformatics . . . . . . .
1.1.3 Other . . . . . . . . . . . .
1.2 Preliminaries . . . . . . . . . . . .
1.2.1 Sorting . . . . . . . . . . .
1.2.2 Trees . . . . . . . . . . . . .
1.2.2.1 Binary search tree
1.2.2.2 Trie . . . . . . . .
1.2.3 Hashing . . . . . . . . . . .
1.2.4 Data structure comparison .
1.2.5 Compression . . . . . . . .
1.2.5.1 Entropy . . . . . .
1.2.6 Pigeonhole principle . . . .
1.3 Overview . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
3
4
5
6
7
8
8
9
11
11
12
13
13
2 String Searching
2.1 Problem classification
2.1.1 Error metrics .
2.2 Online searching . . .
2.2.1 Exact . . . . .
2.2.2 Approximate .
2.3 Offline searching . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
15
16
17
19
19
22
24
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3 Full-text Indexes
26
v
vi
Contents
3.1
3.2
3.3
Exact . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Suffix tree . . . . . . . . . . . . . . . .
3.1.2 Suffix array . . . . . . . . . . . . . . .
3.1.2.1 Modifications . . . . . . . . .
3.1.3 Other suffix-based structures . . . . .
3.1.4 FM-index . . . . . . . . . . . . . . . .
3.1.4.1 Burrows–Wheeler Transform
3.1.4.2 Operation . . . . . . . . . . .
3.1.4.3 Efficiency . . . . . . . . . . .
3.1.5 FM-index flavors . . . . . . . . . . . .
3.1.5.1 Binary rank . . . . . . . . .
3.1.6 FM-bloated . . . . . . . . . . . . . . .
3.1.6.1 Superlinear space . . . . . .
3.1.6.2 Linear space . . . . . . . . .
Approximate . . . . . . . . . . . . . . . . . .
Word-based . . . . . . . . . . . . . . . . . . .
3.3.1 BLAST . . . . . . . . . . . . . . . . .
4 Keyword Indexes
4.1 Exact . . . . . . . . . . . . . . . . .
4.1.1 Bloom filter . . . . . . . . . .
4.1.2 Inverted index . . . . . . . .
4.2 Approximate . . . . . . . . . . . . .
4.2.1 The 1-error problem . . . . .
4.2.2 Permuterm index . . . . . . .
4.2.3 Split index . . . . . . . . . .
4.2.3.1 Complexity . . . . .
4.2.3.2 Compression . . . .
4.2.3.3 Parallelization . . .
4.2.3.4 Inverted split index
4.3 Keyword selection . . . . . . . . . .
4.3.1 Minimizers . . . . . . . . . .
4.4 String sketches . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
26
26
27
28
30
30
31
32
32
34
36
36
37
38
39
41
41
.
.
.
.
.
.
.
.
.
.
.
.
.
.
43
43
43
44
45
47
47
48
51
51
52
52
53
53
54
5 Experimental Results
56
5.1 FM-bloated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2 Split index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 String sketches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6 Conclusions
65
6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
A Data Sets
67
B Exact Matching Complexity
69
C Split Index Compression
71
Contents
vii
D String Sketches
73
E English Letter Frequency
74
F Hash Functions
75
Bibliography
76
List of Symbols
92
List of Abbreviations
94
List of Figures
94
List of Tables
97
“Why did the scarecrow win the Nobel Prize?
Because he was out standing in his field.”
Unknown
For everybody who have devoted their precious time to read this
work in its entirety
ix
Chapter 1
Introduction
The Bible, which consists of the Old and the New Testament, is composed of roughly 800
thousand words (in the English language version) [Bib]. Literary works of such stature
were often regarded as good candidates for creating concordances — listings of words
that originated from the specific work. Such collections usually included positions of
the words, which allowed the reader to learn about their frequency and context. Their
assembly was a non-trivial task that required a lot of effort. Under a rather favorable
assumption that a friar (today also referred to as a research assistant) would be able
to achieve a throughput of one word per minute, compilation (do not confuse with code
generation) for the Bible would require over thirteen thousand man-hours, or roughly one
and a half years of constant work. This naturally ignores additional efforts, for instance
printing and dissemination.
Such a listing is one of the earliest examples of a text-based data structure constructed
with the purpose of faster searches at the cost of space and preprocessing. Luckily, today
we are capable of building and using various structures in a much shorter time. With the
aid of silicone, electrons, and capable human minds, we have managed to decrease the
times from years to seconds (indexing) and from seconds to microseconds (searching).
1.1
Applications
String searching has always been ubiquitous in everyday life, most probably since the
very creation of the written word. In the modern world, we encounter text on a regular
basis — on paper, glass, rubber, human skin, metal, cement, and since the 20th century
also on electronic displays. We perform various text-based operations almost all the time,
often subconsciously. This happens in trivial situations such as looking for interesting
1
Introduction
2
news on a website on a slow Sunday afternoon, or trying to locate information in the bus
timetable on a cold Monday morning. Many familiar text-related tasks can be finished
faster thanks to computers, and powerful machines are also crucial to scientific research.
Specific areas are discussed in the following subsections.
1.1.1
Natural language
For years, the main application of computers to textual data was natural language (NL)
processing, which goes back to the work of Alan Turing in the 1950s [Tur50]. The goal was
to understand the meaning as well as the context in which the language was used. One of
the first programs that could actually comprehend and act upon English sentences was
Bobrow’s STUDENT (1967), which solved simple mathematical problems [RN10, p. 19].
The first application to text processing where string searching algorithms could really
shine was spell checking, i.e. determining whether a word is written in a correct form. It
consists in testing whether a word is present in a NL dictionary (a set of words). Such
a functionality is required since spelling errors appear relatively often due to a variety
of reasons, ranging from writer ignorance to typing and transmission errors. Research
in this area started around 1957, and the first spell checker available as an application
is believed to have appeared in 1971 [Pet80]. Today, spell checking is universal, and it
is performed by most programs which accept user input. This includes dedicated text
editors, programming tools, email clients, command-line interfaces, and web browsers.
More sophisticated approaches which try to take the context into account were also
described [CV81, Mit87], due to the fact that checking for dictionary membership is prone
to errors (e.g., mistyping “were” for “where”; Peterson [Pet86] reported that up to 16% of
errors might be undetected). Another familiar scenario is searching for words in a textual
document such as a book or an article, which allows for locating relevant fragments in a
much shorter time than by skimming through the text. Determining positions of certain
keywords in order to learn their context (neighboring words) may be also useful for
plagiarism detection (including the plagiarism of computer programs [PH89]).
With the use of approximate methods, similar words can be obtained from the NL dictionary and correct spelling can be suggested (spelling correction is usually coupled with
spell checking). This may also include proper nouns, for example in the case of shopping catalogs (relevant products) or geographic information systems (specific locations,
e.g., cities). Such techniques are also useful for optical character recognition (OCR)
where they serve as a verification mechanism [EL90]. Other applications are in security,
where it is desirable to check whether a password is not too close to a word from a dictionary [MW94a] and in data cleaning, which consists in detecting errors and duplication in
Introduction
3
data that is stored in the database [CGGM03]. String matching is also employed for preventing the registration of fraudulent websites having similar addresses, the phenomenon
known as “typosquatting” [ME10].
It may happen that the pattern that is searched for is not explicitly specified, as is the
case when we use a web search engine (i.e. we would like to find the entire website, but we
specify only a few keywords), which is an example of information retrieval. For instance,
index-based methods form an important component of the architecture of the Google
engine [BP98].
1.1.2
Bioinformatics
The biological data is commonly represented in a textual form, and for this reason it can
be searched just like any other text. Most popular representations include:
• DNA — the alphabet of four symbols corresponding to nucleobases: A, C, G, and
T. It can be extended with an additional character N, indicating that there might
be any nucleobase at a specified position. This is used, for instance, when the
sequencing method could not determine the nucleobase with a desired certainty.
Sometimes, additional information such as the quality of the read, i.e. the probability that a specific base was determined correctly, is also stored.
• RNA — four nucleobases: A, C, G, and U. Similarly to the DNA, additional information may be present.
• Proteins — 20 symbols corresponding to different amino acids (uppercase letters
from the English alphabet), with 2 additional symbols for amino acids occurring
only in some species (O and U), and 4 placeholders (B, J, X, Z) for situations where
the amino acid is ambiguous. All 26 letters from the English alphabet are used.
Computational information was an integral part of the field of bioinformatics from the
very beginning, and at the end of the 1970s there was a substantial activity in the
development of string (sequence) alignment algorithms (e.g., for RNA structure prediction) [OV03]. Alignment methods allow for finding evolutionary relationships between
genes and proteins and thus construct phylogenetic trees. Sequence similarity in proteins
is important because it may imply structural as well as functional similarity. Researchers
use tools such as BLAST [AGM+ 90], which try to match the string in question with similar ones from the database (of proteins or genomes). Approximate methods play an
Introduction
4
important role here, because related sequences often differ from one another due to mutations in the genetic material. These include point mutations, that is changes at a single
position, as well as insertions and deletions (usually called indels).
Another research area that would not thrive without computers is genome sequencing.
This is caused by the fact that sequencing methods cannot read the whole genome, but
rather they produce hundreds of gigabytes of strings (DNA reads, whose typical length
is from tens to a thousand base pairs [LLL+ 12]) whose exact positions in the genome
are not known. Moreover, the reads often contain mistakes due to the imperfection
of the sequencing itself. The goal of the computers is to calculate the correct order
using complicated statistical and string-based tools, with or without a reference genome
(the latter being called de novo sequencing). This process is well illustrated by its
name — shotgun sequencing, and it can be likened to shredding a piece of paper and
reconstructing the pieces. String searching is crucial here because it allows for finding
repeated occurrences of certain patterns [LTP+ 09, SD10].
1.1.3
Other
Other data can be also represented and manipulated in a textual form. This includes music, where we would like to locate a specific melody, especially using approximate methods
which account for slight variations or imperfections (e.g., singing out of pitch) [Gra11,
p. 77]. Another field where approximate methods play a crucial role is signal processing,
especially in the case of audio signals, which can be processed by speech recognition
algorithms. Such a functionality is becoming more and more popular nowadays, due
to the evolution of multimedia databases containing audiovisual data [Nav01]. String
algorithms can be also used in intrusion detection systems, where their goal is to identify
malicious activities by matching data such as system state graphs, instruction sequences,
or packets with those from the database [KS94, TSCV04]. String searching can be also
applied for the detection of arbitrary two-dimensional shapes in images [BB93], and yet
another application is in compression algorithms, where it is desirable to find repetitive
patterns in a similar way to sequence searching in biological data. Due to the fact that
almost any data can be represented in a textual form many other application areas exist,
see, e.g., Navarro [Nav01] for more information.
This diversity of data causes the string algorithms to be used in very different scenarios.
The pattern size can vary from a few letters (NL words) to a few hundred (DNA reads),
and the input text can be of almost arbitrary size. For instance, Google reported in 2015
that their web search index has reached over 100 thousand terabytes (1017 bytes) [goo].
Massive data is also present in bioinformatics, where the size of the genome of a single
Introduction
5
organism is often measured in gigabytes; one of the largest animal genomes belong to the
lungfish and the salamander, each occupying approximately 120 Gbp [Gre02] (i.e. roughly
28 GB, assuming that each base is coded with 2 bits). As regards proteins, the UniProt
protein database stores approximately 50 million sequences (each composed of roughly
a few hundred symbols) in 2015 and continues to grow exponentially [uni]. It was remarked recently that biological textual databases grow more quickly than the ability to
understand them [BC15].
When it comes to data of such magnitude, it is feasible only to perform an index-based
search (meaning that the data is preprocessed), which is the main focus of this thesis. It
seems most likely that the data sizes will continue to grow, and for this reason there is
a clear need for the development of algorithms which are efficient in practice.
1.2
Preliminaries
This section presents an overview of data structures and algorithms which act as building
blocks for the ones presented later, and it introduces the necessary terminology. String
searching, which is the main topic of this thesis, is described in the following chapter.
Throughout this thesis, data structures are usually approached from two angles: theoretical, which concentrates on the worst-case space and query time, and a practical one. The
latter focuses on performance in real-world scenarios, and it is often heuristically oriented
and focused on cache utilization and reducing slow RAM access. It is worth noting that
state-of-the-art theoretical algorithms sometimes perform very poor in practice because
of certain constant factors which are ignored in the analysis. Moreover, they might not
be even tested or implemented at all. On the other hand, a practical evaluation depends
heavily on the hardware (peculiarities of the CPU cache, instruction prefetching, etc),
properties of the data sets used as input, and most importantly on the implementation.
Moffat and Gog [MG14] provided an extensive analysis of experimentation in the field
of string searching, and they pointed out various caveats. These include for instance a
bias towards certain repetitive patterns when the patterns are randomly sampled from
the input text, or the advantage of smaller data sets which increase the probability that
(at least some of) the data would fit into the cache.
The theoretical analysis of the algorithms is based on the big O family of asymptotic notations, including O, Ω, Θ, and the relevant lower case counterparts (we assume that the
reader is familiar with these tools and with complexity classes). Unless stated otherwise,
the complexity analysis refers to the worst-case scenario, and all logarithms are assumed
Introduction
6
to be base 2 (this might be also stated explicitly as log2 ). When we state that the complexity or the average or the worst case is equal to some value, we mean the running
time of the algorithm. On the other hand, if the time or space is explicitly mentioned,
the word “complexity” might be omitted. Array, string, and vector indexes are always 0based and they are assumed to be contiguous, and collection indexes are 1-based (e.g., a
collection of strings s1 , . . . , sn ). We consider a standard hierarchical memory model with
RAM and a faster CPU cache, and we take for granted that the data always fits into the
main memory, i.e. disk input/output (I/O) is ignored. Moreover, we assume that the size
of the data does not exceed 232 bytes, which means that it is sufficient for each pointer
or counter to occupy 32 bits (4 bytes). Sizes that are given in kilobytes and megabytes
are indicated with abbreviations KB and MB, which refer to standard computer science
quantities, i.e. 210 (rather than 1,000) and 220 .
1.2.1
Sorting
Sorting consists in ordering n elements from a given set S in such a way that the following
holds: ∀i ∈ [0, n − 1) : S[i] 6 S[i + 1], that is the smallest element is always in front.
In reverse sorting, the highest element is in front, and the inequality sign is reversed.
Popular sorting methods include the heapsort and the mergesort with O(n log n) worstcase time guarantees. Another well-known algorithm is the quicksort with average time
O(n log n) (although the worst case is equal to O(n2 )), which is known to be 2–3 times
faster in practice than both heapsort and mergesort [Ski98]. There also exist algorithms
linear in n which can be used in certain scenarios, for instance the radix sort for integers
with time complexity O(wn) (or O(n(w/ log n)) for a byte-wide radix), where w is the
machine word size.
When it comes to sorting n strings of average length m, a comparison sorting method
would take O(n log nm) time (assuming that comparing two strings is linear in time).
Alternatively, we could obtain an O(nm) time bound by sorting each letter column with
a sorting method which is linear for a fixed alphabet (essentially performing a radix
sort), e.g., using a counting sort. Moreover, we can even achieve an O(n) complexity by
building a trie with lexicographically ordered children at each level and performing a preorder, depth-first search (DFS) (see the following subsections for details). When it comes
to suffix sorting (i.e. sorting all suffixes of the input text), dedicated methods which do
not have a linear time guarantee are often used due to reduced space requirements or
good practical performance [MF04, PST07]. Recently, linear methods which are efficient
in practice have also been described [Non13].
Introduction
1.2.2
7
Trees
A tree contains multiple nodes that are connected with each other, with one (top-most)
node designated as the root. Every node contains zero or more children, and a tree is an
undirected graph where any two vertexes are connected by exactly one path (there are no
cycles). Further terminology which is relevant to trees is as follows [CLRS09, Sec. B.5]:
• A parent is a neighbor of a child and it is located closer to the root (and vice
versa).
• A sibling is a node which shares the same parent.
• Leaves are the nodes without children and in the graphical representation they
are always shown at the bottom of the diagram. The leaves are also called external
nodes, and internal nodes are all nodes other than leaves.
• Descendants are the nodes located anywhere in the subtree rooted by the current
node, and ancestors are the nodes anywhere on the path from the root (inclusive)
to the current node. Proper descendants and ancestors exclude the current node.
If n1 is an ancestor of n2 , then n2 is a descendant of n1 , and vice versa.
• The depth of a node is the length of the path from this node to the root.
• The height of a tree is the longest path from the root to any leaf (i.e. the depth
of the deepest node).
The maximum number of children can be limited for each node. Many structures are
binary trees, which means that every node has at most two children, and a generic term
is a k-ary tree (or multiary for k > 2). A full (complete) k-ary tree is a structure where
every node has exactly 0 (in the case of leaves) or k (in the case of internal nodes)
children. A perfect tree is a tree where all leaves have the same depth. A historical note:
apparently, a binary tree used to be called a bifurcating arborescence in the early years
of computer science [Knu97, p. 363]. A balanced (height-balanced, self-balancing) tree
is a tree whose height is maintained with respect to its total size irrespective of possible
updates and deletions. The height of a balanced binary tree is logarithmic, i.e. O(log n)
(logk n for a k-ary tree). It is often desirable to maintain such a balance because otherwise
a tree may lose its properties (e.g., worst-case search complexity). This is caused by the
fact that the time complexity of various algorithms is proportional to the height of the
tree.
There exist many kinds of trees, and they are characterized by some additional properties
which make them useful for a certain purpose.
8
Introduction
1.2.2.1
Binary search tree
The binary search tree (BST) is used for determining a membership in the set or for
storing key-value pairs. Every node stores one value V ; the value of its right child is
always bigger than V , and the value of its left child is always smaller than V . The lookup
operation consists in traversing the tree towards the leaves until either the value is found
or there are no more nodes to process, which indicates that the value is not present. The
BST is often used to maintain a collection of numbers, however, the values can also be
strings (they are ordered alphabetically), see Figure 1.1. It is crucial that the BST is
balanced. Otherwise, in the scenario where every node had exactly one child, its height
would be linear (basically forming a linked list) and thus the complexity for the traversal
would degrade from O(log n) to O(n). The occupied space is clearly linear (there is one
node per value) and the preprocessing takes O(n log n) time because each insertion costs
O(log n).
Karen
Alan
Bob
Tom
Erin
Sasha
Zelda
Figure 1.1: A binary search tree (BST) storing strings from the English alphabet.
The value of the right child is always bigger than the value of the parent, and the value
of the left child is always smaller than the value of the parent.
1.2.2.2
Trie
The trie (digital tree) [Fre60] is a tree in which the position of a node (more specifically, a
path from the root to the node) describes the associated value, see Figure 1.2. The nodes
often store IDs or flags which indicate whether a given node has a word, which is required
because some nodes may be only intermediary and not associated with any value. The
values are often strings, and the paths may correspond to the prefixes of the input text. A
trie supports basic operations such as searching, insertion, and deletion. For the lookup,
we check whether each consecutive character from the query is present in the trie while
moving towards the leaves, hence the search complexity is directly proportional to the
length of the pattern. In order to build a trie, we have to perform a full lookup for each
9
Introduction
word, thus the preprocessing complexity is equal to O(n) for words of total length n.
The space is linear because there is at most one node per input character.
t
t
A (5)
e
o
te
n
ten (1)
e
tea (2)
A
i
i
n
to (4)
d
in (6)
n
ted (3)
inn (7)
Figure 1.2: A trie, which is one of the basic structures used in string searching, constructed for strings from the set {A, in, inn, tea, ted, ten, to}. Each edge corresponds
to one character, and the strings are stored implicitly (here shown for clarity). Additional information such as IDs (here shown inside the parentheses) is sometimes kept
in the nodes.
Various modifications of the regular trie exist — an example could be the Patricia
trie [Mor68], whose aim is to reduce the occupied space. The idea is to merge every
node which has no siblings with its parent, thus reducing the total number of nodes,
and resulting edge labels include the characters from all edges that were merged (the
complexities are unchanged).
1.2.3
Hashing
A hash function H transforms data of arbitrary size into the data of fixed size. Typical
output sizes include 64, 128, 256, and 512 bits. The input can be in principle of any
type, although hash functions are usually designed so that they work well for a particular
kind of data, e.g., for strings or for integers. Hash functions often have certain desirable
properties such as a limited number of collisions, where for two chunks of data d1 and d2 ,
the probability that H(d1 ) = H(d2 ) should be relatively low (e.g., H is called universal
if P r(H(d1 ) = H(d2 )) 6 1/n for an n-element hash table [CW77]). There exists a group
of cryptographic hash functions, which offer certain guarantees regarding the number
of collisions. They also provide non-reversibility, which means that it is hard (in the
Introduction
10
mathematical sense, for example the problem may be NP-hard) to deduce the value of
the input string from the hash value. Such properties are provided at the price of reduced
speed, and for this reason cryptographic hash functions are usually not used for string
matching.
A perfect hash function guarantees no collisions, e.g., the two-level FKS scheme with O(n)
space [FKS84]. All keys have to be usually known beforehand, although dynamic perfect
hashing was also considered [DKM+ 94]. A minimal perfect hash function (MPHF)
uses every bucket in the hash table, i.e. there is one value per bucket (the lower space
bound for describing an MPHF is equal to roughly 1.44n bits for n elements [BBD09]).
The complexity of a hash function is usually linear in the input length, although it is
sometimes assumed that it takes constant time.
A hash function is an integral part of a hash table HT , which is a data structure that
associates the values with buckets based on the key, i.e. the hash of the value. This can
be represented with the following relation: HT [H(v)] = v for any value v. Hash tables
are often used in string searching because they allow for quick membership queries, see
Figure 1.3. The size of the hash table is usually much smaller than the number of
all possible hash values, and it is often the case that a collision occurs (the same key
is produced for two different values). There exist various methods of resolving such
collisions, and the most popular ones are as follows:
• Chaining — each bucket holds a list of all values which hashed to this bucket.
• Probing — if a collision occurs, the value is inserted into the next unoccupied
bucket. This may be linear probing, where the consecutive buckets are scanned
linearly until an empty bucket is found, or quadratic probing, where the gaps
between consecutive buckets are formed by the results of a quadratic polynomial.
• Double hashing — gaps between consecutive buckets are determined by another
hash function. A simple approach could be for instance to locate the next bucket
index i using the formula i = H1 (v)+iH2 (v) mod |HT | for any two hash functions
H1 and H2 .
In order to resolve the collisions, the keys have to be usually stored as well. The techniques which try to locate an empty bucket (as opposed to chaining) are referred to as
open addressing. A key characteristic of the hash table is its load factor (LF), which
is defined as the number of entries divided by the number of buckets. Let us note that
0 6 LF 6 1.0 for open addressing (the performance degrades rapidly as LF → 1.0),
however, in the case of chaining it holds that 0 6 LF 6 n for n entries.
11
Introduction
keys
hash
function
buckets
00
John Smith
Lisa Smith
Sandra Dee
01
521-8976
02
521-1234
03
:
:
13
14
521-9655
15
Figure 1.3: A hash table for strings; reproduced from Wikimedia Commons1 .
1.2.4
Data structure comparison
In the previous subsections, we introduced data structures which are used by more sophisticated algorithms (described in the following chapters). Still, they can be also used
on their own for exact string searching. In Figure 1.1 we present a comparison of their
complexities together with a linear, direct access array. It is to be noted that even though
the worst case of a hash table lookup is linear (iterating over one bucket which stores all
the elements), it is extremely unlikely, and any popular hash function offers reasonable
guarantees against building such a degenerate hash table.
Data structure
Array
Balanced BST
Hash table
Trie
Lookup
O(n)
O(log n)
O(m) avg, O(n) worst-case
O(m)
Preprocessing
O(1)
O(n log n)
O(n)
O(n)
Space
O(n)
O(n)
O(n)
O(n)
Table 1.1: A comparison of the complexities of basic data structures which can be
used for exact string searching. Here, we assume that string comparison takes constant
time.
1.2.5
Compression
Compression consists in representing the data in an alternative (encoded) form with
the purpose of reducing the size. After compression, the data can be decompressed
(decoded) in order to obtain the original representation. Typical applications include
reducing storage sizes and saving bandwidth during transmission. Compression can be
1
Jorge Stolfi, available at http://en.wikipedia.org/wiki/File:Hash_table_3_1_1_0_1_
0_0_SP.svg, CC A-SA 3.0.
12
Introduction
either lossless or lossy, depending on whether the result of decompression matches the
original data. The former is useful especially when it comes to multimedia (frequently
used by domain-specific methods such as those based on human perception of images),
where the lower quality may be acceptable or even indiscernible, and storing the data in
an uncompressed form is often infeasible. For instance, the original size of a 2-hour Full
HD movie with 32 bits per pixel and 24 frames per second would amount to more than
one terabyte. Data that can be compressed is sometimes called redundant.
One of the most popular compression methods is character substitution, where the selected symbols (bit q-grams) are replaced with ones that take less space. A classic algorithm is called Huffman coding [Huf52], and it offers an optimal substitution method.
Based on q-gram frequencies, it produces a codebook which maps more frequent characters to shorter codes, in such a way that every code is uniquely decodable (e.g., 00 and
1 are uniquely decodable, but 00 and 0 are not). For real-world data, Huffman coding
offers compression rates that are close to the entropy (see the following subsection), and
it is often used as a component of more complex algorithms. We refer the reader to
Salomon’s monograph [Sal04] for more information on data compression.
1.2.5.1
Entropy
We can easily determine the compression ratio by taking the size (a number of occupied
bits) of the original data and dividing it by the size of the compressed data, r = n/nc .
It may seem that the following should hold: r > 1, however, certain algorithms might
actually increase the data size after compressing it when operating on an inconvenient
data set (which is of course highly undesirable). A related problem is how to determine
the “compressibility” of the data, i.e. the optimal compression ratio (the highest r). This
brings us to the notion of entropy, sometimes also called Shannon’s entropy after the
name of the author [Sha48]. It describes the amount of information which is contained
in a message, and in the case of strings it determines the average number of bits which
is required in order to encode an input symbol under a specified alphabet and frequency
distribution. This means that the entropy describes a theoretical bound on data compression (one that cannot be exceeded by any algorithm). Higher entropy means that
it is more difficult to compress the data (e.g., when multiple symbols appear with equal
frequency). The formula is presented in Figure 1.4.
E = −K
n
X
pi log pi .
i=1
Figure 1.4: The formula for Shannon’s entropy, where E is the entropy function, pi
is the probability that symbol i occurs, and K is any constant.
Introduction
13
A variation of entropy which is used in the context of strings is called a k-th order
entropy. It takes the context of k preceding symbols into account and it allows for the
use of different codes based on this context (e.g., ignoring a symbol c2 which always
appears after the symbol c1 ). Shannon’s entropy corresponds to the case of k = 0
(denoted as H0 , or Hk in general). When we increase the k value, we also increase the
theoretical bound on compressibility, although the size of the data required for storing
context information may at some point dominate the space [Gog11].
1.2.6
Pigeonhole principle
Let us consider a situation where we have x buckets and n items which are to be positioned inside those buckets. The pigeonhole principle (often also called Dirichlet principle) states that if n > x, then at least one of the buckets must store more than one
item. The name comes from an intuitive representation of the buckets as boxes and
items as pigeons. Despite its simplicity, this principle has been successfully applied to
various mathematical problems. It is also often used in computer science, for example to
describe the number of collisions in a hash table. Later we will see that the pigeonhole
principle is also useful in the context of string searching, especially when it comes to
string partitioning and approximate matching.
1.3
Overview
This thesis is organized as follows:
• Chapter 2 provides an overview of the field of string searching, deals with the
underlying theory, introduces relevant notations, and discusses related work in the
context of online search algorithms.
• Chapter 3 includes related work and discusses current state-of-the-art algorithms
for full-text indexing as well as our contribution to this area.
• Chapter 4 does the same for keyword indexes.
• Chapter 5 describes the experimental setup and presents practical results.
• Chapter 6 contains conclusions and pointers to the possible future work.
• Appendix A offers information regarding the data sets which were used for the
experimental evaluation.
Introduction
14
• Appendix B discusses the complexity of exact string comparison.
• Appendix C discusses the q-gram-based compression of the split index (Section 4.2.3) in detail.
• Appendix D contains experimental results for string sketches (Section 4.4) when
used for the alphabet with uniform letter frequencies.
• Appendix E presents the frequencies of English alphabet letters.
• Appendix F contains Internet addresses where the reader can obtain the code for
hash functions which were used to obtain experimental results for the split index
(Section 5.2).
Chapter 2
String Searching
This thesis deals with strings, which are sequences of symbols over a specified alphabet.
The string is usually denoted as S, the alphabet as Σ, and the length (size) as n or |S|
for strings, and σ or |Σ| for the alphabet. An arbitrary string S is sometimes called
a word (which is not to be confused with the machine word, i.e. a basic data unit in
the processor), and it is defined over a given alphabet Σ, that is S belongs to the set
of all words specified over the said alphabet, S ∈ Σ∗ . Both strings and alphabets are
assumed to be finite and well-defined, and alphabets are totally ordered. A string with a
specified value is written with the teletype font, as in abcd. The brackets are usually used
to indicate the character at a specified position and the index is 0-based, for instance
if string S = text, then S[1] = e. A substring (sometimes referred to as a factor) is
written as S[i0 , i1 ] (an inclusive range); for the previous example, S[1, 2] = ex, and a
single character is a substring of length 1 (usually denoted with c). The last character is
indicated with -1 and P = P [0, −1]. S1 ⊂ S2 indicates that the string S1 is a substring
of S2 (conversely, S1 6⊂ S2 indicates that S1 is not a substring of S2 ). The subscripts
are usually used to distinguish multiple strings, and two strings may be concatenated
(merged into one), recorded as S = S1 S2 or S = S1 + S2 , in which case |S| = |S1 | + |S2 |.
Removing one substring from another is indicated with the subtraction sign, S = S1 −S2 ,
provided that S2 ⊂ S1 , and as a result |S| = |S1 | − occ · |S2 | for occ occurrences of S2
in S1 . The equality sign indicates that the strings match exactly, which means that the
following relation always holds: S1 = S2 → |S1 | = |S2 |.
String searching, or string matching, refers to locating a substring (a pattern, P , or a
query, Q, with length m) in a longer text T . The textual data T that is searched is
called the input (input string, input text, text, database), and its length is denoted by n
(e.g., O(n) indicates that the complexity is linear with respect to the size of the original
data). The pattern is usually much smaller than the input, often in multiple orders of
15
String Searching
16
magnitude (m n). Based on the position of the pattern in the text, we write that P
occurs in T with a shift s, i.e. P [0, m − 1] = T [s, s + m − 1] ∧ s 6 n − m. As mentioned
before, the applications may vary (see Section 1.1), and the data itself can come from
many different domains. Still, string searching algorithms can operate on any text while
being oblivious to the actual meaning of the data. The field concerning the algorithms
for string processing is sometimes called stringology.
Two important notions are prefixes and suffixes, where the former is a substring S[0, i]
and the latter is a substring S[i, |S| − 1] for any 0 6 i < |S|. Let us observe at this
point that every substring is a prefix of one of the suffixes of the original string, as well
as a suffix of one of the prefixes. This simple statement is a basis for many algorithms
which are described in the following chapters. A proper prefix or suffix is not equal to
the string itself. The strings can be lexicographically ordered, which means that they are
sorted according to the ordering of the characters from the given alphabet (e.g., for the
English alphabet, letter a comes before b, b comes before c, etc). Formally, S1 < S2 for
two strings of respective lengths n1 and n2 , if ∃s : ∀i ∈ [0, s) : S1 [i] = S2 [i] ∧ S1 [s] <
S2 [s] ∧ 0 6 s < min(n1, n2) or n1 < n2 ∧ S1 = S2 [0, n1 − 1] [CLRS09, p. 304]. When
it comes to strings, we often mention q-grams and k-mers, which are lists of contiguous
characters (strings or substrings). The former is usually used in general terms and the
latter is used for biological data (especially DNA reads). A 5-gram or 5-mer is a q-gram
of length 5.
2.1
Problem classification
The match between the pattern and the substring of the input text is determined according to the specified similarity measure, which allows us to divide the algorithms into
two categories: exact and approximate. The former refers to direct matching, where the
length as well as all characters at the corresponding positions must be equal to each
another. This relation can be represented formally for two strings S1 and S2 of the same
length n, which are equal if and only if ∀i ∈ [0, n) : S1 [i] = S2 [i] (or simply, S1 = S2 ). In
the case of approximate matching, the similarity is measured with a specified distance,
also called an error metric, between the two strings. It is to be noted that the word approximation is not used here strictly in the mathematical sense, since approximate search
is actually harder than the exact one when it comes to strings [Nav01]. In general, given
two strings S1 and S2 , the distance D(S1 , S2 ) is the minimum cost of edit operations
that would transform S1 into S2 or vice versa. The edits are usually defined as a finite,
well-defined set of rules E = {r : r(S) = S 0 }, and each rule can be associated with a
different cost. When error metrics are used, the results of string matching are limited to
String Searching
17
those substrings which are close to the pattern. This is defined by the threshold k, that
is we report all substrings s for which D(s, P ) 6 k. For metrics with fixed penalties for
errors, k is called the maximum allowed number of errors. This value may depend both
on the data set and the pattern length, for instance for spell checking a reasonable number of errors is higher for longer words. It should hold that k < m, since otherwise the
pattern could match any string, and k = 0 corresponds to the exact matching scenario.
See Subsection 2.1.1 for detailed descriptions of the most popular error metrics.
The problem of searching (also called a lookup) can vary depending on the kind of answer
that is provided. This includes the following operations:
• Match — determining the membership, i.e. deciding whether P ⊂ T (a decision
problem). When we consider the search complexity, we usually implicitly mean the
match query.
• Count — stating how many times P occurs in T . This refers to the cardinality
of the set containing all indexes i s.t. T [i, i + m − 1] is equal to P . Specific values
of i are ignored in this scenario. The time complexity of the count operation often
depends on the number of occurrences (denoted with occ).
• Locate — reporting all occurrences of P in T , i.e. returning all indexes i s.t.
T [i, i + m − 1] is equal to P .
• Display — showing k characters which are located before and after each match,
that is for all aforementioned indexes i we display substrings T [i − k, i − 1] and
T [i + m, i + m + k − 1]. In the case of approximate matching, it might refer to
showing all text substrings or keywords s s.t. D(s, P ) 6 k.
String searching algorithms can be also categorized based on whether the data is preprocessed. One such classification adapted from Melichar et al. [MHP05, p. 8] is presented
in Table 2.1. Offline searching is also called index-based searching because we preprocess
the text and build a data structure which is called an index. This is opposed to online
searching, where no preprocessing of the input text takes place. For detailed descriptions
of the examples from these classes, consult Chapters 3 and 4 (offline) and Section 2.2
(online).
2.1.1
Error metrics
The motivation behind error metrics is to minimize the score between the strings which
are somehow related to each other. Character differences that are more likely to occur
18
String Searching
Text prepr.
No
No
Yes
Yes
Pattern prepr.
No
Yes
No
Yes
Algorithm type
Online
Online
Offline
Offline
Examples
Naive, dynamic programming
Pattern automata, rolling hash
Index-based methods
Signature methods
Table 2.1: Algorithm classification based on whether the data is preprocessed.
should carry a lower penalty depending on the application area, for instance in the case
of DNA certain mutations appear in the real world much more often than others. The
most popular metrics include:
• Hamming distance — relevant for two strings of equal length n, calculates the
number of differing characters at corresponding positions (hence it is sometimes
called the k-mismatches problem). Throughout this thesis we denote the Hamming
distance with Ham, and given that |S1 | = |S2 | = n, Ham(S1 , S2 ) = |E|, where
E = {i : i ∈ [0, n) ∧ S1 [i] 6= S2 [i]}, and Ham(S1 , S2 ) 6 n. Without preprocessing,
calculating the Hamming distance takes O(n) time. Applications of the Hamming distance include, i.a., bioinformatics [KCO+ 01, LSS01], biometrics [DFM98],
cheminformatics [Flo98], circuit design [GLPS97], and web crawling [MJDS07].
• Levenshtein distance — measures the minimum number of edits, here defined as
insertions, deletions, and substitutions. It was first described in the context of
error correction for data transmission [Lev66]. It must hold that Lev(S1 , S2 ) 6
max(|S1 |, |S2 |). The calculation using the dynamic programming algorithm takes
O(|S1 ||S2 |) time using O(min(|S1 |, |S2 |)) space (see Subsection 2.2.2). Ukkonen
[Ukk85a] recognized certain properties of the DP matrix and presented an algorithm
with O(k min(|S1 |, |S2 |)) time for k errors, and an approximation algorithm in a
near-linear time was also described [AKO10]. Levenshtein distance is sometimes
called simply the edit distance. When the distance for approximate matching is
not explicitly specified, we assume the Levenshtein distance.
• Other edit distances. These may allow only a subset of edit actions, e.g., longest
common subsequence (LCS) which is restricted to indels [NW70] or the Episode
distance with deletions [DFG+ 97]. Another approach is to introduce additional
actions; examples include the Damerau–Levenshtein distance which counts a transposition as one edit operation [Bar07], a distance which allows for matching one
character with two and vice versa (specifically designed for OCR) [CB08], or a
distance which has weights for substitutions based on the probability that a user
may mistype one character for another [BM00].
String Searching
19
• Sequence alignment — there may exist gaps (other characters) in between substrings of S1 and S2 . Moreover, certain characters may match each other even
though they are not strictly equal. The gaps themselves (e.g., their lengths or
positions [NR03, AE86]) as well as the inequality between individual characters
are quantified. The score is calculated using a similarity matrix which is constructed based on statistical properties of the elements from the domain in question
(e.g., the BLOSUM62 matrix for the sequence alignment of proteins [Edd04]). The
problem can be also formulated as (δ, α)-matching, where the width of the gaps
is at most α and for a set P of positions of corresponding characters, it should
hold that ∀(i1 , i2 ) ∈ P : |S1 [i1 ] − S2 [i2 ]| 6 δ [CIM+ 02]. This means that absolute values of numerical differences between certain characters cannot exceed a
specified threshold. Sequence alignment is a generalization of the edit distance,
and it can be also performed for multiple sequences, although this is known to be
NP-complete [WJ94].
• Regular expression matching — the patterns may contain certain metacharacters with various meanings. These can specify ranges of characters which can
match at certain positions or use additional constructs such as the wildcard symbol
which matches 0 or more consecutive characters of any type.
2.2
Online searching
In this section, we present selected algorithms for online string searching, and we divide
them into exact and approximate ones. Online algorithms do not preprocess the input
text, however, the pattern may be preprocessed. We assume that the preprocessing time
complexity is equal to O(1), and the time required for pattern preprocessing is subsumed
under search complexity, which means that we consider a scenario where the patterns
are not known beforehand. Search time refers to the match query.
2.2.1
Exact
Faro and Lecroq [FL13] provided a survey on online algorithms for exact matching and
remarked that over 80 algorithms have been proposed since the 1970s. They categorized
the algorithms into the following three groups:
• Character comparisons
• Automata
String Searching
20
• Bit parallelism
The naive algorithm attempts to match every possible substring of T of length m with the
pattern P . This means that it iterates from left to right and checks whether T [i, m−1] =
P for each 0 6 i 6 n−m. Right to left iteration would be also possible and the algorithm
would report the same results. Time complexity is equal to O(nm) in the worst case
(although to O(n) on average, see Appendix B for more information), and there is no
preprocessing or space overhead.
Even without text preprocessing, the performance of the naive algorithm can be improved
significantly by taking advantage of the information provided by the mismatches between
the text and the pattern. Classical solutions for single-pattern matching include the
Knuth–Morris–Pratt (KMP) [KMP77] and the Boyer–Moore (BM) [BM77] algorithm.
The KMP uses information regarding the characters that appear in the pattern in order
to avoid repeated comparisons known from the naive approach. It reduces the time
complexity from O(nm) to O(n + m) in the worst case at the cost of O(m) space. When
a mismatch occurs at position i in the pattern (P [i] 6= T [s + i]), the algorithm shifts P
by i − l, where l is the length of the longest proper prefix of PS = P [0, i − 1] which is
also a suffix of PS (instead of just 1 position), and it starts matching from the position
i = l instead of i = 0. Information regarding PS is precomputed and stored in a table of
size m. Let us observe that the algorithm does not skip any characters from the input
string. Interestingly, Baeza-Yates and Navarro [BYN04] reported that in practice the
KMP algorithm is roughly two times slower than the brute-force search (although this
depends on the alphabet size).
The BM algorithm, on the other hand, omits certain characters from the input. It
begins the matching from the end of the pattern and allows for forward jumps based on
mismatches. Thanks to the preprocessing, the size of each shift can be determined in
constant time. One of the two rules for jumping is called a bad character rule, which,
given that P [i] 6= T [s + i] ∧ T [s + i] = c, aligns T [s + i] with the rightmost occurrence
of c in P (P [j] = c, where j < i), or shifts the pattern by m if c ∈
/ P . The other rule
is a complex good suffix rule, whose description we omit here, and which is also not a
part of the Boyer–Moore–Horspool (BMH) [Hor80] algorithm which uses only the bad
character rule — this is because the good suffix rule requires extra cost to compute and
it is often not practical. The worst-case time complexity of the BM algorithm is equal to
O(nm + σ), with O(n/ min(m, σ) + m + σ) average (the same holds for BMH), and the
average number of comparisons is equal to roughly 3n [Col94]. This can be improved to
achieve a linear time in the worst case by introducing additional rules [Gal79].
21
String Searching
One of the algorithms developed later is the Rabin–Karp (RK) [KR87] algorithm. It
starts with calculating the hash value of the pattern in the preprocessing stage, and then
compares this hash with every substring of the text, sliding over it in a similar way to the
naive algorithm. Character-wise verification takes place only if two hashes are equal to
each other. The trick is to use a hash function which can be computed in constant time for
the next substring given its output for the previous substring and the next character, a socalled rolling hash [Kor06], viz. H(T [s, s+m−1], T [s+m]) = H(T [s+1, s+m]). A simple
n−1
P
example would be to simply add the values of all characters, H(S) =
S[i]. There
i=0
exist other functions such as the Rabin fingerprint [Rab81], which treats the characters as
n
P
ci xi−1 . The
polynomial variables ci , and the indeterminate x is a fixed base, R(S) =
i=1
RK algorithm is suitable for multiple-pattern matching, since we can quickly compare the
hash of the current substring with the hashes of all patterns using any efficient set data
structure. In this way, we obtain the average time complexity of O(n + m) (assuming
that hashing takes linear time), however, it is still equal to O(nm) in the worst case
when the hashes do match and verification is required.
Another approach is taken by the Aho–Corasick [AC75] algorithm. It builds a finite state
machine (FSM), i.e. an automaton which has a finite number of states. The structure
of the automaton resembles a trie and it contains edges between certain nodes which
represent the transitions. It is constructed from the queries and attempts to match all
queries at once when sliding over the text. The transitions indicate the next possible
pattern which can be still fully matched after a mismatch at a specified position occurs.
The search complexity is equal to O(n log σ + m + z), which means that it is linear with
respect to the input length n, the length of all patterns (m, for building the automaton),
and the number of occurrences z.
An example of a bit-parallel algorithm is the shift-or algorithm by Baeza-Yates and
Gonnet [BYG92], which aims to speed up the comparisons. The pattern length should
be smaller than the machine word size, which is usually equal to 32 or 64 bits. During the
preprocessing, a mismatch mask M is computed for each character c from the alphabet,
where M [i] = 0 if P [i] = c and M [i] = 1 otherwise. Moreover, we maintain a state
mask R, initially set to all 1s, which holds information about the matches so far. We
proceed in a similar manner to the naive algorithm, trying to match the pattern with
every substring, but instead of character-wise comparisons we use bit operations. At each
step, we shift the state mask to the left and OR it with the M for the current character
T [i]. A match is reported if the most significant bit of R is equal to 0. Provided that
m 6 w, the time complexity is equal to O(n) and the masks occupy O(σ) space.
Based on the practical evaluation, Faro and Lecroq [FL13] reported that there is no
superior algorithm and the effectiveness depends heavily on the size of the pattern and
String Searching
22
the size of the alphabet. The differences in performance are substantial — algorithms
which are the fastest for short patterns are often among the slowest for long patterns,
and vice versa.
2.2.2
Approximate
In the following paragraphs, we use δ to denote the complexity of calculating the distance
function between two strings (consult Subsection 2.1.1 for the description of the most
popular metrics). Navarro [Nav01] presented an extensive survey regarding approximate
online matching, where he categorizes the algorithms into four categories which resemble
the ones presented for the exact scenario:
• Dynamic programming
• Automata
• Bit parallelism
• Filtering
The naive algorithm works in a similar manner to the one for exact searching, that is it
compares the pattern with every possible substring of the input text. It forms a generic
idea which can be adapted depending on the edit distance which is used, and for this
reason the time complexity is equal to O(nδ).
The oldest algorithms are based on the principle of dynamic programming (DP). This
means that they divide the problem into subproblems — these are solved and their
answers are stored in order to avoid recomputing these answers (i.e. it is applicable when
the subproblems overlap) [CLRS09, p. 359]. One of the most well-known examples is the
Needleman–Wunsch (NW) [NW70] algorithm, which was originally designed to compare
biological sequences. Starting from the first character of both strings, it successively
considers all possible actions (insertion, (mis)match, deletion) and constructs a matrix
which holds all alignment scores. It calculates the global alignment, and it can use a
substitution matrix which specifies alignment scores (penalties). The situation where the
scores triplet is equal to (−1, 1, −1) (for gaps, matches, and mismatches, respectively)
corresponds directly to the Levenshtein distance (consult Figure 2.1). The NW method
can be invoked with the input text as one string and the pattern as the other.
A closely related variation of the NW algorithm is the Smith–Waterman (SW) [SW81]
algorithm, which can also identify local (and not just global) alignments by not allowing
negative scores. This means that the alignment does not have to cover the entire length
23
String Searching
of the text, and it is therefore more suitable for locating a pattern as a substring. Both
algorithms can be adapted to other distance metrics by manipulating the scoring matrix,
for example by assigning infinite costs in order to prohibit certain operations. The time
complexity of the NW and SW approaches is equal to O(nm) and it possible to calculate
them using O(min(n, m)) space [Hir75]. Despite their simplicity, both methods are still
popular for sequence alignment because they might be relatively fast in practice and they
report the true answer to the problem, which is crucial when the quality of an alignment
matters.
t
a
x
i
0
-1
-2
-3
-4
t
-1
1
0
-1
-2
e
-2
0
0
-1
-2
x
-3
-1
-1
1
0
t
-4
-2
-2
0
0
Figure 2.1: Calculating an alignment with Levenshtein distance using the Needleman–
Wunsch (NW) algorithm. We follow the path from the top-left to the bottom-right
corner selecting the highest possible score (underlined); the optimal global alignment
is as follows: text → taxi (no gaps).
Multiple other dynamic programming algorithms were proposed over the years, and they
gradually tightened the theoretical bounds. The difference lies mostly in their flexibility,
that is a possibility of being adapted to other distance metrics, as well as practical performance. Notable results for the edit distance include the Chang–Lampe [CL92] algorithm
√
with O(kn/ σ) average time using O(m) space, and the Cole–Hariharan [CH02] algorithm with the worst-case time O(n + m + k c n/m) with c = 3 for non-k-break periodic
patterns and c = 4 otherwise, taking O(m + os ) space, where os refers to occurrences
of certain substrings of the pattern in the text (the analysis is rather lengthy). For the
LCS metric, Grabowski [Gra14] provided the algorithm with O(nm log log n/ log2 n) time
bound and linear space.
A significant achievement in the automata category is the Wu–Manber–Myers [WMM96]
algorithm that uses the Four Russians technique, which consists in partitioning the matrix into fixed-size blocks, precomputing the values for each possible block, and then using
a lookup table. It implicitly constructs the automaton, where each state corresponds to
the values in the DP matrix. They obtain an O(kn/ log n) expected time bound using
O(n) space. As regards bit parallelism, Myers [Mye99] presented the calculation of the
DP matrix in O(dk/wen) average time.
An important category is formed by the filtering algorithms, which try to identify parts
of the input text where it is not possible to match any substrings with the pattern. After
String Searching
24
parts of the text are rejected, a non-filtering algorithm is used on the remaining parts.
Numerous filtering algorithms have been proposed, and one of the most significant is the
Chang–Marr [CM94] algorithm with the time bound of O(n(k + logσ m)/m) for the error
√
level α when it holds that α < 1 − e/ σ (for very large σ).
As regards the k-mismatches problem, a notable example is the Amir–Lewenstein–
√
Porat [ALP04] algorithm which can answer the locate query in O(n k log k) time. This
p
was refined to O(n+n k/w log k) in the word RAM model, where w = Ω(log n) [FG09].
Recently, Clifford et al. [CFP+ 15] described an algorithm with search time complexity
O(nk 2 log k/m + n polylog m).
2.3
Offline searching
An online search is often infeasible for real-world data, since the time required for one
lookup might be measured in the order of seconds. This is caused by the fact that
any online method has to access at least n/m characters from the input text [Gra11,
p. 155], and it normally holds that m n. This thesis is focused on index-based (offline)
methods, where a data structure (an index, pl. indexes or indices, we opt for the former
term) is built based on the input text in order to speed up further searches, which is a
classic example of data preprocessing. This is justified even if the preprocessing time is
long, since the same text is often queried with multiple patterns. The indexes can be
divided into two following categories:
• Full-text indexes
• Keyword (dictionary) indexes
The former means that we can search for any substring in the input text (string matching,
text matching), whereas the latter operates on individual words (word matching, keyword
matching, dictionary matching, matching in dictionaries). Keyword indexes are usually
appropriate where there exist well-defined boundaries between the keywords (which are
often simply called words), for instance in the case of a natural language dictionary or
individual DNA reads (k-mers). It is worth noting that the number of distinct words is
almost always smaller than the total number of words in a dictionary D, all of which are
taken from a document or a set of documents. Heaps’ law states that |D| = O(nβ ), where
n is the text size and β is an empirical constant (usually in the interval [0.4, 0.6]) [Hea78].
Full-text and keyword indexes are actually related to each other, because they are often
based on similar concepts (e.g., the pigeonhole principle) and they may even use the
other kind as the underlying data structure.
String Searching
25
The indexes can be divided into static and dynamic ones, depending on whether updates
are allowed after the initial construction. Another category is formed by external indexes
— these are optimized with respect to disk I/O, and they aim to be efficient for the data
which does not fit into the main memory. We can also distinguish compressed indexes
(see Subsection 1.2.5.1 for more information on compression), which store the data in
an encoded form. One goal is to reduce storage requirements while still allowing fast
searches, especially when compared to the scenario where a naive decompression of the
whole index has to be performed. On the other hand, it is also possible to achieve
both space saving and a speedup with respect to the uncompressed index. This can
be achieved mostly due to reduced I/O and (rather surprisingly) fewer comparisons
required for the compressed data [Gra11, p. 131]. Navarro and Mäkinen [NM07] note
that the most successful indexes can nowadays obtain both almost optimal space and
query time. A compressed data structure usually also falls into the category of a succinct
data structure. This is a rather loose term which is commonly applied to algorithms
which employ efficient data representations with respect to space, often close to the
theoretic bound. Thanks to reduced storage requirements, succinct data structures can
process texts which are an order of magnitude bigger than ones suitable for classical
data structures [GP14]. The term succinct may also suggest that the we are not required
to decompress the entire structure in order to perform a lookup operation. Moreover,
certain indexes can be classified as self-indexes, which means that they implicitly store
the input string S. In other words, it is possible to transform (decompress) the index
back to S, and thus the index can essentially replace the text.
The main advantage of indexes when compared to the online scenario are fast queries,
however, this naturally comes at a price. Indexes might occupy a substantial amount
of space (sometimes even orders of magnitude more than the input), they are expensive
to construct, and it is often problematic to support functionality such as approximate
matching and updates. Still, Navarro et al. [NBYST01] point out that in spite of the
existence of very fast (both from a practical and a theoretical point of view) online
algorithms, the data size often renders online algorithms infeasible (which is even more
relevant in the year 2015).
Index-based methods are explored in detail in the following chapters: full-text indexes
in Chapter 3 and keyword indexes in Chapter 4. Experimental evaluation of our contributions can be found in Chapter 5.
Chapter 3
Full-text Indexes
Full-text indexes allow for searching for an arbitrary substring from the input text.
Formally, for a string T of length n, having a set of x substrings S = {s1 , . . . , sx } over a
given alphabet Σ, I(S) is a full-text index supporting matching with a specified distance
D. For any query pattern P , it returns all substrings s from T s.t. D(P, s) 6 k (with
k = 0 for exact matching). In the following sections, we describe data structures from
this category, divided into exact (Section 3.1) and approximate ones (Section 3.2). Our
contribution in this field is presented in Subsection 3.1.6, which describes a variant of
the well-known FM-index called FM-bloated.
3.1
3.1.1
Exact
Suffix tree
The suffix tree (ST) was introduced by Weiner in 1973 [Wei73]. It is a trie (see Subsection 1.2.2.2) which stores all suffixes of the input string, that is n suffixes in total for the
string of length n. Moreover, the suffix tree is compressed, which in this context means
that each node which has only one child is merged with this child, as shown in Figure 3.1.
Searching for a pattern takes O(m) time, since we proceed in a way similar to the search
in a regular trie. Suffix trees offer a lot of additional functionality beyond string searching
such as calculating the Lempel–Ziv compression [Gus97, p. 166] or searching for string
repeats [AKO02]. The ST takes linear space with respect to the total input size (O(n2 )
if uncompressed), however, it occupies significantly more space than the original string
— in a space-efficient implementation around 10.1n bytes on average in practice and
even up to 20n in the worst case [Kur99, AOK02], which might be a bottleneck when
dealing with massive data. Moreover, the space complexity given in bits is actually equal
26
27
Full-text Indexes
to O(n log n) (which is also the case for the suffix array) rather than O(n log σ) required
to store the original text. When it comes to preprocessing, there exist algorithms which
construct the ST in linear time [Ukk95, Far97].
As regards the implementation, an important consideration is how to represent the children of each node. A straightforward approach such as storing them in a linked list would
degrade the search time, since in order to achieve the overall time of O(m), we have to
be able to locate each child in constant time. This can be accomplished for example with
a hash table which offers an O(1) average time for a lookup.
a
a
$
banana$
banana$
na
a$
$
ana
$
ana$
na
na$
na$
nana$
na$
anana$
Figure 3.1: A suffix tree (ST) which stores all suffixes of the text banana with an
appended terminating character $, which prevents a situation where a suffix could be
a prefix of another suffix.
A common variation is called generalized suffix tree and it refers to a ST which stores
multiple strings, that is all suffixes for each string S1 , . . . , Sx . Additional information
which identifies the string Si is stored in the nodes, and the complexities are the same
as for a regular ST [Gus97, p. 116]. Compressed suffix trees which reduce the space
requirements were also described (they are usually based on a compressed suffix array) [Gog11, GNPS15].
3.1.2
Suffix array
The suffix array (SA) comes from Manber and Myers [MM93] and it stores indexes of
sorted suffixes of the input text, see Figure 3.2 for an example. According to Kärkkäinen [Kär95], suffix arrays perform comparably to suffix trees when it comes to string
28
Full-text Indexes
matching, however, they are slower for other kinds of searches such as regular expression
matching. Even though the SA takes more space than the original string (4n bytes in
its basic form, and the original string has to be stored as well), it is significantly smaller
than the suffix tree (5n < 10.1n) and it has better locality properties [AKO02]. The
search over a SA takes O(m log n) time, since we perform a binary search over n suffixes and each comparison takes at most m time (although the comparison is constant
on average, see Appendix B). The space complexity is equal to O(n) (there are n suffixes and we store one index per suffix), and it is possible to construct a SA in linear
time. See Puglisi et al. [PST07] for an extensive survey of multiple construction algorithms with a practical evaluation, which concludes that the algorithm by Maniscalco
and Puglisi [MP08] is the fastest one. Parallel construction algorithms which use the
GPU were also considered [DK13]. Let us point out a similarity of SA to the ST, since
sorted suffixes correspond to the depth-first traversal over the ST. Main disadvantage
with respect to the ST is a lack of additional functionality such as this mentioned in the
previous subsection.
Suffix
$
a$
ana$
anana$
banana$
na$
nana$
Index
6
5
3
1
0
4
2
Figure 3.2: A suffix array (SA) which stores indexes (0-based) of sorted suffixes of
the text banana$. The suffixes are not stored explicitly, although the entire input text
has to be stored.
3.1.2.1
Modifications
Multiple modifications of the original SA have been proposed over the years. Their aim
is to either speed up the searches by storing additional information or to reduce space
requirements by compressing the data or omitting a subset of the data. Most notable
examples are presented below.
The enhanced suffix array (ESA) is a variant where additional information in the form
of a longest common prefix (LCP) table is stored [AKO02]. For a suffix array SA over
the string of length n, the LCP table L holds integers from the range [0, n], and it has
the following properties: L[0] = 0, and L[i] holds the length of the longest common
prefix of suffixes from SA [i] and SA [i − 1]. ESA can essentially replace the suffix tree
since it offers the same functionality, and it can deal with the same problems in the same
29
Full-text Indexes
time complexity (although constant alphabet size is assumed in the analysis) [AKO04].
For certain applications, it also required to store the Burrows–Wheeler transform (see
−1
Subsection 3.1.4.1) of the input string and an inverse suffix array (SA
[SA [i]] = i).
The size of the index can be always reduced using any compression method. However,
such a naive approach would certainly have a negative impact on search performance
because of the overhead associated with decompression, and a much better approach is
to use a dedicated solution. Mäkinen [Mäk00] presented a compact suffix array (CoSA)
with average search time O(((2n−n0 )/n0 )2 (m+log n)) where n0 is the length of the CoSA
and practical space reduction of up to 50% by replacing repetitive suffixes with links to
other suffixes. Grossi and Vitter [GV05] introduced a compressed suffix array (CSA)
which uses O(n log σ) bits instead of O(n log n) bits. It is based on a transformation of
the SA into the array which points to the position of the next suffix in the text. For
instance, for the text banana$ and the suffix ana$, the next suffix is na$, see Figure 3.3.
These transformed values are compressible because of certain properties such as the fact
that number of increasing sequences is in O(σ). The search takes O(m/ logσ n + logσ n)
time, and the relation between search time and space can be fine-tuned using certain
parameters.
For more information on compressed indexes, including the modifications of the SA, we
refer the reader to the survey by Navarro and Mäkinen [NM07]. The FM-index which
is presented in Subsection 3.1.4 can be also regarded as a compressed variant of the
SA [FM05].
i
0
1
2
3
4
5
6
Suffix
$
a$
ana$
anana$
banana$
na$
nana$
SA index
6
5
3
1
0
4
2
CSA index
4
0
5
6
3
1
2
Figure 3.3: A compressed suffix array (CSA) for the text banana$ which stores indexes
pointing to the next suffix from the text (the SA is shown for clarity and it is not stored
along with the CSA).
The sparse suffix array stores only suffixes which are located at the positions in the
form iq for a fixed q value [KU96]. In order to answer a query, q searches and q −
1 explicit verifications are required, and it must hold that m > q. Another notable
example of a modified suffix array which stores only a subset of data is the sampled
suffix array [CNP+ 12]. The idea is to select a subset of the alphabet (denoted with
ΣS ) and extract corresponding substrings from the text. The array is constructed only
Full-text Indexes
30
over those suffixes which start with a symbol from the chosen subalphabet (although
the sorting is performed on full suffixes). Only the part of the pattern which contains a
character c ∈ ΣS is searched for (i.e. there is one search in total) and the matches are
verified by comparing the rest of the pattern with the text. The disadvantage is that
the following must hold, ∃c ∈ P : c ∈ ΣS . Practical reduction in space in the order of
50% was reported. Recently, Grabowski and Raniszewski [GR14] proposed an alternative
sampling technique based on minimizers (see Section 4.3.1) which allows for matching
all patterns P s.t. |P | > q where q is the minimizer window length and requires only one
search.
3.1.3
Other suffix-based structures
The suffix tray combines — just as the name suggests — the suffix tree with the suffix
array [CKL06]. The top-level structure is a ST whose nodes are divided into heavy and
light, depending on whether their subtrees have more or fewer leaves than some predefined
threshold. Light children of heavy nodes store their corresponding SA interval. The query
time equals O(m + log σ), and preprocessing and space complexities are equal to O(n).
The authors also described a dynamic variant which is called a suffix trist and allows
updates.
Yet another modification of the classical ST is called suffix cactus (SC) [Kär95]. Here,
Kärkkäinen reworks the compaction procedure, which is a part of the construction of
the ST. Instead of collapsing only the nodes which have only one child, every internal
node is combined with one of its children. Various methods of selecting such a child exist
(e.g., alphabetical ordering), and thus the SC can take multiple forms for the same input
string. The original article reports the best search times for the DNA, whereas the SC
performed worse than both ST and SA for the English language and random data. The
space complexity is equal to O(n).
3.1.4
FM-index
The FM-index is a compressed (succinct) full-text index which was introduced by Ferragina and Manzini [FM00] in the year 2000. It was applied in a variety of situations, for instance for sequence assembly [LTP+ 09, SD10] or for ranked document retrieval [CPS12]. Multiple modifications of the FM-index were described throughout the
years (some are introduced in the following subsections). The strength of the original
FM-index lies in the fact that it occupies less space than the input text while still allowing fast queries. The search time of its unmodified version is linear with respect to the
pattern length (although a constant-size alphabet is assumed), and the space complexity
Full-text Indexes
31
is equal to O(Hk (T ) + log log n/ log n) bits per input symbol. Taking the alphabet size
into account, Grabowski et al. [GNP+ 06] provide a more accurate total size bound of
O(Hk (T )n + (σ log σ + log log n) logn n + nγ σ σ+1 ) bits for 0 < γ < 1.
3.1.4.1
Burrows–Wheeler Transform
FM-index is based on the Burrows–Wheeler transform (BWT) [BW94], which is an ingenious method of transforming a string S in order to reduce its entropy. BWT permutes
the characters of S in such a way that duplicated characters often appear next to each
other, which allows for easier processing using methods such as run-length or moveto-front encoding (as is the case in, e.g., the bzip2 compressor [Sew, PJT09]). Most
importantly, this transformation is reversible (as opposed to straightforward sorting),
which means that we can extract the original string from the permuted order. BWT
could be also used for compression based on the k-th order entropy (described in Subsection 1.2.5.1) since basic context information can be extracted from BWT, however,
the loss of speed renders such an approach impractical [Deo05].
In order to calculate the BWT, we first append a special character (we describe it with
$, but in practice it can be any character c ∈
/ S) to S in order indicate its end. The
character $ is lexicographically smaller than all {c : c ∈ S}. The next step is to take all
rotations of S (|S| rotations in total) and sort them in a lexicographic order, thus forming
the BWT matrix, where we denote the first column (sorted characters) with F and the
last column (the result of the BWT, i.e. T bwt ) with L. In order to finish the transform,
we take the last character of each rotation, as demonstrated in Figure 3.4. Let us note
the similarities between the BWT and the suffix array described in Subsection 3.1.2,
since the sorted rotations correspond directly to sorted suffixes (see Figure 3.5). The
calculation takes O(n) time, assuming that the prefixes can be sorted in linear time,
and the space complexity of the naive approach is equal to O(n2 ) (but it is linear if
optimized).
In order to reverse the BWT, we first sort all characters and thus obtain the first column
of the matrix. At this point, we have two columns, namely the first and the last one,
which means that we also have all character 2-grams from the original string S. Sorting
these 2-grams gives us the first and the second column, and we proceed in this manner
(later we sort 3-grams, 4-grams, etc) until we reach |S|-grams and thus reconstruct the
whole transformation matrix. At this point, S can be found in the row where the last
character is equal to $.
32
Full-text Indexes
R1
R2
R3
R4
R5
R6
R7
R8
$
a
e
n
p
r
t
t
p
t
r
$
a
n
e
t
a
t
n
p
t
$
r
e
t
e
$
a
t
p
n
r
t
r
p
t
e
a
$
n
e
n
a
t
r
t
p
$
r
$
t
e
n
t
a
p
n
p
t
r
$
e
t
a
Figure 3.4: Calculating a Burrows–Wheeler transform (BWT) for the string pattern
with an appended terminating character $ (it is required for reversing the transform). The rotations are already sorted and the result is in the last column, i.e.
BWT(pattern$) = nptr$eta.
3.1.4.2
Operation
Important aspects of the FM-index are as follows:
• Count table C, which describes the number of occurrences of lexicographically
smaller characters for all c ∈ S (see Figure 3.6).
• Rank operation, which counts the number of set bits in a bit vector v before a
certain position i (we assume that v[i] is included as well), that is rank(i, v) =
|{i0 : 0 6 i0 6 i ∧ v[i0 ] = 1}|.
• Select operation (used only in some variants, e.g., the RLFM [MN05b]), which
reports the position of the i-th set bit in the bit vector v, that is select(i, v) = p if
and only if |{i0 : 0 6 i0 < p ∧ v[i0 ] = 1}| = i − 1.
Note that both rank and select operations can be generalized to any finite alphabet Σ.
When we perform the search using the FM-index, we iterate the pattern characterwise
in a reverse order while maintaining a current range r = [s, e]. Initially, i = m − 1 and
r = [0, n − 1], that is we start from the last character in the pattern and the range covers
the whole input string (here the input string corresponds to T bwt , that is a text after the
BWT). At each step we update s and e using the formulae presented in Figure 3.7. The
size of the range after the last iteration gives us the number of occurrences of P in T ,
or it turns out that P 6⊂ T if s > e at any point. This mechanism is also known as the
LF-mapping.
3.1.4.3
Efficiency
We can see that the performance of C lookup and rank is crucial to the complexity of
the search procedure. In particular, if these operations are constant, the search takes
33
Full-text Indexes
i
0
1
2
3
4
5
6
7
BWT
n
p
t
r
$
e
t
a
SA
7
1
4
6
0
5
3
2
suffix
$
attern$
ern$
n$
pattern$
rn$
tern$
ttern$
Figure 3.5: A relation between the BWT and the SA for the string pattern with
an appended terminating character $. Let us note that BW T [i] = S[SA[i] − 1] (where
S[−1] corresponds to the last character in S), that is a character at the position i in
BWT is a character preceding a suffix which is located at the same position in the SA.
c
C
$
0
i
1
m
5
p
6
s
8
Figure 3.6: Count table C which is a part of the FM-index for the text mississippi$.
The entries describe the number of occurrences of lexicographically smaller characters
for all c ∈ S. For instance for the letter m, there are 4 occurrences of i and 1 occurrence
of $ in S, hence C[m] = 5. It is worth noting that C is actually a compact representation
of the F column.
s = C[P [i]] + rank(s − 1, P [i])
e = C[P [i]] + rank(e, P [i]) − 1
Figure 3.7: Formulae for updating the range during the search procedure in the FMindex, where P [i] is the current character and C is the count table. Rank is invoked
on T bwt and it counts occurrences of the current character P [i].
O(m) time. For the count table, we can simply precompute the values and store them
an array of size σ with O(1) lookup. As regards rank, a naive implementation which
would iterate the whole array would clearly take O(n) time. On the other hand, if we
were to precompute all values, we would have to store a table of size O(nσ). One of the
popular solutions for an efficient rank uses two structures which are introduced in the
following paragraphs.
The RRR (from authors’ names: Raman, Raman, and Rao) is a data structure which
can answer the rank query in O(1) time for bit vectors (i.e. where Σ = {0, 1}), while
providing compression at the same time [RRR02]. It divides a bit vector v of size n into
n/b blocks each of size b, and groups each consecutive s blocks into one superblock (see
Figure 3.8). For each block, we store a weight w which describes the number of set bits
and offset o which describes its position in a table TR (the maximum value of o depends
on w). In TR , for each w and each o, we store a value of rank for each index i, where
34
Full-text Indexes
0 6 i 6 b (see Figure 3.9). This means that we have to keep
b
w
entries each of size b for
each of the b + 1 consecutive weights. Such a scheme provides compression with respect
to storing all n bits explicitly. We achieve the O(1) query time by storing a rank value
for each superblock, and thus during the search we only iterate at most s blocks (s is
constant). The space complexity is equal to nH0 (v) + O(n log log n/ log n) bits.
001
001
101
000
Figure 3.8: An example of RRR blocks for b = 3 and s = 2, where the first superblock
is equal to 001001 and the second superblock is equal to 101000.
offset
0
1
2
block value
011
101
110
rank
012
112
122
Figure 3.9: An example of an RRR table for w = 2 and b = 3, where 32 = 3 (i.e. the
number of all block values of length 3 with weight 2 is equal to 3), with rank presented
for successive indexes i ∈ [0, 2]. Block values do not have to be stored explicitly.
The wavelet tree (WT) from Grossi et al. [GGV03] is a balanced tree data structure that
stores a hierarchy of bit vectors instead of the original string, which allows the use of
RRR (or any other bit vector with an efficient rank operation). Starting from the root,
we recursively partition the alphabet into two subsets of equal length (if the number
of distinct characters is even), until we reach single symbols which are stored as leaves.
Characters belonging to the first subset are indicated with 0s, and characters belonging
to the second subset are indicated with 1s (consult Figure 3.10 for an example).
Thanks to the WT we can implement a rank query for any fixed size alphabet in O(log σ)
time (assuming that a binary rank is calculated in constant time), since the height of
the tree is equal to log σ. For a given character c, we query the WT at each node and
proceed left or right depending on the subset to which c belongs. Each subsequent rank
is called in the form rank(c, p), where p is the result of the rank at the previous level.
Ferragina et al. [FMMN07] described generalized WTs, for instance a multiary WT with
O(log σ/ log log n) traversal time (consult Bowe’s thesis [Bow10] for more information
and a practical evaluation).
3.1.5
FM-index flavors
Multiple flavors of the FM-index were proposed over the years with the goal of decreasing
the query time, e.g., having O(m) time without the dependence on σ, or reducing the
occupied space. The structures which provide asymptotically optimal bounds are often
35
Full-text Indexes
abracadabra
00101010010
{a,b}
abaaaba
0100010
{a}
a
{c,d,r}
rcdr
1011
{b}
b
{c}
{d,r}
rdr
101
c
{d}
d
{r}
r
Figure 3.10: A wavelet tree (WT) over the string abracadabra. The alphabet is
divided into two subsets at each level, with 0 corresponding to one subset and 1 to the
other.
not practical due to the very large constants which are involved [Vig08]. For this reason,
many authors focus on practical performance, and these structures are usually based
on a fast rank operation and take advantage of compressed representations of bit vectors [NM07]. The following paragraphs present selected examples, consult Navarro and
Mäkinen [NM07] for an extensive survey of compressed full-text indexes which discusses
the whole family of FM-indexes.
One of the notable examples where the query time does not depend on σ is the alphabetindependent FM-index by Grabowski et al. [GMN04]. The idea is to first compress the
text using Huffman coding and then apply the BWT transform over it, obtaining a bit
vector v. This vector is then used for searching in a manner corresponding to the FMindex — the array C stores the number of zeros up to a certain position, and the relation
C[c] + rank(T bwt , i, c) is replaced with i − rank(v, i) if c = 0 and n0 − rank(v, n0 ) +
rank(v, i) if c = 1, where n0 is the length of the text compressed with Huffman. The
space complexity is equal to O(n(H0 (T ) + 1)) bits, and the average search time is equal
to O(m(H0 (T ) + 1)) under “reasonable assumptions”.
On the practical front, Grabowski et al. [GRD15] recently described a cache-aligned rank
with 1 cache miss. Moreover, they proposed the so-called FM-dummy index with several
Full-text Indexes
36
variants, for instance one which stores a separate bit vector for each alphabet symbol
(σ vectors in total). Other variants include using certain dense codes as well as using
Huffman-shaped multiary wavelet trees with different arity values. A Huffman-shaped
wavelet tree is unbalanced, and the paths for frequent characters are shorter (which
translates to a smaller number of rank queries on bit vectors). Moreover, the operations
(which are performed in the same manner as for the regular wavelet tree) are faster on
average [CNO14]. They reported search times which are 2–3 times faster than those for
other state-of-the-art methods at the cost of using 1.5–5 times more space.
Data structures which concentrate on reducing space requirements rather than the query
time include the compressed bit vectors from Kärkkäinen et al. [KKP14], where different
compression methods are used for blocks depending on the type of the block, for instance
run-length encoding for blocks with a small number of runs. Another notable example is
a data-aware FM-index by Huo et al. [HCZ+ 15], which encodes the bit vectors (resulting
from the WT) using Gamma coding (a kind of variable-length coding), and thus obtain
one of the best compression ratios in practice.
3.1.5.1
Binary rank
As described in the previous subsection, in order to achieve good overall performance, it is
sufficient to design a data structure which supports an efficient rank query for bit vectors
thanks to the use of a wavelet tree (RRR being a notable example). Jacobson [Jac89]
originally showed that it is possible to obtain a constant-time rank operation using o(n)
extra bits for |v| = n (the same holds for select [Cla96]). Vigna [Vig08] proposed to
interleave (i.e. store next to one another) blocks and superblocks (concepts which were
introduced for the RRR structure) for uncompressed bit vectors in order to reduce the
number of cache and translation lookaside buffer (TLB) misses from 3 to 2. This was
extended by Gog and Petri [GP14], who showed better practical performance by using
a slightly different layout with 64-bit counters. Gonzalez and Navarro [GN09] provided
a discussion of the dynamic scenario where insertions and deletions to the vector are
allowed, and they obtain a space bound of nH0 (v) + o(n log σ) bits and O(log n(1 +
log σ/ log log n)) time for all operations (i.e. queries and updates).
3.1.6
FM-bloated
One of the crucial issues when it comes to the performance of the FM-index is the
number of CPU cache misses which occur during the search. This comes from the fact
that in order to calculate the LF-mapping, non-local access to the BWT sequence is
often required (in the order of Ω(m) misses during the search for a pattern of length m),
37
Full-text Indexes
even for a small alphabet. The problem of cache misses during the FM-index backward
search was identified as the main performance limiter by Chacón et al. [CMEH13], who
proposed to perform the LF-mapping with several symbols at a time (in practice, at most
4 for the 4-symbol DNA alphabet for which the scheme was described). This solution
allowed, for example, to improve the search speed by a factor of 1.5 for the price of
occupying roughly 2 times the size of the 1-step FM-index. Here, we address the problem
of cache misses during the pattern search (count query) in a way related to the Chacón
et al. solution — we also work on q-grams, yet the algorithmic details are different. Two
following subsections describe two variants of our approach, and experimental results can
be found in Section 5.1.
3.1.6.1
Superlinear space
FM-bloated is a variation of the FM-index which aims to speed up the queries at the cost
of additional space. We start by calculating the BWT for the input string in the same
way as for the regular FM-index, however, the difference is that we operate on q-grams
rather than on individual characters, and the count table stores results for each q-gram
sampled from the BWT matrix. This is the case for all q, where q is the power of 2,
up to some predefined value qmax (for instance 128). Namely, for each suffix T [i, n − 1],
we take all q-grams in the following form: T [i − 1] (1-gram), T [i − 2, i − 1] (2-gram),
T [i − 4, i − 1] (4-gram), etc. The q-grams are extracted until we reach qmax or one
of the q-grams contains the terminating character (such a q-gram with the terminating
character is discarded), consult Figure 3.11 for an example. Let Q denote a collection of
all q-grams for all i. For each distinct item s from Q, we create a list Ls of its occurrences
in the sorted suffix order (simply called SA order). This resembles an inverted index on
q-grams, yet the main difference is that the elements in the lists are arranged in SA rather
than the text order, e.g., for the q-gram t in Figure 3.11, the 1-based list of occurrences
corresponding to rows would be as follows: {3, 7}.
row
R1
R2
R3
R4
R5
R6
R7
R8
BWT
$pattern
attern$p
ern$patt
n$patter
pattern$
rn$patte
tern$pat
ttern$pa
q-grams
n, rn, tern
p
t, tt, patt
r, er, tter
e, te, atte
t, at
a, pa
Figure 3.11: Q-gram extraction in the FM-bloated structure with superlinear space
for the text T = pattern$. All q-grams are extracted, i.e. qmax > dn/2e.
Full-text Indexes
38
For a given pattern P , we start the LF-mapping with its longest suffix Ps s.t. |Ps | 6
qmax ∧ |Ps | = 2c for some c ∈ Z. The following backward steps deal with the remaining
prefix of P in a similar way. Note that the number of LF-mapping steps is equal to the
number of 1s in the binary representation of m, i.e. it is in the order of O(log m), and if m
is a power of two, then the result for match and count queries can be reported in constant
time (we simply return |Ls |). When qmax is bigger, the overall index size is bigger,
but the search is faster (for patterns of sufficient length) because it allows for farther
jumps towards the beginning of the pattern. In our representation, each LF-mapping
step translates to performing two predecessor queries on a list Ls . A naive solution is
a binary search with O(log n) worst-case time (or even a linear search, which may be
faster if the list is short), yet the predecessor query can be also handled in O(log log n)
time using a y-fast trie [Wil83]. Hence, the overall average search complexity is equal
to O(m + log m log log n), with O(m/CL + log m log log n) cache misses where CL is the
cache line size in bytes (provided that each symbol from the pattern occupies one byte).
As regards the space complexity, there is a total of n log n q-gram occurrences (log n
q-gram positions for each of n rows of the BWT matrix). Hence, the total length of all
occurrence lists is equal to n log n, and the total complexity is equal to O(n log2 n) bits
(since we need log n bits to store one position from the BWT matrix of n rows).
As regards the implementation (in the C++ language), our focus is on data compaction.
Each q-gram acts as a key in a hash table where collisions are resolved with chaining,
and the q-grams are stored implicitly, i.e. as a (pointer, length) pair, where the pointer
refers to the original string. The values in the hash table include the count and the
list of occurrences which are stored in one, contiguous array. We use a binary search
for calculating rank on lists whose length is greater than or equal to 16 (an empirically
determined value) and a linear search otherwise.
3.1.6.2
Linear space
In this variant, instead of extracting all 1-, 2-, etc, q-grams for each row of the BWT
matrix, we extract only selected q-grams with the help of minimizers (consult Subsection 4.3.1 for the description of minimizers). The first step is to calculate all (α, q)minimizers for the input text T , with some fixed α and q parameters and lexicographic
ordering, where ties are resolved in favor of the leftmost of the smallest substrings. Next,
we store both the count table and the occurrence lists for all single characters in the same
way as for the regular FM-index (using, e.g., a wavelet tree). Moreover, we store information about the counts and occurrences of all q-grams which are located in between
the minimizers from the set M (T ) — these q-grams are referred to as phrases. For the
set of minimizer indexes I(T ), consecutive phrases pi are constructed in the following
39
Full-text Indexes
|I|−1
manner: ∀i=1 pi = T [I[i], I[i + 1] − 1], consult Figure 3.12. It is worth noting that this
approach resembles the recently proposed SamSAMi index, a sampled suffix array on
minimizers [GR14].
T
M
I
Phrases
Phrase ranges
appearance
ap, ar, an
0, 4, 6
appe, ar
[0, 3], [4, 5]
Figure 3.12: Constructing FM-bloated phrases for the text appearance with the use
of (4, 2)-minimizers.
The search proceeds as follows:
1. We calculate all minimizers for the pattern.
2. We search for the pattern suffix PS = P [sr , −1], where sr is the starting position
of the rightmost minimizer using the regular FM-index mechanism, i.e. processing
1 character at a time.
3. We operate on the phrases between the minimizers rather than individual characters and the search for these q-grams is performed in the same way as for the
superlinear variant. If it turns out that the phrase is a 1-gram, a faster FM-index
mechanism for single characters can be used.
4. We search for the pattern prefix PP = P [0, sl − 1], where sl is the starting position
of the leftmost minimizer using the regular FM-index mechanism, i.e. processing 1
character at a time.
The use of minimizers ensures that the phrases are selected from P in the same way
as they are selected from T during the index construction. The overall average search
complexity is equal to O(m log log n) (again, assuming that a y-fast trie [Wil83] is used),
and the space complexity is linear.
3.2
Approximate
Navarro et al. [NBYST01] provided an extensive survey of full-text indexes for approximate string matching. They categorized the algorithms into three categories based on
the search procedure:
Full-text Indexes
40
• Neighborhood generation — all strings in {S : S ∈ Σ∗ ∧ D(S, P ) 6 k} for a
given pattern P are searched for directly.
• Partitioning into exact searching (PiES) — substrings of the pattern are
searched for in an exact manner and these matches are extended into approximate
matches.
• Intermediate partitioning — substrings of the pattern are searched for approximately but with a fewer number of errors. This method lies in between the two
other ones.
In the neighborhood generation approach, we generate the k-neighborhood K of the
pattern, which contains all strings which could be possible matches over a specified
alphabet Σ (if the alphabet is finite, the amount of such strings is finite as well). These
strings can be searched for using any exact index such as a suffix tree or a suffix array. The
main issue is the fact that the size of K grows exponentially, |K| = O(mk σ k ) [Ukk85b],
which means that basically all factors (and especially k) should be small. When the
suffix tree is used as an index for the input text, Cobbs [Cob95] proposed a solution
which reduces the amount of nodes that have to be processed. It runs in O(mq + |o|)
time and occupies O(q) space, where q 6 n (q depends on the problem instance) and |o|
is the size of the output.
When the pattern is partitioned and searched for exactly (PiES), we have to again store
the index which can answer these exact queries. Let us note that this approach is based
on the pigeonhole principle. In the context of approximate string searching this means
that for a given k, at least one of k + 1 parts of average length |P |/(k + 1) must match
the text exactly (more generally, s parts match if k + s parts are created). The value
of k should not be too large, otherwise it could be the case that a substantial part
of the input text has to be verified (especially if the pattern is small). Alternatively,
the pattern can be divided into m − q + 1 overlapping q-grams and these q-grams are
searched for (using the locate query) against the index of q-grams extracted from the
text (see Figure 3.13 for an example of q-gram extraction). These q-grams which are
stored by the index are situated at fixed positions with an interval h and it must hold
that h 6 b(m−k −q +1)/(k +s)c for occurrences of P in T to contain s samples. Sutinen
and Tarhio [ST95] suggested that the optimal value for q is in the order of Θ(logσ m).
If it turns out that the positions of subsequent q-grams may correspond to a match,
explicit verification is performed. Similarly to the k-neighborhood scenario, any index
can be used in order to answer the exact queries. Let us note that this approach with
pattern substring lookup and verification can be also used for exact searching.
Full-text Indexes
41
In the case of intermediate partitioning, we split the pattern into s pieces, and we search
for these pieces in an approximate manner using neighborhood generation. The case
of s = 1 corresponds to pure neighborhood generation, whereas the case of s = k + 1
is almost like PiES. In general, this method requires more searching but less verification when compared to PiES, and thus lies in between the two approaches which were
previously described.
Consult Maaß and Nowak [MN05a] in order to see a detailed comparison of the complexities of modern text indexing methods for approximate matching. Notable structures
from the theoretical point of view include the k-errata trie by Cole et al. [CGL04] which
is based on the suffix tree and the LCP structure (see Subsection 3.1.2.1 for a description
of the LCP). It can be used in various contexts, including full-text and keyword indexing
as well as wildcard matching. For full-text indexing and the k-mismatches problem, it
uses O(n logk n/k!) space and offers O(m + logk n/k! + occ) query time. This was extended by Tsur [Tsu10] who described a structure similar to the one from Cole et al. with
time complexity O(m + log log n + occ) (for constant k) and O(n1+ ) space for a constant
> 0. As regards a solution which is dedicated for the Hamming distance, Gabriele et
al. [GMRS03] provided an index with average search time O(m + occ) and O(n logl n)
space (for some l).
Let us note that these full-text indexes can be usually easily adapted to the keyword
matching scenario which is described in the following chapter.
3.3
Word-based
An interesting category of data structures are word-based indexes, which can be used
for approximate matching and especially sequence alignment. They employ heuristic
approaches in order to speed up the searching, and for this reason they are not guaranteed to find the optimal match. This means that they are also approximate in the
mathematical sense, i.e. they do not return the true answer to the problem. Their popularity is especially widespread in the context of bioinformatics, where the massive sizes of
the databases often force the programmers to use efficient filtering techniques. Notable
examples include BLAST [AGM+ 90] and FASTA [LP85] tools.
3.3.1
BLAST
BLAST stands for Basic Local Alignment Search Tool, and it was published by Altschul
et al. [AGM+ 90] in 1990 with the purpose of comparing biological sequences (see Subsection 1.1.2 for more information about biological data). The name may refer to the
42
Full-text Indexes
algorithm or to the whole suite of string searching tools for bioinformatics which are based
on the said algorithm. BLAST relies heavily on various heuristics and for this reason it
is highly domain specific. In fact there exist various flavors of BLAST for different data
sets, for instance one for protein data (blastp) and one for the DNA (blastn). Another
notable modification is the PSI-BLAST which is combined with dynamic programming
in order to identify distant protein relationships.
The basic algorithm proceeds as follows:
1. Certain regions are removed from the pattern. These include repeated substrings
and regions of low complexity (measured statistically using, e.g., DUST for DNA
[MGSA06]).
2. We create a set Q containing q-grams with overlaps (that is all available q-grams,
see Figure 3.13) which are extracted from the pattern. Each s ∈ Q is scored against
all possible q-grams (these can be precomputed) and ones with the highest scores
are retained creating a candidate set QC .
3. Each word from QC is searched for in an exact manner against the database (using
for instance an inverted index, see Subsection 4.1.2). These exact matches create
the seeds which are later used for extending the matches.
4. The seeds are extended to the left and to the right as long as the alignment score
is increasing.
5. Alignment significance is assessed using domain-specific statistical tools.
size
2
3
4
5
q-grams
te, ex, xt, ti, in, ng
tex, ext, xti, tin, ing
text, exti, xtin, ting
texti, extin, xting
Figure 3.13: Selecting all overlapping q-grams (with the shift of 1) from the text
T = texting. It must always hold that q 6 |T |.
In general, BLAST is faster than other alignment algorithms such as the SW algorithm
(see Subsection 2.2.2) due to its heuristic approach. However, this comes at a price of reduced accuracy, and Shpaer et al. [SRY+ 96] state that there is a substantial chance that
BLAST will miss a distant sequence similarity. Moreover, hardware-oriented implementations of the SW have been created, and in certain cases they can match the performance
of BLAST [MV08]. Still, BLAST is currently the most common tool for sequence alignment using massive biological data, and it is openly available via its website [BLA], which
means that it can be conveniently run without consuming local resources.
Chapter 4
Keyword Indexes
Keywords indexes operate on individual words rather than the whole input string. Formally, for a collection D = {d1 , . . . , dx } of x strings (words, q-grams) of total length n
over a given alphabet Σ, I(D) is a keyword index supporting matching with a specified
distance D. For any query pattern P , it returns all words d from D s.t. D(P, d) 6 k
(with k = 0 for exact matching). Approximate dictionary matching was introduced by
Minsky and Papert in 1969 [MP69, CGL04].
In the following sections, we describe algorithms from this category, divided into exact
(Section 4.1) and approximate ones (Section 4.2). Our contribution in this field is presented in Subsection 4.2.3, which describes an index for approximate matching with few
mismatches (especially 1 mismatch).
4.1
Exact
If the goal were to support only the match query for a finite number of keywords, we
could use any efficient set data structure such as a hash table or a trie (see Subsections 1.2.3 and 1.2.2.2) in order to store all those keywords. Boytsov [Boy11] reported
that depending on the data set, either one of these two may be faster. In order to reduce
space requirements we could use minimal perfect hashing (see Subsection 1.2.3), and we
could also compress the entries in the buckets.
4.1.1
Bloom filter
Alternatively, we could provide only approximate answers (in a mathematical sense) in
order to occupy even less space. A relevant data structure is the Bloom filter (BF) [Blo70],
43
44
Keyword Indexes
which is a space-efficient, probabilistic data structure with possible false positive matches,
but no false negatives and an adjustable error rate. The BF uses a bit vector A of size n,
where no bits are initially set. Each element e is hashed with k different hash functions Hi
in the form H(e) = i, where i ∈ Z ∧ 0 6 i < n, and ∀ki=1 A[Hi (e)] = 1. When the lookup
is performed, the queried element is hashed with the same functions and it is checked
whether A[i] = 1 for all i, and if that is the case a possible match is reported (consult
Figure 4.1). Broder and Mitzenmacher [BM04] provided the following formula for the
expected false positive rate: FP = 0.5k > 0.5m ln 2/n , where m is the size of the filter in
bits and n is the number of elements. They note that for example when m = 8n, the false
positive probability is slightly above 0.02. Recently, Fan et al. [FAKM14] described a
structure based on cuckoo hashing which takes even less space than the BF and supports
deletions (unlike the BF).
{x, y, z}
0
1
0
1
1
1
0
0
0
0
0
1
0
1
0
0
1
0
w
Figure 4.1: A Bloom filter (BF) for approximate membership queries with n = 18
and k = 3, holding the elements from the set {x, y, z}. The element w is not in the set
since H2 (w) = 15 and A[15] = 0; reproduced from Wikimedia Commons1 .
4.1.2
Inverted index
An inverted index is a keyword index which contains a mapping from words d ∈ D to the
lists which store all positions pi of their occurrences in the text (d → p1 , . . . , pn ). These
positions can be for instance indexes in a string of characters, or if a more coarse-grained
approach were sufficient, they could identify individual documents or databases. See
Figure 4.2 for an example with a single input string. The positions allow a search on the
whole phrase (i.e. multiple words) by searching for each word separately and checking
whether the positions describe consecutive words in the text (that is by looking for list
intersections with a shift). It could be also used for searching for a query which may cross
the boundaries of the words by searching for substrings of a pattern and comparing the
respective positions (consult Section 4.3 for more information). This means that the goal
of an inverted index is to support various kinds of queries (e.g., locate, see Section 2.1)
efficiently.
1
David Eppstein, available at http://en.wikipedia.org/wiki/Bloom_filter#/media/
File:Bloom_filter.svg, in public domain.
45
Keyword Indexes
Word
This
is
a
banana
Occurrence list
0
12
15
5, 17
Figure 4.2: An inverted index which stores a mapping from words to their positions
(0-based) in the text “This banana is a banana”.
Main advantage of the inverted index are fast single-word queries, which can be answered
in constant average time using for example a hash table (an inverted index is a rather
generic idea, which means that it could be also implemented with other data structures
such as binary trees). On the other hand, there is a substantial space overhead (in the
order of O(n)) and the original string has to be stored as well. For this reason, one of the
key challenges for inverted indexes is how to succinctly represent the lists of positions
while still allowing fast access. Multiple methods were proposed, and they are often
combined with one other [AM05].
The most popular one is to store gaps, that is differences between subsequent positions.
For the index in Figure 4.2, the list for banana would be equal to {5, 12} instead of {5, 17}
(17 − 5 = 12). The values of the gaps are usually smaller than the original positions and
for this reason they can stored using a fewer amount of bits. Another popular approach
is to use byte-aligned coding. Here, each byte contains a one-bit flag which is set if the
number is bigger or equal to 27 (that is when it does not fit into 7 bits), and the other
seven bits are used for the data. If the number does not fit, 7 least-significant bits are
stored in the original byte, and the algorithm tries to store the remaining bits in the
next byte, proceeding until the whole number has been exhausted. In order to reduce
the average length (in bits) of the occurrence list, one could also divide the original
text into multiple blocks of fixed size. Instead of storing exact positions only block
indexes are stored, and after the index is retrieved the word is searched for explicitly
within the block [MW94b]. If the size of the data is so massive that it is infeasible to
construct a single index (as is often the case for web search engines), sometimes only
the most relevant data is selected for being stored in the index (thus forming a pruned
index) [NC07].
4.2
Approximate
Boytsov [Boy11] presented an extensive survey of keyword indexes for approximate
searching (including a practical evaluation). He divided the algorithms into two following categories:
46
Keyword Indexes
• Direct methods — like neighborhood generation (see Section 3.2), where certain
candidates are searched for exactly.
• Sequence-based filtering methods — the dictionary is divided into many (disjoint or overlapping) clusters. During the search, a query is assigned to one or
several clusters containing candidate strings and thus an explicit verification is
performed only on a fraction of the original dictionary.
Notable results from the theoretical point of view include the k-errata trie by Cole
et al. [CGL04] which was already mentioned in the previous chapter. For the Hamk
ming distance and dictionary matching, it uses O(n + d (logk!d) ) space and offers O(m +
(log d)k
k!
log log n + occ) query time, where d = |D| (this also holds for the edit distance
but with larger constants). Another theoretical work describing the algorithm which is
similar to our split index (which we describe in Subsection 4.2.3) was given by Shi and
Widmayer [SW96], who obtained O(n) preprocessing time and space complexity and
O(n) expected time if k is bounded by O(m/ log m). They introduced the notion of
home strings for a given q-gram, which is the set of strings in D that contain the q-gram
in the exact form (the value of q is set to |P |/(k + 1)). In the search phase, they partition
P into k + 1 disjoint q-grams and use a candidate inspection order to speed up finding
the matches with up to k edit distance errors.
On the practical front, Bocek et al. [BHSH07] provided a generalization of the Mor–
Fraenkel (MF) [MF82] algorithm for k > 1 which is called FastSS. To check if two strings
S1 and S2 match with up to k errors, we first delete all possible ordered subsets of k 0
symbols for all 0 6 k 0 6 k from S1 and S2 . Then we conclude that S1 and S2 may be
in edit distance at most k if and only if the intersection of the resulting lists of strings is
non-empty (explicit verification is still required). For instance, if S1 = abbac and k = 2,
then its neighborhood is as follows: abbac, bbac, abac, abac, abbc, abba, abb, aba, abc,
aba, abc, aac, bba, bbc, bac and bac (of course, some of the resulting strings are repeated
and they may be removed). If S2 = baxcy, then its respective neighborhood for k = 2 will
contain, e.g., the string bac, but the following verification will show that S1 and S2 are
in edit distance greater than 2. If, however, Lev(S1 , S2 ) 6 2, then it is impossible not to
have in the neighborhood of S2 at least one string from the neighborhood of S1 , hence we
will never miss a match. The lookup requires O(kmk log nk ) time (where m is the average
word length from the dictionary) and the index occupies O(nk ) space. Another practical
filter was presented by Karch et al. [KLS10] and it improved on the FastSS method.
They reduced space requirements and query time by splitting long words (similarly to
FastBlockSS which is a variant of the original method) and storing the neighborhood
implicitly with indexes and pointers to original dictionary entries. They claimed to be
faster than other approaches such as the aforementioned FastSS and the BK-tree [BK73].
Keyword Indexes
47
Recently, Chegrane and Belazzougui [CB14] described another practical index and they
reported better results when compared to Karch et al. Their structure is based on the
dictionary by Belazzougui [Bel09] for the edit distance of 1 (see the following subsection).
An approximate (in the mathematical sense) data structure for approximate matching
which is based on the Bloom filter (see Subsection 4.1.1) was also described [MW94a].
4.2.1
The 1-error problem
It is important to consider methods for detecting a single error, since over 80% of errors (even up to roughly 95%) are within k = 1 for the edit distance with transpositions [Dam64, PZ84]. Belazzougui and Venturini [BV12] presented a compressed index
whose space is bounded in terms of the k-th order empirical entropy of the indexed dictionary. It can be based either on perfect hashing, having O(m + occ) query time or
on a compressed permuterm index with O(m min(m, logσ n log log n) + occ) time (when
σ = logc n for some constant c) but improved space requirements. The former is a
compressed variant of a dictionary presented by Belazzougui [Bel09] which is based on
neighborhood generation and occupies O(n log σ) space and can answer queries in O(m)
time. Chung et al. [CTW14] showed a theoretical work where external memory is used,
and their focus is on I/O operations. They limited the number of these operations to
O(1 + m/wB + occ/B), where w is the size of the machine word and B is the number of
words within a block (a basic unit of I/O), and their structure occupies O(n/B) blocks.
In the category of filters, Mor and Fraenkel [MF82] described a method which is based
on the deletion-only 1-neighborhood.
For the 1-mismatch problem, Yao and Yao [YY95] described a data structure for binary strings of fixed length m with O(m log log |D|) query time and O(|D|m log m) space
requirements. This was later improved by Brodal and Gąsieniec [BG96] with a data
structure with O(m) query time which occupies O(n) space. This was improved with a
structure with O(1) query time and O(|D| log m) space in the cell probe model (where
only memory accesses are counted) [BV00]. Another notable example is a recent theoretical work of Chan and Lewenstein [CL15], who introduced the index with the optimal
query time, i.e. O(m/w+occ), which uses additional O(wd log1+ d) bits of space (beyond
the dictionary itself), assuming a constant-size alphabet.
4.2.2
Permuterm index
A permuterm index is a keyword index which supports queries with one wildcard symbol [Gar76]. The idea is store all rotations of a given word appended with the terminating
Keyword Indexes
48
character, for instance for the word text, the index would consist of the following permuterm vocabulary: text$, ext$t, xt$te, t$tex, $text. When it comes to searching, the query is first rotated so that the wildcard appears at the end, and subsequently
its prefix is searched for using the index. This could be for example a trie, or any other
data structure which supports a prefix lookup.
The main problem with the standard permuterm index is its space usage, as the number of strings inserted into the data structure is the number of words multiplied by
the average string length. Ferragina and Venturini [FV10] proposed a compressed permuterm index in order to overcome the limitations of the original structure with respect
to space. They explored the relation between the permuterm index and the BWT (see
Subsection 3.1.4.1), which is applied to the concatenation of all strings from the input
dictionary, and they provided a modification of the LF-mapping known from FM-indexes
in order to support the functionality of the permuterm index.
4.2.3
Split index
One of the practical approximate indexes was described by Cisłak (thesis author) and
Grabowski [CG15]. Experimental results for this structure can be found in Section 5.2.
As indexes supporting approximate matching tend to grow exponentially in k, the maximum number of allowed errors, it is also a worthwhile goal to design efficient indexes
supporting only a small k. For this reason, we focus on the problem of dictionary matching with few mismatches, especially one mismatch, where Ham(di , P ) 6 k for a collection
of words d ∈ D, a pattern P , and the Hamming distance Ham. The algorithm that we
are going to present is uncomplicated and based on the Dirichlet principle, ubiquitous
in approximate string matching techniques. We partition each word d into k + 1 disjoint
pieces p1 , . . . , pk+1 , of average length |d|/(k + 1) (hence the name “split index”), and
each such piece acts as a key in a hash table HT . The size of each piece pi of word d
is determined using the following formula: |pi | = b|d|/(k + 1)e (or |pi | = b|d|/(k + 1)c,
P
depending on the practical evaluation) and |pk+1 | = |d| − ki=1 |pi |, i.e. the piece size is
rounded to the nearest integer and the last piece covers the characters which are not in
other pieces. This means that the pieces might be in fact unequal in length, e.g., 3 and
2 for |d| = 5 ∧ k = 1. The values in HT are the lists of words which have one of their
pieces as the corresponding key. In this way, every word occurs on exactly k + 1 lists.
This seemingly bloats the space usage, still, in the case of small k the occupied space is
acceptable. Moreover, instead of storing full words on the respective lists, we only store
their “missing” prefix or suffix. For instance for the word table and k = 1, we would
Keyword Indexes
49
have a relation tab → le on one list (i.e. tab would be the key and le would be the
value) and le → tab on the other.
In the case of k = 1, we first populate each list with the pieces without the prefix and
then with the pieces without the suffix; additionally we store the position on the list (as
a 16-bit index) where the latter part begins. In this way, we traverse only a half of a
list on average during the search. We can also support k larger than 1 — in this case,
we ignore the piece order on a list, and we store dlog2 (k + 1)e bits with each piece that
indicate which part of the word is the list key. Let us note that this approach would also
work for k = 1, however, it turned out to be less efficient.
As regards the implementation (in the C++ language), our focus is on data compactness.
In the hash table, we store the buckets which contain word pieces as keys (e.g., le) and
pointers to the lists which store the missing pieces of the word (e.g., tab, ft). These
pointers are always located right next to the keys, which means that unless we are
very unlucky, a specific pointer should already be present in the CPU cache during the
traversal. The memory layouts of these substructures are fully contiguous. Successive
strings are represented by multiple characters with a prepended 8-bit counter which
specifies the length, and the counter with the value 0 indicates the end of the list.
During the traversal, each length can be compared with the length of the piece of the
pattern. As mentioned before, the words are partitioned into pieces of fixed length. This
means that on average we calculate the Hamming distance for only half of the pieces
on the list, since the rest can be ignored based on their length. Any hash function for
strings can be used, and two important considerations are the speed and the number of
collisions, since a high number of collisions results in longer buckets, which may in turn
have a negative effect on the query time (this subject is explored in more detail along
with the results in Chapter 5). Figure 4.3 illustrates the layout of the split index.
The preprocessing stage proceeds as follows:
1. Duplicate words are removed from the dictionary D.
The following steps refer to each word d from D:
2. The word d is split into k + 1 pieces.
3. For each piece pi : if pi ∈
/ HT , we create a new list ln containing the missing pieces
(later simply referred to as a missing piece; in the case of k = 1, this is always one
contiguous piece) P = {pj : j ∈ [1, k + 1] ∧ j 6= i} and add it to the hash table (we
append pi and the pointer to ln to the bucket). Otherwise, if pi ∈ HT , we append
the missing pieces P to the already existing list li .
Keyword Indexes
50
Figure 4.3: Split index for keyword indexing which shows the insertion of the word
table for k = 1. The index also stores the words left and tablet (only selected lists
containing pieces of these two words are shown), and L1 and L2 indicate pointers to
the respective lists. The first cell of each list indicates a 1-based word position (i.e. the
word count from the left) where the missing prefixes begin (k = 1, hence we deal with
two parts, namely prefixes and suffixes), and 0 means that the list has only missing
suffixes. Adapted from Wikimedia Commons1 .
As regards the search:
1. The pattern P is split into k + 1 pieces.
2. We search for each piece pi (the prefix and the suffix if k = 1): the list li is retrieved
from the hash table or we continue if pi ∈
/ HT . We traverse each missing piece pj
from li . If |pj | = |P | − |pi |, the verification is performed and the result is returned
if Ham(pj , P − pi ) 6 k.
3. The pieces are combined into one word in order to form the answer.
1
Jorge Stolfi, available at http://en.wikipedia.org/wiki/File:Hash_table_3_1_1_0_1_
0_0_SP.svg, CC A-SA 3.0.
Keyword Indexes
4.2.3.1
51
Complexity
P|D|
Let us consider the average word length |d|, where |d| = ( i=1 |di |)/|D|. The average
time complexity of the preprocessing stage is equal to O(kn), where k is the allowed
number of errors, and n is the total input dictionary size (i.e. the length of the concateP|D|
nation of all words from D, n = i=1 |di |). This is because for each word and for each
piece pi we either add the missing pieces to a new list or append them to the already
existing one in O(|d|) time (let us note that |D| · |d| = n). We assume that adding a new
element to the bucket takes constant time on average, and the calculation of all hashes
takes O(n) time in total. This is true irrespective of which list layout is used (there are
two layouts for k = 1 and k > 1, see the preceding paragraphs). The occupied space is
equal to O(kn), because each piece appears on exactly k lists and in exactly 1 bucket.
The average search complexity is equal to O(kt), where t is the average length of the
list. We search for each of k + 1 pieces of the pattern of length m, and when the
list corresponding to the piece p is found, it is traversed and at most t verifications
are performed. Each verification takes at most O(min(m, |dmax |)) time where dmax is
the longest word in the dictionary (or O(k) time, in theory, using the old technique
from Landau and Vishkin [LV89], after O(n log σ)-time preprocessing), but O(1) time
on average. Again, we assume that determining a location of the specific list, that is
iterating a bucket, takes O(1) time on average. As regards the list, its average length t is
higher when there is a higher probability that two words d1 and d2 from D have two parts
of the same length l which match exactly, i.e. P r(d1 [s1 , s1 + l − 1] = d2 [s2 , s2 + l − 1]).
Since all words are sampled from the same alphabet Σ, t depends on the alphabet size,
that is t = f (σ). Still, the dependence is rather indirect; in real-world dictionaries which
store words from a given language, t will be rather dependent on the k-th order entropy
of the language.
4.2.3.2
Compression
In order to reduce storage requirements, we apply a basic compression technique. We
find the most frequent q-grams in the word collection and replace their occurrences on
the lists with unused symbols, e.g., byte values 128, . . . , 255. The values of q can be
specified at the preprocessing stage, for instance q = 2 and q = 4 are reasonable for
the English alphabet and DNA, respectively. Different q values can be also combined
depending on the distribution of q-grams in the input text, i.e. we may try all possible
combinations of q-grams up to a certain q value and select ones which provide the best
compression. In such a case, longer q-grams should be encoded before shorter ones. For
example, a word compression could be encoded as #p*s\ using the following substitution
Keyword Indexes
52
list: com → #, re → *, co → $, om → &, sion → \ (note that not all q-grams from the
substitution list are used). Possibly even a recursive approach could be applied, although
this would certainly have a substantial impact on the query time. See Section 5.2 for the
experimental results and a further discussion.
The space usage could be further reduced by the use of a different character encoding.
For the DNA (assuming 4 symbols only) it would be sufficient to use 2 bits per character,
and for the basic English alphabet 5 bits. In the latter case there are 26 letters, which in
a simplified text can be augmented only with a space character, a few punctuation marks,
and a capital letter flag. Such an approach would be also beneficial for space compaction,
and it could have a further positive impact on cache usage. The compression naturally
reduces the space while increasing the search time, and a sort of a middle ground can
be achieved by deciding which additional information to store in the index. This can
be for instance the length of an encoded (compressed) piece after decoding, which could
eliminate some pieces based on their size without performing the decompression and
explicit verification.
4.2.3.3
Parallelization
The algorithm could be sped up by means of parallelization, since index access during the
search procedure is read-only. In the most straightforward approach, we could simply
distribute individual queries between multiple threads. A more fine-grained variation
would be to concurrently operate on word pieces after the word has been split up (with
the number of pieces being dependent on the k parameter). We could even access in
parallel the lists which contain missing pieces (prefixes and suffixes for k = 1), although
the gain would be probably limited since these lists usually store at most a few words.
If we had a sufficient amount of threads at our disposal, these approaches could be
combined. Still, it is to be noted that the use of multiple threads has a negative effect
on cache utilization.
4.2.3.4
Inverted split index
The split index could be extended in order to include the functionality of an inverted
index for approximate matching. As mentioned in Subsection 4.1.2, the inverted index
could be in practice any data structure which supports an efficient word lookup. Let us
consider the compact list layout of the split index presented in Figure 4.3, where each
piece is located right next to other pieces. Instead of storing only the 8-bit counter which
specifies the length of the piece, we could also store (right next to this piece) its position
in the text. Such an approach would increase the average length of the list only by
Keyword Indexes
53
constant factor and it would not break the contiguity of the lists, while also keeping the
O(kn) space complexity. Moreover, the position should be already present in the CPU
cache during the list traversal.
4.3
Keyword selection
Keyword indexes can be also used in the scenario where there are no explicit boundaries
between the words. In such a case, we would like to select the keywords according to
a well-defined set of rules and form a dictionary D from the input text T . Such an
index which stores q-grams sampled from the input text may be referred to as a q-gram
index. It is useful for answering keyword rather than full-text queries, which might be
required for example due to time requirements (i.e. when we would like to trade space
for speed). Examples of the input which cannot be easily divided into words include
some natural languages (e.g., Chinese) where it is not possible to clearly distinguish the
words (their boundaries depend on the context) or other kinds of data such as a complete
genome [KWLL05].
Let us consider the input text T which is divided into n − q + 1 q-grams. The issue
lies in the amount of space which is occupied by all q-gram tuples (s, li ), where s is
the q-gram and li identifies its positions, which is in the order of O(n) or O(nqmax )
for all possible q-grams up to some qmax value. General compression techniques are
usually not sufficient and thus a dedicated solution is required. This is especially the
case in the context of bioinformatics where data sets are substantial; the applications
could be for instance retrieving the seeds in the seed-and-extend algorithm described in
Section 3.3. One of the approaches was proposed by Kim et al. [KWLL05], and it aims to
eliminate the redundancy in position information. Consecutive q-grams are grouped into
subsequences, and each q-gram is identified by the position of the subsequence within the
documents and the position of a q-gram within the subsequence, which forms a two-level
index structure. This concept was also extended by the original authors to include the
functionality of approximate matching [KWL07].
4.3.1
Minimizers
The idea of minimizers was introduced by Roberts et al. [RHH+ 04] (with applications in, e.g., genome sequencing with de Bruijn graphs [CLJ+ 14] and k-mer counting [DKGDG15]), and it consists in storing only selected rather than all q-grams from
the input text. Here, the goal is to choose such q-grams from a given string S (a set
M (S)), so that for two strings S1 and S2 , if it holds for a pattern P that P ⊂ S1 ∧P ⊂ S2
54
Keyword Indexes
and |P | is above some threshold, then it should also hold that |M (S1 ) ∩ M (S2 )| > 1. In
order to find a (α, q)-minimizer, we slide a window of length q + α − 1 (α consecutive
q-grams) over T , shifting it by 1 character at a time, and at each window position we
select a q-gram which is the smallest one lexicographically (ties may be resolved for instance in favor of the leftmost of the smallest substrings). Figure 4.4 demonstrates this
process.
T
W1
W2
W3
W4
t
t
e
e
e
x
x
x
x
t
t
t
t
t
i
n
g
i
i
i
n
n
g
Figure 4.4: Selecting (3, 2)-minimizers (underlined), that is choosing 2-grams while
sliding a window of length 4 (3 + 2 − 1 = 4) over the text texting. The results belong
to the following set: {ex, in}.
Let us repeat an important property of the minimizers which makes them useful in
practice. If for two strings S1 and S2 it holds that P ⊂ S1 ∧ P ⊂ S2 ∧ |P | > q + α − 1,
then it is guaranteed that S1 and S2 share an (α, q)-minimizer (because they share one
full window). This means that for certain applications we can still ensure that no exact
matches are overlooked by storing the minimizers rather than all q-grams.
4.4
String sketches
We introduce the concept of string sketches, whose goal is to speed up string comparisons
at the cost of additional space. For a given string S, a sketch S 0 is constructed as
S 0 = f (S) using some function f which returns a fixed-sized block of data. In particular,
for two strings S1 and S2 , we would like to determine with certainty that S1 6= S2 or
Ham(S1 , S2 ) > k when comparing only sketches S10 and S20 . There exists a similarity
between sketches and hash functions, however, hash comparison would work only in the
context of exact matching. When the sketch comparison is not decisive, we still have
to perform an explicit verification on S1 and S2 , but the sketches allow for reducing
the number of such verifications. Since the sketches refer to individual words, they are
relevant in the context of keyword indexes. Assuming that each word d ∈ D is stored
along with d0 , sketches could be especially useful if the queries are known in advance or
|D| is relatively high, since sketch calculation might be time-consuming.
Sketches use individual bits in order to store information about q-gram frequencies in
the string. Various approaches exist, and main properties of the said q-grams include:
Keyword Indexes
55
• Size — for instance individual letters (1-grams) are sensible for the English alphabet but pairs (2-grams) might be better for the DNA.
• Frequency — we can store binary information in each bit that indicates whether a
certain q-gram appears in the string (8 q-grams in total for a 1-byte sketch, we call
this approach an occurrence sketch), or we can store their count (at most 3) using
2-bits per q-gram (4 q-grams in total for a 1-byte sketch, we call this approach a
count sketch).
• Selection — which q-grams should be included in the sketches. These could be
for instance q-grams which occur most commonly in the sample text.
For instance, let us consider an occurrence sketch which is built over 8 most common
letters of the English alphabet, namely {e, t, a, o, i, n, s, h} (consult Appendix E to see
the frequencies). For the word instance, the 1-byte sketch where each bit corresponds
to one of the letters from the aforementioned set would be as follows: 11101110.
We can quickly compare two sketches by taking a binary xor operation and counting
the number of bits which are set in the result (calculating the Hamming weight, HW ).
Note that HW can be determined in constant time using a lookup table of size 28n
bytes, where n is the sketch size in bytes. We denote the sketch difference with HS ,
and HS = HW (S10 ⊕ S20 ). Let us note that HS does not determine the number of
mismatches, for instance for S1 = run and S2 = ran, HS (S1 , S2 ) might be equal to 2
(occurrence differences in a and u) but there is still only one mismatch. On the other
extreme, for two strings of length n where each string consists of a repeated occurrence
of one different letter, HS might be equal to 1, but the number of mismatches is n. In
general, HS can be used to provide a lower bound on the true number of errors. For
sketches which record information about single characters (1-grams), the following holds:
Ham(S1 , S2 ) > dHS (S10 , S20 )/2e (the right-hand side can be calculated quickly using a
lookup table, since 0 6 HS 6 8). The true number of mismatches is underestimated
especially by count sketches, since we calculate the Hamming weight instead of comparing
the counts. For instance, for the count of 3 (bits 11) and the count of 1 (bits 01), the
difference is 1 instead of 2. Still, even though the true error is higher than HS , sketches
can be used in order to speed up the comparisons because certain strings will be compared
(and rejected) in constant time using fast bitwise operations and array lookups. As
regards the space overhead incurred by the sketches, it is equal to O(|D| + σ), since we
have to store one constant-size sketch per word together with the lookup tables which
are used to speed up the processing. Consult Section 5.3 in order to see the experimental
results.
Chapter 5
Experimental Results
The results were obtained on the machine with the following specifications:
• Processor — Intel i5-3230M running at 2.6 GHz
• RAM — 8 GB DDR3 memory
• Operating system — Ubuntu 14.04 64-bit (kernel version 3.16.0-41)
Programs were written in the C++ programming language (with certain prototypes in
the Python language) using features from the C++11 standard [cpp11]. They use the
C++ Standard Library, Boost libraries (version 1.57) [Kar05], and Linux system libraries.
Correctness was analyzed using Valgrind [NS03], a tool for error checking and profiling
(no errors or memory leaks were reported). The source code was compiled (as a 32-bit
version) with clang compiler v. 3.4-1, which turned out to be produce a slightly faster
executable than the gcc when checked under the -O3 optimization flag.
5.1
FM-bloated
For the description of the FM-bloated structure consult Subsection 3.1.6. Here, we
present experimental results for the superlinear index version. As regards the hash
function, xxhash was used (available on the Internet, consult Appendix F), and the load
factor was equal to 2.81.
The length of the pattern has a crucial impact on the search time, since the number
LF-mapping steps is equal the number of 1s in the binary representation of m. This
means that the search will be the fastest for m in the form 2c (constant time for m up
56
57
Experimental Results
to a certain maximum value) and the slowest for m in the form 2c − 1, where c ∈ Z. We
can see in Figure 5.1 that the query time also generally decreases as the pattern length
increases, mostly due to the fact that the times are given per character. The results are
the average times calculated for one million queries which were extracted from the input
text.
160
Query time per char (ns)
140
120
100
80
60
40
20
0
0
20
40
60
80
Pattern length
100
120
140
Figure 5.1: Query time per character vs pattern length (16, 20, 22, 24, 31, 32, 64, 80,
82, 85, 90, 100, 102, 126, 127, and 128) for the English text of size 30 MB. Let us point
out notable differences between pattern lengths 31 and 32, and 127 and 128.
We also compare our approach with other FM-index-based structures (consult Figure 5.2). We used the implementations from the sdsl library [GBMP14] (available on
the Internet [sds]) and the implementations of FM-dummy structures by Grabowski et
al. [GRD15] (available on the Internet [ran]). As regards the space, the FM-bloated structure (just as the name suggests) is roughly two order of magnitude bigger than other
indexes. The index size for other methods ranged from approximately 0.6n to 4.25n,
where n is the input text size. FM-bloated, on the other hand, occupied the amount of
space equal to almost 85n (for qmax = 128).
5.2
Split index
In this section we present the results which appeared in a preprint by Cisłak (thesis
author) and Grabowski [CG15]. For the description of the split index consult Subsection 4.2.3.
58
Experimental Results
FM-bloated superlinear
FM-index Huffman WT
CSA compressed bit vector
FM-dummy 2 256c
FM-dummy WT4 512
Query time per char (ns)
104
103
102
101
20
40
60
80
Pattern length
100
120
Figure 5.2: Query time per character vs pattern length (16, 24, 64, 80, and 120) for
different methods for the English text of size 30 MB. Note the logarithmic y-scale.
One of the crucial components of the split index is a hash function. Ideally, we would
like to minimize the average length of the bucket (let us recall that we use chaining for
collision resolution), however, the hash function should be also relatively fast because
it has to be calculated for each of k + 1 parts of the pattern (of total length m). We
investigated various hash functions, and it turned out that the differences in query times
are not negligible, although the average length of the bucket was almost the same in
all cases (relative differences were smaller than 1%). We can see in Table 5.1 that the
fastest function was the xxhash (available on the Internet, consult Appendix F), and for
this reason it was used for the calculation of other results.
Hash function
xxhash
sdbm
FNV1
FNV1a
SuperFast
Murmur3
City
FARSH
SpookyV2
Farm
Query time (µs)
0.93
0.95
0.95
0.95
0.96
0.97
0.99
1.00
1.04
1.04
Table 5.1: Evaluated hash functions and search times per query for the English
dictionary of size 2.67 MB and k = 1. A list of common English misspellings was used
as queries, max LF = 2.0.
59
Experimental Results
Decreasing the value of the load factor did not strictly provide a speedup in terms of the
query time, as demonstrated in Figure 5.3. This can be explained by the fact that even
though the relative reduction in the number of collisions was substantial, the absolute
difference was equal to at most a few collisions per list. Moreover, when the LF was
higher, pointers to the lists could be possibly closer to each other, which might have had
a positive effect on cache utilization. The best query time was reported for the maximum
LF value of 2.0, hence this value was used for the calculation of other results.
1.00
0.98
0.96
0.94
Index size (MB)
Query time ( s)
1.02
0
1
2
3
6.2
6.0
5.8
5.6
5.4
5.2
5.0
0
1
2
3
4
5
6
7
8
4
5
6
7
8
Load factor
Load factor
Figure 5.3: Query time and index size vs the load factor for the English dictionary
of size 2.67 MB and k = 1. A list of common English misspellings was used as queries.
The value of LF can be higher than 1.0 because we use chaining for collision resolution.
In Table 5.2 we can see a linear increase in the index size and an exponential increase
in query time with growing k. Even though we concentrate on k = 1 and the most
promising results are reported for this case, our index might remain competitive also for
higher k values.
k
1
2
3
Query time (µs)
0.51
11.49
62.85
Index size (KB)
1,715
2,248
3,078
Table 5.2: Query time and index size vs the error value k for the English language
dictionary of size 0.79 MB. A list of common English misspellings was used as queries.
Q-gram substitution coding provided a reduction in the index size, at the cost of increased
query time. Q-grams were generated separately for each dictionary D as a list of 100
60
Experimental Results
q-grams which provided the best compression for D, i.e. they minimized the size of
P|D|
all encoded words, SE =
i=1 |Enc(di )|. For the English language dictionaries, we
also considered using only 2-grams or only 3-grams, and for the DNA only 2-grams (a
maximum of 25 2-grams) and 4-grams, since mixing the q-grams of various sizes has a
further negative impact on the query time. For the DNA, 5,000 queries were generated
randomly by introducing noise into words sampled from dictionary, and their length was
equal to the length of the particular word. Up to 3 errors were inserted, each with a 50%
probability. For the English dictionaries we opted for the list of common misspellings,
and the results were similar to the case of randomly generated queries. The evaluation
was run 100 times and the results were averaged.
We can see the speed-to-space relation for the English dictionaries in Figure 5.4 and for
the DNA in Figure 5.5. In the case of English, using the optimal (from the compression
point of view, i.e. minimizing the index size) combination of mixed q-grams provided almost the same index size as using only 2-grams. Substitution coding methods performed
better for the DNA (where σ = 5) because the sequences are more repetitive. Let us note
that the compression provided a higher relative decrease in index size with respect to
the original text as the size of the dictionary increased. For instance, for the dictionary
of size 627.8 MB the compression ratio was equal to 1.93 and the query time was still
Index size (MB)
Query time ( s)
around 100 µs. Consult Appendix C for more information about the compression.
7
6
5
4
3
2
1
0
12
10
8
6
4
2
0
1
2
3
4
5
6
3
4
5
6
Dictionary size (MB)
No q-grams
2-grams
3-grams
Mixed q-grams
1
2
Dictionary size (MB)
Figure 5.4: Query time and index size vs dictionary size for k = 1, with and without
q-gram coding. Mixed q-grams refer to the combination of q-grams which provided the
best compression, and for the three dictionaries these were equal to ([2-, 3-, 4-] grams):
[88, 8, 4], [96, 2, 2], and [94, 4, 2], respectively. English language dictionaries and the
list of common English misspellings were used.
61
Index size (MB)
Query time ( s)
Experimental Results
120
100
80
60
40
20
0
800
700
600
500
400
300
200
100
0
0
100
200
300
400
500
600
300
400
500
600
Dictionary size (MB)
No q-grams
2-grams
4-grams
Mixed q-grams
0
100
200
Dictionary size (MB)
Figure 5.5: Query time and index size vs dictionary size for k = 1, with and without
q-gram coding. Mixed q-grams refer to the combination of q-grams which provided
the best compression, and these were equal to ([2-, 3-, 4-] grams): [16, 66, 18] (due to
computational constraints, they were calculated only for the first dictionary, but used
for all four dictionaries). DNA dictionaries and the randomly generated queries were
used.
Tested on the English language dictionaries, promising results were reported when compared to methods proposed by other authors. Others consider the Levenshtein distance
as the edit distance, whereas we use the Hamming distance, which puts us at the advantageous position. Still, the provided speedup is significant, and we believe that the
more restrictive Hamming distance is also an important measure of practical use (see
Subsection 2.1.1 for more information). The implementations of other authors are available on the Internet [boy, che]. As regards the results reported for the MF and Boytsov’s
Reduced alphabet neighborhood generation, it was not possible to accurately calculate
the size of the index (both implementations by Boytsov), and for this reason we used
rough ratios based on index sizes reported by Boytsov for similar dictionary sizes. Let
us note that we compare our algorithm with Chegrane and Belazzougui [CB14], who
published better results when compared to Karch et al. [KLS10], who in turned claimed
to be faster than other state-of-the-art methods. We have not managed to identify any
practice-oriented indexes for matching in dictionaries over any fixed alphabet Σ dedicated for the Hamming distance, which could be directly compared to our split index.
The times for the brute-force algorithm are not listed, since they were roughly 3 orders
of magnitude higher than the ones presented. Consult Figure 5.6 for details.
We also evaluated different word splitting schemes. For instance for k = 1, one could
62
Experimental Results
35
Our method
Our method w. compression
Chegrane and Belazzougui
Boytsov
Mor-Fraenkel
30
Query time ( s)
25
20
15
10
5
0
2
4
6
Index size (MB)
8
10
12
Figure 5.6: Query time vs index size for different methods. The method with compression encoded mixed q-grams. We used the Hamming distance, and the other authors
used the Levenshtein distance for k = 1. English language dictionaries of size 0.79 MB,
2.67 MB, and 5.8 MB were used as input, and the list of common misspellings was used
for queries.
split the word into two parts of different sizes, e.g., 6 → (2, 4) instead of 6 → (3, 3), however, unequal splitting methods caused slower queries when compared to the regular one.
As regards Hamming distance calculation, it turned out that the naive implementation
(i.e. simply iterating and comparing each character) was the fastest one. The compiler with automatic optimization was simply more efficient than other implementations
(e.g., ones based directly on SSE instructions) that we have investigated.
5.3
String sketches
String sketches which were introduced in Section 4.4 allow for faster string comparison,
since in certain cases we can deduce for two strings S1 and S2 that D(S1 , S2 ) > k for some
k without performing an explicit verification. In our implementation, a sketch comparison
requires performing one bitwise operation and one array lookup, i.e. 2 constant operations
in total. We analyze the comparison time between two strings using various sketch
types versus an explicit verification. The sketch is calculated once per query and it is
then reused for the comparison with consecutive words, i.e. we examine the situation
where a single query is compared against a dictionary of words. The dictionary size
for which a speedup was reported was around 100 words or more, since in the case of
fewer words sketch construction was too slow in relation with the comparisons. When the
63
Experimental Results
sketch comparison was not decisive a verification was performed and it contributed to the
elapsed time. The words were generated over the English alphabet (consult Appendix E
in order to see letter frequencies), and each sketch occupied 2 bytes (1-byte sketches were
not effective). Figures 5.7 and 5.8 contain the results for occurrence and count sketches,
respectively. Consult Appendix D for more information regarding the letter distribution
in the alphabet.
8
Comparison time (ns)
7
No sketches
Occurrence sketch common
Occurrence sketch mixed
Occurrence sketch rare
6
5
4
3
2
101
Word size
102
Figure 5.7: Comparison time vs word size for 1 mismatch using occurrence sketches
for words generated over the English alphabet. Each sketch occupies 2 bytes, and time
refers to average comparison time between a pair of words. Common sketches use 16
most common letters, rare sketches use 16 least common letters, and mixed sketches
use 8 most common and 8 least common letters. Note the logarithmic x-scale.
64
Experimental Results
7.0
6.5
Comparison time (ns)
6.0
No sketches
Count sketch common
Count sketch mixed
Count sketch rare
5.5
5.0
4.5
4.0
3.5
3.0
2.5
101
Word size
102
Figure 5.8: Comparison time vs word size for 1 mismatch using count sketches for
words generated over the English alphabet. Each sketch occupies 2 bytes, and time
refers to average comparison time between a pair of words. Common sketches use 16
most common letters, rare sketches use 16 least common letters, and mixed sketches
use 8 most common and 8 least common letters. Note the logarithmic x-scale.
Chapter 6
Conclusions
String searching algorithms are ubiquitous in computer science. They are used for common tasks performed on home PCs such as searching inside text documents or spell
checking as well as for industrial projects, e.g., genome sequencing. Strings can be defined very broadly, and they usually contain natural language and biological data (DNA,
proteins), but they can also represent various kinds of data such as music or images.
An interesting aspect of string matching is the diversity and complexity of the solutions
which have been presented over the years (both theoretical and practical), despite the
simplicity of problem formulation (one of the most common ones being “check if pattern
P exists in text T ”).
We investigated string searching methods which preprocess the input text and construct
a data structure called an index. This allows to reduce the time required for searching,
and it is often indispensable when it comes to massive sizes of modern data sets. The
indexes are divided into full-text ones which operate on the whole input text and can
answer arbitrary queries, and keyword indexes which store a dictionary of individual
q-grams (these can corresponds to, e.g., words in a natural language dictionary or DNA
reads).
Key contributions include the structure called FM-bloated, which is a modification of
the FM-index (a compressed, full-text index) that trades space for speed. Two variants of the FM-bloated were described — one using O(n log2 n) bits of space with
O(m + log m log log n) average query time, and one with linear space and O(m log log n)
average query time, where n is the input text length and m is the pattern length. We
experimentally show that by operating on q-grams in addition to individual characters a
significant speedup can be achieved (albeit at the cost of very high space requirements,
hence the name “bloated”).
65
Conclusions
66
The split index is a keyword index for the k-mismatches problem with a focus on the
1-error case. It performed better than other solutions for the Hamming distance, and
times in the order of 1 microsecond were reported for one mismatch for a few-megabyte
natural language dictionary on a medium-end PC. A minor contribution includes string
sketches which aim to speed up approximate string comparison at the cost of additional
space (O(1) per string).
6.1
Future work
We presented results for the superlinear variant of the FM-bloated index in order to
demonstrate its potential and capabilities. Multiple modifications and implementations
of this data structure can be introduced. Let us recall that we store the count table and
occurrence lists for selected q-grams in addition to individual characters from the regular
FM-index. This q-gram selection process can be fine-tuned — the more q-grams we store,
the faster the search should be, but the index size grows as well. For instance, the linear
space version could be augmented with additional 1-, 2-, etc q-grams which start at the
position of each minimizer, up to an s-gram where s is the maximum gap size between
two minimizers. This would eliminate two phases of the search (for prefixes and suffixes,
cf. Subsection 3.1.6.2) where individual characters have to be used for the LF-mapping
mechanism. Moreover, the comparison with other methods could be augmented with an
inverted index on q-grams, whose properties should be more similar to FM-bloated than
those of FM-index variants, especially when it comes to space requirements.
As regards the split index, we describe possible extensions in Subsections 4.2.3.3 and
4.2.3.4. These include using multiple threads and introducing the functionality of an
inverted index on q-grams. Moreover, the algorithm could be possibly extended to handle
the Levenshtein distance as well, although this would certainly have a substantial impact
on space usage. Another desired functionality could include a dedicated support for a
binary alphabet (σ = 2). In such a case, individual characters could be stored with bits,
which should have a positive effect on cache usage thanks to further data compaction
and possibly an alignment with the cache line size.
Appendix A
Data Sets
The following tables present information regarding the data sets that were used in
this work. Table A.1 describes data sets from the popular Pizza & Chili (P&C) corpus [FGNV09, piz], which were used for full-text indexes (English.30 was extracted
from English.50). Table A.2 describes data sets which were used for keyword indexes.
The English dictionaries come from Linux packages and the webpage by Foster [fos],
and the list of common misspellings which were used as queries was obtained from the
Wikipedia [typ]. The DNA dictionaries contain 20-mers which were extracted from the
genome of Drosophila melanogaster that was collected from the FlyBase database [fly].
The provided sizes refer to the size of the dictionary after preprocessing — for keyword
indexes, duplicates as well as delimiters (usually newline characters) are removed. The
abbreviation NL refers to natural language.
Name
English.30
English.50
English.200
Source
P&C
P&C
P&C
Type
NL (English)
NL (English)
NL (English)
Size
30 MB
50 MB
200 MB
Table A.1: A summary of data sets which were used for the experimental evaluation
of full-text indexes.
67
68
Data Sets
Name
iamerican
foster
iamerican-insane
misspellings
dmel-tiny
dmel-small
dmel-medium
dmel-big
Source
Linux package
Foster
Linux package
Wikipedia
FlyBase
FlyBase
FlyBase
FlyBase
NL
NL
NL
NL
Type
(English)
(English)
(English)
(English)
DNA
DNA
DNA
DNA
Size
0.79 MB
2.67 MB
5.8 MB
42.2 KB (4,261 words)
6.01 MB
135.89 MB
262.78 MB
627.80 MB
Table A.2: A summary of data sets which were used for the experimental evaluation
of keyword indexes.
Appendix B
Exact Matching Complexity
In the theoretical analysis we often mention exact string comparison, i.e. determining
whether S1 = S2 . It must hold that |S1 | = |S2 |, and the worst-case complexity of this
operation is equal to O(n) (all characters have to be compared when the two strings
match). On the other hand, the average complexity depends on the alphabet Σ. If, for
instance, σ = 2, we have 1/2 probability that characters S1 [0] and S2 [0] match, 1/4 that
characters S1 [1] and S2 [1] match as well, etc, in the case of uniform letter frequencies.
More generally, the probability that there is a match between all characters up to a 0based position i is equal to 1/σ i+1 , and the average number of required comparisons AC
is equal to 1+1/(σ −1) for any σ > 2. We can derive the following relation: lim AC = 1,
σ→∞
and hence treating the average time required for exact comparison of two random strings
from the same alphabet Σ as O(1) is justified for any σ. In Figure B.1 we present the
relation between the average number of comparisons and the σ value.
In the case of real-world Σ such as the English language alphabet, context information
in the form of k-th order entropy should be taken into account. In a simplified analysis,
let us consider the frequencies from Appendix E; the probability that two characters
sampled at random match is equal to 0.1272 (for a) + 0.0912 (for t), etc. Proceeding in
this manner, the probability for the match between the first pair of characters is equal to
6.55%, for the first and the second pair 0.43%, etc. As regards an empirical evaluation
on the English.200 text, the average number of comparisons between a random pair of
strings was equal to approximately 1.075.
69
70
Exact Matching Complexity
Avg number of comparisons
2.0
1.8
1.6
1.4
1.2
1.0
0
5
10
15
Alphabet size ( )
20
25
30
Figure B.1: Average number of character comparisons when comparing two random
strings (for exact matching) from the same alphabet with uniform letter frequency vs
the alphabet size σ.
Appendix C
Split Index Compression
This appendix presents additional information regarding the q-gram-based compression
of the split index (consult Subsection 4.2.3 for the description of this data structure and
Section 5.2 for the experimental results). In Figures C.1 and C.2 we can see the relation
between the index size and the selection of 100 2-grams and 3-grams for the English
alphabet (where the 2-grams clearly provided a better compression) and 100 3-grams
and 4-grams for the DNA.
4.9
Index size (MB)
4.8
4.7
4.6
4.5
4.4
4.3
0
20
40
60
2-grams count
80
100
Figure C.1: Index size vs the number of 2-grams used for the compression for the
English dictionary. 100 q-grams were used, and the remaining q-grams were 3-grams.
71
72
Split Index Compression
12.2
Index size (MB)
12.0
11.8
11.6
11.4
11.2
11.0
0
20
40
60
3-grams count
80
100
Figure C.2: Index size vs the number of 3-grams used for the compression for the
DNA dictionary. 100 q-grams were used, and the remaining q-grams were 4-grams.
Appendix D
String Sketches
In Section 5.3 we discussed the use of string sketches for the English alphabet, where
we could take advantage of the varying letter frequency. Here, we present the results for
the alphabet with uniform distribution and σ = 26. Instead of selecting the most or the
least common letters, the 2-byte sketches contain information regarding 16 (occurrence)
or 8 (count) randomly selected letters. We can see in Figure D.1 that in this case the
sketches do not provide the desired speedup.
6.5
No sketches
Occurrence sketch
Count sketch
Comparison time (ns)
6.0
5.5
5.0
4.5
4.0
3.5
3.0
101
Word size
102
Figure D.1: Comparison time vs word size for 1 mismatch using various string sketches
generated over the alphabet with uniform letter frequency and σ = 26. Each sketch
occupies 2 bytes, and time refers to average comparison time between a pair of words.
Note the logarithmic x-scale.
73
Appendix E
English Letter Frequency
Frequencies presented in Table E.1 [Lew00, p. 36] were used for the generation of random
queries where the letter distribution corresponded to the real-world English use.
Letter
e
t
a
o
i
n
s
h
r
d
l
c
u
Frequency
12.702%
9.056%
8.167%
7.507%
6.966%
6.749%
6.327%
6.094%
5.987%
4.253%
4.025%
2.782%
2.758%
Letter
m
w
f
g
y
p
b
v
k
j
x
q
z
Frequency
2.406%
2.361%
2.228%
2.015%
1.974%
1.929%
1.492%
0.978%
0.772%
0.153%
0.150%
0.095%
0.074%
Table E.1: Frequencies of English alphabet letters.
74
Appendix F
Hash Functions
Table F.1 contains Internet addresses of hash functions which were used to obtain experimental results for the split index (Section 5.2). If the hash function is not listed, it
means that our own implementation was used.
Name
City
Farm
FARSH
Murmur3
SpookyV2
SuperFast
xxhash
Address
https://code.google.com/p/cityhash/
https://code.google.com/p/farmhash/
https://github.com/Bulat-Ziganshin/FARSH
https://code.google.com/p/smhasher/wiki/MurmurHash3
http://burtleburtle.net/bob/hash/spooky.html
http://www.azillionmonkeys.com/qed/hash.html
https://code.google.com/p/xxhash/
Table F.1: A summary of Internet addresses of hash functions.
75
Bibliography
[AC75] Alfred V. Aho and Margaret J. Corasick. Efficient string matching: an aid
to bibliographic search. Communications of the ACM, 18(6):333–340, 1975.
[AE86] Stephen F. Altschul and Bruce W. Erickson. Optimal sequence alignment
using affine gap costs. Bulletin of Mathematical Biology, 48(5-6):603–616,
1986.
[AGM+ 90] Stephen F. Altschul, Warren Gish, Webb Miller, Eugene W. Myers, and
David J. Lipman. Basic local alignment search tool. Journal of Molecular
Biology, 215(3):403–410, 1990.
[AKO02] Mohamed Ibrahim Abouelhoda, Stefan Kurtz, and Enno Ohlebusch. The
enhanced suffix array and its applications to genome analysis. In Algorithms
in Bioinformatics, pages 449–463. Springer, 2002.
[AKO04] Mohamed Ibrahim Abouelhoda, Stefan Kurtz, and Enno Ohlebusch. Replacing suffix trees with enhanced suffix arrays. Journal of Discrete Algorithms, 2(1):53–86, 2004.
[AKO10] Alexandr Andoni, Robert Krauthgamer, and Krzysztof Onak. Polylogarithmic approximation for edit distance and the asymmetric query complexity.
In Foundations of Computer Science, 2010 51st Annual IEEE Symposium
on, pages 377–386. IEEE, 2010.
[ALP04] Amihood Amir, Moshe Lewenstein, and Ely Porat. Faster algorithms for
string matching with k mismatches. Journal of Algorithms, 50(2):257–275,
2004.
[AM05] Vo Ngoc Anh and Alistair Moffat. Inverted index compression using wordaligned binary codes. Information Retrieval, 8(1):151–166, 2005.
[AOK02] Mohamed Ibrahim Abouelhoda, Enno Ohlebusch, and Stefan Kurtz. Optimal exact string matching based on suffix arrays. In String Processing and
Information Retrieval, pages 31–43. Springer, 2002.
76
77
Bibliography
[Bar07] Gregory V. Bard. Spelling-error tolerant, order-independent pass-phrases
via the Damerau-Levenshtein string-edit distance metric. In Proceedings
of the fifth Australasian symposium on ACSW frontiers-Volume 68, pages
117–124. Australian Computer Society, Inc., 2007.
[BB93] Horst Bunke and Urs Bühler. Applications of approximate string matching
to 2D shape recognition. Pattern Recognition, 26(12):1797–1812, 1993.
[BBD09] Djamal Belazzougui, Fabiano C. Botelho, and Martin Dietzfelbinger. Hash,
displace, and compress. In Algorithms - ESA 2009, 17th Annual European Symposium, Copenhagen, Denmark, September 7-9, 2009. Proceedings, pages 682–693, 2009.
[BC15] Djamal Belazzougui and Fabio Cunial. Space-efficient detection of unusual
words. arXiv preprint arXiv:1508.02968, 2015.
[Bel09] Djamal Belazzougui. Faster and space-optimal edit distance “1” dictionary.
In Combinatorial Pattern Matching, pages 154–167. Springer, 2009.
[BG96] Gerth Stølting Brodal and Leszek Gąsieniec.
Approximate dictionary
queries. In Combinatorial Pattern Matching, pages 65–74. Springer, 1996.
[BHSH07] Thomas Bocek, Ela Hunt, Burkhard Stiller, and Fabio Hecht. Fast similarity search in large dictionaries. Technical Report ifi-2007.02, Department
of Informatics, University of Zurich, Switzerland, 2007.
[Bib] The Holy Bible. King James Version.
[BK73] Walter A. Burkhard and Robert M. Keller. Some approaches to best-match
file searching. Communications of the ACM, 16(4):230–236, 1973.
[BLA] BLAST: Basic local alignment search tool. http://blast.ncbi.nlm.nih.
gov/Blast.cgi. [Online; accessed 24-June-2015].
[Blo70] Burton H. Bloom. Space/time trade-offs in hash coding with allowable
errors. Communications of the ACM, 13(7):422–426, 1970.
[BM77] Robert S. Boyer and J. Strother Moore. A fast string searching algorithm.
Communications of the ACM, 20(10):762–772, 1977.
[BM00] Eric Brill and Robert C. Moore. An improved error model for noisy channel
spelling correction. In Proceedings of the 38th Annual Meeting on Association for Computational Linguistics, pages 286–293. Association for Computational Linguistics, 2000.
78
Bibliography
[BM04] Andrei Broder and Michael Mitzenmacher. Network applications of Bloom
filters: A survey. Internet Mathematics, 1(4):485–509, 2004.
[Bow10] Alexander Bowe. Multiary wavelet trees in practice. Honours thesis, RMIT
University, Australia, 2010.
[boy] Leonid Boytsov’s software.
http://searchivarius.org/personal/
software. [Online; accessed 12-March-2015].
[Boy11] Leonid Boytsov. Indexing methods for approximate dictionary searching:
Comparative analysis. Journal of Experimental Algorithmics, 16:1–1, 2011.
[BP98] Sergey Brin and Lawrence Page. The anatomy of a large-scale hypertextual
web search engine. Computer Networks and ISDN Systems, 30(1):107–117,
1998.
[BV00] Gerth Stølting Brodal and Srinivasan Venkatesh. Improved bounds for dictionary look-up with one error. Information Processing Letters, 75(1):57–
59, 2000.
[BV12] Djamal Belazzougui and Rossano Venturini. Compressed string dictionary
look-up with edit distance one. In Combinatorial Pattern Matching, pages
280–292. Springer, 2012.
[BW94] Michael Burrows and David J. Wheeler. A block-sorting lossless data compression algorithm. Technical Report 124, Systems Research Center, 1994.
[BYG92] Ricardo Baeza-Yates and Gaston H. Gonnet. A new approach to text
searching. Communications of the ACM, 35(10):74–82, 1992.
[BYN04] Ricardo Baeza-Yates and Gonzalo Navarro. Text searching: theory and
practice. In Formal languages and applications, pages 565–597. Springer,
2004.
[CB08] Manolis Christodoulakis and Gerhard Brey. Edit distance with singlesymbol combinations and splits. In Prague Stringology Conference 2008,
pages 208–217, 2008.
[CB14] Ibrahim Chegrane and Djamal Belazzougui. Simple, compact and robust
approximate string dictionary. Journal of Discrete Algorithms, 28:49–60,
2014.
[CFP+ 15] Raphaël Clifford, Allyx Fontaine, Ely Porat, Benjamin Sach, and Tatiana Starikovskaya. The k-mismatch problem revisited. arXiv preprint
arXiv:1508.00731, 2015.
79
Bibliography
[CG15] Aleksander Cisłak and Szymon Grabowski.
A practical index for ap-
proximate dictionary matching with few mismatches.
arXiv preprint
arXiv:1501.04948, 2015.
[CGGM03] Surajit Chaudhuri, Kris Ganjam, Venkatesh Ganti, and Rajeev Motwani.
Robust and efficient fuzzy match for online data cleaning. In Proceedings of
the 2003 ACM SIGMOD international conference on Management of data,
pages 313–324. ACM, 2003.
[CGL04] Richard Cole, Lee-Ad Gottlieb, and Moshe Lewenstein. Dictionary matching and indexing with errors and don’t cares. In Proceedings of the thirtysixth annual ACM symposium on Theory of computing, pages 91–100.
ACM, 2004.
[CH02] Richard Cole and Ramesh Hariharan. Approximate string matching: A
simpler faster algorithm. SIAM Journal on Computing, 31(6):1761–1782,
2002.
[che] Simple, compact and robust approximate string dictionary. https://code.
google.com/p/compact-approximate-string-dictionary/. [Online; accessed 12-March-2015].
[CIM+ 02] Maxime Crochemore, Costas S. Iliopoulos, Christos Makris, Wojciech Rytter, Athanasios K. Tsakalidis, and T. Tsichlas. Approximate string matching with gaps. Nordic Journal of Computing, 9(1):54–65, 2002.
[CKL06] Richard Cole, Tsvi Kopelowitz, and Moshe Lewenstein. Suffix trays and
suffix trists: structures for faster text indexing. In Automata, Languages
and Programming, pages 358–369. Springer, 2006.
[CL92] William I. Chang and Jordan Lampe. Theoretical and empirical comparisons of approximate string matching algorithms. In Combinatorial Pattern
Matching, pages 175–184. Springer, 1992.
[CL15] Timothy Chan and Moshe Lewenstein. Fast string dictionary lookup with
one error. In Combinatorial Pattern Matching, pages 114–123. Springer,
2015.
[Cla96] David Clark. Compact Pat Trees. PhD thesis, University of Waterloo,
Canada, 1996.
[CLJ+ 14] Rayan Chikhi, Antoine Limasset, Shaun Jackman, Jared T. Simpson, and
Paul Medvedev. On the representation of de Bruijn graphs. In Research in
Computational Molecular Biology, pages 35–55. Springer, 2014.
80
Bibliography
[CLRS09] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford
Stein. Introduction to Algorithms. The MIT Press, 3rd edition, 2009.
[CM94] William I. Chang and Thomas G. Marr. Approximate string matching
and local similarity. In Combinatorial Pattern Matching, pages 259–273.
Springer, 1994.
[CMEH13] Alejandro Chacón, Juan Carlos Moure, Antonio Espinosa, and Porfidio
Hernández. n-step FM-index for faster pattern matching. Procedia Computer Science, 18:70–79, 2013.
[CNO14] Francisco Claude, Gonzalo Navarro, and Alberto Ordóñez. Efficient compressed wavelet trees over large alphabets. arXiv preprint arXiv:1405.1220,
2014.
[CNP+ 12] Francisco Claude, Gonzalo Navarro, Hannu Peltola, Leena Salmela, and
Jorma Tarhio. String matching with alphabet sampling. Journal of Discrete
Algorithms, 11:37–50, 2012.
[Cob95] Archie L. Cobbs. Fast approximate matching using suffix trees. In Combinatorial Pattern Matching, pages 41–54. Springer, 1995.
[Col94] Richard Cole. Tight bounds on the complexity of the Boyer–Moore string
matching algorithm. SIAM Journal on Computing, 23(5):1075–1091, 1994.
[cpp11] Standard for programming language C++. Technical Report 14882:2011,
ISO/IEC, 2011.
[CPS12] J. Shane Culpepper, Matthias Petri, and Falk Scholer. Efficient in-memory
top-k document retrieval. In Proceedings of the 35th international ACM
SIGIR conference on Research and development in information retrieval,
pages 225–234. ACM, 2012.
[CTW14] Chin-Wan Chung, Yufei Tao, and Wei Wang. I/O-efficient dictionary search
with one edit error. In String Processing and Information Retrieval, pages
191–202. Springer, 2014.
[CV81] Lorinda L. Cherry and William Vesterman. Writing tools - the STYLE and
DICTION programs. Technical report, Rutgers University, USA, 1981.
[CW77] J. Lawrence Carter and Mark N. Wegman. Universal classes of hash functions. In Proceedings of the ninth annual ACM symposium on Theory of
computing, pages 106–112. ACM, 1977.
81
Bibliography
[Dam64] Fred J. Damerau. A technique for computer detection and correction of
spelling errors. Communications of the ACM, 7(3):171–176, 1964.
[Deo05] Sebastian Deorowicz.
Context exhumation after the Burrows–Wheeler
transform. Information Processing Letters, 95(1):313–320, 2005.
[DFG+ 97] Gautam Das, Rudolf Fleischer, Leszek Gąsieniec, Dimitris Gunopulos, and
Juha Kärkkäinen. Episode matching. In Combinatorial Pattern Matching,
pages 12–27. Springer, 1997.
[DFM98] George I. Davida, Yair Frankel, and Brian J. Matt. On enabling secure
applications through off-line biometric identification. In Security and Privacy, 1998. Proceedings. 1998 IEEE Symposium on, pages 148–157. IEEE,
1998.
[DK13] Mrinal Deo and Sean Keely. Parallel suffix array and least common prefix
for the GPU. In ACM SIGPLAN Notices, volume 48, pages 197–206. ACM,
2013.
[DKGDG15] Sebastian Deorowicz, Marek Kokot, Szymon Grabowski, and Agnieszka
Debudaj-Grabysz. KMC 2: Fast and resource-frugal k-mer counting. Bioinformatics, 31(10):1569–1576, 2015.
[DKM+ 94] Martin Dietzfelbinger, Anna Karlin, Kurt Mehlhorn, Friedhelm Meyer Auf
Der Heide, Hans Rohnert, and Robert E. Tarjan. Dynamic perfect hashing:
Upper and lower bounds. SIAM Journal on Computing, 23(4):738–761,
1994.
[Edd04] Sean R. Eddy. Where did the BLOSUM62 alignment score matrix come
from? Nature Biotechnology, 22(8):1035–1036, 2004.
[EL90] D.G. Elliman and I.T. Lancaster. A review of segmentation and contextual
analysis techniques for text recognition. Pattern Recognition, 23(3):337–
346, 1990.
[FAKM14] Bin Fan, Dave G. Andersen, Michael Kaminsky, and Michael D. Mitzenmacher. Cuckoo filter: Practically better than Bloom. In Proceedings of the
10th ACM International Conference on Emerging Networking Experiments
and Technologies, pages 75–88. ACM, 2014.
[Far97] Martin Farach. Optimal suffix tree construction with large alphabets. In
38th Annual Symposium on Foundations of Computer Science, FOCS ’97,
Miami Beach, Florida, USA, October 19-22, 1997, pages 137–143, 1997.
82
Bibliography
[FG09] Kimmo Fredriksson and Szymon Grabowski. Fast convolutions and their
applications in approximate string matching. In Combinatorial Algorithms,
pages 254–265. Springer, 2009.
[FGNV09] Paolo Ferragina, Rodrigo González, Gonzalo Navarro, and Rossano Venturini. Compressed text indexes: From theory to practice. Journal of
Experimental Algorithmics, 13:12, 2009.
[FKS84] Michael L Fredman, János Komlós, and Endre Szemerédi. Storing a sparse
table with O(1) worst case access time. Journal of the ACM, 31(3):538–544,
1984.
[FL13] Simone Faro and Thierry Lecroq. The exact online string matching problem: a review of the most recent results.
ACM Computing Surveys,
45(2):13, 2013.
[Flo98] Darren R. Flower. On the properties of bit string-based measures of chemical similarity. Journal of Chemical Information and Computer Sciences,
38(3):379–386, 1998.
[fly] Flybase homepage.
http://flybase.org/.
[Online; accessed 28-July-
2015].
[FM00] Paolo Ferragina and Giovanni Manzini. Opportunistic data structures with
applications. In Foundations of Computer Science, 2000. Proceedings. 41st
Annual Symposium on, pages 390–398. IEEE, 2000.
[FM05] Paolo Ferragina and Giovanni Manzini. Indexing compressed text. Journal
of the ACM, 52(4):552–581, 2005.
[FMMN07] Paolo Ferragina, Giovanni Manzini, Veli Mäkinen, and Gonzalo Navarro.
Compressed representations of sequences and full-text indexes.
ACM
Transactions on Algorithms, 3(2):20, 2007.
[fos] /foster/dictionary.
http://www.math.sjsu.edu/~foster/dictionary.
txt. [Online; accessed 28-July-2015].
[Fre60] Edward Fredkin. Trie memory. Communications of the ACM, 3(9):490–499,
1960.
[FV10] Paolo Ferragina and Rossano Venturini. The compressed permuterm index.
ACM Transactions on Algorithms, 7(1):10, 2010.
[Gal79] Zvi Galil. On improving the worst case running time of the Boyer–Moore
string matching algorithm. Communications of the ACM, 22(9):505–508,
1979.
83
Bibliography
[Gar76] Eugene Garfield. The permuterm subject index: An autobiographical review. Journal of the American Society for Information Science, 27(5):288–
291, 1976.
[GBMP14] Simon Gog, Timo Beller, Alistair Moffat, and Matthias Petri. From theory to practice: Plug and play with succinct data structures. In 13th
International Symposium on Experimental Algorithms, (SEA 2014), pages
326–337, 2014.
[GGV03] Roberto Grossi, Ankur Gupta, and Jeffrey Scott Vitter.
High-order
entropy-compressed text indexes. In Proceedings of the fourteenth annual
ACM-SIAM symposium on Discrete algorithms, pages 841–850. Society for
Industrial and Applied Mathematics, 2003.
[GLPS97] Patrick Girard, Christian Landrault, Serge Pravossoudovitch, and Daniel
Severac. Reduction of power consumption during test application by test
vector ordering. Electronics Letters, 33(21):1752–1754, 1997.
[GMN04] Szymon Grabowski, Veli Mäkinen, and Gonzalo Navarro.
First Huff-
man, then Burrows-Wheeler: A simple alphabet-independent FM-index.
In String Processing and Information Retrieval, pages 210–211. Springer,
2004.
[GMRS03] Alessandra Gabriele, Filippo Mignosi, Antonio Restivo, and Marinella
Sciortino. Indexing structures for approximate string matching. In Algorithms and Complexity, pages 140–151. Springer, 2003.
[GN09] Rodrigo González and Gonzalo Navarro.
Rank/select on dynamic
compressed sequences and applications. Theoretical Computer Science,
410(43):4414–4422, 2009.
[GNP+ 06] Szymon Grabowski, Gonzalo Navarro, Rafał Przywarski, Alejandro
Salinger, and Veli Mäkinen. A simple alphabet-independent FM-index.
International Journal of Foundations of Computer Science, 17(06):1365–
1384, 2006.
[GNPS15] Travis Gagie, Gonzalo Navarro, Simon J. Puglisi, and Jouni Sirén. Relative
compressed suffix trees. arXiv preprint arXiv:1508.02550, 2015.
[Gog11] Simon Gog. Compressed suffix trees: Design, construction, and applications. PhD thesis, University of Ulm, Germany, 2011.
[goo] How search works - the story. http://www.google.com/insidesearch/
howsearchworks/thestory/. [Online; accessed 21-April-2015].
84
Bibliography
[GP14] Simon Gog and Matthias Petri. Optimized succinct data structures for
massive data. Software: Practice and Experience, 44(11):1287–1314, 2014.
[GR14] Szymon Grabowski and Marcin Raniszewski. Sampling the suffix array
with minimizers. arXiv preprint arXiv:1406.2348, 2014.
[Gra11] Szymon Grabowski.
New Algorithms for Exact and Approximate Text
Matching. Zeszyty Naukowe Nr 1051. Politechnika Łódzka, available at:
http://cybra.lodz.pl/dlibra/doccontent?id=3528&from=FBC, 2011.
[Gra14] Szymon Grabowski. New tabulation and sparse dynamic programming
based techniques for sequence similarity problems. In Proceedings of the
Prague Stringology Conference 2014, Prague, Czech Republic, September
1-3, 2014, pages 202–211, 2014.
[GRD15] Szymon Grabowski, Marcin Raniszewski, and Sebastian Deorowicz. FMindex for dummies. arXiv preprint arXiv:1506.04896, 2015.
[Gre02] T. Ryan Gregory. Genome size and developmental complexity. Genetica,
115(1):131–146, 2002.
[Gus97] Dan Gusfield. Algorithms on strings, trees and sequences: computer science
and computational biology. Cambridge University Press, 1997.
[GV05] Roberto Grossi and Jeffrey Scott Vitter. Compressed suffix arrays and
suffix trees with applications to text indexing and string matching. SIAM
Journal on Computing, 35(2):378–407, 2005.
[HCZ+ 15] Hongwei Huo, Longgang Chen, Heng Zhao, Jeffrey Scott Vitter, Yakov
Nekrich, and Qiang Yu. A data-aware FM-index. In Proceedings of the Seventeenth Workshop on Algorithm Engineering and Experiments, ALENEX
2015, San Diego, CA, USA, January 5, 2015, pages 10–23, 2015.
[Hea78] Harold Stanley Heaps. Information retrieval: Computational and theoretical aspects. Academic Press, Inc., 1978.
[Hir75] Daniel S. Hirschberg. A linear space algorithm for computing maximal
common subsequences. Communications of the ACM, 18(6):341–343, 1975.
[Hor80] R. Nigel Horspool. Practical fast searching in strings. Software: Practice
and Experience, 10(6):501–506, 1980.
[Huf52] David A. Huffman. A method for the construction of minimum redundancy
codes. Proceedings of the IRE, 40(9):1098–1101, 1952.
85
Bibliography
[Jac89] Guy Jacobson. Space-efficient static trees and graphs. In Foundations
of Computer Science, 1989, 30th Annual Symposium on, pages 549–554.
IEEE, 1989.
[Kär95] Juha Kärkkäinen. Suffix cactus: A cross between suffix tree and suffix
array. In Combinatorial Pattern Matching, pages 191–204. Springer, 1995.
[Kar05] Björn Karlsson. Beyond the C++ standard library: an introduction to
Boost. Addison-Wesley, 2005.
[KCO+ 01] Stefan Kurtz, Jomuna V. Choudhuri, Enno Ohlebusch, Chris Schleiermacher, Jens Stoye, and Robert Giegerich. REPuter: the manifold applications of repeat analysis on a genomic scale. Nucleic Acids Research,
29(22):4633–4642, 2001.
[KKP14] Juha Kärkkäinen, Dominik Kempa, and Simon J. Puglisi. Hybrid compression of bitvectors for the FM-index. In Data Compression Conference,
2014, pages 302–311. IEEE, 2014.
[KLS10] Daniel Karch, Dennis Luxen, and Peter Sanders. Improved fast similarity
search in dictionaries. In String Processing and Information Retrieval, pages
173–178. Springer, 2010.
[KMP77] Donald E. Knuth, James H. Morris, and Vaughan R. Pratt. Fast pattern
matching in strings. SIAM Journal on Computing, 6(2):323–350, 1977.
[Knu97] Donald E. Knuth. The Art of Computer Programming, volume 2. AddisonWesley, 1997.
[Kor06] Jesse Kornblum. Identifying almost identical files using context triggered
piecewise hashing. Digital Investigation, 3:91–97, 2006.
[KR87] Richard M. Karp and Michael O. Rabin. Efficient randomized patternmatching algorithms.
IBM Journal of Research and Development,
31(2):249–260, 1987.
[KS94] Sandeep Kumar and Eugene H Spafford. A pattern matching model for
misuse intrusion detection. Technical Report 94-071, Department of Computer Science, Purdue University, USA, 1994.
[KU96] Juha Kärkkäinen and Esko Ukkonen. Sparse suffix trees. In Computing
and Combinatorics, pages 219–230. Springer, 1996.
[Kur99] Stefan Kurtz. Reducing the space requirement of suffix trees. Software:
Practice and Experience, 29(13):1149–71, 1999.
86
Bibliography
[KWL07] M. Kim, K Whang, and J. Lee. n-gram/2L-approximation: a two-level ngram inverted index structure for approximate string matching. Computer
Systems Science and Engineering, 22(6):365, 2007.
[KWLL05] Min-Soo Kim, Kyu-Young Whang, Jae-Gil Lee, and Min-Jae Lee.
n-
gram/2L: A space and time efficient two-level n-gram inverted index structure. In Proceedings of the 31st international conference on Very large data
bases, pages 325–336. VLDB Endowment, 2005.
[Lev66] Vladimir I. Levenshtein. Binary codes capable of correcting deletions, insertions, and reversals. In Soviet Physics Doklady, volume 10, pages 707–710,
1966.
[Lew00] Robert Lewand. Cryptological mathematics. MAA, 2000.
[LLL+ 12] Lin Liu, Yinhu Li, Siliang Li, Ni Hu, Yimin He, Ray Pong, Danni Lin,
Lihua Lu, and Maggie Law. Comparison of next-generation sequencing
systems. BioMed Research International, 2012, 2012.
[LP85] David J. Lipman and William R. Pearson. Rapid and sensitive protein
similarity searches. Science, 227(4693):1435–1441, 1985.
[LSS01] Gad M. Landau, Jeanette P. Schmidt, and Dina Sokol. An algorithm for
approximate tandem repeats. Journal of Computational Biology, 8(1):1–18,
2001.
[LTP+ 09] Ben Langmead, Cole Trapnell, Mihai Pop, Steven L. Salzberg, et al. Ultrafast and memory-efficient alignment of short DNA sequences to the human
genome. Genome Biology, 10(3):R25, 2009.
[LV89] Gad M. Landau and Uzi Vishkin. Fast parallel and serial approximate
string matching. Journal of Algorithms, 10(2):157–169, 1989.
[Mäk00] Veli Mäkinen. Compact suffix array. In Combinatorial Pattern Matching,
pages 305–319. Springer, 2000.
[ME10] Tyler Moore and Benjamin Edelman. Measuring the perpetrators and funders of typosquatting. In Financial Cryptography and Data Security, pages
175–191. Springer, 2010.
[MF82] Moshe Mor and Aviezri S. Fraenkel. A hash code method for detecting and
correcting spelling errors. Communications of the ACM, 25(12):935–938,
1982.
87
Bibliography
[MF04] Giovanni Manzini and Paolo Ferragina. Engineering a lightweight suffix
array construction algorithm. Algorithmica, 40(1):33–50, 2004.
[MG14] Alistair Moffat and Simon Gog. String search experimentation using massive data.
Philosophical Transactions of the Royal Society of London
A: Mathematical, Physical and Engineering Sciences, 372(2016):20130135,
2014.
[MGSA06] Aleksandr Morgulis, E. Michael Gertz, Alejandro A. Schäffer, and Richa
Agarwala. A fast and symmetric DUST implementation to mask lowcomplexity DNA sequences. Journal of Computational Biology, 13(5):1028–
1040, 2006.
[MHP05] Bořivoj Melichar, Jan Holub, and Tomáš Polcar. Text searching algorithms.
Department of Computer Science and Engineering, Czech Technical University in Prague, Czech Republic, 2005.
[Mit87] Roger Mitton. Spelling checkers, spelling correctors and the misspellings of
poor spellers. Information Processing & management, 23(5):495–505, 1987.
[MJDS07] Gurmeet Singh Manku, Arvind Jain, and Anish Das Sarma. Detecting
near-duplicates for web crawling. In Proceedings of the 16th international
conference on World Wide Web, pages 141–150. ACM, 2007.
[MM93] Udi Manber and Gene Myers. Suffix arrays: a new method for on-line
string searches. SIAM Journal on Computing, 22(5):935–948, 1993.
[MN05a] Moritz G. Maaß and Johannes Nowak. Text indexing with errors. In
Combinatorial Pattern Matching, pages 21–32. Springer, 2005.
[MN05b] Veli Mäkinen and Gonzalo Navarro.
Succinct suffix arrays based on
run-length encoding. In Combinatorial Pattern Matching, pages 45–56.
Springer, 2005.
[Mor68] Donald R. Morrison. PATRICIA – practical algorithm to retrieve information coded in alphanumeric. Journal of the ACM, 15(4):514–534, 1968.
[MP69] Marvin Minsky and Seymour Papert. Perceptrons. MIT Press, Cambridge,
Massachusetts, 1969.
[MP08] Michael A. Maniscalco and Simon J. Puglisi. An efficient, versatile approach
to suffix sorting. Journal of Experimental Algorithmics, 12:1–2, 2008.
[MV08] Svetlin A. Manavski and Giorgio Valle. CUDA compatible GPU cards as
efficient hardware accelerators for Smith-Waterman sequence alignment.
BMC Bioinformatics, 9(Suppl 2):S10, 2008.
88
Bibliography
[MW94a] Udi Manber and Sun Wu. An algorithm for approximate membership checking with application to password security. Information Processing Letters,
50(4):191–197, 1994.
[MW94b] Udi Manber and Sun Wu. GLIMPSE: A tool to search through entire file
systems. In Usenix Winter, pages 23–32, 1994.
[Mye99] Gene Myers. A fast bit-vector algorithm for approximate string matching
based on dynamic programming. Journal of the ACM, 46(3):395–415, 1999.
[Nav01] Gonzalo Navarro. A guided tour to approximate string matching. ACM
Computing Surveys, 33(1):31–88, 2001.
[NBYST01] Gonzalo Navarro, Ricardo A. Baeza-Yates, Erkki Sutinen, and Jorma
Tarhio. Indexing methods for approximate string matching. IEEE Data
Engineering Bulletin, 24(4):19–27, 2001.
[NC07] Alexandros Ntoulas and Junghoo Cho. Pruning policies for two-tiered inverted index with correctness guarantee. In Proceedings of the 30th annual
international ACM SIGIR conference on Research and development in information retrieval, pages 191–198. ACM, 2007.
[NM07] Gonzalo Navarro and Veli Mäkinen. Compressed full-text indexes. ACM
Computing Surveys, 39(1):2, 2007.
[Non13] Ge Nong. Practical linear-time o(1)-workspace suffix sorting for constant
alphabets. ACM Transactions on Information Systems, 31(3):15, 2013.
[NR03] Gonzalo Navarro and Mathieu Raffinot. Fast and simple character classes
and bounded gaps pattern matching, with applications to protein searching.
Journal of Computational Biology, 10(6):903–923, 2003.
[NS03] Nicholas Nethercote and Julian Seward. Valgrind: A program supervision
framework. Electronic Notes in Theoretical Computer Science, 89(2):44–66,
2003.
[NW70] Saul B. Needleman and Christian D. Wunsch. A general method applicable
to the search for similarities in the amino acid sequence of two proteins.
Journal of Molecular Biology, 48(3):443–453, 1970.
[OV03] Christos A. Ouzounis and Alfonso Valencia. Early bioinformatics: the birth
of a discipline — a personal view. Bioinformatics, 19(17):2176–2190, 2003.
[Pet80] James L. Peterson.
Computer programs for detecting and correcting
spelling errors. Communications of the ACM, 23(12):676–687, 1980.
89
Bibliography
[Pet86] James L. Peterson. A note on undetected typing errors. Communications
of the ACM, 29(7):633–637, 1986.
[PH89] Alan Parker and O. Hamblen, James. Computer algorithms for plagiarism
detection. IEEE Transactions on Education, 32(2):94–99, 1989.
[piz] Pizza&Chili corpus – compressed indexes and their testbeds.
http:
//pizzachili.dcc.uchile.cl/. [Online; accessed 09-June-2015].
[PJT09] Victor Pankratius, Ali Jannesari, and Walter F. Tichy. Parallelizing bzip2:
A case study in multicore software engineering. Software, IEEE, 26(6):70–
77, 2009.
[PST07] Simon J. Puglisi, William F. Smyth, and Andrew H. Turpin. A taxonomy
of suffix array construction algorithms. ACM Computing Surveys, 39(2):4,
2007.
[PZ84] Joseph J. Pollock and Antonio Zamora. Automatic spelling correction in
scientific and scholarly text. Communications of the ACM, 27(4):358–368,
1984.
[Rab81] Michael O. Rabin. Fingerprinting by random polynomials. Technical Report TR-15-81, Department of Mathematics, The Hebrew University of
Jerusalem, Israel, 1981.
[ran] /indexes/fm. http://ranisz.iis.p.lodz.pl/indexes/fm/. [Online; accessed 24-July-2015].
[RHH+ 04] Michael Roberts, Wayne Hayes, Brian R. Hunt, Stephen M. Mount, and
James A. Yorke. Reducing storage requirements for biological sequence
comparison. Bioinformatics, 20(18):3363–3369, 2004.
[RN10] Stuart Russell and Peter Norvig. Artificial Intelligence: a Modern Approach. Prentice Hall, 3rd edition, 2010.
[RRR02] Rajeev Raman, Venkatesh Raman, and S. Srinivasa Rao. Succinct indexable dictionaries with applications to encoding k-ary trees and multisets.
In Proceedings of the thirteenth annual ACM-SIAM symposium on Discrete
algorithms, pages 233–242. Society for Industrial and Applied Mathematics,
2002.
[Sal04] David Salomon. Data compression: the complete reference. Springer Science
& Business Media, 2004.
90
Bibliography
[SD10] Jared T. Simpson and Richard Durbin. Efficient construction of an assembly string graph using the FM-index. Bioinformatics, 26(12):i367–i373,
2010.
[sds] sdsl-lite. https://github.com/simongog/sdsl-lite. [Online; accessed
24-July-2015].
[Sew] Julian Seward. bzip2. http://bzip.org/. [Online; accessed 13-July-2015].
[Sha48] Claude Elwood Shannon. A mathematical theory of communication. The
Bell Systems Technical Journal, 27(3):379–423, 1948.
[Ski98] Steven S. Skiena. The algorithm design manual, volume 1. Springer Science
& Business Media, 1998.
[SRY+ 96] Eugene G. Shpaer, Max Robinson, David Yee, James D. Candlin, Robert
Mines, and Tim Hunkapiller. Sensitivity and selectivity in protein similarity
searches: a comparison of Smith–Waterman in hardware to BLAST and
FASTA. Genomics, 38(2):179–191, 1996.
[ST95] Erkki Sutinen and Jorma Tarhio. On using q-gram locations in approximate
string matching. In Algorithms - ESA ’95, Third Annual European Symposium, Corfu, Greece, September 25-27, 1995, Proceedings, pages 327–340,
1995.
[SW81] Temple F. Smith and Michael S. Waterman. Identification of common
molecular subsequences. Journal of Molecular Biology, 147(1):195–197,
1981.
[SW96] Fei Shi and Peter Widmayer. Approximate multiple string searching by
clustering. Genome Informatics, 7:33–40, 1996.
[TSCV04] Nathan Tuck, Timothy Sherwood, Brad Calder, and George Varghese. Deterministic memory-efficient string matching algorithms for intrusion detection. In INFOCOM 2004. Twenty-third Annual Joint Conference of the
IEEE Computer and Communications Societies, volume 4, pages 2628–
2639. IEEE, 2004.
[Tsu10] Dekel Tsur. Fast index for approximate string matching. Journal of Discrete
Algorithms, 8(4):339–345, 2010.
[Tur50] Alan Turing. Computing machinery and intelligence. Mind, 59:433–460,
1950.
91
Bibliography
[typ] Lists of common misspellings.
http://en.wikipedia.org/wiki/
Wikipedia:Lists_of_common_misspellings/For_machines. [Online; accessed 28-July-2015].
[Ukk85a] Esko Ukkonen. Algorithms for approximate string matching. Information
and Control, 64(1):100–118, 1985.
[Ukk85b] Esko Ukkonen. Finding approximate patterns in strings. Journal of Algorithms, 6(1):132–137, 1985.
[Ukk95] Esko Ukkonen.
On-line construction of suffix trees.
Algorithmica,
14(3):249–260, 1995.
[uni] UniProt. http://www.uniprot.org/. [Online; accessed 21-April-2015].
[Vig08] Sebastiano Vigna. Broadword implementation of rank/select queries. In
Experimental Algorithms, pages 154–168. Springer, 2008.
[Wei73] Peter Weiner. Linear pattern matching algorithms. In Switching and Automata Theory, 1973. SWAT’08. IEEE Conference Record of 14th Annual
Symposium on, pages 1–11. IEEE, 1973.
[Wil83] Dan E. Willard. Log-logarithmic worst-case range queries are possible in
space θ(n). Information Processing Letters, 17(2):81–84, 1983.
[WJ94] Lusheng Wang and Tao Jiang. On the complexity of multiple sequence
alignment. Journal of Computational Biology, 1(4):337–348, 1994.
[WMM96] Sun Wu, Udi Manber, and Gene Myers. A subquadratic algorithm for
approximate limited expression matching. Algorithmica, 15(1):50–67, 1996.
[YY95] Andrew Chi-Chih Yao and F. Frances Yao. Dictionary look-up with small
errors. In Combinatorial Pattern Matching, pages 387–394, 1995.
List of Symbols
B block — either an I/O unit or a fixed-size piece of a bit vector
c character (a string of length 1)
C count table in the FM-index
CL cache line size
D distance metric
δ time required for calculating D(S1 , S2 ) for two strings over the same alphabet Σ
D dictionary of keywords (for keyword indexes)
d word (string) from a dictionary
Enc(d) encoded (compressed) word d
F first column of the BWT matrix
H hash function
Hk (S) k-th order entropy of string S
HT hash table
HW Hamming weight (number of 1s in a bit vector)
Ham Hamming distance
bie number i rounded to the nearest integer
I index for string matching
k number of errors (in approximate matching)
L last column of the BWT matrix
LF load factor
92
List of Symbols
Lev Levenshtein distance
m pattern size, m = |P |
n input size
occ number of occurrences of the pattern
M (S) set of minimizers of string S
P pattern
p piece of a word in the case of word partitioning
P r(e) probability of event e
Q collection of q-grams
S string
S 0 string sketch over S
s substring
S set of substrings (for full-text indexes)
SA suffix array
Σ alphabet
Σ∗ set of all strings over the alphabet Σ
σ alphabet size, σ = |Σ|
T input string (text)
TR RRR table
T bwt input text T after applying the BWT
w size of the machine word (typically 32 or 64 bits)
v bit vector
93
List of Abbreviations
BF Bloom filter
LF load factor
BM Boyer–Moore algorithm
MF Mor–Fraenkel algorithm
BMH Boyer–Moore–Horspool algorithm
MPHF minimal perfect hash function
BST binary search tree
NL natural language
BWT Burrows–Wheeler transform
NW Needleman–Wunsch algorithm
CSA compressed suffix array
OCR optical character recognition
CoSA compact suffix array
P&C Pizza & Chili corpus
DFS depth-first search
PiES partitioning into exact searching
DP dynamic programming
RK Rabin–Karp algorithm
ESA enhanced suffix array
SA suffix array
FSM finite state machine
SC suffix cactus
I/O input/output
TLB translation lookaside buffer
KMP Knuth–Morris–Pratt algorithm
ST suffix tree
LCP longest common prefix
SW Smith–Waterman algorithm
LCS longest common subsequence
WT wavelet tree
94
List of Figures
1.1
1.2
1.3
1.4
A binary search tree (BST) storing strings from the English alphabet.
A trie which is one of the basic structures used in string searching. . .
A hash table for strings. . . . . . . . . . . . . . . . . . . . . . . . . . .
The formula for Shannon’s entropy. . . . . . . . . . . . . . . . . . . . .
2.1
Calculating an alignment with Levenshtein distance using the Needleman–
Wunsch (NW) algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.1
3.2
3.3
3.8
3.9
3.10
3.11
3.12
3.13
A suffix tree (ST) which stores all suffixes of a given text. . . . . . . . . .
A suffix array (SA) which stores indexes of sorted suffixes of a given text.
A compressed suffix array (CSA) which stores indexes pointing to the next
suffix from the text. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Calculating the Burrows–Wheeler transform (BWT). . . . . . . . . . . . .
A relation between the BWT and the SA. . . . . . . . . . . . . . . . . . .
Count table C which is a part of the FM-index. . . . . . . . . . . . . . . .
Formulae for updating the range during the search procedure in the FMindex. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
An example of RRR blocks. . . . . . . . . . . . . . . . . . . . . . . . . . .
An example of an RRR table. . . . . . . . . . . . . . . . . . . . . . . . . .
A wavelet tree (WT). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Q-gram extraction in the FM-bloated structure with superlinear space. . .
Constructing FM-bloated phrases with the use of minimizers. . . . . . . .
Selecting all overlapping q-grams from a given text. . . . . . . . . . . . . .
33
34
34
35
37
39
42
4.1
4.2
4.3
4.4
A Bloom filter (BF) for approximate membership queries. . .
An inverted index which stores a mapping from words to their
Split index for keyword indexing. . . . . . . . . . . . . . . . .
Selecting minimizers from a given text. . . . . . . . . . . . . .
44
45
50
54
5.1
Query time per character vs pattern length for the English text of size
30 MB (FM-bloated). . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Query time per character vs pattern length for different methods for the
English text of size 30 MB (FM-bloated). . . . . . . . . . . . . . . . . . .
Query time and index size vs the load factor (split index). . . . . . . . .
Query time and index size vs dictionary size, with and without q-gram
coding for English dictionaries (split index). . . . . . . . . . . . . . . . .
Query time and index size vs dictionary size, with and without q-gram
coding for DNA dictionaries (split index). . . . . . . . . . . . . . . . . .
Query time vs index size for different methods (split index). . . . . . . .
3.4
3.5
3.6
3.7
5.2
5.3
5.4
5.5
5.6
95
.
.
.
.
. . . . . .
positions.
. . . . . .
. . . . . .
. 8
. 9
. 11
. 12
.
.
.
.
27
28
29
32
33
33
. 57
. 58
. 59
. 60
. 61
. 62
List of Figures
5.7
5.8
96
Comparison time vs word size for 1 mismatch using occurrence sketches
for words generated over the English alphabet. . . . . . . . . . . . . . . . 63
Comparison time vs word size for 1 mismatch using count sketches for
words generated over the English alphabet. . . . . . . . . . . . . . . . . . 64
B.1 Average number of character comparisons when comparing two random
strings from the same alphabet with uniform letter frequency vs the alphabet size. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
C.1 Index size vs the number of 2-grams used for the compression for the
English dictionary (split index). . . . . . . . . . . . . . . . . . . . . . . . . 71
C.2 Index size vs the number of 3-grams used for the compression for the DNA
dictionary (split index). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
D.1 Comparison time vs word size for 1 mismatch using various string sketches
generated over the alphabet with uniform letter frequency and σ = 26. . . 73
List of Tables
1.1
A comparison of the complexities of basic data structures which can be
used for exact string searching. . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1
Algorithm classification based on whether the data is preprocessed. . . . . 18
5.1
5.2
Evaluated hash functions and search times per query (split index). . . . . 58
Query time and index size vs the error value k (split index). . . . . . . . . 59
A.1 A summary of data
of full-text indexes.
A.2 A summary of data
of keyword indexes.
sets which were used
. . . . . . . . . . . .
sets which were used
. . . . . . . . . . . .
for
. .
for
. .
the experimental
. . . . . . . . . .
the experimental
. . . . . . . . . .
evaluation
. . . . . . . 67
evaluation
. . . . . . . 68
E.1 Frequencies of English alphabet letters. . . . . . . . . . . . . . . . . . . . . 74
F.1 A summary of Internet addresses of hash functions. . . . . . . . . . . . . . 75
97
| 8cs.DS
|
arXiv:1511.06967v1 [math.AC] 22 Nov 2015
ARTIN APPROXIMATION PROPERTY AND THE GENERAL
NERON DESINGULARIZATION
DORIN POPESCU
Abstract. This is an exposition on the General Neron Desingularization and its
applications. We end with a recent constructive form of this desingularization in
dimension one.
Key words : Artin approximation, Neron Desingularization, Bass-Quillen Conjecture, Quillen’s Question, smooth morphisms, regular morphisms, smoothing ring
morphisms.
2010 Mathematics Subject Classification: Primary 1302, Secondary 13B40, 13H05,
13H10, 13J05, 13J10, 13J15, 14B07, 14B12, 14B25.
Introduction
Let K be a field and R = Khxi, x = (x1 , . . . , xm ) be the ring of algebraic power
series in x over K, that is the algebraic closure of the polynomial ring K[x] in the
formal power series ring R̂ = K[[x]]. Let f = (f1 , . . . , fq ) in Y = (Y1 , . . . , Yn ) over
R and ŷ be a solution of f in the completion R̂ of R.
Theorem 1 (M. Artin [Artin69]). For any c ∈ N there exists a solution y (c) in R
such that y (c) ≡ ŷ mod (x)c .
In general we say that a local ring (A, m) has the Artin approximation property
if for every system of polynomials f = (f1 , . . . , fq ) ∈ A[Y ]q , Y = (Y1 , . . . , Yn ), a
solution ŷ of f in the completion  and c ∈ N there exists a solution y (c) in A of f
such that y (c) ≡ ŷ mod mc . In fact A has the Artin approximation property if every
finite system of polynomial equations over A has a solution in A if and only if it
has a solution in the completion  of A. We should mention that M. Artin proved
already in [Artin68] that the ring of convergent power series with coefficients in C
has the Artin approximation property as it was later called.
A ring morphism u : A → A′ of Noetherian rings has regular fibers if for all prime
ideals P ∈ Spec A the ring A′/P A′ is a regular ring, i.e. its localizations are regular
local rings. It has geometrically regular fibers if for all prime ideals P ∈ Spec A and
all finite field extensions K of the fraction field of A/P the ring K ⊗A/P A′/P A′ is
regular.
A flat morphism of Noetherian rings u is regular if its fibers are geometrically
regular. If u is regular of finite type then u is called smooth. A localization of a
smooth algebra is called essentially smooth.
We gratefully acknowledge the support from the project ID-PCE-2011-1023, granted by the
Romanian National Authority for Scientific Research, CNCS - UEFISCDI.
1
A Henselian Noetherian local ring A is excellent if the completion map A → Â
is regular. For example, a Henselian discrete valuation ring V is excellent if the
completion map V → V̂ induces a separable fraction field extension.
Theorem 2 (M. Artin [Artin69]). Let V be an excellent Henselian discrete valuation
ring and V hxi the ring of algebraic power series in x over V , that is the algebraic
closure of the polynomial ring V [x] in the formal power series ring V [[x]]. Then
V hxi has the Artin approximation property.
The proof used the so called the Néron Desingularization, which says that an
unramified extension V ⊂ V ′ of valuation rings inducing separable field extensions
on the fraction and residue fields, is a filtered inductive union of essentially finite
type subextensions V ⊂ A, which are regular local rings, even essentially smooth
V -subalgebras of V ′ .
Néron Desingularization is extended by the following theorem.
Theorem 3 (General Neron Desingularization, Popescu [P85], [P86], [P90], André
[An91], Teissier [Te94], Swan [Sw98], Spivakovski [Sp99]). Let u : A → A′ be a
regular morphism of Noetherian rings and B an A-algebra of finite type. Then any
A-morphism v : B → A′ factors through a smooth A-algebra C, that is v is a
composite A-morphism B → C → A′ .
The smooth A-algebra C given for B by the above theorem is called a General
Neron Desingularization. Note that C is not uniquely associated to B and so we
better speak about a General Neron Desingularization.
The above theorem gives a positive answer to a conjecture of M. Artin [Artin70].
Theorem 4 ([P86], [P00]). An excellent Henselian local ring has the Artin approximation property.
This paper is a survey on the Artin approximation property, the General Neron
Desingularization and their applications. It relies mainly on some lectures given by
us within the special semester on Artin Approximation of the Chaire Jean Morlet at
CIRM, Luminy, Spring 2015 (see http://hlombardi.free.fr/Popescu-Luminy2015.pdf).
1. Artin approximation properties
First we show how one recovers Theorem 4 from Theorem 3. Indeed, let f be a
finite system of polynomial equations over A in Y = (Y1 , . . . , Yn ) and ŷ a solution
of f in Â. Set B = A[Y ]/(f ) and let v : B → Â be the morphism given by Y → ŷ.
By Theorem 3, v factors through a smooth A-algebra C, that is v is a composite
A-morphism B → C → A′ . Thus changing B by C we may reduce the problem to
the case when B is smooth over A. Since A′ is local, changing B by Bb for some
b ∈ B \ v −1 (mA′ ) we may assume that 1 ∈ (g) : I MB for somepolynomials
∂g
g = (g1 , . . . , gr ) from (f ) and a r × r-minor M of the Jacobian matrix
. Thus
∂Y
g(ŷ) = 0 and M(ŷ) is invertible. By the Implicit Function Theorem there exists
y ∈ A such that y ≡ ŷ modulo mÂ.
2
The following consequence of Theorem 3 was noticed and hinted by N. Radu to
M. André. This was the origin of André’s interest to read our theorem and to write
later [An91].
Corollary 5. Let u : A → A′ be a regular morphism of Noetherian rings. Then the
differential module ΩA′ /A is flat.
For the proof note that by Theorem 3 it follows that A′ is a filtered inductive limit
of some smooth A-algebras C and so ΩA′ /A is a filtered inductive limit of A′ ⊗C ΩC/A ,
the last modules being free modules.
Definition 6. A Noetherian local ring (A, m) has the strong Artin approximation
property if for every finite system of polynomial equations f in Y = (Y1 , . . . , Yn ) over
A there exists a map ν : N → N with the following property:
If y ′ ∈ An satisfies f (y ′) ≡ 0 modulo mν(c) , c ∈ N, then there exists
a solution y ∈ An of f with y ≡ y ′ modulo mc .
M. Greenberg [Gr66] proved that excellent Henselian discrete valuation rings have
the strong Artin approximation property and ν is linear in this case.
Theorem 7 (M. Artin [Artin69]). The algebraic power series ring over a field has
the strong Artin approximation property.
Note that in general ν is not linear as it is showed in [Ron05]. The following
theorem was conjectured by M. Artin in [Artin70].
Theorem 8 ([P75]). Let A be an excellent Henselian discrete valuation ring and
Ahxi the ring of algebraic power series in x over A. Then Ahxi has the strong Artin
approximation property.
Theorem 9 (Pfister-Popescu [PfPo75] (see also [KMPPR78], [P79])). The Noetherian complete local rings have the strong Artin approximation property. In particular,
A has the strong Artin approximation property if it has the Artin approximation
property.
Thus Theorem 8 follows from Theorem 2 and Theorem 4 gives that excellent
Henselian local rings have the strong Artin approximation property. An easy direct
proof of this fact is given in [P86, 4.5] using Theorem 3 and the ultrapower methods.
What about the converse implication in Theorem 4? It is clear that A is Henselian
if it has the Artin approximation property. On the other hand, if A is reduced and
it has the Artin approximation property, then  is reduced, too. Indeed, if ẑ ∈  is
nonzero and satisfies ẑ r = 0 then choosing c ∈ N such that ẑ 6∈ mc  we get a z ∈ A
such that z r = 0 and z ≡ ẑ modulo mc Â. It follows that z 6= 0, which contradicts
our hypothesis. It is easy to see that a local ring B which is finite as a module
over A has the Artin approximation property if A has it. It follows that if A has
the Artin approximation property, then it has so called reduced formal fibers. In
particular, A must be a so called universally japanese ring.
Using also the strong Artin approximation property it is possible to prove that
given a system of polynomial equations f ∈ A[Y ]r , Y = (Y1 , . . . , Yn ) and another
3
one g ∈ A[Y, Z]t , Z = (Z1 , . . . , Zs ) then the sentence
LA := there exists y ∈ An such that f (y) = 0 and g(y, z) 6= 0 for all z ∈ As
holds in A if and only if L holds in  provided that A has the Artin approximation
property. In this way it was proved in [BNP81] that if A has the Artin approximation
property, then A is a normal domain if and only if  is a normal domain, too (this was
actually the starting point of the quoted paper). Later, Cipu and myself [CP81] used
this fact to show that the formal fibers of A are the so called geometrically normal
domains if A has the Artin approximation property. Finally, Rotthaus [Rot90]
proved that A is excellent if A has the Artin approximation property.
Next, let (A, m) be an excellent Henselian local ring, Â its completion and MCM(A)
(resp. MCM(Â)) be the set of isomorphism classes of maximal Cohen Macaulay
modules over A (resp. Â). Assume that A is an isolated singularity. Then a
maximal Cohen-Macaulay module is free on the punctured spectrum. Since  is
also an isolated singularity we see that the map ϕ :MCM(A) →MCM(Â) given by
M → Â ⊗A M is surjective by a theorem of Elkik [El73, Theorem 3].
Theorem 10 (Popescu-Roczen, [PR90]). ϕ is bijective.
Proof. Let M, N be
two finite A-modules. P
We may suppose that M = An /(u),
P
N = An /(v), uk = j∈[n] ukj ej , k ∈ [t], vr = j∈[n] vrj ej , r ∈ [p], where ukj , vrj ∈ A
and (ej ) is the canonical basis of An . Let f : An → An be an A-linear map defined
by an invertible n × n-matrix (xij ) with respect to (ej ). Then f induces a bijection
M → N if and only if f maps (u) onto (v), that is there exist ykr , zrk ∈ A, k ∈ [t],
r ∈ [p] such that
P
1) f (uk ) = r∈[p] ykr vr , k ∈ [t] and
P
2) f ( k∈[t] zrk uk ) = vr , r ∈ [p].
Note
Pthat 1), 2) are
Pequivalent to
′
1 ) i∈[n] uki xij = r∈[p] ykr vrj , k ∈ [t], j ∈ [n],
P
P
2′ ) k∈[t] zrk ( i∈[n] uki xij ) = vrj , r ∈ [p], j ∈ [n].
Therefore, if  ⊗A M ∼
=  ⊗A N there exist (x̂ij ), (ŷkr ), (ẑrk ) in  such that
det(x̂ijP
) 6∈ m and
P
1′′ ) i∈[n] uki x̂ij = r∈[p] ŷkr vrj , k ∈ [t], j ∈ [n],
P
P
2′′ ) k∈[t] ẑrk ( i∈[n] uki x̂ij ) = vrj , r ∈ [p], j ∈ [n].
Then by the Artin approximation property there exists a solution of 1′′ ), 2′′ ), let
us say (xij ), (ykr ), (zrk ) in A, such that xij ≡ x̂ij , ykr ≡ ŷkr , zrk ≡ ẑrk modulo mÂ.
It follows that det(xij ) ≡ det(x̂ij ) 6≡ 0 modulo m and so M ∼
= N.
Corollary 11. In the hypothesis of the above theorem if M ∈ MCM(A) is indecomposable, then  ⊗A M is indecomposable, too.
Proof. Assume that Â⊗A M = N̂1 ⊕ N̂2 . Then N̂i ∈ MCM(Â) and by the surjectivity
of ϕ we get N̂i =  ⊗A Ni for some Ni ∈ MCM(A). Then  ⊗A M ∼
= (Â ⊗A N1 ) ⊕
∼
(Â ⊗A N2 ) and the injectivity of ϕ gives M = N1 ⊕ N2 .
4
Remark 12. If A is not Henselian then the above corollary is false. For example
let A = C[X, Y ](X,Y ) /(Y 2 − X 2 − X 3 ). Then M = (X, Y )A is indecomposable in
√
MCM(A) but  ⊗A M is decomposable. Indeed, for û = 1 + X ∈  we have
 ⊗A M = (Y − ûX) ⊕ (Y − ûX)Â.
Remark 13. Let Γ(A), Γ(Â) be the so called AR-quivers of A, Â. Then ϕ induces
also an inclusion Γ(A) ⊂ Γ(Â) (see [PR90]).
Remark 14. It is known that MCM(Â) is finite if and only if  is a simple singularity. What about a complex unimodal singularity R? Certainly in this case
MCM(R) is infinite but maybe there exists a special property which characterizes
the unimodal singularities. For this purpose it would be necessary to describe somehow MCM(R) at least in some special cases. Small attempts are done by Andreas
Steenpass [St14].
For most of the cases when we need the Artin approximation property, it is enough
to apply Artin’s Theorem 1. Sometimes we might need a special kind of Artin approximation, the so called Artin approximation in nested subring condition, namely
the following result which was also considered as possible by M. Artin in [Artin70].
Theorem 15 ([P86], [P00, Theorem 3.6]). Let K be a field, A = Khxi, x =
(x1 , . . . , xm ), f = (f1 , . . . , fr ) ∈ Khx, Y ir , Y = (Y1 , . . . , Yn ) and 0 ≤ s1 ≤ . . . ≤ sn ≤
m, c be some non-negative integers. Suppose that f has a solution ŷ = (ŷ1, . . . , ŷn )
in K[[x]] such that ŷi ∈ K[[x1 , . . . , xsi ]] for all 1 ≤ i ≤ n. Then there exists a
solution y = (y1 , . . . , yn ) of f in A such that yi ∈ Khx1 , . . . , xsi i for all 1 ≤ i ≤ n
and y ≡ ŷ mod (x)c K[[x]].
Corollary 16. The Weierstrass Preparation Theorem holds for the ring of algebraic
power series over a field.
Proof. Let f ∈ Khxi, x = (x1 , . . . , xm ) be an algebraic power series such that
f (0, . . . , 0, xm ) 6= 0. By Weierstrass Preparation Theorem f is associated in diP
i
visibility with a monic polynomial ĝ = xpm + p−1
i=0 ẑi xm ∈ K[[x1 , . . . , xm−1 ]][xm ]
for some p ∈ N, ẑi ∈ (x1 , . . . , xm−1 )K[[x1 , . . . , xm−1 ]]. Thus the system F1 =
P
i
f − Y (xpm + p−1
i=0 Zi xm ), F2 = Y U − 1 has a solution ŷ, û, ẑi in K[[x]] such that
ẑi ∈ K[[x1 , . . . , xm−1 ]]. By Theorem 15 there exists a solution y, u, zi in Khxi such
that zi ∈ Khx1 , . . . , xm−1 i and is congruent modulo (x) with the previous one. Thus
P
i
y is invertible and f = yg, where g = xpm + p−1
i=0 zi xm ∈ Khx1 , . . . , xm−1 i[xm ]. By
the unicity of the (formal) Weierstrass Preparation Theorem it follows that y = ŷ
and g = ĝ.
Now, we see that Theorem 15 is useful to get algebraic versal deformations (see
[Artin74]) . Let D = KhZi, A = KhT i/J, Z = (Z1 , . . . , Zs ), T = (T1 , . . . , Tn ) and
N = D/(f1 , . . . , fd ). A deformation of N over A is a
P = KhT, Zi/(J) ∼
= ((A ⊗K D)(T,Z) )h -module L such that
1) L ⊗A K ∼
= N,
2) L is flat over A,
5
where above B h denotes the Henselization of a local ring B. The condition 1) says
that L has the form P/(F1 , . . . , Fd ) with Fi ∈ KhT, Zi, Fi ∼
= fi modulo (T ) and 2)
says that
2′ ) TorA
1 (L, K) = 0
by the Local Flatness Criterion, since L is (T )-adically ideal separated because P is
local Noetherian. Let
ν
Pe −
→ Pd → P → L → 0
be part of a free resolution of L over P , where the map P d → P is given by
(F1 , . . . , Fd ). Then 2′ ) says that tensorizing with K ⊗A − the above sequence we get
an exact sequence
D e → D d → D → N → 0,
because P is flat over A. Therefore, 2′ ) is equivalent to
P
2′′ ) For all g ∈ D d with di=1 gi fi = 0 there exists G ∈ KhT, Zid with G ≡ g
P
modulo (T ) such that G modulo J ∈ Im ν, that is di=1 Gi Fi ∈ (J).
We would like to construct a versal deformation L (see [KMPPR78, pages 157159]), that is for any A′ = KhUi/J ′ , U = (U1 , . . . , Un′ ), P ′ = ((A′ ⊗K D)(U,Z))h and
L′ = KhUi/(F ′ ) a deformation of N to A′ there exists a morphism α : A → A′ such
that P ′ ⊗P L ∼
= L′ , where the structural map of P ′ over P is given by α. If we
replace above the algebraic power series with formal power series then this problem
is solved by Schlessinger in the infinitesimal case followed by some theorems of Elkik
and M. Artin. Set  = K[[T ]]/(J), P̂ = (( ⊗K D)(T,Z))h . We will assume that we
have already L such that L̂ = P̂ ⊗P L is versal in the frame of complete local rings.
How to get the versal property for L in the frame of algebraic power series?
Let A′ , P ′ , L′ be as above. Since L̂ is versal in the frame of complete local rings,
there exists α̂ : Â → Â′ such that P̂ ′ ⊗P̂ L̂ ∼
= L̂′ = P̂ ′ ⊗P ′ L′ , where the structure of
P̂ ′ as a P̂ -algebra is given by α̂. Assume that α̂ is given by T → t̂ ∈ (U)K[[U]]n .
Then we have
i) J(t̂) ≡ 0 modulo (J ′ ).
On the other hand, we may suppose that α̂ induces an isomorphism P̂ ′ ⊗P̂ L̂ → L̂′
which is given by (T, Z) → (t̂, ẑ) for some ẑ ∈ (U, Z)K[[U, Z]]s with ẑ ≡ Z modulo
(U, Z)2 and the ideals (F (t̂, ẑ)), (F ′ ) of K[[U, Z]] coincide. Thus there exists an
invertible d × d-matrix Ĉ = (Ĉij ) over K[[U, Z]] with
P
ii) Fi′ = dj=1 Ĉij Fj (t̂, ẑ)).
By Theorem 15 we may find t ∈ (U)KhUin and z ∈ (U, Z)KhU, Zis, Cij ∈
KhU, Zi satisfying i), ii) and such that t ≡ t̂, z ≡ ẑ, Cij ≡ Ĉij modulo (U, Z)2.
Note that det(Cij ) ≡ det Ĉ modulo (U, Z)2 and so (Cij ) is invertible. It follows that
α : A → A′ given by T → t is the wanted one, that is P ′ ⊗P L ∼
= L′ , where the
′
structure of P as a P -algebra is given by α.
Next we give an idea of the proof of Theorem 15 in a particular, but essential
case.
Proposition 17. Let K be a field, A = Khxi, x = (x1 , . . . , xm ), f = (f1 , . . . , fr ) ∈
Khx, Y ir , Y = (Y1 , . . . , Yn ) and 0 ≤ s ≤ m, 1 ≤ q ≤ n, c be some non-negative
6
integers. Suppose that f has a solution ŷ = (ŷ1 , . . . , ŷn ) in K[[x]] such that ŷi ∈
K[[x1 , . . . , xs ]] for all 1 ≤ i ≤ q. Then there exists a solution y = (y1 , . . . , yn ) of f
in A such that yi ∈ Khx1 , . . . , xs i for all 1 ≤ i ≤ q and y ≡ ŷ mod (x)c K[[x]].
Proof. Note that B = K[[x1 , . . . , xs ]]hxs+1 , . . . , xm i is excellent Henselian and so it
has the Artin approximation property. Thus the system of polynomials
f ((ŷi )1≤i≤q , Yq+1, . . . , Yn ) has a solution (ỹj )q<j≤n in B with ỹj ≡ ŷj modulo (x)c .
Now it is enough to apply the following lemma for A = Khx1 , . . . , xs i.
Lemma 18. Let (A, m) be an excellent Henselian local ring, Â its completion, A[x]h ,
x = (x1 , . . . , xm ), Â[x]h be the Henselizations of A[x](m,x) respectively Â[x](m,x) , f =
(f1 , . . . , fr ) a system of polynomials in Y = (Y1 , . . . , Yn ) over A[x]h and 1 ≤ q < n,
c be some positive integers. Suppose that f has a solution ŷ = (ŷ1 , . . . , ŷn ) in Â[x]h
such that ŷi ∈ Â for all i ≤ q. Then there exists a solution y = (y1 , . . . , yn ) of f in
A[x]h such that yi ∈ A for all i ≤ q and y ≡ ŷ mod mc Â[x]h .
Proof. Â[x]h is a union of etale neighborhoods of Â[x]. Take an etale neighborhood
B of Â[x](m,x) such that ŷi ∈ B for all q < i ≤ n. Then B ∼
= (Â[x, T ]/(ĝ))(m,x,T ) for
some monic polynomial ĝ in T over Â[x] with ĝ(0) ∈ (m, x) and ∂ĝ/∂T (0) 6∈ (m, x),
let us say
e−1
X
X
e
ĝ = T +
(
ẑjk xk )T j ,
j=0 k∈Nm ,|k|<u
for some u high enough and ẑjk ∈ Â. Note that ẑ00 ∈ m and ẑ10 6∈ mÂ. Changing
if necessary u, we may suppose that
ŷi ≡
e−1
X
X
(
ŷijk xk )T j mod ĝ
j=0 k∈Nm ,|k|<u
for some ŷijk ∈ Â, q < i ≤ n. Actually, we should take ŷi as a fraction but for an
easier expression we will skip the denominator. Substitute Yi , q < i ≤ n by
Yi+
=
e−1
X
X
(
Yijk xk )T j
j=0 k∈Nm ,|k|<u
in f and divide by the monic polynomial
G = Te +
e−1
X
X
(
Zjk xk )T j
j=0 k∈Nm ,|k|<u
in Â[x, T, Y1 , . . . , Yt , (Yij ), (Zj )], where (Yijk ), (Zjk ) are new variables.
We get
+
fp (Y1 , . . . , Yq , Y ) ≡
e−1
X
(
X
Fpjk (Y1 , . . . , Yq , (Yijk ), (Zjk ))xk T j mod G,
j=0 k∈Nm ,|k|<u
7
1 ≤ p ≤ r. Then ŷ is a solution of f in B if and only if (ŷ1 , . . . , ŷq , (ŷijk ), (ẑjk )) is a
solution of (Fpjk ) in Â. As A has the Artin approximation property we may choose
a solution (y1 , . . . , yq , (yijk ), (zjk )) of (Fpjk ) in A which coincides modulo mc  with
the former one. Then
e−1
X
X
yi =
(
(yijk ))xk T j ,
j=0 k∈Nm ,|k|<u
q < i ≤ n together with yi , 1 ≤ i ≤ q form a solution of f in the etale neighborhood
B ′ = (A[x, T ]/(g))(m,x,T ),
e
g=T +
e−1
X
X
(
zjk xk )T j
i=0 k∈Nm ,|k|<u
of A[x](m,x) , which is contained in A[x]h . Clearly, y is the wanted solution.
2. Applications to the Bass-Quillen Conjecture
Let R[T ], T = (T1 , . . . , Tn ) be a polynomial algebra in T over a regular local
ring (R, m). An extension of Serre’s Problem proved by Quillen and Suslin is the
following
Conjecture 19 (Bass-Quillen). Every finitely generated projective module over
R[T ] is free.
Theorem 20 (Lindel [Li81]). The Bass-Quillen Conjecture holds if R is essentially
of finite type over a field.
Swan’s unpublished notes on Lindel’s paper (see [P89, Proposition 2.1]) contain
two interesting remarks.
1) Lindel’s proof works also when R is essentially of finite type over a DVR A
such that its local parameter p 6∈ m2 .
2) The Bass-Quillen Conjecture holds if (R, m) is a regular local ring containing
a field, or p =char R/m 6∈ m2 providing that the following question has a positive
answer.
Question 21 (Swan). Is a regular local ring a filtered inductive limit of regular
local rings essentially of finite type over Z?
Indeed, suppose for example that R contains a field and R is a filtered inductive
limit of regular local rings Ri essentially of finite type over a prime field P . A
finitely generated projective R[T ]-module M is an extension of a finitely generated
projective Ri [T ]-module Mi for some i, that is M ∼
= R[T ] ⊗Ri [T ] Mi . By Theorem
20 we get Mi free and so M is free, too.
Theorem 22 ([P89]). Swan’s Question 21 holds for regular local rings (R, m, k)
which are in one of the following cases:
(1) R contains a field,
8
(2) the characteristic p of k is not in m2 ,
(3) R is excellent Henselian.
Proof. (1) Suppose that R contains a field k. We may assume that k is the prime
field of R and so a perfect field. Then the inclusion u : k → R is regular and by
Theorem 3 it is a filtered inductive limit of smooth k morphisms k → Ri . Thus Ri is
a regular ring of finite type over k and so over Z. Therefore R is a filtered inductive
limit of regular local rings essentially of finite type over Z. Similarly we may treat
(2).
(3) First assume that R is complete. By the Cohen Structure Theorem we may
also assume that R is a factor of a complete local ring of type A = Z(p) [[x1 , . . . , xm ]]
for some prime integer p. By (2) we see that A is a filtered inductive limit of regular
local rings Ai essentially of finite type over Z. Since R, A are regular local rings we
see that R = A/(x) for a part x of a regular system of parameters of A. Then there
exists a system of elements x′ of a certain Ai which is mapped into x by the limit
map Ai → A. It follows that x′ is part of a regular system of parameters of At for all
t ≥ j for some j > i and so Rt = At /(x′ ) are regular local rings. Now, it is enough
to see that R is a filtered inductive limit of Rt , t ≥ j.
Next assume that R is excellent Henselian and let R̂ be its completion. Using
[AD83], or [Sw98] it is enough to show that given a finite type Z-subalgebra E of
R the inclusion α : E → R factors through a regular local ring E ′ essentially of
finite type over Z, that is there exists β : E ′ → R such that α is the composite map
β
E → E′ −
→ R.
As above, R̂ is a filtered inductive limit of regular local rings and so the composite
α
map α̂ : E −
→ R → R̂ factors through a regular local ring F essentially of finite type
over Z. We may choose a finite type Z-subalgebra D ⊂ F such that F ∼
= Dq for
some q ∈ Spec D and the map E → F factors through D, i.e. D is an E-algebra.
As D is excellent, its regular locus Reg D is open and so there exists d ∈ D \ q such
that Dd is a regular ring. Changing D by Dd we may assume that D is regular.
Let E = Z[b1 , . . . , bn ] for some bi ∈ E ⊂ R and let D = Z[Y ]/(h), Y = (Y1 , . . . , Yn )
for some polynomials h. Since D is an E-algebra we may write bi ≡ Pi (Y ) modulo h,
i = 1, . . . , n for some polynomials Pi ∈ E[Y ] ⊂ R[Y ]. Note that there exists ŷ ∈ R̂n
such that bi = Pi (ŷ), h(ŷ) = 0 because α̂ factors through D. As R has the Artin
approximation property, by Theorem 4 there exists y ∈ Rn such that bi = Pi (y),
h(y) = 0. Let ρ : D → R be the map given by Y → y. Clearly, α factors through D
and we may take E ′ = Dρ−1 m . More precisely, we have the following diagram which
is commutative except in the right square,
α //
ROO
❄❄
❄❄
❄❄ ρ
❄
E❄
D
R̂OO
//
//
F
9
Corollary 23 ([P89]). The Bass-Quillen Conjecture holds if R is a regular local
ring in one of the cases (1), (2) of the above theorem.
Remark 24. Theorem 22 is not a complete answer to Question 21, but (3) says
that a positive answer is expected in general. Since there exists no result similar
to Lindel’s saying that the Bass-Quillen Conjecture holds for all regular local rings
essentially of finite type over Z we decided to wait with our further research. So we
have waited already 25 years.
Another problem is to replace in the Bass-Quillen Conjecture the polynomial
algebra R[T ] by other R-algebras. The tool is given by the following theorem.
Theorem 25 (Vorst [V83]). Let A be a ring, A[x], x = (x1 , . . . , xm ) a polynomial
algebra, I ⊂ A[x] a monomial ideal and B = A[x]/I. Then every finitely generated
projective B-module M is extended from a finitely generated projective A-module N,
that is M ∼
= B ⊗A N, if for all n ∈ N every finitely generated projective A[T ]-module,
T = (T1 , . . . , Tn ) is extended from a finitely generated projective A-module.
Corollary 26 ([P15]). Let R be a regular local ring in one of the cases (1), (2) of
Theorem 22, I ⊂ R[x] be a monomial ideal with x = (x1 , . . . , xm ) and B = R[x]/I.
Then any finitely generated projective B-module is free.
For the proof, apply the above theorem using Corollary 23.
The Bass-Quillen Conjecture could also hold when R is not regular as the following
corollary shows.
Corollary 27 ([P15]). Let R be a regular local ring in one of the cases of Theorem
22, I ⊂ R[x] be a monomial ideal with x = (x1 , . . . , xm ) and B = R[x]/I. Then
every finitely generated projective B[T ]-module, T = (T1 , . . . , Tn ) is free.
This result holds because B[T ] is a factor of R[x, T ] by the monomial ideal
IR[x, T ].
Remark 28. If I is not monomial, then the Bass-Quillen Conjecture may fail when
replacing R by B. Indeed, if B = R[x1 , x2 ]/(x21 − x32 ) then there exist finitely
generated projective B[T ]-modules of rank one which are not free (see [La78, (5.10)]).
Now, let (R, m) be a regular local ring and f ∈ m \ m2 .
Question 29 (Quillen [Q76]). Is free a finitely generated projective module over
Rf ?
Theorem 30 (Bhatwadeckar-Rao, [BR83]). Quillen’s Question has a positive answer if R is essentially of finite type over a field.
Theorem 31 ([P02]). Quillen’s Question has a positive answer if R contains a field.
This goes similarly to Corollary 23 using Theorem 30 instead of Theorem 20.
Remark 32. The paper [P02] was not accepted for publication in many journals
since the referees said that ”relies on a theorem [that is Theorem 3] which is still
not recognized by the mathematical community”. Since our paper was quoted as an
unpublished preprint in [Sw98] we published it later in the Romanian Bulletin and
it was noticed and quoted by many people (see for instance [KaVi14]).
10
3. General Neron Desingularization
Using Artin’s methods from [Artin68], Ploski gave the following theorem, which
is the first form of a possible extension of Neron Desingularization in dim > 1.
Theorem 33 ([Pl74]). Let C{x}, x = (x1 , . . . , xm ), f = (f1 , . . . , fs ) be some convergent power series from C{x, Y }, Y = (Y1 , . . . , Yn ) and ŷ ∈ C[[x]]n with ŷ(0) = 0
be a solution of f = 0. Then the map v : B = C{x, Y }/(f ) → C[[x]] given
by Y → ŷ factors through an A-algebra of type B ′ = C{x, Z} for some variables
Z = (Z1 , . . . , Zs ), that is v is a composite map B → B ′ → C[[x]].
Using Theorem 3 one can get an extension of the above theorem.
Theorem 34 ([P15]). Let (A, m) be an excellent Henselian local ring, Â its completion, B a finite type A-algebra and v : B → Â an A-morphism. Then v factors
through an A-algebra of type A[Z]h for some variables Z = (Z1 , . . . , Zs ), where A[Z]h
is the Henselization of A[Z](m,Z) .
Suppose that B = A[Y ]/I , Y = (Y1 , . . . , Yn ). If f = (f1 , . . . , fr ), r ≤ n is a system
of polynomials from I, then
by ∆f the ideal generated by all r × r−minors
denote
∂fi
of the Jacobian matrix
. After Elkik [El73], let HB/A be the radical of the
∂Yj
X
(f ) : I ∆f B, where the sum is taken over all systems of polynomials f
ideal
f
from I with r ≤ n. Then BP , P ∈ Spec B is essentially smooth over A if and only if
P 6⊃ HB/A by the Jacobian criterion for smoothness. Thus HB/A measures the non
smooth locus of B over A.
In the linear case we may easily get cases of Theorem 3 when dim A > 1.
Lemma 35 ([P85, (4.1)]). Let A be a ring and a1 , a2 a weak regular sequence of A,
that is a1 is a non-zero divisor of A and a2 is a non-zero divisor of A/(a1 ). Let A′
be a flat A-algebra and set B = A[Y1 , Y2 ]/(f ), where f = a1 Y1 + a2 Y2 . Then HB/A
is the radical of (a1 , a2 ) and any A-morphism B → A′ factors through a polynomial
A-algebra in one variable.
Proof. Note that all solutions of f = 0 in A are multiples of (−a2 , a1 ). By flatness,
any solution of f in A′ is a linear combinations of some solutions of f in A and so
again a multiple of (−a2 , a1 ). Let h : B → A′ be a map given by Yi → yi ∈ A′ .
Then (y1 , y2 ) = z(−a2 , a1 ) and so h factors through A[Z], that is h is the composite
map B → A[Z] → A′ , the first map being given by (Y1 , Y2) → Z(−a2 , a1 ) and the
second one by Z → z.
Pn
Proposition 36 ([P85, Lemma 4.2]). Let fi = i=1 aij Yj ∈ A[Y1 , . . . , Yn ], i ∈ [r] be
(k)
(k)
a system of linear homogeneous polynomials and y (k) = (y1 , . . . , yn ), k ∈ [p] be a
complete system of solutions of f = (f1 , . . . , fr ) = 0 in A. Let b = (b1 , . . . , br ) ∈ Ar
and c a solution of f = b in A. Let A′ be a flat A-algebra and
B = A[Y1 , . . . , Yn ]/(f − b). Then any A-morphism B → A′ factors through a polynomial A-algebra in p variables.
11
Proof. Let h : B → A′ be a map given by Y → y ′ ∈ A′n . Since A′ is flat over A we see
that y ′ −h(c) is a linear combinations
of y (k) , that is there exists z = (z1 , . . . , zp ) ∈ A′p
P
such that y ′ − h(c) = pk=1 zk h(y (k) ). Therefore, h factors through A[Z1 , . . . , Zp ],
that is h is the compositePA-morphism B → A[Z1 , . . . , Zp ] → A′ , where the first
map is given by Y → c + pk=1 Zk y (k) and the second one by Z → z.
Another form of Theorem 3 is the following theorem which is a positive answer
to a conjecture of M. Artin [Artin82].
Theorem 37 (Cipu-Popescu [CP84]). Let u : A → A′ be a regular morphism of
Noetherian rings, B an A-algebra of finite type, v : B → A′ an A-morphism and
D ⊂ Spec B the open smooth locus of B over A. Then there exist a smooth A-algebra
C and two A-morphisms t : B → C, w : C → A′ such that v = wt and C is smooth
over B at t∗ −1 (D), t∗ : Spec C → Spec B being induced by t.
There exists also a form of Theorem 3 recalling us the strong Artin approximation
property.
Theorem 38 ([P01], [P15]). Let (A, m) be a Noetherian local ring with the completion map A → Â regular, B an A-algebra of finite type and ν the Artin function over
 associated to the system of polynomials f defining B. Then there exists a function λ : N → N, λ ≥ ν such that for every positive integer c and every morphism
v : B → A/mλ(c) there exists a smooth A-algebra C and two A-algebra morphisms
v
t : B → C, w : C → A/mc such that wt is the composite map B −
→ A/mλ(c) → A/mc .
Sometimes, we may find some information about λ (and so about ν). Let A be
a discrete valuation ring, x a local parameter of A, A′ = Â its completion and
B = A[Y ]/I, Y = (Y1 , . . . , Yn ) an A-algebra of finite type. If f = (f1 , . . . , fr ),
r ≤ n is a system of polynomials from I then we consider a r × r-minor M of the
Jacobian matrix (∂fi /∂Yj ). Let c ∈ N. Suppose that there exists an A-morphism
v : B → A′ /(x2c+1 ) and N ∈ ((f ) : I) such that v(NM) 6∈ (x)c /(x2c+1 ), where for
simplicity we write v(NM) instead of v(NM + I).
Theorem 39 ([P15, Theorem 10]). There exists a B-algebra C which is smooth
over A such that every A-morphism v ′ : B → A′ with v ′ ≡ v modulo x2c+1 (that is
v ′ (Y ) ≡ v(Y ) modulo x2c+1 ) factors through C.
Corollary 40 ([P15, Theorem 15]). In the assumptions and notation of Corollary
39 there exists a canonical bijection
for some s ∈ N.
A′s → {v ′ ∈ HomA (B, A′ ) : v ′ ≡ v modulo x2c+1 }
Let k be a field and F a k-algebra of finite type, let us say F = k[U]/J, U =
(U1 , . . . , Un ). An arc Spec k[[x]] → Spec F is given by a k-morphism F → A′ = k[[x]].
Assume that HF/k 6= 0 (this happens for example when F is reduced and k is
perfect). Set A = k[x](x) , B = A ⊗k F . Let f = (f1 , . . . , fr ), r ≤ n be a system
of polynomials from J and M a r × r-minor of the Jacobian matrix (∂fi /∂Uj ).
Let c ∈ N. Assume that there exists an A-morphism g : F → A′ /(x2c+1 ) and
12
N ∈ ((f ) : J) such that g(NM) 6∈ (x)c /(x2c+1 ). Note that A ⊗k − induces a
bijection Homk (F, A′ ) → HomA (B, A′ ) by adjunction.
Corollary 41 ([P15, Corollary 16]). The set {g ′ ∈ Homk (F, A′ ) : g ′ ≡ g modulo x2c+1 }
is in bijection with an affine space A′s over A′ for some s ∈ N.
Next we give a possible extension of Greenberg’s result on the strong Artin approximation property [Gr66]. Let (A, m) be a Cohen-Macaulay local ring (for example a reduced ring) of dimension one, A′ = Â the completion of A, B = A[Y ]/I,
Y = (Y1 , . . . , Yn ) an A-algebra of finite type and c, e ∈ N. Suppose that there exists
f = (f1 , . . . , fr ) in I, a r×r-minor M of the Jacobian matrix (∂f /∂Y ), N ∈ ((f ) : I)
and an A-morphism v : B → A/m2e+c such that (v(MN)) ⊃ me /m2e+c . Then we
may construct a General Neron Desingularization in the idea of Theorem 39, which
could be used to get the following theorem.
Theorem 42 (A. Popescu-D. Popescu, [APDP15]). There exists an A-morphism
v ′ : B → Â such that v ′ ≡ v modulo mc Â, that is v ′ (Y + I) ≡ v(Y + I) modulo mc Â.
Moreover, if A is also excellent Henselian there exists an A-morphism v ′ : B → A
such that v ′ ≡ v modulo mc .
Remark 43. The above theorem could be extended for Noetherian local rings of
dimension one (see [PfPo15]). In this case the statement depends also on a reduced
primary decomposition of (0) in A.
Using [APDP15] we end this section with an algorithmic attempt to explain the
proof of Theorem 3 in the frame of Noetherian local domains of dimension one. Let
u : A → A′ be a flat morphism of Noetherian local domains of dimension 1. Suppose
that A ⊃ Q and the maximal ideal m of A generates the maximal ideal of A′ . Then
u is a regular morphism. Moreover, we suppose that there exist canonical inclusions
k = A/m → A, k ′ = A′ /mA′ → A′ such that u(k) ⊂ k ′ .
If A is essentially of finite type over Q, then the ideal HB/A can be computed
in Singular
by following its definition but it is easier to describe only the ideal
P
(f
)
:
I
∆f B defined above. This is the case considered in our algorithmic part,
f
let us say A ∼
= k[x]/F for some variables x = (x1 , . . . xm ), and the completion of
′
′
A is k JxK/(F ). When v is defined by polynomials y from k ′ [x] then our problem
is easy. Let L be the field obtained by adjoining to k all coefficients of y. Then
R = L[x]/(F ) is a subring of A′ containing Im v which is essentially smooth over A.
Then we may take B ′ as a standard smooth A-algebra such that R is a localization
of B ′ . Consequently we suppose usually that y is not in k ′ [x].
We may suppose that v(HB/A ) 6= 0. Indeed, if v(HB/A ) = 0 then v induces an Amorphism v ′ : B ′ = B/HB/A → A′ and we may replace (B, v) by (B ′ , v ′). Applying
this trick several times we reduce to the case v(HB/A ) 6= 0. However, the fraction
field of Im v is essentially smooth over A by separability, that is HIm v/A A′ 6= 0 and
in the worst case our trick will
change B by Im v after several steps.
′
Choose P ∈ ∆f ((f ) : I) \ I for some system of polynomials f = (f1 , . . . , fr )
from I and d′ ∈ v(P ′ )A′ ∩ A, d′ 6= 0. Moreover, we may choose P ′ to be from
13
∂f
. Then d′ = v(P ′ )z ∈ v(HB/A ) ∩ A
M (f ) : I where M is a r × r-minor of
∂Y
for some z ∈ A′ . Set B1 = B[Z]/(fr+1 ), where fr+1 = −d′ +P ′ Z and let v1 : B1 → A′
be the map of B-algebras given by Z → z. It follows that d′ ∈ (f, fr+1 ) : (I, fr+1 )
and d′ ∈ ∆f , d′ ∈ ∆fr+1 . Then d = d′2 ≡ P modulo (I, fr+1 ) for P = P ′2 Z 2 ∈ HB1 /A .
Replace
B byB1 and the Jacobian matrix J = (∂f /∂Y ) will be now the new J given
J 0
by
. Thus we reduce to the case when d ∈ HB/A ∩ A.
∗ P′
But how to get d with a computer if y is not polynomial defined over k ′ ? Then
the algorithm is complicated because we are not able to tell the computer who is y
and so how to get d′ . We may choose an element a ∈ m and find a minimal c ∈ N
such that ac ∈ (v(M)) + (a2c ) (this is possible because dim A = 1). Set d′ = ac . It
follows that d′ ∈ (v(M)) + (d′2 ) ⊂ (v(M)) + (d′4 ) ⊂ . . . and so d′ ∈ (v(M)), that is
d′ = v(M)z for some z ∈ A′ . Certainly we cannot find precisely z, but later it is
enough to know just a kind of truncation of it modulo d′6 .
Thus we may suppose that there exist f = (f1 , . . . , fr ), r ≤ n a system of polynomials from I, a r × r-minor M of the Jacobian matrix (∂fi /∂Yj ), N ∈ ((f ) : I) such
that 0 6= d ≡ P = MN modulo I. We may assume that M = det((∂fi /∂Yj )i,j∈[r]).
Set Ā = A/(d3 ), Ā′ = A′ /d3A′ , ū = Ā ⊗A u, B̄ = B/d3 B, v̄ = Ā ⊗A v. Clearly, ū is a
regular morphism of Artinian local rings and it is easy to find a General Neron Desingularization in this frame. Thus there exists a B̄-algebra C, which is smooth over
Ā such that v̄ factors through C. Moreover we may suppose that C ∼
= (Ā[U]/(ω))τ
for some polynomials ωτ ∈ k[U] which are not in m(Ā[U]/(ω)) (note that k ⊂ A).
Then D ∼
= (A[U]/(ω))τ is smooth over A and u factors through D. Usually, v does
not factor through D, though v̄ factors through C ∼
= Ā ⊗A D.
′
n
Let y ∈ D be such that the composite map B̄ → C → D̄ is given by Y →
y ′ + d3 D. Thus I(y ′) ≡ 0 modulo d3 D. We have d ≡ P modulo I and so P (y ′) ≡ d
modulo d3 . Thus P (y ′) = ds for a certain s ∈ D with s ≡ 1 modulo d. Let H
be the n × n-matrix obtained by adding down to (∂f /∂Y ) as a border the block
(0|Idn−r ). Let G′ be the adjoint matrix of H and G = NG′ . We have
GH = HG = NMIdn = P Idn
and so
dsIdn = P (y ′)Idn = G(y ′ )H(y ′).
Set h = s(Y − y ′ ) − dG(y ′)T , where T = (T1 , . . . , Tn ) are new variables. Since
s(Y − y ′) ≡ dG(y ′)T modulo h
and
f (Y ) − f (y ′) ≡
X ∂f
(y ′)(Yj − yj′ )
∂Y
j
j
modulo higher order terms in Yj − yj′ , by Taylor’s formula we see that for p =
maxi deg fi we have
sp f (Y ) − sp f (y ′) ≡ sp−1 dP (y ′)T + d2 Q
14
modulo h, where Q ∈ T 2 D[T ]r . This is because (∂f /∂Y )G = (P Idr |0). We have
f (y ′) = d2 b for some b ∈ dD r . Set gi = sp bi + sp Ti + Qi , i ∈ [r]. Then we may take
B ′ to be a localization of (D[Y, T ]/(I, h, g))s.
Remark 44. An algorithmic proof in the frame of all Noetherian local rings of
dimension one is given in [PfPo15].
References
[An91] M. André, Cinq exposes sur la desingularisation, Handwritten manuscript Ecole Polytechnique Federale de Lausanne, (1991).
[Artin68] M. Artin, On the solutions of analytic equations, Invent. Math., 5, (1968), 277-291.
[Artin69] M. Artin, Algebraic approximation of structures over complete local rings, Publ. Math.
IHES, 36, (1969), 23-58.
[Artin70] M. Artin, Constructions techniques for algebraic spaces, Actes Congres. Intern. Math., t
1, (1970), 419-423-291.
[Artin74] M. Artin, Versal deformations and algebraic stacks, Invent. Math., 27, 1974, 165-189.
[Artin82] M. Artin, Algebraic structure of power series rings, Contemp. Math. AMS, Providence,1982, 223-227.
[AD83] M. Artin, J. Denef, Smoothing of a ring homomorphism along a section, Arithmetic and
Geometry, vol. II, Birkhäuser, Boston, (1983), 5-32.
[BNP81] S. Basarab, V. Nica, D. Popescu, Approximation properties and existential completeness
for ring morphisms, Manuscripta Math. 33, (1981), 227-282.
[BR83] S. M. Bhatwadeckar, R. A. Rao, On a question of Quillen, Trans. Amer. Math. Soc., 279,
(1983),801-810.
[CP81] M. Cipu, D. Popescu, Some extensions of Neron’s p-desingularization and approximation,
Rev. Roum. Math. Pures et Appl., 26(1981),1299-1304.
[CP84] M. Cipu, D. Popescu, A desingularization theorem of Neron type, Ann. Univ. Ferrara, 30
(1984), 63-76.
[Sing12] W. Decker, G.-M. Greuel, G. Pfister, H. Schönemann: Singular 3-1-6 — A computer
algebra system for polynomial computations.http://www.singular.uni-kl.de (2012).
[El73] R. Elkik, Solutions d’equations a coefficients dans un anneaux henselien, Ann. Sci. Ecole
Normale Sup., 6 (1973), 553-604.
[Gr66] M. Greenberg, Rational points in henselian discrete valuation rings, Publ. Math. IHES, 31,
(1966), 59-64.
[EGA66] A. Grothendieck, J. Dieudonne, Elements de geometrie algebrique, IV, Part 1, Publ.
Math. IHES, 1966.
[KaVi14] M. Kashiwara, K. Vilonen, Microdifferential systems and the codimension-three conjecture, Ann. of Math., (2) 180, (2014), 573-620.
[KMPPR78] H. Kurke, T. Mostowski, G. Pfister, D. Popescu, M. Roczen, Die Approximationseigenschaft lokaler Ringe, Springer Lect. Notes in Math., 634, Springer-Verlag, Berlin-New
York, (1978).
[La78] T. Y. Lam, Serre’s Conjecture, Springer Lect. Notes in Math., 635, Berlin, 1978.
[Li81] H. Lindel, On the Bass-Quillen conjecture concerning projective modules over polynomial
rings, Invent. Math., 65, (1981), 319-323.
[Ne64] A. Néron, Modeles minimaux des varietes abeliennes sur les corps locaux et globaux, Publ.
Math. IHES, 21, 1964.
[PfPo75] G. Pfister, D. Popescu, Die strenge Approximationseigenschaft lokaler Ringe, Inventiones
Math. 30, (1975), 145-174.
[PfPo15] G. Pfister, D. Popescu, Constructive General Neron Desingularization for one dimensional local rings, 2015, in preparation.
15
[Pl74] A. Ploski, Note on a theorem of M. Artin, Bull. Acad. Polon. des Sci., t. XXII, 11 (1974),
1107-1110.
[APDP15] A. Popescu, D. Popescu, A method to compute the General Neron Desingularization in
the frame of one dimensional local domains, arXiv:AC/1508.05511.
[P75] D. Popescu, A strong approximation theorem over discrete valuation rings, Rev. Roum.
Math. Pures et Appl., 20, (1975), 659-692.
[P79] D. Popescu, Algebraically pure morphisms, Rev.Roum.Math.Pures et Appl., 24, (1979), 947977.
[P85] D. Popescu, General Neron Desingularization, Nagoya Math. J., 100 (1985), 97-126.
[P86] D. Popescu, General Neron Desingularization and approximation, Nagoya Math. J., 104,
(1986), 85-115.
[P89] D. Popescu, Polynomial rings and their projective modules, Nagoya Math. J., 113, (1989),
121-128.
[P90] D. Popescu, Letter to the Editor. General Neron Desingularization and approximation,
Nagoya Math. J., 118 (1990), 45-53.
[P00] D. Popescu, Artin Approximation, in ”Handbook of Algebra”, vol. 2, Ed. M. Hazewinkel,
Elsevier, 2000, 321-355.
[P01] D. Popescu, Variations on Néron desingularization, in: Sitzungsberichte der Berliner Mathematischen Gesselschaft, Berlin, 2001, 143-151.
[P02] D. Popescu, On a question of Quillen, Bull. Math. Soc. Sci. Math. Roumanie, 45(93),
(2002), 209-212.
[P15] D. Popescu, Around General Neron Desingularization, arXiv:AC/1504.06938.
[PR90] D. Popescu, M. Roczen, Indecomposable Cohen-Macaulay modules and irreducible maps,
Compositio Math. 76, (1990), 277-294.
[Q76] D. Quillen, Projective modules over polynomial rings, Invent. Math., 36, (1976), 167-171.
[Ron05] G. Rond, Sur la linearité de la fonction de Artin, Ann. Sci. Ecole Norm. Sup., (4), 38,
(2005), 979-988.
[Rot90] C. Rotthaus, Rings with the property of approximation, Math. Ann., 287, (1990), 455-466.
[Sp99] M. Spivakovski, A new proof of D. Popescu’s theorem on smoothing of ring homomorphisms,
J. Amer. Math. Soc., 294, (1999), 381-444.
[St14] A. Steenpass, Algorithms in SINGULAR: Parallelization, Syzygies, and Singularities, PhD
Thesis, 2014, Kaiserslautern.
[Sw98] R. Swan, Neron-Popescu desingularization, in ”Algebra and Geometry”, Ed. M. Kang,
International Press, Cambridge, (1998), 135-192.
[Te94] B. Teissier, Resultats recents sur l’approximation des morphismes en algebre commutative
[d’apres Artin, Popescu et Spivakovski], Sem. Bourbaki, 784, (1994), 259-282.
[V83] T. Vorst, The Serre Problem for discrete Hodge algebras, Math. Z., 184, (1983), 425-433.
Dorin Popescu, Simion Stoilow Institute of Mathematics of the Romanian Academy, Research unit 5, University of Bucharest, P.O.Box 1-764, Bucharest 014700,
Romania
E-mail address: [email protected]
16
| 0math.AC
|
RANK THREE GEOMETRY AND POSITIVE CURVATURE
arXiv:1607.04134v1 [math.DG] 14 Jul 2016
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
Abstract. An axiomatic characterization of buildings of type C3 due to Tits is used to prove
that any cohomogeneity two polar action of type C3 on a positively curved simply connected
manifold is equivariantly diffeomorphic to a polar action on a rank one symmetric space. This
includes two actions on the Cayley plane whose associated C3 type geometry is not covered by
a building.
The rank (or size) of a Coxeter matrix M coincides with the number of generators of its
associated Coxeter system. The basic objects in Tits’ local approach to buildings [Ti2] are the
so-called chamber systems C of type M (see also [Ro]). Indeed, if any so-called (spherical)
residue (subchamber system) of C of rank 3 is covered by a building, so is C .
Recall that a polar G action on a Riemannian manifold M is an isometric action with a socalled section Σ, i.e., an immersed submanifold of M that meets all G orbits orthogonally. Since
the action by the identity component of G is polar as well, we assume throughout without stating
it that G is connected.
It is a key observation of [FGT] that the study of polar G actions on 1-connected positively
curved manifolds M in essence is the study of a certain class of (connected) chamber systems
C (M; G). Moreover, when the universal (Tits) cover of C (M; G) is a building it has the structure
of a compact spherical building in the sense of Burns and Spatzier [BSp]. This was utilized in
[FGT] to show:
Theorem A. Any polar G action of cohomogeneity at least two on a simply connected closed
positively curved manifold M is equivariantly diffeomorphic to a polar G action on a rank one
symmetric space if the associated chamber system C (M; G) is not of type C3 .
We note here, that when the action has no fixed points, the rank of C (M; G) is dim(M/ G)+1,
i.e., one more than the cohomogeneity of the action. In the above theorem the Cayley plane
emerges only in cohomogeneity two and when G has fixed points. Moreover, there are indeed
chamber systems with type M = C3 whose universal cover is NOT a building (see, e.g., [Ne],
[FGT], [Ly], [KL] and below). In our case, a polar G action on M is of type C3 if and only if its
orbit space M/ G is a geodesic 2-simplex with angles π/2, π/3 and π/4.
Our aim here is to take care of this exceptional case and prove
Theorem B. Any polar G action on a simply connected positively curved manifold M of
type C3 is equivariantly diffeomorphic to a polar action on a rank one symmetric space. This
The first author is supported in part by an NSFC grant and he is grateful to the University of Notre Dame for
its hospitality. The second author is supported in part by an NSF grant, a research chair at the Hausdorff Center at
the University of Bonn and by a Humboldt research award. The third author is grateful to the University of Notre
Dame and the Capital Normal University in Beijing for their hospitality.
1
2
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
includes two actions on the Cayley plane where the universal covers of the associated chamber
systems are not buildings.
Combining these results of course establishes, the
Corollary. Any polar G action of cohomogeneity at least two on a simply connected closed
positively curved manifold M is equivariantly diffeomorphic to a polar G action on a rank one
symmetric space.
This is in stark contrast to the case of cohomogeneity one, where in dimensions seven and
thirteen there are infinitely many non-homogeneous manifolds (even up to homotopy). The
classification work in [GWZ] also lead to the discovery and construction of a new example of a
positively curved manifold (see [De] and [GVZ]).
By necessity, as indicated above, the proof of Theorem B is entirely different from the proof
of Theorem A. In general, the geometric realization of our chamber systems C (M; G) utilized
in the proof of Theorem A are not simplicial. However, in [FGT] it was proved that in fact
Theorem C. The geometric realization |C (M, G)| of a chamber system C (M, G) of type A3
or C3 associated with a simply connected polar G-manifold M is simplicial.
When the geometric realization of a chamber system of type M is simplicial it is called a Tits
geometry of type M. This allows us to use an axiomatic characterization of C3 geometries that
are buildings (see [Ti2], Proposition 9). So rather than considering the universal cover C˜(M; G)
directly, we construct in all but two cases a suitable cover of C (M; G) (possibly C (M; G) itself),
and prove that it satisfies the C3 building axiom of Tits. The two cases where this methods fails,
are then recognized as being equivalent to two C3 type polar actions on the Cayley plane OP2
(cf. [PTh, GK]).
We note, that since all our chamber systems C (M, G) are homogeneous and those of type C3
(and A3 ) are Tits geometries an independent alternate proof of Theorem B follows from [KL].
1. Preliminaries
The purpose of this section is threefold. While explaining the overall approaches to the
strategies needed in the proof of Theorem B, we recall the basic concepts and establish notation.
Throughout G denotes a compact connected Lie group acting on a closed 1- connected positively curved manifold M in a polar fashion and of type C3 .
Fix a chamber C in a section Σ for the action. Then C is isometric to the orbit spaces M/ G
and Σ/W, where W is the reflection group of Σ and W acts simply transitively on the chambers of
Σ. Since the action is of type C3 , C is a convex positively curved 2-simplex with geodesic sides
= faces, ℓr , ℓt and ℓq opposite its vertices r, t and q with angles π/2, π/3 and π/4 respectively.
By the Reconstruction Theorem of [GZ] recall that any polar G manifold M is completely
determined by its so-called polar data. In our case, this data consist of G and all its isotropy
groups, together with their inclusions along a chamber C (cf. also Lemma 1.5 in [Go]). We
denote the principal isotropy group by H, and the isotropy groups at vertices and opposite faces
by Gr , Gt , Gq and Gℓr , Gℓt , Gℓq respectively. What remains after removing G from this data will
be referred to as the local data for the action.
RANK THREE GEOMETRY AND POSITIVE CURVATURE
3
With two exceptions, it turns out that only partial data are needed to show that the action
indeed is equivalent to a polar action on a rank one symmetric space. Since the data in the
two exceptional cases coincide with those of the exceptional C3 actions on the Cayley plane,
this will then complete the proof of Theorem A. In addition, it is worth noting, that since the
groups G derived from those data (in 7.2 and 8.1) are maximal connected subgroups of F4 , the
identity component of the isometry group of the Cayley plane OP2 , their actions are uniquely
determined and turn out to be polar.
The proof of Theorem A in all but the two exceptional cases is based on showing that the
universal cover, C˜ of the chamber system C = C (M, G) associated to the polar action is a
spherical Tits building [FGT]. Here, the homogeneous chamber system C (M, G) is the union
∪g∈G g C of all chambers with three adjacency relations one for each face: Specifically g1 C and
g2 C are i adjacent if their respective i faces are the same in M. This chamber system with the
thin topology, i.e., induced from the its path metric is a simplicial complex by Theorem C, and
hence C (M, G) is a so-called C3 geometry.
As indicated, the Fundamental Theorem of Tits used in [FGT] to show that C˜ is a building
yields nothing for rank three chamber systems as well as rank three geometries. Instead we will
show that C , or a cover we construct of C is a C3 building (and hence simply connected) by
verifying an axiomatic incidence characterization (see section 3) of such buildings due also to
Tits.
The construction of chamber system covers we utilize is equivalent in our context to the
principal bundle construction of [GZ] (Theorem 4.5) for Coxter polar actions and manifolds.
Specifically for our case:
• Given the data, H, Gℓi , G j , i, j ∈ {t, r, q} and G for (M, G), the data for (P, L × G) consists
of graphs Ĥ, Ĝℓi , Ĝ j in Ĝ := L × G of compatible homomorphisms from H, Gℓi , G j , i, j ∈
t, r, q to L. In particular, the local data for (P, L × G) are isomorphic to the local data for
(M, G).
• Clearly L acts freely as a group of automorphisms, and C (P, L × G)/L = C (M, G), i.e.,
Cb(M; G) := C (P, L × G) is a chamber system covering of C (M; G).
In our case L will be S1 (or in one case S3 ).
2. Basic tools and obstructions
The aim of this section is to establish a number of properties and restrictions of the data to
be used throughout. Unless otherwise stated G will be a compact connected Lie group and M a
closed simply connected positively curved manifold.
Without any curvature assumptions we have the possibly well known
Lemma 2.1 (Orbit equivalence). Let M be a simply connected polar G manifold. Then the
slice representation of any isotropy group is orbit equivalent to that of its identity component.
Proof. Recall that the slice representation of an isotropy group K = G p ⊂ G restricted to the
orthogonal complement T p⊥ of the fixed point set of K inside the normal space to the orbit G p
is a polar representation. Clearly the finite group K/K0 acts isometrically on the orbit space
S(T p⊥ )/K0 , which is isometric to a chamber C of the polar K0 action on the sphere S(T p⊥ ). Since
C is convex with non-empty boundary its soul point (the unique point at maximal distance to
4
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
the boundary) is fixed by K/K0 . This soul point, however, corresponds to a principal K0 orbit,
and hence to an exceptional K orbit unless K/K0 acts trivially on C. However, by Theorem 1.5
[AT] there are no exceptional orbits of a polar action on a simply connected manifold.
Because of this, when subsequently talking casually about a slice representation we refer to
the slice representation of its identity component unless otherwise stated.
Using positive curvature the following basic fact was derived in [FGT], Theorem 3.2:
Lemma 2.2 (Primitivity). The group G is generated by the (identity components) of the face
isotropy groups of any fixed chamber.
Naturally, the slice representations of Gt , Gq and Gr play a fundamental role. We denote the
respective kernels of these representations by Kt , Kq and Kr and their quotients by Ḡt , Ḡq and
Ḡr . Since in particular the slice representation of Gt is of type A2 it follows that the multiplicity
triple of the polar G manifold M, i.e, the dimensions of the unit spheres in the normal slices
along the edges ℓq , ℓr , ℓt is (d, d, k) ∈ Z3+ , where d = 1, 2, 4 or 8.
For the kernels Kt and Kq , which are usually large groups, we have:
Lemma 2.3 (Slice Kernel). Let M be a simply connected polar G-manifold of type C3 . If
G acts effectively, then the kernel Kt , respectively Kq acts effectively on the slices T q⊥ and T r⊥ ,
respectively T t⊥ and T r⊥ .
Proof. Note that Kt fixes all sections through t since Kt acts trivially on the slice T t⊥ . We must
prove that Kt ∩ Kq = {1}, Kt ∩ Kr = {1} and Kq ∩ Kr = {1}. We consider only Kt ∩ Kq , since the
arguments for the remaining cases are similar.
Note that since G is assumed to act effectively on M, and Kt ∩ Kq is contained in the principal
isotropy group, it suffices to prove that Kt ∩ Kq is normal in G. By the primitivity (see 2.2),
−1
G = hp−1
q (Ḡq,0 ), pt (Ḡt,0 )i, where pq : Gq → Ḡq is the quotient homomorphism and Ḡq,0 is the
identity component of Ḡq and similarly for pt . Thus, it suffices to show that Kt ∩ Kq is normal
−1
in each of p−1
t (Ḡt,0 ) and pq (Ḡq,0 ). In each case, assuming the effective vertex isotropy group is
connected does not alter the proof only simplifies notation. Accordingly, we proceed to assume
that Ḡt is connected, i.e., Ḡt = Ḡt,0 and will show that Kt ∩ Kq is a normal subgroup of Gt .
Note that Kt ∩ Kq is a normal subgroup of Kt acting trivially on both the slices T t⊥ and T q⊥ .
By assumption the quotient map Gt,0 ⊂ Gt → Ḡt is surjective when restricted to the identity
˜
component Gt,0 of Gt . A finite central cover G̃t,0 of Gt,0 is isomorphic to the product K̃t,0 × Ḡ
t
˜ is locally isomorphic
where K̃t,0 is locally isomorphic to the identity component Kt,0 of Kt and Ḡ
t
˜ ) covering Ḡ , where π :
to Ḡt . In particular, Gt contains a connected and closed subgroup π( Ḡ
t
t
˜ ) commutes with
G̃t,0 → Gt,0 is the cover map. Moreover, every element of the subgroup π( Ḡ
t
˜ ), the conjugation by h gives rise to an
the elements in Kt,0 . On the other hand, for every h ∈ π( Ḡ
t
element in the automorphism group Aut(Kt ) since Kt is normal, hence defines a homomorphism
˜ ) → Aut(K ). Since φ(π( Ḡ
˜ )) has a trivial image in Aut(K ) under the forgetful
φ : π( Ḡ
t
t
t
t,0
˜
homomorphism Aut(Kt ) → Aut(Kt,0 ), the group φ(π( Ḡt )) is finite, and hence trivial because
˜ )) is connected. This implies that the elements of π( Ḡ
˜ ) commute with the elements of
φ(π( Ḡ
t
t
RANK THREE GEOMETRY AND POSITIVE CURVATURE
5
˜ )i and K ∩ K is normal in K , it then follows that K ∩ K is a normal
Kt . Since Gt = hKt , π( Ḡ
t
t
q
t
t
q
subgroup of Gt .
As mentioned above, the same arguments show that Kt ∩ Kq is normal in p−1
t (Ḡt,0 ) in case Ḡt
is not connected. The same arguments also show that Kt ∩ Kq is normal in p−1
q (Ḡq,0 ).
Remark 2.4. It turns out that in all cases Ḡt is connected. In fact, this is automatic whenever
d , 2, since Ḡt acts transitively on a projective plane. Up to local isomorphism its identity
component is one of the groups SO(3), SU(3), Sp(3), or F4 corresponding to d = 1, 2, 4 and 8
respectively, and the slice representation is its standard polar representation of type A2 (see also
Table 4.3). In view of the Transversality Lemma 2.5 below, Gt is connected whenever k ≥ 2. In
the (2, 2, 1) case, the connectedness of Gr (again by Lemma 2.5) implies that also in this case
Ḡt is connected (see Proposition 5.5).
The following simple topological consequence of transversality combined with the fact that
the canonical deformation retraction of the orbit space triangle minus any side to its opposite
vertex lifts to M (or alternatively of the work [Wie]) will also be used frequently:
Lemma 2.5 (Transversality). Given a multiplicity triple (d, d, m). Then the inclusion maps
G / Gr ⊂ M, G / Gq ⊂ M and G / Gℓt ⊂ M are d-connected, G / Gℓr ⊂ M, and G / Gℓr ⊂ M are
min{d, m} connected, and G / Gt ⊂ M is m-connected.
Recall here that a continuos map is said to be k - connected if the induced map between the
ith homotopy groups is an isomorphism for i < k and a surjection for i = k.
Another Connectivity Theorem [Wi3] (Theorem 2.1) using positive curvature á la Synge is
very powerful:
Lemma 2.6 (Wilking). Let M be a positively curved n-manifold and N a totally geodesic
closed codimension k submanifold. Then the inclusion map N → M is n − 2k + 1 connected.
If in addition N is fixed by an isometric action of a compact Lie group K with principal orbit
of dimension m(K), then the inclusion map is n − 2k + 1 + m(K) connected.
We conclude this section with two severe restrictions on G stemming from positive curvature.
The first follow from the well known Synge type fact, that an isometric Tk action has orbits
with dim ≤ 1 in odd dimensions and 0 in even dimensions, when M has positive curvature
(cf. [Su]). In particular, since Gq has maximal rank among the isotropy groups, and the Euler
characteristic χ(G / Gq ) > 0 if and only if rk(G) = rk(Gq ) ([HS] page 248) we conclude
Lemma 2.7 (Rank Lemma). The dimension of M is even if and only if rk(G) = rk(Gq ), and
otherwise rank rk(G) = rk(Gq ) + 1.
When adapting Wilking’s Isotropy Representation Lemma 3.1 from [Wi2] for positively
curved G manifolds to polar manifolds of type C3 we obtain:
Lemma 2.8 (Sphere Transitive Subrepresentations). Let Li ⊳ Gℓi , i ∈ {q, r, t} be a simple
normal subgroup and U an irreducible isotropy subrepresentation of G /Li . Then (U, Li ) is
isomorphic to a standard defining representation. In particular, Li acts transitively on the sphere
S(U).
6
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
Proof. Let U be an irreducible isotropy subrepresentation of G /Li not isomorphic to a summand
of the slice representation of Li on T i⊥ . By [Wi2], U is isomorphic to a summand of the isotropy
representation of L∗i /Li , where L∗i is a vertex isotropy group. On the other hand, the almost
effective factor of L∗i is well understood (cf. the tables 4.3 and 4.4), which are all the standard
defining representation. The desired result follows.
3. The C3 building axiom
Recall that Tits has provided an axiomatic characterization of buildings of irreducible type M
when the geometric realization |C | (C with the thin topology) of the associated chamber system
C , is a simplicial complex. This characterization is given in terms of the incidence geometry
associated with C .
The purpose of this section is to describe this characterization when M = C3 and translate it
to our context.
Here, by definition
• Vertices x, y ∈ |C | are incident, denoted x ∗ y, if and only if x and y are contained in a
closed chamber of |C |.
Clearly, the incidence relation (not an equivalence relation) is preserved by the action of G in
our case.
To describe the needed characterization we will use the following standard terminology:
• The shadow of a vertex x on the set of vertices of type i ∈ I, denoted Shi (x), is the union
of all vertices of type i incident to x.
Following Tits [Ti2], when M = C3 , we call the vertices of type q, r and t, points, lines, and
planes respectively. We denote by Q, R and T the set of points, lines, and planes in C (M; G).
Notice that G acts transitively on Q, R and T . With this terminology the axiomatic characterization [Ti2] (cf. Proposition 9 and the proof of the C3 case on p. 544) alluded to above
states:
Theorem 3.1 (C3 Axiom). A connected Tits geometry of type C3 is a building if and only if
the following axiom holds:
• (LL) If two lines are both incident to two different points, they coincide.
Equivalently:
• If ShQ (r) ∩ ShQ (r′ ) has cardinality at least two, then r = r′ .
or:
• For any q, q′ ∈ Q, with q , q′ , ShR (q) ∩ ShR (q′ ) has cardinality at most one.
In our case, if r ∈ R and q ∈ Q are incident, (LL) is clearly equivalent to
• For any r′ ∈ Gq (r), r′ , r, we have Gr (q) ∩ Gr′ (q) = q
or,
• For any q′ ∈ Gr (q), q′ , q, we have Gq (r) ∩ Gq′ (r) = r
We proceed to interpret (LL) in terms of the isotropy groups data. This will be used either
directly for C (M; G) or for a suitably constructed cover C˜(M; G) as described at the end of
section 1. For notational simplicity we will describe it here only for C (M; G) (for the general
case see remark 3.5 below).
RANK THREE GEOMETRY AND POSITIVE CURVATURE
7
Proposition 3.2. If C (M; G) is a building of type C3 , then the following holds:
⋆ for any pair of different points q, q′ ∈ Q both incident to an r ∈ R, we have
Gq ∩ Gq′ ⊂ Grq ∩ Grq′
where Grq denotes the isotropy group of the unique edge between r and q (cf. Theorem C).
Proof. Note that every line in the orbit Gq ∩ Gq′ (r) is incident to both q and q′ . Axiom (LL)
implies that the orbit contains only one line, r and hence Gq ∩ Gq′ ⊂ Gr . Since C (M; G) is a
building, we have Gr ∩ Gq = Grq and Gr ∩ Gq′ = Grq′ . The desired result follows.
We will see that the condition ⋆ together with an assumption on a suitable reduction of the
G action implies that C (M; G) is a building of type C3 .
To describe the reduction, let r ∈ R be a line, and let S⊥r,Q be the normal sphere in the summand
in the slice T r⊥ . Then the shadow of r in Q is exp( π4 S⊥r,Q ). Moreover, the isotropy group Gr acts
transitively on S⊥r,Q .
Let Kr,Q denote the identity component of the kernel of the transitive Gr action on S⊥r,Q .
It is clear that the fixed point connected component M Kr,Q (containing r) is a cohomogeneity
one N0 (Kr,Q ) submanifold of M, where N0 (Kr,Q ) is the identity component of the normalizer
N(Kr,Q ) of Kr,Q in G. The corresponding chamber system denoted C (M Kr,Q ) is a subcomplex of
C (M) := C (M; G) that inherits an incidence structure, which gives rise to a Tits geometry of
rank 2.
Lemma 3.3 (Reduction). The connected chamber system C (M, G) of type C3 is a building if
for any r ∈ R, the reduction C (M Kr,Q ) is a C2 -building and ⋆ holds.
Proof. If not, by Axiom (LL) there are two points q , q′ ∈ Q which are both incident to two
different lines r, r′ ∈ R. By ⋆ we know that Gq ∩ Gq′ ⊂ Grq ∩ Grq′ and Gq ∩ Gq′ ⊂ Gr′ q ∩ Gr′ q′ .
Therefore, the configuration {rq, rq′, r′ q, r′q′ } is contained in the fixed point set M Gq ∩ Gq′ . Since
by definition clearly Kr,Q is a subgroup of Gq ∩ Gq′ , we have that M Gq ∩ Gq′ ⊂ M Kr,Q . This implies
that there is a length 4 circuit in the C2 building C (M Kr,Q ). A contradiction.
The following technical criterion will be more useful to us:
Lemma 3.4 (C3 Building Criterion). The connected chamber system C (M, G) is a building if
for any r ∈ R, the reduction C (M Kr,Q ) is a C2 -building and the following Property (P) holds:
(P) For any q ∈ Sh Q (r), and any Lie group L with Kr,Q ⊂ L ⊂ Gq but L 1 Grq , the normalizer
N(Kr,Q ) ∩ L is not contained in Grq either.
Proof. By the previous lemma it suffices to verify ⋆. Suppose ⋆ is not true. Then there is
an r ∈ R and a pair of points q , q′ both incident to r such that Gq ∩ Gq′ is not a subgroup
of Grq . Let L = Gq ∩ Gq′ . By Assumption (P), there is an α ∈ N(Kr,Q ) ∩ L so that α < Grq .
However, Gr ∩ Gq ∩N(Kr,Q ) = Grq ∩N(Kr,Q ) since M Kr,Q is an C2 building. In particular, α <
Gr , and so there is a length 4 circuit {rq, qα(r), α(r)q′ , q′ r} in the C2 building C (M Kr,Q ). A
contradiction.
Remark 3.5. For an S1 cover C˜ := C (P, S1 × G) of C (M, G) constructed as above note
that the property ⋆ is inherited from (M, G). Likewise, the group K̂ being the graph of the
8
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
homomorphism Gℓt ⊂ Gr to S1 restricted to K := Kr,Q satisfies Property (P) when K does.
For this note that by construction the local data for the reduction PK̂ are isomorphic to the
local data for M K . It then follows as in the proofs above, that if a component of the reduction
C (PK̂ ) ⊂ C (P, S1 × G) is a C2 -building, then the corresponding component of C˜ will be a C3
building covering C (M, G), and our main result, Theorem 4.10, from the [FGT] applies.
Remark 3.6. If K′ = K′r,Q ⊂ Kr,Q = K is a subgroup, then the assumption of C (M K ) being a C2
′
building in the above criterion may be replaced by, the fixed point component C (M K ) ⊃ C (M K )
being a C2 building, or a rank 3 building. For the latter, we notice that, by Charney-Lytchak
[CL] Theorem 2, a rank 3 spherical building is a CAT(1) space, hence any two points of distance
less than π are joined by a unique geodesic. This clearly excludes a length 4 circuit in the above
proof, since its perimeter is π.
Remark 3.7. Note that clearly Kt ⊂ Kℓq ∩ Kℓr and similarly for the other kernels of vertex and
edge isotropy groups. In particular, for the identity component K′ of Kt we have K′ ⊂ K, where K
(= Kr,Q ) is the identity component of the kernel of Gr acting on Sd . Consequently, the reduction
′
M K is a cohomogeneity two manifold of type either A3 , or C3 containing the cohomogeneity
one manifold M K (cf. 3.6 above).
4. Classification outline and organization
The subsequent sections are devoted to a proof of the following main result of the paper:
Theorem 4.1. Let M be a compact, simply connected positively curved polar G-manifold
with associated chamber system C (M; G) of type C3 . Then the universal cover C˜ of C (M; G)
is a building if and only if (M, G) is not equivariantly diffeomorphic to one of the exceptional
polar actions on OP2 by G = SU(3) · SU(3) or G = SO(3) · G2 .
This combined with the main result of [FGT] proves Theorem B in the introduction.
The purpose of this section is to describe how the proof is organized according to four types
of scenarios driven by the possible compatible types of slice representations for Gt and Gq at
the vertices t and q of a chamber C.
The common feature in each scenario and all cases is the determination of all local data.
The basic input for this is indeed knowledge of the slice representations at the vertices t and
q of a chamber C, and Lemma 2.3. The local data identifies the desired K ⊂ Gr reduction
M K with its cohomogeneity one action by N(K) referred to in the Building Criteria Lemma 3.4,
with Property (P) being essentially automatic. The main difficulty is to establish that C (M K ) ⊂
C (M; G) or the corresponding reduction in a cover (which by construction has the same local
data) is a C2 building. The first step for this frequently uses the following consequence of the
classification work on positively curved cohomogeneity one manifolds in [GWZ] and [Ve].
Lemma 4.2. Any simply connected positively curved cohomogeneity one manifold with multiplicity pair different from (1, 1), (1, 3) and (1, 7) is equivariantly diffeomorphic to a rank one
symmetric space.
As already pointed out and used, there are only four possible (effective) slice representations
at t, in particular forcing the codimensions of the orbit strata corresponding to ℓq , ℓr , and ℓt to
RANK THREE GEOMETRY AND POSITIVE CURVATURE
9
be d + 1, d + 1 and k + 1, where d = 1, 2, 4 or 8. In Table 4.3, L± , respectively H are the singular,
respectively principal isotropy groups for the effective slice representation, χ by Ḡt restricted to
the unit sphere, and l± + 1 are the codimensions of the singular orbits.
n
Ḡt
4
SO(3)
7
PSU(3)
L−
L+
H
(l− , l+ ) W
S(O(2) O(1))
S(O(1) O(2))
Z2 ⊕ Z2
(1, 1) A2
T2 /Z3
(2, 2) A2
χ
Ad S(U(2) U(1))/∆(Z3 ) S(U(1) U(2))/∆(Z3 )
13 Sp(3)/∆(Z2 ) ψ14 Sp(2) Sp(1)/∆(Z2 ) Sp(1) Sp(2)/∆(Z2 ) Sp(1)3 /∆(Z2 ) (4, 4) A2
25
F4
Spin(9)
ψ26
Spin(9)
Spin(8)
(8, 8) A2
Table 4.3. Effective t-slice representations on S⊥t = Sn
Similarly (see Table 4.4), the identity component (Ḡq )0 of possible effective C2 type slice representations at q which are compatible with the multiplicity restrictions in Table 4.3 are known
as well (see e.g. Table E of [GWZ] in which we have corrected an error for the exceptional
SO(2) Spin(7) representation (see also [GKK] (Main Theorem)).
n
(Ḡq )0
8k + 15, k ≥ 0
Sp(2) Sp(k+2)
4k + 7 , k ≥ 1 even
SU(2) SU(k+2)
∆(Z2 )
∆(Z2 )
L−
χ
ˆ k+2
ν2 ⊗ν
Sp(2) Sp(k)
ˆ k+2
µ2 ⊗µ
△ SU(2) SU(k)
∆(Z2 )
L+
Sp(1)2
Sp(k+1)
∆(Z2 )
∆(Z2 )
S1 · SU(k + 1)
H
Sp(1)2
Sp(k)
∆(Z2 )
S1 · SU(k)
∆(Z2 )
(l− , l+ )
W
(4, 4k + 3) C2
(2, 2k + 1) C2
ˆ k+2 △ SU(2) SU(k) S1 · SU(k + 1) S1 · SU(k) (2, 2k + 1) C2
4k + 7 , k ≥ 1 odd SU(2) SU(k + 2) µ2 ⊗µ
4k + 7, k ≥ 1
7
2k + 3, k ≥ 1 even
U(2) SU(k+2)
∆(Zk )
U(2) SU(2)
∆(Z2 )
SO(2) SO(k+2)
∆(Z2 )
ˆ k+2
µ2 ⊗µ
△ U(2) SU(k)
∆(Zk )
T2 · SU(k+1)
T2 · SU(k)
ˆ 2
µ2 ⊗µ
△ SO(3)
∆(Zk )
2
T
∆(Zk )
1
S
(2, 1)
C2
ˆ k+2
ρ2 ⊗ρ
△ SO(2) SO(k)
∆(Z2 )
SO(k + 1)
SO(k)
(1, k)
C2
(1, k)
C2
ˆ k+2 △ SO(2) SO(k) Z2 · SO(k + 1) Z2 · SO(k)
2k + 3, k ≥ 1 odd SO(2) SO(k + 2) ρ2 ⊗ρ
(2, 2k + 1) C2
13
SO(2) G2
ˆ 7 △ SO(2) SU(2)
ρ2 ⊗φ
Z2 · SU(3)
Z2 · SU(2)
(1, 5)
C2
15
SO(2) Spin(7)
ˆ 7 △ SO(2) SU(3)
ρ2 ⊗∆
G2
SU(3)
(1, 6)
C2
9
SO(5)
ad
U(2)
SO(3) SO(2)
T2
(2, 2)
C2
19
SU(5)
Λ2 µ5
Sp(2)
SU(2) SU(3)
SU(2)2
(4, 5)
C2
Λ2 µ5
1
1
1
∆(Z2 )
U(5)
S · Sp(2)
S · SU(2) SU(3) S · SU(2)2
Table 4.4. Effective q-slice representation on S⊥q = Sn
Aside from a few exceptional representations, they are the isotropy representations of the
Grassmannians G2,m+2 (k) of 2-planes in km+2 , where k = R, C, or H. The pairs of multiplicities
10
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
that occur for the exceptional representations are (1, 6), (1, 5), (4, 5), (2, 2), corresponding to
Ḡq = SO(2) Spin(7), SO(2) G2 , SU(5), U(5), or SO(5).
Note that effectively, there are only four exceptional Gq slice representations, corresponding to the last four rows of Table 4.4. However, special situations occur also when the slice
representation of Ḡq is the isotropy representation of the real Grassmann manifold, when its
multiplicity (1, k) happen to have k = d = 1, 2, 4 or 8. We will refer to these as flips. As may be
expected, the low multiplicity cases (1, 1, 1), (1, 1, 5) and (2, 2, 3) play important special roles.
The latter two are where the exceptional Cayley plane emerges, the only cases where complete
information about the polar data are required.
Accordingly we have organized the proof of 4.1 into four sections depending on the type of
slice representations we have along Q: Three Grassmann flips, three Grassmann series (two non
minimal), two minimal Grassmann representations, and four exceptional representations.
5. Grassmann Flip Gq slice representation
This section will deal with the multiplicity cases (d, d, 1) with d = 2, 4 and 8, leaving d = 1
(minimal and odd) for section 7. We have the following common features:
Lemma 5.1. The isotropy groups Gq and Gr are connected, and the reducible Ḡr slice representation on S1 ∗ Sd = Sd+2 is the standard action by SO(2) SO(d + 1). For the kernels of the
slice representations we have that Kt = {1}, Kq = Kℓr and Kr = Kℓq .
Proof. The Transversality Lemma 2.5 implies that the orbits Q = G q and R = G r are simply
connected since M is. In particular, Gq and Gr are connected since G is. The second claim
follows since d is even (cf. Appendix in [FGT] for a description of reducible polar representations).
Since (Ḡq )ℓr (cf. Table 4.4) as well as (Ḡr )ℓq act effectively on the respective normal spheres
d
S , we see that Kq = Kℓr and Kr = Kℓq . Also since Kt ⊂ Kℓr we have Kt ∩ Kq = Kt but Kt ∩ Kq = {1}
by the Kernel Lemma 2.3 and hence Kt = {1}.
Recall that K is the identity component of the kernel of the Gr action restricted to Sd .
Lemma 5.2. Clearly K⊳Gr , and K ⊂ Gℓt acts transitively on the corresponding normal sphere
S1 with kernel identity component of Kr . Moreover, K ∩ Kq = {1} and hence K ⊂ Gq → Ḡq is
injective.
The reduction M K is a positively curved irreducible cohomogeneity one N0 (K) manifolds with
multiplicity pair (d, 1).
Proof. Note that K ∩ Kq acts trivially on S1 ∗ Sd , so K ∩ Kq ⊂ Kr . The second claim follows since
Kr ∩ Kq = {1}.
Since K ⊳ Gℓt → Ḡq is injective, we see from Table 4.4 that N(K)∩ Gq / Gℓt = N(K)∩ Ḡq /Ḡℓt =
1
S , and hence M K is cohomogeneity one with multiplicity pair (d, 1).
To complete the proof assume by contradiction that the action is reducible, i.e., that the action
by N0 (K)/K on M K is equivalent to the sum action of SO(2) SO(d + 1) on S1 ∗ Sd , where the
isotropy (N0 (K)/K)q is SO(2) SO(d). In all cases, it is easy to see that, the center of Gq intersects
the center of N0 (K) in a nontrivial subgroup S1 . This, together with primitivity implies that, S1
RANK THREE GEOMETRY AND POSITIVE CURVATURE
11
is in the center of G. Notice that, as a subgroup of Gq , S1 can not be in Kq because Kq ⊂ H,
and the factor SO(2) ⊳ Ḡq acts freely on the unit sphere of the slice T q⊥ . Thus, the fixed point
1
set M S coincides with the orbit G q = G / Gq . From the classification of positively curved
homogeneous spaces we get immediately that, G is the product of S1 (or T2 if d = 2) with one
of a few orthogonal groups or unitary groups, each of which is not big enough to contain the
simple group Gt . The desired result follows.
Although what remains is in spirit the same for all the flip cases, we will cary out the arguments for each case individually, beginning with d = 8.
Proposition 5.3. In the Flip (8,8,1) case, C (M, G) or an S1 covering is a building, with the
isotropy representation of E7 / E6 × S1 as a linear model.
Proof. From Lemma 5.1 and Tables 4.4 and 4.3 we obtain the following information about the
local data: Gt = F4 ⊃ Spin(8) = H, Gℓq = Spin(9), Gℓr = Spin(9), Gq = S1 · Spin(10) ⊃
∆(S1 ) · Spin(8) = Gℓt , and Gr = S1 · Spin(9).
Also Gr ⊲K = ∆(S1 ) ⊂ Gℓt ⊂ Gq , and from Lemma 5.2 and Lemma 4.2 we see that the
corresponding reduction, M K is S19 , S19 /Zm , or S19 / S1 = CP9 with the tensor product representation by SO(2) SO(10) of type C2 or induced by it. It is easily seen that the Assumption
(P) in Lemma 3.4 is satisfied as well. In particular, if M K = S19 , the associated chamber system
C (M K ) is the a building of type C2 and by Lemma 3.4 we conclude that C (M, G) is a building.
For the latter two cases, we will use the bundle construction for polar actions to obtain a free
S1 covering of C (M; G). Guided by our knowledge of the cohomogeneity one diagrams , i.e.,
data for the cohomogeneity one manifolds S19 /Zm or CP9 we proceed as follows:
Note that since Gt , Gℓr and Gℓq are simple groups, only the trivial homomorphism to S1 exists.
Now let Ĝq , Ĝr be the graphs of the projection homomorphisms Gq → S1 , and Gr → S1 . We
denote the total space of the corresponding principal S1 bundle over M by P. Then P is a polar
S1 · G manifold, and C (P; S1 · G) covers C (M; G).
Let K̂ ⊂ Ĝℓt be the graph of K in S1 · G. From 3.5 and our choice of data in S1 · G it follows
that PK̂ → M K is the Hopf bundle if M K = CP9 , and the bundle S1 ×Zm S19 → S19 /Zm if M K =
S19 /Zm . In the former case, C (PK̂ ) is the C2 building C (S19 , SO(2) SO(10)) and we are done by
Lemma 3.4 via 3.5. In the latter case, the action on the reduction PK̂ is not primitive, so C (PK̂ ) is
not connected. However, each connected component is the C2 building C (S19 , SO(2) SO(10))
and hence by 3.5 the corresponding component of C (P) is a C3 building covering C (M). When
combined with the previous section, this in turn shows that M K cannot be a lens space when M
is simply connected.
Proposition 5.4. In the Flip (4,4,1) case C (M, G) or an S1 covering is a building, with the
isotropy representation of SO(12)/ U(6) as a linear model.
Proof. From Lemma 5.1 and Tables 4.4 and 4.3 we obtain the following information about the
local data modulo a common Z2 kernel: Gt = Sp(3) ⊃ Sp(1)3 = H, Gℓq = Sp(1) Sp(2), Gℓr =
Sp(2) Sp(1), Gq = S1 Spin(6) Sp(1) ⊃ ∆(S1 ) · Spin(4) Sp(1) = Gℓt , and Gr = S1 Spin(5) Sp(1).
In this case Gr ⊲K = ∆(S1 ) Sp(2)⊳Gℓt ⊂ Gq , and from Lemma 5.2 and Lemma 4.2 we see that
the corresponding reduction, M K is S11 , S11 /Zm , or S11 / S1 = CP5 with the linear tensor product
12
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
representation by SO(2) SO(6) of type C2 or induced by it. It is easily seen that the Assumption
(P) in Lemma 3.4 is satisfied as well. In particular, if M K = S11 , the associated chamber system
C (M K ) is the a building of type C2 and by Lemma 3.4 we conclude that C (M, G) is a building.
If M K = CP5 or a lens space S11 /Zm , we proceed as above with an S1 bundle construction.
Again only the trivial homomorphism to S1 exists from Gt , Gℓr and Gℓq , and we choose Ĝq , Ĝr
to be the graphs of the projection homomorphisms Gq → S1 , and Gr → S1 . We denote the
total space of the corresponding principal S1 bundle over M by P. As above, P is a polar S1 · G
manifold, and C (P; S1 · G) covers C (M; G).
From 3.5 and our choice of data in S1 · G it follows that PK̂ → M K is the Hopf bundle if
M K = CP5 , and the bundle S1 ×Zm S11 → S11 /Zm if M K = S11 /Zm . The proof is completed as
above.
Proposition 5.5. In the Flip (2,2,1) case C (M, G) or an S1 covering is a building, with the
isotropy representation of SU(6)/ S(U(3) U(3)) as a linear model .
Proof. We begin by verifying our earlier claim (see 2.4) that Ḡt is connected also in this case.
From (5.1) we already know that Gr and hence Ḡr is connected, and that its slice representation
is the product action of Ḡr = SO(3) × SO(2) on R3 ⊕ R2 . The singular isotropy group along R2
(away from origin) is SO(3). Hence, the isotropy group Ḡℓq = SO(3).
On the other hand, suppose Ḡt is not connected. Then, by 5.1 Ḡt = Gt = PSU(3) ⋊ Z2
and Gℓq = (S(U(2) U(1))/Z3 ) ⋊ Z2 . In particular the slice representation along ℓq is by Ḡℓq =
PSU(2) ⋊ Z2 acting on S2 = CP1 where Z2 acts by complex conjugation. Contradicting Ḡℓq =
SO(3).
The above and Tables 4.4 and 4.3 yield the following information about the local data modulo
the Z3 kernel: Gt = SU(3) ⊃ T2 = H, Gℓq = S(U(2) U(1)) = U(2), Gℓr = S(U(1) U(2)) = U(2),
Gq = U(2) U(2). Moreover, Gℓt = T3 and Gr = S1 · U(2), where the U(2) factor in Gr is the face
isotropy group of Gℓq .
Here, Gr ⊲K = T2 ⊳ Gℓt ⊂ Gq , and from Lemma 5.2 and Lemma 4.2 we see that the corresponding reduction, M K is S7 , S7 /Zm , or S7 / S1 = CP3 with the linear tensor product representation by SO(2) SO(4) of type C2 or induced by it. Again, the Assumption (P) in Lemma 3.4
is easily checked to hold. In particular, if M K = S7 , we conclude as above that C (M, G) is a
building.
For the latter two cases, we are again guided by the reduction for our bundle construction.
For Ĝt we have no choice but Ĝt = {1} · Gt . We let Ĝq be the graph of the homomorphism
U(2) U(2) → S1 defined by sending (A, B) to det(A) det(B)−1 , and Ĝr the graph of the projection
homomorphism Gr = S1 · U(2) → S1 . This yields a compatible choice of data for a polar S1 · G
action on a principal S1 bundle P over M whose corresponding chamber system C (P; S1 · G) is
a free S1 cover of C (M, G).
Again from 3.5 and our choice of data in S1 · G it follows that PK̂ → M K is the Hopf bundle
if M K = CP3 , and the bundle S1 ×Zm S7 → S7 /Zm if M K = S7 /Zm , and the proof is completed as
above.
Remark 5.6. The tensor representation of SU(3) SU(3) on C3 ⊗ C3 is not polar, but it is polar
on the projective space P(C3 ⊗ C3 ). On the other hand, it is necessary in the above construction
RANK THREE GEOMETRY AND POSITIVE CURVATURE
13
of the covering that both Gq and Gr have T2 factors, since the face isotropy groups Gℓr Gℓq
U(2) which are subgroups in Gt = SU(3), hence a compatible homomorphism to S1 will be
trivial on the face isotropy groups.
6. Non minimal Grassmann Series for Gq slice representation
Recall that there are three infinite families of cases (1, 1, k), k ≥ 1, (2, 2, 2k + 1), k ≥ 1 and
(4, 4, 4k + 3), k ≥ 0 corresponding the real, complex and quaternion Grassmann series for the
Gq slice representation.
We point out that (1, 1, 1) is special in two ways: There are two scenarios. One of them
corresponding to the “Flip” case of d = 1 not covered in the previous subsection, the other
being “standard”. Yet the standard (1, 1, 1) does not appear as a reduction in any of the general
cases (1, 1, k), k ≥ 2. For the (2, 2, 3) case, there are two scenarios as well, both with the same
local data(!): One of them belonging to the family, the other not. Moreover, each of the cases
(2, 2, 2k + 1) with k ≥ 2 admit a reduction to the “Flip” (2, 2, 1) case, whereas (2, 2, 3) does not.
For the reasons just provided, this subsection will deal with the multiplicity cases (1, 1, k), k ≥
2, (2, 2, 2k + 1), k ≥ 2 and (4, 4, 4k + 3), k ≥ 0, each of which has a uniform treatment.
Although the case (2, 2, 3) is significantly different from the other general cases to be treated
here, we begin by pointing out some common features for all the cases (1, 1, k), k ≥ 2, (2, 2, 2k +
1), k ≥ 1 and (4, 4, 4k + 3), k ≥ 0, i.e., including the case (2, 2, 3).
To describe the information we have about the local data in a uniform fashion, we use Gd (k)
to denote SO(k), SU(k) and Sp(k), k ≥ 1, according to d = 1, 2 and d = 4, with the exceptional
convention that G1 (−1) = Z2 , G2 (−1) = S1 or T2 , depending on whether the center of Kt is finite
or not, and G4 (0) = G4 (−1) = {1}. Also, we use the symbol ”=” to mean ”isomorphic” up to a
finite connected covering.
Lemma 6.1. In all cases Gt is connected as are Gq and Gr when d , 1. Moreover Kt = Gd (k)
with the additional possibility that Kt = Gd (k) · S1 when d = 2.
For the q and r vertex isotropy groups we have: Gq = Gd (2) Gd (k + 2) · Gd (−1), Gr =
Gd (2) Gd (k + 1) · Gd (−1). Moreover, the normal subgroup K ⊳ Gr is Gd (k + 1) · Gd (−1), where
Gd (k + 1) is a block subgroup of Gd (k + 2) ⊳ Gq , and if d = 1, “· Gd (−1)” denotes a nontrivial
extension. In particular, Gq = S(O(2) O(k + 2)).
Proof. The connectedness claim is a direct consequence of transversality. The proof follows the
same strategy in all cases, just simpler when all vertex isotropy groups are conneceted. The two
possibilities for Gt when d = 2 correspond to the different rank possibilities for Ḡq , cf. Table
4.4. For these reasons we only provide the proof in the most subtle case of d = 1.
First, notice that the effective slice representation Ḡt = SO(3) on T t⊥ is of type A2 with
principal isotropy group H̄ = Z22 . Hence, H is an extension of Z22 by the kernel Kt . On the
other hand, (Ḡq )0 = SO(2) SO(k + 2) (cf. Table 4.4), and Ḡq ⊂ O(2) O(k + 2), up to a possible
quotient by a diagonal Z2 in the center if k is even. Therefore, H is also an extension of SO(k),
SO(k) · Z2 or SO(k) · Z22 by Kq . This together with Lemma 2.3, implies that Kt = SO(k) and
hence Gt = SO(3) SO(k). In particular, H = SO(k) × Z22 .
We conclude that Gℓr = O(2) SO(k), and similarly, Gℓq = O(2) SO(k), acting on the normal
sphere S1 with principal isotropy group H. Thus Kℓr = SO(k) × Z2 = Kt × Z2 . Since Kq ⊳ Kℓr ,
14
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
we get easily that Kq = {1} or Z2 , since Kq ∩ Kt = {1}. On the other hand, as a subgroup of Gq ,
Gℓr = ∆(O(2)) SO(k). Hence Gq contains exactly two connected components, whose identity
component is SO(2) SO(k + 2) ⊃ (Gℓr )0 . All in all it follows that, Gq = S(O(2) O(k + 2)). The
rest of the proof is straightforward.
Note that Kt contains Gd (k) as a normal subgroup. The fact that the reduction M Gd (k) with the
action by the identity component of its normalizer, N0 (Gd (k)), will give a geometry of type A3
or C3 will play an important role in the d = 1, 2 cases below (cf. 3.6).
′
In what follows we will consider the reduction M K by K′ = Gd (k + 1) ⊳ Gd (k + 1) · Gd (1) =
K ⊂ Gr rather than the one by K.
′
Lemma 6.2. The cohomogeneity one N(K′ ) manifold M K has multiplicity pair (d, 2d − 1),
and the action is not equivalent to the reducible cohomogeneity one action on S2d−1 ∗ Sd .
Proof. For simplicity we give a proof for d = 2, all other cases are the same.
First note that the orbit space of the cohomogeneity one N(K′ )-action is rq, and the two
singular isotropy groups (mod kernel) are SU(2) · S1 and SU(2) · T2 respectively, with principal
isotropy group T2 . Hence the multiplicity pair is (2, 3).
′
To prove that it is not reducible, we argue by contradiction. Indeed, if M K is equivariantly
diffeomorphic to S2 ∗ S3 with the product action of SU(2) U(2), it follows that the normal subgroup SU(2) ⊳ Gq is also normal in N(K′ ). By primitivity G = hGr , Gq i = hN, Gq i and hence
SU(2) is normal in G. On the other hand, the face isotropy group Gℓr ⊂ Gt contains a subgroup
SU(2) which sits as ∆(SU(2)) ⊂ Gq . Therefore, the projection homomorphism p : G → SU(2)
is an epimorphism on ∆(SU(2)). However, since it sits in SU(3) ⊳ Gt it must be trivial, because
any homomorphism from SU(3) to SU(2) is trivial. A contradiction.
When d = 1 this is not immediately of much help since there are several positively curved
irreducible cohomogeneity one manifolds with multiplicity pair (1, 1) (cf. Tables A and E in
[GWZ]) whose associated chamber system is not of type C2 . However, when d = 2, respectively d = 4 corresponding to multiplicity pairs (2, 3), respectively (4, 7) we read off from the
classification in [GWZ] that
′
Corollary 6.3. The universal covering of M K is equivariantly diffeomorphic to a linear
action of type C2 on S11 , CP5 or HP2 when d = 2, and on S23 when d = 4.
We are now ready to deal with each family individually, beginning with d = 1, i.e. with
the standard (1, 1, k ≥ 2) case, where the (almost) effective slice representation at q ∈ Q is the
defining tensor product representation of SO(2) SO(k + 2).
Proposition 6.4. In the standard (1, 1, k) case with k ≥ 2, the associated chamber system
C (M; G) is a building, with the isotropy representation of SO(k + 3)/ SO(3) SO(k) as a linear
model.
Proof. By Lemma 6.1 Kt = SO(k), which is a normal subgroup of the principal isotropy group
H. Consider the reduction M Kt with the action of its normalizer N(Kt ), once again a polar action
with the same section Σ. By Lemma 6.1 it is clear that the identity component of N(Kt ) ∩ Gq
is T2 . Hence, the subaction by N0 (Kt ), the identity component of N(Kt ), is of type A3 , with a
RANK THREE GEOMETRY AND POSITIVE CURVATURE
15
right angle at q. Therefore, from the classification of A3 geometries (cf. section 7 in [FGT])
it is immediate that, the universal cover of M Kt is equivariantly difffeomorphic to S8 with the
linear action of SO(3) SO(3). In particular, if the section Σ = S2 , then M Kt = S8 and the
chamber complex for the subaction is a building of type A3 , and we are done by Remark 3.6,
since Property (P) is clearly satisfied for K = SO(k + 1) ⊳ Gr .
It remains to prove that M Kt is simply connected. Consider the normal subgroup SO(2) ⊳ Gq ,
and the fixed point component M SO(2) , a homogeneous manifold of positive curvature with dimension at least two, since M SO(2) ∩ M Kt ⊂ M Kt is of dimension 2. Since the identity component
of the isotropy group, (Gq )0 = SO(2) SO(k + 2), we see that M SO(2) = Sk+2 or RPk+2 , according to M Kt ∩ M SO(2) = S2 or RP2 , equivalently, according to M Kt = S8 or RP8 . We argue by
contradiction. If M SO(2) = RPk+2 , then the identity connected component of the normalizer
N(SO(2)) acts transitively on it with principal isotropy group SO(2) O(k + 2) ⊂ Gq . Hence
Gq = SO(2) O(k + 2), a contradiction, since Gq = S(O(2) O(k + 2)).
Proposition 6.5. In the standard (2, 2, 2k + 1) case, with k ≥ 2 the chamber system C (M; G)
is covered by a building, with the isotropy representation of U(k + 3)/ U(k) U(3) as a linear
model.
Proof. First note that the reduction M SU(k) , where SU(k) ⊳ Kt , k ≥ 2, is a positively curved
cohomogeneity two manifold of type C3 with multiplicity triple (2, 2, 1). Moreover, SU(k) is a
′
block subgroup in K′ ⊂ K, where K′ = SU(k + 1) ⊂ SU(k + 2) ⊳ Gq and of course M K ⊂ M SU(k) .
We will prove that both reductions above are simply connected, by appealing to the Con′
nectivity Lemma 2.6 of Wilking [Wi3]. To do this we now proceed to prove that codimM K ⊂
M SU(k) = 6, and codimM SU(k) ⊂ M = 6k.
By the Spherical isotropy Lemma 2.8, every irreducible isotropy subrepresentation of K′ =
SU(k + 1) is the defining representation µk+1 . From Table B in [GWZ] and the above fact that
SU(k + 2) ⊃ K′ it follows that, there is a simple normal subgroup L ⊳ G such that, SU(k + 2) ⊳ G q
projects to a block subgroup of L where L = SU(n) if k ≥ 4, L = SU(n) or SO(n) if k = 3, and
finally L = SU(n), SO(n) or one of the exceptional Lie groups F4 ⊂ E6 ⊂ E7 ⊂ E8 , if k = 2.
On the other hand, by the Flip Proposition 5.5 the normalizer N(SU(k)) is either SU(3) SU(3)
or U(3) SU(3) modulo Kt . Since SU(k) ⊳ Kt is a block subgroup in K′ , this together with the
above implies that in fact L = SU(k + 3) for all k ≥ 2, and only one such factor exist. In
particular, the K′ -isotropy representation along ℓt contains exactly 3 copies of µk+1 , one copy
along the normal slice T ℓ⊥t , and two copies along the orbit G / Gℓt . Therefore, the codimension
′
of M K in M is 6(k + 1), and hence the codimension of M SU(k) in M is 6k. By the Connectivity
lemma 2.6 of Wilking, we conclude that πi (M) πi (M SU(k) ) for i ≤ 2, by induction on k. In
particular, M SU(k) is simply connected and hence S17 if dim(M) is odd and CP8 if dim(M) is
even, by the Flip Proposition 5.5. Since Assumption (P) in Lemma 3.4 is satisfied we conclude
from 3.6 that C (M; G) is a building if dim(M) is odd.
It remains to prove that C (M; G) is covered by a building if dim(M) is even. In this case, by
the above we know that π2 (M) π2 (M SU(k) ) Z. On the other hand, from the Transversality
Lemma 2.5 it follows that π2 (M) π2 (G / Gt ), and hence Gt contains at least an S1 in its center,
i.e, SU(3) U(k) ⊳ Gt . By Lemma 6.1 we get that, both Gq and Gr have at least a T2 factor, and
we are now in the same situation as in the proof of Lemma 5.5 above. As a consequence we
16
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
can proceed with the same construction of a principal S1 bundle P over M and conclude that its
associated chamber system is a building covering C (M; G).
Proposition 6.6. In the standard (4, 4, 4k+3) case where k ≥ 0, the chamber system C (M; G)
is a building, with the isotropy representation of Sp(k + 3)/ Sp(k) Sp(3) as a linear model.
Proof. Since the Assumption (P) for K′ = Sp(k + 1) in Lemma 3.4 is easily seen to be satisfied,
′
it suffices by Corollary 6.3 to prove that M K is simply connected. As in the proof of the general
(2, 2, 2k + 1) case above this is achieved via Wilkings Connectivity Lemma 2.6.
Consider the normal subgroup Sp(2)⊳ Gq . It is clear that M Sp(2) is a homogeneous space with
a transitive action by the identity component of its normalizer N0 (Sp(2)) with isotropy group
Gq . By the classification of positively curved homogeneous spaces we get that M Sp(2) is either
S4(k+3)−1 or RP4(k+3)−1 . Moreover, the universal cover Ñ0 (Sp(2)) is Sp(k + 3) Sp(2) Sp(1), and in
particular has the same rank as G by the Rank Lemma.
On the other hand, by Lemma 2.8 and Table B in [GWZ] it follows that, G contains a normal
subgroup isomorphic to Sp(n) so that K′ ⊂ Sp(k + 2) ⊂ Sp(k + 3) ⊂ Sp(n) is in a chain of block
subgroups. Up to a finite cover, we let G = Sp(n) · L. On the other hand, by Corollary 6.3 we
know that N0 (K′ ) = Sp(2) Sp(3)K′ . This together with the information on Ñ(Sp(2)) implies that
G = Sp(k + 3) · L. As in the proof of the (2, 2, 2k + 1) case we see that the isotropy representation
of K′ , along ℓt contains exactly three copies of νk+1 , one copy along the normal slice T ℓ⊥t , and
′
two copies along the orbit G / Gℓt . In particular, the codimension of M K in M is 12(k + 1).
′
Recalling that the dimension of M K is 23, it follows again by connectivity and induction on k
′
as before that M K is simply connected.
7. Minimal Grassmann Gq slice representation
This section will deal with the multiplicity cases (1, 1, 1) and (2, 2, 3), including the appearance of an exceptional Cayley plane action. In all previous cases all reductions considered have
been irreducible polar actions. Here, however, we will encounter reductions, that are reducible
cohomogeneity two actions, and we will rely on the independent classification of such actions
in sections 6 and 7 of [FGT].
′
We begin with the d = 2 case, where by 6.3 we know that the universal covering M̃ K of the
′
reduction M K is diffeomorphic to S11 , CP5 or HP2 . The first two scenarios follow the outline of
the general (2, 2, 2k + 1) case, whereas the latter is significantly different.
Proposition 7.1. In the case of multiplicities (2, 2, 3), C (M; G) is covered by a building, with
′
the isotropy representation of U(7)/ U(4) U(3) as a linear model, provided M K is not diffeomorphic to HP2 .
Proof. By Lemma 6.1, Gt is either SU(3) or U(3) depending on whether Kt is finite or S1 . In
the latter case, the reduction M Kt is a positively curved cohomogeneity two manifold of type
C3 with multiplicity triple (2, 2, 1), as in the general (2, 2, 2k + 1) case, where k ≥ 2 (cf. 6.5).
Therefore, N0 (Kt )/Kt = SU(3) · SU(3) or U(3) · SU(3), by the Flip Proposition 5.5. The desired
result follows, as in the proof of Proposition 6.5.
RANK THREE GEOMETRY AND POSITIVE CURVATURE
17
From now on we assume that, up to finite kernel, Gt = SU(3), and correspondingly, Gq =
U(2) SU(3), and Gr = U(2) SU(2). Moreover, K′ = SU(2), and from our assumption on the re′
duction M K , by Corollary 6.3 the normalizer N(K′ ) contains SU(2) SU(2) SU(3) as its semisimple part. On the other hand, by the Rank Lemma 2.7 we know that rk(G) = 5 (resp. rk(G) = 4)
if dim(M) is odd (resp. even). In particular, SU(2) SU(2) SU(3) is a maximal rank subgroup of
G if rk(G) = 4. In this case, it is immediate, by Borel and de Siebenthal [BS] (see the Table on
page 219), that G is not a simple group of rank 4. Similarly, we claim that G is not a simple
group when its rank is 5: Indeed if so, by Lemma 2.8 and Table B in [GWZ], it would follow
that G = SU(6) and K′ = SU(2) ⊂ SU(3) ⊳ Gq is a block subgroup. This, however, is not
possible, since then N(K′ ) would contain SU(4). Thus, G = L1 · L2 , where L1 , L2 are nontrivial
Lie groups. Without loss of generality, we assume that the projection of SU(3) ⊳ Gq to L2 has
nontrivial image. But then SU(3) must be contained in L2 , because otherwise, the normalizer
N(K′ ) would be much smaller than SU(2) SU(2) SU(3). By Primitivity 2.2 it is easy to see that
Gt is diagonally imbedded in L1 · L2 , since G = hGt , Gℓt i = hGt , K′ i. In particular, both L1
and L2 have rank at least two since the projections from Gt are almost imbeddings, i,e,, have
finite kernel. If both L1 and L2 have rank two, it is easy to see that, L1 = SU(3) and K′ ⊂ L2 ,
where L2 = SU(3) or G2 . Neither scenario is possible: For the latter since, by the primitivity,
G = h∆(SU(3)), K′ i = SU(3) · SU(3), while for the former the semisimple part of N(K′ ) is L1 .
Therefore rk(G) = 5 and once again by Lemma 2.8 and Table B in [GWZ], G = SU(3) SU(4).
Note that dimM = 21 and the principal orbit of K′ in M is of dimension at least 2. In particu′
lar, it follows from Wilkings Connectivity Lemma2.6 that M K is simply connected. Thus, as in
the general case the desired result follows from Lemma 3.4.
Proposition 7.2. In the case of multiplicities (2, 2, 3), M is equivariantly diffeomorphic to
′
the Cayley plane OP2 with an isometric polar action by SU(3) · SU(3), provided M K is diffeomorphic to HP2 .
Proof. Recall that K′ = SU(2) ⊳ Gℓt . By Lemma 2.8 and the slice representation of Gℓt it
′
follows that, every irreducible subrepresentation of K′ on the normal space to M K is the standard
′
representation µ2 on C2 . In particular, the codimension of M K is a multiple of 4, and so M has
dimension divisible by 4. By 6.1 the isotropy group Gt = SU(3) or U(3), and correspondingly,
Gq = U(2) SU(3) or U(2) U(3), and Gr = U(2) SU(2) or U(2) U(2). By the Rank Lemma
rk(G) = rk(Gq ) = 4 or 5.
By Lemma 2.8 the isotropy representations of K′ ⊂ SU(3) ⊳ Gq , as well as of SU(2) ⊂
Gℓq ⊂ Gt , are spherical transitive. By Table B in [GWZ] it follows that, G can not be a simple
group of rank 5, and moreover, G can not contain F4 , Sp(4), SO(8) and SO(9) as a normal
subgroup, since if so, the semisimple part of N0 (K′ )/K′ would not be SU(3), a contradiction to
′
our assumption on the reduction M K , for which N0 (K′ )/K′ = SU(3)· S1 . On the other hand, note
that the identity component of the normalizer N0 (Gt ) = Gt since Gt is a maximal isotropy group
and hence N0 (Gt )/ Gt acts freely on the positively curved fixed point set M Gt of even dimension.
Therefore, G can not contain SU(5) as a normal subgroup, since otherwise, Gt would be a block
subgroup in SU(5) and hence N0 (Gt )/ Gt would not be trivial. Consequently, G is not a simple
group, and moreover, G = L1 · L2 , where SU(3) ⊳ Gt is diagonally imbedded in G. In particular,
both L1 and L2 contain SU(3) as subgroups. It is easy to see that, SU(3) ⊳ Gq ⊂ G = L1 · L2 is
a subgroup in either L1 or L2 , say in L2 . Hence, K′ ⊂ L2 , and L1 ⊳ N(K′ ). It follows that L1 =
18
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
SU(3). Furthermore, L2 can neither be a semi-simple group of rank 3 or G2 , since otherwise,
N0 (K′ )/K′ contains a rank 3 semisimple group. Hence, L2 is SU(3) or U(3). The latter, however,
is impossible: Indeed, in this case Gt = U(3), and the center S1 ⊂ Z(G) would be contained in
Kt , and hence in every principal isotropy groups (the center is invariant under conjugation) thus
1
M S = M.
In summary we have proved that G = SU(3) · SU(3) (indeed a quotient group by ∆(Z3 )),
with Gt = SU(3) diagonally imbedded in G. We claim that this combined with the above
analysis of the isotropy groups modulo conjugation will force the polar data (Gt , Gq , Gr ) ⊂ G
(noting that face isotropy groups are intersections of vertex isotropy groups) to be (Gt , Gq , Gr ) =
(∆(SU(3)), U(2) · SU(3), S(U(2) U(2))), where U(2) ⊂ SU(3) is the upper 2 × 2 block subgroup
in SU(3), and S(U(2) U(2)) ⊂ SU(3) · SU(3) is the product of the lower 2 × 2 block subgroups.
In other words, by the recognition theorem for polar actions [GZ] there is at most one such polar
action. - On the other hand the unique action by the maximal subgroup SU(3) · SU(3) ⊂ F4 , the
isometry group of the Cayley plane OP2 is indeed polar of type C3 [PTh].
To prove the above claim, by conjugation we may assume that Gt = ∆(SU(3)) and Gq =
U(2) · SU(3) as claimed. Moreover, up to conjugation by an element of the face isotropy group
Gℓr = Gt ∩ Gq , we may further assume that K′ ⊳ Gℓt ⊂ Gq is the lower 2×2 block subgroup in the
second factor SU(3) ⊳ G. Note that K′ is a normal subgroup of Gr , indeed the second factor of
SU(2)· SU(2)⊳ Gr ⊂ SU(3)· SU(3). Since Gℓq = ∆(SU(3))∩ Gr , it follows that SU(2)· SU(2)⊳ Gr
is the product of the lower 2 × 2 block subgroups. Since Gr = hSU(2) · SU(2), Hi where
H = ∆(T2 ) is the principal isotropy group, the desired assertion follows.
Next we deal with the case of multiplicity (1, 1, 1), where there are two scenarios: One is
naturally viewed as part of the infinite family (1, 1, k), whereas the other should be viewed as
the flip case with d = 1.
We point out that unlike all other cases an S3 chamber system cover arises in the first case,
corresponding to a polar action of SO(3) SO(3) on HP2 .
Proposition 7.3. For the multiplicity (1, 1, 1) case, the chamber system C (M; G) is covered
by a building, with the isotropy representation of either SO(7)/ SO(4) SO(3), or of Sp(3)/ U(3)
as a linear model.
Proof. Recall that, Ḡt = SO(3), and H̄ = Z22 . We first claim that the identity component (Gt )0 =
SO(3). To see this, recall that the kernel Kt ⊂ Kℓr , and Ḡq is either SO(2) SO(3) or S(O(2) O(3)).
The claim follows since, if dim Kt ≥ 1 or (Gt )0 = S3 , then Kt ∩ Kq is nontrivial, a contradiction
to Lemma 2.3. From this we also conclude that (Gq )0 is not S1 × S3 , since otherwise again
Kt ∩ Kq is non-trivial. Hence it is isomorphic to either SO(2) SO(3) (the “standard” case) or to
the 2 fold covering U(2) of SO(2) SO(3) (the “flip” case). By the Rank Lemma 2.7 it follows
that rank G ≤ 3.
We start with the following observation:
• Let z be cyclic subgroup of the principal isotropy group H with non-trivial image [z] ⊂ H̄.
Then the action by N0 (z) on the reduction M z is a reducible polar action of cohomogeneity 2.
To see this note that the type t orbit in the reduction is no longer a vertex. Indeed the normalizer
of [z] ⊂ H̄ ⊂ SO(3) is O(2).
RANK THREE GEOMETRY AND POSITIVE CURVATURE
19
In addition, note that the identity component of every face isotropy group is S1 . By the Dual
Generation Lemma 7.2 in [FGT] we conclude that
• 7.3.1. The semisimple part of N0 (z) has rank at most one.
To proceed we will prove that
(a). G is not a simple group of rank 3.
This is a direct consequence of 7.3.1 combined with the following algebraic fact: If G is
a rank 3 simple group, i.e., one of SO(6) = SU(4), SO(7) or Sp(3) (up to center), then, the
normalizer of any order 2 subgroup Z2 ⊂ SO(3) = (Gt )0 contains a semisimple subgroup
of rank at least 2. The algebraic fact is easily established by noticing that the inclusion map
SO(3) → G either can be lifted to a homomorphism into one of the four matrix groups, or
SO(3) sits in the quotient image of a diagonally imbedded SU(2) in one of the matrix groups.
Next we are going to prove that
(b). If G is a rank 2 group, then either (M, G) = (CP5 , SU(3)) or (HP2 , SO(3) SO(3)) up to
equivariant diffeomorphism.
Exactly as in Case (a), we can exclude G being SO(5) since a subgroup Z2 ⊂ H̄ ⊂ (Gq )0
will have a normalizer containing SO(4). We now exclude G being the exceptional group G2 .
Otherwise, (Gq )0 must be U(2), and contained in either an SO(4) ⊂ G2 or an SU(3) ⊂ G2 by
Borel-Siebenthal [BS]. The center Z2 ⊂ U(2) is in Kq . For the same reason as above, U(2) is
not in SO(4) ⊂ G2 . Finally, if U(2) ⊂ SU(3) ⊂ G2 , the q orbit (G q)Z2 in the reduction M Z2
contains (G2 / SU(3))Z2 = (S6 )Z2 = S2 . Again by the Dual Generation Lemma 7.2 of [FGT] this
is impossible, since the identity component of the isotropy group of the face opposite of q is a
circle, which cannot act transitively on the orbit (G q)Z2 . Therefore, up to local isomorphism,
G is SO(3) SO(3) or SU(3) respectively. One checks that the corresponding isotropy group
data are given by Gt = ∆(SO(3)) ⊂ SO(3) SO(3), and Gq = O(2) SO(3) ⊂ SO(3) SO(3),
respectively by Gt = SO(3) ⊂ SU(3) (inclusion induced by the field homomorphism), and
Gq = U(2) ⊂ SU(3) as a block subgroup. The recognization theorem then yields (b).
(c). Now suppose G = L1 · L2 , where Li is a rank i Lie group.
If L1 acts freely on M, then L1 = S1 , SO(3), or S3 , and L2 acts on M/L1 in a polar fashion of
type C3 . Hence, M/L1 is even dimensional and thus CP5 or HP2 by (b). In either case, we know
that the universal cover C˜ of the chamber system C (M/L1, L2 ) is a building. Since C (M, G) is
a connected chamber system covering C (M/L1, L2 ) it follows that C˜ is the universal cover of
C (M, G).
Now consider the remaining case where
• L1 does not act freely on M, and we let Zm ⊂ L1 be a cyclic group such that M Zm , ∅.
Note that G can not be SO(3) · T2 , since, then Gt and SO(3) ⊳ Gq would be the same simple
group factor, which is absurd. In particular, the semi-simple part of G has rank at least two.
Thus from now on we may assume that L2 is a rank two semi-simple group. Moreover, by the
argument in Case (b) it is immediate that in fact L2 is either SO(4) or SU(3).
Notice that:
• If Kt is not trivial, then (M Kt , N0 (Kt )) is a polar manifold with the same section, which is
of type A3 . By the Connectivity Lemma 2.6 it follows that M Kt is simply connected. Hence,
from the classification of A3 geometries, M Kt is diffeomorphic to S8 , and the chamber system of
(M Kt , N0 (Kt )) is a building. By 3.6 C (M, G) is a building.
20
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
Therefore, we may assume in the following that Kt = {1}, hence Gt = SO(3). It follows that,
Gq is either S(O(2) O(3)) or U(2).
We split the rest of the proof according to L1 abelian or not. In either case note that the
normalizer N0 (Zm ) is S1 ·L2 . From this we get immediately that Zm 1 H, by appealing to 7.3.1.
(ci) G = S1 ·L2 .
It suffices to prove that the S1 action is free, since then the situation reduces to the previous
rank 2 case.
Note that Zm is normal in G. From this and the above it follows that Zm is neither in Gt nor
in Gℓt . To see this, if Zm ⊂ Gℓt then (Ḡq )ℓt ⊂ Ḡq would contain a non-trivial normal subgroup
of Ḡq contradicting Table 4.4. The proof in the other case is similar but simpler. Hence, M Zm is
either the orbit G r or G q.
Assuming M Zm = G / Gr , it is immediate that L2 = SU(3) from the list of positively curved
homogeneous spaces. On the other hand, notice that Gr is not connected, indeed (Gr )0 = T2 and
Gr ⊃ Gℓq ⊃ O(2), it follows that G / Gr is not simply connected. However, G / Gr is a totally
geodesic submanifold in M which has dimension 11. A contradiction to Wilking’s Connectivity
Lemma 2.6.
Assuming M Zm = G / Gq , corresponding to L2 = SU(3) or SO(4), the universal cover of
G / Gq is a sphere of dimension either 5 or 3. The latter case is ruled out as follows: If Gq = U(2)
then Kq = Z2 is in the center of U(2) hence also in the center of G. This is impossible, since
Kq ⊂ H and G acts effectively on M by assumption. If Gq = S(O(2) O(3)) there are no nontrivial homomorphisms to S1 , hence Gq ⊂ SO(4), which is impossible. For the former case,
Gq = U(2) and G = U(3), with action on G / Gq equivalent to the standard linear action on a 5dimensional spherical space form with Zm in the kernel. Thus, Gq ⊃ Zm × U(2), a contradiction.
(cii) G = L1 · L2, where L1 is a simple rank one group, i.e., either S3 or SO(3).
We will show that in this case G = SO(3) SO(4), with local data Gq = S(O(2) O(3)) ⊂ G,
and Gt = ∆(SO(3)) ⊂ G forcing all data to coincide with those of the isotropy representation of
SO(7)/ SO(3) SO(4), and hence M with the action of G is determined via recognition.
We first prove that L2 = SO(4). If not, we start with an observation that, L1 = SO(3),
and moreover, Gt is a diagonally imbedded subgroup in L1 · L2 . Indeed, otherwise, an order 2
element z ∈ H ⊂ Gt will have a normalizer N0 (z) which contains a rank 2 semisimple subgroup,
contradicting 7.3.1. For the same reason, as above, we see that Gq , U(2) and hence, Gq =
S(O(2) O(3)). Similarly by 7.3.1, SO(3) ⊳ Gq must be diagonally embedded in L1 · L2 . This is
impossible since then N(SO(3))/ SO(3) is finite, but (Gq )0 ⊂ N(SO(3)).
Finally, given that L2 = SO(4) it follows as above that Gq , U(2), hence Gq = S(O(2) O(3)).
Since Gt = SO(3) and Gℓr = O(2) sits diagonally in Gq it follows that Gt sits diagonally in
L1 · L2 , in particular L1 = SO(3). Using the same arguments as above we see that SO(3) ⊳ Gq is
in L2 . All together, all isotropy data are determined.
8. Exceptional Gq slice representation
This section will deal with the remaining cases, all of which are exceptional with multiplicities (1, 1, 5), (2, 2, 2), (4, 4, 5), and (1, 1, 6). All but the latter will occur, and the case of (1, 1, 5)
will include an exceptional action on the Cayley plane.
RANK THREE GEOMETRY AND POSITIVE CURVATURE
21
Proposition 8.1. In the case of the multiplicities (1, 1, 5) where the (effective) slice representation at T q⊥ is the tensor representation of SO(2) G2 on R2 ⊗ R7 , either M is equivariantly
diffeomorphic to the Cayley plane OP2 with an isometric polar action by SO(3) · G2 or C (M, G)
is a building, with the tensor product representation of SO(3) Spin(7) on R3 ⊗ R8 as a linear
model.
Proof. By the Transversality Lemma 2.5 we conclude that Gt is connected since G / Gt is simply
connected. The kernel Kt is a normal subgroup in Gt , as well as of the principal isotropy
group H with quotients Gt /Kt = SO(3), and H /Kt = Z2 ⊕ Z2 respectively (cf. Table 4.3). By
the Slice Lemma 2.3 Kt acts effectively on the q-slice. Combining this with Table 4.4 where
(Ḡq )0 = SO(2) G2 it follows that, the identity components (Kt )0 = H0 = S3 . Thus, Gt = SO(4),
or Spin(4) = S3 × S3 . The latter, however, is impossible, since then Kt = S3 ×Z2 ⊳ S3 ×Q8 =
H where Q8 is the quaternion group of order 8. On the other hand, by Table 4.4 the slice
representation at q is the natural tensor representation of O(2) G2 on R2 ⊗ R7 , where the center
Z2 ⊂ Q8 is in the kernel Kq and so in Kt ∩ Kq . A contradiction. Therefore, Gt = SO(4), and
consequently, Gq = O(2) · G2 , Gr = O(2) · SU(3) and Gℓt = SU(3) · Z22 .
By Lemma 6.3 we have 3 ≤ rk(G) ≤ 4.
Case (i). Assume rk(G) = 3:
By Lemma 6.3 again dimM is even. By [BS] (table on page 219) SO(2) G2 is not a subgroup
in any rank 3 simple group. Therefore, G = L · G2 , where L is a rank one group. By Table 4.4 the
face isotropy group (Gℓr )0 = SU(2)·∆(SO(2)) is diagonally embedded in SO(2) G2 ⊳ Gq ⊂ L·G2 .
It follows that, the composition homomorphism Gt ⊂ G → L is nontrivial, hence surjective onto
L, because Gt = SO(4). Hence L = SO(3) and G = SO(3) G2 since the only proper nontrivial
normal Lie subgroup of SO(4) is S3 with quotient SO(3). By the above, we already know
that, Gt = SO(4) is a diagonal subgroup given by an epimorphism SO(4) → SO(3) and a
monomorphism SO(4) → G2 . It is clear that, up to conjugation, Gq = O(2) · G2 ⊂ G where
O(2) ⊂ SO(3) is the standard upper 2 × 2 block matrices subgroup.
As in the proof of Proposition 7.2 we now claim that there is at most one polar action with
the data as above. Since we are dealing with a non classical Lie group, however, we proceed as
follows:
Given another C3 type polar action of G = SO(3) G2 with isomorphic local data along
a chamber C ′ with vertices t′ , q′ , r′. Without loss of generality we may assume that Gq =
Gq′ ⊂ SO(3) G2 , and moreover, Gt = Gt′ since any two SO(4) subgroups in G2 are conjugate. Moreover, we can further assume that Gℓt = Gℓt′ since the singular isotropy groups
pair for the slice representation at q is unique up to conjugation. In particular, the principal
isotropy groups H = H′ . We prove now (Gℓq )0 = (Gℓq′ )0 = SO(2) SU(2). This clearly implies the assertion since Gr is generated by (Gℓq )0 and Gℓt . Recall that Gt = ∆(SO(4)) ⊂ G,
its composition with the projection to G2 ⊳ G is a monomorphism, so is the composition of
(Gℓq )0 ⊂ (Gr )0 = SO(2) · SU(3) to G2 , hence, (Gℓq )0 is a diagonal subgroup of Gr , whose
projection to the factor SU(3) is injective. Hence it suffices to show that the projection images
of (Gℓq )0 and (Gℓq′ )0 in SU(3) = (Gℓt )0 = (Gℓt′ )0 coincide. On the other hand, note that the
projection image of (Gℓq )0 in SU(3) is the normalizer N0 (H0 ) in SU(3) = (Gℓt )0 , where H0 is the
identity component of the principal isotropy group. The above assertion follows.
22
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
As for existence we again note that SO(3) G2 is a maximal subgroup of the isometry group F4
of the Cayley plane OP2 . The corresponding unique isometric action is indeed polar as proved
in [GK] and of type C3 .
Case (ii). Assume rk(G) = 4:
By Lemma 2.7, dimM is odd. Consider the reduction M H0 with the action of N0 (H0 ), the
identity component of the normalizer. Note that, this is also a C3 type polar action, but the
multiplicity triple is (1, 1, 1). By appealing to Lemma 2.8, the codimension of M H0 is divisible
by 4. Thus from the (1, 1, 1) case it follows that, the universal cover M̃ H0 is S11 , and the identity
component N0 (H0 ) is either U(3) or SO(3) SO(4), modulo kernel.
We are going to prove that M H0 is simply connected. It suffices to show that M H0 ⊂ M is 2connected. This follows trivially by the Connectivity Lemma of Wilking 2.6, if the codimension
of M H0 is at most 12.
If G2 ⊳ Gq is a normal subgroup of G, then G = L · G2 where L is a rank 2 group. Then
N0 (H0 )/ H0 is isomorphic to L · SO(3). Hence L = SO(4). It is easy to count the codimension
to see that it is strictly less than 12.
If G2 is not a normal subgroup, by Lemma 2.8 the isotropy representation of SU(3) ⊂ G2 ⊂ G
is spherical transitive. Hence, G contains a normal simple Lie subgroup L, such that G2 ⊂
Spin(7) ⊂ L is spherical. We claim that L = Spin(7). If not, L contains Spin(8) such that
Spin(7) ⊂ L is a block subgroup in Spin(8), and hence N0 (H0 ) contains Spin(5), which contradicts the above. This proves that G = L1 · Spin(7), where L1 is a rank 1 group. From this we
get that the isotropy subrepresentation of G / H0 contains exactly three copies of the standard
defining representation of SU(2), hence the desired estimate for the codimension.
In summary we conclude that M H0 = S11 , N0 (H0 ) = SO(3) SO(4) and hence, from the multiplicity (1, 1, 1) case, the chamber system for the action of N0 (H0 ) is a building of type C3 . By
remark 3.6 we conclude that C (M, G) is a building.
Proposition 8.2. There is no polar action of type C3 type with multiplicities (1, 1, 6), where
the (effective) slice representation at T q⊥ is the tensor product representation of SO(2) Spin(7)
on R2 ⊗ R8 .
Proof. We will prove that, if there is such a slice representation at q, the chamber system
C (M, G) is a building. The desired claim follows from the classification of C3 buildings, i.e.,
indeed there is no such a building.
To proceed, note that from Table 4.4 Ḡq = SO(2) Spin(7), and the principal isotropy group
H̄ = SU(3). It follows that, up to local isomorphism Gt = SU(3) SO(3) with Kt = SU(3). Notice
that, the reduction (M Kt , N0 (Kt )) is of cohomogeneity 2 with the same section. It is clear that it
is of type A3 since the q vertex is a vertex with angle π/2, because N0 (Kt ) ∩ Gq is T2 . By the
classification of A3 geometries it follows that, M Kt is either S8 or RP8 . We claim that M Kt = S8 ,
and hence the chamber system for (M Kt , N0 (Kt )) is a building. By appealing to 3.6 it follows
that C (M, G) is a building. To see the claim, it suffices to prove that M Kt is orientable and
hence simply connected, thanks to the positive curvature. By 2.8 the isotropy representation of
2
Kt = SU(3) is the defining complex representation. From this it is immediate that, M Kt = M T ,
and hence oriented, where T2 ⊂ Kt is a maximal torus.
RANK THREE GEOMETRY AND POSITIVE CURVATURE
23
Proposition 8.3. When the multiplicity triple is (2, 2, 2), there are two scenarios. In either
case C (M, G) is a building, with linear model the adjoint polar representation of either SO(7)
or of Sp(3) on S20 .
Proof. By Lemma 2.5 we know that all vertex isotropy groups are connected. Notice that, by
Table 4.4, the slice representation at q is the adjoint representation of SO(5) on R10 . Together
with Proposition 2.3, up to local isomorphism, the local isotropy group data are determined
as follows: Gt = U(3), Gq = SO(5) S1 and Gr = SO(3) U(2). Moreover, H = T3 , Gℓt =
SO(3) SO(2) S1 , and K′ = SO(3) ⊳ Gℓt .
Let SO(2) = K′ ∩ H ⊂ K′ . Consider the reduction (M SO(2) , N(SO(2))). It is once again a polar
manifold with the same section. For such a reduction, notice that: the face ℓq has multiplicity
2, the face ℓt is exceptional with normal sphere S0 , and Gq ∩N(SO(2))/ Gℓt ∩N(SO(2)) = S2 .
Therefore, the action of N(SO(2)) is reducible with fundamental chamber rqq′ , where q′ is a
reflection image of q, and rq = ℓt is of exceptional orbit type. In particular, the multiplicities
at q′ are (2, 2), hence the slice representation at q′ for the N(SO(2))-action is again the adjoint
representation of SO(5) on R10 . This clearly implies that q′ is a fixed point. On the other hand,
notice that M SO(2) is orientable and hence simply connected. Therefore, by Theorem 6.2 of
[FGT] we know that M SO(2) = S10 . Since Property (P) holds for SO(2) it follows from Remark
3.6 that C (M, G) is a building.
Remark 8.4. We remark that in the above proof, the chamber system of (M SO(2) , N(SO(2)))
is a building of type A1 × C2 but the one for (M SO(2) , N0 (SO(2))) is not.
Proposition 8.5. In the case of the multiplicities (4, 4, 5), the chamber system C (M; G) is
covered by a building, with the isotropy representation of SO(14)/ U(7) as a linear model.
Proof. By Lemma 2.5 we know that all isotropy groups are connected. Note that Ḡt = Sp(3),
and Ḡq = SU(5) or U(5). By Lemma 2.3, it is easy to see that:
• if Gt is semisimple, then, up to local isomorphism, Gt = Sp(3), Gr = Sp(2) SU(3), Gq =
SU(5) Sp(1) and Gℓt = SU(3) Sp(1)2 , where Sp(1) = Kq is a subgroup of Gt .
• if Gt is not semisimple, then Kt = S1 , and all isotropy groups data are the product of S1
with the corresponding data above.
We now prove that G contains SU(7) as a normal subgroup. By Lemma 2.8 the isotropy
representations of G / Sp(2) and G / SU(3) are both spherical, where Sp(2), SU(3) are normal
factors of face isotropy groups. Hence, a normal factor L of G is either SO(n) or SU(n), by
Table B in [GWZ]. Moreover, the subgroup Kq ⊂ Gt is contained in a block subgroup SO(4) ⊂ L
(resp. a block subgroup SU(2) ⊂ L) if L = SO(n) (resp. L = SU(n)). Since N0 (Kq ) contains
Gq , it follows that n ≥ 14 (resp. n ≥ 7) if L = SO(n) (resp. SU(n)). To rule out the former
case, consider the fixed point set M Kq with the polar action of N0 (Kq ). It is clearly a reducible
cohomogeneity 2 action with q a vertex of angle π/4. By the Dual Generation Lemma 7.2 of
[FGT] it follows that N0 (Kq ) is either Gq (the fixed point case) or the product of SU(5) ⊳ Gq with
the face isotropy group opposite to q in the reduction M Kq /N0 (Kq ). From this it is immediate
that L = SU(7).
Note that if Gt is semisimple, or dim(M) is even, then rank G ≤ 6, by the Rank Lemma,
and hence G = SU(7). For the remaining case, i.e., dim(M) being odd and Gt = S1 · Sp(3),
24
FUQUAN FANG, KARSTEN GROVE, AND GUDLAUGUR THORBERGSSON
we now prove that G = U(7), up to local isomorphism. Indeed, it is clear that rank G = 7,
and hence G = SU(7) · L2 , where L2 is a rank 1 group. It suffices to prove that L2 = S1 . Let
K′ = SU(3) ⊳ Gℓt . It is clear that the projection p2 : G → L2 is trivial, when restricted to either
of Sp(3) ⊳ Gt and K′ ⊂ Gq . By primitivity 2.2, G = hGt , Gℓt i = hGt , K′ i. Therefore, p2 (Gt ) = L2
and hence, L2 = S1 .
To complete the proof, we split into two cases, i.e, dim(M) being even or odd. For the
former, Kt = S1 and G = SU(7). It is clear that Gt = Sp(3) S1 is a subgroup of U(6) ⊂ SU(7)
and Gq = SU(5) Sp(1)· S1 is the normalizer N(Sp(1)) in G, where Sp(1)⊳ Gℓr ⊂ Gt . This forces
all isotropy groups data to be the same as for the linear cohomogeneity 2 polar action on CP20
induced from the isotropy representation of SO(14)/ U(7). Hence, in particular, the chamber
system C (M; G) is covered by a building. For the latter, G = SU(7) or U(7) depending on
′
Kt = {1} or S1 . The fixed point set M K is odd dimensional, since the isotropy representation of
′
K′ is the defining complex representation. Note that N0 (K′ ) = SU(4) Ti ·K′ , i = 1, 2, and M K is
equivariantly diffeomorphic to S11 with a standard linear cohomogeneity one action of type C2 .
Hence, by Lemma 3.4, C (M; G) is a building.
References
[AT]
M. Alexandrino and D. Töben, Singular Riemannian foliations on simply connected spaces, Differential
Geom. Appl. 24 (2006), 383–397.
[BS]
A. Borel and J. De Siebenthal, Les sous-groupes fermes de rang maximum des groupes de Lie clos,
Comment. Math. Helv. 23(1949), 200–221.
[BSp] K. Burns and R. Spatzier, On topological Tits buildings and their classification, Inst. Hautes Études Sci.
Publ. Math. 65 (1987), 5–34.
[CL]
R. Charney and A. Lytchak, Metric characterizations of spherical and Euclidean buildings, Geom. Topol.
5 (2001), 521–550.
[De]
O. Dearricott, A 7-manifold with positive curvature, Duke Math. J. 158 (2011), 307–346.
[EH]
J.-H. Eschenburg and E. Heintze, On the classification of polar representations, Math. Z. 232 (1999),
391–398.
[FGT] F. Fang, K. Grove and G. Thorbergsson, Tits geometry and positive curvature, preprint.
[GK]
C. Gorodski and A. Kollross, Some remarks on polar actions, preprint .
[Go]
F. J. Gozzi, Low Dimensional Polar Actions, arXiv 1407.0638
[GVZ] K. Grove, L. Verdiani and W. Ziller, An exotic T 1 S4 with positive curvature, Geom. Funct. Anal. 21
(2011), 499–524.
[GWZ] K. Grove, B. Wilking and W. Ziller, Positively curved cohomogeneity one manifolds and 3-Sasakian
geometry, J. Differential Geom. 78 (2008), 33–111.
[GZ]
K. Grove and W. Ziller, Polar Actions and Manifolds, The Journal of Fixed Point Theory and Applications
11 (2012), 279–313
[GKK] T. Grundhöfer, N. Knarr, and L. Kramer, Flag-homogeneous compact connected polygons. II, Geom.
Dedicata 83 (2000), 1–29.
[HS]
H. Hopf and H. Samelson, Ein Satz über die Wirkungsräume geschlossener Liescher Gruppen, Comment.
Math. Helv. 13 (1941), 240–251.
[KL]
L. Kramer and A. Lytchak, Homogeneous compact geometries, Transform. Groups 19 (2014), 793–852.
[Ly]
A. Lytchak, Polar foliations of symmetric spaces, Geom. Funct. Anal. 24 (2014), 1298–1315.
[Ne]
A. Neumaier, Some sporadic geometries related to PGL(3, 2), Arch. Math. 42 (1984), 89–96.
[PTh] F. Podestà and G. Thorbergsson, Polar actions on rank-one symmetric spaces, J. Differential Geom. 53
(1999), 131–175.
[Ro]
M. Ronan, Lectures on buildings. Perspectives in Mathematics 7. Academic Press, Inc., Boston, MA,
1989.
RANK THREE GEOMETRY AND POSITIVE CURVATURE
[Su]
[Ti1]
[Ti2]
[Ve]
[Wie]
[Wi1]
[Wi2]
[Wi3]
25
K. Sugahara, The isometry group and the diameter of a Riemannian manifold with positive curvature,
Math. Japon. 27 (1982), 631–634.
J. Tits, Buildings of spherical type and finite BN-pairs.. Lecture Notes in Mathematics 386. SpringerVerlag, Berlin-New York, 1974.
J. Tits, A local approach to buildings, The geometric vein. The Coxeter Festschrift. Edited by C. Davis,
B. Grünbaum, and F. A. Sherk, 519–547, Springer, New York-Berlin, 1981.
L. Verdiani, Cohomogeneity one manifolds of even dimension with strictly positive sectional curvature,
J. Differential Geom., 68 (2004), 31–72.
S. Wiesendorf, Taut submanifolds and foliations, J. Differential Geom. 96 (2014), 457–505.
B. Wilking, Nonnegatively and positively curved manifolds, Surveys in differential geometry. Vol. XI,
25–62, Surv. Differ. Geom. 11, Int. Press, Somerville, MA, 2007.
B. Wilking, Positively curved manifolds with symmetry, Ann. of Math. 163 (2006), 607–668.
B. Wilking, Torus actions on manifolds of positive sectional curvature, Acta Math. 191 (2003), no. 2,
259–297.
Department of Mathematics, Capital Normal University, Beijing 100048, China
E-mail address: fuquan [email protected]
Department of Mathematics, University of Notre Dame, Notre Dame, IN 46556, USA
E-mail address: [email protected]
Mathematisches Institut, Universität zu Köln, Weyertal 86-90, 50931 Köln, Germany
E-mail address: [email protected]
| 4math.GR
|
Real-Time Rejection and Mitigation of
Time Synchronization Attacks on the
Global Positioning System
arXiv:1802.01649v1 [cs.SY] 5 Feb 2018
Ali Khalajmehrabadi, Student Member, IEEE, Nikolaos Gatsis, Member, IEEE,
David Akopian, Senior Member, IEEE, and Ahmad F. Taha, Member, IEEE
Abstract—This paper introduces the Time Synchronization Attack Rejection and Mitigation (TSARM) technique for
Time Synchronization Attacks (TSAs) over the Global Positioning System (GPS). The technique estimates the clock
bias and drift of the GPS receiver along with the possible
attack contrary to previous approaches. Having estimated
the time instants of the attack, the clock bias and drift of
the receiver are corrected. The proposed technique is computationally efficient and can be easily implemented in real
time, in a fashion complementary to standard algorithms
for position, velocity, and time estimation in off-the-shelf
receivers. The performance of this technique is evaluated
on a set of collected data from a real GPS receiver. Our
method renders excellent time recovery consistent with the
application requirements. The numerical results demonstrate that the TSARM technique outperforms competing
approaches in the literature.
Index Terms—Global Positioning System, Time Synchronization Attack, Spoofing Detection
I. I NTRODUCTION
I
NFRASTRUCTURES such as road tolling systems, terrestrial digital video broadcasting, cell phone and air traffic
control towers, real-time industrial control systems, and Phasor
Measurement Units (PMUs) [1] heavily rely on synchronized
precise timing for consistent and accurate network communications to maintain records and ensure their traceability. The
Global Positioning System (GPS) provides time reference of
microsecond precision for these systems [2]–[5].
The GPS-based time-synchronization systems use the civilian GPS channels, which are open to the public [6], [7]. The
unencrypted nature of these signals makes them vulnerable
to unintentional interference and intentional attacks. Thus,
unauthorized manipulation of GPS signals leads to disruption
of correct readings of GPS-based time references, and thus,
is called Time Synchronization Attack (TSA). To address the
impact of malicious attacks, for instance on PMU data, the
Electric Power Research Institute published a technical report
that recognizes the vulnerability of PMUs to GPS spoofing
under its scenario WAMPAC.12: GPS Time Signal Compromise [8]. These attacks introduce erroneous time stamps which
are eventually equivalent to inducing wrong phase angle in
Authors are with the Electrical and Computer Engineering Department, University of Texas at San Antonio, San Antonio, TX, 78249,
USA (e-mails: {ali.khalajmehrabadi, nikolaos.gatsis, david.akopian, ahmad.taha}@utsa.edu).
the PMU measurements [9], [10]. The impact of TSAs on
generator trip control, transmission line fault detection, voltage
stability monitoring, disturbing event locationing, and power
system state estimation has been studied and evaluated both
experimentally [11] and through simulations [12]–[14].
Intentional unauthorized manipulation of GPS signals is
commonly referred to as GPS spoofing, and can be categorized
based on the spoofer mechanism as follows:
• Jamming (blocking): The spoofer sends high power signals
to jam the normal operation of the receiver by disrupting
the normal operation of the victim receiver, often referred
to as loosing lock. Then, the victim receiver may lock onto
the spoofer signal after jamming [9], [15]–[17].
• Data level spoofing: The spoofer manipulates the navigation
data such as orbital parameters (ephemerides) that are used
to compute satellite locations [13], [15], [18].
• Signal level spoofing: The spoofer synthesizes GPS-like
signals that carry the same navigation data as concurrently
broadcasted by the satellites [11].
• Record-and-replay attack: The spoofer records the authentic
GPS signals and retransmits them with selected delays at
higher power [9], [19]. Typically the spoofer starts from
low power transmission and increases its power to force
the receiver to lock onto the spoofed (delayed) signal. The
spoofer may change the transmitting signal properties such
that the victim receiver miscalculates its estimates.
Common off-the-shelf GPS receivers lack proper mechanisms to detect these attacks. A group of studies have been
directed towards evaluating the requirements for successful
attacks, theoretically [16] and experimentally [11], [29]–[31].
For instance, the work in [30] has designed a real spoofer as
a Software Defined Radio (SDR) that records authentic GPS
signals and retransmits fake signals. It provides the option of
manipulating various signal properties for spoofing.
A. Spoofing Detection Techniques in the Literature
The first level of countermeasures to reduce the effect of
malicious attacks on GPS receivers typically relies on the
Receiver Autonomous Integrity Monitoring (RAIM) [4]. Offthe-shelf GPS receivers typically apply RAIM consistency
checks to detect the anomalies exploiting measurement redundancies. For example, RAIM may evaluate the variance of
GPS solution residuals and consequently generate an alarm
if it exceeds a predetermined threshold. Similar variance
authentication techniques have been proposed in [22], [32]
TABLE I
GPS S POOFING D ETECTION T ECHNIQUES : D ETECTION D OMAIN AND I MPLEMENTATION A SPECTS
Method
EKF
CUSUM [20]
Attack Detection Domain
GPS Navigation domain
GPS baseband signal domain
Attack
Not Estimated
Not Estimated
Ref. [21]
GPS baseband & power grid domains
Not Estimated
SPREE [22]
Ref. [23], [24]
Ref. [25], [26]
Ref. [27]
Ref. [28]
TSARM
GPS baseband signal domain
GPS baseband signal domain
GPS navigation domain
GPS navigation domain
GPS navigation domain
GPS navigation domain
Not Estimated
Not Estimated
Not Estimated
Not Estimated
Not Estimated
Estimated
based on hypothesis testing on the Kalman filter innovations;
however, they are vulnerable to smarter attacks that pass RAIM
checks or the innovation hypothesis testing.
A plethora of countermeasures have been designed to
make the receivers robust against more sophisticated attacks
[9], [15], [17]–[19], [21]–[25], [27], [28], [33]–[35]. Vector
tracking exploits the signals from all satellites jointly and
feedbacks the predicted position, velocity, and time (PVT) to
the internal lock loops [23], [24], [33]. If an attack occurs, the
lock loops become unstable which is an indication of attack.
Cooperative GPS receivers can perfrom authentication check
by analyzing the integrity of measurements through peer-topeer communications [24], [25], [34], [35]. Also, a quick
sanity check for stationary time synchronization devices is to
monitor the estimated location. As the true location can be
known a priori, any large shift that exceeds the maximum
allowable position estimation error can be an indication of
attack [28]. The receiver carrier-to-noise receiver can be used
as an indicator of spoofing attack [17]. In [21], the difference
between the carrier-to-noise ratios of two GPS antennas has
been proposed as a metric of PMU trustworthiness. In addition, some approaches compare the receiver’s clock behavior
against its statistics in normal operation [19], [28], [33].
B. Existing Literature Gaps
As discussed above, prior research studies addressed a
breadth of problems related to GPS spoofing. However, there
are certain gaps that should still be addressed: 1) Most of the
works do not provide analytical models for different types of
spoofing attacks. The possible attacking procedure models are
crucial for designing the countermeasures against the spoofing
attacks. 2) Although some countermeasures might be effective
for a certain type of attack, a comprehensive countermeasure
development is still lacking for defending the GPS receiver.
This is practically needed as the receiver cannot predict the
type of attack. 3) The main effort in the literature is in
detection of possible spoofing attacks. However, even with
the spoofing detection, the GPS receiver cannot resume its
normal operation, especially in PMU applications where the
network’s normal operation cannot be interrupted. So, the
spoofing countermeasures should not only detect the attacks
but also mitigate their effects so that the network can resume
its normal operation. 4) There is a need for simpler solutions
which can be integrated with current systems.
Implementation Aspects
Benchmark for most common GPS receivers
Applies hypothesis testing on packets of received signal
Combines the statistics of carrier-to-noise ratio difference
between two GPS antennas
Applies auxiliary peak tracking in the correlators of receiver
Applies a position-information-aided vector tracking loop
Needs collaboration among multiple GPS receivers
Applies an anti-spoofing particle filter
Applies hypothesis testing on a GPS clock signature
Applies a real-time optimization technique
Relevant
Yes
No
No
No
No
No
Yes
Yes
-
C. Contributions of This Work
This work addresses the previously mentioned gaps for
stationary time synchronization systems. To the best of our
knowledge, this is the first work that provides the following
major contributions: 1) The new method is not a mere spoofing
detector; it also estimates the spoofing attack. 2) The spoofed
signatures, i.e., clock bias and drift, are corrected using the
estimated attack. 3) The new method detects the smartest
attacks that maintain the consistency in the measurement set.
A descriptive comparison between our solution and representative works in the literature is provided in Table I. A review
of the spoofing detection domain shows that most of the prior
art operates at the baseband signal processing domain, which
necessitates manipulation of the receiver circuitry. Hence, the
approach in the present paper is compared only to those works
whose detection methodology lies in navigation domain.
The proposed TSA detection and mitigation approach in
this paper consists of two parts. First, a dynamical model
is introduced which analytically models the attacks in the
receiver’s clock bias and drift. Through a proposed novel Time
Synchronization Attack Rejection and Mitigation (TSARM)
approach, the clock bias and drift are estimated along with
the attack. Secondly, the estimated clock bias and drift are
modified based on the estimated attacks so that the receiver
would be able to continue its normal operation with corrected
timing for the application. The proposed method detects and
mitigates the effects of the smartest and most consistent
reported attacks in which the position of the victim receiver is
not altered and the attacks on the pseudoranges are consistent
with the attacks on pseudorange rates.
Different from outlier detection approaches in [36], [37],
the proposed method detects the anomalous behavior of the
spoofer even if the measurement integrity is preserved. The
spoofing mitigation scheme has the following desirable attributes: 1) It solves a small quadratic program, which makes
it applicable to commonly used devices. 2) It can be easily integrated into existing systems without changing the receiver’s
circuitry or necessitating mulitple GPS receivers as opposed
to [21]–[24], [33], [34]. 3) It can run in parallel with current
systems and provide an alert if spoofing has occurred. 4)
Without halting the normal operation of the system, corrected
timing estimates can be computed.
The proposed anti-spoofing technique has been evaluated
using a commercial GPS receiver with open-source measurements access [38]. These measurements have been perturbed
with spoofing attacks specific to PMU operation. Applying the
proposed anti-spoofing technique shows that the clock bias of
the receiver can be corrected within the maximum allowable
error in the PMU IEEE C37.118 standard [39].
Paper Organization: A brief description of the GPS is
described in Section II. Then, we provide the models for
possible spoofing attacks in Section III. Section IV elaborates
on the proposed solution to detect and modify the effect of
these attacks. Our solution is numerically evaluated in Section
V followed by the conclusions in Section VI.
II. GPS PVT E STIMATION
In this section, a brief overview of the GPS Position,
Velocity, and Time (PVT) estimation is presented.
The main idea of localization and timing through GPS is
trilateration, which relies on the known location of satellites
as well as distance measurements between satellites and the
GPS receiver. In particular, the GPS signal from satellite n
contains a set of navigation data, comprising the ephemeris and
the almanac (typically updated every 2 hours and one week,
respectively), together with the signal’s time of transmission
(tn ). This data is used to compute the satellite’s position
pn = [xn (tn ), yn (tn ), zn (tn )]T in Earth Centered Earth Fixed
(ECEF) coordinates, through a function known to the GPS
receiver. Let tR denote the time that the signal arrives at the
GPS receiver. The distance between the user (GPS receiver)
and satellite n can be found by multiplying the signal propagation time tR − tn by the speed of light c. This quantity is
called pseudorange: ρn = c(tR − tn ), n = 1, . . . , N , where
N is the number of visible satellites. The pseudorange is not
the exact distance because the receiver and satellite clocks
are both biased with respect to the absolute GPS time. Let
the receiver and satellite clock biases be denoted by bu and
bn , respectively. Therefore, the time of reception tR and tn
are related to their absolute values in GPS time as follows:
+ bn , n = 1, . . . , N. The
+ bu ; tn = tGPS
tR = tGPS
n
R
bn ’s are computed from the received navigation data and are
considered known. However, the bias bu must be estimated
and should be subtracted from the measured tR to yield the
receiver absolute GPS time tGPS
R , which can be used as a time
reference used for synchronization. Synchronization systems
time stamp their readings based on the Coordinated Universal
Time (UTC) which has a known offset with the GPS time as
tUTC
= tGPS
− ∆tUTC , where ∆tUTC is available online.1
R
R
Let pu = [xu , yu , zu ]T be the coordinates of the GPS
receiver, and dn its true range to satellite n. This distance
is expressed via the locations pu , pn and the times tGPS
R ,
tGPS
as dn = kpn − pu k2 = c(tGPS
− tGPS
). Therefore, the
n
n
R
measurement equation becomes
ρn = kpn − pu k2 + c(bu − bn ) + ρn
(1)
where n = 1, . . . , N , and ρn represents the noise. The
unknowns in (1) are xu , yu , zu , bu and therefore measurements
from at least four satellites are needed to estimate them.
1 https://confluence.qps.nl/qinsy/en/utc-to-gps-time-correction-\
32245263.html (accessed Jan. 16, 2018).
Furthermore, the nominal carrier frequency (fc =
1575.42 MHz) of the transmitted signals from the satellite
experiences a Doppler shift at the receiver due to the relative
motion between the receiver and the satellite. Hence, in
addition to pseudoranges, pseudorange rates are estimated
from the Doppler shift and are related to the relative satellite
velocity vn and the user velocity vu via
pn − pu
+ ḃu + ρ̇n
ρ̇n = (vn − vu )T
(2)
kpn − pu k
where ḃu is the clock drift.
In most cases, there are more than four visible satellites,
resulting in an overdetermined system of equations in (1)
and (2). Typical GPS receivers use nonlinear Weighted Least
Squares (WLS) to solve (1) and (2) and provide an estimate of
the location, velocity, clock bias, and clock drift of the receiver,
often referred to as PVT solution. To additionally exploit the
consecutive nature of the estimates, a dynamical model is used.
The conventional dynamical model for stationary receivers is
a random walk model [3, Chap. 9]
xu [l + 1]
xu [l]
yu [l]
yu [l + 1]
03×2
I3×3
zu [l + 1] =
zu [l] + w[l] (3)
1
∆t
02×3
bu [l]
bu [l + 1]
0 1
ḃu [l]
ḃu [l + 1]
where l is the time index, ∆t is the time resolution (typically
1 sec), and w is the noise. The dynamical system (3) and
measurement equations (1) and (2) are the basis for estimating
the user PVT using the Extended Kalman Filter (EKF).
Previous works have shown that simple attacks are able
to mislead the solutions of WLS or EKF. Stationary GPSbased time synchronization systems are currently equipped
with the position-hold mode option which can potentially
detect an attack if the GPS position differs from a known
receiver location by a maximum allowed error [40]. This can
be used as the first indication of attack. But, more advanced
spoofers, such as the ones developed in [30], have the ability to
manipulate the clock bias and drift estimates of the stationary
receiver without altering its position and velocity (the latter
should be zero). So, even with EKF on the conventional
dynamical models, perturbations on the pseudoranges in (1)
and pseudorange rates in (2) can be designed so that they
directly result in clock bias and drift perturbations without
altering the position and velocity of the receiver.
III. M ODELING T IME S YNCHRONIZATION ATTACKS
This section puts forth a general attack model that encompasses the attack types discussed in the literature. This model
is instrumental for designing the anti-spoofing technique discussed in the next section.
While TSAs have different physical mechanisms, they manifest themselves as attacks on pseudorange and pseudorange
rates. These attacks can be modeled as direct perturbations on
(1) and (2) as
ρs [l] = ρ[l] + sρ [l]
(4)
ρ̇s [l] = ρ̇[l] + sρ̇ [l]
5000
50000
0
0
-5000
-10000
-50000
-15000
0
50
100
150
200
250
0
300
50
100
150
200
250
300
350
400
250
300
350
400
(a)
(a)
0
500
-2000
-4000
0
-6000
-8000
0
50
100
150
200
250
300
-500
0
(b)
Time (s)
100
150
200
(b)
Time(s)
Fig. 1. Type I attack on (a) pseudorange and (b) pseudorange rate
versus local observation time.
where sρ and sρ̇ are the spoofing perturbations on pseudoranges and pseudorange rates, respectively; and ρs and ρ̇s are
respectively the spoofed pseudorange and pseudorange rates.
A typical spoofer follows practical considerations to introduce feasible attacks. These considerations can be formulated
as follows: 1) An attack is meaningful if it infringes the
maximum allowed error defined in the system specification.
For instance in PMU applications, the attack should exceed
the maximum allowable error tolerance specified by the IEEE
C37.118 Standard, which is 1% Total Variation Error (TVE),
equivalently expressed as 0.573◦ phase angle error, 26.65 µs
clock bias error, or 7989 m of distance-equivalent bias error
[39]. On the other hand, CDMA cellular networks require
timing accuracy of 10 µs.2 2) Due to the peculiarities of
the GPS receivers, the internal feedback loops may loose
lock on the spoofed signal if the spoofer’s signal properties
change rapidly [11], [29]. 3) The designed spoofers have the
ability to manipulate the clock drift (by manipulating the
Doppler frequency) and clock bias (by manipulating the code
delay) [30]. These perturbations can be applied separately,
however, the smartest attacks maintain the consistency of the
spoofer’s transmitted signal. This means that the pertubations
on pseudoranges, sρ , are the integration of perturbations over
pseudorange rates, sρ̇ , in (4).
Here, distinguishing between two attack procedures is advantageous as the literature includes very few research reports
on the technical intricacies of the spoofer constraints:
• Type I: The spoofer manipulates the authentic signal so
that the bias abruptly changes in a very short time [13],
[15], [28]. Fig. 1 illustrates this attack. The attack on the
pseudoranges suddenly appears at t = 30 s and perturbs
the pseudoranges by 8000 m. The equivalent attack on
pseudorange rates is a Dirac delta function.
• Type II: The spoofer gradually manipulates the authentic
signals and changes the clock bias through time [11], [17],
[19], [28], [29], [41]. This attack can be modeled by
sρ [l] = sρ [l − 1] + sρ̇ [l]∆t
(5)
sρ̇ [l] = sρ̇ [l − 1] + ṡρ̇ [l]∆t
2 http://www.endruntechnologies.com/cdma
2017).
50
(accessed
Sept.
11,
Fig. 2. Type II attack on (a) pseudorange and (b) pseudorange rate
versus local observation time.
where sρ̇ and ṡρ̇ are respectively called distance equivalent
velocity and distance equivalent acceleration of the attack.
To maintain the victim receiver lock on the spoofer’s signals,
the attack should not exceed a certain distance equivalent
velocity. Two such limiting numbers are reported in the literature, namely, |sρ̇ | ≤ 400 m/s in [29] and |sρ̇ | ≤ 1000 m/s
in [11]. The acceleration to reach the maximum spoofing
velocity is reported to be |ṡρ̇ | ≤ 5 m/s2 . The spoofer acceleration ṡρ̇ can be random, which makes Type II attack quite
general. The distance equivalent velocity can be converted
to the equivalent bias change rate (in s/s) through dividing
the velocity by the speed of light. Fig. 2 illustrates this
attack. The attack on the pseudoranges starts at t = 30 s and
perturbs the pseudoranges gradually with distance equivalent
velocity not exceeding 400 m/s and maximum distance
equivalent random acceleration satisfying |ṡρ̇ | ≤ 5 m/s2 .
The introduced attack models are quite general and can
mathematically capture most attacks on the victim receiver’s
measurements (pseudoranges and pseudorange rates) discussed
in Section I. In another words, Type I and Type II attacks
can be the result of data level spoofing, signal level spoofing,
record-and-replay attack, or a combination of the aformentioned attacks. The main difference between Type I and Type
II attacks is the spoofing speed. The speed of the attack
depends on the capabilities of the spoofer with respect to
manipulating various features of the GPS signals. Indeed,
attacks of different speeds have been reported in the literature
provided earlier in the present section. This work does not
deal with jamming, which disrupts the navigation functionality
completely whereas spoofing misleads it.
In the next section, a dynamical model for the clock bias
and drift is introduced which incorporates these attacks. Based
on this dynamical model, an optimization problem to estimate
these attacks along with the clock bias and drift is proposed.
IV. TSA-AWARE DYNAMICAL M ODEL , TSA R EJECTION
AND M ITIGATION
This section introduces a dynamical model to accommodate
the spoofing attack and a method to estimate the attack.
Afterwards, a procedure for approximately nullifing the effects
of the attack on the clock bias and drift is introduced.
A. Novel TSA-aware Dynamical Model
Modeling of the attack on pseudoranges and pseudorange
rates is motivated by the attack types discussed in the previous
section. These attacks do not alter the position or velocity,
but only the clock bias and clock drift. Our model does
not follow the conventional dynamical model for stationary
receivers which allows the position of the receiver to follow
a random walk model (3). Instead, the known position and
velocity of the victim receiver are exploited jointly. The state
vector contains the clock bias and clock drift, and the attacks
are explicitly modeled on these components, leading to the
following dynamical model:
cbu [l]
cbu [l + 1]
1 ∆t
csb [l]
cwb [l]
=
+
+
0 1
csḃ [l]
cwḃ [l] (6)
cḃu [l]
cḃu [l + 1]
{z
} | {z } | {z } | {z } | {z }
|
xl+1
xl
sl
wl
F
where sb and sḃ are the attacks on clock bias and clock
drift and wb and wḃ are colored Gaussian noise samples with
covariance function defined in [3, Chap. 9]. Here, both sides
are multiplied with c, which is a typically adopted convention.
The state noise covariance matrix, Ql , is particular to the
crystal oscillator of the device.
Similarly, define ρ[l] = [ρ1 [l], . . . , ρN [l]]T and ρ̇[l] =
[ρ̇1[l], .
. . , ρ̇N
[l]]T . The measurement
equation can be as
cbu [l]
ρ[l]
1N ×1 0N ×1
=
+
ρ̇[l]
0N ×1 1N ×1
cḃu [l]
| {z } |
{z
} | {z }
yl
xl
H
kp1 [l] − pu [l]k
cb1 [l]
ρ1 [l]
..
. .
. .
.
. .
cb [l]
kp
[l]
−
p
[l]k
N
u
N
+ ρN [l] .
(v1 [l] − vu [l])T . p1 [l]−pu [l] −
c
ḃ
[l]
[l]
1
kp1 [l]−pu [l]k
ρ̇1
..
.. ..
. .
.
pN [l]−pu [l]
T
ρ̇N [l]
cḃN [l]
(vN [l] − vu [l]) . kpN [l]−pu [l]k
|
{z
} | {z }
l
cl
(7)
Explicit modeling of pu and vu in cl indicates that the
dynamical model benefits from using the stationary victim
receiver’s known position and velocity (the latter is zero). The
measurement noise covariance matrix, Rl , is obtained through
the measurements in the receiver. Detailed explanation of how
to obtain the state and measurement covariance matrices, Ql
and Rl , is provided in Section V. It should be noted that
the state covariance Ql only depends on the victim receiver’s
clock behavior and does not change under spoofing. However,
the measurement covariance matrix, Rl , experiences contraction. The reason is that to ensure that the victim receiver
maintains lock to the fake signals, the spoofer typically applies
a power advantage over the real incoming GPS signals at the
victim receiver’s front end [17].
Comparing (5), (6) and (7), TSAs which do not alter the
position and velocity transfer the attack on pseudoranges and
pseudorange rates directly to clock bias and clock drift. Thus,
it holds that sρ̇ = csb and ṡρ̇ = csḃ .
B. Attack Detection
Let l = k, . . . , k + L − 1 define the time index within the
observation window of length L, where k is the running time
index. The solution to the dynamical model of (6) and (7) is
obtained through stacking L measurements and forming the
following optimization problem:
( k+L−1
1 X
kyl − Hxl − cl k2R−1
(x̂, ŝ) = argmin
l
2
x,s
l=k
) (8)
k+L−1
k+L−1
X
1 X
2
+
kxl+1 − Fxl − sl kQ−1 +
λkDsl k1
l
2
l=k
l=k
kxk2M
T
where
= x Mx, x̂ = [x̂1 , . . . , x̂L ]T are the estimated
states, ŝ = [ŝ1 , . . . , ŝL ]T are the estimated attacks, λ is a
regularization coefficient, and D is an L × 2L total variation
matrix which forms the variation of the signal over time as [42]
−1
0
D=
..
.
0
0
−1
..
.
...
1
0
..
.
0
0
1
..
.
−1
...
...
..
.
0
0
0
..
.
.
1
(9)
The first term is the weighted residuals in the measurement
equation, and the second term is the weighted residuals of the
state equation. The last regularization term promotes sparsity
over the total variation of the estimated attack.
In (8), the clock bias and clock drift are estimated jointly
with the attack. Here, the model of the two introduced attacks
should be considered. In Type I attack, a step attack is
applied over the pseudoranges. The solution to the clock bias
equivalently P
experiences
a step at the attack time. The term
k+L−1
kDsl k1 =
|s
b [l] − sb [l − 1]| + |sḃ [l] − sḃ [l − 1]|
l=k+1
indicates a rise as it tracks the significant differences between
two subsequent time instants. If the magnitude of the estimated
attack in two adjacent times does not change significantly, the
total variation of the attack is close to zero. Otherwise, in the
presence of an attack, the total variation of the attack includes
a spike at the attack time.
In Type II attack, the total variation of the attack does not
show significant changes as the attack magnitude is small at
the beginning and the sparsity is not evident initially. Although
we explained why it is meaningful to expect only few nonzero
entries in the total variation of the attacks in general, this is not
a necessary condition for capturing the attacks during initial
small total variation magnitudes. This means that explicit
modeling of the attacks in (6) and estimation through (8) does
not require the attacks to exhibit sparsity over the total variation. Furthermore, when the bias and bias drift are corrected
using the estimated attack (we will provide one mechanism
in Section IV-C), sparsity over the total variation appears for
subsequent time instants. In these time instants, the attack
appears to be more prominent, and in effect, the low dynamic
behavior of the attack is magnified, a fact that facilitates the
attack detection and will also be verified numerically. This
effect is a direct consequence of (8) and the correction scheme
discussed in the next section.
The optimization problem of (8) boils down to solving a
simple quadratic program. Specifically, the epigraph trick in
convex optimization can be used to transform the `1 -norm into
linear constraints [43]. The observation window L slides for
a lag time Tlag < L, which can be set to Tlag = 1 for realtime operation. The next section details the sliding window
operation of the algorithm, and elaborates on how to use the
solution of (8) in order to provide corrected bias and drift.
C. State Correction
In observation window of length L, the estimated attack ŝ
is used to compensate the impact of the attack on the clock
bias, clock drift, and measurements.
Revisiting the attack model in (6), the bias at time l + 1
depends on the clock bias and clock drift at time l. This dependence successively traces back to the initial time. Therefore,
any attack on the bias that occurred in the past is accumulated
through time. A similar observation is valid for the clock
drift. The clock bias at time l is therefore contaminated by
the cumulative effect of the attack on both the clock bias and
clock drift in the previous times. The correction method takes
into account the previously mentioned effect and modifies the
bias and drift by subtracting the cumulative outcome of the
clock bias and drift attacks as follows:
!
l−1
l
X
X
cb̃u [l]
cb̂u [l]
ŝḃ [l0 ]∆t 1
=
−
ŝb [l0 ] −
ρ̃[l]
ρ[l]
0
0
l =k
l =k
(10)
!
!
!
l
˜
X
ˆ
cḃu [l] = cḃu [l] −
0
ŝḃ [l ] 1
˜
ρ̇[l]
ρ̇[l]
l0 =k
˜
where b̃u and ḃu are respectively the corrected clock bias and
˜ are respectively the corrected pseudorange
clock drift, ρ̃ and ρ̇
and pseudorange rates, and 1 is an all one vector of length
N + 1. In (10), l = 1, . . . , L for the first observation window
(k = 1) and k + L − Tlag ≤ l ≤ k + L − 1 for the observation
windows afterwards. This ensures that the measurements and
states are not doubly corrected. The corrected measurements
are used for solving (8) for the next observation window.
The overall attack detection and modification procedure
is illustrated in Algorithm 1. After the receiver collects L
measurements, problem (8) is solved. Based on the estimated
attack, the clock bias and clock drift are cleaned using (10).
This process is repeated for a sliding window and only the
clock bias and drift of the time instants that have not been
cleaned previously are corrected. In another words, there is
no duplication of modification over the states.
The proposed technique boils down to solving a simple
quadratic program with only few variables and can thus be
performed in real time. For example, efficient implementations
of quadratic programming solvers are readily available in lowlevel programming languages. The implementation of this
technique in GPS receivers and electronic devices is thus
straightforward and does not necessitate creating new libraries.
V. N UMERICAL R ESULTS
We first describe the data collection device and then assess
three representative detection schemes in the literature that fail
to detect the TSA attacks. These attacks mislead the clock bias
and clock drift, while maintaining correct location and velocity
estimates. The performance of our detection and modification
technique over these attacks is illustrated afterwards.
Algorithm 1 : TSA Rejection and Mitigation (TSARM)
1: Set k = 1
2: while True do
3:
Batch yl ∀l = k, . . . , k + L − 1
4:
Construct H, cl , F ∀l = k, . . . , k + L − 1
5:
Compute Ql and Rl (details provided in Section V)
6:
Estimate x̂, ŝ via (8)
ˆ
7:
Assign cb̂u [l] = x̂[m], m = 2l − 1 and cḃu [l] =
x̂[m], m = 2l ∀l = k, . . . , k + L − 1
8:
Assign ŝb [l] = ŝ[m], m = 2l−1 and ŝḃ [l] = ŝ[m], m =
2l ∀l = k, . . . , k + L − 1
ˆ
9:
Modify b̂u [l], ḃu [l], ρ[l] and ρ̇[l] via (10) ∀l =
1, . . . , L for the first window and k + L − Tlag ≤ l ≤
k + L − 1 forthe
windows afterwards
ρ̃[l]
10:
Set yl = ˜
∀l = k, . . . , k + L − 1
ρ̇[l]
11:
Output tUTC
[l] = tR [l] − b̃u [l] − ∆tU T C ∀l =
R
k, . . . , k + L − 1 to the user for time stamping
12:
Slide the observation window by setting k = k + Tlag
13: end while
A. GPS Data Collection Device
A set of real GPS signals has been recorded with a Google
Nexus 9 Tablet at the University of Texas at San Antonio on
June, 1, 2017.3 The ground truth of the position is obtained
through taking the median of the WLS position estimates for
a stationary device. This device has been recently equipped
with a GPS chipset that provides raw GPS measurements. An
android application, called GNSS Logger, has been released
along with the post-processing MATLAB codes by the Google
Android location team [38].
Of interest here are the two classes of the
Android.location package. The GnssClock4 provides the
GPS receiver clock properties and the GnssMeasurement5
provides the measurements from the GPS signals
both with sub-nanosecond accuracies. To obtain the
pseudorange measurements, the transmission time is
subtracted from the time of reception. The function
getReceivedSvTimeNanos() provides the transmission
time of the signal which is with respect to the current GPS
week (Saturday-Sunday midnight). The signal reception
time is available using the function getTimeNanos(). To
translate the receiver’s time to the GPS time (and GPS
time of week), the package provides the difference between
the device clock time and GPS time through the function
getFullBiasNanos().
The receiver clock’s covariance matrix, Ql , is dependent
on the statistics of the device clock oscillator. The following
3 Our data is available at: https://github.com/Alikhalaj2006/UTSA
GPS DATA.git
4 https://developer.android.com/reference/android/location/
GnssClock.html (accessed Feb. 20, 2017).
5 https://developer.android.com/reference/android/location/
GnssMeasurement.html (accessed Feb. 20, 2017).
105
Bias (m)
1
0
EKF Normal
EKF Spoofed
Anti-Spoofing PF
-1
-2
0
50
100
150
200
250
300
350
400
Time(s)
Drift (m/s)
-60
EKF Normal
EKF Spoofed
-62
-64
-66
0
50
100
150
200
250
300
350
400
Time(s)
Fig. 3. The effect of Type II attack on the EKF and the anti-spoofing
particle filter [27] on (a) clock bias and (b) clock drift. The attack started
at t = 30 s. Panel (b) does not include the drift.
No attack
500
0
Consistent attack Inconsistent attack
-500
0
50
1000
100
150
200
100
150
200
150
200
(a)
0
Attack started
-1000
0
50
(b)
500
0
Attack started
-500
0
50
100
(c)
Time (s)
Fig. 4. Performance of hypothesis testing based on statistic (V-B) [28]
under Type I attack for different false alarm probabilities: (a) No attack,
(b) inconsistent attack, (c) consistent attack.
model is typically adopted:
3
Ql =
c2 σb2 ∆t + c2 σḃ2 ∆t3
2
c2 σḃ2 ∆t2
2
c2 σḃ2 ∆t2
c2 σḃ2 ∆t
!
(11)
where σb2 = h20 and σḃ2 = 2π 2 h−2 ; and we select h0 =
8 × 10−19 and h−2 = 2 × 10−20 [44, Chap. 9]. For calculating
the measurement covariance matrix, Rl , the uncertainty of the
pseuodrange and pseudorange rates are used. These uncertainties are available from the device together with the respective
measurements.5 In the experiments, we set λ = 5 × 10−10 ,
because the distance magnitudes are in tens of thousands of
meters. The estimated clock bias and drift through EKF in
normal operation is considered as the ground truth for the
subsequent analysis. In what follows, reported times are local.
B. Failure of Prior Work in Detecting Consistent Attacks
This section demonstrates that three relevant approaches
from Table I may fail to detect consistent attacks, that is,
attacks where sρ is the integral of sρ̇ in (4).
The performance of the EKF and the anti-spoofing particle
filter of [27] subject to a Type II attack is reported first. The
perturbations over GPS measurements are the same as in Fig. 2
and are used as input to the EKF and the particle filter. The
attack starts at t = 30 s. Fig. 3 depicts the effect of attack on
the clock bias and drift. The EKF on the dynamical model in
(6) and (7) blindly follows the attack after a short settling time.
The anti-spoofing particle filter only estimates the clock bias
and assumes the clock drift is known from WLS. Similarly to
the EKF, the particle filter is not able to detect the consistent
spoofing attack. The maximum difference between the receiver
estimated position obtained from the EKF on (3) under Type
II attack and under normal operation is xdiff = 67 m, ydiff =
112 m, and zdiff = 71 m. The position estimate has thus not
been considerably altered by the attack.
The third approach to be evaluated has been proposed
in [28] and monitors the statistics of the receiver clock,
as a typical spoofing detection technique [33]. Considering that off-the-shelf GPS receivers compute the bias
at regular ∆t intervals, a particular approach is to estimate the GPS time after k time epochs, and confirm that the time elapsed is indeed k∆t [28]. To this
hend, the following statistic can bePformulated: iD(k) =
k
ˆ
S
S
tGP
(k) − tGP
(1) − (k − 1)∆t − k0 =1 ḃ[k 0 ]∆t c. The
R
R
test statistic D is normally distributed with mean zero when
there is no attack and may have nonzero mean depending on
the attack, as will be demonstrated shortly. Its variance needs
to be estimated from a few samples under normal operation.
The detection procedure relies on statistical hypothesis testing.
For this, a false alarm probability, PF A , is defined. Each PF A
corresponds to a threshold γ to which D(k) is compared
against [45, Chap. 6]. If |D(k)| ≥ γ, the receiver is considered
to be under attack.
The result of this method is shown in Fig. 4 for different
false alarm probabilities. Fig. 4 (a) depicts D(k) when the
system is not under attack. The time signature lies between
the thresholds only for low false alarm probabilities. The
system can detect the attack in case of an inconsistent Type I
attack, in which sρ , is not the integration of perturbations over
pseudorange rates, sρ̇ , and only pseudoranges are attacked.
Fig. 4 (b) shows that the attack is detected right away.
However, for smart attacks, where the spoofer maintains the
consistency between the pseudorange and pseudorange rates,
Fig. 4 (c) illustrates that the signature D(k) fails to detect the
attack. This example shows that the statistical behavior of the
clock can remain untouched under smart spoofing attacks. In
addition, even if an attack is detected, the previous methods
cannot provide an estimate of the attack.
C. Spoofing Detection on Type I Attack
Fig. 5 shows the result of solving (8) using the GPS
measurements perturbed by the Type I attack of Fig. 1. The
spoofer has the capability to attack the signal in a very short
time so that the clock bias experiences a jump at t = 30 s. The
estimated total variation of bias attack renders a spike right at
the attack time. The modification procedure of (10) corrects
the clock bias using the estimated attack.
Bias Spoofed
Bias Normal
10
15
20
25
(a)
30
5000
0
-5000
-10000
35
40
45
Attack started
1
5
10
15
20
1
25
(b)
30
35
40
45
Attack started
-1
5
10
15
20
5000
25
(c)
30
35
40
45
50
Attack started
-5000
1
5
10
15
20
25
-10000
1
5
10
15
20
30
35
40
45
50
-200000
12 Normal
12 Attacked
-300000
5 Normal
5 Attacked
150
200
250
300
350
400
(a)
450
6 Normal
6 Attacked
Attack started
2 Normal
2 Attacked
12 Normal
12 Attacked
5 Normal
-500
5 Attacked
-1000
0
50
100
50
40
45
50
40
45
50
40
45
50
-200
1
5
10
15
20
25
30
35
(b)
1
Attack started
0
-1
1
5
10
15
20
25
30
35
5000
Attack started
Bias Modified
Bias Normal
0
-5000
5
10
15
20
25
30
35
Fig. 7. The result of attack detection and modification over Type II attack
for t = 1 s through t = 50 s. The attack started at t = 30 s. From
top to bottom: (a) Normal clock bias (blue) and spoofed bias (red), (b)
estimated bias attack ŝb , (c) total variation of the estimated bias attack,
and (d) true bias (blue) and modified bias (magenta).
150
200
250
300
350
400
(b)
Time (s)
Fig. 6. Comparison of (a) normal pseudorange change (ρ(k) − ρ(1))
and spoofed pseudoranges change (ρs (k) − ρs (1)), and (b) normal
pseudorange rates (ρ̇) and spoofed pseudorange rates (ρ˙s ) under Type
II attack for some of the visible satellites. The attack started at t = 30 s.
Bias (m)
2 Normal
2 Attacked
0
45
Attack started
1
MBias (m) Drift TV (m/s) Bias TV(m)
0
-100000
100
40
6 Normal
6 Attacked
Attack started
50
35
(d)
Time (s)
Fig. 5. The result of attack detection and modification over a Type I
attack that started at t = 30 s. From top to bottom: (a) Normal clock bias
(blue) and spoofed bias (red), (b) total variation of the estimated bias
attack ŝb , (c) total variation of the estimated drift attack ŝḃ , and (d) true
bias (blue) and modified bias (magenta).
0
30
0
(d)
Time (s)
100000
25
(a)
200
(c)
Bias Modified
Bias Normal
0
Attack started
Bias Spoofed
Bias Normal
0
50
0
1
10000
50
Bias TV(m)
5
Drift TV (m/s)
-10000
1
MBias (m) Drift TV (m/s) Bias TV (m)
Bias (m)
Attack started
0
MBias (m)
Bias (m)
10000
10000
-10000
11
Attack started
Bias Spoofed
Bias Normal
0
15
20
25
30
40
45
50
55
60
40
45
50
55
60
40
45
50
55
60
Attack started
0
-2000
11
35
(a)
2000
15
20
25
30
2
35
(b)
Attack started
0
-2
11
15
20
25
30
5000
35
(c)
Attack started
Bias Modified
Bias Normal
0
-5000
11
15
20
25
30
35
40
45
50
55
60
(d)
Time (s)
D. Spoofing Detection on Type II Attack
Fig. 8. The result of attack detection and modification over Type II
attack for t = 11 s through t = 60 s. From top to bottom: (a) Normal
clock bias (blue) and spoofed bias (red), (b) estimated bias attack ŝb , (c)
total variation of the estimated bias attack, and (d) true bias (blue) and
modified bias (magenta).
The impact of Type II attack on the pseudoranges and
pseuodrange rates is shown in Fig. 6. Specifically, Fig. 6
(a) illustrates the normal and spoofed pseudorange changes
with respect to their initial value at t = 0 s for some of the
visible satellites in the receiver’s view. Fig. 6 (b) depicts the
corresponding pseudorange rates. The tag at the end of each
line indicates the satellite ID and whether the pseudorange
(or pseudorange rate) corresponds to normal operation or
operation under attack. The spoofed pseudoranges diverge
quadratically starting at t = 30 s following the Type II attack.
For the Type II attack, Algorithm 1 is implemented for an
sliding window with L = 50 s with Tlag = 10 s. Fig. 7
shows the attacked clock bias starting at t = 30 s. Since
the attack magnitude is small at initial times of the spoofing,
neither the estimated attack ŝb nor the total variation do not
show significant values. The procedure of sliding window is
to correct the current clock bias and clock drift for all the
times that have not been modified previously. Hence, at the
first run the estimates of the whole window are modified.
Fig. 8 shows the estimated attack and its corresponding total
variation after one Tlag . As is obvious from the figure, the
modification of the previous clock biases transforms the low
dynamic behavior of the spoofer to a large jump at t = 50 s
which facilitates the detection of attack through the total
variation component in (8). The clock bias and drift have been
modified for the previous time instants and need to be cleaned
only for t = 50 s − 60 s.
1200
In the present work, the set of GPS signals are obtained from
an actual GPS receiver in a real environment, but the attacks
are simulated based on the characteristics of real spoofers
reported in the literature. Experimentation on the behavior
of the proposed detection and mitigation approach under real
spoofing scenarios is the subject of future research.
RMSE (m)
1000
800
600
R EFERENCES
400
200
10
20
30
40
50
60
70
L
Fig. 9. The RMSE of TSARM for various values of L and Tlag .
E. Analysis of the Results
Let K be the total length of the observation time (in this
experiment, K = 386). The q
root mean square error (RMSE)
PK−1
2
is introduced: RMSE = Kc
k=0 (b̃u [k] − b̌u [k]) , which
shows the average error between the clock bias that is output
from the spoofing detection technique, b̃u , and the estimated
clock bias from EKF under normal operation, b̌u , which is
considered as the ground truth. Comparing the results of the
estimated spoofed bias from the EKF and the normal bias
shows that RMSEEKF = 3882 m. This error for the antispoofing particle filter is RMSEPF = 3785 m. Having applied
TSARM, the clock bias has been modified with a maximum
error of RMSETSARM = 258 m. Fig. 9 illustrates the RMSE
of TSARM for a range of values for the window size, L,
and the lag time, Tlag . When the observation window is
smaller, fewer measurements are used for state estimation.
On the other hand, when L exceeds 40 s, the number of
states to be estimated grows although more measurements
are employed for estimation. The numerical results illustrate
that (6) models the clock bias and drift attacks effectively,
which are subsequently estimated using (8) and corrected
through (10).
VI. C ONCLUDING R EMARKS AND F UTURE W ORK
This work discussed the research issue of time synchronization attacks on devices that rely on GPS for time tagging their measurements. Two principal types of attacks are
discussed and a dynamical model that specifically models
these attacks is introduced. The attack detection technique
solves an optimization problem to estimate the attacks on
the clock bias and clock drift. The spoofer manipulated clock
bias and drift are corrected using the estimated attacks. The
proposed method detects the behavior of spoofer even if the
measurements integrity is preserved. The numerical results
demonstrate that the attack can be largely rejected, and the
bias can be estimated within 0.86 µs of its true value, which
lies within the standardized accuracy in PMU and CDMA
applications. The proposed method can be implemented for
real-time operation.
[1] “Energy.gov, Office of Electricity Delivery & Energy Reliability,” http:
//energy.gov/oe/services/technology-development/smart-grid, accessed:
2016-08-08.
[2] “Gps.gov,
Official U.S. government information about
the Global Positioning System (GPS) and related topics,”
http://www.gps.gov/applications/timing/.
[3] B. W. Parkinson, J. J. Spilker, P. Axelrad, and P. Enge, Global Positioning System: Theory and Applications. American Institute of Aeronautics
and Astronautics, 1996, vol. I.
[4] ——, Global Positioning System: Theory and Applications. American
Institute of Aeronautics and Astronautics, 1996, vol. II.
[5] P. Misra and P. Enge, Global Positioning System: Signals, Measurements,
and Performance, 2nd ed. Ganga-Jamuna Press, Lincoln MA, 2006.
[6] I. Yaesh and U. Shaked, “Neuro-adaptive hinf inity estimation and its
application to improved tracking in GPS receivers,” IEEE Trans. Ind.
Electron., vol. 56, no. 3, pp. 642–647, Mar. 2009.
[7] X. Li and Q. Xu, “A reliable fusion positioning strategy for land vehicles
in GPS-denied environments based on low-cost sensors,” IEEE Trans.
Ind. Electron., vol. 64, pp. 3205–3215, Apr. 2017.
[8] “Electric sector failure scenarios and impact analyses - version 3.0,”
Electric Power Research Institute, Tech. Rep., Dec. 2015.
[9] D. Schmidt, K. Radke, S. Camtepe, E. Foo, and M. Ren, “A survey
and analysis of the GNSS spoofing threat and countermeasures,” ACM
Computer Survey, vol. 48, no. 4, pp. 64:1–64:31, May 2016.
[10] B. Moussa, M. Debbabi, and C. Assi, “Security assessment of time
synchronization mechanisms for the smart grid,” IEEE Commun. Surveys
Tut., vol. 18, no. 3, pp. 1952–1973, thirdquarter 2016.
[11] D. P. Shepard, T. E. Humphreys, and A. A. Fansler, “Evaluation of the
vulnerability of phasor measurement units to GPS spoofing attacks,” Int.
J. Crit. Infrastruct. Protect., vol. 5, pp. 146–153, Dec. 2012.
[12] Z. Zhang, S. Gong, A. D. Dimitrovski, and H. Li, “Time synchronization
attack in smart grid: Impact and analysis,” IEEE Trans. Smart Grid,
vol. 4, no. 1, pp. 87–98, Mar. 2013.
[13] X. Jiang, J. Zhang, B. J. Harding, J. J. Makela, and A. D. DominguezGarcı́a, “Spoofing GPS receiver clock offset of phasor measurement
units,” IEEE Trans. Power Systems, vol. 28, pp. 3253–3262, Aug. 2013.
[14] P. Risbud, N. Gatsis, and A. Taha, “Assessing power system state
estimation accuracy with GPS-spoofed PMU measurements,” in IEEE
Trans. Smart Grid, to be published.
[15] T. Nighswander, B. Ledvina, J. Diamond, R. Brumley, and D. Brumley,
“GPS software attacks,” in Proc. of the ACM Conf. on Comput. and
Commun. Security, Oct. 2012, pp. 450–461.
[16] N. O. Tippenhauer, C. Pöpper, K. B. Rasmussen, and S. Čapkun, “On
the requirements for successful GPS spoofing attacks,” in Proc. of the
18th ACM Conf. on Comput. and Commun. Security, Oct. 2011, pp.
75–86.
[17] K. D. Wesson, J. N. Gross, T. E. Humphreys, and B. L. Evans, “GNSS
signal authentication via power and distortion monitoring,” IEEE Trans.
on Aeros. and Elect. Systems, vol. PP, no. 99, pp. 1–1, 2017.
[18] M. L. Psiaki and T. E. Humphreys, “GNSS spoofing and detection,”
Proc. of the IEEE, vol. 104, no. 6, pp. 1258–1270, June 2016.
[19] P. Papadimitratos and A. Jovanovic, “GNSS-based Positioning: Attacks
and Countermeasures,” in Proc. of the IEEE Military Commun. Conf.,
San Diego, CA, USA, Nov. 2008, pp. 1–7.
[20] Q. Zeng, H. Li, and L. Qian, “GPS spoofing attack on time synchronization in wireless networks and detection scheme design,” in IEEE
Military Communications Conference, Oct. 2012, pp. 1–5.
[21] Y. Fan, Z. Zhang, M. Trinkle, A. D. Dimitrovski, J. B. Song, and
H. Li, “A cross-layer defense mechanism against GPS spoofing attacks
on PMUs in smart grids,” IEEE Trans. Smart Grid, vol. 6, no. 6, pp.
2659–2668, Nov. 2015.
[22] A. Ranganathan, H. Ólafsdóttir, and S. Capkun, “SPREE: A spoofing
resistant GPS receiver,” in Proc. of the 22Nd Annual Int. Conf. on Mobile
Comput. and Netw., 2016, pp. 348–360.
[23] D. Chou, L. Heng, and G. X. Gao, “Robust GPS-based timing for phasor
measurement units: A position-information-aided approach,” in Proc.
of the 27th Int. Tech. Meeting of The Sat. Division of the Institute of
Navigation, Sept. 2014, pp. 1261 – 1269.
[24] Y. Ng and G. X. Gao, “Advanced multi-receiver position-informationaided vector tracking for robust GPS time transfer to PMUs,” in Proc.
of the Institute of Navigation GNSS+ Conf. (ION GNSS+ 2015), Sept.
2015, pp. 3443 – 3448.
[25] D.-Y. Yu, A. Ranganathan, T. Locher, S. Čapkun, and D. Basin, “Short
paper: Detection of GPS spoofing attacks in power grids,” in Proc. of
the ACM Conf. on Security and Privacy in Wireless Mobile Netw., 2014,
pp. 99–104.
[26] K. Jansen, N. O. Tippenhauer, and C. Pöpper, “Multi-receiver GPS
spoofing detection: Error models and realization,” in Proc. of the 32nd
Annual Conf. on Comput. Security Appl., 2016, pp. 237–250.
[27] S. Han, D. Luo, W. Meng, and C. Li, “A novel anti-spoofing method
based on particle filter for GNSS,” in Proc. of the IEEE Int. Conf. on
Commun. (ICC), June 2014, pp. 5413–5418.
[28] F. Zhu, A. Youssef, and W. Hamouda, “Detection techniques for datalevel spoofing in GPS-based phasor measurement units,” in Proc. of the
2016 Int. Conf. on Selected Topics in Mobile Wireless Netw. (MoWNeT),
Apr. 2016, pp. 1–8.
[29] D. P. Shepard and T. E. Humphreys, “Characterization of receiver
response to a spoofing attacks,” in Proc. of the 24th Int. Tech. Meeting
of The Sat. Division of the Institute of Navigation (ION GNSS 2011),
Portland, OR, Sept. 2011, pp. 2608 – 2618.
[30] T. E. Humphreys, B. M. Ledvina, M. L. Psiaki, B. W. O’Hanlon, and
P. M. Kintner, “Assessing the spoofing threat: Development of a portable
GPS civilian spoofer,” in Proc. of the 21st Int. Tech. Meeting of the Sat.
Division of The Institute of Navigation (ION GNSS 2008), Savannah,
GA, Sept. 2008, pp. 2314–2325.
[31] B. Motella, M. Pini, M. Fantino, P. Mulassano, M. Nicola, J. FortunyGuasch, M. Wildemeersch, and D. Symeonidis, “Performance assessment of low cost GPS receivers under civilian spoofing attacks,” in 5th
ESA Workshop on Sat. Nav. Tech. and European Workshop on GNSS
Signals and Signal Proc., Dec. 2010, pp. 1–8.
[32] P. Teunissen, “Quality control in integrated navigation systems,” IEEE
Aeros. and Elect. Sys. Magazine, vol. 5, pp. 35 – 41, July 1990.
[33] L. Heng, J. J. Makela, A. D. Dominguez-Garcı́a, R. B. Bobba, W. H.
Sanders, and G. X. Gao, “Reliable GPS-based timing for power systems:
A multi-layered multi-receiver architecture,” in Power and Energy Conf.
at Illinois (PECI), Feb. 2014, pp. 1–7.
[34] L. Heng, D. B. Work, and G. X. Gao, “GPS signal authentication from
cooperative peers,” IEEE Trans. on Intell. Transportation Syst., vol. 16,
no. 4, pp. 1794–1805, Aug. 2015.
[35] D. Radin, P. F. Swaszek, K. C. Seals, and R. J. Hartnett, “GNSS
spoof detection based on pseudoranges from multiple receivers,” in
Proceedings of the 2015 International Technical Meeting of The Institute
of Navigation, Jan. 2015, pp. 657–671.
[36] C. Masreliez and R. Martin, “Robust Bayesian estimation for the linear
model and robustifying the Kalman filter,” IEEE Trans. Autom. Control,
vol. 22, no. 3, pp. 361–371, June 1977.
[37] S. Farahmand, G. B. Giannakis, and D. Angelosante, “Doubly robust
smoothing of dynamical processes via outlier sparsity constraints,” IEEE
Trans. Signal Process., vol. 59, pp. 4529–4543, Oct. 2011.
[38] “Android GNSS,” https://developer.android.com/guide/topics/sensors/
gnss.html, accessed: 2017-02-20.
[39] “IEEE standard for synchrophasor measurements for power systems,”
IEEE Std C37.118.1-2011 (Revision of IEEE Std C37.118-2005), pp.
1–61, Dec. 2011.
[40] “Model 1088b GPS satellite clock (40 ns),” http://www.arbiter.com/
catalog/product/model-1088b-gps-satellite-precision-time-clock-40ns.
php, accessed: 2017-02-20.
[41] Z. Zhang, S. Gong, A. D. Dimitrovski, and H. Li, “Time synchronization
attack in smart grid: Impact and analysis,” IEEE Trans. on Smart Grid,
vol. 4, no. 1, pp. 87–98, March 2013.
[42] F. I. Karahanoglu, . Bayram, and D. V. D. Ville, “A signal processing approach to generalized 1-D total variation,” IEEE Trans. Signal Process.,
vol. 59, no. 11, pp. 5265–5274, Nov. 2011.
[43] S. Boyd and L. Vandenberghe, Convex Optimization.
Cambridge
University Press, 2004.
[44] R. G. Brown and P. Y. C. Hwang, Introduction to random signals and
applied Kalman filtering: with MATLAB exercises and solutions; 3rd ed.
New York, NY: Wiley, 1997.
[45] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume II:
Detection Theory. Prentice-Hall Inc, 1993.
Ali Khalajmehrabadi (S’16) received the B.Sc.
degree from the Babol Noshirvani University of
Technology, Iran, in 2010, and the M.Sc. degree
from University Technology Malaysia, Malaysia,
in 2012, where he was awarded the Best Graduate Student Award. He is currently pursuing the
Ph.D. degree with the Department of Electrical
and Computer Engineering, University of Texas
at San Antonio. His research interests include indoor localization and navigation systems, collaborative localization, and global navigation satellite system. He is a Student Member of the Institute of Navigation and
the IEEE.
Nikolaos Gatsis (S’04-M’05) received the
diploma (with Hons.) degree in electrical and
computer engineering from the University of Patras, Greece, in 2005, and the M.Sc. degree in
electrical engineering and the Ph.D. degree in
electrical engineering with minor in mathematics
from the University of Minnesota, in 2010 and
2012, respectively. He is currently an Assistant
Professor with the Department of Electrical and
Computer Engineering, University of Texas at
San Antonio. His research interests lie in the
areas of smart power grids, communication networks, and cyberphysical
systems, with an emphasis on optimal resource management and statistical signal processing. He has co-organized symposia in the area of
smart grids in IEEE GlobalSIP 2015 and IEEE GlobalSIP 2016. He has
also served as a co-guest editor for a special issue of the IEEE Journal
on Selected Topics in Signal Processing on critical infrastructures.
David Akopian (M’02-SM’04) received the
Ph.D. degree in electrical engineering in 1997.
He is a Professor with the University of Texas
at San Antonio. He was a Senior Research Engineer and a Specialist with Nokia Corporation
from 1999 to 2003. From 1993 to 1999, he
was a Researcher and an Instructor with the
Tampere University of Technology, Finland. He
has authored and coauthored over 30 patents
and 140 publications. His current research interests include digital signal processing algorithms
for communication and navigation receivers, positioning, dedicated
hardware architectures and platforms for software defined radio and
communication technologies for healthcare applications. He served in
organizing and program committees of many IEEE conferences and
co-chairs annual SPIE Multimedia on Mobile Devices conferences. His
research has been supported by the National Science Foundation,
National Institutes of Health, USAF, U.S. Navy, and Texas foundations.
Ahmad F. Taha (S’07–M’15) received the B.E.
and Ph.D. degrees in Electrical and Computer
Engineering from the American University of
Beirut, Lebanon in 2011 and Purdue University,
West Lafayette, Indiana in 2015. In Summer
2010, Summer 2014, and Spring 2015 he was
a visiting scholar at MIT, University of Toronto,
and Argonne National Laboratory. Currently he
is an assistant professor with the Department
of Electrical and Computer Engineering at The
University of Texas, San Antonio. Dr. Taha is interested in understanding how complex cyber-physical systems operate,
behave, and misbehave. His research focus includes optimization and
control of power system, observer design and dynamic state estimation,
and cyber-security.
| 3cs.SY
|
1
Information capacity of direct detection optical
transmission systems
arXiv:1711.01888v1 [cs.IT] 6 Nov 2017
Antonio Mecozzi, Fellow, OSA, Fellow, IEEE and Mark Shtaif, Fellow, OSA, Fellow, IEEE
Abstract—We show that the spectral efficiency of a direct
detection transmission system is at most 1 bit/s/Hz less than the
spectral efficiency of a system employing coherent detection with
the same modulation format. Correspondingly, the capacity per
complex degree of freedom in systems using direct detection is
lower by at most 1 bit.
(a)
Tx
Noise
ݐ ܧ
ܧ ݐ
propagation
channel
ݐ ܧ
+
Rx
Index Terms—Channel capacity, Optical detection, Modulation.
I. I NTRODUCTION
R
ECENTLY, the field of optical communications is witnessing a revival of interest in direct detection receivers,
which are often viewed as a promising low-cost alternative to
their expensive coherent counterparts [1]–[12]. This process
stimulates an interesting fundamental question to whose answer the present paper is dedicated: What is the difference
between the information capacity of a direct detection system
and that of a system using coherent detection?
In order to answer this question we consider the channel
schematic illustrated in Fig. 1a, which consists of a transmitter
that is capable of generating any desirable complex waveform
whose spectrum is contained within a bandwidth B, a noise
source of arbitrary spectrum and statistics, a propagation
channel, and a receiver. Although the linearity of the channel
and the additivity of the noise are immaterial to our analysis,
we will assume these properties in the beginning, while
postponing the generalization of our discussion to Sec. IV. The
direct detection receiver in our definition is one that recovers
the communicated data from the intensity (i.e. absolute square
value) of the received electric field, while using a single photodiode. As illustrated in Fig. 1b, it consists of a square optical
band-pass filter of width B that rejects out of band noise,
a photo-diode whose output current is proportional to the
received optical intensity1 , and a processing unit that recovers
the information. The benchmark to which we compare the
direct detection receiver, is the coherent receiver, in whose case
the complex-valued received optical field is reconstructed.
Intuitively, it is tempting to conclude that since a direct
detection receiver ignores one of the two degrees of freedom
that are necessary for uniquely characterizing the electric field,
its capacity should be close to half of the capacity of a coherent
system. Surprisingly, this notion turns out to be incorrect, and
as we show in this paper, the capacity per complex degree of
freedom in systems using direct detection is lower by not more
A. Mecozzi is with the Department of Physical and Chemical Sciences,
University of L’Aquila, L’Aquila 67100, Italy. M. Shtaif is with the Department of Physical Electronics, Tel Aviv University, Tel Aviv 69978, Israel.
1 For simplicity, in what follows we will assume that the proportionality
coefficient is 1.
(b)
ݐ ܧ
OBPF
ܤ
PD
Processing and
Data Recovery
Fig. 1. (a) The setup considered in this paper. It consists of an I/Q transmitter
(Tx) that can generate any complex waveform whose spectrum is contained
in a bandwidth B, a stationary noise source of arbitrary spectrum and
distribution, a propagation channel, and a receiver (Rx). (b) The schematic
of a direct detection receiver. The incoming optical field is filtered with an
optical band-pass filter (OBPF) to reject out of band noise and square law
detected (without any manipulation of the field) by a single photodiode (PD).
The photodiode and subsequent electronics are assumed to have bandwidth of
at least 2B so as to accommodate the bandwidth of the intensity waveform.
than 1 bit than that of fully coherent systems. Correspondingly,
the loss in terms of spectral efficiency is limited to be no
greater than 1 bit/s/Hz.
Throughout the paper, in order to simplify the notation, we
will assume that the transmitted field is scalar. This assumption
does not affect the generality of the results, as the transmission
of orthogonal polarization components through linear channels
is independent.
II. R ELATION TO PRIOR WORK
Our result, stating that a direct detection channel is characterized by almost the same capacity as a coherent channel,
requires some clarification in view of its being in an apparent
contradiction to prior work, where the capacity of a seemingly
similar channel was found to be lower by approximately a
factor of two. This work consists of Ref. [13] published by
the authors of the current manuscript, as well as a number of
more recent works, the most relevant of which are contained
in Refs. [14]–[17]. In order to avoid confusion, we will adopt
the terminology of [16] and refer to the channels studied
in those papers as various flavors of the intensity channel,
whereas the term direct detection channel will only be used
in reference to the channel that we study here. The reason for
this apparent contradiction boils down to the fact that all the
versions of the intensity channel assume that the information
2
is encoded at a given rate B directly onto the intensity of
the transmitted optical signal and then it is recovered by
sampling the received signal’s intensity at exactly the same
rate. In all cases, the channel is assumed to be memoryless and
the optical bandwidth (and hence also the spectral efficiency)
do not play any role. The studies in [14]–[17] can be given
practical justification when considering very short-reach, or
very old optical systems that used a low-coherence optical
source (such as a light-emitting diode). Indeed, with such
sources the optical phase is far too noisy to be used for
transmitting information and the source linewidth is so much
greater than the modulation bandwidth that relating to spectral
efficiency in the modern sense is not meaningful.
In contrast to the above, the direct detection channel is
inspired by modern fiber-optic communications systems, the
vast majority of which relies on a highly coherent laser
source – one whose linewidth is substantially smaller than
the bandwidth of the modulation. This is the reason for our
assumption that the transmitter in the direct detection channel
can encode information into any complex waveform with the
only constraint being that its spectrum is contained in some
bandwidth B. In addition, since the process of photo-detection
involves frequency doubling, the spectrum of the measured
intensity is contained in a two-sided bandwidth of 2B, and
hence sampling at the rate of 2B is imperative in order to
extract the information present in the photo-detection current.
In order to further clarify the difference between the direct
detection channel and the intensity channel, we denote the
complex-valued field received after optical filtering, by E(t).
Since the spectrum of E(t) is contained in a bandwidth B, it
can be rigorously expressed as
E(t) =
∞
X
Ek sinc [π(Bt − k)] ,
(1)
k=−∞
where sinc(x) = sin(x)/x, and where Ek = E(t = k/B) are
the complex-field samples carrying the transmitted information. The detected photo-current is proportional to I = |E(t)|2 .
If this photocurrent were to be sampled at a rate of B, as in
[13]–[17], the samples at t = n/B would have been equal to
In = |En |2 , and the phase information would have been lost.
In this case the drop in the amount of extracted information
(and hence the capacity) would have been roughly a factor of
2, similarly to the results obtained in [13]–[17]. Yet, in the
direct detection channel the sampling of the photo-current is
done at a rate of 2B, so that the middle-point samples that
are taken at t = (n + 0.5)/B are also obtained. These middle
samples are given by
∞
X
1
In+1/2 =
sinc π n − m +
2
k,m=−∞
1
∗
×sinc π n − k +
Em
Ek ,
(2)
2
and they are clearly affected by the phase differences between
the various complex-field samples. In fact, as our final result
indicates, knowledge of all intensity samples (In and In+1/2
for all n) allows one to collect almost all of the information
contained in the complex optical field.
We note that the idea of increasing the information rate by
sampling the received analog signal at a rate that is higher than
B, has been considered previously [18]. This is a natural idea
in cases where the receiver contains a nonlinear element that
expands the analog bandwidth of the received waveform so
that sampling at B is no longer sufficient in order to collect
all the information from the analog waveform. In our case
the nonlinearity is that of square-law detection and it expands
the analog bandwidth by exactly a factor of 2. Hence, unlike
the case studied in [18], where the doubling of the sampling
rate produced only minuscule benefits, here sampling at 2B
is sufficient in order to extract all the information contained
in the analog intensity waveform, and there is no benefit in
increasing the sampling rate farther.
Finally, it is instructive to relate to the most widespread
example, where the additive noise of Fig. 1a is white Gaussian.
In this case, our theory implies that the capacity of the
direct detection channel is within 1 bit of log(1 + SNR)
[19], where the SNR is the ratio between the average power
of the information carrying signal and the variance of the
filtered noise (summed over both quadratures). Conversely, as
demonstrated in [13] and [17], the capacity of the intensity
channel (i.e. one that samples the received intensity at the rate
B) in the limit of high SNR, is 0.5 log(SNR/2) — roughly
half of the direct detection channel’s capacity.2
III. T HE INFORMATION CAPACITY OF A DIRECT
DETECTION RECEIVER
A. The definition of distinguishable waveforms
Usually, in engineering practice, two waveforms E1 (t) and
E2 (t) are said to be distinguishable when the energy of the
difference between them is greater than 0,
Z ∞
|E1 (t) − E2 (t)|2 dt > 0.
(3)
−∞
In the context of optical communications this definition is too
restrictive, because in all cases of interest, optical receivers (including coherent receivers) are unable to distinguish between
waveforms that differ only by a constant (i.e. time independent) phase.3 Owing to this reality, we define waveforms to be
distinguishable only when they can be told apart by an ideal
coherent receiver. Formally, this means that E1 (t) and E2 (t)
are distinguishable only when they remain distinguishable
according to Eq. (3), even if one of them is rotated in the
complex plane by some arbitrary constant phase, i.e.
Z ∞
|E1 (t) − eiθ E2 (t)|2 dt > 0.
(4)
−∞
2 In Ref. [13], the capacity in the high SNR limit is written as
0.5 log(SNR/4), but the difference is only in the SNR definition, which
relates to the noise variance in one quadrature.
3 In order to overcome this limitation, the transmitter and reciever would
have to share an exact time-reference (on the scale of a fraction of a single
optical cycle!). In principle, this can be achieved by means of an atomic clock.
However, the costs of such a solution on the one hand, and the minuscule
potential benefit in terms of information rate, on the other hand, ensure that
this solution isn’t deployed.
3
where Ω =
2πB
M .
B
Fk =
M
Z
The Fourier coefficients are given by
M/B
E(t)e
ikΩt
0
M −1
2πkn
1 X
dt =
En ei M ,
M n=0
(6)
with En = E(t = n/B), and where the second equality in
(6) takes advantage of the relation between the Fourier series
coefficients and the discrete Fourier transform in band-limited
periodic signals. We now assign
A(Z) =
M
−1
X
Fk Z k
(7)
k=0
−1
to be the Z-transform of the Fourier
coefficients {Fk }M
k=0 .
Clearly, E(t) = A exp(−iΩt) , and hence special attention
needs to be paid to the cases where the value of Z is on
the unit circle. When FM −1 6= 0, A(Z) is an M − 1 degree
polynomial which admits M −1 zeros, and it can be expressed
as
A(Z) = C
M
−1
Y
F0
.
(Z − Zk ), with C = QM −1
k=0 (−Zk )
k=0
(8)
Consider now the functions
Uk (Z) =
Fig. 2. An example of 8 different waveforms in the case of M = 4 (so that
2M −1 = 8), all band-limited to [0, B] (with B = 1/T ) and all having the
intensity shown in (a). In (b) the phases φ(t) of these eight waveforms are
plotted. As explained in the text, this is the largest number of distinguishable
waveforms that are band-limited to [0, B] and have the same intensity. We
stress that waveforms that only differ by a constant (time-independent) phase
are not counted as distinguishable in our definition.
for all values of θ. Notice that distinguishability by means of
a coherent receiver (4) doesn’t necessarily imply distinguishability by means of a direct detection receiver. The gap between
the two is the subject of the subsection that follows.
B. The multiplicity of complex waveforms having the same
intensity
We consider a complex signal E(t), whose spectrum is
contained within a bandwidth B and which is periodic in time,
with a period M/B, with M being an integer. The assumption
of periodicity is not a limiting factor in our arguments, as once
the main results are established, the non-periodic case can be
addressed by assuming the limit of M → ∞. A direct detection receiver can only exploit the intensity I(t) = |E0 (t)|2 in
order to extract the transmitted data. Our first claim, which
is key to proving the main arguments of this paper, is that
there are at most 2M −1 distinguishable legitimate waveforms
Ej (t) (with j = 0, 1, . . . , 2M −1 ), whose intensities |Ej (t)|2
are equal to I(t). An illustration of this idea in the case of
M = 4 can be found in Figure 2.
In order to formally prove our statements, we express E(t)
as a Fourier series having at most M non-zero elements,
E(t) =
M
−1
X
k=0
Fk e−ikΩt ,
(5)
ZZk∗ − 1
,
Z − Zk
(9)
one for each zero of A(Z). Since |Uk exp(iθ) | = 1, these
functions have the property that the action of Uk exp(−iΩt)
on E(t) produces a pure phase modulation and hence they
can be considered as the dual of all-pass filters where time
and frequency are interchanged. When any combination of
these functions multiplies A(Z), it does not change the degree
of the resulting polynomial, as the corresponding zeros of
A(Z) are simply reflected with respect to the unit circle, as
illustrated below. For example, if we multiply A(Z) by the
product of U1 (Z), U3 (Z) and U7 (Z), the zeros Z1 , Z3 , and
Z7 are replaced by 1/Z1∗ , 1/Z3∗ , and 1/Z7∗ , respectively. Yet,
the modulus of the product A(Z)U1 (Z)U3 (Z)U7 (Z) remains
identical to the modulus of A(Z) when Z is on the unit circle
(and in particular, when Z = exp(−iΩt)). Since there is a
total of M − 1 functions Uk (Z), there are 2M −1 functions
Aj (Z) that have the same modulus on the unit circle. Thus, we
end up with 2M −1 time waveforms Ej (t) = Aj exp(−iΩt)
whose intensity is I(t), i.e. identical to the intensity of E(t).
−1
Note that {Uk (Z)}M
k=0 are the only functions applying a pure
phase modulation to E(t) that also preserve the number of
elements in the Z-transform (the degree of the polynomial)
and consequently the spectral width of the resulting time
M −1
waveforms. For this reason {Ej (t)}2j=1 are the only temporal
waveforms whose intensity is I(t), and whose spectrum is
fully contained within the bandwidth B. A further discussion
M −1
of the uniqueness of the waveforms {Ej (t)}2j=1 is provided
in the appendix.
Prior to concluding this section, it is interesting to stress that
while 2M −1 is the highest possible number of distinguishable
band-limited waveforms whose intensity equals I(t), the actual
number of such waveforms is 2N0 , where N0 is the number
of zeros that are not located on the unit circle. That is
because when a zero Zl falls on the unit circle, Ul (Z) can
4
be easily verified to be a constant (i.e. Z-independent) phasefactor whose application to A(Z) does not produce a new
waveform. Note also that in this situation E(tl ) = 0 with
tl = arg(Zl )/Ω.4
C. The implications to capacity
We now prove the following relation between Cd – the
information capacity of the direct detection channel, and the
capacity Cc of a system using coherent detection
Cc − 1 ≤ Cd ≤ Cc ,
(10)
where in all cases, we are referring to the capacity per complex
degree of freedom.5 We denote by X the input alphabet of
our channel and by Y 0 the output alphabet available to a
coherent receiver. The output alphabet that is available to a
direct detection receiver is denoted by Y . Since no constraints
are imposed on the transmitter, the alphabet X contains all
complex waveforms without restriction. The alphabet Y 0 , on
the other hand, contains only those complex waveforms that
are band-limited to B, whereas Y contains all real-valued
waveforms that can be obtained by squaring the absolute value
of the waveforms contained in Y 0 . Communication requires
that a probability px (x) is prescribed to the transmission of
each individual waveform x ∈ X.6 The effect of the communications channel (noise distortions etc.) is characterized
by the conditional probabilities of detecting a given element
y 0 ∈ Y 0 (in the case of coherent detection), or y ∈ Y (in
the case of direct detection), given that a particular element
x ∈ X was transmitted. These conditional probabilities are
denoted by py0 |x (y 0 |x) and py|x (y|x), respectively. The mutual
information per complex degree of freedom between the
transmitter and each of the two receivers equals [19]
1
[H(Y 0 ) − H(Y 0 |X)] ,
(11)
M
1
I(X; Y ) =
[H(Y ) − H(Y |X)] ,
(12)
M
where the entropy H(Y ) and the conditional entropy H(Y |X)
are given by7
X
py (y) log2 py (y)
(13)
H(Y ) = −
I(X; Y 0 )
=
y
X
X
H(Y |X) = − px (x)
py|x (y|x) log2 py|x (y|x) ,(14)
x
y
and where the corresponding equations H(Y 0 ) and H(Y 0 |X)
are obtained by replacing y with y 0 in all places. The capacities
Cc and Cd are obtained by maximizing the mutual information
4 Interestingly,
when E(t) equals 0 exactly M − 1 times within the time
period M/B, then it is also the only waveform that is band-limited to B and
has that particular intensity.
5 The number of complex degrees of freedom is M , which is the product
of the temporal duration of the signal M/B and the bandwidth B.
6 Since x is an element of the alphabet X it represents a time dependent
waveform. Nonetheless, in order to keep the notation simple, we avoid writing
x(t), leaving the time dependence of x implicit. Additionally, in order to avoid
over-cluttering the notation, we denote the probability distribution of x simply
by px (x). A similar practice is used with the elements of Y and Y 0 .
7 In line with our simplified notation, summation over x and y should be
P
interpreted in a generalized sense. In addition py (y) = x px (x)py|x (y|x).
values of Eqs. (11) and (12) with respect to the transmitted
distribution px (x). In order to derive Eq. (10), we take
advantage of the relation
I(X; Y 0 )
= I(X; Y 0 , Y ) = I(X; Y ) + I(X; Y 0 |Y )
M −1
,
(15)
≤ I(X; Y ) +
M
where the first equality follows from the fact that
I(X; Y |Y 0 ) = 0, and the second equality follows from the
relations
1
[H(X) − H(X|Y 0 , Y )] ,
(16)
I(X; Y 0 , Y ) =
M
1
I(X; Y ) =
[H(X) − H(X|Y )] ,
(17)
M
1
I(X; Y 0 |Y ) =
[H(X|Y ) − H(X|Y 0 , Y )] . (18)
M
The last inequality is true because Y 0 can take no more than
2M −1 functional values for any given Y .
In the limit of large M , Eq. (15) reduces to
I(X; Y 0 ) − 1 ≤ I(X; Y ) ≤ I(X; Y 0 ).
(19)
Note that expressions (15) and (19) hold for any distribution
of the transmitted alphabet PX (x). This means that, for any
modulation format, the information per complex degree of
freedom that can be extracted when using a direct detection
receiver is at most one bit less than the information per channel
use that can be extracted with coherent detection.
When px (x) is set to be the distribution that maximizes
I(X; Y 0 ), we arrive at
Cc − 1 ≤ Ip (X; Y ) ≤ Cc ,
(20)
where Ip (X; Y ) is the mutual information I(X; Y ) that corresponds to the distribution px (x) for which Cc is attained.
Clearly, Cd ≥ Ip (X; Y ) and hence Cd ≥ Cc − 1. Nonetheless,
Cd remains smaller or equal to Cc , as follows from the rightside inequality of Eq. (15). This concludes the proof of Eq.
(10).
Finally, we note that the capacity per complex degree of
freedom, which we have evaluated in Sec. III-C is identical to
the spectral efficiency, which is the more commonly used term
in the context of fiber communications. Hence the spectral
efficiency of a direct detection system is at most 1 bit/sec/Hz
smaller than that of a system using coherent detection. In order
to see that the two are exactly the same, note that B is both
the bandwidth of the optical signal as well as the number of
complex degrees of freedom that are transmitted per second.
IV. E XTENSION TO N ONLINEAR SYSTEMS OR TO
NON - ADDITIVE NOISE
Fiber-optic communications are often affected by the nonlinear propagation phenomena taking place in optical fibers.
Their effect is not only to distort the signal itself, but also to
cause a nonlinear interaction between the signal and noise, in
which case the noise can no longer be modeled as additive.
From the standpoint of our current study, the only difficulty
that is imposed by this situation is that it is impossible to
relate to the spectrum occupied by the data-carrying signal
5
as a constant, and hence the definition of spectral efficiency
becomes problematic. Nonetheless, it must be stressed that our
analysis of the received waveforms in Sec. III did not explicitly
assume anything about the type of noise, or propagation.
Therefore, our results with respect to the capacity of the
optically filtered signal in Fig. 1b, remain perfectly valid.
In other words, after band-pass filtering, the information per
degree of freedom that is contained in the received complex
optical signal is at most one bit larger than the information
contained in its intensity. With this said, it must be noted
that we do not claim that positioning of a square filter in
front of the receiver is an optimal practice in the nonlinear
case. Nonetheless, in practical situations encountered in fiber
communications, the inclusion of such a filter is practically
unavoidable.
V. D ISCUSSION
While Eq. (10) corresponds to the only relevant case of
M → ∞, the opposite limit of M = 1, which can be
deduced from Eq. (15) may challenge one’s physical intuition,
as it predicts equality between the mutual information values
corresponding to direct and coherent detection. For this reason
the discussion of this special case is interesting in spite of
the fact that it is of no practical importance whatsoever. In
order to resolve this apparent conundrum, note that the case
M = 1 represents a situation in which the complex field E is
time independent. In particular, the phase difference between
any two possible fields is also time independent, implying
that the fields are distinguishable only provided that their
intensities differ. Hence, in this artificial situation, the coherent
receiver has no advantage over the direct detection receiver and
therefore their capacities are identical.
Another curious point related to the assumption of periodicity is that it is not the only convenient choice for arriving
at the result of Sec. III.B. Since E(t) is band limited and its
spectrum is contained within ω ∈ [0, B], it can be written as
E(t) =
∞
X
En e−i(Bt−n)π sinc [π(Bt − n)] ,
(21)
n=−∞
where, as noted earlier, En = E(t = n/B), and where
sinc(x) = sin(x)/x. If we impose the requirement that
En = 0 for n < 0 and for n ≥ M , we end up with a bandlimited, but non-periodic E(t). Nonetheless, the number of
waveforms Ej (t) whose intensity equals that of E(t) remains
at most 2M −1 . In order to see that, consider a time interval of
M 0 /B that contains the interval M/B at its center. Assume
also that M 0 M , so that the tails of the various sinc
functions decay to the extent that the signal within M 0 /B can
be extended periodically without introducing any bandwidth
broadening. We may now apply the reasoning of Sec. III to the
signal in the interval M 0 /B, according to which the number
of equal intensity waveforms is 2 to the power of the number
of zeros in A(Z) that do not coincide with the unit circle.
Evidently, the number of such zeros is at most M −1, because
there are at least M 0 − M zeros that fall on the unit circle.
These are the zeros of E(t) at the times tl = l/B (with l
being outside of the range of 0 to M − 1), which correspond
to zeros in A(Z) at Zl = exp(iΩtl ), i.e. on the unit circle.
Finally, it is important to stress the consequences of our
definition of direct detection, which requires that the incoming
optical signal is detected by a single photo-diode per polarization, and without any manipulation of the signal prior to
photo-detection.8 This definition excludes not only the use of
a local oscillator, as in coherent detection, but also all selfcoherent schemes, such as the ones proposed in [20], and
phase-reconstruction schemes of the kind considered in [21],
[22].
VI. ACKNOWLEDGEMENT
A. Mecozzi acknowledges financial support from the Italian
Government under project INCIPICT. M. Shtaif acknowledges financial support from Israel Science Foundation (grant
1401/16).
A PPENDIX A
Our discussion in Sec. III involved the statement that
the number of distinguishable complex waveforms that are
characterized by a bandwidth B and a period M/B, cannot
be greater than 2M −1 . One justification for this claim is
that the only functions in Z-domain that produce pure phase
modulation and do not increase the order of the polynomial
A(Z) (and hence they do not increase the bandwidth of E(t))
are of the form ck (Z Z̃k − 1)/(Zk − 1), where Zk is one of the
zeros of A(Z) and ck a constant. In order for these functions
not to change the time-domain intensity waveform of E(t),
one must have Z̃k = Zk∗ and |ck | = 1, so that the amplitude
of the function’s transfer function on the unit circle is 1. This
means that the only functions that can be applied to A(Z)
without changing neither the order of the polynomial, nor the
intensity waveform, are the M − 1 functions specified in Eq.
(9). Indeed, the number of such function combinations does
not exceed 2M −1 . Here, we also present an alternative proof
that is based on the uniqueness of minimum-phase functions.
Given a periodic band-limited waveform, we have shown
that by reflecting any of its zeros with respect to the unit
circle, as described in Sec. III.B, one obtains a different
waveform having the same intensity, bandwidth, and timeperiod.
If we look at an arbitrary given waveform E0 (t) =
p
I(t) exp(iφ0 (t)) of the above specified characteristics, and
identify all of its zeros, we can chose to reflect only the
zeros that are inside thep unit circle, thereby producing a
new waveform Em (t) = I(t) exp(φm (t)), having the same
intensity I(t), but a different phase. Since the spectrum of
Em (t) is contained between 0 and B, and since all of its zeros
in Z-domain are outside the unit circle, it belongs to a special
class of functions that is famously known as functions of
minimum-phase [23]. One of the most well known properties
of such functions is that up to an immaterial constant, their
8 The requirement that there is no manipulation of the signal prior to photodetection can be replaced by the requirement that no manipulation other than
all-pass filtering (e.g. dispersion) is applied prior to photo-detection. The
reason is that all pass filtering can also be done at the transmitter and hence
it does not affect the assumption of this work.
6
phase is uniquely determined by their intensity
means
h by
i of
p
the the Hilbert transform. Namely, φ(t) = H log I(t) + c,
where H[·] designates the Hilbert transform and where c is
an unknown constant. Since waveforms differing only by a
constant phase are indistinguishable in our definition (see Sec.
III.B), we conclude that the minimum phase function that
corresponds to a given intensity profile is unique.
The uniqueness of the minimum-phase function implies that
each waveform in the set of distinguishable equal intensity
waveforms that are band-limited to B and periodic in M/B,
can be obtained from any other waveform in the set by means
of functions of the form given in Eq. (9), whose effect is
to reflect the zeros of the waveforms that it acts upon in
Z-domain. Had it not been so, different waveforms in the
set would have produced different minimum phase functions.
Therefore the total number of distinguishable waveforms in
the set cannot exceed 2M −1 .
R EFERENCES
[1] S. Randel, F. Breyer, S.C.J. Lee, and J.W. Walewski, “Advanced Modulation Schemes for Short-Range Optical Communications,” IEEE J. Sel.
Topics Quantum Electron. 16, 1280–1289 (2010).
[2] T. Takahara, T. Tanaka, M. Nishihara, Y. Kai, L. Li, Z. Tao, and J. Rasmussen, “Discrete Multi-Tone for 100 Gb/s Optical Access Networks,” in
Optical Fiber Communication Conference, OSA Technical Digest (online)
(Optical Society of America, 2014), paper M2I.1.
[3] A. Weiss, A. Yeredor, and M. Shtaif, “Iterative Symbol Recovery For
Power Efficient DC Biased Optical OFDM Systems, IEEE J. of Lightwave
Technol. 34, 2331–2338 (2016).
[4] A.J. Lowery and J. Armstrong, “Orthogonal-frequency-division multiplexing for dispersion compensation of long-haul optical systems,” Opt.
Express 14, 2079–2084 (2006).
[5] B.J.C. Schmidt, A.J. Lowery, and J. Armstrong, “Experimental Demonstrations of Electronic Dispersion Compensation for Long-Haul Transmission Using Direct-Detection Optical OFDM,” J. Lightwave Technol.
26, 196–203 (2008).
[6] A. Li, D. Che, V. Chen, and W. Shieh, “Spectrally efficient optical
transmission based on Stokes vector direct detection,” Opt. Express 22,
15662–15667 (2016).
[7] S. Randel, D. Pilori, S. Chandrasekhar, G. Raybon, and P.J. Winzer
“100-Gb/s Discrete-Multitone Transmission Over 80-km SSMF Using Single-Sideband Modulation With Novel Interference-Cancellation
Scheme,” Proc. of European Conference of Optical Communications 2015
(ECOC15), Valencia – Spain, Paper 0697 (2015).
[8] M. Schuster, S. Randel, C.A. Bunge, S.C.J. Lee, F. Breyer, B. Spinnler,
and K. Petermann, “Spectrally Efficient Compatible Single-Sideband
Modulation for OFDM Transmission With Direct Detection,” IEEE
Photon. Technol. Letters 20, 670–672 (2008).
[9] A. Mecozzi, C. Antonelli, and M. Shtaif, “KK coherent receiver,” Optica
3, 1220-1227 (2016).
[10] C. Antonelli, A. Mecozzi, and M. Shtaif, “Kramers-Kronig PAM
Transceiver,” in Optical Fiber Communication Conference, OSA Technical Digest (online) (Optical Society of America, 2017), paper Tu3I.5.
[11] X. Chen, C. Antonelli, S. Chandrasekhar, G. Raybon, J. Sinsky, A.
Mecozzi, M. Shtaif, and P. Winzer, “218-Gb/s Single-Wavelength, SinglePolarization, Single- photo-diode Transmission Over 125-km of Standard
Singlemode Fiber Using Kramers-Kronig Detection,” in Optical Fiber
Communication Conference, OSA Technical Digest (online) (Optical
Society of America, 2017), post deadline paper Th5B.6
[12] Z. Li, S. Erkilinc, K. Shi, E. Sillekens, L. Galdino, B. Thomsen, P.
Bayvel, and R. Killey, “SSBI Mitigation and Kramers-Kronig Scheme
in Single-Sideband Direct-Detection Transmission with Receiver-based
Electronic Dispersion Compensation,” J. Lightwave Technol. 35, 1887–
1893 (2017).
[13] A. Mecozzi and M. Shtaif, “On the Capacity of Intensity Modulated
Systems Using Optical Amplifiers,” IEEE Photon. Technol. Lett. 13,
1029–1031 (2001).
[14] A. Lapidoth, “On phase noise channels at high SNR,” in Proc. IEEE
Information Theory Workshop (ITW), Bangalore, India, Oct. 2002, pp.
1–4.
[15] S. Hranilovic and F. R. Kschischang, “Capacity bounds for power- and
band-limited optical intensity channels corrupted by Gaussian noise,”
IEEE Transactions on Information Theory 50, 784–795 (2004).
[16] M. Katz and S. Shamai, “On the capacity-achieving distribution of the
discrete-time noncoherent and partially coherent AWGN channels,” IEEE
Transactions on Information Theory 50, 2257–2270 (2004).
[17] A. Lapidoth, S. M. Moser, and M. A. Wigger, “On the Capacity of FreeSpace Optical Intensity Channels,” IEEE Transactions on Information
Theory 55, 4449–4461 (2009).
[18] E.N. Gilbert, “Increased information rate by oversampling,” IEEE Transactions on Information Theory 39, 1973–1976 (1993).
[19] C. E. Shannon, “Mathematical theory of communications,” Bell System
Technical Journal, 27, pp. 379–423 & 623–656, July & October, 1948.
[20] Xiang Liu, S. Chandrasekhar, and Andreas Leven, “Digital self-coherent
detection,” Opt. Express 16, 792–803 (2008).
[21] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and
M. Segev, “Phase Retrieval with Application to Optical Imaging: A
contemporary overview,” IEEE Signal Processing Magazine 32, 87–109
(2015).
[22] Gang Wang, Georgios B. Giannakis, Yonina C. Eldar, “Solving Systems of Random Quadratic Equations via Truncated Amplitude Flow,”
arXiv:1605.08285, available at https://arxiv.org/abs/1605.08285.
[23] Alan V. Oppenheim and Ronald W. Schafer, Discrete-Time Signal Processing, 3nd ed, Upper Saddle River, NJ: Prentice-Hall Signal Processing
Series, (2010).
| 7cs.IT
|
Extracting Three Dimensional Surface Model of Human Kidney from the
Visible Human Data Set using Free Software
Kirana Kumara P*
Centre for Product Design and Manufacturing
Indian Institute of Science
Bangalore, INDIA - 560012
E-mail: [email protected]
*
Corresponding author: Phone: +91 80 2293 2284; Fax: +91 80 2360 1975
Abstract
Three dimensional digital model of a representative human kidney is needed
for a surgical simulator that is capable of simulating a laparoscopic surgery
involving kidney. Buying a three dimensional computer model of a
representative human kidney, or reconstructing a human kidney from an
image sequence using commercial software, both involve (sometimes
significant amount of) money. In this paper, author has shown that one can
obtain a three dimensional surface model of human kidney by making use of
images from the Visible Human Data Set and a few free software packages
(ImageJ, ITK-SNAP, and MeshLab in particular). Images from the Visible
Human Data Set, and the software packages used here, both do not cost
anything. Hence, the practice of extracting the geometry of a representative
human kidney for free, as illustrated in the present work, could be a free
alternative to the use of expensive commercial software or to the purchase of a
digital model.
Keywords
Visible; Human; Data; Set; Kidney; Surface; Model; Free.
Introduction
Laparoscopic surgery is often a substitute for a traditional open surgery, when human
kidney is the organ that is to be operated upon [1]; choosing laparoscopic surgery over an
open surgery reduces trauma and shortens the recovery time for the patient [1]. But since
laparoscopic surgery needs highly skilled surgeons, it is preferable to use a surgical simulator
for training and evaluating surgeons [2]. A surgical simulator that can simulate a laparoscopic
surgery over a human kidney needs a virtual kidney, i.e., a computer (digital) three
dimensional (3D) model of a representative human kidney.
Currently, mainly two approaches are being practiced to obtain the geometry of a
representative human kidney. The first approach is to buy a readily available 3D model of a
human kidney from an online store (e.g., [3]). The second approach is to use commercial
software packages (such as 3D-DOCTOR [4], Mimics [5], Amira [6]) to reconstruct a 3D
geometry of a kidney from a two dimensional (2D) image sequence. One can see that both of
these approaches cost (sometimes significant amount of) money.
Present work shows that it is possible to obtain a three dimensional surface model of a
representative human kidney, completely for free. Present approach is to make use of a few
free software packages to extract 3D geometry of human kidney from CT-scan images from
the Visible Human Data Set (VHD) (also known as The Visible Human Project Image Data
Set, or The Visible Human Project Data Sets) [7]. The free software packages used are ImageJ
[8-10], ITK-SNAP [11, 12], and MeshLab [13, 14]. One can note that images from VHD may
be downloaded for free, after obtaining a free license [15] from National Library of Medicine
(NLM) [16] which is a part of National Institutes of Health (NIH) [17]; VHD is a part of the
more ambitious Visible Human Project (VHP) [18].
Approaches similar to the approach presented in the present paper may be found in the
present author’s previous works [19, 20]. Although the free software packages used in those
works are the same as the ones used in the present work, the images were not from VHD;
images used in [19] and [20] were downloaded from [21] (the images are no longer accessible
as of now; but images were downloadable sometime back). Also, [19] and [20] discuss the
reconstruction of a pig liver, while the present work deals with the reconstruction of a human
kidney.
Upon conducting literature review, one can see that there are authors who have used
VHD together with commercial software packages, e.g., [22]. Also, there are authors who
have used images from sources other than VHD and have used commercial software packages
to perform 3D reconstruction of biological organs [23-26]. Also, there are authors who have
used free software packages to extract geometry of biological organs [27]. But the present
author could not find any source in the literature where the three free software packages –
ImageJ, ITK-SNAP and MeshLab – were used to obtain 3D surface model of human kidney
from CT-scan images from the VHD.
The practice of extracting the geometry of a representative human kidney for free, as
presented in the present work, could be a free alternative to the use of expensive commercial
software or to the purchase of a digital model.
Material and Method
For the present work, CT-scan images from VHD and the three free and open source
software packages – ImageJ, ITK-SNAP, and MeshLab – form the material. As far as method
is concerned, the three software packages are used to reconstruct 3D models of human kidney
from the CT-scan images from the VHD.
VHD contains CT, MRI and cryosection images. In this work, only normal CT-scan
images of visible human male and female are used. Present work uses images in the ‘png’
format since this is the format recommended by VHP. File size of CT-scan images is small,
and the images are good enough for reconstructing a 3D model of whole kidney (i.e., inner (or
finer) details of kidney are not present; reconstructed 3D model represents just the outer
surface of kidney). One can easily identify human kidney in the CT-scan images of VHD.
In the present work, ImageJ 1.42q is used to form an image stack which contains
kidney. ITK-SNAP Version 2.0.0 is used for segmentation and 3D reconstruction to the
correct scale. MeshLab v1.2.2 is used to control the level of detail in the reconstructed 3D
model; it also serves as a tool to smoothen the 3D model and reduce its file size.
Now the method is explained in a bit detail in the following subsections.
Using ImageJ to Form an Image Stack
CT-scan images for visible human male and female are available from head to toe. Out
of these images, one has to identify the images which belong to kidney. Upon viewing
individual images in ImageJ, and upon consulting [28] and [29], one can conclude that for
visible human male, both left are right kidneys are contained between the images
‘cvm1551f.png’ and ‘cvm1692f.png’ (47 images in total). Similarly, for visible human
female, both left and right kidneys are contained between the images ‘cvf1564f.png’ and
‘cvf1693f.png’ (130 images in total). Now these 47 images for male, and 130 images for
female, have to be copied into two separate empty folders. Now, to form an image stack for
male, select the menu item ‘File -> Import -> Image Sequence…’, browse to the location of
the folder containing 47 images and select the first image in the folder and follow the prompts
(with default options); all 47 images are now displayed in ImageJ as an image stack; now,
select the menu item ‘File -> Save As -> Raw Data…’ to save the image stack in the ‘*.raw’
format (where * is the name given). Similar procedure may be followed to obtain an image
stack for the female.
Using ITK-SNAP to Perform Segmentation and 3D Reconstruction
ITK-SNAP does the segmentation and 3D reconstruction to the correct scale. Hence
header information for the images in the image stack is essential. VHD contains header
information for each of the images in its database. Upon going through the header files of
each of the 47 images of male, one can note that the following header information is identical
for all the 47 images: Image matrix size – X = 512, Image matrix size – Y = 512, Image
dimension – X = 460 mm, Image dimension –Y = 460 mm, Image pixel size – X = 0.898438,
Image pixel size – Y = 0.898438, Screen Format = 16 bit, Spacing between scans = 3 mm.
Similarly, the following header information is identical for all the 130 female images: Image
matrix size – X = 512, Image matrix size – Y = 512, Image dimension – X = 480 mm, Image
dimension –Y = 480 mm, Image pixel size – X = 0.9375, Image pixel size – Y = 0.9375,
Screen Format = 16 bit, Spacing between scans = 1 mm.
Now, the method of reconstructing the left kidney of the male is explained in detail,
with illustrations. The same method may be employed to reconstruct the right kidney of the
male, and the left and right kidneys of the female.
Select the menu item ‘File -> Open Greyscale Image…’, browse to the location of the
image stack for male, follow the prompts and supply the header information. As noted in the
first paragraph of this subsection, the ‘missing header information’ to be supplied for the
image stack (for male) is: Image dimensions: X: 512, Y: 512, Z: 47, Voxel spacing: X =
0.898438, Y = 0.898438, Z = 3, Voxel representation: 16 bit, unsigned. Once the header
information is supplied, image stack is displayed in ITK-SNAP. One can browse through all
the 47 images in the image stack. For illustration purposes, 17th image and 33rd image in the
image stack (i.e., ‘cvm1602f.png’ and ‘cvm1650f.png’ in the VHD) are shown in Figure 1
and Figure 2 respectively; also, the left and right kidneys are identified in Figure 1 and Figure
2, by making use of illustrations from [28] and [29].
Right Kidney
Left Kidney
Figure 1. The 17th image in the image stack
Right Kidney
Left Kidney
Figure 2. The 33rd image in the image stack
Now the task is to do the segmentation. Select ‘Polygon tool’ from the ‘IRIS
Toolbox’, for slice-by-slice manual segmentation. Select ‘continuous’ radio button under
‘Polygon Tool’. Now click and drag the mouse cursor along the edge of the left kidney (as
seen in the axial view (window)), carefully. This draws the contour of the edge of the left
kidney. Now, right click on the image, and select the ‘accept’ button to create the
segmentation for the image on display. This process has to be repeated for all images in the
image stack, which contain pixels that belong to the left kidney. For illustration purposes,
17th image and 33rd image in the image stack, after segmentation, are shown in Figure 3 and
Figure 4 respectively.
Segmented
Left Kidney
Figure 3. The 17th image in the image stack (after segmentation)
Segmented
Left Kidney
Figure 4. The 33rd image in the image stack (after segmentation)
Once the segmentation is over, 3D reconstruction is to be carried out. This is
accomplished by the menu item ‘Segmentation -> Save As Mesh…’, following prompts,
browsing to the location where the reconstructed model is to be stored, and giving a name in
the format ‘path\*.stl’ for the file that represents the reconstructed 3D model (where ‘path’ is
the complete path, e.g., C:\Users\PC\Desktop\*.stl, and * is any file name).
Now, a 3D reconstruction of the left kidney for the visible human male is over. Similar
process may be followed to reconstruct the right kidney of the visible human male, and the
left and right kidneys of the visible human female.
Using MeshLab to Reduce the Total Number of Faces Describing the 3D
Model
The 3D model of kidney obtained through the use of ITK-SNAP typically is of very
large size and typically is described by a very large number of surface triangles. MeshLab
could be very helpful in reducing the total number of surface triangles that are needed to
describe the 3D model satisfactorily. It also serves as a tool to smoothen the reconstructed 3D
geometry; after using smoothing features provided by MeshLab, it may be necessary to scale
the reconstructed 3D models to the correct dimensions, if the original dimensions are to be
strictly retained. MeshLab can also improve the triangle quality of surface triangles of the 3D
model. It can also reduce the file size.
The 3D models of kidney, after undergoing processing with MeshLab, are shown in
the next section.
Results
Reconstructed left kidney of the male, after undergoing processing through MeshLab,
is shown in Figure 5. Similarly, reconstructed right kidney of the male, after undergoing
processing through MeshLab, is shown in Figure 6. Reconstructed left kidney of the female is
shown in Figure 7. Reconstructed right kidney of the female is shown in Figure 8. All the four
3D models are made up of 1000 surface triangles. While obtaining these four models, job of
MeshLab is to smoothen the 3D models reconstructed through ITK-SNAP and to reduce the
total number of surface triangles to 1000.
Figure 5. Reconstructed left kidney of male
Figure 6. Reconstructed right kidney of male
Figure 7. Reconstructed left kidney of female
Figure 8. Reconstructed right kidney of female
Discussion
In this work, 3D model of human kidney is extracted from CT-scan images from the
VHD, using free software packages. The free software packages used are: 1) ImageJ 2) ITKSNAP 3) MeshLab. The organs reconstructed are: 1) left kidney of visible human male 2)
right kidney of visible human male 3) left kidney of visible human female 4) right kidney of
visible human female. All the four models are in ‘stl’ format.
Use of free software packages together with images that may be obtained for free, as
has been done in the present work, makes it possible to obtain the geometry of a
representative human kidney, completely for free. Buying a 3D model of a human kidney, or
using a commercial software package to extract 3D models from image sequences, cost
(sometimes significant amount of) money. Also, in the present approach, user can control how
finely the geometry should be described (using the free software package MeshLab). Since
MeshLab can improve the quality of the surface mesh that describes a reconstructed 3D
model, the reconstructed 3D model that has undergone processing with MeshLab can be used
in a finite element analysis after converting the surface model to a solid model using software
packages like Rhinoceros. Also, the method used to extract the geometry of a kidney, as
illustrated in the present work, may possibly be used to extract other whole biological organs
from VHD.
It may be noted that the method given here to obtain the 3D models of human kidney
need not be followed rigidly. It is good to read the documentation for the software packages
used here, and one can experiment with the various options provided by the software
packages instead of rigidly following the method illustrated in this work. For example, instead
of tracing the boundary of the kidney in each of the images through the mouse pointer, the
‘Paintbrush tool’ provided by ITK-SNAP can be tried out to carry out the segmentation; ITKSNAP also provides a tool that can do semi-automatic segmentation.
As to the limitations, present work uses only CT-scan images. Although these are
found to be sufficient to obtain the geometry of a whole kidney, whenever the reconstructed
geometry should include the finer details of the kidney, or whenever some other organ is to be
extracted from VHD, there is a possibility that other types of images (e.g., MRI images) are
more suited in some cases. Also, multiple software packages need to be downloaded, installed
and used here.
Future work is to extract other biological organs from VHD, using free software
packages. Aim is to reconstruct biological organs with inner details (not obtaining just the
outer surface of the organs) and to use other types of images (like MRI, cryosection images)
from the VHD if need be.
Conclusion
It is possible to obtain the 3D surface model of a representative human kidney from
CT-scan images from the VHD, using free software packages only; the free software
packages needed are: ImageJ, ITK-SNAP, and MeshLab. The practice of extracting the
geometry of a representative human kidney completely for free, as illustrated in the present
work, could be a free alternative to the use of expensive commercial software packages or to
the purchase of a digital model.
Acknowledgements
Author is grateful to the Robotics Lab, Department of Mechanical Engineering &
Centre for Product Design and Manufacturing, Indian Institute of Science, Bangalore, INDIA,
for providing the necessary infrastructure to carry out this work.
Author acknowledges Prof. Ashitava Ghosal, Robotics Lab, Department of
Mechanical Engineering & Centre for Product Design and Manufacturing, Indian Institute of
Science, Bangalore, INDIA, for providing the CT-scan images from the Visible Human Data
Set (VHD).
Author acknowledges National Library of Medicine (NLM) and Visible Human
Project (VHP) for providing free access to the Visible Human Data Set (VHD) to Prof.
Ashitava Ghosal.
Visible Human Data Set (VHD) is an anatomical data set developed under a contract
from the National Library of Medicine (NLM) by the Departments of Cellular and Structural
Biology, and Radiology, University of Colorado School of Medicine.
References
1.
Jay T. Bishoff, Louis R. Kavoussi (online), LAPAROSCOPIC SURGERY OF THE
KIDNEY. Available at: http://www.pkdiet.com/pdf/Laparoscopic%20nephrectomy.pdf
(accessed July 14, 2012)
2.
Issenberg SB, McGaghie WC, Hart IR, Mayer JW, Felner JM, Petrusa ER, Waugh RA,
Brown DD, Safford RR, Gessner IH, Gordon DL, Ewy GA., Simulation Technology for
Health Care Professional Skills Training and Assessment, The Journal of the American
Medical Association, 1999, 282(9), p. 861-6.
3.
http://www.turbosquid.com (accessed July 14, 2012)
4.
http://www.ablesw.com/3d-doctor/ (accessed July 14, 2012)
5.
http://biomedical.materialise.com/mimics (accessed July 14, 2012)
6.
http://www.amira.com/ (accessed July 14, 2012)
7.
http://www.nlm.nih.gov/research/visible/getting_data.html (accessed July 14, 2012)
8.
http://rsbweb.nih.gov/ij/ (accessed July 14, 2012)
9.
Rasband W.S., ImageJ, U. S. National Institutes of Health, Bethesda, Maryland, USA,
http://rsb.info.nih.gov/ij/ , 1997-2009.
10. Abramoff M.D., Magelhaes P.J., Ram S.J., Image Processing with ImageJ,
Biophotonics International, 2004, volume 11, issue 7, pp. 36-42.
11. http://www.itksnap.org/pmwiki/pmwiki.php (accessed July 14, 2012)
12. Paul A. Yushkevich, Joseph Piven, Heather Cody Hazlett, Rachel Gimpel Smith, Sean
Ho, James C. Gee, Guido Gerig, User-guided 3D active contour segmentation of
anatomical structures: Significantly improved efficiency and reliability, Neuroimage,
2006, 31(3), 1116-28.
13. http://meshlab.sourceforge.net/ (accessed July 14, 2012)
14. MeshLab, Visual Computing Lab – ISTI – CNR
15. http://www.nlm.nih.gov/research/visible/vhpagree.pdf (accessed July 14, 2012)
16. http://www.nlm.nih.gov/ (accessed July 14, 2012)
17. http://www.nih.gov/ (accessed July 14, 2012)
18. http://www.nlm.nih.gov/research/visible/visible_human.html (accessed July 14, 2012)
19. Kirana Kumara P, Ashitava Ghosal, A Procedure for the 3D Reconstruction of
Biological Organs from 2D Image Sequences, Proceedings of BEATS 2010, 2010,
International Conference on Biomedical Engineering and Assistive Technologies
(BEATS 2010), Dr B R Ambedkar National Institute of Technology, Jalandhar, INDIA.
20. Kirana Kumara P (online), Reconstructing Solid Model from 2D Scanned Images of
Biological
Organs
for
Finite
Element
Simulation.
Available
at:
http://knolskiranakumarap.wordpress.com/2011/07/03/reconstructing-solid-modelfrom-2d-3rc2kfwq179j2-5/ (accessed July 14, 2012)
21. http://www.biorobotics.harvard.edu/ (accessed July 14, 2012)
22. Aimee Sergovich, Marjorie Johnson, Timothy D. Wilson, Explorable ThreeDimensional Digital Model of the Female Pelvis, Pelvic Contents, and Perineum for
Anatomical Education, Anatomical Sciences Education, 2010, 3:127–133.
23. Dong Sun Shin, Jin Seo Park, Byeong-Seok Shin, Min Suk Chung, Surface models of
the male urogenital organs built from the Visible Korean using popular software,
Anatomy & Cell Biology, 2011, 44(2), p. 151-159.
24. Amy Elizabeth Kerdok, Characterizing the Nonlinear Mechanical Response of Liver to
Surgical Manipulation, Ph.D Thesis, 2006, The Division of Engineering and Applied
Sciences, Harvard University.
25. Li Lou, Shu Wei Liu, Zhen Mei Zhao, Pheng Ann Heng, Yu Chun Tang, Zheng Ping Li,
Yong Ming Xie, Yim Pan Chui, Segmentation and reconstruction of hepatic veins and
intrahepatic portal vein based on the coronal sectional anatomic dataset, Surgical and
Radiologic Anatomy, 2009, 31:763-768.
26. Chen G, Li XC, Wu GQ, Zhang SX, Xiong XF, Tan LW, Yang RG, Li K, Yang SZ,
Dong JH., Three-dimensional reconstruction of digitized human liver: based on Chinese
Visible Human, Chinese Medical Journal, 2010, 123(2), p. 146-50.
27. Hong-jian Gao, 3-D Reconstruction of Liver Slice Images Based on MITK Framework,
The 1st International Conference on Bioinformatics and Biomedical Engineering
(ICBBE 2007), 2007, DOI: 10.1109/ICBBE.2007.247
28. http://www.voxel-man.de/vm/io_correlation.en.html (accessed July 14, 2012)
29. Henry Gray, Anatomy of the Human Body, Philadelphia, Lea & Febiger, 1918.
| 5cs.CE
|
FACE RINGS OF CYCLES, ASSOCIAHEDRA,
AND STANDARD YOUNG TABLEAUX
arXiv:1503.06243v3 [math.CO] 22 Aug 2016
ANTON DOCHTERMANN
Abstract. We show that Jn , the Stanley-Reisner ideal of the n-cycle, has a free resolution
supported on the (n − 3)-dimensional simplicial associahedron An . This resolution is not
minimal for n ≥ 6; in this case the Betti numbers of Jn are strictly smaller than the f -vector
of An . We show that in fact the Betti numbers βd of Jn are in bijection with the number
of standard Young tableaux of shape (d + 1, 2, 1n−d−3 ). This complements the fact that
the number of (d − 1)-dimensional faces of An are given by the number of standard Young
tableaux of (super)shape (d + 1, d + 1, 1n−d−3 ); a bijective proof of this result was first
provided by Stanley. An application of discrete Morse theory yields a cellular resolution of
Jn that we show is minimal at the first syzygy. We furthermore exhibit a simple involution
on the set of associahedron tableaux with fixed points given by the Betti tableaux, suggesting
a Morse matching and in particular a poset structure on these objects.
1. Introduction
In this paper we study some intriguing connections between basic objects from commutative algebra and combinatorics. For K an arbitrary field we let R = K[x1 , x2 , . . . , xn ] denote
the polynomial ring in n variables. We let Jn denote the edge ideal of the complement of the
n-cycle Cn . By definition, Jn is the ideal generated by the degree 2 monomials corresponding
to the diagonals of Cn . One can also realize Jn as the Stanley-Reisner ideal of the cycle Cn
(now thought of as a one-dimensional simplicial complex).
1
2
6
3
5
4
Figure 1. J6 = hx1 x3 , x1 x4 , x1 x5 , x2 x4 , x2 x5 , x2 x6 , x3 x5 , x3 x6 , x4 x6 i.
The ideals Jn are of course very simple algebraic objects and their homological properties
are well-understood. One can verify that R/Jn is a Gorenstein ring, the dimension of R/Jn
is 2, and (hence) the projective dimension of R/Jn is n − 2. In fact a minimal free resolution
can be described explicitly, and cellular realizations have been provided by Biermann [3] and
more recently by Sturgeon [20].
Date: August 23, 2016.
1
We wish to further investigate the combinatorics involved in the resolutions of Jn . Our
original interest in cellular resolutions of Jn came from the fact that the ideal Jn has an
‘almost linear’ resolution, in the sense that the nonzero entries in the differentials of its minimal resolution are linear forms from R, except at the last syzygy where the nonzero entries
are all degree 2. Recent work in combinatorial commutative algebra has seen considerable
interest in cellular resolutions of monomial and binomial ideals (see for example [2], [4], [8],
[9], [12], [13], [15], [17]), but in almost all cases the ideals under consideration have linear
resolutions. Here we seek to extend some of these constructions.
In the construction of any cellular resolution, one must construct a CW -complex with
faces labeled by monomials that generate the ideal. In the case of Jn there is a well known
geometric object whose vertices are labeled by the diagonals of an n-gon, namely the (simplicial) associahedron An . By definition, An is the simplicial complex with vertex set given
by diagonals of an n-gon, with faces given by collections of diagonals that are non-crossing.
The facets of An are triangulations of the n-gon, of which there are a Catalan number many.
It is well known that An is spherical, and in fact can be realized as the boundary of a convex
polytope. In addition there is a natural way to associate a monomial to each face of An , and
in the first part of the paper we show that this labeled facial structure of An (considered as
a polytope with a single interior cell) encodes the syzygies of Jn .
Theorem 3. With its natural monomial labeling, the complex An supports a free resolution
of the ideal Jn .
The resolution of Jn supported on the associahedron An is not minimal for n > 5, and in
particular in this case we have faces F ( G with the same monomial labeling. The f -vector
of An is completely understood (a closed form can be written down), and in fact the number
of (d − 1)-dimensional faces of An is equal to the number of standard Young tableaux of
shape (d + 1, d + 1, 1n−d−3 ); a bijective proof of this was first provided by Stanley [19].
Since a resolution of Jn is supported on An we know that the f -vector of An provides an
upper bound on the Betti numbers βd (R/Jn ), with equality in the case of β1 (R/Jn ) = f0 (An ).
In the second part of the paper we show that the Betti numbers βd of R/Jn are given by
standard Young tableaux on a set of subpartitions involved in the Stanley bijection.
Theorem 6. The total Betti numbers βd of the module R/Jn are given by the number of
standard Young tableaux of shape (d + 1, 2, 1n−d−3 ).
This bijection along with an application of the hook formula leads to a closed form expression
for the Betti numbers of R/Jn . In addition, the fact that the partition (d + 1, 2, 1n−d−3 ) is
conjugate to (n−d−1, 2, 1d+1 ) provides a nice combinatorial interpretation of the palindromic
property βd = βn−d−2 for the Betti numbers of the Gorenstein ring R/Jn .
The fact that we can (in theory) identify the Betti numbers of R/Jn with certain faces
of An suggests that it may be possible to collapse away faces of An to obtain a minimal
resolution of Jn , employing an algebraic version of Morse theory due to Batzies and Welker
([1]). Indeed certain geometric properties of any subdivision of an n-gon (along with the
almost linearity of Jn ) imply that certain faces must be matched away. For d = 2 we are
able to write down a Morse matching involving the edges and 2-faces of An such that the
number of unmatched (critical) cells is precisely β2 (corresponding to the first syzygy module
of R/Jn ), see Proposition 12. This leads to minimal resolutions of Jn for the cases n ≤ 7.
2
In addition, our identification of both the Betti numbers of R/Jn and the faces of An with
standard Young tableaux leads us to consider a partial matching on the set of associahedron
tableaux such that the unmatched elements correspond to the Betti numbers. The hope
would be to import a poset structure from the face poset of An to extend this matching to
a Morse matching. The trouble with this last step is that the Stanley bijection does not
give us an explicit labeling of the faces of An by standard Young tableaux; there are choices
involved and the bijection itself is recursively defined. However, we can define a very simple
partial matching on the set of standard Young tableaux of shape (d + 1, d + 1, 1n−d−3 ) such
that the unmatched elements can naturally be thought of as standard Young tableaux of
shape (d + 1, 2, 1n−d−3 ) (by deleting the largest entries); see Proposition 16. This suggests a
poset structure on the set of standard Young tableax that extends this covering relation.
The rest of the paper is organized as follows. We begin in Section 2 with some basics regarding the commutative algebra involved in our study. In Section 3 we discuss associahedra
and their role in resolutions of Jn . We turn to standard Young tableaux in Section 4 and
here establish our results regarding the Betti numbers of R/Jn . In Section 5 we discuss our
applications of discrete Morse theory and related matchings of stand Young tableaux. We
end with some open questions.
2. Some commutative algebra
As above we let Jn denote the Stanley-Reisner ideal of the n-cycle, by definition the ideal
in R = K[x1 , x2 , . . . , xn ] generated by degree 2 monomials corresponding to the diagonals.
We are interested in combinatorial interpretations of certain homological invariants of Jn ,
and in particular the combinatorial structure of its minimal free resolution. Recall that a
free resolution of an R-module M is an exact sequence of R-modules
0 ← M ← F1 ← F2 ← · · · ← Fp ← 0,
where each Fd ' ⊕j R(−j)βd,j is free and the differential maps are graded. The resolution
is minimal if each of the βd,j are minimum among all resolutions, in which case the βd,j
are called the (graded) Betti numbers of M . Also in this case the number p (length of the
minimal resolution) is called the projective dimension of M .
Our main tool in calculating Betti numbers will be Hochster’s formula (see for example
[14]), which gives a formula for the Betti numbers of the Stanley-Reisner ring R/I∆ associated
to a simplicial complex ∆.
Theorem 1 (Hochster’s formula). For a simplicial complex ∆ on vertex set [n] we let R/I∆
denote its Stanley-Reisner ring. Then for d ≥ 1 the Betti numbers βd,j of R/I∆ are given by
X
(1)
βd,j (∆) =
dimk H̃j−d−1 (∆[W ]; k).
[n]
W ∈( j )
Here ∆[W ] denotes the simplicial complex induced on the vertex set W .
A cellular resolution of M is a CW-complex X with a monomial labeling of its faces,
such that the algebraic chain complex computing the cellular homology of X ‘supports’ a
resolution of M . We refer to Section 3 for details and more precise definitions.
We next collect some easy observations regarding the Betti numbers of Jn . Since Jn is the
Stanley-Reisner ideal of a triangulated 1-dimensional sphere, we see that R/Jn is Gorenstein
3
and has (Krull) dimension 2. The Auslander-Buchsbaum formula then implies that the
projective dimension of R/Jn is n − 2, which says that βd,j = 0 whenever d > n − 2. An easy
application of Hochster’s formula also implies that a minimal resolution of Jn is linear until
the last nonzero term, by which we mean if 0 ≤ d < n − 2, then βd,j = 0 for all j 6= d + 1.
Also, we have βn−2,n = 1 and βn−2,j = 0 for j 6= n. In this sense the ideals Jn have an ‘almost
linear’ resolution, as mentioned in the introduction.
Convention: Since for any d we have βd,j 6= 0 for at most one value of j, we we will
(without loss of generality) sometimes drop the j and use βd = βd,j to denote the Betti
numbers of Jn .
3. Asssociahedra
For each n, we let An denote the (dual) associahedron, the (n − 4)-dimensional simplicial
complex whose vertices are given by diagonals of a labeled regular n-gon, with facets given
by triangulations (collections of diagonals that do not intersect in their interior). It is well
known that An is homeomorphic to a sphere and in fact is polytopal, and several embeddings
(most often of the dual simple polytope) are described throughout the literature (see [6] for a
good account of the history). From here on we will use An to denote the (n − 3)-dimensional
simplicial polytope (i.e., including the interior).
We wish to describe a monomial labeling of the faces of An . Recall that each vertex of
An corresponds to some diagonal {i, j} of an n-gon, so we simply label that vertex with the
monomial xi xj . We label the higher-dimensional faces of An with the least common multiple
of the vertices contained in that face. We wish to show that, with this simple labeling, the
associahedron An supports a resolution of Jn .
x1x3
x1x5
1
2
6
x1x4
x1x3x5
x1x2x4x5
3
5
4
x2x5
x2x4
x3x5
x4x6
x3x6
x2x6
Figure 2. The complex A6 with its monomial labeling (partially indicated).
Let us first clarify our terms. To simplify notation we will associate to any monomial
· · · xinn ∈ R the vector (i1 , i2 , . . . , in ) ∈ Nn and will freely move between notations. We
xi11 xi22
4
define a labeled polyhedral complex to be a polyhedral complex X together with an assignment
aF ∈ Nn to each face F ∈ X such that for all i = 1, 2, . . . , n we have
(aF )i = max{(aG )i : G ⊂ F }.
If X is a labeled polyhedral complex we can consider the ideal M = MX ⊂ K[x1 , x2 , . . . , xn ]
generated by the monomials corresponding to its vertices (as usual we identify an element
α ∈ Nn as the exponent vector of a monomial). The topological space underlying X (with
a chosen orientation) has an associated chain complex FX of k-vector spaces that computes
cellular homology. Since X has monomial labels on each of its cells, we can homogenize
the differentials with respect to this basis and in this way FX becomes a complex of free
modules over the polynomial ring R = K[x1 , x2 , . . . , xn ]. We say that the polyhedral complex
X supports a resolution of the ideal M if FX is in fact a graded free resolution of M . For
more details and examples of cellular resolutions we refer to [14].
For any σ ∈ Nn we let X≤σ denote the subcomplex of X consisting of faces F for which
aF ≤ σ componentwise. We then have the following criteria, also from [14].
Lemma 2. Let X be a labeled polyhedral complex and let M = MX ⊂ K[x1 , . . . , xn ] denote
the associated monomial ideal generated by the vertices. Then X supports a cellular resolution
of M if and only if the complex X≤σ is K-acyclic (or empty) for all σ ∈ Nn . Futhermore,
the resolution is minimal if and only if aF 6= aG for any pair of faces F ( G.
With this criteria in place we can establish the following.
Theorem 3. For each n ≥ 4 the associahedron An , with the monomial labeling described
above, supports a cellular resolution of the edge ideal Jn .
Proof. Let An denote the n-dimensional simplicial associahedron with this monomial labeling. By construction, the vertices of An correspond to the generators of Jn . To show that An
supports a resolution of Jn , according to Lemma 2 it is enough to show that for any σ ∈ Nn
we have that the subcomplex (An )≤σ is K-acyclic.
Let σ ∈ Nn and let (An )≤σ denote the subcomplex of An consisting of all faces with a
monomial labeling that divides σ (as usual, thinking of σ as the exponent vector of the
monomial xσ1 1 xσ2 2 · · · xσnn ). In particular, a face F ∈ An is an element of (An )≤σ if and only if
for every diagonal xi xj ∈ F we have σi > 0 and σj > 0. We claim that (An )≤σ is contractible
(and hence K-acyclic).
Note that since Jn is squarefree we may assume σ has 0/1 entries, and hence we can
identify σ with a subset of [n]. Also, if σi = 1 for all i (so that σ = [n]) then we have
(An )≤σ = An , which is a convex polytope and hence contractible. If σ has fewer than 2
nonzero entries then (An )≤σ is empty. Without loss of generality, we may then assume that
σ1 = 1 and σn = 0. Let j be the largest integer such that j > 2 and σj = 1.
Now, since j < n and {1, j} ⊂ σ we see that the diagonal (1, j) is a vertex of the simplicial
complex (An )≤σ . In fact, (1, j) is an element of every facet of (An )≤σ since no other diagonal
picked up by the elements of σ intersects (1, j). We conclude that (An )≤σ is a cone and
hence contractible.
For n = 5 one can check that this resolution is in fact minimal, but for n ≥ 6 this is no
longer the case. In particular for n ≥ 6 we have faces F ( G in An with the same monomial
label.
5
4. Standard Young Tableaux
It turns out that the number of i-dimensional faces of the associahedron An (the entries of
the face vector of An ) are given by the number of standard Young tableau (SYT) of certain
shapes. Recall that if λ = (λ1 , λ2 , . . . , λk ) is a partition of n, a standard Young tableaux of
shape λ is a filling of the Young diagram of λ with distinct entries {1, 2, . . . , n} such that
rows and columns are increasing (see Example 5).
For 0 ≤ d ≤ n − 3, we let f (n, d) denote the number of ways to choose d diagonals in a
convex n-gon such that no two diagonals intersect in their interior. We see that f (n, d) is
precisely the number of (d − 1)-dimensional faces of the polytope An . A result attributed to
Cayley (according to [19]) asserts that
1
n+d n−3
(2)
f (n, d) =
.
n+d d+1
d
Using the hook length formula one can see that this number is also the number of standard
Young tableaux of shape
(d + 1, d + 1, 1n−d−3 ),
where as usual
1n−d−3 = 1, . . . , 1
| {z }
(n−d−3)-times
denotes a sequence of n − d − 3 entries with value 1. This fact was apparently first observed
by O’Hara and Zelevinsky (unpublished), and a simple bijection was given by Stanley [19].
Example 4. If we take d = n − 3 we obtain
2n − 3 n − 3
1
2(n − 2)
1
=
,
f (n, n − 3) =
2n − 3 n − 2
n−3
(n − 2) + 1 n − 2
the (n − 2)nd Catalan number.
Example 5. If n = 5 and d = 1 the f (5, 1) =
shape λ = (2, 2, 1) are given by
1 2
3 4
5
1 2
3 5
4
1 3
2 4
5
1 6
6 2
2
1
1 3
2 5
4
= 5 standard Young tableaux of
1 4
2 5
3
These correspond to the 5 diagonals of a 5-gon.
It turns out the Betti numbers of the rings R/Jn are also counted by the number of
standard Young tableaux of certain related (sub)shapes. To establish this result we will
employ Hochster’s formula (Theorem 1 from above). Recall that the ring R/Jn can be
recovered as the Stanley-Reisner ring of the n-cycle, thought of as a 1-dimensional simplicial
complex. Note that when |W | < n the only nonzero contribution to Equation (1) comes from
0-dimensional reduced homology, i.e. the number of connected components of the induced
complex on W (minus one).
n
For n ≥ 4, let βd,j
denote the Betti numbers of the ring R/Jn . Equation (1) implies that
n
for d ≥ 1 we have βd,j
= 0 unless d = n − 2 and j = n, or 1 ≤ d < n − 2 and j = d + 1.
6
n
Another application of Equation (1) gives β1,2
=
cases we have the following result.
n
2
n
− n and βn−2,n
= 1. For the remaining
Theorem 6. For all n ≥ 5 and 1 ≤ d ≤ n − 3 the Betti numbers of R/Jn are given by
(3)
n
βdn = βd,d+1
= the number of standard Young tableau of shape (d + 1, 2, 1n−d−3 ).
Proof. We will establish the equality in Equation (3) by showing that for n ≥ 6 and 1 < d <
n − 4 both sides of the equation satisfy the recursion
n−2
(4)
F (n, d) = F (n − 1, d − 1) + F (n − 1, d) +
.
d+1
For the Betti numbers (the left hand side), we use Hochster’s formula. For each d, the
n
via Equation (1) involves subcomplexes given by subsets W of [n]
computation of βd,d+1
of size d + 1. First suppose we have chosen W ⊂ [n] with n ∈ W . Then we recover the
contribution to Equation (1) from the homology of induced subsets of size d in the cycle on
n−1
the vertices [n − 1], namely βd−1,d
. However, if 1 and n − 1 are both not in W then we get
an additional contribution given by the isolated point n. There are n−3
such instances.
d
Next suppose n ∈
/ W . Then again we recover contribution from the homology of induced
n−1
subsets of size d + 1 in the cycle [n − 1]; this quantity is given by βd,d+1
. In this case we have
an additional contribution coming from the subsets W including both
1 and n − 1, since
n−3
as subsets of the n-cycle these will be disconnected. There are d−1 of these. Putting this
together, we have
n−3
n−3
n−2
n−1
n−1
n−1
n−1
n
βd,d+1 = βd−1,d + βd,d+1 +
+
= βd−1,d + βd,d+1 +
,
d
d−1
d
recovering Equation (4).
We next consider the right hand side of Equation (3), namely the number of standard
Young tableaux of shape (d + 1, 2, 1n−d−3 ). Recall that the fillings involve picking entries
one each from the set {1, 2, . . . , n}. If n is an entry in the first row (necessarily in the last
column) then we recover all such fillings from standard Young tableaux of shape
(d, 2, 1n−d−3 ) = ((d − 1) + 1, 2, 1(n−1)−(d−1)−3 ).
If n is the (only) entry in the last row, then we recover all such fillings from standard Young
tableaux of shape (d + 1, 2, 1n−d−4 ) = (d + 1, 2, 1(n−1)−d−3 ). With these counts we miss the
standard tableaux with n as the entry in the second row (necessarily in the second column).
In this case we must have 1 as the entry in the first row, first column, but are free to choose
any increasing sequence of length d to fill the
remaining entries of the first row (with the rest
n−2
of the entries determined). There are d such choices. Adding these three counts gives
us the desired recursion from Equation (4).
5
We next check the initial conditions. For n = 5 Hochster’s formula again gives us β1,2
=
5
β2,3 = 5. One can check (see Example 5) that there are precisely 5 standard Young tableau
of shape (2, 2, 1) and of its conjugate shape (3, 2).
n
For arbitrary n and d = 1 we have β1,2
= n2 − n, given by the number of generators of
Jn . On the other hand, in a standard Young tableau of shape (2, 2, 1n−4 ) we can have any
pair (i, j) with i < j occupy the second row except
(1, 2), (1, 3), . . . , (1, n), or (2, 3). Hence
the number of such fillings is also given by n2 − n.
7
Similarly, for arbitrary n and d = n−3, Hochster’s formula implies that the Betti numbers
n
βn−3,n−2
are given by all choices of n − 2 vertices of the n-gon corresponding to complements
of diagonals (since
these remaining pair of vertices with be disconnected). Hence again
n
n
βn−3,n−2 = 2 − n (this also follows from the fact that the ring R/Jn is Gorenstein and
therefore has a palindromic sequence of Betti numbers). In terms of tableaux, we see that
the shape (n − 2, 2) is conjugate to (2, 2, 1n−4 ) and hence both shapes have the same number
of fillings.
Remark 7. An application of the hook length formula gives an explicit value for the Betti
numbers of R/Jn :
n
d(n − d − 2)
n
(5)
βd,d+1 =
n−1
d+1
After a version this paper was posted on the ArXiv it was pointed out to the author that
this formula had previously been established in [5], with a combinatorial proof given in [7].
Remark 8. As we have seen, the rings R/Jn are Gorenstein and hence the Betti numbers of
R/Jn are palindromic in the sense that
n
βdn = βn−d−2
.
The realization of the Betti numbers of R/Jn in terms of standard Young tableaux (Theorem 6) provides a nice combinatorial interpretation of this property. The partition (d +
1, 2, 1n−d−3 ) is conjugate to the partition ((n−d−2)+1, 2, 1n−(n−d−2)−3 ) = (n−d−1, 2, 1d−1 )
and hence they have the same number of fillings.
Example 9. For n = 6 the resolution of R/J6 can be represented as
0 ← R ←
←
←
← R ← 0.
In each homological degree we have a basis for the free module given by all standard Young
tableaux of the indicated shape. Note that (2, 2, 1, 1) is conjugate to (4, 2).
5. Discrete Morse theory and matchings
As we have seen, the associahedron An (with the monomial labeling described above)
supports a resolution of the ideal Jn . We have also seen that the resolution is not minimal,
and in particular the labeling of An produces distinct faces F ( G with the same monomial
labeling. In fact as n increases the resolution becomes further and further from minimal in
n
the sense that the number of facets of An (a Catalan number, on the order of n43/2 ) dominates
the dimension of the second highest syzygy module of R/Jn (which is on the order of n2 ).
Example 10. Face numbers versus Betti numbers for n = 6, 7, 8, 9 are indicated below.
Here f (n, j) refers to the number of j-dimensional faces in the Associahedron An .
d=0 1 2 3 4
1 9 16 9 1
f (6, d − 1) 1 9 21 14 1
βd6
8
d=0 1 2 3 4 5
1 14 35 35 14 1
f (7, d − 1) 1 14 56 84 42 1
βd7
d=0 1
2
3
4
5 6
1 20 64 90 64 20 1
f (8, d − 1) 1 20 120 300 330 132 1
βd8
d=0 1
2
3
4
5
6 7
1 27 105 189 189 105 27 1
f (9, d − 1) 1 27 225 825 1485 1287 429 1
βd9
5.1. Morse matchings and first syzygies. Batzies and Welker and others (see [1] and
[18]) have developed a theory of algebraic Morse theory that allows one to match faces of a
labeled complex in order to produce resolutions that become closer to minimal. In the usual
combinatorial description of this theory, one must match elements in the face poset of the
labeled complex that have the same monomial labeling. The matching must also satisfy a
certain acyclic condition, described below. We refer to [1] for further details.
A closer analysis of our monomial labeling of An reveals certain faces that must be matched
away in any minimal resolution, in the sense that the associated monomial has the wrong
degree. In particular, since we know that R/Jn has an ‘almost’ linear resolution (as described
above) it must be the case that in any minimal cellular resolution X , each j-dimensional face
of X is labeled by a monomial of degree j + 2 (for j < n − 3). Our labeling of An has the
property that the monomial m associated to a face F is given by the product of the variables
involved in the choice of diagonals, and in particular a ‘properly’ labeled j-dimensional face
corresponds to a subdivision of Cn with j + 1 diagonals involving precisely j + 2 vertices.
This motivates the following.
Definition 11. Suppose S is a subdivision of the n-gon Cn , by which we mean a collection of
d non-crossing diagonals. We will say that S is proper if the set of endpoints of the diagonals
has exactly d + 1 elements (as vertices of Cn ). We will say that S is superproper if uses more
than d + 1 vertices and subproper if it uses less.
1
2
1
3
6
5
4
2
6
1
3
5
4
2
1
3
6
5
4
1
2
6
3
5
4
2
3
6
5
4
Figure 3. For n = 6, the three superproper subdivisions with d = 2 and the
two subproper subdivisions with d = 3. All other subdivisions of the 6-gon are
proper.
In fact we can explicitly describe a (partial) Morse matching on the monomial-labeled face
poset of An that is ‘perfect’ for rank d = 2. A superproper subdivision of an n-cycle with
d = 2 is simply a pair of disjoint diagonals, say E = {ij, k`} with i < j, k < `, and i < k.
9
In the face poset of An we match this 2-face with the 3-face F , where
{ij, k`, j`} if j < k
F =
{ij, k`, i`} otherwise.
A subproper subdivision with d = 3 is an inscribed triangle, say with diagonals {ij, ik, jk},
i < j < k. We match this face with the (proper) 1-face {ij, jk}. Recall that the Hasse
diagram of the face poset of An is a graph with vertices given by all faces of An , and with
edges given by all cover relations X ≺ Y . It is easy to that our association is a matching
of the Hasse diagram of the face poset of An , and it is clearly ‘algebraic’ in the sense that
matched faces have the same monomial labeling.
As is typical we think of the Hasse diagram as a directed graph with the orientation on
a matched edge pointing up (increasing dimension), and with all unmatched edges pointing
down. The collection of faces not involved in the matching are called the ‘critical cells’, they
form a subposet of the original poset. The main theorem of (algebraic) discrete Morse theory
[1] says that if we have an acyclic (algebraic) matching on the Hasse diagram of a cellular
resolution, then the critical cells form a monomial-labeled CW-complex that also supports a
cellular resolution. In this way one can obtains a resolution that is closer to being minimal.
In our case we have the following result.
Proposition 12. For all n ≥ 6, the matching on the monomial labeled face poset of An
described above is acyclic. Furthermore, the number of unmatched (critical) edges is given
by β2n .
Proof. We first make the simple observation that if F is any 2-face of An corresponding
to a subproper subdivision (in other words an inscribed triangle), there for any 1-face E
with E ≺ F we must have that E is a path of length 2 (a proper subdivision with d = 2).
Similarly, if E is a path of length 2 and E ≺ F is an upward oriented edge then it must
be the case that F is an inscribed triangle with the same vertex set as E. This implies
that there cannot be any cycles in the oriented face poset involving proper subdivisions with
d = 2 (paths of length 2).
Next suppose E ≺ F is an upward oriented edge in the face poset of An where E consists
of two disjoint diagonals (a superproper subdivision with d = 2). Then according to our
matching it must be the case that F is a path of length 3. To form a cycle in the face poset
there must be some downward edge from F to E 0 with E 0 ≺ F . But then according to our
matching it must be the case that E 0 is a path of length 2. Hence our observation from the
previous paragraph implies that no cycles exist. We conclude that the matching is acyclic.
We next count the unmatched edges. First observe that the number of proper subdivisions
of an n-gon with 2 diagonals is given by n(n−3)(n−4)
. To see this, note that the diagonals
2
involved in such a subdivision must form a path of length 3. Once
we designate the middle
n−3
vertex in this path (of which there are n choices) we have 2 choices for the remaining
two vertices. Next we claim that the the number of subproper subdivisions of an n-gon with
d = 3 diagonals (necessarily forming an inscribed triangle) is given by n(n−4)(n−5)
. To see
6
this, we first count inscribed triangles with ordered vertex set (v1 , v2 , v3 ). We are free to
choose the first vertex v1 from among the n nodes of the cycle. For the second vertex we
have two cases. If we choose v2 from among the two vertices that are distance 2 from v1 ,
we are left with (n − 5) choices for v3 . If we choose v2 from among the vertices more than
distance 2 from v1 (of which there are n − 5 choices), we are then left with n − 6 choices for
10
v3 . In total there are
n(2(n − 5) + (n − 5)(n − 6)) = n(n − 4)(n − 5)
inscribed triangles with the ordered vertex set. Dividing out by 6 to forget the ordering
gives us the desired count. As described above, we match each of the d = 2 superproper
subdivisions with a d = 3 proper subdivision, and we match each of the d = 3 subproper
subdivisions with a d = 2 proper subdivision. Hence after matching the number of critical
edges is given by
n(n − 3)(n − 4) n(n − 4)(n − 5)
n 2(n − 4)
−
=
,
2
6
3
n−1
which is precisely β2n (see Remark 7). This completes the proof.
Hence our simple matching leaves precisely the number of critical 1-cells that we require.
The rank of the first free module in the resulting cellular resolution will be equal to the rank
of the first syzygy module of R/Jn .
Example 13. For n = 6 this matching in fact leads to a minimal resolution of J6 . In this
case we have three superproper subdivisions with d = 2 (namely {13, 46}, {15, 24}, and
{26, 35}), and two subproper subdivisions with d = 3 (namely {13, 15, 35} and {24, 26, 46}).
x1x3
x1x5
x3x5
x1x4
x2x5
x3x6
x2x4
x4x6
x2x6
Figure 4. The monomial labeled A6 with five pairs of faces matched (the
shaded faces are the improper subdivisions). The resulting complex on the
right supports a minimal resolution of J6 .
We remark that the procedure described above can be extended to the case n = 7. We
leave the details to the reader but point out that in this case we have:
• 14 superproper subdivisions with d = 2 (pairs of disjoint diagonals), corresponding
to 14 edges of A7 that we match with 2-faces,
• 7 subproper subdivisions with d = 3 (inscribed triangles in a 7-gon), corresponding
to seven 2-faces that each get matched with an edge,
• 14 superproper subdivisions with d = 3 (forests consisting of three edges and two
components), corresponding to seven 2-faces that get matched to a 3-face,
11
• 14 subproper subdivisions with d = 4 (inscribed triangles with a pendant edge),
corresponding to fourteen 3-faces that get matched down to a 2-face.
The resulting CW-complex has 56 − 14 − 7 = 35 edges, 84 − 14 − 7 − 14 − 14 = 35
two-dimensional faces, and 42 − 28 = 14 three-dimensional faces, as desired. Unfortunately
we do not know how to extend this matching procedure in general; see the next Section for
some comments regarding this.
5.2. An involution of the associahedron tableaux. Recall that the faces of the associahedron An are counted by standard Young tableaux of certain shapes, while the Betti
numbers of Jn are counted by standard Young tableaux of certain subshapes. Again motivated by discrete Morse theory this leads to ask whether we can find a matching on the
set of associahedron tableaux such that the unmatched elements correspond to the Betti
numbers of Jn . This matching should have the property that two matched tableaux differ in
cardinality by one. Let us emphasize that since we do not have a poset structure on these
elements we are not at this pointing searching for a M orse matching. Let us first fix some
notation.
Definition 14. For fixed n ≥ 4 and 1 ≤ d ≤ n − 3, we call the collection of standard Young
tableaux of shape (d + 1, d + 1, 1n−d−3 ) the associahedron tableaux (denoted by An ), and the
n−d−3
standard S
Young tableaux
) the syzygy tableaux (denoted by Sn ).
S of shape (d + 1, 2, 1
Let A = An and S = Sn .
Note that an element of An has n + d − 1 boxes, whereas an element of Sn has n boxes. If
X ∈ An is an associahedron tableux with largest entries in the second row in the positions
(∗, ∗, n + 1, n + 2, . . . , n + d − 1) then it naturally becomes a syzygy tableau by just removing
those boxes. In particular we say that these particular associahedron tableau restrict to
syzygy tableaux, and in this way we have a natural inclusion Sn ⊂ An .
Example 15. The associahedron tableau on the left restricts to a syzygy tableau, whereas
the associahedron tableau on the right does not. Here n = 7 and d = 3.
∗ ∗ ∗ ∗
∗ ∗ ∗ ∗
∗ ∗ 8 9
∗ ∗
∗
→ ∗
∗ ∗ ∗ ∗
∗ ∗ ∗ 8
9
We next describe an involution on the set A such that the fixed elements are precisely
the elements that restrict to S. If X is a standard young tableau we use |X| to denote the
number of boxes in the underlying partition.
Proposition 16. There exists an involution σ on the set A such that the fixed point set of
σ is precisely the set of tableaux in A that restrict to S. Furthermore, if X ∈ A such that
σ(X) 6= X, we have |σ(X)| = |X| ± 1.
Proof. Suppose X ∈ An is an associahedron tableau. If X restricts to a syzygy tableau we
set σ(X) = X. Otherwise some element of {n + 1, n + 2, . . . , n + d − 1} is not in the second
row of X; let i be the largest element with this property. Then i must be the last element
of the first row, or else the bottom element in the first column.
In the latter case (i is the bottom most element of first column) we bring that element i
to the first row, and add the element n + d to the end of the second row. This defines σ(X).
In the former case (i is the last element of the first row) we obtain σ(X) by bringing that
12
element down to the bottom of the first column and deleting the last element of the second
row (which must be n + d − 1). It is clear that σ(σ(X)) = X.
Example 17. An example of the involution matching an associahedron tableau of shape
(3, 3, 1, 1) with one of shape (4, 4, 1) is given by the following.
∗ ∗ ∗
∗ ∗ ∗
∗ ∗ ∗ 8
∗
∗ ∗ ∗ 9
8
↔ ∗
6. Further questions
We end with a number of questions that arise from our study. As we have seen in Section
4, the number f (n, d) of dissections of an n-gon using d (non-crossing) diagonals is well
understood, and is given by the number of standard Young tableaux of shape (d + 1, d +
1, 1n−d−3 ). In the context of enumerating the Betti numbers of the ideal Jn we were interested
in subdivisions that involved a fixed number of vertices. Define f (n, d, j) to be the number
of ways to choose d non-crossing diagonals in a convex n-gon such that the set of endpoints
consists of precisely j vertices of the n-gon.
Question 18. Is there a nice formula for f (n, d, j)? Can it be related to the standard Young
tableaux of shape (d + 1, d + 1, 1n−d−3 )?
We note that if we take d = n − 3 then varying j gives a refinement of the Catalan numbers
which (as far as we know) has not appeared elsewhere. The first few refinements are
14 = 2 + 12, 42 = 14 + 28, 132 = 4 + 64 + 64, . . .
A related question would be to consider those subdivisions for which the collection of
diagonals forms a (connected) tree, since this is likely the more relevant property in the
context of syzygies. For n ≤ 7 it so happens that the proper subdivisions correspond to
those collections of diagonals that form a tree. However, for n = 8 there exist proper
subdivisions that are not trees: for example if d = 4 we can take 3 diagonals to form a
triangle with vertices {1, 3, 5} along with one disconnected diagonal {6, 8}, in total using 5
vertices of the 8-gon.
Question 19. How many dissections of an n-gon with d diagonals have the property that
the set of diagonals forms a tree?
In our quest for a Morse matching on the monomial labeled face poset of the associahedron
An we were unable to employ Stanley’s bijection between faces of An and standard Young
tableaux. As mentioned above, the difficulty arises as the bijection given in [19] is recursively
defined and involves certain choices. However, the fact that the face poset of An is labeled
by standard Young tableaux suggests that there might be a meaningful poset structure on
the set of all standard Young tableaux (or at least the set of Associahedron tableaux). The
hope would be that this poset structure extends the partial order given by the involution on
A described in the proof of Proposition 16. Hence the poset should be graded by the number
of boxes in the underlying partition, but will not restrict to Young’s lattice if one forgets
the fillings. We refer to Example 17 for a example of a cover relation between two standard
Young tableaux such that the underlying partitions are not related in Young’s lattice.
13
Question 20. Does there exist a meaningful poset structure on the set of standard Young
tableaux, consistent with the conditions described above?
Finally, we see in Figure 4 that a minimal resolution of J6 is supported on a 3-dimensional
polytope. As we mentioned the construction there was a bit ad hoc but it does lead us to
following:
Question 21. Does the ideal Jn have a minimal cellular resolution supported on a (necessarily (n − 3)-dimensional) polytope?
Work in this direction (along with some further generalizations) is currently being pursued
by Engström and Linusson [10].
Acknowledgements. We thank Ken Baker for his assistance with the figures, and Alex
Engström, Jakob Jonsson, and Michelle Wachs for helpful conversations. Alex and I first
realized the potential connection to standard Young tableaux after inputting the Betti numbers of Jn into OEIS [16] some years ago. Thanks also to the anonymous referee for a careful
reading.
References
[1] E. Batzies, V. Welker, Discrete Morse theory for cellular resolutions, J. Reine Angew. Math. 543 (2002),
147–168.
[2] D. Bayer, B. Sturmfels, Cellular resolutions of monomial modules, J. Reine Angew. Math. 502 (1998),
123–140.
[3] J. Biermann, Cellular structure on the minimal resolution of the edge ideal of the complement of the
n-cycle, submitted.
[4] B. Braun, J. Browder, and S. Klee, Cellular resolutions of ideals defined by nondegenerate simplicial
homomorphisms, Israel J. Math. (2012), 1–24.
[5] W. Bruns , T. Hibi, Cohen-Macaulay partially ordered sets with pure resolutions, European J. Combin.
19 (1998), 779–785.
[6] C. Ceballos, F. Santos, G. Ziegler, Many non-equivalent realizations of the associahedron, Combinatorica
35 (2015): 513. doi:10.1007/s00493-014-2959-9
[7] S. Choi, J.S. Kim, A combinatorial proof of a formula for Betti numbers of a stacked polytope, Electron.
J. Combin., 17 (2010), Research Paper # R9.
[8] A. Dochtermann, A. Engström, Cellular resolutions of cointerval ideals, Math. Z. 270 (2012), no. 1-2,
145–163.
[9] A. Dochtermann, F. Mohammadi, Cellular resolutions from mapping cones, J. Combin. Theory Ser. A
128 (2014), 180–206.
[10] A. Engström, S. Linusson, personal communication, 2016.
[11] C. Francisco, J. Mermin, J. Schweig, Catalan numbers, binary trees, and pointed pseudotriangulations,
European J. Combin. 45 (2015), 85–96.
[12] A. Goodarzi, Cellular structure for the Herzog-Takayama Resolution, J. Algebr. Comb. 41 (1) (2015),
21–28.
[13] J. Mermin, The Eliahou-Kervaire resolution is cellular, J. Commut. Algebra 2 (2010), no. 1, 55–78.
[14] E. Miller, B. Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, vol. 227.
Springer, New York (2005).
[15] U. Nagel, V. Reiner, Betti numbers of monomial ideals and shifted skew shapes, Electron. J. Combin.
16 (2009), no. 2, Special volume in honor of Anders Björner, Research Paper 3, 59 pp.
[16] The On-Line Encyclopedia of Integer Sequences, published electronically at http://oeis.org, 2014.
[17] A. Sinefakopoulos, On Borel fixed ideals generated in one degree, J. Algebra 319 (2008), no. 7, 2739–
2760.
[18] E. Sköldberg, Morse theory from an algebraic viewpoint, Trans. Amer. Math. Soc. 358 (2006), 115–129.
14
[19] R. Stanley, Polygon dissections and standard Young tableaux, J. Combin. Theory Ser. A 76 (1996),
175–177.
[20] S. Sturgeon, personal communication, 2013.
15
| 0math.AC
|
arXiv:1605.00050v1 [cs.SY] 30 Apr 2016
DSP IMPLEMENTATION OF A DIRECT ADAPTIVE FEEDFOWARD CONTROL
ALGORITHM FOR REJECTING REPEATABLE RUNOUT IN HARD DISK DRIVES
Jinwen Pan
Prateek Shah
Roberto Horowitz
Department of Mechanical Engineering Department of Mechanical Engineering Department of Mechanical Engineering
University of California, Berkeley
University of California, Berkeley
University of California, Berkeley
Berkeley, California 94720
Berkeley, California 94720
Berkeley, California 94720
Email: [email protected]
Email: [email protected]
Email: [email protected]
ABSTRACT
A direct adaptive feedforward control method for tracking repeatable runout (RRO) in bit patterned media recording
(BPMR) hard disk drives (HDD) is proposed. The technique estimates the system parameters and the residual RRO simultaneously and constructs a feedforward signal based on a known regressor. An improved version of the proposed algorithm to avoid
matrix inversion and reduce computation complexity is given.
Results for both MATLAB simulation and digital signal processor (DSP) implementation are provided to verify the effectiveness
of the proposed algorithm.
1
are briefly listed here: (1) RRO profile is unknown; (2) RRO
frequency spectrum can spread beyond the bandwidth of servo
system, therefore, it will be amplified by the feedback controller;
(3) RRO spectrum contains many harmonics of the spindle frequency (e.g. ∼ 200 harmonics) that should be attenuated, which
increases the computational burden in the controller; (4) RRO
profile is changing from track to track (i.e. it is varying on the
radial direction); (5) HDD servo dynamics changes from drive to
drive and by temperature [9].
The remainder of this paper is organized as follows. Section 2 presents our direct adaptive feedforward control algorithm
and section 3 shows the real time DSP implementation results.
INTRODUCTION
Data bits are ideally written on concentric circular tracks in conventional HDDs that use magnetic disks with continuous media. This process is different in bit patterned media recording
since data should be written on tracks with predetermined shapes,
which are created by lithography on the disk [1, 2]. As shown
in Fig. 1 the trajectories that are required to be followed by the
servo system in BPMR are servo tracks, which are characterized
by the servo sectors written on the disk. Deviation of a servo
track from an ideal circular shape is called RRO. Therefore, the
servo controller in BPMR has to follow the RRO which is unknown in the time of design, and as a result the servo control
methodologies used for conventional drives [3] cannot be applied to BPMR directly. In our prior works, we proposed indirect
adaptive control methods for mechatronic devices to compensate
for unknown disturbances (such as RRO) [4, 5, 6] and dynamics
mismatches [7, 8]. In this paper, we propose a direct adaptive
control method to address challenges specific to BPMR which
Servo tracks
Conventional media
Data tracks
Bit-patterned media
FIGURE 1: Servo Track (dotted blue) and data track (solid red)
in conventional and bit-patterned media
2
Control Design
The architecture that is considered for the servo control system
is shown in Fig. 2. An add-on feedforward controller is designed
for HDD. Without loss of generality, we chose VCM as an example. Here, R is the transfer function from VCM input to PES,
0 Submitted to ASME 2016 Conference on Information Storage and Processing Systems
1
Copyright c 2016 by ASME
uf
e
R
ua
ua
ue
r
R
−n̂b with n̂ the estiwhere B̂ q−1 = b̂1 (k) q−1
b
+ · · · + b̂Tn̂b (k) q
−1
mate of nb . Since A q
r (k) = θR̄ φr (k) with θR̄ an unknown
vector and φr (k) a nonzero vector, from Eq. (5) we have
e
θ̂D (k) + θR̄ (k) = D−1
(k) θ̂M (k) ,
B̂
where DB̂ (k)
can be formed based on the magnitude and phase
of B̂ e− jωi , i = 1, 2, · · · , nr with nr the number of frequencies to
cancel. The updating law for θ̂D (k) is
FIGURE 2: Control architecture.
ue is an exogenous excitation signal, ua is the feedforward signal, r is the unknown RRO with known frequencies and e is the
PES. We aim to design an adaptive controller that generates ua
in order to fade the frequency contents of the error signal e at selective frequencies which correspond to the harmonics of spindle
frequency (120Hz and its harmonics in our case).
θ̂D (k + 1) = θ̂D (k) − αD−1
(k) θ̂M (k) .
B̂
(6)
The inverse of DB̂ (k) involves inverting the estimated magnitudes that might be very small in transition, especially when θ̂B
is initialized by zeros. In that case, any small fluctuation of θ̂B
can cause large transient
error. A smoothing on magnitude and
phase of B̂ e− jωi has to be designed to relax transient errors.
The basic direct adaptive feedforward control algorithm is summarized in Table. 1.
2.1
Basic Direct Adaptive Feedforward Control
From Fig. 2, the PES can be written as
e (k) = R q−1 (ue (k) + ua (k)) + r (k) ,
where R q−1 = B q−1 /A q−1 and can be expanded as
T
e (k) = A∗ q−1 e (k) + B q−1 ue (k) + θM
φr (k) ,
(1)
where A
q−1 = 1−A∗ q−1 , A∗ q−1 = a1 q−1 +· · ·+ana q−na ,
B q−1 = b1 q−1 + · · · + bnb q−nb and the residual error
T
θM
φr (k) = B q−1 ua (k) + A q−1 r (k) ,
(2)
0.
1.
2.
3.
4.
5.
where φr (k) is the regressor for RRO with known frequencies.
In regressor form, PES is
Initialize the regressors φe (k), φue (k) and φr (k);
Apply ue (k) and ua (k) to VCM;
Subtract ê(k) from PES to determine the estimate error ẽ(k);
Update the parameters θ̂ (k) using Eq. (3);
Update the matrix DB̂ (k) from θ̂B (k) and compute its inverse;
Update θ̂D (k) using Eq. (6) and compute ua (k) from Eq. (4).
TABLE 1: Basic Direct Adaptive Feedforward Control
T
e (k) = θAT φe (k) + θBT φue (k) + θM
φr (k) ,
and its estimation
ê (k) = θ̂AT
(k) φe (k) + θ̂BT
2.2
Improved Direct Adaptive Feedforward Control
As mentioned earlier, computational complexity of inverting DB̂ (k) grows as the number of frequencies increases which
is a crucial burden in DSP implementation. In this section, we
will provide an improved version to avoid matrix inversion. By
applying “swapping lemma” to (5), we have
T
(k) φue (k) + θ̂M
(k) φr (k) ,
where θAT = [a1 , a2 , · · · , ana ], φe (k), θBT = b1 , b2 , · · · , bnb and
φue (k) are regressors for e(k) and ue (k). θ̂A , θ̂B and θ̂M are the
estimates of θA , θB and θM and the updating law is
θ̂ (k + 1) = θ̂ (k) + K (k) φ (k) ẽ (k) / 1 + φ T (k) φ (k) , (3)
T (k) T ,
where ẽ (k) = e (k) − ê (k), θ̂ (k) = θ̂AT (k) , θ̂BT (k) , θ̂M
T (k) , φ T (k) T and K(k) is a decreasing gain.
φ (k) = φeT (k) , φue
r
Eq. (3) indicates that both of the system and the residual RRO
are estimated simultaneously. The feedforward control signal is
constructed using the same regressor as the RRO yielding
T
(k) φr (k) ,
θ̂DT (k) B̂ q−1 φr (k) + A q−1 r (k) = θ̂M
(7)
therefore, the updating law for θ̂D (k) is
T
(k) φr (k) .
θ̂D (k + 1) = θ̂D (k) − α B̂ q−1 φr (k) θ̂M
(8)
Note that in (8) no matrix inverse is required. The improved direct adaptive feedforward control algorithm is summarized in Table. 2 where the first three steps are the same as those in Table. 1.
To be noted here, the proposed direct adaptive feedforward control algorithm and its improved version can be directly extended
to the second-stage actuator which is mili-actuator (MA) responsible for high frequency RRO.
ua (k) = θ̂DT (k) φr (k) .
(4)
In Eq. (2), using θ̂M (k) and B̂ q−1 instead of θM and B(q−1 ),
approximately we have
T
B̂ q−1 θ̂DT (k) φr (k) + A q−1 r (k) = θ̂M
(k) φr (k) ,
(5)
2
Copyright c 2016 by ASME
Acknowledgment
Am plitude Spectrum (nm )
0.5
Adaptive Controller Off
Adaptive Controller On
0.4
Financial support for this study was provided by a grant from the
Advanced Storage Technology Consortium (ASTC).
0.3
REFERENCES
[1] Shahsavari, B., Keikha, E., Zhang, F., and Horowitz, R.,
2015. “Adaptive repetitive control design with online secondary path modeling and application to bit-patterned media
recording”. Magnetics, IEEE Transactions on, 51(4), pp. 1–
8.
[2] Keikha, E., Shahsavari, B., Al-Mamun, A., and Horowitz,
R., 2013. “A probabilistic approach to robust controller design for a servo system with irregular sampling”. In Control
and Automation (ICCA), 2013 10th IEEE International Conference on, IEEE, pp. 1790–1795.
[3] Kempf, C., Messner, W., Tomizuka, M., and Horowitz, R.,
1993. “Comparison of four discrete-time repetitive control algorithms”. IEEE Control Systems Magazine, 13(6),
pp. 48–54.
[4] Shahsavari, B., Keikha, E., Zhang, F., and Horowitz, R.,
2014. “Repeatable runout following in bit patterned media
recording”. In ASME 2014 Conference on Information Storage and Processing Systems, American Society of Mechanical Engineers, pp. V001T03A001–.
[5] Shahsavari, B., Keikha, E., Zhang, F., and Horowitz, R.,
2014. “Adaptive repetitive control using a modified filteredx lms algorithm”. In ASME 2014 Dynamic Systems and
Control Conference, American Society of Mechanical Engineers, pp. V001T13A006–.
[6] Shahsavari, B., Pan, J., and Horowitz, R., 2016. “Adaptive
rejection of periodic disturbances acting on linear systems
with unknown dynamics”. arXiv preprint arXiv:1603.05361.
[7] Zhang, F., Keikha, E., Shahsavari, B., and Horowitz, R.,
2014. “Adaptive mismatch compensation for vibratory gyroscopes”. In Inertial Sensors and Systems (ISISS), 2014
International Symposium on, IEEE, pp. 1–4.
[8] Zhang, F., Keikha, E., Shahsavari, B., and Horowitz, R.,
2014. “Adaptive mismatch compensation for rate integrating vibratory gyroscopes with improved convergence
rate”. In ASME 2014 Dynamic Systems and Control
Conference, American Society of Mechanical Engineers,
pp. V003T45A003–V003T45A003.
[9] Bagherieh, O., Shahsavari, B., and Horowitz, R., 2015. “Online identification of system uncertainties using coprime factorizations with application to hard disk drives”. In ASME
2015 Dynamic Systems and Control Conference, American Society of Mechanical Engineers, pp. V002T23A006–
V002T23A006.
0.2
0.1
0
20
40
60
80
100
Ha rm o nic
120
140
160
VCM Inp. (V)
FIGURE 3: Spectrum comparison.
0.5
0
-0.5
0
50
100
150
200
250
300
350
200
250
300
350
MA Inp. (V)
Step
1
0
-1
0
50
100
150
Step
FIGURE 4: Feedforward signal for VCM and MA.
..
.
4. Construct the matrix B̂ q−1 (k) φr (k) and compute residual
T (k) φ (k);
error θ̂M
r
5. Update θ̂D (k) using Eq. (8) and compute ua (k) from Eq. (4).
TABLE 2: Improved Dreict Adaptive Feedforward Control
3
Experiment Results and Conclusion
We implement both of the two algorithms in MATLAB simulation and the real time experiment setup on dual-stage HDD. In
simulation, R and RRO together with NRRO are modeled from
real system measurement data. Since the simulation and experiment results were very close, only experiment results using the
improved version are shown in Fig. 3, where RRO is reduced
to NRRO level. In simulation as well as in experiments, VCM
was responsible for the low frequency RRO (harmonics up to 58),
while MA was responsible for the high frequency RRO (harmonics from 59 to 173). As a result, the feedforward control signal
in one disk revolution shown in Fig. 4 for the VCM consists of
low frequency contents while for the MA it has high frequency
components.
3
Copyright c 2016 by ASME
| 3cs.SY
|
c Springer-Verlag, 2014. This is the author’s version of the work. It is posted here by permission
of Springer-Verlag for your personal use. Not for redistribution. The final publication is published
in the proceedings of the 3rd Conference on Principles of Security and Trust (POST ’14), and is
available at link.springer.com
arXiv:1401.4339v2 [cs.CR] 21 Jan 2014
Information Flow Control in WebKit’s
JavaScript Bytecode
Abhishek Bichhawat1 , Vineet Rajani2 , Deepak Garg2 , and Christian Hammer1
1
Saarland University, Germany
2
MPI-SWS, Germany
Abstract. Websites today routinely combine JavaScript from multiple sources, both trusted and untrusted. Hence, JavaScript security is
of paramount importance. A specific interesting problem is information
flow control (IFC) for JavaScript. In this paper, we develop, formalize
and implement a dynamic IFC mechanism for the JavaScript engine of a
production Web browser (specifically, Safari’s WebKit engine). Our IFC
mechanism works at the level of JavaScript bytecode and hence leverages years of industrial effort on optimizing both the source to bytecode
compiler and the bytecode interpreter. We track both explicit and implicit flows and observe only moderate overhead. Working with bytecode
results in new challenges including the extensive use of unstructured
control flow in bytecode (which complicates lowering of program context
taints), unstructured exceptions (which complicate the matter further)
and the need to make IFC analysis permissive. We explain how we address these challenges, formally model the JavaScript bytecode semantics
and our instrumentation, prove the standard property of terminationinsensitive non-interference, and present experimental results on an optimized prototype.
Keywords: Dynamic information flow control, JavaScript bytecode, taint
tracking, control flow graphs, immediate post-dominator analysis
1
Introduction
JavaScript (JS) is an indispensable part of the modern Web. More than 95% of
all websites use JS for browser-side computation in Web applications [1]. Aggregator websites (e.g., news portals) integrate content from various mutually
untrusted sources. Online mailboxes display context-sensitive advertisements.
All these components are glued together with JS. The dynamic nature of JS
permits easy inclusion of external libraries and third-party code, and encourages
a variety of code injection attacks, which may lead to integrity violations. Confidentiality violations like information stealing are possible wherever third-party
code is loaded directly into another web page [2]. Loading third-party code into
separate iframes protects the main frame by the same-origin policy, but hinders
interaction that mashup pages crucially rely on and does not guarantee absence
of attacks [3]. Information flow control (IFC) is an elegant solution for such
problems. It ensures security even in the presence of untrusted and buggy code.
IFC for JS differs from traditional IFC as JS is extremely dynamic [3,1], which
makes sound static analysis difficult.
Therefore, research on IFC for JS has focused on dynamic techniques. These
techniques may be grouped into four broad categories. First, one may build
an IFC-enabled, custom interpreter for JS source [4,5]. This turns out to be extremely slow and requires additional code annotations to handle semi-structured
control flow like exceptions, return-in-the-middle, break and continue. Second,
we could use a black-box technique, wherein an off-the-shelf JS interpreter is
wrapped in a monitor. This is nontrivial, but doable with only moderate overhead and has been implemented in secure multi-execution (SME)[6,7]. However,
because SME is a black-box technique, it is not clear how it can be generalized
beyond non-interference [8] to handle declassification [9,10]. Third, some variant
of inline reference monitoring (IRM) might inline taint tracking with the client
code. Existing security systems for JS with IRM require subsetting the language
in order to prevent dynamic features that can invalidate the monitoring process.
Finally, it is possible to instrument the runtime system of an existing JS engine,
either an interpreter or a just-in-time compiler (JIT), to monitor the program
on-the-fly. While this requires adapting the respective runtime, it incurs only
moderate overhead because it retains other optimizations within the runtime
and is resilient to subversion attacks.
In this work, we opt for the last approach. We instrument a production
JS engine to track taints dynamically and enforce termination-insensitive noninterference [11]. Specifically, we instrument the bytecode interpreter in WebKit,
the JS engine used in Safari and other open-source browsers. The major benefit
of working in the bytecode interpreter as opposed to source is that we retain
the benefits of these years of engineering efforts in optimizing the production
interpreter and the source to bytecode compiler.
We describe the key challenges that arise in dynamic IFC for JS bytecode (as
opposed to JS source), present our formal model of the bytecode, the WebKit JS
interpreter and our instrumentation, present our correctness theorem, and list
experimental results from a preliminary evaluation with an optimized prototype
running in Safari. In doing so, our work significantly advances the state-of-the-art
in IFC for JS. Our main contributions are:
– We formally model WebKit’s bytecode syntax and semantics, our instrumentation for IFC analysis and prove non-interference. As far as we are aware,
this is the first formal model of bytecode of an in-production JS engine. This
is a nontrivial task because WebKit’s bytecode language is large (147 bytecodes) and we built the model through a careful and thorough understanding
of approximately 20,000 lines of actual interpreter code.3
3
Unlike some prior work, we are not interested in modeling semantics of JS specified
by the ECMAScript standard. Our goal is to remain faithful to the production
bytecode interpreter. Our formalization is based on WebKit build #r122160, which
was the last build when we started our work.
– Using ideas from prior work [12], we use on-the-fly intra-procedural static
analysis of immediate post-dominators to restrict overtainting, even with
bytecode’s pervasive unstructured conditional jumps. We extend the prior
work to deal with exceptions. Our technique covers all unstructured control
flow in JS (including break and continue), without requiring additional code
annotations of prior work [5] and improves permissiveness.
– To make IFC execution more permissive, we propose and implement a bytecode-specific variant of the permissive-upgrade check [13].
– We implement our complete IFC mechanism in WebKit and observe moderate overheads.
Limitations We list some limitations of our work to clarify its scope. Although
our instrumentation covers all WebKit bytecodes, we have not yet instrumented
or modeled native JS methods, including those that manipulate the Document
Object Model (DOM). This is ongoing work, beyond the scope of this paper.
Like some prior work [4], our sequential non-interference theorem covers only
single invocations of the JS interpreter. In reality, JS is reactive. The interpreter
is invoked every time an event (like a mouse click) with a handler occurs and
these invocations share state through the DOM. We expect that generalizing to
reactive non-interference [14] will not require any instrumentation beyond what
we already plan to do for the DOM. Finally, we do not handle JIT-compilation
as it is considerably more engineering effort. JIT can be handled by inlining our
IFC mechanism through a bytecode transformation.
Due to lack of space, several proofs and details of the model have been omitted
from this paper. They can be found in the technical appendix (Section 8).
2
Related Work
Three classes of research are closely related to our work: formalization of JS semantics, IFC for dynamic languages, and formal models of Web browsers. Maffeis
et al. [15] present a formal semantics for the entire ECMA-262 specification, the
foundation for JS 3.0. Guha et al. [16] present the semantics of a core language
which models the essence of JS and argue that all of JS 3.0 can be translated to
that core. S5 [17] extends [16] to include accessors and eval. Our work goes one
step further and formalizes the core language of a production JS engine (WebKit), which is generated by the source-to-bytecode compiler included in WebKit.
Recent work by Bodin et al. [18] presents a Coq formalization of ECMAScript
Edition 5 along with an extracted executable interpreter for it. This is a formalization of the English ECMAScript specification whereas we formalize the JS
bytecode implemented in a real Web browser.
Information flow control is an active area of security research. With the
widespread use of JS, research in dynamic techniques for IFC has regained momentum. Nonetheless, static analyses are not completely futile. Guarnieri et
al. [19] present a static abstract interpretation for tracking taints in JS. However, the omnipresent eval construct is not supported and this approach does
not take implicit flows into account. Chugh et al. propose a staged information
flow approach for JS [20]. They perform server-side static policy checks on statically available code and generate residual policy-checks that must be applied
to dynamically loaded code. This approach is limited to certain JS constructs
excluding dynamic features like dynamic field access or the with construct.
Austin and Flanagan [21] propose purely dynamic IFC for dynamically-typed
languages like JS. They use the no-sensitive-upgrade (NSU) check [22] to handle implicit flows. Their permissive-upgrade strategy [13] is more permissive
than NSU but retains termination-insensitive non-interference. We build on the
permissive-upgrade strategy. Just et al. [12] present dynamic IFC for JS bytecode
with static analysis to determine implicit flows precisely even in the presence of
semi-unstructured control flow like break and continue. Again, NSU is leveraged to prevent implicit flows. Our overall ideas for dealing with unstructured
control flow are based on this work. In contrast to this paper, there was no formalization of the bytecodes, no proof of correctness, and implicit flow due to
exceptions was ignored.
Hedin and Sabelfeld propose a dynamic IFC approach for a language which
models the core features of JS [4], but they ignore JS’s constructs for semi-structured control flow like break and continue. Their approach leverages a dynamic
type system for JS source. To improve permissiveness, their subsequent work [23]
uses testing. It detects security violations due to branches that have not been
executed and injects annotations to prevent these in subsequent runs. A further
extension introduces annotations to deal with semi-structured control flow [5].
Our approach relies on analyzing CFGs and does not require annotations.
Secure multi-execution (SME) [6] is another approach to enforcing noninterference at runtime. Conceptually, one executes the same code once for each
security level (like low and high) with the following constraints: high inputs are
replaced by default values for the low execution, and low outputs are permitted
only in the low execution. This modification of the semantics forces even unsafe
scripts to adhere to non-interference. FlowFox [7] demonstrates SME in the context of Web browsers. Executing a script multiple times can be prohibitive for
a security lattice with multiple levels. Further, all writes to the DOM are considered publicly visible output, while tainting allows persisting a security label
on DOM elements. It is also unclear how declassification may be integrated into
SME. Austin and Flanagan [24] introduce a notion of faceted values to simulate
multiple executions in one run. They keep n values for every variable corresponding to n security levels. All the values are used for computation as the program
proceeds but the mechanism enforces non-interference by restricting the leak of
high values to low observers.
Browsers work reactively; input is fed to an event queue that is processed
over time. Input to one event can produce output that influences the input to
a subsequent event. Bohannon et al. [14] present a formalization of a reactive
system and compare several definitions of reactive non-interference. Bielova et
al. [25] extend reactive non-interference to a browser model based on SME. This
is currently the only approach that supports reactive non-interference for JS. We
will extend our work to the reactive setting as the next step.
Finally, Featherweight Firefox [26] presents a formal model of a browser based
on a reactive model that resembles that of Bohannon et al. [14]. It instantiates
the consumer and producer states in the model with actual browser objects like
window, page, cookie store, mode, connection, etc. Our current work entirely
focuses on the formalization of the JS engine and taint tracking to monitor
information leaks. We believe these two approaches complement each other and
plan to integrate such a model into our future holistic enforcement mechanism
spanning JS, the DOM and other browser components.
3
Background
We provide a brief overview of basic concepts in dynamic enforcement of information flow control (IFC). In dynamic IFC, a language runtime is instrumented
to carry a security label or taint with every value. The taint is an element of a
pre-determined lattice and is an upper bound on the security levels of all entities
that have influenced the computation that led to the value. For simplicity of exposition, we use throughout this paper a three-point lattice {L, H, ?} (L = low
or public, H = high or secret, ? = partially leaked secret), with L v H v ? [13].
For now, readers may ignore ?. Our instrumentation works over a more general
powerset lattice, whose individual elements are Web domains. We write r` for a
value r tagged with label `.
Information flows can be categorized as explicit and implicit [27]. Explicit
flows arise as a result of variables being assigned to others, or through primitive
operations. For instance, the statement x = y + z causes an explicit flow from
values in both z and y to x. Explicit flows are handled in the runtime by updating
the label of the computed value (x in our example) with the least upper bound
of the labels of the operands in the computation (y, z in our example).
Implicit flows arise from control dependencies. For example, in the program
l = 0; if (h) {l = 1;}, there is an implicit flow from h to the final value
of l (that value is 1 iff h is 1). To handle implicit flows, dynamic IFC systems
maintain the so-called pc label (program-context label), which is an upper bound
on the labels of values that have influenced the control flow thus far. In our last
example, if the value in h has label H, then pc will be H within the if branch.
After l = 1 is executed, the final value of l inherits not only the label of 1
(which is L), but also of the pc; hence, that label is also H. This alone does not
prevent information leaks: When h = 0, l ends with 0L ; when h = 1, l ends
with 1H . Since 0L and 1H can be distinguished by a public attacker, this program
leaks the value of h despite correct propagation of implicit taints. Formally, the
instrumented semantics so far fail the standard property of non-interference [8].
This problem can be resolved through the well-known no-sensitive-upgrade
(NSU) check [22,21], which prohibits assignment to a low-labeled variable when
pc is high. This recovers non-interference if the adversary cannot observe program
termination (termination-insensitive non-interference). In our example, when h
= 0, the program terminates with l = 0L . When h = 1, the instruction l = 1
gets stuck due to NSU. These two outcomes are deemed observationally equivalent for the low adversary, who cannot determine whether or not the program
has terminated in the second case. Hence, the program is deemed secure.
Roughly, a program is termination-insensitive non-interferent if any two terminating runs of the program starting from low-equivalent heaps (i.e., heaps
that look equivalent to the adversary) end in low-equivalent heaps. Like all
sound dynamic IFC approaches, our instrumentation renders any JS program
termination-insensitive non-interferent, at the cost of modifying semantics of
programs that leak information.
4
Design, Challenges, Insights and Solutions
We implement dynamic IFC for JS in the widely used WebKit engine by instrumenting WebKit’s bytecode interpreter. In WebKit, bytecode is generated by a
source-code compiler. Our goal is to not modify the compiler, but we are forced
to make slight changes to it to make it compliant with our instrumentation. The
modification is explained in Section 6. Nonetheless, almost all our work is limited
to the bytecode interpreter.
WebKit’s bytecode interpreter is a rather standard stack machine, with several additional data structures for JS-specific features like scope chains, variable
environments, prototype chains and function objects. Local variables are held in
registers on the call stack. Our instrumentation adds a label to all data structures, including registers, object properties and scope chain pointers, adds code
to propagate explicit and implicit taints and implements a more permissive variant of the NSU check. Our label is a word size bit-set (currently 64 bits); each
bit in the bit-set represents taint from a distinct domain (like google.com). Join
on labels is simply bitwise or.
Unlike the ECMAScript specification of JS semantics, the actual implementation does not treat scope chains or variable environments like ordinary objects. Consequently, we model and instrument taint propagation on all these
data structures separately. Working at the low-level of the bytecode also leads
to several interesting conceptual and implementation issues in taint propagation
as well as interesting questions about the threat model, all of which we explain
in this section. Some of the issues are quite general and apply beyond JS. For
example, we combine our dynamic analysis with a bit of static analysis to handle
unstructured control flow and exceptions.
Threat model and compiler assumptions We explain our high-level threat model.
Following standard practice, our adversary may observe all low-labeled values
in the heap (more generally, an adversary at level ` in a lattice can observe
all heap values with labels ≤ `). However, we do not allow the adversary to
directly observe internal data structures like the call stack or scope chains. This
is consistent with actual interfaces in a browser that third-party scripts can
access. In our non-interference proofs we must also show low-equivalence of these
internal data structures across two runs to get the right induction invariants,
but assuming that they are inaccessible to the adversary allows more permissive
program execution, which we explain in Section 4.1.
The bytecode interpreter executes in a shared space with other browser components, so we assume that those components do not leak information over side
channels, e.g., they do not copy heap data from secret to public locations. This
also applies to the compiler, but we do not assume that the compiler is functionally correct. Trivial errors in the compiler, e.g., omitting a bytecode could
result in a leaky program even when the source code has no information leaks.
Because our IFC works on the compiler’s output, such compiler errors are not a
concern. Formally, we assume that the compiler is an unspecified deterministic
function of the program to compile and of the call stack, but not of the heap.
This assumption also matches how the compiler works within WebKit: It needs
access to the call stack and scope chain to optimize generated bytecode. However, the compiler never needs access to the heap. We ignore information leaks
due to other side channels like timing.
4.1
Challenges and Solutions
IFC for JS is known to be difficult due to JS’s highly dynamic nature. Working
with bytecode instead of source code makes IFC harder. Nonetheless, solutions
to many JS-specific IFC concerns proposed in earlier work [4] also apply to
our instrumentation, sometimes in slightly modified form. For example, in JS,
every object has a fixed parent, called a prototype, which is looked up when a
property does not exist in the child. This can lead to implicit flows: If an object
is created in a high context (when the pc is high) and a field missing from it,
but present in the prototype, is accessed later in a low context, then there is
an implicit leak from the high pc. This problem is avoided in both source- and
bytecode-level analysis in the same way: The “prototype” pointer from the child
to the parent is labeled with the pc where the child is created, and the label
of any value read from the parent after traversing the pointer is joined with
this label. Other potential information flow problems whose solutions remain
unchanged between source- and bytecode-level analysis include implicit leaks
through function pointers and handling of eval [12,4].
Working with bytecode both leads to some interesting insights, which are,
in some cases, even applicable to source code analysis and other languages, and
poses new challenges. We discuss some of these challenges and insights.
Unstructured control flow and CFGs To avoid overtainting pc labels, an important goal in implicit flow tracking is to determine when the influence of a control
construct has ended. For block-structured control flow limited to if and while
commands, this is straightforward: The effect of a control construct ends with
its lexical scope, e.g., in (if (h) {l = 1;}; l = 2), h influences the control
flow at l = 1 but not at l = 2. This leads to a straightforward pc upgrading
and downgrading strategy: One maintains a stack of pc labels [22]; the effective
pc is the top one. When entering a control flow construct like if or while, a new
pc label, equal to the join of labels of all values on which the construct’s guard
depends with the previous effective pc, is pushed. When exiting the construct,
the label is popped.
Unfortunately, it is unclear how to extend this simple strategy to non-blockstructured control flow constructs such as exceptions, break, continue and
return-in-the-middle for functions, all of which occur in JS. For example, consider the program l = 1; while(1) {... if (h) {break;}; l = 0; break;}
with h labeled H. This program leaks the value of h into l, but no assignment
to l appears in a block-scope guarded by h. Indeed, the pc upgrading and downgrading strategy just described is ineffective for this program. Prior work on
source code IFC either omits some of these constructs [4,28], or introduces additional classes of labels to address these problems — a label for exceptions [4],
a label for each loop containing break or continue and a label for each function [5]. These labels are more restrictive than needed, e.g., the code indicated
by dots in the example above is executed irrespective of the condition h in the
first iteration, and thus there is no need to raise the pc before checking that
condition. Further, these labels are programmer annotations, which we cannot
support as we do not wish to modify the compiler.
Importantly, unstructured control flow is a very serious concern for us, because WebKit’s bytecode has completely unstructured branches like jump-iffalse. In fact, all control flow, except function calls, is unstructured in bytecode.
To solve this problem, we adopt a solution based on static analysis of generated bytecode [29,12]. We maintain a control flow graph (CFG) of known bytecodes and for each branch node, compute its immediate post-dominator (IPD).
The IPD of a node is the first instruction that will definitely be executed, no
matter which branch is taken. Our pc upgrading and downgrading strategy now
extends to arbitrary control flow: When executing a branch node, we push a
new pc label on the stack along with the node’s IPD. When we actually reach
the IPD, we pop the pc label. In [30,31], the authors prove that the IPD marks
the end of the scope of an operation and hence the security context of the operation, so our strategy is sound. In our earlier example, the IPD of if(h) ...
is the end of the while loop because of the first break statement, so when h
== 0, the assignment l = 1 fails due to the NSU check and the program is
termination-insensitive non-interference secure.
JS requires dynamic code compilation. We are forced to extend the CFG and
to compute IPDs whenever code for either a function or an eval is compiled.
Fortunately, the IPD of a node in the CFG lies either in the same function as
the node or some function earlier in the call-chain (the latter may happen due to
exceptions), so extending the CFG does not affect computation of IPDs of earlier
nodes. This also relies on the fact that code generated from eval cannot alter
the CFG of earlier functions in the call stack [12]. In the actual implementation,
we optimize the calculation of IPDs further by working only intra-procedurally,
as explained below. At the end, our IPD-based solution works for all forms of
unstructured control flow, including unstructured branches in the bytecode, and
semi-structured break, continue, return-in-the-middle and exceptions in the
source code.
Exceptions and synthetic exit nodes Maintaining a CFG in the presence of exceptions is expensive. An exception-throwing node in a function that does not
catch that exception should have an outgoing control flow edge to the next exception handler in the call-stack. This means that (a) the CFG is, in general,
inter-procedural, and (b) edges going out of a function depend on its calling context, so IPDs of nodes in the function must be computed every time the function
is called. Moreover, in the case of recursive functions, the nodes must be replicated for every call. This is rather expensive. Ideally, we would like to build the
function’s CFG once when the function is compiled and work intra-procedurally
(as we would had there been no exceptions). We explain how we attain this goal
in the sequel.4
In our design, every function that may throw an unhandled exception has
a special, synthetic exit node (SEN), which is placed after the regular return
node(s) of the function. Every exception-throwing node, whose exception will
not be caught within the function, has an outgoing edge to the SEN, which is
traversed when the exception is thrown. The semantics of SEN (described below)
correctly transfer control to the appropriate exception handler. By doing this, we
eliminate all cross-function edges and our CFGs become intra-procedural. The
CFG of a function can be computed when the function is compiled and is never
updated. (In our implementation, we build two variants of the CFG, depending
on whether or not there is an exception handler in the call stack. This improves
efficiency, as we explain later.)
Control flows to the SEN when the function returns normally or when an
exception is thrown but not handled within the function. If no unhandled exception occurred within the function, then the SEN transfers control to the caller
(we record whether or not an unhandled exception occurred). If an unhandled
exception occurred, then the SEN triggers a special mechanism that searches
the call stack backward for the first appropriate exception handler and transfers
control to it. (In JS, exceptions are indistinguishable, so we need to find only the
first exception handler.) Importantly, we pop the call-stack up to the frame that
contains the first exception handler but do not pop the pc-stack, which ensures
that all code up to the exception handler’s IPD executes with the same pc as
the SEN, which is indeed the semantics one would expect if we had a CFG with
cross-function edges for exceptions. This prevents information leaks.
If a function does not handle a possible exception but there is an exception
handler on the call stack, then all bytecodes that could potentially throw an
exception have the SEN as one successor in the CFG. Any branching bytecode
will thus need to push to the pc-stack according to the security label of its
condition. However, we do not push a new pc-stack entry if the IPD of the
current node is the same as the IPD on the top of the pc-stack (this is just
4
This problem and our solution are not particular to JS; they apply to dynamic IFC
analysis in all languages with exceptions and functions.
an optimization) or if the IPD of the current node is the SEN, as in this case
the real IPD, which is outside of this method, is already on the pc-stack. These
semantics emulate the effect of having cross-function exception edges.
For illustration, consider the following two functions f and g. The at the
end of g denotes its SEN. Note that there is an edge from throw 9 to because
throw 9 is not handled within g. denotes the IPD of the handler catch(e)
{ l = 1; }.
function g() = {
function f() = {
if (h) {throw 9;}
l = 0;
return 7;
try { g(); } catch(e) { l = 1; }
}
return l;
}
It should be clear that in the absence of instrumentation, when f is invoked
with pc = L, the two functions together leak the value of h (which is assumed to
have label H) into the return value of f. We show how our SEN mechanism prevents this leak. When invoking g() we do not know if there will be an exception
in this function. Depending on the outcome of this method call, we will either
jump to the exception handler or continue at . Based on that branch, we push
the current pc and IPD (L, ) on the pc-stack. When executing the condition
if (h) we do not push again, but merely update the top element to (H, ). If h
== 0, control reaches without an exception but with pc = H because the IPD
of if (h) is . At this point, returns control to f, thus pc = H, but at , pc
is lowered to L, so f ends with the return value 0L . If h == 1, control reaches
with an unhandled exception. At this point, following the semantics of SEN, we
find the exception handler catch(e) { l = 1; } and invoke it with the same
pc as the point of exception, i.e., H. Consequently, NSU prevents the assignment
l = 1, which makes the program termination-insensitive non-interferent.
Because we do not wish to replicate the CFG of a function every time it is
called recursively, we need a method to distinguish the same node corresponding
to two different recursive calls on the pc-stack. For this, when pushing an IPD
onto the pc-stack, we pair it with a pointer to the current call-frame. Since the
call-frame pointer is unique for each recursive call, the CFG node paired with
the call-frame identifies a unique merge point in the real control flow graph.
In practice, even the intra-procedural CFG is quite dense because many JS
bytecodes can potentially throw exceptions and, hence, have edges to the SEN.
To avoid overtainting, we perform a crucial common-case optimization: When
there is no exception handler on the call stack we do not create the SEN and
the corresponding edges from potentially exception-throwing bytecodes at all.
This is safe as a potentially thrown exception can only terminate the program
instantly, which satisfies termination-insensitive non-interference if we ensure
that the exception message is not visible to the attacker. Whether or not an
exception handler exists is easily tracked using a stack of Booleans that mirrors
the call-stack; in our design we overlay this stack on the pc-stack by adding an
extra Boolean field to each entry of the pc-stack. In summary, each entry of our
pc-stack is a quadruple containing a security label, a node in the intraprocedural
CFG, a call-frame pointer and a Boolean value. In combination with SENs, this
design allows us to work only with intraprocedural CFGs that are computed
when a function is compiled. This improves efficiency.
Permissive-upgrade check, with changes The standard NSU check halts program
execution whenever an attempt is made to assign a variable with a low-labeled
value in a high pc. In our earlier example, l = 0; if (h) {l = 1;}, assuming
that h stores a H-labeled value, program execution is halted at the command
l = 1. As Austin and Flanagan (AF in the sequel) observe [13], this may be
overly restrictive when l will not, in fact, have observable effects (e.g., l may
be overwritten by a constant immediately after if (h) {l = 1;}). So, they
propose propagating a special taint called ? into l at the instruction l = 1 and
halting a program when it tries to use a value labeled ? in a way that will be
observable (AF call this special taint P for “partially leaked”). This idea, called
the permissive-upgrade check, allows more program execution than NSU would,
so we adopt it. In fact, this additional permissiveness is absolutely essential for us
because the WebKit compiler often generates dead assignments within branches,
so execution would pointlessly halt if standard NSU were used.
We differ from AF in what constitutes a use of a value labeled ?. As expected,
AF treat occurrence of ? in the guard of a branch as a use. Thus, the program l
= 0; if (h) {l = 1;}; if (l) {l’ = 2} is halted at the command if (l)
when h == 1 because l obtains taint ? at the assignment l = 1 (if the program
is not halted, it leaks h through l’). However, they allow ?-tainted values to
flow into the heap. Consider the program l = 0; if (h) {l = 1;}; obj.a =
l. This program is insecure in our model: The heap location obj.a, which is
accessible to the adversary, ends with 0L when h == 0 and with 1? when h
== 1. AF deem the program secure by assuming that any value with label ? is
low-equivalent to any other value (in particular, 0L and 1? are low-equivalent).
However, this definition of low-equivalence for dynamic analysis is virtually impossible to enforce if the adversary has access to the heap outside the language:
After writing 0L to obj.a (for h == 0), a dynamic analysis cannot determine
that the alternate execution of the program (for h == 1) would have written a
?-labeled value and, hence, cannot prevent the adversary from seeing 0L .
Consequently, in our design, we use a modified permissive-upgrade check,
which we call the deferred NSU check, wherein a program is halted at any construct that may potentially flow a ?-labeled value into the heap. This includes
all branches whose guard contains a ?-labeled value and any assignments whose
target is a heap location and whose source is ?-labeled. However, we do not
constrain flow of ?-labeled values in data structures that are invisible to the adversary in our model, e.g., local registers and variable environments. This design
critically relies on treating internal data structures differently from ordinary JS
objects, which is not the case, for instance, in the ECMAScript specification.
5
Formal Model and IFC
We formally model WebKit’s JS bytecode and the semantics of its bytecode
interpreter with our instrumentation of dynamic IFC. We prove termination-
ins := prim-ins | obj-ins
| func-ins
| scope-ins | exc-ins
prim-ins := prim dst:r src1:r src2:r
| mov dst:r src:r
| jfalse cond:r target:offset
| loop-if-less src1:r src2:r target:offset
| typeof dst:r src:r
| instanceof dst:r value:r cProt:r
func-ins := enter
| ret result:r
| end result:r
| call func:r args:n
| call-put-result res:r
| call-eval func:r args:n
| create-arguments dst:r
| new-func dst:r func:f
| create-activation dst:r
| construct func:r args:n
| create-this dst:r
obj-ins := new-object dst:r
| get-by-id dst:r base:r prop:id
| put-by-id base:r prop:id value:r direct:b
| del-by-id dst:r base:r prop:id
| get-pnames dst:r base:r i:n size:n breaktarget:offset
| next-pname dst:r base:r i:n size:n iter:n target:offset
| put-getter-setter base:r prop:id getter:r setter:r
scope-ins := resolve dst:r prop:id
| resolve-skip dst:r prop:id skip:n
| resolve-global dst:r prop:id
| resolve-base dst:r prop:id isStrict:bool
| resolve-with-base bDst:r pDst:r prop:id
| get-scoped-var dst:r index:n skip:n
| put-scoped-var index:n skip:n value:r
| push-scope scope:r
| pop-scope
| jmp-scope count:n target:offset
exc-ins := throw ex:r
| catch ex:r
Fig. 1. Instructions
insensitive non-interference for programs executed through our instrumented
interpreter. We do not model the construction of the CFG or computation of
IPDs; these are standard. To keep presentation accessible, we present our formal model at a somewhat high-level of abstraction. Details are resolved in our
technical appendix.
5.1
Bytecode and Data Structures
The version of WebKit we model uses a total of 147 bytecodes or instructions,
of which we model 69. The remaining 78 bytecodes are redundant from the
perspective of formal modeling because they are specializations or wrappers on
other bytecodes to improve efficiency. The syntax of the 69 bytecodes we model
is shown in Fig. 1. The bytecode prim abstractly represents 34 primitive binary
and unary (with just the first two arguments) operations, all of which behave
similarly. For convenience, we divide the bytecodes into primitive instructions
(prim-ins), instructions related to objects and prototype chains (obj-ins), instructions related to functions (func-ins), instructions related to scope chains
(scope-ins) and instructions related to exceptions (exc-ins). A bytecode has the
form hinst_name list_of _argsi. The arguments to the instruction are of the
form hvari:htypei, where var is the variable name and type is one of the following:
r, n, bool, id, prop and offset for register, constant integer, constant Boolean,
identifier, property name and jump offset value, respectively.
In WebKit, bytecode is organized into code blocks. Each code block is a
sequence of bytecodes with line numbers and corresponds to the instructions for
a function or an eval statement. A code block is generated when a function
is created or an eval is executed. In our instrumentation, we perform control
flow analysis on a code block when it is created and in our formal model we
abstractly represent a code block as a CFG, written ζ. Formally, a CFG is a
directed graph, whose nodes are bytecodes and whose edges represent possible
control flows. There are no cross-function edges. A CFG also records the IPD of
each node. IPDs are computed using an algorithm by Lengauer and Tarjan [32]
when the CFG is created. If the CFG contains uncaught exceptions, we also
create a SEN. For a CFG ζ and a non-branching node ι ∈ ζ, Succ(ζ, ι) denotes
ι’s unique successor. For a conditional branching node ι, Left(ζ, ι) and Right(ζ, ι)
denote successors when the condition is true and false, respectively.
The bytecode interpreter is a standard stack machine, with support for JS
features like scope chains and prototype chains. The state of the machine (with
our instrumentation) is a quadruple hι, θ, σ, ρi, where ι represents the current
node that is being executed, θ represents the heap, σ represents the call-stack
and ρ is the pc-stack.
We assume an abstract, countable set A = {a, b, . . .} of heap locations, which
are references to objects. The heap θ is a partial map from locations to objects.
An object O may be:
– An ordinary JS object N = ({pi 7→ vi }ni=0 , __proto__ 7→ a`p , `s ), containing properties named p0 , . . . , pn that map to labeled values v0 , . . . , vn , a
prototype field that points to a parent at heap location a, and two labels `p
and `s . `p records the pc where the object was created. `s is the so-called
structure label, which is an upper bound on all pcs that have influenced
which fields exist in the object.5
– A function object F = (N, ζ, Σ), where N is an ordinary object, ζ is a CFG,
which corresponds to the the function stored in the object, and Σ is the
scope chain (closing context) of the function.
A labeled value v = r` is a value r paired with a security label `. A value r
in our model may be a heap location a or a JS primitive value n, which includes
integers, Booleans, regular expressions, arrays, strings and the special JS values
undefined and null.
The call-stack σ contains one call-frame for each incomplete function call. A
call-frame µ contains an array of registers for local variables, a CFG ζ for the
function represented by the call-frame, the return address (a node in the CFG of
the previous frame), and a pointer to a scope-chain that allows access to variables
in outer scopes. Additionally, each call-frame has an exception table which maps
each potentially exception-throwing bytecode in the function to the exception
handler within the function that surrounds the bytecode; when no such exception
handler exists, it points to the SEN of the function (we conservatively assume
that any unknown code may throw an exception, so bytecodes call and eval
are exception-throwing for this purpose). |σ| denotes the size of the call-stack
and !σ its top frame. Each register contains a labeled value.
A scope chain, Σ, is a sequence of scope chain nodes (SCNs), denoted S,
paired with labels. In WebKit, a scope chain node S may either be an object
or a variable environment V , which is an array of labeled values. Thus, Σ ::=
(S1 , `1 ) : . . . : (Sn , `n ) and S ::= O | V and V ::= v1 : . . . : vn .
5
The __proto__ field is the parent of the object; it is not the same as the prototype
field of a function object, which is an ordinary property. Also, in our actual model,
fields pi map to more general property descriptors that also contain attributes along
with labeled values. We elide attributes here to keep the presentation simple.
Each entry of the pc-stack ρ is a triple (`, ι, p), where ` is a security label,
ι is a node in a CFG, and p is a pointer to some call-frame on the call stack
σ. (For simplicity, we ignore a fourth Boolean field described in Section 4.1 in
this presentation.) When we enter a new control context, we push the new pc `
together with the IPD ι of the entry point of the control context and a pointer
p to current call-frame. The pair (ι, p) uniquely identifies where the control of
the context ends; p is necessary to distinguish the same branch point in different
recursive calls of the function [12]. In our semantics, we use the meta-function
isIPD to pop the stack. It takes the current instruction, the current pc-stack and
the call stack σ, and returns a new pc-stack.
(
ρ.pop() if !ρ = (_, ι, !σ)
isIPD(ι, ρ, σ) :=
ρ
otherwise
As explained in Section 4.1, as an optimization, we push a new node (`, ι, σ)
onto ρ only when (ι, σ) (the IPD) differs from the corresponding pair on the top
of the stack and, to handle exceptions correctly, we also require that ι not be
the SEN. Otherwise, we just join ` with the label on the top of the stack. This
is formalized in the function ρ.push(`, ι, σ), whose obvious definition we elide.
If x is a pair of any syntactic entity and a security label, we write Υ (x) for the
entity and Γ (x) for the label. In particular, for v = r` , Υ (v) = r and Γ (v) = `.
5.2
Semantics and IFC with Intra-procedural CFGs
We now present the semantics, which faithfully models our implementation using intra-procedural CFGs with SENs. The semantics is defined as a set of state
transition rules that define the judgment: hι, θ, σ, ρi
hι0 , θ0 , σ 0 , ρ0 i. Fig. 2 shows
rules for selected bytecodes. For reasons of space we omit rules for other bytecodes and formal descriptions of some meta-function like opCall that are used
in the rules. C ⇒ A B is shorthand for a meta-level (if (C) then A else B).
prim reads the values from two registers src1 and src2, performs a binary
operation generically denoted by ⊕ on the values and writes the result into the
register dst. dst is assigned the join of the labels in src1, src2 and the head of
the pc-stack (!ρ). To implement deferred NSU (Section 4.1), the existing label
in dst is compared with the current pc. If the label is lower than the pc, then
the label of dst is joined with ?. Note that the premise ρ0 = isIPD(ι0 , ρ, σ) pops
an entry from the pc-stack if its IPD matches the new program node ι0 . This
premise occurs in all semantic rules.
jfalse is a conditional jump. It skips offset number of successive nodes
in the CFG if the register cond contains false, else it falls-through to the next
node. Formally, the node it branches to is either Right(ζ, ι) or Left(ζ, ι), where ζ
is the CFG in !σ. In accordance with deferred NSU, the operation is performed
only if cond is not labeled ?. jfalse also starts a new control context, so a new
node is pushed on the top of the pc-stack with a label that is the join of Γ (cond)
and the current label on the top of the stack (unless the IPD of the branch point
is already on top of the stack or it is the SEN, in which case we join the new
ι = “op-prim dst:r src1:r src2:r”,
L := Γ (!σ(src1)) t Γ (!σ(src2)) t Γ (!ρ),
V := Υ (!σ(src1)) ⊕ Υ (!σ(src2))
(Γ (!σ(dst)) ≥ Γ (!ρ))
h =⇒ (L :=iL) (L := ?)
σ 0 := σ
0
prim:
0
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
0
,
ι := Succ(!σ .CFG, ι), ρ := isIPD(ι0 , ρ, σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
ι = “op-jfalse cond:r target:offset”,
Γ (!σ(cond)) 6= ?, L := Γ (!σ(cond)) t Γ (!ρ),
Υ (!σ(cond)) = false =⇒ ι0 := Left(!σ.CFG, ι)
ι0 := Right(!σ.CFG, ι),
ρ00 := ρ.push(L, IPD(ι), CF (ι)), ρ0 := isIPD(ι0 , ρ00 , σ)
jfalse:
ι, θ, σ, ρ
ι0 , θ, σ, ρ0
ι = “op-push-scope scope:r”,
σ 0 := pushScope(Γ (!ρ), σ, scope),
ι := Succ(!σ 0 .CFG, ι), ρ0 := isIPD(ι0 , ρ, σ 0 )
push-scope:
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
0
ι = “op-call func:r args:n”,
Γ (func) 6= ?, (ι0 , σ 0 , `f ) := opCall(σ, ι, func, args),
L = `f t Γ (!σ(func)) t Γ (!ρ),
ρ00 := ρ.push(L, IPD(ι), CF (ι)), ρ0 := isIPD(ι0 , ρ00 , σ 0 )
call:
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
ret:
ι = “op-put-by-id base:r prop:id value:r direct:b”,
Γ (!σ(value)) 6= ?, direct = true =⇒
θ0 := putDirect(Γ (!ρ), σ, θ, base, prop, value)
θ0 := putIndirect(Γ (!ρ), σ, θ, base, prop, value),
ι0 := Succ(!σ.CFG, ι), ρ0 := isIPD(ι0 , ρ, σ)
put-by-id:
ι, θ, σ, ρ
ι0 , θ0 , σ, ρ0
ι = “op-ret res:r”,
(ι0 , σ 0 , γ) := opRet(σ, res), ρ0 := isIPD(ι0 , ρ, σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
throw:
ι = “op-throw ex:r”, excValue := Υ (!σ(ex)),
(σ 0 , ι0 ) := throwException(σ, ι), ρ0 := isIPD(ι0 , ρ, σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
Fig. 2. Semantics, selected rules
label with the previous). Traversed from bottom to top, the pc-stack always has
monotonically non-decreasing labels.
put-by-id updates the property prop in the object pointed to by register
base. As explained in Section 4.1, we allow this only if the value to be written is
not labeled ?. The flag direct states whether or not to traverse the prototype
chain in finding the property; it is set by the compiler as an optimization. If the
flag is true, then the chain is not traversed (meta-function putDirect handles this
case). If direct is false, then the chain is traversed (meta-function putIndirect).
Importantly, when the chain is traversed, the resulting value is labeled with the
join of prototype labels `p and structure labels `s of all traversed objects. This
is standard and necessary to prevent implicit leaks through the __proto__
pointers and structure changes to objects.
push-scope, which corresponds to the start of the JS construct with(obj),
pushes the object pointed to by the register scope into the scope chain. Because
pushing an object into the scope chain can implicitly leak information from the
program context later, we also label all nodes in the scope-chain with the pc’s at
which they were added to the chain. Further, deferred NSU applies to the scope
chain pointer in the call-frame as it does to all other registers.
call invokes a function of the target object stored in the register func. Due
to deferred NSU, the call proceeds only if Γ (func) is not ?. The call creates a new
call-frame and initializes arguments, the scope chain pointer (initialized with the
function object’s Σ field), CFG and the return node in the new frame. The CFG
in the call-frame is copied from the function object pointed to by func. All this
is formalized in the meta-function opCall, whose details we omit here. Call is a
branch instruction and it pushes a new label on the pc-stack which is the join of
the current pc, Γ (func) and the structure label `f of the function object (unless
the IPD of the current node is the SEN or already on the top of the pc-stack, in
which case we join the new pc-label with the previous). call also initializes the
new registers’ labels to the new pc. A separate bytecode, not shown here and
executed first in the called function, sets register values to undefined. eval is
similar to call but the code to be executed is also compiled.
ret exits a function. It returns control to the caller, as formalized in the
meta-function opRet. The return value is written to an interpreter variable (γ).
throw throws an exception, passing the value in register ex as argument to
the exception handler. Our pc-stack push semantics ensure that the exception
handler, if any, is present in the call-frame pointed to by the top of the pc-stack.
The meta-function throwException pops the call-stack up to this call-frame and
transfers control to the exception handler, by looking it up in the exception
table of the call-frame. The exception value in the register ex is transferred to
the handler through an interpreter variable.
The semantics of other bytecodes have been described in Section 8.3.
Correctness of IFC We prove that our IFC analysis guarantees terminationinsensitive non-interference [11]. Intuitively, this means that if a program is run
twice from two states that are observationally equivalent for the adversary and
both executions terminate, then the two final states are also equivalent for the
adversary. To state the theorem formally, we formalize equivalence for various
data structures in our model. The only nonstandard data structure we use is the
CFG, but graph equality suffices for it. A well-known complication is that low
heap locations allocated in the two runs need not be identical. We adopt the
standard solution of parametrizing our definitions of equivalence with a partial
bijection β between heap locations. The idea is that two heap locations are
related in the partial bijection if they were created by corresponding allocations
in the two runs. We then define a rather standard relation hι1 , θ1 , σ1 , ρ1 i ∼β`
hι2 , θ2 , σ2 , ρ2 i, which means that the states on the left and right are equivalent
to an observer at level `, up to the bijection β on heap locations. The details
have been presented in Section 8.4.
Theorem 1 (Termination-insensitive non-interference) Suppose:
(1) hι1 , θ1 , σ1 , ρ1 i ∼β` hι2 , θ2 , σ2 , ρ2 i, (2) hι1 , θ1 , σ1 , ρ1 i ∗ hend, θ10 , [], []i, and
0
(3) hι2 , θ2 , σ2 , ρ2 i ∗ hend, θ20 , [], []i. Then, ∃β 0 ⊇ β such that θ10 ∼β` θ20 .
6
Implementation
We instrumented WebKit’s JS engine (JavaScriptCore) to implement the IFC
semantics of the previous section. Before a function starts executing, we generate
its CFG and calculate IPDs of its nodes by static analysis of its bytecode. We
modify the source-to-bytecode compiler to emit a slightly different, but functionally equivalent bytecode sequence for finally blocks; this is needed for accurate
computation of IPDs. For evaluation purposes, we label each source script with
the script’s domain of origin; each seen domain is dynamically allocated a bit in
our bit-set label. In general, our instrumentation terminates a script that violates
Normalized
JS
Execu2on
Time
(%)
350
Interpreter
JIT
Basic
Op@mized
300
250
200
150
100
50
0
3d
access
bitops
controlflow
crypto
date
math
regexp
string
SunSpider
Tests
Fig. 3. Overheads of basic and optimized IFC in SunSpider benchmarks
IFC. However, for the purpose of evaluating overhead of our instrumentation,
we ignore IFC violations in all experiments described here.
We also implement and evaluate a variant of sparse labeling [21] which optimizes the common case of computations that mostly use local variables (registers
in the bytecode). Until a function reads a value from the heap with a label different from the pc, we propagate taints only on heap-writes, but not on in-register
computations. Until that point, all registers are assumed to be implicitly tainted
with pc. This simple optimization reduces the overhead incurred by taint tracking significantly in microbenchmarks. For both the basic and optimized version,
our instrumentation adds approximately 4,500 lines of code to WebKit.
Our baseline for evaluation is the uninstrumented interpreter with JIT disabled. For comparison, we also include measurements with JIT enabled. Our
experiments are based on WebKit build #r122160 running in Safari 6.0. The
machine has a 3.2GHz Quad-core Intel Xeon processor with 8GB RAM and
runs Mac OS X version 10.7.4.
Microbenchmark We executed the standard SunSpider 1.0.1 JS benchmark suite
on the uninstrumented interpreter with JIT disabled and JIT enabled, and on
the basic and the optimized IFC instrumentations with JIT disabled. Results are
shown in Figure 3. The x-axis ranges over SunSpider tests and the y-axis shows
the average execution time, normalized to our baseline (uninstrumented interpreter with JIT disabled) and averaged across 100 runs. Error bars are standard
deviations. Although the overheads of IFC vary from test to test, the average
overheads over our baseline are 121% and 45% for basic IFC and optimized
IFC, respectively. The test regexp has almost zero overhead because it spends
most time in native code, which we have not yet instrumented. We also note
that, as expected, the JIT-enabled configuration performs extremely well on the
SunSpider benchmarks.
Normalized
JavaScript
Execu5on
Time
(%)
160
Interpreter
JIT
Basic
instrumentaGon
OpGmized
InstrumentaGon
140
120
100
80
60
40
20
0
Google
Yahoo
Amazon
Wikipedia
eBay
Websites
Bing
Linkedin
Live
Twi?er
Fig. 4. Overheads of basic and optimized IFC in real websites
Macrobenchmarks We measured the execution time of the intial JS on 9 popular
English language Websites. We load each Website in Safari and measure the
total time taken to execute the JS code without user interaction. This excludes
time for network communication and internal browser events and establishes a
very conservative baseline. The results, normalized to our baseline, are shown in
Fig. 4. Our overheads are all less than 42% (with an average of around 29% in
both instrumentations). Interestingly, we observe that our optimization is less
effective on real websites indicating that real JS accesses the heap more often
than the SunSpider tests. When compared to the amount of time it takes to
fetch a page over the network and to render it, these overheads are negligible.
Enabling JIT worsens performance compared to our baseline indicating that, for
the code executed here, JIT is not useful.
We also experimented with JSBench [33], a sophisticated benchmark derived
from JS code in the wild. The average overhead on all JSBench tests (a total 23
iterations) is approximately 38% for both instrumentations. The average time
for running the benchmark tests on the uninstrumented interpreter with JIT
disabled was about 636.11ms with a standard deviation of about 0.30% of the
mean. The average time for running the same benchmark tests on the instrumented interpreter and the optimized version was about 874.31ms and 880.85ms
respectively, with a standard deviation of about 4.09% and 5.04% of the mean
in the two cases.
7
Conclusion and Future Work
We have explored dynamic information flow control for JS bytecode in WebKit,
a production JS engine. We formally model the bytecode, its semantics, our
instrumentation and prove the latter correct. We identify challenges, largely
arising from pervasive use of unstructured control flow in bytecode, and resolve
them using very limited static analysis. Our evaluation indicates only moderate
overheads in practice.
In ongoing work, we are instrumenting the DOM and other native JS methods. We also plan to generalize our model and non-interference theorem to
take into account the reactive nature of Web browsers. Going beyond noninterference, the design and implementation of a policy language for representing
allowed information flows looks necessary.
Acknowledgments This work was funded in part by the Deutsche Forschungsgemeinschaft (DFG) grant “Information Flow Control for Browser Clients” under
the priority program “Reliably Secure Software Systems” (RS3) and the German Federal Ministry of Education and Research (BMBF) within the Centre for
IT-Security, Privacy and Accountability (CISPA) at Saarland University.
References
1. Richards, G., Hammer, C., Burg, B., Vitek, J.: The eval that men do – a large-scale
study of the use of eval in JavaScript applications. In Mezzini, M., ed.: ECOOP
’11. Volume 6813 of LNCS. (2011) 52–78
2. Jang, D., Jhala, R., Lerner, S., Shacham, H.: An empirical study of privacyviolating information flows in JavaScript web applications. In: Proc. 17th ACM
Conference on Computer and Communications Security. (2010) 270–283
3. Richards, G., Hammer, C., Zappa Nardelli, F., Jagannathan, S., Vitek, J.: Flexible access control for javascript. In: Proc. 2013 ACM SIGPLAN International
Conference on Object Oriented Programming Systems Languages & Applications.
OOPSLA ’13 (2013) 305–322
4. Hedin, D., Sabelfeld, A.: Information-flow security for a core of JavaScript. In:
Proc. 25th IEEE Computer Security Foundations Symposium. (2012) 3–18
5. Hedin, D., Birgisson, A., Bello, L., Sabelfeld, A.: JSFlow: Tracking information
flow in JavaScript and its APIs. In: Proc. 29th ACM Symposium on Applied
Computing. (2014)
6. Devriese, D., Piessens, F.: Noninterference through secure multi-execution. In:
Proc. 2010 IEEE Symposium on Security and Privacy. (2010) 109–124
7. De Groef, W., Devriese, D., Nikiforakis, N., Piessens, F.: Flowfox: a web browser
with flexible and precise information flow control. In: Proc. 2012 ACM Conference
on Computer and Communications Security. (2012) 748–759
8. Goguen, J.A., Meseguer, J.: Security policies and security models. In: Proc. 1982
IEEE Symposium on Security and Privacy. (1982) 11–20
9. Myers, A.C., Liskov, B.: A decentralized model for information flow control. In:
Proc. 16th ACM Symposium on Operating Systems Principles. (1997) 129–142
10. Zdancewic, S., Myers, A.C.: Robust declassification. In: Proc. 14th IEEE Computer
Security Foundations Workshop. (2001) 15–23
11. Volpano, D., Irvine, C., Smith, G.: A sound type system for secure flow analysis.
J. Comput. Secur. 4(2-3) (January 1996) 167–187
12. Just, S., Cleary, A., Shirley, B., Hammer, C.: Information flow analysis for JavaScript. In: Proc. 1st ACM SIGPLAN International Workshop on Programming
Language and Systems Technologies for Internet Clients. (2011) 9–18
13. Austin, T.H., Flanagan, C.: Permissive dynamic information flow analysis. In:
Proc. 5th ACM SIGPLAN Workshop on Programming Languages and Analysis
for Security. (2010) 3:1–3:12
14. Bohannon, A., Pierce, B.C., Sjöberg, V., Weirich, S., Zdancewic, S.: Reactive noninterference. In: Proc. 16th ACM Conference on Computer and Communications
Security. (2009) 79–90
15. Maffeis, S., Mitchell, J.C., Taly, A.: An operational semantics for JavaScript. In:
Proc. 6th Asian Symposium on Programming Languages and Systems. APLAS ’08
(2008) 307–325
16. Guha, A., Saftoiu, C., Krishnamurthi, S.: The essence of JavaScript. In: Proc.
24th European Conference on Object-Oriented Programming. (2010) 126–150
17. Politz, J.G., Carroll, M.J., Lerner, B.S., Pombrio, J., Krishnamurthi, S.: A tested
semantics for getters, setters, and eval in JavaScript. In: Proceedings of the 8th
Dynamic Languages Symposium. (2012) 1–16
18. Bodin, M., Chargueraud, A., Filaretti, D., Gardner, P., Maffeis, S., Naudziuniene,
D., Schmitt, A., Smith, G.: A trusted mechanised javascript specification. In:
Proc. 41st ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages. (2014)
19. Guarnieri, S., Pistoia, M., Tripp, O., Dolby, J., Teilhet, S., Berg, R.: Saving the
world wide web from vulnerable javascript. In: Proc. 2011 International Symposium
on Software Testing and Analysis. ISSTA ’11 (2011) 177–187
20. Chugh, R., Meister, J.A., Jhala, R., Lerner, S.: Staged information flow for JavaScript. In: Proc. 2009 ACM SIGPLAN Conference on Programming Language
Design and Implementation. (2009) 50–62
21. Austin, T.H., Flanagan, C.: Efficient purely-dynamic information flow analysis. In:
Proc. ACM SIGPLAN Fourth Workshop on Programming Languages and Analysis
for Security. (2009) 113–124
22. Zdancewic, S.A.: Programming Languages for Information Security. PhD thesis,
Cornell University (August 2002)
23. Birgisson, A., Hedin, D., Sabelfeld, A.: Boosting the permissiveness of dynamic
information-flow tracking by testing. In: Computer Security – ESORICS 2012.
Volume 7459 of LNCS. Springer Berlin Heidelberg (2012) 55–72
24. Austin, T.H., Flanagan, C.: Multiple facets for dynamic information flow. In: Proc.
39th annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages. (2012) 165–178
25. Bielova, N., Devriese, D., Massacci, F., Piessens, F.: Reactive non-interference for a
browser model. In: 5th International Conference on Network and System Security
(NSS),. (2011) 97–104
26. Bohannon, A., Pierce, B.C.: Featherweight firefox: formalizing the core of a web
browser. In: Proc. 2010 USENIX conference on Web application development.
WebApps’10 (2010) 11–22
27. Denning, D.E.: A lattice model of secure information flow. Commun. ACM 19(5)
(May 1976) 236–243
28. Dhawan, M., Ganapathy, V.: Analyzing information flow in JavaScript-based
browser extensions. In: Proc. 2009 Annual Computer Security Applications Conference. ACSAC ’09 (2009) 382–391
29. Robling Denning, D.E.: Cryptography and Data Security. Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA (1982)
30. Xin, B., Zhang, X.: Efficient online detection of dynamic control dependence. In:
Proc. 2007 International Symposium on Software Testing and Analysis. (2007)
185–195
31. Masri, W., Podgurski, A.: Algorithms and tool support for dynamic information
flow analysis. Information & Software Technology 51(2) (2009) 385–404
32. Lengauer, T., Tarjan, R.E.: A fast algorithm for finding dominators in a flowgraph.
ACM Trans. Program. Lang. Syst. 1(1) (January 1979) 121–141
33. Richards, G., Gal, A., Eich, B., Vitek, J.: Automated construction of javascript
benchmarks. In: Proceedings of the 2011 ACM International Conference on Object
Oriented Programming Systems Languages and Applications. (2011) 677–694
8
8.1
Appendix
Data Structures
The formal model described in Section 5 was typechecked in C++. The various data structures used for defining the functions used in the semantics of the
language are given in Figure 5. The source-code of the JavaScript program is
represented as a structure containing the source and a Boolean flag indicating
the strict mode is set or not. The instruction as indicated before is a structure
consisting of the opcode and the list of operands. The opcode is a string indicating
the operation and the operand is a union of registerIndex, immediateValue, identifier, Boolean, funcIndex and offset. The immediateValue denotes the directly
supplied value to an opcode, registerIndex is the index of the register containing
the value to be operated upon, identifier represents the string name directly
used by the opcode, Boolean is a often a flag indicating the truth value of some
parameter and offset represents the offset where the control jumps to. Similarly,
functionIndex indicates the index of the function object being invoked.
The function’s source code is represented in the form of a control flow graph
(CFG). Formally, it is defined as a struct with a list of CFG nodes, each of which
contain the instructions that are to be performed and the edges point to the
next instruction in the program. Multiple outgoing edges indicate a branching
instruction. It also contains variables indicating the number of variables used by
the function code and a reference to the globalObject.
The labels are interpreted as a structure consisting of long integer label. The
label represents the value of the label, which are interpreted as bit vectors. A
special label star which represents partially leaked data, is used for deferred nosensitive-upgrade check. The program counter (pc) is represented as a stack of
pc-nodes, each of which contains the context label and the IPD of the operation
that pushed the node, the callframe of the current node and the handler flag
indicating the presence of an exception handler.
Different types of values are used as operands for performing the operations.
They include Boolean, integer, string, double and objects or special values like
NaN or undefined. These values are associated with a label each and are wrapped
by the JSValue class. All the values used in the data structures have the type
JSValue. The objects consist of properties, a prototype chain pointer with an
associated label and a structure label for the object. The properties are represented as a structure of the propertyName and its descriptor. The descriptor of
the property contains the value, some Boolean flags and a property label. The
struct SourceCode{
String programSrc;
bool strictMode;
};
struct JSValue{
valueTemplate data;
JSLabel label;
};
typedef char* Opcode;
union Operand{
int immediateValue;
String identifier;
int registerIndex;
int funcIndex;
bool flag;
int offset;
};
struct Instruction{
Opcode opc;
Operand* opr;
};
struct CFGNode{
Instruction* inst;
struct CFGNode* left;
struct CFGNode* right;
struct CFGNode* succ;
};
struct PropertyDescriptor{
JSValue value;
bool writable;
bool enumerable;
bool configurable;
JSLabel structLabel;
};
struct Property{
String propertyName;
PropertyDescriptor* pDesc;
};
struct PropertySlot{
Property prop;
PropertySlot* next;
};
struct Register{
JSValue value;
};
struct CallFrameNode{
Register rf[REGISTER_FILE_SIZE];
CFG cfg;
CFGNode* returnAddress;
ScopeChainNode* sc;
JSFunctionObject* callee;
JSLabel calleeLabel;
int argCount;
bool getter;
int dReg;
};
struct CallFrameStack{
CallFrameNode cFN;
CallFrameStack* previous;
};
struct PCNode{
JSLabel l;
struct JSObject{
CFGNode* ipd;
Property property[MAX_PROPS];
CallFrameNode* cFN;
struct proto{
bool handler;
JSLabel l;
};
struct CFG{
JSObject* __proto__;
struct CFGNode* cfgNode;
} prototype;
struct PCStack{
JSGlobalObject* globalObject; JSLabel structLabel;
PCNode node;
int numVars;
};
PCStack* previous;
int numFns;
};
bool strictMode;
struct Heap{
};
unsigned location;
struct JSActivation{
JSObject o;
CallFrameNode* callFrameNode;
struct JSLabel{
}[HEAP_SIZE];
JSLabel structLabel;
uint64_t label;
};
};
enum FunctionType{
JSFunction, HostFunction
enum ScopeChainObjectType{
enum Specials{
};
LexicalObject, VariableObject
NaN, undefined
};
};
struct JSFunctionObject:
JSObject{
union SChainObject{
union ValueType{
CFG* funcCFG;
JSObject* obj;
bool b;
ScopeChainNode* scopeChain;
JSActivation* actObj;
int n;
FunctionType fType;
};
String s;
};
double d;
struct ScopeChainNode{
JSObject* o;
struct JSGlobalObject:
SChainObject Object;
};
JSObject{
ScopeChainObjectType scObjType;
JSFunctionObject evalFunction;
ScopeChainNode* next;
union valueTemplate{
JSObject* objectPrototype;
JSLabel scopeLabel;
Specials s;
JSObject* functionPrototype;
};
ValueType v;
};
};
Fig. 5. Data Structures
heap is a collection of objects with an associated memory address. It is essentially
a map from location to object.
There are subtypes of JSObject that define the function object and the global
object. The function object contains a pointer to the associated CFG and the
scope chain. It also contains a field defining the type of function it represents,
namely, host or user-defined.
The call-stack is made up of various call-frame nodes, each of which contains a
set of registers, the associated CFG, the return address of the function, a pointer
to the scope chain, and an exception table. The registers store values and objects
and are used as operands for performing the operations. The exception table
contains the details about the handlers associated with different instructions in
the CFG of the call-frame. The scope chain is a list of nodes containing objects or
activation objects along with a label indicating the context in which the object
in that node was added. The activation object is a structure containing a pointer
to the call-frame node for which the activation object was created.
The next section defines the different procedures used in the semantics of the
language. The statement stop implies that the program execution hangs.
8.2
Algorithms
The different meta-functions used in the semantics presented in Section 8.3 are
described below:
procedure isInstanceOf (JSLabel context, JSValue obj, JSValue protoVal)
oProto := obj.prototype.__proto__
while oProto do
if oProto = protoVal then
ret := JSValue::construct(true)
ret.label := context
return ret
end if
oProto := oProto.prototype.__proto__
context := context.Join(oProto.prototype.l)
end while
ret := JSValue::construct(false)
ret.label := context
return ret
end procedure
procedure opRet(CallFrameStack* callStack, int ret)
JSValue retValue = callStack.cFN.rf[ret].value
if hostCallFrameFlag then
callStack.pop()
return nil, callStack, retValue
end if
callStack.pop()
return callStack.cFN.returnAddress, callStack, retValue
end procedure
procedure opCall(CallFrameStack* callStack, CFGNode* ip, int func, int argCount)
JSValue funcValue := callStack.cFN.rf[func].value
JSFunctionObject fObj
CallFrameNode *sigmaTop := new CallFrameNode()
CallFrameNode *prevTop := callStack.top()
callStack.push(sigmaTop)
CallType callType := getCallData(funcValue, &fObj)
if callType = CallTypeJS then
ScopeChainNode* sc := fObj.scopeChain
callStack.cFN.cfg := fObj.funcCFG
callStack.cFN.returnAddress := ip.Succ
callStack.cFN.sc := sc
args.noOfArgs := argCount
callStack.cFN.argCount := argCount
for i ← 0, argCount do
callStack.cFN.rf[sigmaTop.baseRegister() + i].value :=
prevTop.rf[prevTop.headRegister()-i].value
end for
ip := callStack.cFN.cfg.cfgNode
else if callType = CallTypeHost then
stop
end if
retState.ip := ip
retState.sigma := callStack
return retState
end procedure
. Not modeled
procedure opCallEval(JSLabel contextLabel, CallFrameStack* callStack, CFGNode* ip, int func,
int argCount)
JSValue funcValue := callStack.cFN.rf[func].value
JSFunctionObject* fObj
JSObject* variableObject
Argument* arguments
if isHostEval(funcValue) then
ScopeChainNode* sc := fObj.scopeChain
callStack.cFN.returnAddress := ip + 1
callStack.cFN.sc := sc
callStack.cFN.argCount := argCount
SourceCode progSrc := funcValue.getSource()
Compiler::preparse(progSrc)
CFG* evalCodeBlock := Compiler::compile(progSrc)
unsigned numVars := evalCodeBlock.numVariables()
unsigned numFuncs := evalCodeBlock.numFuncDecls()
if numVars || numFuncs then
if evalCodeBlock.strictMode then
JSActivation* variableObject := new JSActivation()
variableObject.create(callStack)
SChainObject* scObj
scObj.actObj := variableObject
sc.push(scObj, variableObject, contextLabel)
else
for (ScopeChainNode* n := sc;; n := n.next) do
if n.isVariableObject() && !n.isLexicalObject() then
variableObject := n.getObject()
break
end if
end for
end if
for i ← 0, numVars do
Identifier iden := evalCodeBlock.variable(i)
if !variableObject.hasProperty(iden) then
variableObject.insertVariable(iden)
end if
end for
for i ← 0, numFuncs do
JSFunctionObject* fObj := evalCodeBlock.funcDecl(i)
variableObject.insertFunction(fObj)
end for
end if
callStack.cFN.cfg := evalCodeBlock
ip := evalCodeBlock.cfgNode
retState.ip := ip
retState.sigma := callStack
return retState
else
return opCall(contextLabel, callStack, ip, func, argCount)
end if
end procedure
procedure createArguments(Heap* h, CallFrameStack* callStack)
JSObject* jsArgument := JSArgument::create(h, callStack)
h.o[++(h.location)] := *jsArgument
retState.theta := h
retState.val := JSValue::construct(jsArgument)
return retState
end procedure
procedure newFunc(CallFrameStack* callStack, Heap* heap, int funcIndex, JSLabel context)
CFG* cBlock := callStack.cFN.cfg
SourceCode fcCode := cBlock.getFunctionSrc(funcIndex)
CFG* fcBlock := Compiler::compile(fcCode,callStack.cFN)
JSFunctionObject* fObj := JSFunctionObject::create( fcBlock,callStack.cFN.sc)
fObj.structLabel := context
heap.o[++(heap.location)] := *fObj
retState.theta := heap
retState.val := JSValue::construct(fObj)
return retState
end procedure
procedure createActivation(CallFrameStack* callStack, JSLabel contextLabel)
JSActivation* jsActivation := new JSActivation()
jsActivation.create(callStack)
jsActivation.structLabel := contextLabel
SChainObject* scObj
scObj.actObj := jsActivation
JSValue vActivation := JSValue::jsValuefromActivation (jsActivation)
if callStack.cFN.scopeLabel ≥ contextLabel then
callStack.cFN.sc.push(scObj, VariableObject, contextLabel)
callStack.cFN.scopeLabel := contextLabel
else
stop
end if
return retState
end procedure
procedure createThis(JSLabel contextLabel, CallFrameStack* callStack, Heap* h)
JSFunctionObject* callee := callStack.cFN.callee
PropertySlot p(callee)
String str := "prototype"
JSValue proto := p.getValue(str)
JSObject* obj := new JSObject()
obj.structLabel := contextLabel
obj.prototype.__proto__ := proto.toObject()
obj.prototype.l := proto.toObject().structLabel.join( contextLabel)
h.o[++(h.location)] := *obj
retState.theta := h
retState.val := JSValue::construct(obj)
return retState
end procedure
procedure newObject(Heap* h, JSLabel contextLabel)
JSObject* obj := emptyObject()
obj.structLabel := contextLabel
obj.prototype.__proto__ := ObjectPrototype::create()
obj.prototype.l := contextLabel
h.o[++(h.location)] := *obj
retState.theta := h
retState.val := JSValue::construct(obj)
return retState
end procedure
procedure getPropertyById(JSValue v, String p, int dst)
JSObject* O := v.toObject()
JSLabel label := O.structLabel
JSValue ret := jsUndefined()
if O.isUndefined() then
ret.label := label
return ret
end if
while O 6= null do
if O.containsProperty(p) then
if p.isGetter() then
JSValue v = p.getValue()
JSFunctionObject* funcObj = (JSFunctionObject*) v.toObject()
CallFrameNode *sigmaTop = new CallFrameNode()
callStack.push(sigmaTop)
ScopeChainNode* sc = fObj.scopeChain
CFG* newCodeBlock = fObj.funcCFG
callStack.cFN.cfg = *newCodeBlock
callStack.cFN.returnAddress = ip + 1
callStack.cFN.sc = sc
callStack.cFN.getter = true
callStack.cFN.dReg = dst
ip = newCodeBlock.cfgNode
interpreter.iota = ip
interpreter.sigma = callStack
else
ret := getProperty(p).getValue()
ret.label := label
end if
return ret
else
O := O.prototype.__proto__
end if
label := label.join(label)
end while
end procedure
procedure putDirect(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int base, String
property, int propVal)
JSValue baseValue := callStack.cFN.rf[base]. value
JSValue propValue := callStack.cFN.rf[propVal]. value
JSObject* obj := baseValue.toObject()
PropertyDescriptor dataPD := PropertyDescriptor::createPD(true, true, true)
dataPD.value := propValue
obj.setProperty(property, dataPD)
obj.structLabel := obj.structLabel.join(contextLabel)
h.o[++(h.location)] := *obj
return h
end procedure
procedure putIndirect(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int base,
String property, int val)
JSValue baseValue := callStack.cFN.rf[base].value
JSValue propValue := callStack.cFN.rf[val].value
JSObject* obj := baseValue.toObject()
bool isStrict := callStack.cFN.cfg.isStrictMode()
contextLabel := obj.structLabel.join(contextLabel)
if obj.containsPropertyInItself(property) && obj. getProperty(property).isDataProperty() &&
!isStrict && obj.isWritable() then
obj.getProperty(property).setValue(propValue)
h.o[++(h.location)] := *obj
return h
end if
return putDirect(contextLabel, callStack, h, base, property, val)
end procedure
procedure delById(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int base, Identifier property)
JSValue baseValue := callStack.cFN.rf[base].value
JSObject* obj := baseValue.toObject()
int loc := h.findObject(obj)
Property prop := obj.getProperty(property)
PropertyDescriptor pd := prop.getPropertyDescriptor()
if obj.getPropertyValue(prop).label ≥ contextLabel then
if !obj.containsPropertyInItself(property) then
retState.theta := h
retState.val := JSValue::construct(true)
return retState
end if
if obj.containsPropertyInItself(property) && prop. isConfigurable() then
if !(callStack.cFN.cfg.isStrictMode()) then
pd.value := JSValue::constructUndefined()
obj.setProperty(property, pd)
h.o[loc] := *obj
retState.theta := h
retState.val := JSValue::construct(true)
return retState
end if
end if
retState.theta := h
retState.val := JSValue::construct(false)
return retState
else
stop
end if
end procedure
procedure putGetterSetter(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int base,
Identifier property, JSValue getterValue, JSValue setterValue)
JSValue baseValue := callStack.cFN.rf[base].value
JSObject* obj := baseValue.toObject()
int loc := h.findObject(obj)
JSFunctionObject *getterObj, *setterObj
JSFunctionObject *getterFuncObj := null, *setterFuncObj := null
if !getterValue.isUndefined() then
getterFuncObj := getterValue.toFunctionObject (callStack.cFN.cfg, callStack.cFN.sc)
end if
if !setterValue.isUndefined() then
setterFuncObj := setterValue.toFunctionObject (callStack.cFN.cfg, callStack.cFN.sc)
end if
if getterFuncObj 6= null then
obj.setGetter(property, getterObj)
end if
if setterFuncObj 6= null then
obj.setSetter(setterObj)
end if
PropertyDescriptor accessor := PropertyDescriptor ::createPD(false, false, false, true)
JSValue v := JSValue::constructUndefined()
v.label := contextLabel
accessor.value := v
obj.setProperty(property, accessor)
obj.structLabel := contextLabel
h.o[loc] := *obj
return h
end procedure
procedure getPropNames(CallFrameStack* callStack, Instruction* ip, int base, int i, int size, int
breakOffset)
JSValue baseVal := callStack.cFN.rf[base].value
JSObject* obj := baseVal.toObject()
PropertyIterator* propItr := obj.getProperties()
if baseVal.isUndefined() || baseVal.isNull() then
retState.v1 := jsUndefined()
retState.v2 := jsUndefined()
retState.v3 := jsUndefined()
retState.ip := ip + breakOffset
return retState
end if
retState.v1 := JSValue::construct(propItr)
retState.v2 := JSValue::construct(0)
retState.v3 := JSValue::construct(propItr.size())
retState.ip := ip + 1
return retState
end procedure
procedure getNextPropName(CallFrameStack* cStack, Instruction* ip, JSValue base, int i, int
size, int iter, int offset, int dst)
JSObject* obj := base.toObject()
PropertyIterator* propItr := cStack.cFN.rf[iter].value. toPropertyIterator()
int b := rFile[i].value.toInteger()
int e := rFile[size].value.toInteger()
while b < e do
String key := propItr.get(b)
retState.value1 := JSValue::construct(b + 1)
if !(key.isNull()) then
retState.value2 := JSValue::construct(key)
ip := ip + offset
break
end if
b++
end while
return retState
end procedure
procedure resolveInSc(JSLabel contextLabel, ScopeChainNode* scopeHead, String property)
JSValue v
JSLabel l
ScopeChainNode* scn := scopeHead
while scn 6= NULL do
PropertySlot pSlot := scn.getPropertySlot()
if pSlot.contains(property) then
v := pSlot.getValue(property)
v.label := contextLabel
return v
end if
scn := scn.next
if scn.scObjType = VariableObject then
contextLabel = contextLabel.join(scn.Object. actObj.structLabel)
else if scn.scObjType = LexicalObject then
contextLabel = contextLabel.join(scn.Object. obj.structLabel)
end if
contextLabel := contextLabel.join(scn. scopeNextLabel)
end while
v := jsUndefined()
v.label := contextLabel
return v
end procedure
procedure resolveInScWithSkip(JSLabel contextLabel, ScopeChainNode* scopeHead, String property, int skip)
JSValue v
JSLabel l
ScopeChainNode* scn := scopeHead
while skip−− do
scn := scn.next
if scn.scObjType = VariableObject then
contextLabel := contextLabel.join( scn.Object. actObj.structLabel)
else if scn.scObjType = LexicalObject then
contextLabel := contextLabel.join(scn.Object. obj.structLabel)
end if
contextLabel := contextLabel.join(scn. scopeNextLabel)
end while
while scn 6= null do
PropertySlot pSlot := scn.getPropertySlot()
if pSlot.contains(property) then
v := pSlot.getValue(property)
v.label := contextLabel
return v
end if
scn := scn.next
if scn.scObjType = VariableObject then
contextLabel := contextLabel.join(scn.Object. actObj.structLabel)
else if scn.scObjType = LexicalObject then
contextLabel := contextLabel.join(scn.Object. obj.structLabel)
end if
contextLabel := contextLabel.join(scn. scopeNextLabel)
end while
v := jsUndefined()
v.label := contextLabel
return v
end procedure
procedure resolveGlobal(JSLabel contextLabel, CallFrameStack* cStack, String property)
JSValue v
struct CFG* cBlock := cStack.cFN.cfg
JSGlobalObject* globalObject := cBlock. getGlobalObject()
PropertySlot pSlot(globalObject)
if pSlot.contains(property) then
v := pSlot.getValue(property)
v.label := contextLabel
return v
end if
v := jsUndefined()
v.label := contextLabel
return v
end procedure
procedure resolveBase(JSLabel contextLabel, CallFrameStack* cStack, ScopeChainNode* scopeHead, String property, bool strict)
JSValue v
ScopeChainNode* scn := scopeHead
CFG *cBlock := cStack.cFN.cfg
JSGlobalObject *gObject := cBlock.globalObject
while scn 6= null do
JSObject* obj := scn.get()
contextLabel := obj.structLabel.join(contextLabel)
PropertySlot pSlot(obj)
if scn.next = null && strict && !pSlot.contains (property) then
v := emptyJSValue()
v.label := contextLabel
return v
end if
if pSlot.contains(property) then
v := JSValueContainingObject(obj)
v.label := contextLabel
return v
end if
scn := scn.next
if scn 6= null then
contextLabel := contextLabel.join(scn. scopeNextLabel)
end if
end while
v := JSValue::construct(gObject)
v.label := contextLabel
return v
end procedure
procedure resolveBaseAndProperty(JSLabel contextLabel, CallFrameStack cStack, int bRegister, int pRegister, String property)
JSValue v
ScopeChainNode* scn := cStack.cFN.sc
while scn 6= null do
JSObject* obj := scn.get()
contextLabel := obj.structLabel.join(contextLabel)
PropertySlot pSlot(obj)
if pSlot.contains(property) then
v := pSlot.getValue(property)
v.label := contextLabel
ret.val1 := v
v := JSValueContainingObject(obj)
v.label := contextLabel
ret.val2 := v
return ret
end if
scn := scn.next
if scn 6= null then
contextLabel := contextLabel.join(scn. scopeNextLabel)
end if
end while
end procedure
procedure getScopedVar(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int index,
int skip)
JSValue v
ScopeChainNode* scn := callStack.cFN.sc
while skip−− do
if scn.scObjType = VariableObject then
contextLabel := contextLabel.join(scn.Object.actObj. structLabel)
else if scn.scObjType = LexicalObject then
contextLabel := contextLabel.join(scn.Object.obj. structLabel)
end if
contextLabel := contextLabel.join(scn. scopeLabel)
scn := scn.next
end while
v := scn.registerAt(index)
if scn.scObjType = VariableObject then
v.label := contextLabel.join(scn.Object.actObj. structLabel)
else if scn.scObjType = LexicalObject then
v.label := contextLabel.join(scn.Object.obj. structLabel)
end if
return v
end procedure
procedure putScopedVar(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int index,
int skip, int value)
CallFrameStack* cStack
ScopeChainNode* scn := callStack.cFN.sc
JSValue val := callStack.cFN.rf[value].value
while skip−− do
if scn.scObjType = VariableObject then
contextLabel := contextLabel.join(scn.Object. actObj.structLabel)
else if scn.scObjType = LexicalObject then
contextLabel := contextLabel.join(scn.Object. obj.structLabel)
end if
contextLabel := contextLabel.join(scn. scopeLabel)
scn := scn.next
end while
cStack := scn.setRegisterAt(contextLabel, index, val)
return cStack
end procedure
procedure pushScope(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int scope)
ScopeChainNode* sc := callStack.cFN.sc
JSValue v := callStack.cFN.rf[scope].value
JSObject* o := v.toObject()
SChainObject* scObj
if sc.scopeLabel ≥ contextLabel then
scObj.obj := o
sc.push(scObj, LexicalObject, contextLabel)
callStack.cFN.sc := sc
else if sc.scopeLabel = star then
scObj.obj := o
sc.push(scObj, LexicalObject, star)
callStack.cFN.sc := sc
end if
return callStack
end procedure
procedure popScope(JSLabel contextLabel, CallFrameStack* callStack, Heap* h)
ScopeChainNode* sc := callStack.cFN.sc
JSLabel l := sc.scopeLabel
if l ≥ contextLabel then
sc.pop()
callStack.cFN.sc := sc
else
stop
end if
return callStack
end procedure
procedure jmpScope(JSLabel contextLabel, CallFrameStack* callStack, Heap* h, int count)
ScopeChainNode* sc := callStack.cFN.sc
while count−− > 0 do
JSLabel l := sc.scopeLabel
if l ≥ contextLabel then
sc.pop()
callStack.cFN.sc := sc
else
stop
end if
end while
return callStack
end procedure
procedure throwException(CallFrameStack* callStack, CFGNode* iota)
CFGNode* handler
while callStack.cFN.hasHandler()==false do
callStack.pop()
end while
while callStack.cFN.sc.length() - callStack.cFN.getHandlerScopeDepth() do
callStack.cFN.sc.pop()
end while
handler := callStack.cFN.getHandler(iota)
interpreter.iota := handler
interpreter.sigma := callStack
end procedure
8.3
Semantics
ι = “prim dst:r src1:r src2:r”
L := Γ (!σ(src1)) t Γ (!σ(src2)) t Γ (!ρ)
V := Υ (!σ(src1)) ⊕ Υ (!σ(src2))
(Γ
(!σ(dst))
≥
Γ
(!ρ))
⇒
(L
:=
L) (L := ?)
σ 0 := σ
prim:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
prim reads the values from two registers (src1 and src2), performs the binary
operation generically denoted by ⊕, and writes the result into the dst register.
The label assigned to the value in dst register is the join of the label of value in
src1, src2 and the head of the pc-stack (!ρ). In order to avoid implicit leak of
information, the label of the existing value in dst is compared with the current
context label. If the label is lower than the context label, the label of the value
in dst is set to ?.
σ 0 := σ
mov:
ι = “mov dst:r src:r”
L := Γ (!σ(src)) t Γ (!ρ)
V = Υ (!σ(src))
(Γ (!σ(dst))
≥
Γ
(!ρ))
⇒
(L
:=
L) (L := ?)
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
mov copies the value from the src register to the dst register. The label assigned
to the value in dst register is the join of the label of value in src and the head
of the pc-stack (!ρ). In order to avoid implicit leak of information, the label of
the existing value in dst is compared with the current context label. If the label
is lower than the context label, the label of the value in dst is joined with ?.
jfalse:
ι = “jfalse cond:r target:offset”
Γ (!σ(cond)) 6= ?
ρ00 := ρ.push(Γ (!σ(cond)) t Γ (!ρ), IPD(ι), CF(ι), false)
Υ (!σ(cond)) = false ⇒ ι0 := Left(!σ.CFG, ι) ι0 := Right(!σ.CFG, ι)
ρ0 := isIPD(ι0 , ρ00 , σ)
ι, θ, σ, ρ
ι0 , θ, σ, ρ0
jfalse is a branching instruction. Based on the value in the cond register, it
decides which branch to take. The operation is performed only if the value in
cond is not labelled with a ?. If it contains a ?, we terminate the execution to
prevent possible leak of information. The push function defined in the rule does
the following: A node is pushed on the top of the pc-stack containing the IPD
of the branching instruction and the label of the value in cond joined with the
context, to define the context of this branch. If the IPD of the instruction is SEN
or the same as the top of the pc-stack, then we just join the label on top of the
pc-stack with the context label determined by the cond register.
loop-if-less:
ι = “loop-if-less src1:r src2:r target:offset”
Γ (!σ(src1)) 6= ?
Γ (!σ(src2)) 6= ?
L := Γ (!σ(src1)) t Γ (!σ(src2)) t Γ (!ρ)
Υ (!σ(src1)) < Υ (!σ(src2)) ⇒ ι0 := Left(!σ.CFG, ι) ι0 := Right(!σ.CFG, ι)
00
ρ := ρ.push(L, IPD(ι), CF(ι), false)
ρ0 := isIPD(ι0 , ρ00 , σ)
ι, θ, σ, ρ
ι0 , θ, σ, ρ0
loop-if-less is another branching instruction. If the value of src1 is less than
src2, then it jumps to the target, else continues with the next instruction. The
operation is performed only if the values in src1 and src2 are not labelled with
a ?. If any one of them contains a ?, we abort the execution to prevent possible
leak of information. The push function defined in the rule does the following: A
node is pushed on the top of the pc-stack containing the IPD of the branching
instruction and the join of the label of the values in src1 and src2 joined with
the context, to define the context of this branch. If the IPD of the instruction is
SEN or the same as the top of the pc-stack, then we just join the label on top
of the pc-stack with the context label determined above.
σ 0 := σ
typeof:
ι = “typeof dst:r src:r”
L := (Γ (src) t Γ (!ρ))
V := determineType(!σ(src))
(Γ (!σ(dst))
≥
Γ
(!ρ))
⇒ (L := L) (L := ?)
Υ (!σ(dst)):=V
ι0 := Succ(!σ 0 .CFG, ι)
Γ (!σ(dst)):=L
0
ι, θ, σ, ρ
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
typeof determines the type string for src according to ECMAScript rules, and
puts the result in register dst. We do a deferred NSU check on dst before writing
the result in it. The determineType function returns the data type of the value
passed as the parameter.
ι = “instanceof dst:r value:r cProt:r”
v := isInstanceOf (Γ (!ρ), !σ(value), cProt)
L = Γ (v)
(Γ
(!σ(dst))
≥
Γ
(!ρ))
⇒
(L
:=
L)
(L := ?),
σ 0 := σ
instanceof:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
0
ι, θ, σ, ρ
0
ι , θ, σ , ρ
V = Υ (v),
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
instanceof tests whether the cProt is in the prototype chain of the object in
register value and puts the Boolean result in the dst register after deferred NSU
check.
enter:
ι = “enter”
ι0 := Succ(!σ.CFG, ι)
ρ0 := isIPD(ι0 , ρ, σ)
ι0 , θ, σ, ρ0
ι, θ, σ, ρ
enter marks the beginning of a code block.
ret:
ι = “ret res:r”
(ι0 , σ 0 , γ) := opRet(σ, res)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
ret is the last instruction to be executed in a function. It pops the call-frame and
returns the control to the callee’s call-frame. The return value of the function is
written to a local variable in the interpreter (γ), which can be read by the next
instruction being executed.
end:
ι = “end res:r”
ι, θ, σ, ρ
opEnd(σ, res)
−
end marks the end of a program. opEnd passes the value present in res register
to the caller (the native function that invoked the interpreter).
00
call:
ρ
ι = “op-call func:r args:n”
Γ (func) 6= ?
(ι0 , σ 0 , H, `f ) := opCall(σ, ι, func, args)
:= ρ.push(`f t Γ (!σ(func)) t Γ (!ρ), IPD(ι), CF(ι), H)
ρ0 := isIPD(ι0 , ρ00 , σ 0 )
ι0 , θ, σ 0 , ρ0
ι, θ, σ, ρ
call, initially, checks the function object’s label for ? and if the label contains a
?, the program execution is aborted. The reason for termination is the possible
leak of information as explained above. If not, call creates a new call-frame,
copies the arguments, initializes the registers, scope-chain pointer, codeblock
and the return address. The registers are initialized to undefined and assigned
a label obtained by joining the label of the context in which the function was
created and the label of the function object itself. We treat call as a branching
instruction and hence, push a new node on the top of the pc-stack with the
label determined above along with its IPD and call-frame. The field H in the
push function is determined by looking up the exception table. If it contains
an associated exception handler, it sets the field to true else it is set to false.
If the IPD is the SEN then we just join the label on the top of the stack with
the currently calculated label. It then points the instruction pointer to the first
instruction of the new code block.
σ 0 := σ
call-put-result:
ι = “call-put-result res:r”
L := Γ (γ) t Γ (!ρ)
V := Υ (γ)
(Γ (!σ(res))
≥ Γ (!ρ)) ⇒ L := L L := ?
Υ (!σ(res)):=V
ι0 := Succ(!σ 0 .CFG, ι)
Γ (!σ(res)):=L
ι, θ, σ, ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
call-put-result copies the return value γ to the res register. The label assigned
to the value in res register is the join of the label of the return value and the
head of the pc-stack. In order to avoid implicit leak of information, deferred
no-sensitive-upgrade is performed.
call-eval:
ρ00
ι = “call-eval func:r args:n”
Γ (!σ(func)) 6= ?
(ι0 , σ 0 , H, `f ) := opCallEval(Γ(!ρ), σ, ι, func, args)
:= ρ.push(`f t Γ (!σ(func)) t Γ (!ρ), IPD(ι), CF(ι), H)
ρ0 := isIPD(ι0 , ρ00 , σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
call-eval calls a function with the string passed as an argument converted to a
code block. If func register contains the original global eval function, then it is
performed in local scope, else it is similar to call.
ι = “create-arguments dst:r”
(θ 0 , v) := createArguments(θ, σ)
L := Γ (!ρ)
V := Υ (v)
(Γ
(!σ(dst))
≥
Γ
(!ρ))
⇒
L
:=
L
L
:=
?
σ 0 := σ
create-arguments:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
0
ι ,θ ,σ ,ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
create-arguments creates the arguments object and places its pointer in the local
dst register after the deferred NSU check. The label of the arguments object is
set to the context.
ι = “new-func dst:r funcIndex:f”
(θ 0 , v) := newFunc(σ, θ, funcIndex, Γ(!ρ))
L := Γ (v) t Γ (!ρ)
(Γ (!σ(dst))
≥ Γ (!ρ)) ⇒ L := L L := ?
σ 0 := σ
new-func:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
V := Υ (v)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ 0 , σ 0 , ρ0
new-func constructs a new function instance from function at funcIndex and the
current scope chain and puts the result in dst after deferred NSU check.
ι = “create-activation dst:r”
(σ 0 , v) := createActivation(σ, Γ(!ρ))
L := Γ (v) t Γ (!ρ)
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ L := L L := ?
σ 0 := σ
create-activation:
Υ (!σ00 (dst)):=V
Γ (!σ 00 (dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
V := Υ (v)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
create-activation creates the activation object for the current call-frame if it has
not been already created and writes it to the dst after the deferred NSU check
and pushes the object in the scope-chain. If the label of the head of the existing
scope-chain is less than the context, then the label of the pushed node is set to
?, else it is set to the context.
00
construct:
ρ
ι = “construct func:r args:n”
Γ (!σ(func)) 6= ?
(ι0 , σ 0 , H, `f ) := opCall(σ, ι, func, args)
0
:= ρ.push(`f t Γ (!σ (func)) t Γ (!ρ), IPD(ι), CF(ι), H)
ρ0 := isIPD(ι0 , ρ00 , σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
construct invokes register func as a constructor and is similar to call. For JavaScript functions, the this object being passed (the first argument in the list of
arguments) is a new object. For host constructors, no this is passed.
ι = “create-this dst:r”
(θ 0 , v) := createThis(Γ(!ρ), σ, θ)
L := Γ (v) t Γ (!ρ)
(Γ (!σ(dst))
≥ Γ (!ρ)) ⇒ L := L L := ?
σ 0 := σ
create-this:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
V := Υ (v)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ 0 , σ 0 , ρ0
create-this creates and allocates an object as this used for construction later in
the function. The object is labelled the context and placed in dst after deferred
NSU check. The prototype chain pointer is also labelled with the context label.
ι = “new-object dst:r”
(θ 0 , v) := newObject(θ, Γ (!ρ))
L := Γ (v) t Γ (!ρ)
V := Υ (v)
(Γ (!σ(dst))
≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
new-object:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ 0 , σ 0 , ρ0
new-object constructs a new empty object instance and puts it in dst after deferred NSU check. The object is labelled with the context label and the prototype
chain pointer is also labelled with the context.
ι = “get-by-id dst:r base:r prop:id vdst:r”
v := getPropertyById(!σ(base), prop, vdst)
L := Γ (v) t Γ (!ρ)
(Γ (!σ(dst))
≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
get-by-id:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
V := Υ (v)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
get-by-id gets the property named by the identifier prop from the object in the
base register and puts it into the dst register after the deferred NSU check. If
the object does not contain the property, it looks up the prototype chain to
determine if any of the proto objects contain the property. When traversing the
prototype chain, the context is joined with the structure label of all the objects
and the prototype chain pointer labels until the property is found or the end of
the chain. It then joins the property label to the context. If the property is not
found, it returns undefined. The joined label of the context is the label of the
property put in the dst register.
If the property is an accessor property, it calls the getter function, sets the
getter flag in the call-frame and updates the destination register field with the
register where the value is to be inserted. It then transfers the control to the
first instruction in the getter function.
put-by-id:
ι = “put-by-id base:r prop:id value:r direct:b”
Γ (!σ(value)) 6= ?
(direct = true ⇒ θ 0 := putDirect(Γ (!ρ), σ, θ, base, prop, value)
θ 0 := putIndirect(Γ (!ρ), σ, θ, base, prop, value))
ι0 := Succ(!σ.CFG, ι)
ρ0 := isIPD(ι0 , ρ, σ)
ι, θ, σ, ρ
ι0 , θ 0 , σ, ρ0
put-by-id writes into the heap the property of an object. We check for ? in
the label of value register. If it contains a ?, the program aborts as this could
potentially result in an implicit information flow. If not, it writes the property
into the object. The basic functionality is to search for the property in the object
and its prototype chain, and change it. If the property is not found, a new
property for the current object with the property label as the context is created.
Based on whether the property is in the object itself (or needs to be created in
the object itself) or in the prototype chain of the object, it calls putDirect and
putIndirect, respectively.
ι = “del-by-id dst:r base:r prop:id”
Γ (!σ(base)) 6= ?
(θ 0 , v) := delById(Γ (!ρ), σ, θ, base, prop)
L := Γ (v) t Γ (!ρ)
V := Υ (v) (Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
del-by-id:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
0
ι ,θ ,σ ,ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
del-by-id deletes the property specified by prop in the object contained in base.
If the structure label of the object is less than the context, the deletion does not
happen. If the property is found, the property is deleted and Boolean value true
is written to dst, else it writes false to dst. The label of the Boolean value is the
structure label of the object joined with the property label.
getter-setter:
ι = “put-getter-setter base:r prop:id getter:r setter:r”,
?∈
/ Γ (!σ(getter)), ? ∈
/ Γ (!σ(setter)),
θ 0 := putGetterSetter(Γ(!ρ), σ, base, prop, !σ(getter), !σ(setter)),
ι0 := Succ(!σ.CFG, ι), ρ0 := isIPD(ι0 , ρ)
ι0 , θ 0 , σ, ρ0
ι, θ, σ, ρ
getter-setter puts the accessor descriptor to the object in register base. It initially
checks if the structure label of the object is greater or equal to the context. The
property for which the accessor properties are added is given in the register prop.
The property label of the accessor functions is set to the context. putGetterSetter
calls putIndirect internally and sets the getter/setter property of the object with
the specified value.
ι = “get-pnames dst:r base:r i:r size:r breakTarget:offset”
Γ (!σ(base)) 6= ?
(v1 , v2 , v3 , ι0 ) := getPropNames(σ, ι, base, i, size, breakTarget)
Ln := Γ (!σ(base)) t Γ (vn ) t Γ (!ρ)
Vn := Υ (vn ), n := 1, 2, 3
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L1 := L1 ) (L1 := ?)
(Γ (!σ(i)) ≥ Γ (!ρ)) ⇒ (L2 := L2 ) (L2 := ?)
(Γ (!σ(size)) ≥ Γ (!ρ)) ⇒ (L3 := L3 ) (L3 := ?)
σ 00 := σ
get-pnames:
Υ (!σ(dst)):=V1
Γ (!σ(dst)):=L1
σ 000 := σ 00
Υ (!σ00 (i)):=V2
Γ (!σ 00 (i)):=L2
σ 0 := σ 000
Υ (!σ000 (size)):=V3
Γ (!σ 000 (size)):=L3
vn = undefined ⇒ (L = Γ (!σ(base)))
(L = Γ (!σ(base)) t Γ(θ(!σ(base))) t (∀p ∈ Prop(θ(!σ(base))).Γ(p)))
ρ00 = ρ.push(L, IPD(ι), CF(ι), false)
ρ0 := isIPD(ι0 , ρ00 , σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
get-pnames creates a property name list for object in register base and puts it in
dst, initializing i and size for iteration through the list, after the deferred NSU
check. If base is undefined or null, it jumps to breakTarget. It is a branching
instruction and pushes the label with join of all the property labels and the
structure label of the object along with the IPD on the pc-stack. If the IPD of
the instruction is SEN or the same as the top of the pc-stack, then we just join
the label on top of the pc-stack with the context label determined above.
ι = “next-pname dst:r base:r i:n size:n iter:n target:offset”
(v1 , v2 , ι0 ) := getNextPropNames(σ, ι, base, i, size, iter, target)
Ln := Γ (vn ) t Γ (!ρ)
Vn := Υ (vn ), n := 1, 2
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L1 := L1 ) (L1 := ?)
(Γ (!σ(i)) ≥ Γ (!ρ)) ⇒ (L2 := L2 ) (L2 := ?)
σ 00 := σ
next-pname:
Υ (!σ(dst)):=V1
Γ (!σ(dst)):=L1
σ 0 := σ 00
ι, θ, σ, ρ
Υ (!σ00 (i)):=V2
Γ (!σ 00 (i)):=L2
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
next-pname copies the next name from the property name list created by getpnames in iter to dst after deferred NSU check, and jumps to target. If there are
no names left, it continues with the next instruction. Although, it behaves as a
branching instruction, the context pertaining to this opcode is already pushed
in get-pnames. Also, the IPD corresponding to this instruction, is the same as
the one determined by get-pnames. Thus, we do not push on the pc-stack in this
instruction.
L := Γ (v)
t Γ (!ρ)
σ 0 := σ
resolve:
ι = “resolve dst:r prop:id”
v := resolveInSc(Γ(!ρ), !σ.sc, prop)
V := Υ (v)
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
resolve searches for the property in the scope chain and writes it into dst register,
if found. The label of the property written in dst is a join of the context label, all
the nodes (structure label of the object contained in it) traversed in the scope
chain and the label associated with the pointers in the chain until the node
(object) where the property is found. If the initial label of the value contained in
dst was lower than the context label, then the label of the value in dst is joined
with ?. In case the property is not found, the instruction throws an exception
(similar to throw, as described later).
ι = “resolve-skip dst:r prop:id skip:n”
v := resolveInScWithSkip(Γ(!ρ), !σ.sc, prop, skip)
L := Γ (v)
t Γ (!ρ)
V := Υ (v)
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
resolve-skip:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
resolve-skip looks up the property named by prop in the scope chain similar to
resolve, but it skips the top skip levels and writes the result to register dst. If
the property is not found, it also raises an exception and behaves similarly to
resolve.
L := Γ (v)
t Γ (!ρ)
σ 0 := σ
resolve-global:
ι = “resolve-global dst:r prop:id”
v := resolveGlobal(Γ(!ρ), σ, prop)
V := Υ (v)
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
resolve-global looks up the property named by prop in the global object. If the
structure of the global object matches the one passed here, it looks into the
global object. Else, it falls back to perform a full resolve.
ι = “resolve-base dst:r prop:id isStrict:bool”
v := resolveBase(Γ(!ρ), σ, !σ.sc, prop, isStrict)
L := Γ (v)
t Γ (!ρ)
V := Υ (v)
(Γ (!σ(dst)) ≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
resolve-base:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
resolve-base looks up the property named by prop in the scope chain similar to
resolve but writes the object to register dst. If the property is not found and
isStrict is false, the global object is stored in dst.
ι = “resolve-with-base bDst:r pDst:r prop:id”
(bdst, pdst) := resolveBaseAndProperty(Γ(!ρ), σ, baseDst, propDst, prop)
L1 := Γ (!σ(bdst)) t Γ(!ρ)
V1 := Υ(!σ(bdst))
L2 := Γ (!σ(pdst)) t Γ(!ρ)
V2 := Υ(!σ(pdst))
(Γ (!σ(bDst)) ≥ Γ (!ρ)) ⇒ (L1 := L1) (L1 := ?)
(Γ (!σ(pDst)) ≥ Γ (!ρ)) ⇒ (L2 := L2) (L2 := ?)
σ 00 := σ
Υ (!σ(bDst)):=V1
σ 0 := σ 00
Γ (!σ(bDst)):=L1
0
ι0 := Succ(!σ .CFG, ι)
resolve-with-base:
Υ (!σ00 (pDst)):=V2
Γ (!σ 00 (pDst)):=L2
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
ι, θ, σ, ρ
resolve-with-base looks up the property named by prop in the scope chain similar
to resolve-base and writes the object to register bDst. It also, writes the property
to pDst. If the property is not found it raises an exception like resolve.
ι = “get-scoped-var dst:r index:n skip:n”
v := getScopedVar(Γ(!ρ), σ, θ, index, skip)
L := Γ (v) t Γ (!ρ)
(Γ (!σ(dst))
≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
σ 0 := σ
get-scoped-var:
Υ (!σ(dst)):=V
Γ (!σ(dst)):=L
ι0 := Succ(!σ 0 .CFG, ι)
V := Υ (v)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
ι, θ, σ, ρ
get-scoped-var loads the contents of the index local from the scope chain skipping
skip nodes and places it in dst, after deferred NSU. The label of the value in
dst includes the join of the current context along with all the structure label of
objects in the skipped nodes.
put-scoped-var:
ι = “put-scoped-var index:n skip:n value:r”
Γ (!σ(value)) 6= ?
σ 0 := putScopedVar(Γ(!ρ), σ, θ, index, skip, value)
ι0 := Succ(!σ 0 .CFG, ι)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
ι, θ, σ, ρ
put-scoped-var puts the contents of the value in the index local in the scope
chain skipping skip nodes. The label of the value includes the join of the current
context along with the structure label of all the objects in the skipped nodes.
ι = “push-scope scope:r”
σ 0 := pushScope(Γ(!ρ), σ, scope)
ι0 := Succ(!σ 0 .CFG, ι)
push-scope:
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
ι, θ, σ, ρ
push-scope converts scope to object and pushes it onto the top of the current
scope chain. The contents of the register scope are replaced by the created object.
The scope chain pointer label is set to the context.
pop-scope:
σ 0 := popScope(Γ (!ρ), σ)
ι = “pop-scope”
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
ρ0 := isIPD(ι0 , ρ, σ 0 )
0
pop-scope removes the top item from the current scope chain if the scope chain
pointer label is greater than or equal to the context.
jmp-scope:
ι = “jmp-scope count:n target:n”
σ 0 := jmpScope(Γ(!ρ), σ, count)
ι0 := Succ(!σ 0 .CFG, ι)
ι, θ, σ, ρ
0
0
ι , θ, σ , ρ
0
ρ0 := isIPD(ι0 , ρ, σ 0 )
jmp-scope removes the top count items from the current scope chain if the scope
chain pointer label is greater than or equal to the context. It then jumps to offset
specified by target.
throw:
ι = “throw ex:r”
excValue := Υ (!σ(ex))
(σ 0 , ι0 ) := throwException(σ, ι)
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι, θ, σ, ρ
ι0 , θ, σ 0 , ρ0
throw throws an exception and points to the exception handler as the next
instruction to be executed, if any. The exception handler might be in the same
function or in an earlier function. If it is not present, the program terminates. If
it has an exception handler, it has an edge to the synthetic exit node. Apart from
this, throwException pops the call-frames from the call-stack until it reaches the
call-frame containing the exception handler. It writes the exception value to a
local interpreter variable (excValue), which is then read by catch.
L := Γ (excValue)
t Γ (!ρ)
σ 0 := σ
0
catch:
ι = “catch ex:r”
(Γ (!σ(ex))
≥ Γ (!ρ)) ⇒ (L := L) (L := ?)
Υ (!σ(ex)):=Υ (excValue))
Γ (!σ(ex)):=L
0
ι := Succ(!σ .CFG, ι)
ι, θ, σ, ρ
excValue := empty
ρ0 := isIPD(ι0 , ρ, σ 0 )
ι0 , θ, σ 0 , ρ0
catch catches the exception thrown by an instruction whose handler corresponds
to the catch block. It reads the exception value from excValue and writes into
the register ex. If the label of the register is less than the context, a ? is joined
with the label. It then makes the excValue empty and proceeds to execute the
first instruction in the catch block.
8.4
Proofs and Results
The fields in a frame of the pc-stack are denoted by the following symbols: !ρ.ipd
represents the IPD field in the top frame of the pc-stack, Γ (!ρ) returns the label
field in the top frame of the pc-stack, and !ρ.C returns the call-frame field in the
top frame of the pc-stack.
In the definitions and proofs that follow, we assume that the level of attacker
is L in the three-element lattice presented earlier, i.e., in the equivalence relation ∼β` , ` = L. The level of the attacker is omitted for clarity purposes from
definitions and proofs.
Definition 1 (Partial bijection) A partial bijection β is a binary relation on
heap locations satisfying the following properties: (1) if (a, b) ∈ β and (a, b0 ) ∈ β,
then b = b0 , and (2) if (a, b) ∈ β and (a0 , b) ∈ β, then a = a0 .
Using partial bijections, we define equivalence of values, labeled values and
objects.
Definition 2 (Value equivalence) Two values r1 and r2 are equivalent up to
β, written r1 ∼β r2 if either (1) r1 = a, r2 = b and (a, b) ∈ β, or (2) r1 = r1 = v
where v is some primitive value.
Definition 3 (Labeled value equivalence) Two labeled values v1 = r1`1 and
v2 = r2`2 are equivalent up to β, written v1 ∼β v2 if one of the following holds:
(1) `1 = ? or `2 = ?, or (2) `1 = `2 = H, or (3) `1 = `2 = L and r1 ∼β r2 .
The first clause of the above definition is standard for the permissive-upgrade
check. It equates a partially leaked value to every other labeled value.
Objects are formally denoted as N = ({pi 7→ {vi , flags i }}ni=0 , __proto__ 7→
a`p , `s ) . Here pi s correspond to the property name, vi s are their respective
values and flags i represent the writable, enumerable and configurable flags as
described in the PropertyDescriptor structure in the cpp model above. As the
current model does not allow modification of the flags, they are always set to
true. Thus, we do not need to account for the flags i in the equivalence definition
below. __proto__ represents a labelled pointer to the object’s prototype.
Definition 4 (Object equivalence) For ordinary objects N = ({pi 7→ {vi ,
flags i }}ni=0 , __proto__ 7→ a`p , `s ) and N 0 = ({p0i 7→ {vi0 , flags 0i }}m
i=0 , __proto__
0
7→ a0`p , `0s ), we say N ∼β N 0 iff either `s = `0s = H or the following hold:
(1) `s = `0s = L, (2) [p0 , . . . , pn ] = [p00 , . . . , p0m ] (in particular, n = m), (3)
0
∀i. vi ∼β vi0 , and (4) a`p ∼β a`p .
For function objects F = (N, f, Σ) and F 0 = (N 0 , f 0 , Σ 0 ), we say F ∼β F 0 iff
either N.`s = N.`0s = H or N ∼β N 0 , f =β f 0 and Σ ∼β Σ 0 .
The equality f =β f 0 of nodes f, f 0 in CFGs means that the portions of the
CFGs reachable from f, f 0 are equal modulo renaming of operands to bytecodes
under β. Equivalence Σ ∼β Σ 0 of scope chains is defined below. Because we
do not allow ? to flow into heaps, we do not need corresponding clauses in the
definition of object equivalence.
Definition 5 (Heap equivalence) For two heaps θ1 , θ2 , we say that θ1 ∼β θ2
iff ∀(a, b) ∈ β. θ1 (a) ∼β θ2 (b).
Unlike objects, we allow ? to permeate scope chains, so our definition of
scope chain equivalence must account for it. Scope chains are denoted as Σ. A
scope-chain node contains a label ` along with an object S (either JSActivation
or JSObject) represented as (S, `).
Definition 6 (Scope chain equivalence) For two scope chain nodes S, S 0 ,
we say that S ∼β S 0 if one of the following holds: (1) S = O, S 0 = O0 and
O ∼β O0 , or (2) S = v0 : . . . : vn , S 0 = v00 : . . . : vn0 and ∀i. vi ∼β vi0 .
Equivalence of two scope chains Σ, Σ 0 is defined by the following rules. (1)
nil ∼β nil (2) (nil ∼β (S, `)) if ` = H or ` = ? (3) ((S, `) ∼β nil) if ` = H
or ` = ? and (4) ((S, `) : Σ) ∼β ((S 0 , `0 ) : Σ 0 ) if one of the following holds: (a)
` = ? or `0 = ?, (b) ` = `0 = H, or (c) ` = `0 = L, S ∼β S 0 and Σ ∼β Σ 0 .
Definition 7 (Call-frame equivalence) For two call frames µ1 , µ2 , we say
µ1 ∼β µ2 iff (1) #Registers(µ1 ) = #Registers(µ2 ), (2) ∀i. µ1 .Registers[i] ∼β
µ2 .Registers[i], (3) µ1 .CFG =β µ2 .CFG, (4) µ1 .Scopechain ∼β µ2 .Scopechain,
(5) µ1 .ιr = µ2 .ιr (6) (µ1 .`c = µ2 .`c = H) ∨ (µ1 .`c = µ2 .`c = L ∧ µ1 .fcallee ∼β
µ2 .fcallee ) (7) µ1 .argcount = µ2 .argcount (8) µ1 .getter = µ2 .getter and (9)
µ1 .dReg =β µ1 .dReg
Note that a register is simply a labeled value in our semantics, so clause (2)
above is well-defined.
Definition 8 (pc-stack equivalence) For two pc-stacks ρ1 , ρ2 , we say ρ1 ∼ ρ2
iff the corresponding nodes of ρ1 and ρ2 having label L are equal, except for the
call-frame (C) field.
In proofs that follow, two pc-stack nodes are equal if their respective fields
are equal, except for the call-frame (C) field.
Definition 9 (Call-stack equivalence) Given ρ1 ∼ ρ2 , suppose:
e1 is the lowest H-labelled node in ρ1
e2 is the lowest H-labelled node in ρ2
µ1 is the node of σ1 pointed to by e1
µ2 is the node of σ2 pointed to by e2
σ10 is prefix of σ1 up to and including µ1 or
if Γ (!ρ1 ) = L or ρ1 is empty, σ10 = σ1
6. σ20 is prefix of σ2 up to and including µ2 or
if Γ (!ρ2 ) = L or ρ2 is empty, σ20 = σ2
1.
2.
3.
4.
5.
then σ1 ∼βρ1 ,ρ2 σ2 , iff (1) |σ10 | = |σ20 |, and (2) ∀i ≤ |σ10 |.(σ10 [i] ∼β σ20 [i]).
Definition 10 (State equivalence) Two states s1 = hι1 , θ1 , σ1 , ρ1 i and s2 =
hι2 , θ2 , σ2 , ρ2 i are equivalent, written as s1 ∼β s2 , iff ι1 = ι2 , ρ1 ∼ ρ2 , θ1 ∼β θ2 ,
and σ1 ∼βρ1 ,ρ2 σ2 .
Lemma 1 (Confinement Lemma) If hι, θ, σ, ρi
hι0 , θ0 , σ 0 , ρ0 i and Γ (!ρ) =
β
0
0
β 0
H, then ρ ∼ ρ , σ ∼ρ,ρ0 σ and θ ∼ θ where β = {(a, a) | a ∈ θ}
Proof. As Γ (!ρ) = H, the L labelled nodes in the pc-stack will remain unchanged. Branching instructions pushing a new node would have label H due
to monotonicity of pc-stack. Even if ι0 is the IPD corresponding to the !ρ.ipd,
it would only pop the H labelled node. Thus, the L labelled nodes will remain
unchanged. Hence, ρ ∼ ρ0 .
We assume that the !ρ is the first node labelled H in the context stack. For,
other higher labelled nodes above the first node labelled H in the pc-stack, the
call-frames corresponding to the nodes having L label in the pc-stack remain the
same. Hence, σ ∼βρ,ρ0 σ 0 .
By case analysis on the instruction type:
1. prim:
(a) If Γ (!σ(dst)) ≥ Γ (!ρ), then Γ (!σ(dst)) = H.
By premise of prim, Γ (!σ 0 (dst)) = H. By Definition 3, !σ(dst) ∼β
!σ 0 (dst).
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
(b) If Γ (!σ(dst)) < Γ (!ρ), then Γ (!σ 0 (dst)) will contain a ? and by Definition 3, !σ(dst) ∼β !σ 0 (dst).
Only dst changes in the call-frame, so by Definition 7, !σ ∼β !σ 0 . Also, other
call-frames remain unchanged. By Definition 9, σ ∼βρ,ρ0 σ 0 .
θ = θ0 , thus, θ ∼β θ0 .
mov: Similar to prim.
jfalse: σ = σ 0 and θ = θ0 , so, σ ∼βρ,ρ0 σ 0 and θ ∼β θ0 .
loop-if-less: Similar to jfalse.
typeof : Similar to prim.
instanceof : Similar to prim.
enter: σ = σ 0 , so σ ∼βρ,ρ0 σ 0 . θ = θ0 , so θ ∼β θ0 .
ret: If !σ.getter = f alse then only !σ is popped, the call-frames until (!ρ.C)
are unchanged. When !σ.getter = true, then it sets !σ 0 (!σ.dReg) with !σ(res).
Now, let σ1 is the prefix of σ such that !ρ.C =β !σ1 . If !σ 0 ∈
/ σ1 then
changes in !σ 0 does not effect the callframe equivalence and if !σ 0 ∈ σ1 then
Γ (!σ 0 (!σ.dReg)) = ? (when Γ (!σ1 (!σ.dReg))) = L or Γ (!σ1 (!σ.dReg))) = ?)
and Γ (!σ 0 (!σ.dReg))) = H (when Γ (!σ1 (!σ.dReg))) = H), each of the cases
give !σ1 (!σ.dReg) ∼β !σ 0 (!σ.dReg) from Definition 3. So, σ ∼βρ,ρ0 σ 0 , by Definition 9. θ = θ0 , so θ ∼β θ0 .
end: The confinement lemma does not apply.
call: If it pushes on top of pc-stack, !ρ.C is the lowest H-labelled node in
ρ0 . If it joins the label with !ρ, the L labelled nodes remain unchanged and
the !ρ.C = !ρ0 .C. All the call-frames until !ρ.C remain unchanged. So, by
Definition 9, σ ∼βρ,ρ0 σ 0 . θ = θ0 , so θ ∼β θ0 .
call-put-result: Similar to prim.
call-eval: If it is a user-defined eval, it is similar to call.
In strict mode, it pushes a node on scope-chain with label H if Γ (!σ 0 .Σ) = H,
else labels it ?. In non-strict mode, it does not push a node on the scope-chain.
!σ remains equivalent with corresponding call-frame in σ 0 by Definition 7.
As other L call-frames are unchanged, by Definition 9, σ ∼βρ,ρ0 σ 0 .
θ = θ0 , so θ ∼β θ0 .
create-arguments: Over the initial β, by Definition 5, θ ∼β θ0 . If the argument
object is created at x, then β = (x, x) ∪ β after the step is taken.
σ ∼βρ,ρ0 σ 0 (Similar to prim).
new-func: Over the initial β, by Definition 5, θ ∼β θ0 . If the function object
is created at x, then β = (x, x) ∪ β after the step is taken.
σ ∼βρ,ρ0 σ 0 (Similar to prim).
create-activation: Over the initial β, by Definition 5, θ ∼β θ0 . If the argument
objects is created at x, then β = (x, x) ∪ β after the step is taken.
It puts the object in dst with label H or ?, depending on dst value’s initial
label. Also, pushes a node containing the object in the scope chain with a ?,
if Γ (!σ.Σ) = L ∨ ? or with label H, if Γ (!σ.Σ) = H or (!σ.Σ) = nil. Thus,
!σ.Σ ∼β !σ 0 .Σ by Definition 6. By Definition 7, !σ ∼β !σ 0 . Other call-frames
are unchanged, so σ ∼βρ,ρ0 σ 0 by Definition 9.
construct: Similar to call.
17. create-this: Similar to create-arguments.
18. new-object: Over the initial β, by Definition 5, θ ∼β θ0 . If the new object is
created at x, then β = (x, x) ∪ β after the step is taken.
σ ∼βρ,ρ0 σ 0 (Similar to prim).
19. get-by-id: Similar to mov when the property is a data property. If the property is an accessor property then getter is invoked and if the invocation of
getter pushes an entry on top of pc-stack, !ρ.C remains the lowest H-labelled
node in ρ0 . If it joins the label with !ρ, the L labelled nodes remain unchanged
and the !ρ.C = !ρ0 .C. All the call-frames until !ρ.C remain unchanged. So, by
Definition 9, σ ∼βρ,ρ0 σ 0 . θ = θ0 , so θ ∼β θ0 .
20. put-by-id: Sets the property of the object base object to the value with label
H if the structure label of the object `s = H. Thus, the object remains lowequivalent by Definition 4. Thus, θ ∼β θ0 by Definition 5.
Also, σ = σ 0 , so, σ ∼βρ,ρ0 σ 0 .
21. del-by-id : Deletes the property if structure label of object, `s = H. Thus,
the object remains low-equivalent by Definition 4. By Definition 5, θ ∼β θ0 .
σ ∼βρ,ρ0 σ 0 (Similar to mov).
22. getter-setter: Sets accessor property of the object base object with Γ (getter)
and Γ (setter) and label H if the structure label of the object `s = H.
Thus, the object remains low-equivalent by Definition 4. Thus, θ ∼β θ0 by
Definition 5. Also, σ = σ 0 , so, σ ∼βρ,ρ0 σ 0 .
23. get-pnames: Similar to mov and jfalse.
24. next-pname: Similar to mov.
25. resolve: If the property exists, it is similar to mov. If it does not, it is similar
to throw.
26. resolve-skip: Similar to resolve.
27. resolve-global: Similar to resolve.
28. resolve-base: Similar to resolve.
29. resolve-with-base: Similar to resolve.
30. get-scoped-var: Similar to mov.
31. put-scoped-var: Writes the value in the indexth register in skipth node. If
Γ (!σ(index)) = H, then, Γ (!σ 0 (index)) = H. Else if Γ (!σ(index)) = L, then,
Γ (!σ 0 (index)) = ?. Other call-frames are unchanged. Thus, σ ∼βρ,ρ0 σ 0 by
Definition 7 and 9.
θ = θ0 , so θ ∼β θ0 .
32. push-scope: Pushes node on scope-chain with label H if Γ (!σ.Σ) = H or
(!σ.Σ) = nil. Else, assigns a ? as the label. Thus, !σ.Σ ∼β !σ 0 .Σ. Registers remain unchanged. By Definition 7, !σ ∼β !σ 0 . Other call-frames are
unchanged, so by Definition 9, σ ∼βρ,ρ0 σ 0 . θ = θ0 , so θ ∼β θ0 .
33. pop-scope: Pops the node from the scope-chain if Γ (!σ.Σ) = H ∨ ?. Registers remain unchanged. By Definition 7, !σ ∼β !σ 0 . Other call-frames are
unchanged, so by Definition 9, σ ∼βρ,ρ0 σ 0 . θ = θ0 , so θ ∼β θ0 .
34. jmp-scope: Similar to pop-scope.
35. throw: Pops the call-frames until the handler is reached, i.e., until (!ρ.C).
The property of IPD ensures that !σ 0 = (!ρ.C). Either !ρ0 .ipd = !ρ.ipd or
ι0 = !ρ.ipd. Thus, !ρ.C is !σ 0 . This call-frame and the ones below remain
unchanged. Thus, σ ∼βρ,ρ0 σ 0 by Definition 9.
θ = θ0 , so θ ∼β θ0 .
36. catch: Similar to mov.
n
Corollary 2 If hι0 , θ0 , σ0 , ρ0 i
hιn , θn , σn , ρn i and ∀(0 ≤ i ≤ n).Γ (!ρi ) =
β
H, then ρ0 ∼ ρn , and σ0 ∼ρ0 ,ρn σn
Proof. To prove: ρ0 ∼ ρn .
Proof by induction on n.
Basis: ρ0 ∼ ρ0
IH : ρ0 ∼ ρn−1
From Definition 8, L labelled nodes of ρ0 and ρn−1 are equal. From Lemma 1,
ρn−1 ∼ ρn so, L labelled nodes of ρn−1 and ρn are equal. Thus, L labelled nodes
of ρ0 and ρn are equal and by Definition 8, ρ0 ∼ ρn .
To prove: σ0 ∼βρ0 ,ρn σn .
Basis: σ0 ∼βρ0 ,ρ0 σ0 .
IH: σ0 ∼βρ0 ,ρn−1 σn−1 .
From Lemma 1, σn−1 ∼βρn−1 ,ρn σn . As ∀(0 ≤ i ≤ n).Γ (ρi ) = H, the lowest Hlabelled node is the same (pc-stack grows monotonically) in ρ0 , ρn−1 , ρn . Let the
call-frames pointed to by lowest H-labelled node be C0 , Cn−1 , Cn with call-stack
size until the call-frames k (from Definition 9 size of the prefix is same and by
transitivity of equality it is the same for all the three cases).
∀µ0 ∈ σ0 , µn−1 ∈ σn−1 , µn ∈ σn until C0 , Cn−1 , Cn respectively with sizes k, the
following conditions hold:
1. ∀(1 ≤ i ≤ k).((µ0 [i].#Registers) = (µn−1 [i].#Registers)) and ∀(1 ≤ i ≤
k).((µn−1 [i].#Registers) = (µn [i].#Registers)).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].#Registers) = (µn [i].#Registers)).
2. As the number of registers is the same, given by r,
∀(1 ≤ i ≤ k).∀r((µ0 [i].Registers[r]) ∼β (µn−1 [i].Registers[r])) and
∀(1 ≤ i ≤ k).∀r.((µn−1 [i].Registers[r]) ∼β (µn [i].Registers[r])).
`n−1
Let v0`0 , vn−1
and vn`n represents the values in the registers for σ0 , σn−1 and
σn respectively. Then from Definition 3
(a) `0 = `n−1 = H: In this case `n = H and v0`0 ∼β vn`n (from Lemma 1 and
Definition 3).
(b) `0 = `n−1 = L and v0 = vn−1 : In this case either:
i. `n = ?
ii. `n = L and vn−1 = vn : In this case, the value remains unchanged.
Thus, from Definition 3 v0`0 ∼β vn`n .
(c) `0 = ? ∨ `n−1 = ?: Now the following cases arise:
i. `0 = ?: v0`0 ∼β vn`n .
ii. `n−1 = ?: By Lemma 1 ln = ?. Thus, v0`0 ∼β vn`n .
3. ∀(1 ≤ i ≤ k).((µ0 [i].CFG) = (µn−1 [i].CFG)) and ∀(1 ≤ i ≤ k).((µn−1 [i].CFG) =
(µn [i].CFG)).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].CFG) = (µn [i].CFG)).
4. ∀(1 ≤ i ≤ k).((µ0 [i].Σ) ∼β (µn−1 [i].Σ)) and ∀(1 ≤ i ≤ k).((µn−1 [i].Σ) ∼β
(µn [i].Σ)).
From Definition 6:
(a) If nil0 and niln−1 be the two scope chains, then due to confinement
(Lemma 1) µn [i].Σ = nil or µn [i].Σ = (S, `n ), where `n = H. In either
case ∀(1 ≤ i ≤ k).((µ0 [i].Σ) ∼β (µn [i].Σ)) from Definition 6.
(b) If ((S0 , `0 ) : Σ0 ), ((Sn−1 , `n−1 ) : Σn−1 ) and ((Sn , `n ) : Σn ) be the three
scope-chains, then for((S0 , `0 ) : Σ0 ) and ((Sn−1 , `n−1 ) : Σn−1 ) one of
the following holds:
i. `0 = ? ∨ `n−1 = ?: Due to confinement (Lemma 1) and Definition 6
`n = ?.
ii. `0 = `n−1 = H: Due to confinement (Lemma 1) and Definition 6
ln = H .
iii. `0 = `n−1 = L ∧ S0 ∼β Sn−1 ∧ Σ0 ∼β Σn−1 : Due to confinement
(Lemma 1) either one should hold:
A. `n = ?: By Definition 6.
B. `n = L ∧ Sn−1 ∼β Sn ∧ Σn−1 ∼β Σn : No additions to the scope
chain.
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].Σ) ∼β (µn [i].Σ)) from Definition 6.
5. ∀(1 ≤ i ≤ k).((µ0 [i].ιr ) = (µn−1 [i].ιr )) and ∀(1 ≤ i ≤ k).((µn−1 [i].ιr ) =
(µn [i].ιr )).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].ιr ) = (µn [i].ιr )).
6. ∀(1 ≤ i ≤ k).(((µ0 [i].`c ) = (µn−1 [i].`c ) = H) ∨ (((µ0 [i].`c ) = (µn−1 [i].`c ) =
L) ∧ ((µ0 [i].fc ) = (µn−1 [i].fc )))) and
∀(1 ≤ i ≤ k).(((µn−1 [i].`c ) = (µn [i].`c ) = H) ∨ (((µn−1 [i].`c ) = (µn [i].`c ) =
L) ∧ ((µn−1 [i].fc ) = (µn [i].fc )))).
Then either:
– ∀(1 ≤ i ≤ k).((µ0 [i].`c ) = (µn [i].`c ) = H) or
– ∀(1 ≤ i ≤ k).(((µ0 [i].`c ) = (µn [i].`c ) = L) ∧ ((µ0 [i].fc ) = (µn [i].fc ))).
7. ∀(1 ≤ i ≤ k).((µ0 [i].argcount) = (µn−1 [i].argcount)) and ∀(1 ≤ i ≤ k).
((µn−1 [i].argcount) = (µn [i].argcount)).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].argcount) = (µn [i].argcount)).
8. ∀(1 ≤ i ≤ k).((µ0 [i].getter) = (µn−1 [i].getter)) and ∀(1 ≤ i ≤ k).((µn−1 [i].getter) =
(µn [i].getter)).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].getter) = (µn [i].getter)).
9. ∀(1 ≤ i ≤ k).((µ0 [i].dReg) =β (µn−1 [i].dReg)) and ∀(1 ≤ i ≤ k).((µn−1 [i].dReg) =β
(µn [i].dReg)).
Thus, ∀(1 ≤ i ≤ k).((µ0 [i].dReg) =β (µn [i].dReg)).
From Definition 7 and Definition 9, σ0 ∼βρ0 ,ρn σn .
Corollary 3 If hι0 , θ0 , σ0 , ρ0 i
H, then θ0 ∼β θn
Proof. By induction on n.
Basis: θ0 ∼β θ0 by Definition 5.
IH: θ0 ∼β θn−1 .
?
hιn , θn , σn , ρn i and ∀(0 ≤ i ≤ n).Γ (!ρi ) =
From IH and Definition 5, ∀(a, b) ∈ β.(θ0 (a) ∼β θn−1 (b). From Lemma 1,
θn−1 ∼β θn . Thus, ∀(b, c) ∈ β.(θn−1 (b) ∼β θn (c)
As (a, b) ∈ β and (b, c) ∈ β, we have (a, c) ∈ β because β is an identity bijection.
Thus, if ∀(a, c) ∈ β.(θ0 (a) ∼β θn (c), then θ0 ∼β θn . If θ0 (a) and θn−1 (b) contain
an ordinary object, then for their respective structure labels `s and `0s , either:
– `s = `0s = H: If `0s = H, then `00s = H by Definition 4, where `00s is the
structure label of the object in θn (c). Thus, θ0 (a) ∼β θn (c).
0
– `s = `0s = L: [p0 , . . . , pn ] = [p00 , . . . , p0m ] (n = m), ∀i. vi ∼β vi0 , and a`p ∼β a`p
for respective properties in θ0 (a) and θn−1 (b).
If `0s = L, then `00s = L and [p00 , . . . , p0m ] = [p000 , . . . , p00k ] (m = k), ∀i. vi0 ∼β vi00 ,
0
00
and a`p ∼β a`p for respective properties in θn−1 (b) and θn (c).
`s = `00s = L, [p0 , . . . , pn ] = [p000 , . . . , p00k ] (n = k). If ∀i. vi ∼β vi0 and ∀i. vi0 ∼β
vi00 , then either `i = `0i = `00i = H or `i = `0i = `00i = L and ri = ri0 = ri00 = n.
0
0
00
00
Also, as a`p ∼β a`p and a`p ∼β a`p , we have a`p ∼β a`p . Thus, by Definition 4
θ0 (a) ∼β θn (c).
If θ0 (a) and θn−1 (b) contain a function object, then for their respective structure
labels `s and `0s , either:
– `s = `0s = H: If `0s = H, then `00s = H by Definition 4, where `00s is the
structure label of the function object in θn (c). Thus, θ0 (a) ∼β θn (c).
– `s = `0s = L: `00s = L is the structure label of the function object in θn (c).
Thus, N ∼β N 00 from the above result for objects. The CFGs f =β f 0 =β f 00
and the scope chains Σ ∼β Σ 00 by Corollary 2. Thus, θ0 (a) ∼β θn (c).
Thus, θ0 ∼β θn .
Lemma 2 (Supporting Lemma 1) Suppose
hι, θ1 , σ1 , ρ1 i
hι01 , θ10 , σ10 , ρ01 i,
hι, θ2 , σ2 , ρ2 i
hι02 , θ20 , σ20 , ρ02 i,
ρ1 ∼ ρ2 , Γ (!ρ1 ) = Γ (!ρ2 ) = L, Γ (!ρ01 ) = Γ (!ρ02 ) and (σ1 ∼βρ1 ,ρ2 σ2 ) ∧ (θ1 ∼β θ2 )
0
0
then ρ01 ∼ ρ02 , and ∃β 0 : ((β 0 ⊇ β) ∧ (σ10 ∼βρ0 ,ρ0 σ20 ) ∧ (θ10 ∼β θ20 )).
1
2
Proof. Every instruction executes isIPD at the end of the operation. If ι0i is the
IPD corresponding to the !ρi .ipd, then it pops the first node on the pc-stack. As
ρ1 ∼ ρ2 and Γ (!ρ1 ) = Γ (!ρ2 ), ι0i would either pop in both the runs or in none.
Thus, ρ01 simρ02 . For instructions that push (branch), we explain in respective
instructions.
Proof by case analysis on the instruction type:
1. prim: No new object is created, so β 0 = β.
As σ1 ∼βρ1 ,ρ2 σ2 , so !σ1 ∼β !σ2 and !σ1 (src i ) ∼β !σ2 (src i ) for i = 1, 2. Case
analysis on the definition of ∼β for src i .
– If (Γ (!σ1 (src 1 )) = ? ∨ Γ (!σ1 (src 2 )) = ? ∨ Γ (!σ2 (src 1 )) = ? ∨ Γ (!σ2 (src 2 )) =
?), then Γ (!σ1 (dst)) = ? ∨ Γ (!σ1 (dst)) = ?. Hence, !σ1 (dst) ∼β !σ2 (dst)
by Definition 3.
– If Γ (!σ1 (src1 )) = Γ (!σ1 (src2 )) = H and Γ (!σ2 (src1 )) = Γ (!σ2 (src2 )) =
H Γ (!σ1 (dst)) = Γ (!σ2 (dst)) = H. So, !σ1 (dst) ∼β !σ2 (dst) by Definition 3.
– If !σ1 (src1 ) = !σ2 (src1 ) ∧ Γ (!σ1 (src2 )) = H ∧ Γ (!σ2 (src2 )) = H, then
Γ (!σ1 (dst)) = Γ (!σ2 (dst)) = H. So, !σ1 (dst) ∼β !σ2 (dst) by Definition 3.
Symmetrical reasoning for !σ1 (src2 ) = !σ2 (src2 ) ∧ Γ (!σ1 (src1 )) = H ∧
Γ (!σ2 (src1 )) = H.
– !σ1 (src1 ) = !σ2 (src1 ) ∧ !σ1 (src2 ) = !σ2 (src2 ):
!σ1 (dst) = !σ2 (dst). So, !σ1 (dst) ∼β !σ2 (dst) by Definition 3.
Only dst changes in the top call-frame of both the call-stacks. Thus, by
Definition 7, !σ10 ∼β !σ20 . Other call-frames in σ10 and σ20 are unchanged. By
0
Definition 9, σ10 ∼βρ0 ,ρ0 σ20 .
1
2.
3.
4.
5.
6.
2
0
θ1 = θ10 and θ2 = θ20 , so, θ10 ∼β θ20 .
mov: Similar reasoning as prim with single source.
jfalse: No new object is created, so β 0 = β.
– !σ1 (cond) = !σ2 (cond) ∧ Γ (!σ1 (cond)) = Γ (!σ2 (cond)) = L: L is the label
to be pushed on ρ.
– Γ (!σ1 (cond)) = Γ (!σ2 (cond)) = H: H is the label to be pushed on ρ.
The IPD of ι would be the same as we have same CFG in both the cases. If
the IPD is SEN, then we join the label of !ρi with the label obtained above,
which is the same in both the runs. Thus, Γ (!ρ01 ) = Γ (!ρ02 ). Because ρ1 ∼ ρ2 ,
ρ01 ∼ ρ02 .
If the IPD is not SEN, then it is some other node in the same call-frame.
Thus the ipd field is also the same. The H field is false in both the cases.
Thus, the pushed node is the same in both the cases and hence, ρ01 ∼ ρ02 .
As, Γ (!ρ01 ) = Γ (!ρ002 ), either ι01 = ι02 = IPD(ι) or ι01 and ι02 may or may not
be equal.
0
0
σ10 = σ1 ∼βρ0 ,ρ0 σ2 = σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
loop-if-less: Similar reasoning as jfalse.
type-of: Similar to mov.
instance-of: No new object is created, so β 0 = β.
The label of the value in the dst is the label of the context joined with
the label of all the prototype chain pointers traversed. As !σ1 (value) ∼β
!σ2 (value), where `s and `0s are the structure labels of objects pointed to by
!σ1 (value) and !σ2 (value) respectively, then by Definition 4:
– If `s = `0s = H, then Γ (!σ10 (dst)) = H and Γ (!σ20 (dst)) = H. So,
!σ10 (dst) ∼β !σ20 (dst) from Definition 3.
– If `s = `0s = L, then the objects have similar properties and prototype
chains. If it is not an instance and none of traversed prototype chain and
objects are H, then Γ (!σ10 (dst)) = Γ (!σ20 (dst)) = L and false. Else if it
is present it has true. So, !σ10 (dst) ∼β !σ20 (dst) from Definition 3. If any
one of traversed prototype chain and objects are H, then Γ (!σ10 (dst)) =
Γ (!σ20 (dst)) = H. So, !σ10 (dst) ∼β !σ20 (dst) from Definition 3.
Only dst changes in the top call-frame of both the call-stacks. Thus, by
Definition 7, !σ10 ∼β !σ20 . Other call-frames are unchanged and by Definition 9,
0
σ10 ∼βρ0 ,ρ0 σ20 .
1
7.
8.
9.
10.
2
0
θ1 = θ10 and θ2 = θ20 , so, θ10 ∼β θ20 .
enter: No new object is created, so β 0 = β.
0
0
σ10 = σ1 ∼βρ0 ,ρ0 σ2 = σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
ret: No new object is created, so β 0 = β.
Since σ1 ∼β σ2 so only two cases arise for the getter flag.
– !σ1 .getter =!σ2 .getter = f alse: σ10 is same as σ1 with !σ1 popped. Similarly, σ20 is same as σ2 with !σ2 popped. As other call-frames are un0
changed by Definition 9, σ10 ∼βρ0 ,ρ0 σ20 .
1 2
– !σ1 .getter =!σ2 .getter = true: only resgister which changes is the σ10 (!σ1 .dReg
and σ20 (!σ2 .dReg). Now, if Γ (!σ10 (!σ1 .dReg)) = Γ (!σ20 (!σ2 .dReg)) = H
then !σ10 (!σ1 .dReg) ∼β !σ20 (!σ1 .dReg) from defintion 3. And if Γ (!σ10 (!σ1 .dReg)) =
Γ (!σ20 (!σ2 .dReg)) = L then !σ10 (!σ1 .dReg)) =!σ1 (res)σ20 (!σ2 .dReg) =
!σ2 (res) and !σ1 (res) ∼β !σ2 (res).
0
0
θ1 = θ1 ∼β θ2 = θ20 .
0
end: No σi and θi0 .
call: No new object is created, so β 0 = β.
Pushes the same node on both ρs (similar to jfalse). The only difference is
the H field. As the CFGs are same, if it has an associated exception handler,
we set the H field to true in both the runs. Else, it is false. Thus, !ρ01 = !ρ02
is the node pushed on ρ and hence, ρ01 ∼ ρ02 .
As !σ1 (f unc) ∼β !σ2 (f unc), if:
– (Γ (!ρ01 ) = H) : As call-frames until !σ1 and !σ2 remain unchanged, which
correspond to the C field in the lowest H-labelled node and σ1 ∼βρ1 ,ρ2 σ2 ,
0
by Definition 9, σ10 ∼βρ0 ,ρ0 σ20 .
1 2
– (Γ (!ρ01 ) = L) : Registers created in the new call-frame contain undefined
with label L and, as θ1 ∼β θ2 so the function objects N ∼β N 0 implying
!σ10 .CFG =!σ20 .CFG and !σ10 .Σ ∼β !σ20 .Σ, also return addresses are the
same and the callee is the same. So !σ10 ∼β !σ20 . Other call-frames are
0
unchanged so, σ10 ∼βρ0 ,ρ0 σ20 .
1
2
0
θ10 = θ1 ∼β θ2 = θ20 .
11. call-put-result: Similar to move.
12. call-eval: ρ01 ∼ ρ02 : Similar to op-call.
In strict mode, it pushes a node on the scope-chain with label L. The pushed
nodes are low-equivalent. Thus, !σ10 .Σ ∼β !σ20 .Σ by Definition 6. In non-strict
0
mode, it does not push anything and is similar to call. Thus, σ10 ∼βρ0 ,ρ0 σ20
1
0
2
and θ10 = θ1 ∼β θ2 = θ20 .
13. create-arguments: Let the argument object be created at x and y in θ1 and
θ2 , then β 0 = β ∪ (x, y). Γ (σ10 (dst)) = Γ (σ10 (dst)) = L and as !σ1 ∼β !σ2 , the
0
0
objects are low-equivalent. Thus, !σ10 ∼β !σ2 by Definition 7 and σ10 ∼βρ0 ,ρ0 σ20
0
1
2
by Definition 9. Also, θ10 ∼β θ20 by Definition 5 as the objects are lowequivalent.
14. new-func: Let the function object be created at x and y in θ1 and θ2 , then
β 0 = β ∪ (x, y). Function objects are low-equivalent as !σ1 .Σ ∼β !σ2 .Σ and
0
!σ1 (func) ∼β !σ2 (func). σ10 (dst) ∼β σ20 (dst) by Definition 3. Thus, !σ10 ∼β !σ2
0
0
by Definition 7 and σ10 ∼βρ0 ,ρ0 σ20 by Definition 9. Also, θ10 ∼β θ20 by Definition
1 2
5 as the objects are low-equivalent.
15. create-activation: Similar to create-arguments.
16. construct: Similar to call.
17. create-this: Similar to create-this.
18. new-object: Similar to create-arguments.
19. get-by-id: No new object is created, so β 0 = β.
As !σ1 (base) ∼β !σ2 (base), either the objects have the same properties or
are labelled H because of Definition 4. In case of data property, either
Γ (σ10 (dst)) = Γ (σ20 (dst)) = H or Γ (σ10 (dst)) = Γ (σ20 (dst)) = L and value of
prop is the same. So, by Definition 3 !σ10 (dst) ∼β !σ20 (dst).
In case of an accessor property, only dst changes in the top call-frame of
both the call-stacks, and σ10 (dst)) ∼β σ20 (dst)) since σ1 ∼β σ2 and θ1 ∼β θ2
Thus, by Definition 7, !σ10 ∼β !σ20 . Other call-frames are unchanged and by
0
Definition 9, σ10 ∼βρ0 ,ρ0 σ20 . For ρ01 ∼ ρ02 , reasoning is similar to call. θ1 = θ10
1
2
0
and θ2 = θ20 , so, θ10 ∼β θ20 .
20. put-by-id: No new object is created, so β 0 = β.
0
σ10 = σ1 ∼βρ0 ,ρ0 σ2 = σ20 .
1
2
Because σ1 ∼βρ1 ,ρ2 σ2 , if value is labelled H, then the properties created or
modified will have label H, and structure labels of the respective objects will
become H. Else if value is labelled L, then the properties created or modified
will have same value and label L. Thus, the objects remain low-equivalent
0
by Definition 4 and hence, by Definition 5, θ10 ∼β θ20 .
21. del-by-id: No new object is created, so β 0 = β.
If the deleted property is H or if the structure label of the object is H , then
(Γ (σ10 (dst)) = Γ (σ20 (dst)) = H). Else if is labelled L, then (Γ (σ10 (dst)) =
Γ (σ20 (dst)) = L) and value is true or false depending on whether the property
0
is deleted or not. σ10 ∼βρ0 ,ρ0 σ20 by Definition 3, 7 and 9. If structure labels of
1 2
the objects are L, they have same properties by Definition 4. If not, they have
structure label as H. Thus, objects remain low-equivalent by Definition 4 and
0
θ10 ∼β θ20 by Definition 5.
22. put-getter-setter: Reasoning similar to put-by-id.
23. get-pnames: No new object is created, so β 0 = β.
As σ1 ∼βρ1 ,ρ2 σ2 , !σ1 (base) ∼β !σ2 (base) and so are the objects (obj1 and
obj2), obj1 ∼β obj2, as θ1 ∼β θ2 . Thus, the structure label of the object is
either H in both the runs or L and have the same properties with values
(Definition 4. The IPD in both the cases is the same and so is the C field.
The mH field is set to false. Thus, ρ01 ∼ ρ02 .
0
For σ10 ∼βρ0 ,ρ0 σ20 , it is similar to mov, but done for dst, i and size.
1
0
2
θ10 = θ1 ∼β θ2 = θ20 .
24. next-pname: Similar to mov, but done for dst and base.
25. resolve: No new object is created, so β 0 = β.
If property is found in a L object and the scope-chain node labels are also
L, then the property value is the same as !σ1 ∼β !σ2 . If it is in H object or
any scope-chain node labels are H or have a ?, then label of the property
0
is H or ?. Thus, !σ10 (dst) ∼β !σ20 (dst) by Definition 3. Thus, σ10 ∼βρ0 ,ρ0 σ20 . If
1 2
property is not found in both runs, it is similar to throw. If property is not
found in second run, then in the first run the property is in H context. So,
the exception thrown is also H. Until, the call-frame of !ρ02 .ipd, call-frames
0
0
are unchanged, so σ10 ∼βρ0 ,ρ0 σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
26. resolve-skip: Similar to resolve.
27. resolve-global: Similar to resolve.
28. resolve-base: Similar to resolve.
29. resolve-with-base: Similar to resolve.
30. get-scoped-var: No new object is created, so β 0 = β.
Reads indexth register in object in skipth node in the scope-chain and writes
into dst. As (!σ1 .Σ ∼β !σ2 .Σ), the value, if labelled L is the same, else is
labelled H or ?. By Definition 3, !σ10 (dst) ∼β !σ20 (dst) and by Definition 7,
0
0
!σ10 ∼β !σ20 . Thus, σ10 ∼βρ0 ,ρ0 σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
31. put-scoped-var: No new object is created, so β 0 = β.
Writes into the scope chain node the same value, if “value” is labelled L. If
it is labelled ? in any of the runs, scope chains remain equivalent. If value
is H, it checks the label of register and puts the value with label H or ?.
Thus, (!σ10 .Σ ∼β !σ20 .Σ) by Definition 6 and (!σ10 ∼β !σ20 ) by Definition 7. By
0
0
Definition 9, σ10 ∼βρ0 ,ρ0 σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
32. push-scope: No new object is created, so β 0 = β.
Pushes in the scope-chain a node containing the object in “scope” with node
label L. As (!σ1 (scope) ∼β !σ2 (scope)) and (!σ1 .Σ ∼β !σ2 .Σ), (!σ10 .Σ ∼β
!σ20 .Σ) by Definition 6. Registers and other call-frames remain the same, so,
0
0
σ10 ∼βρ0 ,ρ0 σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
33. pop-scope: No new object is created, so β 0 = β.
Pops a node from the scope chain if Γ (!σ1 .Σ) = Γ (!σ1 .Σ) 6= (? ∨ H). As
(!σ1 .Σ ∼β !σ2 .Σ), so (!σ10 .Σ ∼β !σ20 .Σ). Other registers remain the same, so,
0
0
σ10 ∼βρ0 ,ρ0 σ20 . θ10 = θ1 ∼β θ2 = θ20 .
1 2
34. jmp-scope: Similar to pop-scope.
35. throw: No new object is created, so β 0 = β.
The property of IPD ensures that !σ10 = (!ρ1 .C) and !σ20 = (!ρ2 .C). The !σ10
0
and !σ20 and the ones below them remain unchanged. Thus, σ10 ∼ρβ0 ,ρ0 σ20 by
1 2
Definition 9.
0
θ10 = θ1 ∼β θ2 = θ20 .
36. catch: Similar to mov.
Lemma 3 (Supporting Lemma 2) Suppose
1. hι0 , θ00 , σ00 , ρ00 i
hι01 , θ10 , σ10 , ρ01 i
n−1
hι0n , θn0 , σn0 , ρ0n i,
2.
3.
4.
5.
m−1
00
00
hι0 , θ000 , σ000 , ρ000 i
hι001 , θ100 , σ100 , ρ001 i
hι00m , θm
, σm
, ρ00m i,
β
(ρ00 ∼ ρ000 ), (σ00 ∼ρ0 ,ρ00 σ000 ), (θ00 ∼β θ000 ),
0 0
(Γ (!ρ00 ) = Γ (!ρ000 ) = L), (Γ (!ρ0n ) = Γ (!ρ00m ) = L),
∀(0 < i < n).(Γ (!ρ0i ) = H) ∧ ∀(0 < j < m).(Γ (!ρ00j ) = H),
00
00
then (ι0n = ι00m ), (ρ0n ∼ ρ00m ), (σn0 ∼βρ0n ,ρ00m σm
), and (θn0 ∼β θm
).
Proof. Starting with the same instruction and high context in both the runs,
we might get two different instructions, ι01 and ι001 . This is only possible if ι was
some branching instruction in the first place and this divergence happened in a
high context. Now,
1. To prove ι0n = ι00m :
From the property of the IPDs we know that if ι0 pushes a H node on top
of pc-stack which was originally L, IPD(ι0 ) pops that node. Since we start
from the same instrucion ι0 , ι0n = ι00m = IPD(ι), where Γ (!ρ) = L.
2. To prove ρ0n ∼ ρ00m :
– n > 1 and m > 1: Γ (!ρ01 ) = Γ (!ρ001 ), because ι0 pushes equal nodes and
ι01 , ι001 are not the IPDs. As ρ00 ∼ ρ000 and Γ (!ρ01 ) = Γ (!ρ001 ), from Lemma 2
we get ρ01 ∼ ρ001 and !ρ01 .ipd =!ρ001 .ipd = IPD(ι0 ), if ι01 6= IPD(ι0 ) and
ι001 6= IPD(ι0 ). As ι0n = ι00m = IPD(ι0 ), it pops the !ρ01 and !ρ001 , which
correspond to ρ0n and ρ00m in the nth and mth step. (IPD is the point
where we pop the final H node on the pc-stack.) Because ρ01 ∼ ρ001 and
from Corollary 2, ρ0n ∼ ρ00m .
– n = 1 and m > 1: If Γ (!ρ01 ) 6= Γ (!ρ001 ) and ι01 = IPD(ι0 ), then Γ (!ρ01 ) = L.
It pops the node pushed by ι0 , i.e., Γ (!ρ0n ) = L. In the other run as
Γ (!ρ001 ) = H and Γ (!ρ00m ) = L, by the property of IPD ι00m = IPD(ι0 ),
which would pop from the pc-stack !ρ001 , the first frame labelled H on the
pc-stack. Thus, ρ0n ∼ ρ00m .
– n > 1 and m = 1: Symmetric case of the above.
00
3. To prove σn0 ∼βρ0n ,ρ00m σm
:
(a) n > 1 and m > 1: From Lemma 2 we get σ10 ∼βρ0 ,ρ00 σ100 . From Corollary 2
1
we get σ10 ∼βρ0 ,ρ0
1
n−1
1
0
0
σn−1
. And from Lemma 1 we have σn−1
∼βρ0
n−1
,ρ0n
σn0 .
00
As, ι0n = ι00m = IPD(ι0 ), we compare all call-frames of σn0 and σm
. As the
IPD of an instruction can lie only in the same call-frame, comparison for
all call-frames in σ00 and σ000 suffice.
σ00 ∼βρ0 ,ρ00 σ000 ⇒ ∀i.((µi ∈ σ00 ∧νi ∈ σ000 ), (µi ∼β νi )∧∀(r ∈ µi , νi ).(µi (r) =
0
0
v1 ∧ νi (r) = v2 , v1 ∼β v2 )).
00
Let v1 and v2 be represented by vn and vm in σn0 and σm
, respectively.
00
The call-frames in σn0 and σm
are represented by µn and νm , respectively.
– We do case analysis on the different cases of Definitions 3 for v1 and
v2 , to show vn ∼β vm . As (∀(1 ≤ i < n).(Γ (!ρ0i ) = H) ∧ ∀(1 ≤ j <
m).(Γ (!ρ00j ) = H)):
• If v1 = v2 ∧ Γ (v1 ) = Γ (v2 ) = L, then either vn = vm or ((? =
Γ (vn )) ∨ (? = Γ (vm ))). By Definition 3(1), vn ∼β vm .
• If Γ (v1 ) = Γ (v2 ) = H, then Γ (vn ) = Γ (vm ) = H. By Definition 3(2), vn ∼β vm .
• If ? = Γ (v1 ), then ? = Γ (vn ) and if ? = Γ (v2 ), then ? = Γ (vm ).
By Definition 3(1), vn ∼β vm .
– Lets S1 and S2 be the scopechains in σ00 and σ000 . And Sn and Sm
00
represent the scopechains in σn0 and σm
, i.e., Sn and Sm are the
respective scope-chains in the nth and mth step of the two runs
and `n and `m are their node labels. For scope chain pointers the
following cases arise:
i. S1 = S2 = nil: In this case Sn and Sm either remain nil or its
head will have a H label, because of the rules of the instructions
that modify the scope-chain.
ii. S1 = (s1 , `1 ) : Σ1 and S2 = (s2 , `2 ) : Σ2 :
A. `1 = ? ∨ `2 = ?: In this case `n and `m will be ? too.
B. `1 = `2 = H: In this case `n and `m will be H too.
C. `1 = `2 = L ∧ Σ1 ∼β Σ2 : In this case `n = ? and `m = ? or
scopechains remain unchanged.
(b) n = 1 and m > 1:
In case of jfalse and loop-if-less, σ00 = σ10 and σ000 = σ100 . And in case
of get-pnames, if n = 1 and m 6= 1, Υ (!σ00 (base)) = undefined and
Υ (!σ000 (base)) 6= undefined. Because σ00 ∼β σ000 , !σ00 (base) ∼β !σ000 (base).
Hence, Γ (!σ00 (base)) = Γ (!σ000 (base)) = H. Thus, Γ (!σ10 (dst)) = Γ (!σ10 (i)) =
Γ (!σ10 (size)) = H and similarly, Γ (!σ100 (dst)) = Γ (!σ100 (i)) = Γ (!σ100 (size)) =
H. Other registers remain unchanged and so do the other call-frames.
Thus, σ10 ∼βρ0 ,ρ00 σ100 . From the case (a) above, we know that if σ10 ∼βρ0 ,ρ00
1
1
1
1
00
.
σ100 , then σn0 ∼βρ0n ,ρ00m σm
(c) n > 1 and m = 1: Symmetric case of the above.
00
:
4. To prove θn0 ∼β θm
(a) n > 1 and m > 1: From Lemma 2 we get θ10 ∼β θ100 . From Corollary 3 we
0
0
. And from Lemma 1 we have θn−1
∼β θn0 . Assume O1 is
get θ10 ∼β θn−1
00
0
an object at x in θ1 and O2 is an object at y in θ1 , such that (x, y) ∈ β
and On and Om are the respective objects in the nth and mth step of
the two runs. We do case analysis on the different cases of Definitions 4
for O1 and O2 , to show On ∼β Om .
– If Γ (O1 ) = Γ (O2 ) = H, then Γ (On ) = Γ (Om ) = H. By Definition 4,
On ∼β Om .
– If O1 = O2 ∧ Γ (O1 ) = Γ (O2 ) = L, then On = Om ∧ Γ (On ) =
Γ (Om ) = L. By Definition 4, On ∼β Om .
Similarly, for function objects, the structure labels would remain H if
they were originally H or will remain L with the same CFGs and scopechains.
(b) n = 1 and m > 1: In case of jfalse, loop-if-less, get-pnames, θ00 = θ10
and θ000 = θ100 . Thus, θ10 ∼β θ100 . From the case (a) above, we know that if
00
θ10 ∼β θ100 , then θn0 ∼β θm
.
(c) n > 1 and m = 1: Symmetric case of the above.
Definition 11 (Trace) A trace is defined as a sequence of configurations or
states resulting from a program evaluation, i.e., for a program evaluation P =
s1
s2
...
sn where si = hιi , θi , σi , ρi i, the corresponding trace is given
as T (P) := s1 :: s2 :: . . . :: sn .
Definition 12 (Epoch-trace) An epoch-trace (E) over a trace T = s1 :: s2 ::
. . . :: sn where si = hιi , θi , σi , ρi i is defined inductively as:
E(nil) := nil
(
si :: E(T )
E(si :: T ) :=
E(T )
if Γ (!ρi ) = L,
else if Γ (!ρi ) = H.
Theorem 1 (Termination-Insensitive Non-interference).
Suppose P and P 0 are two program evaluations.
Then for their respective epoch-traces given by:
E(T (P)) = s1 :: s2 :: . . . :: sn ,
E(T (P 0 )) = s01 :: s02 :: . . . :: s0m ,
if s1 ∼β s01 and n ≤ m,
then
∃βn ⊇ β : sn ∼βn s0n
Proof. Proof proceeds by induction on n.
Basis: s1 ∼β s01 , by assumption.
IH: sk ∼βk s0k where βk ⊇ β.
To prove: ∃βk+1 ⊇ β : sk+1 ∼βk+1 s0k+1 .
Let sk i sk+1 and sk i0 s0k+1 , then:
– i = i0 = 1: From Lemma 2, sk+1 ∼βk+1 s0k+1 where βk+1 ⊇ β.
– i > 1 or i0 > 1: From Lemma 3, sk+1 ∼βk+1 s0k+1 where βk+1 = β.
Corollary 4 Suppose:
1. hι1 , θ1 , σ1 , ρ1 i ∼β hι2 , θ2 , σ2 , ρ2 i
2. hι1 , θ1 , σ1 , ρ1 i ∗ hend, θ10 , [], []i
3. hι2 , θ2 , σ2 , ρ2 i ∗ hend, θ20 , [], []i
0
Then, ∃β 0 ⊇ β such that θ10 ∼β θ20 .
Proof. σ1 , σ2 and ρ1 , ρ2 are empty at the end of ∗ steps. From the semantics,
we know that in L context both runs would push and pop the same number of
nodes. Thus, both take same number of steps in L context. Let k be the number
of states in L context. Then in Theorem 1, n = m = k. Thus, sk ∼βk s0k , where
0
sk = hend, θ10 , [], []i and s0k = hend, θ20 , [], []i. By Definition 10, θ10 ∼β θ20 , where
0
β = βk .
| 6cs.PL
|
1
Energy Storage Sharing in Smart Grid: A Modified
Auction Based Approach
arXiv:1512.07700v1 [cs.SY] 24 Dec 2015
Wayes Tushar, Member, IEEE, Bo Chai, Chau Yuen, Senior Member, IEEE, Shisheng Huang, Member, IEEE,
David B. Smith, Member, IEEE, H. Vincent Poor Fellow, IEEE, and Zaiyue Yang, Member, IEEE
Abstract—This paper studies the solution of joint energy
storage (ES) ownership sharing between multiple shared facility
controllers (SFCs) and those dwelling in a residential community.
The main objective is to enable the residential units (RUs) to
decide on the fraction of their ES capacity that they want to
share with the SFCs of the community in order to assist them
storing electricity, e.g., for fulfilling the demand of various shared
facilities. To this end, a modified auction-based mechanism is
designed that captures the interaction between the SFCs and
the RUs so as to determine the auction price and the allocation
of ES shared by the RUs that governs the proposed joint ES
ownership. The fraction of the capacity of the storage that each
RU decides to put into the market to share with the SFCs and
the auction price are determined by a noncooperative Stackelberg
game formulated between the RUs and the auctioneer. It is shown
that the proposed auction possesses the incentive compatibility
and the individual rationality properties, which are leveraged via
the unique Stackelberg equilibrium (SE) solution of the game.
Numerical experiments are provided to confirm the effectiveness
of the proposed scheme.
Index Terms—Smart grid, shared energy storage, auction
theory, Stackelberg equilibrium, strategy-proof, incentive compatibility.
I. I NTRODUCTION
E
NERGY storage (ES) devices are expected to play a
significant role in the future smart grid due to their
capabilities of giving more flexibility and balance to the grid
by providing a back-up to the renewable energy [1]–[9]. ES
can improve the electricity management in a distribution network, reduce the electricity cost through opportunistic demand
response, and improve the efficient use of energy [10]. The
distinct features of ES make it a perfect candidate to assist in
W. Tushar and C. Yuen are with Singapore University of Technology and
Design (SUTD), 8 Somapah Road, Singapore 487372 (Email: {wayes tushar,
yuenchau}@sutd.edu.sg).
B. Chai is with the State Grid Smart Grid Research Institute, Beijing,
102211, China (Email: [email protected]).
S. Huang is with the Ministry of Home Affairs, Singapore (Email:
[email protected]).
D. B. Smith is with the National ICT Australia (NICTA), ACT 2601,
Australia and adjunct with the Australian National University (Email:
[email protected]).
H. V. Poor is with the School of Engineering and Applied Science at
Princeton University, Princeton, NJ, USA (Email: [email protected]).
Z. Yang is with the State Key Laboratory of Industrial Control Technology
at Zhejiang University, Hangzhou, China (Email: [email protected]).
This work is supported in part by the Singapore University of Technology
and Design (SUTD) through the Energy Innovation Research Program (EIRP)
Singapore NRF2012EWT-EIRP002-045 and IDC grant IDG31500106, and in
part by the U.S. National Science Foundation under Grant ECCS-1549881. D.
B. Smith’s work is supported by NICTA, which is funded by the Australian
Government through the Department of Communications and the Australian
Research Council.
residential demand response by altering the electricity demand
due to the changes in the balance between supply and demand.
Particularly, in a residential community setting, where each
household is equipped with an ES, the use of ES devices can
significantly leverage the efficient flows of energy within the
community in terms of reducing cost, decarbonization of the
electricity grid, and enabling effective demand response (DR).
However, energy storage requires space. In particular for
large consumers like shared facility controllers (SFCs) of large
apartment buildings [11], the energy requirements are very
high, which consequently necessitates the actual installment
of very large energy storage capacity. The investment cost of
such storage can be substantial whereas due to the random
usage of the facilities (depending on the usage pattern of
different residents) some of the storage may remain unused.
Furthermore, the use of ESs for RUs is very limited for two
reasons [10]: firstly, the installation cost of ES devices is very
high and the costs are entirely borne by the users. Secondly, the
ESs are mainly used to save electricity costs for the RUs rather
than offer any support to the local energy authorities, which
further makes their use economically unattractive. Hence, there
is a need for solutions that will capture both the problems
related to space and cost constraints of storage for SFCs and
the benefit to RUs for supporting third parties.
To this end, numerous recent studies have focused on energy
management systems with ES devices as we will see in the
next section. However, most of these studies overlook the
potential benefits that local energy authorities such as SFCs
can attain by jointly sharing the ES devices belonging to the
RUs. Particularly due to recent cost reduction of small-scale
ES devices, sharing of ES devices installed in the RUs by the
SFCs has the potential to benefit both the SFCs and the RUs
of the community as we will see later. In this context, we
propose a scheme that enables joint ES ownership in smart
grid. During the sharing, each RU leases the SFCs a fraction
of its ES device to use, and charges and discharges from the
rest of its ES capacity for its own purposes. On the contrary,
each SFC exclusively uses its portion of ES devices leased
from the RUs. This work is motivated by [10], in which the
authors discussed the idea of joint ownership of ES devices
between domestic customers and local network operators, and
demonstrated the potential system-wide benefits that can be
obtained through such sharing. However, no policy has been
developed in [10] to determine how the fraction of battery
capacity, which is shared by the network operators and the
domestic users, is decided.
Note that, as an owner of an ES device, each RU can decide
IEEE Trans. Smart Grid, Dec. 2015
II. S TATE - OF -T HE A RT
whether or not to take part in the joint ownership scheme with
the SFCs and what fraction of the ES can be shared with the
SFCs. Hence, there is a need for solutions that can capture this
decision making process of the RUs by interacting with the
SFCs of the network. In this context, we propose a joint ES
ownership scheme in which by participating in storage sharing
with the SFCs, both the RUs and SFCs benefit economically.
Due to the interactive nature of the problem, we are motivated
to use auction theory to study this problem [12].
Exploiting the two-way communications aspects, auction
mechanisms can exchange information between users and electricity providers, meet users’ demands at a lower cost, and thus
contribute to the economic and environmental benefits of smart
grid1 [13]. In particular, 1) we modify the Vickrey auction
technique [14] by integrating a Stackelberg game between the
auctioneer and the RUs and show that the modified scheme
leads to a desirable joint ES ownership solution for the RUs
and the SFCs. To do this, we modify the auction price derived
from the Vickrey auction, to benefit the owner of the ES,
through the adaptation of the adopted game as well as keep
the cost savings to the SFCs at the maximum; 2) We study
the attributes of the technique, and show that the proposed
auction scheme possesses both the incentive compatibility and
the individual rationality properties leveraged by the unique
equilibrium solution of the game; 3) We propose an algorithm
for the Stackelberg game that can be executed distributedly
by the RUs and the auctioneer, and the algorithm is shown to
be guaranteed to reach the desired solution. We also discuss
how the proposed scheme can be extended to the time varying
case; and 4) Finally, we provide numerical examples to show
the effectiveness of the proposed scheme.
The importance and necessity of the proposed study with
respect to actual operation of smart grid lies in assisting
the SFCs of large apartment buildings in smart communities
to reduce space requirements and investment costs of large
energy storage units. Furthermore, by participating in storage
sharing with the SFCs, the RUs can benefit economically,
which can consequently influence them to efficiently schedule
their appliances and thus reduce the excess use of electricity.
We stress that multi-agent energy management schemes are
not new in the smart grid paradigm and have been discussed
in [11], [15] and [16]. However, the scheme discussed in the
paper differs from these existing approaches in terms of the
considered system model, chosen methodology and analysis,
and the use of the set of rules to reach the desired solution.
The remainder of the paper is organized as follows. We
provide a comprehensive literature review of the related work
in Section II followed by the considered system model in
Section III. Our proposed modified auction-based mechanism
is demonstrated in Section IV where we also discuss how the
scheme can be adopted in a time varying environment. The
numerical case studies are discussed in Section V, and finally
we draw some concluding remarks in Section VI.
In the recent years, there has been an extensive research
effort to understand the potential of ES devices for residential energy management [17]. This is mainly due to their
capabilities in reducing the intermittency of renewable energy
generation [18] as well as lowering the cost of electricity [19].
The related studies can be divided into two general categories.
The first category of studies consisting of [20], [21], which
assume that the ESs are installed within each RU premises
and are used solely by the owners in order to perform different
energy management tasks such as optimal placement, sizing
and control of charging and discharging of storage devices.
The second type of studies deal with ES devices that are
not installed within the RUs but located in a different location
such as in electric vehicles (EVs). Here, the ESs of EVs are
used to provide ancillary services for RUs [22]–[24] and local
energy providers [25]–[27]. Furthermore, another important
impact of ES devices on residential distribution grids is studied
in [28] and [29]. In particular, these studies focus on how
the use of ES devices can bring benefits for the stakeholders
in external energy markets. In [28], the authors propose a
multi-objective optimization method for siting and sizing of
ESs of a distribution grid to capture the trade-offs between
the storage stakeholders and the distribution system operators.
Furthermore, in [29], optimal storage profiles for different
stakeholders such as distribution grid operators and energy
traders are derived based on case studies with real data. Studies
of other aspects of smart grid can be found in [30]–[36].
As can be seen from the above discussion, the use of
ES devices in smart grid is not only limited to address the
intermittency of renewable generation [18] and assisting users
to take part in energy management to reduce their cost of
electricity [19], [21] but also extends to assisting the grid
(or, other similar energy entities such as an SFC) [37] and
generating revenues for stakeholders [28], [29]. However, one
similarity between most of the above mentioned literature is
that only one entity owns the ES and uses it according to its
requirements. Nonetheless, this might not always be the case if
there are large number of RUs2 in a community. In this regard,
considering the potential benefits of ES sharing, as discussed
in [10], this paper investigates the case in which the SFCs
in a smart community are allowed to share some fraction of
the ESs owned by the RUs through a third party such as an
auctioneer or a community representative.
The proposed modified auction scheme differs from the
existing techniques for multi-agent energy management such
as those in [11], [15], [16] in a number of ways. Particularly, in
contrast to these studies, the proposed auction scheme captures
the interaction between the SFCs and the RUs, whereby the
decision on the auction price is determined via a Stackelberg
game. By exploiting auction rules including the determination
rule, payment rule, and allocation rule the interaction between
the SFCs and RUs is greatly simplified. For instance, the
determination rule can easily identify the number of RUs that
are participating in the auction process, which further leverage
1 Please note that such a technique can be applied in the real distribution
network such as in electric vehicle charging stations by using the two-way
information and power flow infrastructure of smart grids [3].
2 Each RU may participate as a single entity or as a group where RUs
connected via an aggregator [38].
2
IEEE Trans. Smart Grid, Dec. 2015
sicap
either due to the fact that some SFCs do not have their own
ESs [11] or that the ESs of the SFCs are not large enough
to store all the excess energy at that time. It is important to
note that the ES requirement of the SFCs can stem from any
type of intermittent generation profile that the SFCs or RUs
can adopt. For example, one can consider that the proposed
scheme is based on a hybrid generation profile comprising both
solar and wind generation. However, the proposed technique
is equally suitable for other types of intermittent generation
as well. We assume that there are N = |N | RUs, where N is
the set of all RUs in the system, that are willing to share some
parts of their ES with the SFCs of the network. The battery
cap
capacity of each RU i ∈ N is si , and each RU i wants to
put xi fraction of its ES in the market to share with the SFCs,
where
(1)
xi ≤ bi = scap
i − di .
sicap : Total capacity of ES device of RU i.
di : The amount that RU i does not sell.
bi : The maximum ES space the RU i might sell to the SFCs.
bi = sicap − di
di
Sharing price pt
x1 ≤ b1
x2 ≤ b2
Each i decides
ri
ri < pt
Yes
No
x N ≤ bN
Leaves the
et
sharing market
q1
q2
Yes
Each m decides
am
qM
Take
ke part in ES
sharing
am > pt
No
Sharing price pt
Here, bi is the maximum amount of battery space that the RU
can share with the SFCs if the cost-benefit tradeoff for the
sharing is attractive for it. di is the amount of ES that the RU
does not want to share, rather uses for its own needs, e.g., to
run the essential loads in the future if there is any electricity
disruption within the RU or if the price of electricity is very
high.
Fig. 1: The fraction of the ES capacity that an RU i is willing to share with
the SFCs of the community.
the determination of the auction price via the Stackelberg game
in the payment rule. Furthermore, on the one hand the work
here complements the existing works focusing on the potential
of ES for energy management in smart grid. On the other
hand, the proposed work has the potential to open new research
opportunities in terms of control of energy dispatch from ES,
the size of ES, and exploring other interactive techniques such
as cooperative games and bi-level optimization for ES sharing.
To this end, to offer an ES space xi , on the one hand, each
RU i decides on an reservation price ri per unit of energy.
Hereinafter, we will use ES space and energy interchangeably
to refer to the ES space that each RU might share with the
SFCs. However, if the price pt , which each RU received for
sharing its ES, is lower than ri , the RU i removes its ES
space xi from the market as the expected benefit from the
joint sharing of ES is not economically attractive for it. On
the other hand, each SFC m ∈ M, that needs to share ES
space with the RUs to store their energy, decides a reservation
bid am , which represents the maximum unit price the SFC m
is willing to pay for sharing per unit of ES with the RUs in
the smart community, to enter into the sharing market. And,
if am > pt , the SFC removes its commitment of joint ES
ownership with RUs from the market due to the same reason
as mentioned for the RU. A graphical representation of the
concept of ES sharing and their decision making process of
sharing the ES space of each RU i with the SFCs are shown
in Fig. 1. Please note that to keep the formulation simple,
we do not include any specific storage model in the scheme.
However, by suitably modeling some related parameters such
as the storage capacity scap
i , and parameters like di and bi , the
proposed scheme can be adopted for specific ES devices.
III. S YSTEM M ODEL
Let us consider a smart community that consists of a large
number of RUs. Each RU can be an individual home, a
single unit of a large apartment complex, or a large number
of units connected via an aggregator that acts as a single
entity [38]–[40]. Each RU is equipped with an ES device that
the RU can use to store electricity from the main grid or its
renewable energy sources, if there are any, or can perform
DR management according to the real-time price offered by
the grid. The ES device can be a storage device installed
within each RU premises or can be the ES used for the RU’s
electric vehicles. The entire community is considered to be
divided into a number of blocks, where each block consists
of a number of RUs and an SFC. Each SFC m ∈ M, where
M is the set of all SFCs and M = |M|, is responsible for
controlling the electrical equipment and machines such as lifts,
parking lot lights and gates, water pumps, and lights in the
corridor area of a particular block of the community, which
are shared and used by the residents of that block on regular
basis. Each SFC is assumed to have its own renewable energy
generation and is also connected to the main electricity grid
with appropriate communication protocols.
Considering the fact that the nature of energy generation
and consumption is highly sporadic [41], let us assume that
the SFCs in the community need some extra ESs to store
their electricity after meeting the demand of their respected
shared facilities at a particular time of the day. This can be
The interaction that arises from the choice of ES sharing
price between the SFCs and RUs as well as the need of the
SFCs to share the ES space to store their energy and the profits
that the RUs can reap from allowing their ESs to be shared
give rise to a market of ES sharing between the RUs and the
SFCs in the smart grid. In this market, the involved N RUs
and M SFCs will interact with each other to decide as to
how many of them will take part in sharing the ESs between
themselves, and also to agree on the ES sharing parameters
such as the trading price pt and the amount of ES space to
3
IEEE Trans. Smart Grid, Dec. 2015
IV. AUCTION BASED ES OWNERSHIP
Vickrey auction is a type of sealed-bid auction scheme,
where the bidders submit their written bids to the auctioneer without knowing the bids of others participating in the
auction [14]. The highest bidder wins the auction but pays
the second highest bid price. Nevertheless, in this paper, we
modify the classical Vickrey auction [14] to model the joint
ES ownership scheme for a smart community consisting of
multiple customers (i.e., the SFCs) and multiple owners of ES
devices (i.e., the RUs). The modification is motivated by the
following factors: 1) unlike the classical Vickrey auction, the
modified scheme would enable the multiple owners and customers to decide simultaneously and independently whether to
take part in the joint ES sharing through the determination rule
of the proposed auction process, as we will see shortly; 2) the
modification of the auction provides each participating RU i
with flexibility of choosing the amount of ES space that they
may want to share with the SFCs in cases when the auction
price5 pt is lower than their expected reservation price ri ;
and 3) finally, the proposed auction scheme provides solutions
that satisfy both the incentive compatibility and individual
rationality properties, as we will see later, which are desirable
in any mechanism that adopts auction theory [41].
To this end, the proposed auction process, as shown in
Fig. 2, consists of three elements:
Fig. 2: Energy management in a smart community through auction process
consisting of multiple RUs with ES devices, an auctioneer and a number of
SFCs.
be shared. In the considered model, the RUs not only decide
on the reservation prices ri , but also on the amount of ES
space xi that they are willing to share with the SFCs. The
amount of xi is determined by the trade-off between between
the economic benefits that the RU i expects to obtain from
giving the SFCs the joint ownership of its ES device and
the associated reluctance αi of the RU for such sharing. The
reluctance to share ESs may arise from the RUs due to many
factors. For instance, sharing would enable frequent charging
and discharging of ESs that reduce the lifespan3 of an ES
device [42]. Hence, an RU i may set its αi higher so as
to increase its reluctance to participate in the ES sharing.
However, if the RU is more interested in earning revenue rather
than increasing ES life time, it can reduce its αi and thus get
more net benefits from sharing its storage. Therefore, for a
given set of bids am , ∀m and storage requirement qm , ∀m
by the SFCs, the maximum amount of ES xi that each RU
i will decide to put for sharing is strongly affected by the
trading price pt and the reluctance parameter4 αi of each
RU i ∈ N during the sharing process. In this context, we
develop an auction based joint ES ownership scheme in the
next section. We understand that the proposed scheme involves
different types of users such as auctioneers, SFCs, and RUs.
Therefore, the communication protocol used by them could
be asynchronous. However, in our study we assume that the
communication between different entities of the system are
synchronous. This is mainly due to the fact that we assume
our algorithm is executed once in a considered time slot, and
the duration of this time slot can be one hour [43]. Therefore,
synchronization is not a significant issue for the considered
case and the communication complexity is affordable. For
example, the auctioneer can wait for five minutes until it
receives all the data from SFCs and the RUs and then the
algorithm, which is proposed in Section IV-C, can be executed.
1) Owner: The RUs in set N , that own the ES devices,
and expect to earn some economic benefits, e.g., through
maximizing a utility function, by letting the SFCs to share
some fraction of their ES spaces.
2) Customer: The SFCs in set M, that are in need of ESs in
order to store some excess electricity at a particular time
of the day. The SFCs offer the RUs a price with a view
to jointly own some fraction of their ES devices.
3) Auctioneer: A third party (e.g., estate or building manager), that controls the auction process between the
owners and the customers according to some predefined
rules.
The proposed auction policies consist of A) determination
rule, B) payment rule and C) storage allocation rule. Here,
determination rule allows the auctioneer to determine the
maximum limit for the auction price pmax
and the number
t
of SFCs and RUs that will actively take part in the ES sharing
scheme once the auction process is initiated. The payment rule
enables the auctioneer to decide on the price that the customer
needs to pay to the owners for sharing their ES devices, which
allows the RUs to decide how much storage space they will
be putting into the market to share with the SFCs. Finally,
the auctioneer allocates the ES spaces for sharing for each
SFC following the allocation rule of the proposed auction.
It is important to note that although both the customers and
owners do not have any access to others private information
such as the amount of ES to be shared by an RU or the required
energy space by any SFC, the rules of auction are known to
all the participants of the joint ownership process.
3 Please note that the life time degradation due to charging and discharging
may not true for all electromechanical systems such as redox-flow system.
4 Reluctance parameter refers to the opposite of preference parameter [38].
5 Hereinafter, p will be used to refer to auction price instead of sharing or
t
trading price.
4
IEEE Trans. Smart Grid, Dec. 2015
of the SFCs and RUs in the network:
90
Cosumers participating
in auction
Owner
Customer
am < ri ; ∀m ∈ M/{1, 2, . . . , K}, ∀i ∈ N /{1, 2, . . . , J}. (4)
80
Hence, the joint ownership of ES would be a detrimental choice for the RUs and the SFCs within the set
N /{1, 2, . . . , J} and M/{1, 2, . . . , K} respectively, which
consequently remove them from the proposed auction process.
Now, one desirable property of any auction mechanism is that
no participating agents in the auction mechanism will cheat
once the payment and allocation rules are being established.
To this end, we propose that, once J and K are determined,
K − 1 SFCs and J − 1 RUs will be engaged in the joint ES
sharing process, which is a necessary condition for matching
total demand and supply while maintaining a truthful auction
scheme [44]. Nevertheless, if truthful auction is not a necessity,
SFC K and RU J can also be allowed to participate in the
joint ES ownership auction.
Price
70
60
Maximum auction price pmax
t
Vickrey price pmin
t
50
40
Owners participating
in auction
30
0
50
100
150
Storage amount
Fig. 3: Determination of the Vickrey price, the maximum auction price, and
the number of participating RUs and SFCs in the auction process.
The proposed scheme initially determines the set of SFCs
⊂ M and RUs ⊂ N that will effectively take part in the
auction mechanism once the upper bound of the auction price
pmax
is determined. Eventually, the payment and the allocation
t
rules are executed in the course of the auction plan.
B. Payment Rule
We note that the intersection of the demand and supply
curves demonstrates the highest reservation price pmax
for the
t
participating J − 1 RUs. According to the Vickrey auction
mechanism [14], the auction price for sharing the ES devices
would be the second highest reservation price, i.e., the Vickrey
price, which will be indicated as pmin
hereinafter. However, we
t
note that this second highest price might not be considerably
beneficial for all the participating RUs in the auction scheme.
In contrast, if pt is set to pt = pmax
t , the price could be
detrimental for some of the SFCs. Therefore, to make the
auction scheme attractive and beneficial to all the participating
RUs and, at the same time, to be cost effective for all the SFCs,
we strike a balance between the pmax
and pmin
t
t . To do so, we
propose a scheme for deciding on both the auction price pt
and the amount of ESs xi that RUs will put into the market for
sharing according to pt . In particular, we propose a Stackelberg
game between the auctioneer that decides on the auction price
pt to maximize the average cost savings to the SFCs as well as
satisfying their desirable needs of ESs, and the RUs, that decide on the vector of the amount of ES x = [x1 , x2 , . . . , xJ−1 ]
that they would like to put into the market for sharing such that
their benefits are maximized. Please note that the solution of
the proposed problem formulation can also be solved following
other distributed algorithms, e.g., algorithms designed via the
bi-level optimization technique [45].
Stackelberg game: Stackelberg game is a multi-level decision making process, in which the leader of the game takes
the first step to choose its strategy. The followers, on the other
hand, choose their strategy in response to the decision made by
the leader. In the proposed game, we assume the auctioneer
as the leader and the RUs as the followers. Hence, it can
be seen as a single-leader-multiple-follower Stackelberg game
(SLMFSG). We propose that the auctioneer, as a leader of
the SLMFSG Γ, will take the first step to choose a suitable
min
auction price pt from the range [pmin
t , pt ]. Meanwhile, each
RU i ∈ {1, 2, . . . , J − 1}, as a follower of the game, will
play its best strategy by choosing a suitable xi ∈ [0, bi ] in
response to the price pt offered by the auctioneer. The best
A. Determination Rule
The determination rule of the proposed scheme is executed
by the following steps (inspired from [44]):
i) The RUs of set N , i.e., the owners of the ESs, declare
their reservation price ri , ∀i in an increasing order, which
we can consider, without loss of generality, as:
r1 < r2 < . . . < rN .
(2)
The RUs submit the reservation price along with the
amount xi of ES that they are interested to share with
the SFCs to the auctioneer.
ii) The SFCs’ bidding prices, i.e., am , ∀m, are arranged in
a decreasing order, i.e.,
a1 > a2 > . . . > aM .
(3)
The SFCs submit to the auctioneer along with the quantity
qm ∀m of ES that they require.
iii) Once the auctioneer receives the ordered information from
the RUs and the SFCs, it generates the aggregated supply
(reservation price of the RUs versus the amount of ES the
RUs interested to share) and demand curves (reservation
bids am verses the quantity of ES qm needed) using (2)
and (3) respectively.
iv) The auctioneer determines the number of of participating
SFCs K and RUs J that satisfies aK ≥ rJ from
the intersection of the two curves using any standard
numerical method [44].
As soon as the SFC K ≤ M and RU J ≤ N are determined
from the intersection point, as shown in Fig. 3, an important
aspect of the auction mechanism is to determine the number
of SFCs and RUs, which will take part in the joint ownership
of ESs. We note that once the number of SFCs K and RUs
J are determined, the following relationship holds for the rest
5
IEEE Trans. Smart Grid, Dec. 2015
response strategy of each RU i will stem from a utility function
Ui , which captures the benefit that an RU i can gain from
deciding on the amount of ES xi to be shared for the offered
price. Whereas the auctioneer chooses the price pt with a
view to maximize the average cost savings Z of the SFCs
in the network. Now to capture the interaction between the
auctioneer and the RUs, we formally define the SLMFSG Γ
as
Γ = {{{1, 2, . . . , J − 1}, {Auctioneer}}, {Ui }i∈{1,2,...,J−1} ,
{Xi }i∈{1,2,...,J−1} , Z, pt },
which consists of: i) the set of RUs {1, 2, . . . , J − 1} participating in the auction scheme and the auctioneer; ii) the utility
Ui that each RU i reaps from choosing a suitable strategy
xi in response to the price pt announce by the auctioneer;
iii) the strategy set Xi of each RU i ∈ {1, 2, . . . , J − 1};
iv) the average cost savings Z that incurred to each SFC
m ∈ {1, 2, . . . , K − 1} from the strategy
by the
chosen
max
of the
auctioneer, and v) the strategy pt ∈ pmin
t , pt
auctioneer.
In the proposed approach, each RU i iteratively responses to
the strategy pt chosen by the auctioneer independent of other
RUs in set {1, 2, . . . , J − 1}/{i}. The response of i is affected
by the offered price pt , its reluctance parameter αi and the
initial reservation price ri .
However, we note that the auctioneer does not have any
control over the decision making process of the RUs. It only
sets the auction price pt with a view to maximize the cost
savings Z, with respect to the cost with the initial bidding
price, for the SFCs. To this end, the target of auctioneer is
(5)assumed to maximize the average cost savings
! J−1
PK−1
X
(a
−
p
)
m
t
m=1
Z=
xi
(8)
K −1
i=1
by choosing an appropriate price
p to offer to each RU from
PK−1 t
min max
m=1 (am −pt )
is the average savthe range [pt , pt ]. Here,
K−1
ing in auction
price
that
the
SFCs
pay
to
the
RUs for sharing
P
the ESs and i xi is the total amount of ES that all the SFCs
share from the RUs. From Z, we note that the cost savings
will be more if pt is lower for all m ∈ {1, 2, . . . , K − 1}.
However, this is conflicted by that fact that a lower pt may
lead to the choice of lower xi ∀i ∈ {1, 2, . . . , J − 1} by the
RUs, which in turn will affect the cost to the SFCs. Hence,
to reach a desirable solution set (x∗ , p∗t ), the auctioneer and
the RUs continue to interact with each other until the game
reaches a Stackelberg equilibrium (SE).
Now, the utility function Ui , which defines the benefits that
an RU i can attain from sharing xi amount of its ES with the
SFCs, is proposed to be
Ui (xi ) = (pt − ri )xi − αi x2i , xi ≤ bi ,
(6)
where, αi is the reluctant parameter of RU i, and ri is the
reservation price set by RU i. Ui mainly consists of two parts.
The first part (pi − ri )xi is the utility in terms of its revenue
that an RU i obtains from sharing its xi portion of ES device.
The second part αi x2i , on other hand, is the negative impact
in terms of liability on the RU i stemming from sharing its
ES with the SFC. This is mainly due to the fact that once an
RU decides to share its xi amount of storage space with an
SFC, the RU can only use scap
i − xi amount of storage for its
own use. The term αi x2i captures this restriction of the RU
on the usage of its own ES. In (6), the reluctance parameter
αi is introduced as a design parameter to measure the degree
of unwillingness of an RU to take part in energy sharing. In
particular, a higher value of αi refers to the case when an RU
i is more reluctant to take part in the ES sharing, and thus, as
can be seen from (6), even with the same ES sharing attains
a lower net benefit. Thus, Ui can be seen as a net benefit
to RU i for sharing its ES. The utility function is based on
the assumption of a non-decreasing marginal utility, which
is suitable for modeling the benefits of power consumers, as
explained in [46]. In addition, the proposed utility function
also possesses the following properties: i) the utility of any
RU increases as the amount of price pt paid to it for sharing
per unit of ES increases; ii) as the reluctance parameter αi
increases, the RU i becomes more reluctant to share its ES, and
consequently the utility decreases; and iii) for a particular price
pt , the more an RU shares with the SFCs, the less interested it
becomes to share more for the joint ownership. To that end, for
a particular price pt and reluctance parameter αi , the objective
of RU i is
(7)
max (pt − ri )xi − αi x2i , xi ≤ bi .
Definition 1. Let us consider the game Γ as described in (5),
where the utility of each RU i and the average utility per SFC
are described via Ui and Z respectively. Now, Γ will reach a
SE (x∗ , p∗t ), if and only if the solution of the game satisfies
the following set of conditions:
Ui (x∗i , x∗−i , p∗t ) ≥ Ui (xi , x∗−i , p∗t ), ∀i ∈ {1, 2, . . . , J − 1},
max
∀xi ∈ Xi , p∗t ∈ [pmin
t , pt ], (9)
and
PK−1
− p∗t ) X ∗
xi ≥
K −1
i
m=1 (am
PK−1
− pt ) X ∗
xi ,
K −1
i
m=1 (am
(10)
where x−i = [x1 , x2 , . . . , xi−1 , xi+1 , . . . , xJ−1 ].
Hence, according to (9) and (10), both the RUs and the SFCs
achieve their best possible outcomes at the SE. Hence, neither
the RUs nor the auctioneer will have any incentive to change
their strategies as soon as the game Γ reaches the SE. However,
achieving an equilibrium solution in pure strategies is not
always guaranteed in non-cooperative games [38]. Therefore,
we need to investigate whether the proposed Γ possesses an
SE or not.
Theorem 1. There always exists a unique SE solution for
the proposed SLMFSG Γ between the auctioneer and the
participating RUs in set {1, 2, . . . , J − 1}.
Proof: Firstly, we note that the strategy set of the auction
max
.
eer is non-empty and continuous within the range pmin
t , pt
Hence, there will always be a non-empty strategy for the
auctioneer that will enable the RUs to offer some part of their
xi
6
IEEE Trans. Smart Grid, Dec. 2015
ES, within their limits, to the SFCs. Secondly, for any price
pt , the utility function Ui in (6) is strictly concave with respect
2
of xi , ∀i ∈ {1, 2, . . . , J − 1}, i.e., δδxU2i < 0. Hence, for any
i
min max
price pt ∈ pt , pt , each RU will have a unique xi , which
will be chosen from a bounded range [0, bi ] and maximize Ui .
Therefore, it is evident that as soon as the scheme will find
a unique p∗t such that the average utility Z per SFC attains a
maximum value, the SLMFSG Γ will consequently reach its
unique SE.
To this end, first we note that the amount of ES x∗i , at which
the RU i achieves its maximum utility in response to a price
pt can be obtained from (6),
pt − ri
x∗i =
.
(11)
2αi
Algorithm 1: Algorithm for SLMFSG to reach the SE
∗
1: Initialization: p∗t = pmin
t , Z = 0.
2: for Auction price pt from pmin
to pmax
do
t
t
3:
for Each RU i ∈ {1, 2, . . . , (J − 1)} do
4:
RU i adjusts its amount of ES xi to share according to
x∗i = arg
5:
6:
7:
8:
Now, replacing the value of x∗i in (8) and doing some simple
arithmetics, the auction price p∗t , which maximizes the average
cost savings to the SFCs can be found as
P
P
P
J−1 1
K−1
J−1
a
+ i=1 ri (K−1)
m
i=1 2αi
m=1
2αi
∗
pt =
, (12)
PJ−1 K−1
i=1
max
0≤xi ≤bi
[(pt − ri )xi − αi x2i ].
(13)
end for
The auctioneer computes the average cost savings to SFCs
! J −1
PK−1
X ∗
m=1 (am − pt )
xi .
(14)
Z=
K−1
i=1
if Z ≥ Z ∗ then
The auctioneer record the desirable price and maximum
average cost savings
p∗t = pt , Z ∗ = Z.
(15)
9:
end if
10: end for
The SE (x∗ , p∗t ) is achieved.
αi
guaranteed to reach SE of the proposed SLMFSG Γ.
where am for any m ∈ {1, 2, . . . , K − 1} and αi for any
i ∈ {1, 2, . . . , J − 1} is exclusive. Therefore, p∗t is unique for
Γ, and thus Theorem 1 is proved.
Proof: In the proposed algorithm, we note that the choice
of strategies by the RUs emanate from the choice pt of the
auctioneer, which as shown in (12) will always attain a nonempty single value p∗t at the SE due to its bounded strategy set
max
[pmin
t , pt ]. On the other hand, as the Algorithm 1 is designed,
in response to the p∗t , each RU i will choose its strategy xi
from the bounded range [0, bi ] in order to maximize its utility
function Ui . To that end, due to the bounded strategy set and
continuity of Ui with respect to xi , it is confirmed that each
RU i will always reach a fixed point x∗i for the given p∗t .
Therefore, the proposed Algorithm 1 is always guaranteed to
reach the unique SE of the SLMFSG.
C. Algorithm for payment
To attain the SE, the auctioneer, which has the information
of am , m = {1, 2, 3, . . . , K − 1}, needs to communicate with
each RU. It is considered that the auctioneer does not have
any knowledge of the private information of the RUs such
as αi , ∀i. In this regard, in order to decide on a suitable
auction price pt that will be beneficial for both the RUs and
the SFCs, the auctioneer and the RUs interact with one another.
To capture this interaction, we design an iterative algorithm,
which can be implemented by the auctioneer and the RUs in a
distributed fashion to reach the unique SE of the proposed
SLMFSG. The algorithm initiates with the auctioneer who
sets the auction price pt to pmin
and the optimal average
t
cost saving per SFC Z ∗ to 0. Now, in each iteration, after
having the information on the offered auction price by the
auctioneer, each RU i plays its best response xi ≤ bi and
submits its choice to the auctioneer. The auctioneer, on other
hand, receives the information on x = [x1 , x2 , . . . , xJ−1 ]
from all the participating RUs and determines the average
cost savings per SFC Z from its knowledge on the reservation
bids [a1 , a2 , . . . , aK−1 ] and using (8). Then, the auctioneer
compares the Z with Z ∗ . If Z > Z ∗ , the auctioneer updates
the optimal auction price to the one recently offered and
sends a new choice of price to the RUs in the next iteration.
However, if Z ≤ Z ∗ , the auctioneer keeps the same price
and offers another new price to the RUs in the next iteration.
The iteration process continues until the conditions in (9) and
(10) are satisfied, and hence the SLMFSG reaches the SE. We
show the step-by-step process of the proposed algorithm in
Algorithm 1.
D. Allocation Rule
Now, once the the amount of ES x∗i that each RU i ∈
{1, 2, . . . , J − 1} decides to put into the market for sharing in
response to the auction price p∗t is determined, the auctioneer
allocates the quantity Qi to be jointly shared by each RU i
and the SFCs according to following rule [44]:
(
PK−1
PJ−1
x∗i
if i=1 x∗i ≤ m=1 qm ,
Qi (x) =
(16)
P
P
J−1 ∗
(x∗i − ηi )+ if i=1
xi > K−1
m=1 qm ,
where (f )+P = max(0,
Pf ) and ηi is the allotment of the
J−1 ∗
excess ES i=1
xi − K−1
m=1 qm that an RU i must endure.
Essentially, the rule in (16) emphasizes that if the requirements
of the SFCs exceed the available ES space from the RUs, each
RU i will allow the SFCs to share all of the ES xi that it put
into the market. However, if the available ES exceeds the total
demand by the SFCs, then
RU P
i will have to share a
Peach
J−1
K−1
fraction of the oversupply i=1 x∗i − m=1 qm . Nonethless,
this burden, if there is any, can be distributed in different
ways among the participating RUs. For instance, the burden
can be distributed either proportionally to the amount of ES
x∗i that each RU i shared with the SFCs or proportionally to
Theorem 2. The algorithm proposed in Algorithm 1 is always
7
IEEE Trans. Smart Grid, Dec. 2015
the reservation price6 ri of each RU. Alternatively, the total
burden can also be shared equally by the RUs in the auction
scheme [44].
1) Proportional allocation: In proportional allocation [47],
a fraction of the total burden ηi is allocated to each RU i
in
to the reservation
price ri (or, x∗i ) such that
PJ−1
PK−1
P proportion
∗
i ηi =
i=1 xi −
m=1 qm , which can be implemented
as follows:
P
PK−1
J−1 ∗
x
−
i=1 i
m=1 qm ri
P
ηi =
, i = [1, 2, . . . , J − 1]. (17)
i ri
it is clear that all the participants in the proposed auction
scheme are individually rational, which leads to the following
Corollary 1.
Corollary 1. The proposed auction technique possesses the
individual rationality property, in which the J − 1 rational
owners and K − 1 rational customers actively participate in
the mechanism to gain the higher utility.
Theorem 3. The proposed auction mechanism is incentive
compatible, i.e., truthful auction is the best strategy for any
RU i ∈ {1, 2, . . . , J − 1} and SFC m ∈ {1, 2, . . . , K − 1}.
By replacing ri with x∗i in (17), the burden allocation can be
determined in proportion to the shared ES by each RU.
2) Equal allocation: According to equal allocation [44],
each RU bears an equal burden
!
K−1
J−1
X
X
1
∗
ηi =
x −
qm , i = [1, 2, . . . , J − 1] (18)
J − 1 i=1 i m=1
Proof: To validate Theorem 3, first we note that the
choice of strategies by the RUs always guaranteed to converge
to a unique SE, i.e., x∗ = [x∗1 , x∗2 , . . . , x∗J−1 ] as proven in
Theorem 1 and Theorem 2, which confirms the stability of
their selections. Now, according to [44], once the owners of an
auction process, i.e., the RUs in this proposed case, decide on
a stable amount of commodity, i.e., x∗i ∀i ∈ {1, 2, . . . , J − 1},
to supply to or to share with the customers, the auction process
always converges to a strategy-proof auction if the allocation
of commodity is conducted according to the rules described
in (16) and (18). Therefore, neither any RU nor any SFC
will have any intention to falsify their allocation once they
adopt (16) and (18) [44] for sharing the storage space of the
RUs from their SE amount. Therefore, the auction process is
incentive compatible, and thus Theorem 3 is proved.
of the oversupply.
Here it is important to note that, although proportional
allocation allows the distribution of oversupply according to
some properties of the RUs, equal allocation is more suitable
to make the auction scheme strategy proof [44]. Strategy
proofness is important for designing auction mechanisms as
it encourages the participating players not to lie about their
private information such as reservation price [41], which
is essential for the acceptability and sustainability of such
mechanisms in energy markets. Therefore, we will use equal
allocation of (18) for the rest of the paper.
F. Adaptation to Time-Varying Case
To extend the proposed scheme to a time-varying case, we
assume that the ES sharing scheme works in a time-slotted
fashion where each time slot has a suitable time duration based
on the type of application, e.g., 1 hour [43]. It is considered
that in each time slot all the RUs and SFCs take part in the
proposed ES sharing scheme to decide on the parameters such
as the auction price and the amount of ESs that needs to be
shared. However, in a time-varying case, the amount of ES
that an RU shares at time slot t may be affected by the burden
that the RU needed to bear in the previous time slot t − 1. To
this end, first we note that once the number of participating
RUs and SFCs is decided for a particular time slot via the
determination rule, the rest of the procedures, i.e., the payment
and allocation rules are executed following the descriptions in
Section IV-B and IV-D respectively for the respective time
slot. Now, if the total number of RUs and SFCs is fixed, the
RUs and SFCs that participate in the modified auction scheme
in any time slot is determined by their respective reservation
and bidding prices for that time slot. Further, the proposed
auction process may evolve across different time slots based
on the change of the amount of ES that each participating RU
i may want to share and the change in the total amount of
ES required for the SFCs in different time slots. Now, before
discussing how the proposed modified auction scheme can be
extended to a time-varying environment7, first we define the
E. Properties of the Auction Process
We note that once the auction process is executed, there
is always a possibility that the owners of the ES might
cheat on the amount of storage that they wanted put into
the market during auction [13]. In this context, we need to
investigate whether the proposed scheme is beneficial enough,
i.e., individually rational, for the RUs such that they are not
motivated to cheat, i.e., incentive compatible, once the auction
is executed.
Now for the individual rationality property, first we note that
all the players, i.e., the RUs and the auctioneer on behalf of the
SFCs, take part in the SLMFSG to maximize their benefits in
terms of their respected utility from their choice of strategies.
The choice of the RUs is to determine vector of ES x∗ such
that each of the RU can be benefitted at its maximum. On the
other hand, the strategy of the auctioneer is to choose a price pt
to maximize the savings of the SFCs. Accordingly, once both
the RUs and the auctioneer reach such a point of the game
when neither the owners nor the customers can be benefitted
more from choosing another strategy, the SLMFSG reaches
the SE. To this end, it is already proven in Theorem 1 that the
proposed Γ in this auction process must possesses a unique
SE. Therefore, as a subsequent outcome of the Theorem 1,
7 Certain loads such as lifts and water pumps in large apartment buildings
are not easy to schedule as they are shared by different users of the buildings.
Hence, we focus on the time variation of the storage sharing process by the
RUs of the considered system.
6 Please
note that the reservation price ri indicates how much each RU
i wants to be paid forPsharing its ES with the SFCs, and thus affects the
determination of total
x∗i and the total burden.
8
IEEE Trans. Smart Grid, Dec. 2015
which households are equipped with a dedicated battery to sell
the stored electricity to the grid [48]. Nonetheless, xi,t is also
affected by the amount of burden ηi,t−1 that an RU needed to
bear due to an oversupply of ES spaces, if there was any, in
the previous time slot. To this end, the amount of ES space
that an RU i can offer to the SFCs at t can be defined as
(
xi,t−1
if i ∈
/ Jt−1
xi,t =
. (21)
max(bi,t − (xi,t−1 − ηi,t−1 ), 0) otherwise
following parameters:
t: index of time slot.
T : total number of time slot.
ri,t : the reservation price of RU i ∈ N at time slot t.
ri = [ri,1 , ri,2 , . . . , ri,T ]: is the reservation price vector for
RU i ∈ N .
xi,t : the fraction of ES space that the RU i wants to shares
with the SFCs at time slot t.
xi = [xi,1 , xi,2 , . . . , xi,T ]: the vector of ES space shared by
RU i with the SFC during the total considered times.
bi,t : maximum available ES of RU i for sharing at time slot
t.
am,t : the bidding price of each SFC m ∈ M at time slot t.
am = [am,1 , am,2 , . . . , am,T ]: is the reservation price vector
for SFC m ∈ N .
qm,t : the required ES space by each SFC m at time slot t.
pt,t : the auction price at time slot t.
Ui,t : the benefit that each RU i achieves at time slot t.
Zt : the average cost saving per SFC at time slot t.
ηi,t : the burden that is shared by each participating RU at time
slot t.
Kt : number of participating SFCs in the modified auction
scheme at time slot t.
Jt : number of participating RUs in the modified auction
scheme at time slot t.
To this end, the utility function Ui,t of each RU i and the
average cost savings Zt per SFC at time slot t can be defined
as
Ui,t (xi,t ) = (pt,t − ri,t )xi,t − αi x2i,t ,
The SFC m, on the other hand, decides on the amount of ES
qm,t that it needs to share from the RUs at t based on the
random requirement of the shared facilities at t, the available
shared ES space qm,t−1 from time slot t − 1, and the random
generation of renewable energy sources, where appropriate.
Hence,
qm,t = f (qm,t−1 , renewables, facility requirement).
Now, if we assume that the fraction of shared ES available
from previous time slot is negligible, i.e., qm,t−1 ≈ 0, the
requirement qm,t can be assumed to be random for each
time slot t considering the random nature of both renewable
generation and energy requirement of shared facilities. Note
that this assumption is particularly valid if the SFC uses all its
shared ESs from the previous time slot for meeting the demand
of the shared facilities and cannot use them in considered time
slot. Nonetheless, please note that this assumption does not
imply that the inter-temporal relationship between the auction
process across different time slots is non-existent. The auction
process in one time slot still depends on other time slots due
to the inter-temporal dependency of xi,t via (21).
To this end, for the modeled xi,t ∀i ∈ N and qm,t ∀m ∈ M,
the proposed modified auction scheme studied in Section IV
can be adopted in each time slot t = 1, 2, . . . , T with a
view to maximize (19) and (20) ∀t. It is important to note
that the reservation price vector ri of each RU i ∈ N and
the bidding price vector am of each SFC m ∈ M can be
modeled through any existing time-varying pricing schemes
such as time-of-use price [3]. Now, p∗t = [p∗t,1 , p∗t,2 , . . . , p∗t,T ]
and x∗ = [x∗1 , x∗2 , . . . , x∗N ] constitute the solutions of the
proposed modified auction scheme in a time-varying condition,
if the x∗ comprises the solution vector of all ES spaces shared
by the participating RUs in each time slot t = 1, 2, . . . , T
for the auction price vector p∗t . Further, all the auction rules
adopted in each time slot of the proposed time-varying case
will be similar to the rules discussed in Section IV. Hence, the
solution of the proposed modified auction scheme for a timevarying environment also possesses the incentive compatibility
and individual rationality properties for each time slot.
(19)
and
Zt =
PKt −1
(am,t − pt,t )
Kt − 1
m=1
! J−1
X
xi,t
(22)
(20)
i=1
respectively8.
Now, at time slot t, the determination rule of the proposed
scheme determines the number of participating RUs and SFCs
based on their reservation and bidding prices for that time slot.
The number of participation is also motivated by the available
ES space of each RU and the requirement of each SFC.
However, unlike the static case, in a time-varying environment
the offered ES space by an RU at time slot t is influenced by
its contribution to the auction process in the previous time
slot. For instance, if an RU i receives a burden ηi,t−1 in time
slot t − 1, its willingness to share ES space xi,t at time slot
t may reduce. xi,t is also affected by the maximum amount
of ES bi,t available to RU i at t. For simplicity, we assume
that bi,t and αi,t do not change over different time slots.
Therefore, an RU i can offer to share the same amount of
ES space xi,t to the SFCs at time slot t if it did not share
any amount in time slot t − 1. An analogous example of such
arrangement can be found in FIT scheme with ES device in
V. C ASE S TUDY
For numerical case studies, we consider a number of RUs
at different blocks in a smart community that are interested in
allowing the SFCs of the community to jointly share their ES
devices. We stress that when there are a large number of RU
and SFCs in the system, the reservation and bidding prices
will vary significantly from one another. Therefore, it will be
difficult to find an intersection point to determine the highest
8 Please note that in each time slot t, (19) and (20) are related with each
other in a similar manner as (7) and (8) are related for the static case. However,
unlike the static case, the execution of the auction process in each time slot t
is affected by the value of parameters such as xi,t and pt for that particular
time slot.
9
ES shared by each RU
IEEE Trans. Smart Grid, Dec. 2015
TABLE I: Change of average utility achieved by each SFC and each RU in
the network (according to Algorithm 1) due to the change of the reluctance
of each RU for sharing one kWh ES with the SFC.
RU5
300
RU3
RU2
RU4
Reluctance Parameter
!
0.001
0.01
0.1
1
200
RU1
100
0
0
5
10
RU6, RU7, RU8
15
20
Number of iteration
25
30
Average utility per RU
(Net benefit)
3450.6
2883.6 (-16.43%)
1117.3 (-67.6%)
142.37 (-95.3%)
Average utility for SFC
(Average cost savings)
7500
4578.3 (-38.9%)
1671.7 (-77.7%)
259.03 (-96%)
4
Average utility for SFC
8
x 10
6
RU to put into the market for sharing. As can be seen from
the figure, on the one hand, RU 1, RU 2, and RU 3 reach the
SE much quicker than RU 4 and RU 5. On the other hand, no
interest for sharing any ES is observed for RU 4, 5 and 6.
This is due to the fact that as the interaction between the
auctioneer and the RUs continues, the auction price pt is
updated in each iteration. In this regard, once the auction price
for any RU becomes larger than its reservation price, it put all
its reserve ES to the market with an intention to be shared by
the SFCs. Due to this reason, RU 1, RU 2, and RU 3 put their
ESs in the market much sooner, i.e., after the 2nd iteration,
than RU 4 and RU 5 with higher reservation prices, whose
interest for sharing ES reaches the SE once the auction price
is encouraging enough for them to share their ESs after the
5th and 20th iterations. Unfortunately, the utilities of RU 6, 7,
and 8 are not convenient enough to take part in the auction
process, and therefore their shared ES fractions are 0.
We note that the demonstration of the convergence of the
SLMFSG to a unique SE subsequently demonstrates the proofs
of Theorem 1, Theorem 2, Theorem 3 and Corollary 1, which
are strongly related to the SE as explained in the previous
section. Now, we would like to investigate how the reluctance
parameters of the RUs may affect their average utility from
Algorithm 1, and thus affecting their decisions to share ES.
To this end, we first determine the average utility that is
experienced by each RU and SFC for a reluctance parameter of
αi = 0.001 ∀i. Then considering the outcome as a benchmark,
we show the effect of different reluctance parameters on the
achieved average benefits of each SFC and RU in Table I. The
demonstration of this property is necessary in order to better
understand the working principle of the designed technique
for ES sharing.
According to Table I, as the reluctance of each RU increases,
it becomes more uncomfortable, i.e., lower utility, to put its
ES in the market to be jointly owned by the SFCs. As a consequence, it also affects the average utility achieved by each
SFC. As shown in Table I, the reduction in average utilities per
RU are 16.73%, 67.6% and 95.3% respectively compared to
the average utility achieved by an RU at αi = 0.001 for every
ten times reduction in the reluctance parameter. For similar
settings, the reduction of average utility for the SFCs are
38.9%, 77.7% and 96% at αi = 0.01, 0.1 and 1 respectively.
Therefore, the proposed scheme will enable the RUs to put
more storage in the auction market if the related reluctance for
this sharing is small. Note that although the current investment
cost of batteries is very high compared to their relative short
life times, it is expected that battery costs will go down in
the near future [10] and become very popular for addressing
*
Z
4
2
0
0
5
10
15
20
Number of iteration
25
30
Fig. 4: Convergence of Algorithm 1 to the SE. At SE, the average utility per
SFC reaches its maximum and the ES that each RU wants to put into the
market for share reaches a steady state level that maximize their benefits.
reservation price pmax
according to the determination rule. So,
t
in this paper, we limit ourself to around 6 − 10 RUs. However,
having 6-10 RUs can in fact cover a large community, e.g.,
through aggregation such as discussed in [38], [39]. Here,
each RU is assumed to be a group of [5, 25] households,
where each household is equipped with a battery of capacity
25 kilo-Watt hour (kWh) [49]. The reluctance parameter of
all RUs are assumed to be similar, which is taken from range
of [0, 0.1]. It is important to note that αi is considered as a
design parameter in the proposed scheme, which we used to
map the reluctance of each RU to share its ES with the SFCs.
Such reluctance of sharing can be affected by parameters like
ES capacity, the condition of the environment (if applicable)
and the RU’s own requirement. Now, considering the different
system parameters in our proposed scheme, we capture these
two extremes with 0 (not reluctant) and 0.1 (highly reluctant).
The required electricity storage for each SFC is assumed to be
within the range of [100, 500] kWh. Nevertheless, the required
ES for sharing could be different if the usage pattern by the
users changes. Since, the type of ESs (and their associated
cost) used by different RUs can vary significantly [50], the
choices of reservation price to share their ESs with the SFCs
can vary considerably as well. In this context, we consider
that the reservation price set by each RU and SFC is taken
from a range of [20, 70]. It is important to note that all chosen
parameter values are particular to this study only, and may vary
according the availability and number of RUs, requirements of
SFCs, trading policy, time of the day/year and the country.
Now, we first show the convergence of Algorithm 1 to the
SE of the SLMFG in Fig. 4. For this case study, we assume
that there are five SFCs in the smart grid community that are
taking part in an auction process with eight RUs. From Fig. 4,
first we note that the proposed SLMFG reaches the SE after 20
interations when the average cost savings per SFC reaches its
maximum. Hence, the convergence speed, which is just few
seconds, is reasonable. Nonetheless, an interesting property
can be observed when we examine the choice of ES by each
10
IEEE Trans. Smart Grid, Dec. 2015
1100
for the SFCs, it would put a higher burden on the RUs to
carry. As a consequence, the relative utility from auction is
lower. Nevertheless, if the requirement of the SFCs is higher,
the sharing brings significant benefits to the RUs as can be
seen from Fig. 5. On the other hand, for higher reluctance,
RUs tend to share a lower ES amount, which then enables
them to endure a lower burden in case of lower demands from
the SFCs. This consequently enhances their achieved utility.
Nonetheless, if the requirement is higher from the SFCs, their
utility reduces subsequently compared to the RUs with lower
reluctance parameters. Thus, from observing the effects of
different αi ’s on the average utility per RU in Fig. 5, we
understand that, if the total required ES is smaller, RUs with
higher reluctance benefit more and vice versa. This illustrates
the fact that even RUs with high unwillingness to share their
ESs can be beneficial for SFCs of the system if their required
ESs are small. However, for a higher requirement, SFCs would
benefit more from having RUs with lower reluctances as they
will be interested in sharing more to achieve higher average
utilities.
Now, we discuss the computational complexity of the proposed scheme, which is greatly reduced by the determination
rule of the modified auction scheme as this rule determines
the actual number of participating RUs and SFCs in the
auction. We also note that after determining the number of
participating SFCs and RUs, the auctioneer iteratively interacts
with each of the RUs and sets the auction price with a
view to increase the average savings for the SFC. Therefore,
the main computational complexity of the modified auction
scheme stems from the interactions between the auctioneer
and the participating RUs to decide on the auction price. In
this context, the computational complexity of the problem falls
within a category of that of a single leader multiple follower
Stackelberg game, whose computational complexity, which
can be approximated to increase linearly with the number of
followers [38], and is shown to be reasonable in numerous
studies such as in [11] and [38]. Hence, the computational
complexity is feasible for adopting the proposed scheme.
Having an insight into the properties of the proposed auction
scheme, we now demonstrate how the technique can benefit
the RUs of the smart network compared to existing ES
allocation schemes such as equal distribution (ED) [38] and
FIT schemes [48]. ED is essentially an allocation scheme
that allows the SFCs to meet their total storage requirements
by sharing the total requirement equally from each of the
participating RUs. We assume that if the shared ES amount
exceeds the total amount of reservation storage that an RU puts
into the market, the RU will share its full reservation amount.
In FIT, which is a popular scheme for energy trading between
consumers and the grid, we assume that each RU prefers to
sell the same storage amount of energy to the grid at an FIT
price rate, e.g., 22 cents/kWh [52] instead of sharing the same
fraction of storage with the SFC. To this end, the resulting
average utilities that each RU can achieve from sharing its ES
space with the SFCs by adopting the proposed, ED, and FIT
schemes are shown in Table II.
From Table II, first we note that as the amount of required
ES by the SFCs increases the average utility achieved per
α = 0.001 (more willing to share)
Average utility achieved by the RUs
1000
α = 0.01 (less willing to share)
900
800
700
600
500
400
300
η0.001 > 0
η
η0.01 > 0
η
Supply > Demand
200
100
150
200
0.001
0.01
η
>0
0.001
η
=0
0.01
=0
=0
Supply < Demand
250
300
350
400
450
500
Required battery space by the SFCs (kWh)
550
600
Fig. 5: Effect of change of required ES amount by the SFCs on the achieved
average utility per RU.
intermittency of renewables [51]. We have foreseen such a
near future when our proposed scheme will be applicable to
gain the benefit of storage sharing and thus motivate the RUs
to keep their αi ∀i small. According to the observation from
Table I, it can further be said that if the reluctance parameters
of RUs change over either different days or different time slots,
the performance of the system in terms of average utility per
RU and average cost savings per SFC will change accordingly
for the given system parameters.
Once all the participating RUs put their ES amount into
the auction market, they are distributed according to the
allocation rule described in (16) and (18). In this regard,
we investigate how the average utility of each RU is altered as the total storage amount required by the SFCs
changes from in the network. For this particular case, the
considered total ES requirement of the SFCs is assumed to
be 100, 150, 200, 250, 300, 350, 400, 450, 500, 550 and 600. In
general, as shown in Fig. 5, the average utility of each RU
initially increases with the increase required by the SFCs and
eventually becomes saturated to a stable value. This is due to
the fact that as the required amount of ES increases, the RU
can share more of its reserved ES that it put into the market
with the SFCs with the determined auction price from the
SLMFSG. Hence, its utility increases. However, each RU has
a particular fixed ES amount that it puts into the market to
share. Consequently, once the shared ES amount reaches its
maximum, even with the increase of requirement by the SFCs
the RU cannot share more, i.e., ηi = 0. Therefore, its utility
becomes stable without any further increment. Interestingly,
the proposed scheme, as can be seen in Fig. 5, favors the
RUs with higher reluctance more when the ES requirement
by the SFCs is relatively lower and favors the RUs with
lower reluctance during higher demands. This is due to the
way we have designed the proposed allocation scheme, which
is dictated by the burden in (18) and the allocation of ES
through (16). We note that, according to (11), if αi is lower,
the RU i will put a higher amount of ES in the market to
share. However, if the total required amount of ES is lower
11
IEEE Trans. Smart Grid, Dec. 2015
TABLE II: Comparison of the change of average utility per RU in the smart grid system as the required total amount of energy storage required by the SFCs
varies.
Required ES space by the SFCs
Average utility (net benefit) of RU for equal distribution (ED) scheme
Average utility (net benefit) of RU for FIT scheme
Average utility (net benefit) of RU for proposed scheme
Percentage improvement (%) compared to ED scheme
Percentage improvement (%) compared to FIT scheme
ES shared by each RU (kWh)
ES available to share (kWh)
RU also increases for all the cases. The reason for this
increment is explained in Fig. 5. Also, in all the studied
cases, the proposed scheme shows a considerable performance
improvement compared to the ED and FIT schemes. An
interesting trend of performance improvement can be observed
if we compare the performance of the proposed scheme with
the ED and FIT performances for each of the ES requirements.
In particular, the performance of the proposed scheme is higher
as the requirement of the ES increases from 200 to 350.
However, the improvement is relatively less significant as the
ES requirement switches from 400 to 450. This change in
performance can be explained as follows:
In the proposed scheme, as we have seen in Fig. 5, the
amount of ES shared by each participating RU is influenced
by their reluctance parameters. Hence, even the demand of
the SFCs could be larger, the RUs may choose not to share
more of their ES spaces once their reluctance is limited. In
this regard, the RUs in the current case study increase their
share of ES as the requirement by the SFCs increases, which
in turn produces higher revenue for the RUs. Furthermore,
once the RUs choice of ESs reach the saturation, the increase
in demand, i.e., from 200 to 350 in this case, does not affect
their share. As a consequence, their performance improvement
is not as noticeable as the previous four cases. Nonetheless, for
all the considered cases, the auction process performs superior
to the ED scheme with an average performance improvement
of 34.76%, which clearly shows the value of the proposed
methodology to adopt joint ES sharing in smart grid. The
performance improvement with respect to the FIT scheme,
which is 34.34% on average, is due to the difference between
the determined auction price and the price per unit of energy
for the FIT scheme.
Finally, we show how the decision making process of each
RU in the system is affected by its decision in the previous
time slot and the total storage requirement by the SFCs. The
total number of time slots that are considered to show this
performance analysis is four. In this context, we assume that
there are five RUs in the system with ES of 100, 200, 300, 200
and 200 kWh respectively to share with the SFCs. The total
ES requirements of the SFCs for considered four time slot
are 500, 250, 500, and 100. Please note that these numbers are
considered for this case study only and may have different
values for different scenarios. Now, in Fig. 6, we show the
available ES to each of the RUs at the begining of each time
slot and how much they are going to share if the modified
auction scheme is adopted in each time slot. For a simple
analysis, we assume that once an RU shares its total available
ES, it cannot share its ES for the remaining of the time slots.
200
536.52
537.83
629.82
17.4
17.1
250
581.85
583.16
789.82
35.74
35.43
300
624.52
626.83
944.26
51.19
50.63
350
669.85
673.16
960.09
43.32
42.61
400
715.19
717.50
960.09
34.24
33.81
450
757.85
759.16
960.09
26.68
26.46
300
RU1
RU2
RU3
RU4
RU5
200
100
0
1
2
3
Number of time slot
4
300
RU1
RU2
RU3
RU4
RU5
200
100
0
1
2
3
Number of time slot
4
Fig. 6: Demonstration of how the proposed modified auction scheme can be
extended to time varying system. The reservation ES amount varies by the
RUs varies between different time slots based on their sharing amount in the
previous time slot. The total required storage by the SFCs is chosen randomly
due to the reasons explained in Section IV-F.
The reservation prices are considered to change from one time
to the next based on a predefined time of use price scheme.
Now, as can be seen from Fig. 6, in time slot 1, RU1 and RU2
share all their available ESs with the SFC, whereby other RUs
do not share their ESs due to the reasons explained in Fig. 4.
Since, the total requirement is 500, therefore neither of RU1
and RU2 needs to carry any burden. In time slot 3, only RU3
shares its ESs of 300 to meet the requirement. As the SFC’s
requirement is lower than the supply, RU3 needs to carry a
burden of 50 kWh. Similarly, in time slot 3 and 4, all of RU3,
RU4 and RU5 take part in the energy auction scheme as they
have enough ES to share with the SFC. However, the ES to
share in time slot 4 stems from the burden of oversupply from
time slot 3. The scheme is not shown for more than time slot
4 as the available ES from all RUs is already shared by the
SFCs by the end of time slot 4. Thus, the proposed modified
auction scheme can successfully capture the time variation if
the scheme is modified as given in Section IV-F.
VI. C ONCLUSION
In this paper, we have modeled a modified auction based
joint energy storage ownership scheme between a number of
residential units (RUs) and shared facility controllers (SFCs)
in smart grid. We have designed a system and discussed the
determination, payment and allocation rule of the auction,
where the payment rule of this scheme is facilitated by a
Single-leader-multiple-follower Stackelberg game (SLMFSG)
12
IEEE Trans. Smart Grid, Dec. 2015
between the auctioneer and the RUs. The properties of the
auction scheme and the SLMFSG have been studied, and it has
been shown that the proposed auction possesses the individual
rationality and the incentive compatibility properties leveraged
by the unique Stackeberg equilibrium of the SLMFSG. We
have proposed an algorithm for the SLMFSG, which has been
shown to be guaranteed to reach the SE and that also facilitates
the auctioneer and the RUs to decide on the auction price as
well as the amount of ES to be put into the market for joint
ownership.
A compelling extension of the proposed scheme would be
to study of the feasibility of scheduling of loads such as
lifts and water machines in shared space. Another interesting
research direction would be to determine how a very large
number of SFCs or RUs with different reservation and bidding
prices can take part in such a modified auction scheme. One
potential way to look at this problem can be from a cooperative
game-theoretic point-of-view in which the SFCs and RUs may
cooperate to decide on the amount of reservation ES and
bidding price they would like to put into the market so as
to participate in the auction and benefit from sharing. Another
very important, yet interesting, extension of this work would
be to investigate how to quantify the reluctance of each RU to
participate in the ES sharing. Such quantification of reluctance
(or, convenience) will also enable the practical deployment of
many energy management schemes already described in the
literature.
[11] W. Tushar, B. Chai, C. Yuen, D. B. Smith, K. L. Wood, Z. Yang, and
H. V. Poor, “Three-party energy management with distributed energy
resources in smart grid,” IEEE Trans. Ind. Electron., vol. 62, no. 4, Apr.
2015.
[12] P. Klemperer, “Auction theory: A guide to the literature,” Journal of
Economic Surveys, vol. 13, no. 3, pp. 227–286, July 1999.
[13] J. Ma, J. Deng, L. Song, and Z. Han, “Incentive mechanism for demand
side management in smart grid using auction,” IEEE Trans. Smart Grid,
vol. 5, no. 3, pp. 1379–1388, May 2014.
[14] W. Vickrey, “Counterspeculation, auctions, and competitive sealed tenders,” The Journal of Finance, vol. 16, no. 1, pp. 8–37, Mar 1961.
[15] B. Chai, J. Chen, Z. Yang, and Y. Zhang, “Demand response management with multiple utility companies: A two-level game approach,” IEEE
Trans. Smart Grid, vol. 5, no. 2, pp. 722–731, March 2014.
[16] S.Maharjan, Q. Zhu, Y. Zhang, S. Gjessing, and T. Başar, “Dependable
demand response management in the smart grid: A Stackelberg game
approach,” IEEE Trans. Smart Grid, vol. 4, no. 1, pp. 120–132, March
2013.
[17] P. Siano, “Demand response and smart grids-A survey,” Elsevier Renewable and Sustainable Energy Reviews, vol. 30, pp. 461–478, Feb
2014.
[18] P. Denholm, E. Ela, B. Kirby, and M. Milligan, “The role of energy
storage with renewable electricity generation,” National Renewable
Energy Laboratory (NREL), Colorado, USA, Technical Report, Jan
2010.
[19] Y. Cao, T. Jiang, and Q. Zhang, “Reducing electricity cost of smart
appliances via energy buffering framework in smart grid,” IEEE Trans.
Parallel Distrib. Syst., vol. 23, no. 9, pp. 1572–1582, Sep 2012.
[20] M. Sechilariu, B. Wang, and F. Locment, “Building integrated photovoltaic system with energy storage and smart grid communication,” IEEE
Trans. Ind. Electron., vol. 60, no. 4, pp. 1607–1618, April 2013.
[21] G. Carpinelli, G. Celli, S. Mocci, F. Mottola, F. Pilo, and D. Proto,
“Optimal integration of distributed energy storage devices in smart
grids,” IEEE Trans. Smart Grid, vol. 4, no. 2, pp. 985–995, June 2013.
[22] B.-G. Kim, S. Ren, M. van der Schaar, and J.-W. Lee, “Bidirectional
energy trading and residential load scheduling with electric vehicles in
the smart grid,” IEEE J. Sel. Areas Commun., vol. 31, no. 7, pp. 1219–
1234, July 2013.
[23] J. V. Roy, N. Leemput, F. Geth, R. Salenbien, J. Buscher, and J. Driesen,
“Apartment building electricity system impact of operational electric
vehicle charging strategies,” IEEE Trans. Sustain. Energy, vol. 5, no. 1,
pp. 264–272, Jan 2014.
[24] R. Yu, J. Ding, W. Zhong, Y. Liu, and S. Xie, “PHEV charging and
discharging cooperation in v2g networks: A coalition game approach,”
IEEE Internet Things J., vol. 1, no. 6, pp. 578–589, Dec 2014.
[25] J. Lin, K.-C. Leung, and V. Li, “Optimal scheduling with vehicle-to-grid
regulation service,” IEEE Internet Things J., vol. 1, no. 6, pp. 556–569,
Dec 2014.
[26] J. Tan and L. Wang, “Integration of plug-in hybrid electric vehicles into
residential distribution grid based on two-layer intelligent optimization,”
IEEE Trans. Smart Grid, vol. 5, no. 4, pp. 1774–1784, July 2014.
[27] L. Igualada, C. Corchero, M. Cruz-Zambrano, and F.-J. Heredia, “Optimal energy management for a residential microgrid including a vehicleto-grid system,” IEEE Trans. Smart Grid, vol. 5, no. 4, pp. 2163–2172,
July 2014.
[28] F. Geth, J. Tant, E. Haesen, J. Driesen, and R. Belmans, “Integration of
energy storage in distribution grids,” in IEEE Power and Energy Society
General Meeting, Minneapolis, MN, July 2010, pp. 1–6.
[29] S. Nykamp, M. Bosman, A. Molderink, J. Hurink, and G. Smit,
“Value of storage in distribution grids – Competition or cooperation of
stakeholders?” IEEE Trans. Smart Grid, vol. 4, no. 3, pp. 1361–1370,
Sep 2013.
[30] W. Tushar, C. Yuen, D. B. Smith, and H. V. Poor, “Price discrimination
for energy trading in smart grid: A game theoretic approach,” IEEE
Trans. Smart Grid, Dec. 2015, (to appear).
[31] W.-T. Li, C. Yuen, N. U. Hassan, W. Tushar, C.-K. Wen, K. L. Wood,
K. Hu, and X. Liu, “Demand response management for residential smart
grid: From theory to practice,” IEEE Access-Special Section on Smart
Grids: A Hub of Interdisciplinary Research, vol. 3, Nov. 2015.
[32] A. Naeem, A. Shabbir, N. U. Hassan, C. Yuen, A. Ahmed, and
W. Tushar, “Understanding customer behavior in multi-tier demand
response management program,” IEEE Access-Special Section on Smart
Grids: A Hub of Interdisciplinary Research, vol. 3, Nov. 2015.
[33] Y. Liu, C. Yuen, R. Yu, Y. Zhang, and S. Xie, “Queuing-based energy
consumption management for heterogeneous residential demands in
smart grid,” IEEE Trans. Smart Grid, vol. pre-print, June 2015, (doi:
10.1109/TSG.2015.2432571).
R EFERENCES
[1] M. L. D. Silvestre, G. Graditi, and E. R. Sanseverino, “A generalized
framework for optimal sizing of distributed energy resources in microgrids using an indicator-based swarm approach,” IEEE Trans. Ind.
Inform., vol. 10, no. 1, pp. 152–162, Feb 2014.
[2] P. Garcı́a, C. A. Garcı́a, L. M. Fernández, F. Llorens, and F. Jurado,
“ANFIS-based control of a grid-connected hybrid system integrating
renewable energies, hydrogen and batteries,” IEEE Trans. Ind. Inform.,
vol. 10, no. 2, pp. 1107–1117, May 2014.
[3] X. Fang, S. Misra, G. Xue, and D. Yang, “Smart grid - The new and
improved power grid: A survey,” IEEE Commun. Surveys Tuts., vol. 14,
no. 4, pp. 944–980, Oct 2012.
[4] Y. Liu, C. Yuen, S. Huang, N. U. Hassan, X. Wang, and S. Xie, “Peakto-average ratio constrained demand-side management with consumer’s
preference in residential smart grid,” IEEE J. Sel. Topics Signal Process.,
vol. PP, no. 99, pp. 1–14, Jun 2014.
[5] Y. Liu, C. Yuen, N. U. Hassan, S. Huang, R. Yu, and S. Xie, “Electricity
cost minimization for a microgrid with distributed energy resources
under different information availability,” IEEE Trans. Ind. Electron.,
vol. 62, no. 4, pp. 2571–2583, Apr. 2015.
[6] N. U. Hassan, Y. I. Khalid, C. Yuen, and W. Tushar, “Customer
engagement plans for peak load reduction in residential smart grids,”
IEEE Trans. Smart Grid, vol. 6, no. 6, pp. 3029–3041, Nov. 2015.
[7] N. U. Hassan, Y. I. Khalid, C. Yuen, S. Huang, M. A. Pasha, K. L.
Wood, and S. G. Kerk, “Framework for minimum user participation
rate determination to achieve specific demand response management
objectives in residential smart grids,” Elsevier International Journal of
Electrical Power & Energy Systems, vol. 74, pp. 91–103, Jan. 2016.
[8] W. Tushar, C. Yuen, S. Huang, D. B. Smith, and H. V. Poor, “Cost
minimization of charging stations with photovoltaics: An approach with
EV classification,” IEEE Trans. Intell. Transp. Syst., vol. pre-print, Aug.
2015, (doi: 10.1109/TITS.2015.2462824).
[9] S. Huang, W. Tushar, C. Yuen, and K. Otto, “Quantifying economic
benefits in the ancillary electricity market for smart appliances in Singapore households,” Elsevier Sustainable Energy, Grids and Networks,
vol. 1, pp. 53–62, Mar. 2015.
[10] Z. Wang, C. Gu, F. Li, P. Bale, and H. Sun, “Active demand response
using shared energy storage for household energy management,” IEEE
Trans. Smart Grid, vol. 4, no. 4, pp. 1888–1897, Dec 2013.
13
IEEE Trans. Smart Grid, Dec. 2015
[34] X. Wang, C. Yuen, X. Chen, N. U. Hassan, and Y. Ouyang, “Cost-aware
demand scheduling for delay tolerant applications,” Elsevier Journal of
Networks and Computer Applications, vol. 53, pp. 173–182, July 2015.
[35] Y. Zhang, S. He, and J. Chen, “Data gathering optimization by dynamic
sensing and routing in rechargeable sensor network,” IEEE/ACM Trans.
Netw., vol. pre-print, June 2015, (doi: 10.1109/TNET.2015.2425146).
[36] H. Zhang, P. Cheng, L. Shi, and J. Chen, “Optimal DoS attack scheduling
in wireless networked control system,” IEEE Trans. Control Syst.
Technol., vol. pre-print, Aug. 2015, (doi: 10.1109/TCST.2015.2462741).
[37] Z. Wang and S. Wang, “Grid power peak shaving and valley filling
using vehicle-to-grid systems,” IEEE Trans. Power Del., vol. 28, no. 3,
pp. 1822–1829, July 2013.
[38] W. Tushar, W. Saad, H. V. Poor, and D. B. Smith, “Economics of electric
vehicle charging: A game theoretic approach,” IEEE Trans. Smart Grid,
vol. 3, no. 4, pp. 1767–1778, Dec 2012.
[39] L. Gkatzikis, I. Koutsopoulos, and T. Salonidis, “The role of aggregators
in smart grid demand response markets,” IEEE J. Sel. Areas Commun.,
vol. 31, no. 7, pp. 1247–1257, July 2013.
[40] W. Tushar, J. A. Zhang, D. Smith, H. V. Poor, and S. Thiébaux,
“Prioritizing consumers in smart grid: A game theoretic approach,” IEEE
Trans. Smart Grid, vol. 5, no. 3, pp. 1429–1438, May 2014.
[41] W. Saad, Z. Han, H. V. Poor, and T. Başar, “A noncooperative game for
double auction-based energy trading between PHEVs and distribution
grids,” in Proc. IEEE Int’l Conf. Smart Grid Commun. (SmartGridComm), Brussels, Belgium, Oct. 2011, pp. 267–272.
[42] T. H. Bradley and A. A. Frank, “Design, demonstrations and sustainability impact assessments for plug-in hybrid electric vehicles,” Renewable
and Sustainable Energy Reviews, vol. 13, no. 1, pp. 115–128, Jan 2009.
[43] O. Derin and A. Ferrante, “Scheduling energy consumption with local
renewable micro-generation and dynamic electricity prices,” in Proc. 1st
Workshop Green Smart Embedded Syst. Technol.: Infrastruct., Methods,
Tools, Stockholm, Sweden, Apr 2010, pp. 1–6.
[44] P. Huang, A. Scheller-Wolf, and K. Sycara, “Design of a multi-unit
double auction e-market,” Computational Intelligence, vol. 18, no. 4,
pp. 596–617, Feb 2002.
[45] V. Oduguwa and R. Roy, “Bi-level optimisation using genetic algorithm,”
in Proc. IEEE International Conference on Artificial Intelligence Systems, Geelong, Australia, Feb 2002, pp. 322–327.
[46] P. Samadi, A.-H. Mohsenian-Rad, R. Schober, V. Wong, and J. Jatskevich, “Optimal real-time pricing algorithm based on utility maximization
for smart grid,” in Proc. IEEE Int’l Conf. Smart Grid Commun. (SmartGridComm), Gaithersburg, MD, Oct. 2010, pp. 415–420.
[47] X. Guojun, L. Yongsheng, H. Xiaoqin, X. Xicong, W. Qianggang,
and Z. Niancheng, “Study on the proportional allocation of electric
vehicles with conventional and fast charge methods when in distribution network,” in Proc. China International Conference on Electricity
Distribution (CICED), Shanghai, China, Sept 2012, pp. 1–5.
[48] G. Krajačić, N. Duića, A. Tsikalakis, M. Zoulias, G. Caralis, E. Panteri,
and M. da Graça Carvalho, “Feed-in tariffs for promotion of energy
storage technologies,” Energy Policy, vol. 39, no. 3, pp. 1410–1425,
Mar 2011.
[49] Fraunhofer-Gesellschaft, “Breakthrough in electricity storage: New
large and powerful redox flow battery,” Science Daily,
March 2013, retrieved August 31, 2014. [Online]. Available:
www.sciencedaily.com/releases/2013/03/130318105003.htm
[50] H. Ali and S. Ali-Oettinger, “Advancing li-ion,” May
2012,
published
in
pv
magazine.
[Online].
Available:
http://www.pv-magazine.com/archive/articles/beitrag/advancing-li-ion-100006681/501/axzz3Vkp0jXnS
[51] N. Garun, “At 38,000 reservations, Tesla’s powerwall is already sold
out until mid-2016,” 2015, accessed May 7, 2015. [Online]. Available:
http://thenextweb.com/gadgets/2015/05/07/powerwall-is-sold-out/
[52] LIPA, “Proposal concerning modifications to lipa’s tariff for electric
service,” 2010, accessed on April 3, 2015. [Online]. Available:
http://www.lipower.org/pdfs/company/tariff/proposal feedin.pdf
14
| 3cs.SY
|
arXiv:1705.03452v4 [math.AG] 19 Jan 2018
DIRECT SUM DECOMPOSABILITY OF POLYNOMIALS AND
FACTORIZATION OF ASSOCIATED FORMS
MAKSYM FEDORCHUK
Abstract. For a homogeneous polynomial with a non-zero discriminant, we
interpret direct sum decomposability of the polynomial in terms of factorization properties of the Macaulay inverse system of its Milnor algebra. This
leads to an if-and-only-if criterion for direct sum decomposability of such a
polynomial, and to an algorithm for computing direct sum decompositions
over any field, either of characteristic 0 or of sufficiently large positive characteristic, for which polynomial factorization algorithms exist. We also give
simple necessary criteria for direct sum decomposability of arbitrary homogeneous polynomials over arbitrary fields and apply them to prove that many
interesting classes of homogeneous polynomials are not direct sums.
1. Introduction
A homogeneous polynomial f is called a direct sum if, after a linear change
of variables, it can be written as a sum of two or more polynomials in disjoint
sets of variables:
(1.1)
f = f1 (x1 , . . . , xa ) + f2 (xa+1 , . . . , xn ).
When f is a homogeneous polynomial over C defining an isolated hypersurface
singularity in Cn , the geometric significance of such decomposition stems from
the classical Sebastiani-Thom theorem [10] that describes the monodromy operator of the singularity {f = 0} ⊂ Cn as a tensor product of the monodromy
operators of the singularities {f1 = 0} ⊂ Ca and {f2 = 0} ⊂ Cn−a . Direct sums
are also the subject of a well-known symmetric Strassen’s additivity conjecture
postulating that the Waring rank of f in (1.1) is the sum of the Waring ranks
of f1 and f2 (see, for example, [13]).
In this paper, we give a new criterion for recognizing when a smooth form1
is a direct sum over a field either of characteristic 0 or of sufficiently large
positive characteristic. The problem of finding such a criterion for an arbitrary
(smooth or singular) form has been successfully addressed earlier by Kleppe
[9] over an arbitrary field, and Buczyńska-Buczyński-Kleppe-Teitler [3] over
an algebraically closed field. Both works interpret direct sum decomposability
1Hereinafter,
we refer to any homogeneous polynomial as a form, and we call a form f in n
variables smooth if it defines a smooth hypersurface in Pn−1 ; see §1.1 for further terminology
and notational conventions.
1
2
MAKSYM FEDORCHUK
of a form f in terms of its apolar ideal f ⊥ (see §1.5 for more details). In
particular, over an algebraically closed field, [3] gives an effective criterion for
recognizing when f is a direct sum in terms of the graded Betti numbers of f ⊥ .
However, none of these works seem to give an effective method for computing
a direct sum decomposition when it exists, and the criterion of [3] cannot be
used over non-closed fields (see Example 4.4). Although our criterion works
only for smooth forms, it does so over an arbitrary field either of characteristic
0 or of sufficiently large characteristic, and it leads to an algorithm for finding
direct sum decompositions over any such field for which polynomial factorization
algorithms exist. This algorithm is given in Section 4.
Recall that to a smooth form f of degree d + 1 in n variables, one can assign
a degree n(d − 1) form A(f ) in n (dual) variables, called the associated form of
f ([1, 2, 4]). The associated form A(f ) is defined as a Macaulay inverse system
of the Milnor algebra of f [1], which simply means that the apolar ideal of A(f )
coincides with the Jacobian ideal of f :
A(f )⊥ = (∂f /∂x1 , . . . , ∂f /∂xn ).
Such definition leads to an observation that for a smooth form f that is written
as a sum of two forms in disjoint sets of variables, the associated form A(f )
decomposes as a product of two forms in disjoint sets of (dual) variables ([7,
Lemma 2.11]). For example, up to a scalar,
d−1
A(xd+1
+ · · · + xd+1
· · · znd−1 .
1
n ) = z1
The main purpose of this paper is to prove the converse statement, and thus
establish an if-and-only-if criterion for direct sum decomposability of a smooth
form f in terms of the factorization properties of its associated form A(f ) (see
Theorem 1.6).
In Lemma 2.1, we give a simple necessary condition, valid over an arbitrary
field, for direct sum decomposability of an arbitrary form in terms of its gradient
point. It is then applied in Section 3 to prove that a wide class of homogeneous
forms contains no direct sums. In Theorem 1.8, we show that this simple necessary condition is in fact sufficient when a form is GIT stable over an algebraically
closed field.
1.1. Notation and conventions. Let k be a field. The k-linear span of a
subset W of a k-vector space will be denoted by hW i. If W is a representation
of the multiplicative group Gm , then, for every i ∈ Z, we denote by W(i) the
weight-space of the action of weight i.
Let V be a vector space over k with n := dimk V ≥ 2. We set S := Sym V ,
and D := Sym V ∨ . Homogeneous elements of S and D will be called forms.
We have a differentiation action of S on D (also known as the apolar pairing).
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
3
Namely, if x1 , . . . , xn is a basis of V , and z1 , . . . , zn is the dual basis of V ∨ , then
the pairing
Sd′ × Dd → Dd−d′
is given by
g ◦ F = g(∂/∂z1 , . . . , ∂/∂zn )F (z1 , . . . , zn ).
Given a homogeneous nonzero F ∈ Dd , the apolar ideal of F is
F ⊥ := {g ∈ S | g ◦ F = 0} ⊂ S,
and the space of essential variables of F is
E(F ) := {g ◦ F | g ∈ Sd−1 } ⊂ D1 .
If char(k) = 0 or char(k) > d, then the pairing Sd ×Dd → k is perfect, and for
every F ∈ Dd , we have F ∈ Symd E(F ) ⊂ Dd (cf. [3, §2.2]). Furthermore, the
graded k-algebra S/F ⊥ is a Gorenstein Artin local ring with socle in degree d,
and a well-known theorem of Macaulay establishes a bijection between graded
Gorenstein Artin quotients S/I of socle degree d and elements of P(Dd ) (see,
e.g., [8, Lemma 2.12] or [5, Exercise 21.7]).
Let x1 , . . . , xn be a basis of V . The gradient point of f ∈ Sd+1 is defined to
be:
∇f := h∂f /∂x1 , . . . , ∂f /∂xn i ⊂ Sd .
The Jacobian ideal of f is Jf := (∇f ) = (∂f /∂x1 , . . . , ∂f /∂xn ) ⊂ S, and the
Milnor algebra of f is Mf := S/Jf .
Remark 1.2. Even though we allow k to have positive characteristic, we do
not take D to be the divided power algebra (cf. [8, Appendix A]), as the reader
might have anticipated. The reason for this is that at several places we cannot
avoid but to impose a condition that char(k) is large enough (or zero). In this
case, the divided power algebra is isomorphic to D up to the needed degree.
For a homogeneous ideal I ⊂ S, we denote by V(I) the closed subscheme
of PV ∨ ≃ Pn−1 defined by I. We say that a form f ∈ Sd+1 is smooth if the
hypersurface V(f ) ⊂ Pn−1 is smooth over k (this is, of course, equivalent to
V(f ) being non-singular over the algebraic closure of k). The locus of smooth
forms in PSd+1 will be denoted by (PSd+1 )∆ .
1.2. Direct sums and products. Recall from [3] that f ∈ Symd+1 V is called
a direct sum (or a form of Sebastiani-Thom type) if there is a direct sum decomposition V = U ⊕ W and nonzero f1 ∈ Symd+1 U and f2 ∈ Symd+1 W such
that f = f1 + f2 . In other words, f is a direct sum if and only if for some choice
of a basis x1 , . . . , xn of V , we have that
f = f1 (x1 , . . . , xa ) + f2 (xa+1 , . . . , xn ),
4
MAKSYM FEDORCHUK
where 1 ≤ a ≤ n − 1, and f1 , f2 6= 0. Recall also that f ∈ Symd+1 V is called
degenerate if there exists U ( V such that f ∈ Symd+1 U.
By analogy with direct sums, we will call a nonzero form F ∈ D a direct
product if there is a non-trivial direct sum decomposition V ∨ = U ⊕ W and
F = F1 F2 for some F1 ∈ Sym U and F2 ∈ Sym W . In other words, a nonzero
homogeneous F ∈ Sym D is a direct product if and only if for some choice of a
basis z1 , . . . , zn of V ∨ , we have that
(1.3)
F (z1 , . . . , zn ) = F1 (z1 , . . . , za )F2 (za+1 , . . . , zn ),
where 1 ≤ a ≤ n − 1.
Furthermore, we call a direct product decomposition in (1.3) balanced if
(n − a) deg(F1 ) = a deg(F2 ).
Note that a non-trivial factorization F = F1 F2 is a direct product decomposition
if and only if E(F1 ) ∩ E(F2 ) = (0).
Remark 1.4. Note that the roles of S and D are interchangeable in §1.1, and
so for f ∈ S, we can define the apolar ideal f ⊥ ⊂ D and the space of essential
variables E(f ) ⊂ S1 . With this notation, if char(k) = 0 or char(k) > d + 1, then
f ∈ Sd+1 is a direct sum if and only if we can write f = f1 + f2 , where f1 , f2 6= 0
and E(f1 ) ∩ E(f2 ) = (0). Furthermore, if char(k) = 0 or char(k) > d + 1, we
have that E(f ) ⊂ V is dual to (f ⊥ )1 ⊂ V ∨ , and dimk ∇f = dimk E(f ).
We say that [L] ∈ Grass(n, Symd V ) is a balanced direct sum if there is
a non-trivial direct sum decomposition V = U ⊕ W and elements [L1 ] ∈
Grass(dimk U, Symd U) and [L2 ] ∈ Grass(dimk W, Symd W ) such that
L = L1 + L2 ⊂ Symd U ⊕ Symd W ⊂ Symd V.
1.3. Associated forms. We briefly recall the theory of associated forms as
developed in [2]. Let Grass(n, Sd )Res be the affine open subset in Grass(n, Sd )
parameterizing linear spaces hg1 , . . . , gn i ⊂ Sd such that g1 , . . . , gn form a regular
sequence in S. Note that, if char(k) = 0 or char(k) > d + 1, then f ∈ Sd+1 is
smooth if and only if ∇f ∈ Grass(n, Sd )Res .
For every U = hg1 , . . . , gn i ∈ Grass(n, Sd )Res , the ideal IU = (g1 , . . . , gn ) is
a complete intersection ideal, and the k-algebra S/IU is a graded Gorenstein
Artin local ring with socle in degree n(d − 1). Suppose char(k) = 0 or char(k) >
n(d −1). Then, by Macaulay’s theorem, there exists a unique up to scaling form
A(U) ∈ Dn(d−1) such that
A(U)⊥ = IU .
The form A(U) is called the associated form of g1 , . . . , gn by Alper and Isaev,
who systematically studied it in [2, Section 2]. In particular, they showed2 that
2
Although given over C, their proof applies whenever char(k) = 0 or char(k) > n(d − 1).
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
5
the assignment U → A(U) gives rise to an SL(n)-equivariant associated form
morphism
A : Grass(n, Sd )Res → PDn(d−1) .
When U = ∇f for a smooth form f ∈ Sd+1 , we set
A(f ) := A(∇f )
and, following Eastwood and Isaev [4], call A(f ) the associated form of f . The
defining property of A(f ) is that
A(f )⊥ = Jf .
This means that A(f ) is a Macaulay inverse system of the Milnor algebra Mf .
Summarizing, when char(k) = 0 or char(k) > max{n(d − 1), d + 1}, we have
the following commutative diagram of SL(n)-equivariant morphisms:
PSd+1
∇
∆▼
▼▼▼
▼▼A▼
▼▼▼
&
P Dn(d−1)
/
Grass n, Sd
♥♥♥
♥♥♥
♥
♥
♥♥♥ A
♥w
Res
Remark 1.5. In [2], Alper and Isaev define the associated form A(g1 , . . . , gn ) as
an element of Dn(d−1) , which they achieve by choosing a canonical generator of
the socle of S/(g1 , . . . , gn ) given by the Jacobian determinant of g1 , . . . , gn . For
our purposes, it will suffice to consider A(hg1 , . . . , gn i) defined up to a scalar.
1.4. Main results.
Theorem 1.6. Let d ≥ 2. Suppose that either char(k) = 0 or char(k) >
max{n(d − 1), d + 1}. Let f ∈ Sd+1 be a smooth form. Then the following are
equivalent:
(1)
(2)
(3)
(4)
(5)
(6)
f is a direct sum.
∇f is a balanced direct sum.
A(f ) is a balanced direct product.
A(f ) is a direct product.
∇f admits a non-trivial Gm -action defined over k.
A(f ) admits a non-trivial Gm -action defined over k.
Moreover, if z1 , . . . , zn is a basis of V ∨ in which A(f ) factors as A(f ) =
G1 (z1 , . . . , za )G2 (za+1 , . . . , zn ), then f decomposes as
f = f1 (x1 , . . . , xa ) + f2 (xa+1 , . . . , xn )
in the dual basis x1 , . . . , xn of V .
6
MAKSYM FEDORCHUK
Recall from [3, §3.1] that for a non-degenerate form f ∈ Sd+1 , a decomposition
f = f1 + · · · + fr ,
is called a maximally fine direct sum decomposition if V = E(f1 ) ⊕ · · · ⊕ E(fr ),
and fi is not a direct sum in Symd+1 E(fi ), for all i = 1, . . . , r. For nondegenerate forms of degree d + 1 ≥ 3, Kleppe has established that a maximally
fine direct sum decomposition is unique [9, Theorem 3.7]. We use Theorem 1.6
to give an alternate proof of this result for smooth forms, deducing it from the
fact that a polynomial ring over a field is a UFD:
Proposition 1.7. Let d ≥ 2. Suppose that either char(k) = 0 or char(k) >
n(d − 1). Let f ∈ Sd+1 be a smooth form. Then f has a unique maximally fine
direct sum decomposition.
Theorem 1.8. Let d ≥ 2. Suppose k is an algebraically closed field with
char(k) = 0. Then the following are equivalent for a GIT stable f ∈ Sd+1 :
(1) f is a direct sum.
s
(2) The morphism ∇ : PSd+1 → Grass n, Sd has a positive fiber dimension at hf i.
(3) ∇f has a Gm -action.
(4) ∇f is strictly semistable.
(5) dimk SL(V ) · ∇f ≤ n2 − 2.
Consequently, the locus of direct sums is closed in the stable locus.
1.5. Prior works. In [3], Buczyńska, Buczyński, Kleppe, and Teitler prove
that for a non-degenerate form f ∈ Sd+1 over an algebraically closed field, the
apolar ideal f ⊥ has a minimal generator in degree d + 1 if and only if either f
is a direct sum, or f is a limit of direct sums in which case the GL(n)-orbit of
f contains an element of the form
(1.9)
ℓ
X
i=1
xi
∂h(xℓ+1 , . . . , x2ℓ )
+ g(xℓ+1 , . . . , xn ),
∂xℓ+i
where h and g are degree d+1 forms, in ℓ and n−ℓ variables, respectively. Since
the form given by Equation (1.9) is visibly SL(n)-unstable, and in particular
singular, this translates into a computable and effective criterion for recognizing
whether a smooth form f is a direct sum over an algebraically closed field.
In [9], Kleppe uses the quadratic part of the apolar ideal f ⊥ to define an
associative algebra M(f ) of finite dimension over the base field (M(f ) is different
from the Milnor algebra Mf ). He then proves that, over an arbitrary field,
direct sum decompositions of f are in bijection with complete sets of orthogonal
idempotents of M(f ).
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
7
A key step in the proof of the direct sum criterion in [3] is the Jordan normal
form decomposition of a certain linear operator, which in general requires solving a characteristic equation. Similarly, finding a complete set of orthogonal
idempotents requires solving a system of quadratic equations. This makes it
challenging to turn [3] or [9] into an algorithm for finding direct sum decompositions when they exist.
The case of a linear factor in Theorem 3.1 was proved in [3, Proposition 2.12]
using a criterion of Smith and Stong [12] for indecomposability of Gorenstein
Artin algebras into connected sums. Our proof of the linear factor case and the
statement for higher degree factors appear to be new. Corollaries 3.6 and 3.7 are
generalizations of [3, Corollary 1.2], whose proof relies on a theorem of Shafiei
[11] saying that the apolar ideals of the generic determinant and permanent are
generated in degree 2; our approach is independent of Shafiei’s results.
2. Proofs of decomposability criteria
Some implications in the statements of Theorems 1.6 and 1.8 are easy observations, the main of which is separated in Lemma 2.1 below. Others are found
in recent papers [6, 7]. The remaining key ingredient that completes the main
circle of implications is separated into Proposition 2.2 below.
Lemma 2.1 (No restrictions on k). Suppose f ∈ Sd+1 is a direct sum such that
∇f 6= (0). Then the following hold:
(1) There is a non-trivial one-parameter subgroup ρ of SL(V ) that fixes ∇f
and such that we have the following ρ-weight-space decompositions
V = V(n−a) ⊕ V(−a) and ∇f = ∇f((n−a)d) ⊕ ∇f(−ad) .
(2) If dimk ∇f = n, then ∇f ∈ Grass(n, Sd ) is a balanced direct sum.
(3) If dimk ∇f = n, then there is a family {gt | t ∈ k ∗ } of pairwise nonproportional forms in Sd+1 such that ∇gt = ∇f for all t ∈ k ∗ , and
∇(gt − cf ) 6= (0) for all t 6= 1 ∈ k ∗ and c ∈ k.
Proof. This is obvious. Namely, suppose f = f1 (x1 , . . . , xa )+f2 (xa+1 , . . . , xn ) in
some basis x1 , . . . , xn of V , where f1 , f2 6= 0. Then the one-parameter subgroup
acting with weight (n−a) on {xi }ai=1 and weight −a on {xi }ni=a+1 clearly satisfies
(1).
Suppose further that dimk ∇f = n. From
∇f = ∇f1 ⊕ ∇f2 = ∇(f1 + tf2 ), for all t 6= 0 ∈ k,
we see that dimk (∇f1 ) = a and dimk (∇f2 ) = n − a. Thus ∇f = ∇f1 ⊕ ∇f2 is a
balanced direct sum. This proves (2). Taking gt = f1 + tf2 for t 6= 0 proves (3).
8
MAKSYM FEDORCHUK
Proof of Theorem 1.6. The implications (1) =⇒ (2) =⇒ (5) are in Lemma 2.1.
Next we prove (2) =⇒ (1). Suppose ∇f decomposes as a balanced direct sum
in a basis x1 , . . . , xn of V . Then
∇f = hs1 , . . . , sa , t1 , . . . , tn−a i,
for some s1 , . . . , sa ∈ k[x1 , . . . , xa ]d and t1 , . . . , tn−a ∈ k[xa+1 , . . . , xn ]d .
It follows that for every 1 ≤ i ≤ a and a + 1 ≤ j ≤ n, we have
∂2f
∈ k[x1 , . . . , xa ]d−1 ∩ k[xa+1 , . . . , xn ]d−1 = (0).
∂xi ∂xj
Using the assumption on char(k), we conclude that
f ∈ k[x1 , . . . , xa ]d+1 ⊕ k[xa+1 , . . . , xn ]d+1 ,
and so is a direct sum, in the same basis as ∇f .
The equivalence (2) ⇐⇒ (3) is proved in Proposition 2.2 below. This concludes the proof of equivalence for the first three conditions.
Next we prove (4) =⇒ (3). Suppose A(f ) = G1 (z1 , . . . , za )G2 (za+1 , . . . , zn )
is a direct product decomposition in a basis z1 , . . . , zn of V ∨ . Let x1 , . . . , xn
be the dual basis of V . Suppose xd11 · · · xdnn is the smallest with respect to the
graded reverse lexicographic order monomial of degree n(d − 1) that does not lie
in (Jf )n(d−1) . Since z1d1 · · · zndn must appear with a nonzero coefficient in A(f ),
we have that d1 + · · · + da = deg G1 . On the other hand, by [7, Lemma 4.1],
we have that d1 + · · · + da ≤ a(d − 1). It follows that deg G1 ≤ a(d − 1). By
symmetry, we also have that deg G2 ≤ (n − a)(d − 1). We conclude that both
inequalities must be equalities and so A(f ) = G1 G2 is a balanced direct product
decomposition. Alternatively, we can consider a diagonal action of Gm ⊂ SL(V )
on V that acts on V ∨ as follows:
t · (z1 , . . . , zn ) = t(n−a) z1 , . . . , t(n−a) za , t−a za+1 , . . . , t−a zn .
Then A(f ) is homogeneous with respect to this action, and has weight (n −
a) deg G1 − a deg G2 . However, the relevant parts of the proof of [6, Theorem
1.2] go through to show that A(f ) satisfies the Hilbert-Mumford numerical
criterion for semistability. This forces (n − a) deg G1 − a deg G2 = 0.
We now turn to the last two conditions. First, the morphism A is an SL(n)equivariant locally closed immersion by [2, §2.5], and so is stabilizer preserving.
This proves the equivalence (5) ⇐⇒ (6). The implication (5) =⇒ (1) follows
from the proof of [6, Theorem 1.0.1] that shows that for a smooth f , the gradient
point ∇f has a non-trivial Gm -action if and only if f is a direct sum. We
note that even though stated over C, the relevant parts of the proof of [6,
Theorem 1.0.1] use only [6, Lemma 3.5], which remains valid over a field k with
char(k) = 0 or char(k) > d + 1, and the fact that a smooth form over any field
must satisfy the Hilbert-Mumford numerical criterion for stability.
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
9
Proof of Theorem 1.8. By [6, Theorem 1.0.1], for every GIT stable f , the gradient point ∇f is polystable. Furthermore, it admits a Gm -action if and only
if f is a direct sum. Moreover, [6, Proposition 2.1] shows that the morphism of
the GIT quotients
∇// SL(n) : (PSd+1 )s // SL(n) → Grass(n, Sd )ss // SL(n)
is injective. This proves that for every stable f , the fiber dimension of ∇ at hf i
equals to the dimension of the stabilizer of ∇f . This concludes the proof of all
equivalences. The fact that the locus of direct sums is closed in (PSd+1 )s now
follows from the upper semicontinuity (on the domain) of fiber dimensions.
Proposition 2.2. Let d ≥ 2. Suppose k is a field with char(k) = 0 or char(k) >
n(d − 1). Then an element U ∈ Grass(n, Symd V )Res is a balanced direct sum if
and only if A(U) is a balanced direct product. Moreover, if z1 , . . . , zn is a basis
of V ∨ in which A(U) factors as a balanced direct product, then U decomposes
as a balanced direct sum in the dual basis x1 , . . . , xn of V .
Proof. The forward implication is an easy observation. Consider a balanced
direct sum U = hg1, . . . , gn i ∈ Grass(n, k[x1 , . . . , xn ]d )Res , where g1 , . . . , ga ∈
k[x1 , . . . , xa ]d and ga+1 , . . . , gn ∈ k[xa+1 , . . . , xn ]d . Then, up to a nonzero scalar,
A(U) = A(g1, . . . , ga )A(ga+1 , . . . , gn ), where A(g1 , . . . , ga ) ∈ k[z1 , . . . , za ]a(d−1)
and A(ga+1 , . . . , gn ) ∈ k[za+1 , . . . , zn ](n−a)(d−1) ; see [7, Lemma 2.11], which also
follows from the fact that on the level of algebras, we have
k[x1 , . . . , xa ]
k[xa+1 , . . . , xn ]
k[x1 , . . . , xn ]
≃
⊗k
.
(g1 , . . . , gn )
(g1 , . . . , ga )
(ga+1 , . . . , gn )
Suppose now A(U) is a balanced direct product in a basis z1 , . . . , zn of V ∨ :
(2.3)
A(U) = F1 (z1 , . . . , za )F2 (za+1 , . . . , zn ),
where deg(F1 ) = a(d − 1) and deg(F2 ) = (n − a)(d − 1). Let x1 , . . . , xn be the
dual basis of V , and let IU ⊂ k[x1 , . . . , xn ] be the complete intersection ideal
spanned by the elements of U. We have that
IU = A(U)⊥ ⊂ k[x1 , . . . , xn ].
It is then evident from (2.3) and the definition of an apolar ideal that
(2.4)
(x1 , . . . , xa )a(d−1)+1 ⊂ IU
and
(2.5)
(xa+1 , . . . , xn )(n−a)(d−1)+1 ⊂ IU .
We also have the following observation:
10
MAKSYM FEDORCHUK
Claim 2.6.
dimk U ∩ (x1 , . . . , xa ) = a,
dimk U ∩ (xa+1 , . . . , xn ) = n − a.
Proof. By symmetry, it suffices to prove the second statement. Since U is
spanned by a length
n regular sequence of degree d forms, we have that dimk U ∩
(xa+1 , . . . , xn ) ≤ n − a. Suppose we have a strict inequality. Let
R := k[x1 , . . . , xn ]/(IU , xa+1 , . . . , xn ) ≃ k[x1 , . . . , xa ]/I ′ .
Then I ′ is generated in degree d, and has at least a + 1 minimal generators in
that degree. It follows that the top degree of R is strictly less than a(d − 1),
′
and so Ia(d−1)
= k[x1 , . . . , xa ]a(d−1) (cf. [7, Lemma 2.7]). But then
k[x1 , . . . , xa ]a(d−1) ⊂ (xa+1 , . . . , xn ) + IU .
Using (2.5), this gives
k[x1 , . . . , xa ]a(d−1) k[xa+1 , . . . , xn ](n−a)(d−1) ⊂ IU .
Thus every monomial of
k[z1 , . . . , za ]a(d−1) k[za+1 , . . . , zn ](n−a)(d−1)
appears with coefficient 0 in A(U), which contradicts (2.3).
At this point, we can apply [7, Prop. 3.1] to conclude that U ∩ k[x1 , . . . , xa ]d
contains a regular sequence of length a and that U ∩ k[xa+1 , . . . , xn ]d contains a
regular sequence of length n−a. This shows that U decomposes as a balanced direct sum in the basis x1 , . . . , xn of V . However, for the sake of self-containedness,
we proceed to give a more direct argument:
By Claim (2.6), there exists a regular sequence s1 , . . . , sa ∈ k[x1 , . . . , xa ]d such
that
(s1 , . . . , sa ) = g1 (x1 , . . . , xa , 0, . . . , 0), . . . , gn (x1 , . . . , xa , 0, . . . , 0)
and a regular sequence t1 , . . . , tn−a ∈ k[xa+1 , . . . , xn ]d such that
(t1 , . . . , tn−a ) = g1 (0, . . . , 0, xa+1 , . . . , xn ), . . . , gn (0, . . . , 0, xa+1 , . . . , xn ) .
Let
W := hs1 , . . . , sa , t1 , . . . , tn−a i ∈ Grass(n, Sd )Res
and let IW be the ideal generated by W . We are going to prove that U = W ,
which will conclude the proof of the proposition.
Since char(k) = 0 or char(k) > n(d − 1), Macaulay’s theorem applies, and so
to prove that U = W , we need to show that the ideals IU and IW coincide in
degree n(d − 1). For this, it suffices to prove that (IW )n(d−1) ⊂ (IU )n(d−1) .
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
11
Since s1 , . . . , sa is a regular sequence in k[x1 , . . . , xa ]d , we have that
k[x1 , . . . , xa ]a(d−1)+1 ⊂ (s1 , . . . , sa ).
Similarly, we have that
k[xa+1 , . . . , xn ](n−a)(d−1)+1 ⊂ (t1 , . . . , tn−a ).
Together with (2.4) and (2.5), this gives
(2.7)
(x1 , . . . , xa )a(d−1)+1 + (xa+1 , . . . , xn )(n−a)(d−1)+1 ⊂ IU ∩ IW .
Set J := (x1 , . . . , xa )a(d−1)+1 + (xa+1 , . . . , xn )(n−a)(d−1)+1 . It remains to show
that
(IW )n(d−1) ⊂ (IU )n(d−1) + Jn(d−1) .
To this end, consider
a
X
qi si +
i=1
n−a
X
rj tj ∈ (IW )n(d−1) ,
j=1
where q1 , . . . , qa , r1 , . . . , rn−a ∈ Sn(d−1)−d . Since s1 , . . . , sa ∈ k[x1 , . . . , xa ]d , and
we are working modulo J, we can assume that qi ∈ (xa+1 , . . . , xn )(n−a)(d−1) , for
all i = 1, . . . , a. Similarly, we can assume that rj ∈ (x1 , . . . , xa )a(d−1) , for all
j = 1, . . . , n − a.
By construction, we have s1 , . . . , sa ∈ IU + (xa+1 , . . . , xn ) and t1 , . . . , tn−a ∈
IU + (x1 , . . . , xa ). Using this, and (2.7), we conclude that
a
X
i=1
qi si +
n−a
X
rj tj ∈ IU + J.
j=1
This finishes the proof of the proposition.
Proof of Proposition 1.7. If char(k) = 3, the case of (n, d) = (2, 2) is vacuous
since no smooth binary cubic will be a direct sum. In all other cases, char(k) >
n(d − 1) implies char(k) > max{n(d − 1), d + 1}.
Suppose f = f1 + · · · + fs = g1 + · · · + gt are two maximally fine direct sum
decompositions. Then
A(f1 ) · · · A(fs ) = A(g1 ) · · · A(gt ) ∈ PDn(d−1) ,
where V ∨ = ⊕si=1 E(A(fi )) = ⊕tj=1 E(A(gj )). Suppose some A(fi ) shares irreducible factors with more than one A(gj ). Then by the uniqueness of factorization in D, we must have a non-trivial factorization A(fi ) = G1 G2 such
that E(G1 ) ∩ E(G2 ) = (0). Then A(fi ) is a direct product, and so fi must
be a direct sum by Theorem 1.6, contradicting the maximality assumption.
Therefore, no A(fi ) shares an irreducible factor with more than one A(gj );
and, by symmetry, no A(gj ) shares an irreducible factor with more than one
12
MAKSYM FEDORCHUK
A(fi ). It follows that s = t and, up to reordering, A(fi ) = A(gi ), and thus
E(A(fi )) = E(A(gi )), for all i = 1, . . . , t. We conclude that E(fi ) = E(gi ),
which using f1 + · · · + ft = g1 + · · · + gt forces fi = gi , for all i = 1, . . . , t.
3. Necessary conditions for direct sum decomposability
Our next two results give easily verifiable necessary conditions for an arbitrary
form to be a direct sum. They hold over an arbitrary field, with no restriction
on characteristic. We keep notation of §1.1.
Theorem 3.1. Suppose f is a form in S.
⌋, then
(1) Let b = dimk ∇f . If f has a factor g such that dimk ∇g ≤ ⌊ b−1
2
f is not a direct sum.
(2) If f has a repeated factor, then f is not a direct sum.
Corollary 3.2. Suppose f is a form with dimk ∇f ≥ 3. If f has a linear factor,
then f is not a direct sum.
Proof of Theorem 3.1. We apply Lemma 2.1(1). For (1), suppose that in some
basis of V we have
f = gh = f1 (x1 , . . . , xa ) + f2 (xa+1 , . . . , xn ) ∈ S,
and that dimk ∇f = b while dimk ∇g ≤ ⌊ b−1
⌋. Let ρ be the 1-PS subgroup of
2
a
SL(V ) acting with weight (n − a) on {xi }i=1 and weight −a on {xi }ni=a+1 . Then
ρ fixes ∇f , and ∇f = (∇f )((n−a)d) ⊕ (∇f )(−ad) is the decomposition into the
ρ-weight-spaces. Since ∇f ⊂ g∇h + h∇g, we have
b+1
.
dimk (∇f ∩ g∇h) ≥ b − dimk h∇g = b − dimk ∇g ≥
2
It follows by dimension considerations that some nonzero multiple of g belongs
to one of the two weight-spaces of ρ in ∇f . Thus g itself is homogeneous with
respect to ρ. It follows that either g ∈ k[x1 , . . . , xa ] or g ∈ k[xa+1 , . . . , xn ]. This
forces either f2 = 0 or f1 = 0, respectively. A contradiction!
For (2), suppose f is a direct sum with a repeated factor g. Let ρ be the 1-PS
of SL(V ) as above. Since ∇f ⊂ (g), some nonzero multiple of g belongs to a
weight-space of ρ in ∇f , and so we obtain a contradiction as in (1).
Our next result needs the following:
Definition 3.3. Given a basis x1 , . . . , xn of V and a nonzero f ∈ Sd+1 , we
define the state of f to be the set of multi-indices Ξ(f ) ⊂ {(d1 , . . . , dn ) ∈ Zn≥0 |
d1 + · · · + dn = d + 1} such that
X
f=
a(d1 ,...,dn ) xd11 · · · xdnn , where a(d1 ,...,dn ) ∈ k ∗ .
(d1 ,...,dn )∈Ξ(f )
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
13
In other words, the state of f is the set of monomials appearing with nonzero
coefficient in f . We set Ξ(0) = ∅.
Theorem 3.4. Suppose k 6= F2 . Let d ≥ 2. Suppose f ∈ Sd+1 is such that in
some basis x1 , . . . , xn of V the following conditions hold:
(0) Ξ(∂f /∂xi ) 6= ∅ for all 1 ≤ i ≤ n.
(1) Ξ(∂f /∂xi ) ∩ Ξ(∂f /∂xj ) = ∅ for all 1 ≤ i < j ≤ n.
(2)
n
Ξ(f ) = (d1 , . . . , dn ) char(k) ∤ di for some 1 ≤ i ≤ n and
o
∂(xd11 · · · xdnn )
Ξ
⊂ ∪nj=1 Ξ(∂f /∂xj ), for all 1 ≤ i ≤ n .
∂xi
(3) The graph with the vertices in {1, . . . , n} and the edges given by
∂2f
(ij) |
6= 0
∂xi ∂xj
is connected.
Then f is not a direct sum.
Remark 3.5. In words, (1) says that no two first partials of f share a common
monomial, and (2) says that any monomial all of whose nonzero first partials
appear in first partials of f must appear in f .
As an immediate corollary of this theorem, we show that the n × n generic
determinant and permanent polynomials, and the 2n × 2n generic pfaffian polynomials, as well as any other polynomial of the same state, are not direct sums
when n ≥ 3.
Corollary 3.6 (Determinant-like
Ppolynomials are not direct sums). Let n ≥ 3.
Suppose S = k[xi,j ]ni,j=1 and f = σ∈Sn aσ x1,σ(1) · · · xn,σ(n) , where aσ ∈ k ∗ . Then
f is not a direct sum.
Corollary 3.7 (Pfaffian-like polynomials are not direct sums). Let n ≥ 3.
Suppose
S = k[xi,j ]1≤i<j≤2n , where we set xj,i := xi,j for j > i, and f =
P
∗
σ∈S2n aσ xσ(1),σ(2) · · · xσ(2n−1),σ(2n) , where aσ ∈ k . Then f is not a direct sum.
Proof of both corollaries. It is easy to see that f satisfies all conditions of The
orem 3.4.
Proof of Theorem 3.4. If char(k) = p, we set
Ξp := {(d1 , . . . , dn ) | p divides di for all 1 ≤ i ≤ n}
to be the set of all monomials whose gradient point is trivial.
14
MAKSYM FEDORCHUK
Suppose f is a direct sum. Note that Condition 3.4 implies that dimk ∇f =
n. Then by Lemma 2.1(3) there exists a form g such that ∇g = ∇f and
∇(g − cf ) 6= 0 for all c ∈ k. Since ∇g = ∇f , by Condition (2), we must
have Ξ(g) ⊂ Ξ(f ) ∪ Ξp . Then Ξ(∂g/∂xi ) ⊂ Ξ(∂f /∂xi ). Since ∂g/∂xi ∈ ∇f ,
Condition (1) implies that in fact
∂g/∂xi = ci ∂f /∂xi , for some ci ∈ k.
Comparing the second partials, and using Condition (3) we conclude that ci = cj
for all 1 ≤ i < j ≤ n. We obtain ∇(g − c1 f ) = (0), which is a contradiction.
4. Finding a balanced direct product decomposition
algorithmically
In this section, we show how Theorem 1.6 reduces the problem of finding a
direct sum decomposition of a given smooth form f to a polynomial factorization
problem. To begin, suppose that we are given a smooth form f ∈ Symd+1 V in
some basis of V . Then the associated form A(f ) is computed in the dual basis
of V ∨ as the form apolar to (Jf )n(d−1) . To apply Theorem 1.6, we now need to
determine if A(f ) decomposes as a balanced direct product, and if it does, then
in what basis of V ∨ . The following simple lemma explains how to do it (cf. §1.1
for notation):
Lemma 4.1. Suppose char(k) = 0 or char(k) > max{n(d − 1), d + 1}. For a
smooth f ∈ Sd+1 , the associated form A(f ) is a balanced direct product if and
only if there is a non-trivial factorization A(f ) = G1 G2 such that
⊥
∨
(4.2) V = G⊥
1 1 + G2 1 , or, equivalently, E(G1 ) ∩ E(G2 ) = (0) ⊂ V .
⊥
Moreover, in this case, we have G⊥
∩
G
1 1
2 1 = (0) ⊂ V and A(f ) decomposes as a balanced direct product in any basis of V ∨ such that its dual basis is
compatible with the direct sum decomposition in Equation (4.2).
Proof. The equivalence of the two conditions in (4.2) follows from the fact that
E(Gi ) ⊂ V ∨ is dual to (G⊥
i )1 ⊂ V . The claim now follows from definitions and
Theorem 1.6 by observing
that for any non-trivial factorization A(f ) = G1 G2 ,
⊥
we have G⊥
∩
G
⊂
A(f )⊥ 1 = (Jf )1 = (0).
1 1
2 1
4.1. An algorithm for direct sum decompositions. Suppose k is a field,
with either char(k) = 0 or char(k) > max{n(d −1), d + 1}, for which there exists
a polynomial factorization algorithm. Let f ∈ k[x1 , . . . , xn ]d+1 , where d ≥ 2.
Step 1: Compute Jf = (∂f /∂x1 , . . . , ∂f /∂xn ) up to degree n(d − 1) + 1. If
(Jf )n(d−1)+1 6= k[x1 , . . . , xn ]n(d−1)+1 ,
then f is not smooth and we stop; otherwise, continue.
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
15
Step 2: Compute A(f ) as the dual to (Jf )n(d−1) :
hA(f )i = T ∈ k[z1 , . . . , zn ]n(d−1) | g ◦ T = 0, for all g ∈ (Jf )n(d−1) .
Step 3: Compute the irreducible factorization of A(f ) in k[z1 , . . . , zn ] and check
for the existence of balanced direct product factorizations using Lemma 4.1. If
any exist, then f is a direct sum; otherwise, f is not a direct sum.
Step 4: For every balanced direct product factorization of A(f ), Lemma 4.1
gives a basis of V in which f decomposes as a direct sum.
The above algorithm was implemented in a Macaulay 2 package written by
Justin Kim and Zihao Fang (its source code is available upon request). In what
follows, we give a few examples of the algorithm in action.
Remark 4.3. Jaroslaw Buczyński has pointed out that already Step 2 in the
above algorithm is computationally highly expensive when n and d are large.
However, it is reasonably fast when both n and d are small, with Example 4.5
below taking only a few seconds.
Example 4.4 (Binary quartics). Suppose k is an algebraically closed field of
characteristic 0. Then every smooth binary quartic has a standard form
ft = x41 + x42 + tx21 x22 ,
t 6= ±2.
Up to a scalar, the associated form of Ft is
A(ft ) = t(z14 + z24 ) − 12z12 z22 .
Clearly, A(ft ) is singular if and only if t = 0, or t = ±6. For these values of t,
A(ft ) is in fact a balanced direct product, and so ft is a direct sum. Namely,
up to scalars, we have:
A(f0 ) = z12 z22 ,
f0 = x41 + x42 ,
A(f6 ) = (z12 − z22 )2 = (z1 − z2 )2 (z1 + z2 )2 ,
f6 = (x1 − x2 )4 + (x1 + x2 )4 ,
A(f−6 ) = (z12 + z22 )2 = (z1 − iz2 )2 (z1 + iz2 )2 , f−6 = (x1 − ix2 )4 + (x1 + ix2 )4 .
Note that over R, the associated form A(f−6 ) = (z12 + z22 )2 is not a balanced
direct product. Hence f−6 is not a direct sum over R by Theorem 1.6. Since
the apolar ideal of f−6 is the same over R and over C, this example illustrates
that the direct sum decomposability criterion of [3] fails over non-closed fields.
Example 4.5. Consider the following element in Q[x1 , x2 , x3 , x4 ]4 :
f = x41 + 4x31 x2 + 6x21 x22 + 4x1 x32 + 2x42 + 8x31 x3 + 24x21 x2 x3 + 24x1 x22 x3 + 8x32 x3 + 24x21 x23
+ 48x1 x2 x23 + 24x22 x23 + 32x1 x33 + 32x2 x33 + 17x43 − 12x31 x4 − 36x21 x2 x4 − 36x1 x22 x4 − 12x32 x4
−72x21 x3 x4 −144x1 x2 x3 x4 −72x22 x3 x4 −144x1 x23 x4 −144x2 x23 x4 −96x33 x4 +54x21 x24 +108x1 x2 x24
+ 54x22 x24 + 216x1 x3 x24 + 217x2 x3 x24 + 216x23 x24 − 108x1 x34 − 108x2 x34 − 216x3 x34 + 82x44 .
16
MAKSYM FEDORCHUK
Then its associated form is
A(f ) = 9785z18−32316z17z2 +26370z16z22 −260z15z23 +15z14 z24 +24z13z25 −19488z17z3 +40920z16z2 z3
− 25620z15z22 z3 − 180z14z23 z3 + 60z13 z24 z3 − 12z12z25 z3 + 8730z16z32 − 11910z15z2 z32 + 6390z14z22 z32
+30z13z23 z32 −595z15z33 −495z14z2 z33 +15z13z22 z33 −5z12z23 z33 +15z14z34 +120z13z2 z34 +12z13z35 −12z12z2 z35
− 4194z17z4 − 9000z16z2 z4 + 17820z15z22 z4 − 360z14z23 z4 + 90z13 z24 z4 − 7200z16z3 z4 + 21600z15z2 z3 z4
−17280z14z22 z3 z4 +6480z15z32 z4 −8640z14z2 z32 z4 +4320z13z22 z32 z4 −720z14z33 z4 +90z13z34 z4 −7395z16z42
+ 7140z15z2 z42 + 2970z14z22 z42 − 60z13z23 z42 + 15z12z24 z42 + 3120z15z3 z42 − 720z14z2 z3 z42 − 2880z13z22 z3 z42
+ 1080z14z32 z42 − 1440z13z2 z32 z42 + 720z12z22 z32 z42 − 120z13z33 z42 + 15z12z34 z42 − 1800z15z43 + 2880z14z2 z43
+1440z14z3 z43 −1440z13z2 z3 z43 +30z14 z44 +240z13z2 z44 +120z13z3 z44 −120z12z2 z3 z44 +36z13 z45 +2z12z46 .
One checks that A(f ) = G1 G2 , where G1 = z12 with E(G1 ) = z1 , and E(G2 ) =
h3z3 + 2z4 , 3z2 + z4 , 3z1 + z4 i, is a balanced direct product factorization of A(f ).
It follows that f is a direct sum. In fact, f is projectively equivalent to x41 +
(x42 + x43 + x44 + x2 x3 x24 ).
Acknowledgments
The author is grateful to Jarod Alper for an introduction to the subject,
Alexander Isaev for numerous stimulating discussions that inspired this work,
and Zach Teitler for questions that motivated most of the results in Section
3. The author was partially supported by the NSA Young Investigator grant
H98230-16-1-0061 and Alfred P. Sloan Research Fellowship. Justin Kim and
Zihao Fang wrote a Macaulay 2 package for computing associated forms while
supported by the Boston College Undergraduate Research Fellowship grant under the direction of the author.
References
[1] Jarod Alper and Alexander Isaev. Associated forms in classical invariant theory and their
applications to hypersurface singularities. Math. Ann., 360(3-4):799–823, 2014.
[2] Jarod Alper and Alexander Isaev. Associated forms and hypersurface singularities: The
binary case. J. reine angew. Math., 2016. To appear, DOI: 10.1515/crelle-2016-0008.
[3] Weronika Buczyńska, Jaroslaw Buczyński, Johannes Kleppe, and Zach Teitler. Apolarity
and direct sum decomposability of polynomials. Michigan Math. J., 64(4):675–719, 2015.
[4] Michael Eastwood and Alexander Isaev. Extracting invariants of isolated hypersurface
singularities from their moduli algebras. Math. Ann., 356(1):73–98, 2013.
[5] David Eisenbud. Commutative algebra with a view toward algebraic geometry, volume
150 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1995.
[6] Maksym Fedorchuk. GIT semistability of Hilbert points of Milnor algebras. Math. Ann.,
367(1-2):441–460, 2017.
[7] Maksym Fedorchuk and Alexander Isaev. Stability of associated forms, 2017. Preprint,
arXiv:1703.00438.
DECOMPOSABILITY OF POLYNOMIALS AND ASSOCIATED FORMS
17
[8] Anthony Iarrobino and Vassil Kanev. Power sums, Gorenstein algebras, and determinantal loci, volume 1721 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1999.
[9] Johannes Kleppe. Additive splittings of homogeneous polynomials, 2013. Ph.D. thesis,
arXiv:1307.3532.
[10] Marcos Sebastiani and René Thom. Un résultat sur la monodromie. Invent. Math., 13:90–
96, 1971.
[11] Sepideh Masoumeh Shafiei. Apolarity for determinants and permanents of generic matrices. J. Commut. Algebra, 7(1):89–123, 2015.
[12] Larry Smith and R. E. Stong. Projective bundle ideals and Poincaré duality algebras. J.
Pure Appl. Algebra, 215(4):609–627, 2011.
[13] Zach Teitler. Sufficient conditions for Strassen’s additivity conjecture. Illinois J. Math.,
59(4):1071–1085, 2015.
(Fedorchuk) Department of Mathematics, Boston College, 140 Commonwealth
Ave, Chestnut Hill, MA 02467, USA
E-mail address: [email protected]
| 0math.AC
|
GraphVAE: Towards Generation of Small Graphs Using Variational
Autoencoders
Martin Simonovsky 1 Nikos Komodakis 1
arXiv:1802.03480v1 [cs.LG] 9 Feb 2018
Abstract
Deep learning on graphs has become a popular
research topic with many applications. However,
past work has concentrated on learning graph embedding tasks, which is in contrast with advances
in generative models for images and text. Is it
possible to transfer this progress to the domain of
graphs? We propose to sidestep hurdles associated with linearization of such discrete structures
by having a decoder output a probabilistic fullyconnected graph of a predefined maximum size
directly at once. Our method is formulated as
a variational autoencoder. We evaluate on the
challenging task of molecule generation.
1. Introduction
Deep learning on graphs has very recently become a popular research topic (Bronstein et al., 2017), with useful
applications across fields such as chemistry (Gilmer et al.,
2017), medicine (Ktena et al., 2017), or computer vision (Simonovsky & Komodakis, 2017). Past work has concentrated
on learning graph embedding tasks so far, i.e. encoding an
input graph into a vector representation. This is in stark
contrast with fast-paced advances in generative models for
images and text, which have seen massive rise in quality
of generated samples. Hence, it is an intriguing question
how one can transfer this progress to the domain of graphs,
i.e. their decoding from a vector representation. Moreover,
the desire for such a method has been mentioned in the past
by Gómez-Bombarelli et al. (2016).
However, learning to generate graphs is a difficult problem for methods based on gradient optimization, as graphs
are discrete structures. Unlike sequence (text) generation,
graphs can have arbitrary connectivity and there is no clear
best way how to linearize their construction in a sequence
of steps. On the other hand, learning the order for incremen1
Université Paris Est & École des Ponts ParisTech, Champs sur
Marne, France. Correspondence to: Martin Simonovsky <[email protected]>.
tal construction involves discrete decisions, which are not
differentiable.
In this work, we propose to sidestep these hurdles by having
the decoder output a probabilistic fully-connected graph of a
predefined maximum size directly at once. In a probabilistic
graph, the existence of nodes and edges, as well as their
attributes, are modeled as independent random variables.
The method is formulated in the framework of variational
autoencoders (VAE) by Kingma & Welling (2013).
We demonstrate our method, coined GraphVAE, in cheminformatics on the task of molecule generation. Molecular
datasets are a challenging but convenient testbed for our generative model, as they easily allow for both qualitative and
quantitative tests of decoded samples. While our method is
applicable for generating smaller graphs only and its performance leaves space for improvement, we believe our work
is an important initial step towards powerful and efficient
graph decoders.
2. Related work
Graph Decoders. Graph generation has been largely unexplored in deep learning. The closest work to ours is by
Johnson (2017), who incrementally constructs a probabilistic (multi)graph as a world representation according to a
sequence of input sentences to answer a query. While our
model also outputs a probabilistic graph, we do not assume
having a prescribed order of construction transformations
available and we formulate the learning problem as an autoencoder.
Xu et al. (2017) learns to produce a scene graph from an
input image. They construct a graph from a set of object
proposals, provide initial embeddings to each node and edge,
and use message passing to obtain a consistent prediction. In
contrast, our method is a generative model which produces
a probabilistic graph from a single opaque vector, without
specifying the number of nodes or the structure explicitly.
Related work pre-dating deep learning includes random
graphs (Erdos & Rényi, 1960; Barabási & Albert, 1999),
stochastic blockmodels (Snijders & Nowicki, 1997), or state
transition matrix learning (Gong & Xiang, 2003).
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
~
Figure 1. Illustration of the proposed variational graph autoencoder. Starting from a discrete attributed graph G = (A, E, F ) on n nodes
(e.g. a representation of propylene oxide), stochastic graph encoder qφ (z|G) embeds the graph into continuous representation z. Given a
e = (A,
e E,
e Fe) on predefined
point in the latent space, our novel graph decoder pθ (G|z) outputs a probabilistic fully-connected graph G
k ≥ n nodes, from which discrete samples may be drawn. The process can be conditioned on label y for controlled sampling at test time.
e
Reconstruction ability of the autoencoder is facilitated by approximate graph matching for aligning G with G.
Discrete Data Decoders. Text is the most common discrete representation. Generative models there are usually
trained in a maximum likelihood fashion by teacher forcing (Williams & Zipser, 1989), which avoids the need to
backpropagate through output discretization by feeding the
ground truth instead of the past sample at each step. Bengio
et al. (2015) argued this may lead to expose bias, i.e. possibly reduced ability to recover from own mistakes. Recently,
efforts have been made to overcome this problem. Notably,
computing a differentiable approximation using Gumbel distribution (Kusner & Hernández-Lobato, 2016) or bypassing
the problem by learning a stochastic policy in reinforcement
learning (Yu et al., 2017). Our work also circumvents the
non-differentiability problem, namely by formulating the
loss on a probabilistic graph.
Molecule Decoders. Generative models may become
promising for de novo design of molecules fulfilling certain
criteria by being able to search for them over a continuous embedding space (Olivecrona et al., 2017). With that in
mind, we propose a conditional version of our model. While
molecules have an intuitive representation as graphs, the
field has had to resort to textual representations with fixed
syntax, e.g. so-called SMILES strings, to exploit recent
progress made in text generation with RNNs (Olivecrona
et al., 2017; Segler et al., 2017; Gómez-Bombarelli et al.,
2016). As their syntax is brittle, many invalid strings tend to
be generated, which has been recently addressed by Kusner
et al. (2017) by incorporating grammar rules into decoding. While encouraging, their approach does not guarantee
semantic (chemical) validity, similarly as our method.
3. Method
We approach the task of graph generation by devising a neural network able to translate vectors in a continuous code
space to graphs. Our main idea is to output a probabilistic
fully-connected graph and use a standard graph matching
algorithm to align it to the ground truth. The proposed
method is formulated in the framework of variational autoencoders (VAE) by Kingma & Welling (2013), although
other forms of regularized autoencoders would be equally
suitable (Makhzani et al., 2015; Li et al., 2015a). We briefly
recapitulate VAE below and continue with introducing our
novel graph decoder together with an appropriate objective.
3.1. Variational Autoencoder
Let G = (A, E, F ) be a graph specified with its adjacency
matrix A, edge attribute tensor E, and node attribute matrix F . We wish to learn an encoder and a decoder to map
between the space of graphs G and their continuous embedding z ∈ Rc , see Figure 1. In the probabilistic setting
of a VAE, the encoder is defined by a variational posterior qφ (z|G) and the decoder by a generative distribution
pθ (G|z), where φ and θ are learned parameters. Furthermore, there is a prior distribution p(z) imposed on the latent
code representation as a regularization; we use a simplistic isotropic Gaussian prior p(z) = N (0, I). The whole
model is trained by minimizing the upper bound on negative
log-likelihood − log pθ (G) (Kingma & Welling, 2013):
L(φ, θ; G) =
= Eqφ (z|G) [− log pθ (G|z)] + KL[qφ (z|G)||p(z)] (1)
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
The first term of L, the reconstruction loss, enforces high
similarity of sampled generated graphs to the input graph G.
The second term, KL-divergence, regularizes the code space
to allow for sampling of z directly from p(z) instead from
qφ (z|G) later. The dimensionality of z is usually fairly small
so that the autoencoder is encouraged to learn a high-level
compression of the input instead of learning to simply copy
any given input. While the regularization is independent on
the input space, the reconstruction loss must be specifically
designed for each input modality. In the following, we
introduce our graph decoder together with an appropriate
reconstruction loss.
3.2. Probabilistic Graph Decoder
Graphs are discrete objects, ultimately. While this does not
pose a challenge for encoding, demonstrated by the recent
developments in graph convolution networks (Gilmer et al.,
2017), graph generation has been an open problem so far. In
a related task of text sequence generation, the currently dominant approach is character-wise or word-wise prediction
(Bowman et al., 2016). However, graphs can have arbitrary
connectivity and there is no clear way how to linearize their
construction in a sequence of steps1 . On the other hand,
iterative construction of discrete structures during training
without step-wise supervision involves discrete decisions,
which are not differentiable and therefore problematic for
back-propagation.
Fortunately, the task can become much simpler if we restrict the domain to the set of all graphs on maximum k
nodes, where k is fairly small (in practice up to the order
of tens). Under this assumption, handling dense graph representations is still computationally tractable. We propose
to make the decoder output a probabilistic fully-connected
e = (A,
e E,
e Fe) on k nodes at once. This effectively
graph G
sidesteps both problems mentioned above.
In probabilistic graphs, the existence of nodes and edges
is modeled as Bernoulli variables, whereas node and edge
attributes are multinomial variables. While not discussed in
this work, continuous attributes could be easily modeled as
Gaussian variables represented by their mean and variance.
We assume all variables to be independent.
e has thus a probaEach tensor of the representation of G
bilistic interpretation. Specifically, the predicted adjacency
e ∈ [0, 1]k×k contains both node probabilities A
ea,a
matrix A
e
and edge probabilities Aa,b for nodes a 6= b. The edge ate ∈ Rk×k×de indicates class probabilities for
tribute tensor E
edges and, similarly, the node attribute matrix Fe ∈ Rk×dn
contains class probabilities for nodes.
The decoder itself is deterministic. Its architecture is a
simple multi-layer perceptron (MLP) with three outputs in
its last layer. Sigmoid activation function is used to compute
e whereas edge- and node-wise softmax is applied to obtain
A,
e
E and Fe, respectively. At test time, we are often interested
e which can be obtained by
in a (discrete) point estimate of G,
e E,
e and Fe. Note
taking edge- and node-wise argmax in A,
that this can result in a discrete graph on less than k nodes.
3.3. Reconstruction Loss
Given a particular of a discrete input graph G on n ≤ k
e on k nodes,
nodes and its probabilistic reconstruction G
evaluation of Equation 1 requires computation of likelihood
e
pθ (G|z) = P (G|G).
Since no particular ordering of nodes is imposed in either
e or G and matrix representation of graphs is not invariant
G
to permutations of nodes, comparison of two graphs is hard.
However, approximate graph matching described further
in Subsection 3.4 can obtain a binary assignment matrix
e is
X ∈ {0, 1}k×n , where Xa,i = 1 only if node a ∈ G
assigned to i ∈ G and Xa,i = 0 otherwise.
Knowledge of X allows to map information between both
graphs. Specifically, input adjacency matrix is mapped to
the predicted graph as A0 = XAX T , whereas the predicted
node attribute matrix and slices of edge attribute matrix are
e0 =
transferred to the input graph as Fe0 = X T Fe and E
·,·,l
T e
X E·,·,l X. The maximum likelihood estimates, i.e. crossentropy, of respective variables are as follows:
log p(A0 |z) =
X
ea,a + (1 − A0 ) log(1 − A
ea,a ) +
= 1/k
A0a,a log A
a,a
a
+ 1/k(k − 1)
X
ea,b + (1 − A0a,b ) log(1 − A
ea,b )
A0a,b log A
a6=b
log p(F |z) = 1/n
X
T e0
log Fi,·
Fi,·
i
log p(E|z) = 1/(||A||1 − n)
X
T e0
log Ei,j,·
Ei,j,· ,
(2)
i6=j
where we assumed that F and E are encoded in one-hot notation. The formulation considers existence of both matched
and unmatched nodes and edges but attributes of only the
matched ones. Furthermore, averaging over nodes and edges
separately has shown beneficial in training as otherwise the
edges dominate the likelihood. The overall reconstruction
loss is a weighed sum of the previous terms:
1
While algorithms for canonical graph orderings are available
(McKay & Piperno, 2014), Vinyals et al. (2015) empirically found
out that the linearization order matters when learning on sets.
− log p(G|z) = −λA log p(A0 |z) − λF log p(F |z) −
− λE log p(E|z)
(3)
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
3.4. Graph Matching
computing a differentiable loss.
The goal of (second-order) graph matching is to find correspondences X ∈ {0, 1}k×n between nodes of graphs
e based on the similarities of their node pairs
G and G
e It can be
S : (i, j)×(a, b) → R+ for i, j ∈ G and a, b ∈ G.
expressed as integer quadratic programming problem of similarity maximization over X and is typically approximated
by relaxation of X into continuous domain: X ∗ ∈ [0, 1]k×n
(Cho et al., 2014). For our use case, the similarity function
is defined as follows:
S((i, j), (a, b)) =
T e
ea,b A
ea,a A
eb,b [i 6= j ∧ a 6= b] +
= (Ei,j,·
Ea,b,· )Ai,j A
T e
ea,a [i = j ∧ a = b]
+ (Fi,·
Fa,· )A
(4)
The first term evaluates similarity between edge pairs and
the second term between node pairs, [·] being the Iverson
bracket. Note that the scores consider both feature compate and existential compatibility (A),
e which
ibility (Fe and E)
has empirically led to more stable assignments during training. To summarize the motivation behind both Equations 3
and 4, our method aims to find the best graph matching
and then further improve on it by gradient descent on the
loss. Given the stochastic way of training deep network, we
argue that solving the matching step only approximately is
sufficient. This is conceptually similar to the approach for
learning to output unordered sets by (Vinyals et al., 2015),
where the closest ordering of the training data is sought.
In practice, we are looking for a graph matching algorithm
robust to noisy correspondences which can be easily implemented on GPU in batch mode. Max-pooling matching
(MPM) by Cho et al. (2014) is a simple but effective algorithm following the iterative scheme of power methods, see
Appendix A for details. It can be used in batch mode if
similarity tensors are zero-padded, i.e. S((i, j), (a, b)) = 0
for n < i, j ≤ k, and the amount of iterations is fixed.
Max-pooling matching outputs continuous assignment matrix X ∗ . Unfortunately, attempts to directly use X ∗ instead
of X in Equation 3 performed badly, as did experiments with
direct maximization of X ∗ or soft discretization with softmax or straight-through Gumbel softmax (Jang et al., 2016).
We therefore discretize X ∗ to X using Hungarian algorithm
to obtain a strict one-on-one mapping2 . While this operation
is non-differentiable, gradient can still flow to the decoder
directly through the loss function and training convergence
proceeds without problems. Note that this approach is often taken in works on object detection, e.g. (Stewart et al.,
2016), where a set of detections need to be matched to a set
of ground truth bounding boxes and treated as fixed before
2
Some predicted nodes are not assigned for n < k. Our current
implementation performs this step on CPU although a GPU version
has been published (Date & Nagi, 2016).
3.5. Further Details
Encoder. A feed forward network with edge-conditioned
graph convolutions (ECC) (Simonovsky & Komodakis,
2017) is used as encoder, although any other graph embedding method is applicable. As our edge attributes are
categorical, a single linear layer for the filter generating
network in ECC is sufficient. Due to smaller graph sizes
no pooling is used in encoder except for the global one, for
which we employ gated pooling by Li et al. (2015b). As
usual in VAE, we formulate encoder as probabilistic and
enforce Gaussian distribution of qφ (z|G) by having the last
encoder layer outputs 2c features interpreted as mean and
variance, allowing to sample zl ∼ N (µl (G), σl (G)) for
l ∈ 1, .., c using the re-parameterization trick (Kingma &
Welling, 2013).
Disentangled Embedding. In practice, rather than random drawing of graphs, one often desires more control over
the properties of generated graphs. In such case, we follow
Sohn et al. (2015) and condition both encoder and decoder
on label vector y associated with each input graph G. Decoder pθ (G|z, y) is fed a concatenation of z and y, while
in encoder qφ (z|G, y), y is concatenated to every node’s
features just before the graph pooling layer. If the size of
latent space c is small, the decoder is encouraged to exploit
information in the label.
Limitations. The proposed model is expected to be useful
only for generating small graphs. This is due to growth
of GPU memory requirements and number of parameters
(O(k 2 )) as well as matching complexity (O(k 4 )), with small
decrease in quality for high values of k. In Section 4 we
demonstrate results for up to k = 38. Nevertheless, for
many applications even generation of small graphs is still
very useful.
4. Evaluation
We demonstrate our method for the task of molecule generation by evaluating on two large public datasets of organic
molecules, QM9 and ZINC.
4.1. Application in Cheminformatics
Quantitative evaluation of generative models of images and
texts has been troublesome (Theis et al., 2015), as it very
difficult to measure realness of generated samples in an automated and objective way. Thus, researchers frequently
resort there to qualitative evaluation and embedding plots.
However, qualitative evaluation of graphs can be very unintuitive for humans to judge unless the graphs are planar and
fairly simple.
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
Fortunately, we found graph representation of molecules,
as undirected graphs with atoms as nodes and bonds as
edges, to be a convenient testbed for generative models.
On one hand, generated graphs can be easily visualized
in standardized structural diagrams. On the other hand,
chemical validity of graphs, as well as many further properties a molecule can fulfill, can be checked using software
packages (SanitizeMol in RDKit) or simulations. This
makes both qualitative and quantitative tests possible.
Chemical constraints on compatible types of bonds and atom
valences make the space of valid graphs complicated and
molecule generation challenging. In fact, a single addition
or removal of edge or change in atom or bond type can
make a molecule chemically invalid. Comparably, flipping
a single pixel in MNIST-like number generation problem is
of no issue.
To help the network in this application, we introduce three
e
remedies. First, we make the decoder output symmetric A
e
and E by predicting their (upper) triangular parts only, as
undirected graphs are sufficient representation for molecules.
Second, we use prior knowledge that molecules are connected and, at test time only, construct maximum spanning
ea,a ≥ 0.5} in
tree on the set of probable nodes {a : A
order to include its edges (a, b) in the discrete pointwise
ea,b < 0.5 originally. Third,
estimate of the graph even if A
we do not generate Hydrogen explicitly and let it be added
as ”padding” during chemical validity check.
4.2. QM9 Dataset
QM9 dataset (Ramakrishnan et al., 2014) contains about
134k organic molecules of up to 9 heavy (non Hydrogen)
atoms with 4 distinct atomic numbers and 4 bond types, we
set k = 9, de = 4 and dn = 4. We set aside 10k samples
for testing and 10k for validation (model selection).
We compare our unconditional model to the characterbased generator of Gómez-Bombarelli et al. (2016) (CVAE)
and the grammar-based generator of Kusner et al. (2017)
(GVAE). We used the code and architecture in Kusner et al.
(2017) for both baselines, adapting the maximum input
length to the smallest possible. In addition, we demonstrate a conditional generative model for an artificial task of
generating molecules given a histogram of heavy atoms as
4-dimensional label y, the success of which can be easily
validated.
Setup. The encoder has two graph convolutional layers
(32 and 64 channels) with identity connection, batchnorm,
and ReLU; followed by the graph-level output formulation
in Equation 7 of Li et al. (2015b) with auxiliary networks
being a single fully connected layer (FCL) with 128 output
channels; finalized by a FCL outputting (µ, σ). The decoder
has 3 FCLs (128, 256, and 512 channels) with batchnorm
and ReLU; followed by parallel triplet of FCLs to output
graph tensors. We set c = 40, λA = λF = λE = 1, batch
size 32, 75 MPM iterations and train for 25 epochs with
Adam with learning rate 1e-3 and β1 =0.5.
Embedding Visualization. To visually judge the quality
and smoothness of the learned embedding z of our model,
we may traverse it in two ways: along a slice and along a line.
For the former, we randomly choose two c-dimensional orthonormal vectors and sample z in regular grid pattern over
the induced 2D plane. For the latter, we randomly choose
two molecules G(1) , G(2) of the same label from test set
and interpolate between their embeddings µ(G(1) ), µ(G(2) ).
This also evaluates the encoder, and therefore benefits from
low reconstruction error.
We plot two planes in Figure 2, for a frequent label (left)
and a less frequent label in QM9 (right). Both images show
a varied and fairly smooth mix of molecules. The left image
has many valid samples broadly distributed across the plane,
as presumably the autoencoder had to fit a large portion of
database into this space. The right exhibits stronger effect
of regularization, as valid molecules tend to be only around
center.
An example of several interpolations is shown in Figure 3.
We can find both meaningful (1st, 2nd and 4th row) and less
meaningful transitions, though many samples on the lines
do not form chemically valid compounds.
Decoder Quality Metrics. The quality of a conditional
decoder can be evaluated by the validity and variety of generated graphs. For a given label y(l) , we draw ns = 104 samples z(l,s) ∼ p(z) and compute the discrete point estimate
of their decodings Ĝ(l,s) = arg max pθ (G|z(l,s) , y(l) ).
Let V (l) be the list of chemically valid molecules from
Ĝ(l,s) and C (l) be the list of chemically valid molecules
with atom histograms equal to y(l) . We are interested in ratios Valid(l) = |V (l) |/ns and Accurate(l) = |C (l) |/ns .
Furthermore, let Unique(l) = |set(C (l) )|/|C (l) | be the
fraction of unique correct graphs and Novel(l) = 1 −
|set(C (l) ) ∩ QM9|/|set(C (l) )| the fraction of novel out-ofdataset graphs; we define Unique(l) = 0 and Novel(l) = 0
if |C (l) | = 0. Finally, the introduced metrics are aggregated by frequencies of labels in QM9, e.g. Valid =
P
(l)
(l)
l Valid freq(y ). Unconditional decoders are evaluated by assuming there is just a single label, therefore
Valid = Accurate.
In Table 1, we can see that on average 50% of generated
molecules are chemically valid and, in the case of conditional models, about 40% have the correct label which the
decoder was conditioned on. Larger embedding sizes c
are less regularized, demonstrated by a higher number of
Unique samples and by lower accuracy of the conditional
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
N
O
N
N
N
N
NH
N
N
HO
OH
N
O
O
N
OH
N
H2 O
N
HO
N
O
O
OH
N
OH
OH
OH
OH2
N
HO
N
N
N
OH
N
N
N
O
OH2
OH2
OH2
NH
O
O
OH
NH
NH
NH
OH
OH2
NH2
NH
OH
OH2
OH2
OH2
OH2
OH2
OH2
NH
H2 O
H2 O
OH
NH
NH
O
HO
OH
OH2
H2 N
OH
OH
OH
NH2
OH
N
O
HO
OH
O
HO
O
N
OH
OH
NH
O
O
OH
OH
O
OH
N
OH
OH
O
N
OH
NH
O
HO
NH2
O
N
OH
OH
O
O
O
O
O
OH
O
OH
OH
NH
N
HO
O
N
OH
OH
N
O
N
HO
OH
N
N
N
OH
N
O
HO
O
OH
HO
N
N
N
O
OH2
OH
N
OH
O
OH2
O
O
O
O
OH
OH
OH
N
N
N
N
N
OH2
N
OH
OH
OH2
NH2
OH2
HN
OH
N
N
OH
OH
OH
N
N
OH2
OH
N
OH
NH2
OH
HO
O
N
N
OH
N
O
O
NH
OH2
N
O
O
NH
OH2
NH2
OH
O
O
O
NH
N
O
NH2
N
O
N
OH
N
N
OH
O
OH
H2 O
N
NH
O
NH2
N
OH2
O
N
O
H2 N
O
NH
O
NH2
N
O
N
O
OH
OH2
OH
OH2
O
OH2
OH2
NH
OH2
OH
NH
NH
NH
OH2
OH2
O
O
OH2
OH
O
OH2
OH
OH2
OH2
OH2
OH2
HN
O
NH
NH2
NH2
NH2
N
N
O
OH2
OH2
OH2
O
N
O
NH
NH
O
O
OH2
OH2
OH2
NH2
O
O
NH
O
O
OH2
H2 O
NH
O
NH
OH2
OH2
O
N
O
OH
O
OH
O
NH
NH2
N
N
O
OH2
N
OH
N
O
N
OH
H2 O
O
O
O
OH
OH2
N
NH2
OH
O
OH2
O
OH2
H2 O
OH2
OH2
O
O
OH2
O
NH2
O
OH2
OH2
N
O
N
O
NH2
OH
NH
NH
NH
O
NH2
NH2
O
O
OH2
OH
O
OH
OH2
OH
OH2
O
NH
N
O
OH
O
O
O
OH
OH2
NH
NH
O
O
OH2
O
OH2
N
O
NH
OH
O
O
HO
OH2
O
OH
NH
O
O
OH2
N
N
OH
OH
N
NH
OH2
NH
O
OH
O
O
OH2
OH2
OH
OH
N
NH
O
OH2
OH
OH
OH
O
NH2
H2 O
N
O
N
OH
NH
NH
O
OH2
N
OH2
OH2
OH
N
OH2
OH
O
NH2
O
OH2
N
OH
OH2
N
O
O
OH2
N
H2 O
OH
N
OH
OH2
OH
OH2
N
N
NH2
OH2
OH2
HN
OH
OH
OH2
OH2
OH2
NH
OH2
NH
NH
OH2
O
O
OH2
NH
H2 O
OH2
NH2
H2 O
Figure 2. Decodings of latent space points of a conditional model sampled over a random 2D plane in z-space of c = 40 (within 5 units
from center of coordinates). Left: Samples conditioned on 7x Carbon, 1x Nitrogen, 1x Oxygen (12% QM9). Right: Samples conditioned
on 5x Carbon, 1x Nitrogen, 3x Oxygen (2.6% QM9). Color legend as in Figure 3.
model, as the decoder is forced less to rely on actual labels.
The ratio of Valid samples shows less clear behavior, likely
because the discrete performance is not directly optimized
for. For all models, it is remarkable that about 60% of generated molecules are out of the dataset, i.e. the network has
never seen them during training.
Looking at the baselines, CVAE can output only very few
valid samples as expected, while GVAE generates the highest number of valid samples (60%) but of very low variance
(less than 10%). Additionally, we investigate the importance
of graph matching by using identity assignment X instead
and thus learning to reproduce particular node permutations
in the training set, which correspond to the canonical ordering of SMILES strings from rdkit. This ablated model
(denoted as NoGM in Table 1) produces many valid samples
of lower variety and, surprisingly, outperforms GVAE in
this regard. In comparison, our model can achieve good
performance in both metrics at the same time.
Likelihood. Besides the application-specific metric introduced above, we also report evidence lower bound (ELBO)
commonly used in VAE literature, which corresponds to
−L(φ, θ; G) in our notation. In Table 1, we state mean
bounds over test set, using a single z sample per graph.
We observe both reconstruction loss and KL-divergence decrease due to larger c providing more freedom. However,
there seems to be no strong correlation between ELBO and
Valid, which makes model selection somewhat difficult.
Implicit Node Probabilities. Our decoder assumes independence of node and edge probabilities, which allows for
isolated nodes or edges. Making further use of the fact that
molecules are connected graphs, here we investigate the
effect of making node probabilities a function of edge probabilities. Specifically, we consider the probability for node
ea,a = maxb A
ea,b .
a as that of its most probable edge: A
The evaluation on QM9 in Table 2 shows a clear improvement in Valid, Accurate, and Novel metrics in both the
conditional and unconditional setting. However, this is paid
for by lower variability and higher reconstruction loss. This
indicates that while the new constraint is useful, the model
cannot fully cope with it.
4.3. ZINC Dataset
ZINC dataset (Irwin et al., 2012) contains about 250k druglike organic molecules of up to 38 heavy atoms with 9
distinct atomic numbers and 4 bond types, we set k = 38,
de = 4 and dn = 9 and use the same split strategy as
with QM9. We investigate the degree of scalability of an
unconditional generative model.
Setup. The setup is equivalent as for QM9 but with a
wider encoder (64, 128, 256 channels).
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
OH
F
F
H2 N
NH2
NH2
NH2
NH2
F
H2 N
F
N
F
HN
N
F
O
F
H2 N
H2 N
F
H2 N
N
NH
OH
N
NH
NH
N
O
OH2
O
N
O
O
O
N
HN
O
O
NH
OH2
OH
O
NH
NH
O
O
OH
O
NH
O
N
O
O
HO
H2 O
OH
O
O
O
O
O
O
O
OH2
OH2
O
O
O
OH2
OH2
OH2
H2 O
OH2
O
OH
OH
OH2
OH2
O
O
O
OH
OH
O
OH
O
OH
O
N
OH
H2 O
OH
HO
N
H2 O
OH2
OH2
HO
O
HO
O
N
N
N
OH
HO
N
O
N
OH
O
N
OH
O
OH
OH
OH
Figure 3. Linear interpolation between row-wise pairs of randomly chosen molecules in z-space of c = 40 in a conditional model. Color
legend: encoder inputs (green), chemically invalid graphs (red), valid graphs with wrong label (blue), valid and correct (white).
Decoder Quality Metrics. Our best model with c = 40
has archived Valid = 0.135, which is clearly worse than
for QM9. Using implicit node probabilities brought no
improvement. For comparison, CVAE failed to generated
any valid sample, while GVAE achieved Valid = 0.357
(models provided by Kusner et al. (2017), c = 56).
We attribute such a low performance to a generally much
higher chance of producing a chemically-relevant inconsistency (number of possible edges growing quadratically). To
confirm the relationship between performance and graph
size k, we kept only graphs not larger than k = 20 nodes,
corresponding to 21% of ZINC, and obtained Valid = 0.341
(and Valid = 0.185 for k = 30 nodes, 92% of ZINC). To
verify that the problem is likely not caused by our proposed
graph matching loss, we synthetically evaluate it in the following.
Matching Robustness. Robust behavior of graph matching using our similarity function S is important for good
performance of GraphVAE. Here we study graph matching
in isolation to investigate its scalability. To that end, we add
Gaussian noise N (0, A ), N (0, E ), N (0, F ) to each tensor of input graph G, truncating and renormalizing to keep
their probabilistic interpretation, to create its noisy version
GN . We are interested in the quality of matching between
self, P [G, G], using noisy assignment matrix X between G
and GN . The advantage to naive checking X for identity is
the invariance to permutation of equivalent nodes.
In Table 3 we vary k and for each tensor separately and
report mean accuracies (computed in the same fashion as
losses in Equation 3) over 100 random samples from ZINC
with size up to k nodes. While we observe an expected
fall of accuracy with stronger noise, the behavior is fairly
robust with respect to increasing k at a fixed noise level,
the most sensitive being the adjacency matrix. Note that
accuracies are not comparable across tables due to different
dimensionalities of random variables. We may conclude
that the quality of the matching process is not a major hurdle
to scalability.
5. Conclusion
In this work we addressed the problem of generating graphs
from a continuous embedding in the context of variational
autoencoders. We evaluated our method on two molecular datasets of different maximum graph size. While we
achieved to learn embedding of reasonable quality on small
molecules, our decoder had a hard time capturing complex
chemical interactions for larger molecules. Nevertheless,
we believe our method is an important initial step towards
more powerful decoders and will spark interesting in the
community.
There are many avenues to follow for future work. Besides
the obvious desire to improve the current method (for example, by incorporating a more powerful prior distribution
or adding a recurrent mechanism for correcting mistakes),
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
log pθ (G|z)
ELBO
Valid
Accurate
Unique
Novel
Cond.
Ours c = 20
Ours c = 40
Ours c = 60
Ours c = 80
-0.578
-0.504
-0.492
-0.475
-0.722
-0.617
-0.585
-0.557
0.565
0.511
0.520
0.458
0.467
0.416
0.406
0.353
0.314
0.484
0.583
0.666
0.598
0.635
0.613
0.661
Unconditional
Table 1. Performance on conditional and unconditional QM9 models evaluated by mean test-time reconstruction log-likelihood
(log pθ (G|z)), mean test-time evidence lower bound (ELBO), and decoding quality metrics (Section 4.2). Baselines CVAE (GómezBombarelli et al., 2016) and GVAE (Kusner et al., 2017) are listed only for the embedding size with the highest Valid.
Ours c = 20
Ours c = 40
Ours c = 60
Ours c = 80
NoGM c = 80
CVAE c = 60
GVAE c = 20
-0.660
-0.537
-0.486
-0.482
-2.388
–
–
-0.916
-0.744
-0.656
-0.628
-2.553
–
–
0.485
0.542
0.517
0.557
0.810
0.103
0.602
0.485
0.542
0.517
0.557
0.810
0.103
0.602
0.457
0.618
0.695
0.760
0.241
0.675
0.093
0.575
0.617
0.570
0.616
0.610
0.900
0.809
log pθ (G|z)
ELBO
Valid
Accurate
Unique
Novel
Cond.
Ours/imp c = 20
Ours/imp c = 40
Ours/imp c = 60
Ours/imp c = 80
-0.784
-0.671
-0.618
-0.627
-0.919
-0.776
-0.714
-0.713
0.572
0.611
0.566
0.583
0.482
0.518
0.448
0.451
0.238
0.307
0.416
0.475
0.718
0.665
0.710
0.681
Uncond.
Table 2. Performance on conditional and unconditional QM9 models with implicit node probabilities. Improvement with respect to
Table 1 is emphasized in italics.
Ours/imp c = 20
Ours/imp c = 40
Ours/imp c = 60
Ours/imp c = 80
-0.857
-0.737
-0.634
-0.642
-1.091
-0.932
-0.797
-0.777
0.533
0.562
0.587
0.571
0.533
0.562
0.587
0.571
0.228
0.420
0.459
0.520
0.610
0.758
0.730
0.719
Table 3. Mean accuracy of matching ZINC graphs to their noisy
counterparts in a synthetic benchmark as a function of maximum
graph size k.
Noise
k = 15
k = 20
k = 25
k = 30
k = 35
k = 40
A,E,F = 0
99.55
99.52
99.45
99.4
99.47
99.46
A = 0.4
A = 0.8
90.95
82.14
89.55
81.01
86.64
79.62
87.25
79.67
87.07
79.07
86.78
78.69
E = 0.4
E = 0.8
97.11
92.03
96.42
90.76
95.65
89.76
95.90
89.70
95.69
88.34
95.69
89.40
F = 0.4
F = 0.8
98.32
97.26
98.23
97.00
97.64
96.60
98.28
96.91
98.24
96.56
97.90
97.17
we would like to extend it beyond a proof of concept by applying it to real problems in chemistry, such as optimization
of certain properties or predicting chemical reactions. An
advantage of a graph-based decoder compared to SMILESbased decoder is the possibility to predict detailed attributes
of atoms and bonds in addition to the base structure, which
might be useful in these tasks. Our autoencoder might also
be used to pre-train graph encoders for fine-tuning on small
datasets (Goh et al., 2017).
ACKNOWLEDGMENTS
We thank Shell Xu Hu for discussions on variational methods, Shinjae Yoo for project motivation, and anonymous
reviewers for their comments.
References
Barabási, Albert-László and Albert, Réka. Emergence of
scaling in random networks. Science, 286(5439):509–
512, 1999.
Bengio, Samy, Vinyals, Oriol, Jaitly, Navdeep, and Shazeer,
Noam. Scheduled sampling for sequence prediction with
recurrent neural networks. In NIPS, pp. 1171–1179, 2015.
Bowman, Samuel R., Vilnis, Luke, Vinyals, Oriol, Dai,
Andrew M., Józefowicz, Rafal, and Bengio, Samy. Generating sentences from a continuous space. In CoNLL, pp.
10–21, 2016.
Bronstein, Michael M, Bruna, Joan, LeCun, Yann, Szlam,
Arthur, and Vandergheynst, Pierre. Geometric deep learn-
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
ing: going beyond euclidean data. IEEE Signal Processing Magazine, 34(4):18–42, 2017.
Cho, Minsu, Sun, Jian, Duchenne, Olivier, and Ponce, Jean.
Finding matches in a haystack: A max-pooling strategy
for graph matching in the presence of outliers. In CVPR,
pp. 2091–2098, 2014.
Date, Ketan and Nagi, Rakesh. Gpu-accelerated hungarian
algorithms for the linear assignment problem. Parallel
Computing, 57:52–72, 2016.
Erdos, Paul and Rényi, Alfréd. On the evolution of random
graphs. Publ. Math. Inst. Hung. Acad. Sci, 5(1):17–60,
1960.
Gilmer, Justin, Schoenholz, Samuel S., Riley, Patrick F.,
Vinyals, Oriol, and Dahl, George E. Neural message
passing for quantum chemistry. In ICML, pp. 1263–1272,
2017.
Goh, Garrett B., Siegel, Charles, Vishnu, Abhinav, and
Hodas, Nathan O. Chemnet: A transferable and generalizable deep neural network for small-molecule property
prediction. arXiv preprint arXiv:1712.02734, 2017.
Gómez-Bombarelli, Rafael, Duvenaud, David K.,
Hernández-Lobato, José Miguel, Aguilera-Iparraguirre,
Jorge, Hirzel, Timothy D., Adams, Ryan P., and
Aspuru-Guzik, Alán. Automatic chemical design using
a data-driven continuous representation of molecules.
CoRR, abs/1610.02415, 2016.
Gong, Shaogang and Xiang, Tao. Recognition of group
activities using dynamic probabilistic networks. In ICCV,
pp. 742–749, 2003.
Irwin, John J., Sterling, Teague, Mysinger, Michael M.,
Bolstad, Erin S., and Coleman, Ryan G. ZINC: A free tool
to discover chemistry for biology. Journal of Chemical
Information and Modeling, 52(7):1757–1768, 2012.
Jang, Eric, Gu, Shixiang, and Poole, Ben. Categorical reparameterization with gumbel-softmax. CoRR,
abs/1611.01144, 2016.
Johnson, Daniel D. Learning graphical state transitions. In
ICLR, 2017.
Kingma, Diederik P. and Welling, Max. Auto-encoding
variational bayes. CoRR, abs/1312.6114, 2013.
Ktena, Sofia Ira, Parisot, Sarah, Ferrante, Enzo, Rajchl,
Martin, Lee, Matthew C. H., Glocker, Ben, and Rueckert,
Daniel. Distance metric learning using graph convolutional networks: Application to functional brain networks.
In MICCAI, 2017.
Kusner, Matt J. and Hernández-Lobato, José Miguel. GANS
for sequences of discrete elements with the gumbelsoftmax distribution. CoRR, abs/1611.04051, 2016.
Kusner, Matt J., Paige, Brooks, and Hernández-Lobato,
José Miguel. Grammar variational autoencoder. In ICML,
pp. 1945–1954, 2017.
Landrum, Greg. RDKit: Open-source cheminformatics.
URL http://www.rdkit.org.
Li, Yujia, Swersky, Kevin, and Zemel, Richard S. Generative moment matching networks. In ICML, pp. 1718–
1727, 2015a.
Li, Yujia, Tarlow, Daniel, Brockschmidt, Marc, and Zemel,
Richard S. Gated graph sequence neural networks. CoRR,
abs/1511.05493, 2015b.
Makhzani, Alireza, Shlens, Jonathon, Jaitly, Navdeep, and
Goodfellow, Ian J. Adversarial autoencoders. CoRR,
abs/1511.05644, 2015.
McKay, Brendan D. and Piperno, Adolfo. Practical graph
isomorphism, II. Journal of Symbolic Computation, 60
(0):94 – 112, 2014. ISSN 0747-7171.
Olivecrona, Marcus, Blaschke, Thomas, Engkvist, Ola, and
Chen, Hongming. Molecular de novo design through
deep reinforcement learning. CoRR, abs/1704.07555,
2017.
Ramakrishnan, Raghunathan, Dral, Pavlo O, Rupp,
Matthias, and von Lilienfeld, O Anatole. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data, 1, 2014.
Segler, Marwin H. S., Kogej, Thierry, Tyrchan, Christian,
and Waller, Mark P. Generating focussed molecule libraries for drug discovery with recurrent neural networks.
CoRR, abs/1701.01329, 2017.
Simonovsky, Martin and Komodakis, Nikos. Dynamic edgeconditioned filters in convolutional neural networks on
graphs. In CVPR, 2017.
Snijders, Tom A.B. and Nowicki, Krzysztof. Estimation
and prediction for stochastic blockmodels for graphs with
latent block structure. Journal of Classification, 14(1):
75–100, Jan 1997.
Sohn, Kihyuk, Lee, Honglak, and Yan, Xinchen. Learning
structured output representation using deep conditional
generative models. In NIPS, pp. 3483–3491, 2015.
Stewart, Russell, Andriluka, Mykhaylo, and Ng, Andrew Y.
End-to-end people detection in crowded scenes. In CVPR,
pp. 2325–2333, 2016.
GraphVAE: Towards Generation of Small Graphs Using Variational Autoencoders
Theis, Lucas, van den Oord, Aäron, and Bethge, Matthias.
A note on the evaluation of generative models. CoRR,
abs/1511.01844, 2015.
architecture, we train it as unregularized in this section,
i.e. with a deterministic encoder and without KL-divergence
term in Equation 1.
Vinyals, Oriol, Bengio, Samy, and Kudlur, Manjunath. Order matters: Sequence to sequence for sets. arXiv preprint
arXiv:1511.06391, 2015.
Unconditional models for QM9 achieve mean test loglikelihood log pθ (G|z) of roughly −0.37 (about −0.50
for the implicit node probability model) for all c ∈
{20, 40, 60, 80}. While these log-likelihoods are significantly higher than in Tables 1 and 2, our architecture can
not achieve perfect reconstruction of inputs. We were successful to increase training log-likelihood to zero only on
fixed small training sets of hundreds of examples, where
the network could overfit. This indicates that the network
has problems finding generally valid rules for assembly of
output tensors.
Williams, Ronald J. and Zipser, David. A learning algorithm
for continually running fully recurrent neural networks.
Neural Computation, 1(2):270–280, 1989.
Xu, Danfei, Zhu, Yuke, Choy, Christopher Bongsoo, and
Fei-Fei, Li. Scene graph generation by iterative message
passing. In CVPR, 2017.
Yu, Lantao, Zhang, Weinan, Wang, Jun, and Yu, Yong.
Seqgan: Sequence generative adversarial nets with policy
gradient. In AAAI, 2017.
Appendix
A. Max-Pooling Matching
In this section we briefly review max-pooling matching algorithm of Cho et al. (2014). In its relaxed form, a continuous
correspondence matrix X ∗ ∈ [0, 1]k×n between nodes of
e is determined based on similarities of node
graphs G and G
e represented as matrix elements
pairs i, j ∈ G and a, b ∈ G
+
Sia;jb ∈ R .
Let x∗ denote the column-wise replica of X ∗ . The relaxed
graph matching problem is expressed as quadratic
Pn programming task x∗ = arg maxx xT Sx such that i=1 xia ≤ 1,
Pk
kn
a=1 xia ≤ 1, and x ∈ [0, 1] . The optimization strategy
of choice is derived to be equivalent to the power method
with iterative update rule x(t+1) = Sx(t) /||Sx(t) ||2 . The
starting correspondences x(0) are initialized as uniform and
the rule is iterated until convergence; in our use case we run
for a fixed amount of iterations.
In the context of graph matching, the matrix-vector product
Sx can be interpreted as P
sum-pooling
P over match candidates: xia ← xia Sia;ia + j∈Ni b∈Na xjb Sia;jb , where
Ni and Na denote the set of neighbors of node i and a.
The authors argue that this formulation is strongly influenced by uninformative or irrelevant elements and propose a more robust max-pooling version, which considers only the best pairwise
similarity from each neighbor:
P
xia ← xia Sia;ia + j∈Ni maxb∈Na xjb Sia;jb .
B. Unregularized Autoencoder
The regularization in VAE works against achieving perfect
reconstruction of training data, especially for small embedding sizes. To understand the reconstruction ability of our
| 9cs.NE
|
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
arXiv:1802.04196v2 [quant-ph] 19 Feb 2018
MICHEL PLANAT†, RAYMOND ASCHHEIM‡,
MARCELO M. AMARAL‡ AND KLEE IRWIN‡
Abstract. A single qubit may be represented on the Bloch sphere or
similarly on the 3-sphere S 3 . Our goal is to dress this correspondence by
converting the language of universal quantum computing (uqc) to that
of 3-manifolds. A magic state and the Pauli group acting on it define a
model of uqc as a POVM that one recognizes to be a 3-manifold M 3 . E.
g., the d-dimensional POVMs defined from subgroups of finite index of
the modular group P SL(2, Z) correspond to d-fold M 3 - coverings over
the trefoil knot. In this paper, one also investigates quantum information
on a few ‘universal’ knots and links such as the figure-of-eight knot,
the Whitehead link and Borromean rings , making use of the catalog
of platonic manifolds available on SnapPy [4] . Further connections
between POVMs based uqc and M 3 ’s obtained from Dehn fillings are
explored.
PACS: 03.67.Lx, 03.65.Wj, 03.65.Aa, 02.20.-a, 02.10.Kn, 02.40.Pc, 02.40.Sf
MSC codes: 81P68, 81P50, 57M25, 57R65, 14H30, 20E05, 57M12
Keywords: quantum computation, IC-POVMs, knot theory, three-manifolds, branch
coverings, Dehn surgeries.
Manifolds are around us in many guises.
As observers in a three-dimensional world, we are most familiar with twomanifolds: the surface of a ball or a doughnut or a pretzel, the surface of a
house or a tree or a volleyball net...
Three-manifolds may be harder to understand at first. But as actors and
movers in a three-dimensional world, we can learn to imagine them as alternate universes. [1]
1. Introduction
Mathematical concepts pave the way for improvements in technology. As
far as topological quantum computation is concerned, non abelian anyons
have been proposed as an attractive (fault tolerant) alternative to standard
quantum computing which is based on a universal set of quantum gates
[5, 6, 7, 8]. Anyons are two-dimensional quasiparticles with world lines
forming braids in space-time. Whether non abelian anyons do exist in the
real world and/or would be easy to create artificially is still open to discussion. Three-dimensional topological quantum computing beyond anyons [9]
still is not well developed although, as it will be shown in this essay, it is
a straightforward consequence of a set of ideas belonging to standard universal quantum computation (uqc) and simultaneously to three-dimensional
topology. For quantum computation, we have in mind the concepts of magic
states and the related positive-operator valued measures (POVMs) that were
1
MICHEL PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
2
investigated in detail in [10, 11, 12]. For three-dimensional topology, the
starting point consists of the Poincaré and Thurston conjectures, now theorems [1]. Three-dimensional topological quantum computing would federate
the foundations of quantum mechanics and cosmology, a recurrent dream of
many physicists. Three-dimensional topology was already investigated by
several groups in the context of quantum information [13, 14], high energy
physics [15, 16], biology [17] and consciousness studies [18].
1.1. From Poincaré conjecture to uqc. Poincaré conjecture is the elementary (but deep) statement that every simply connected, closed 3-manifold
is homeomorphic to the 3-sphere S 3 [19]. Having in mind the correspondence
between S 3 and the Bloch sphere that houses the qubits ψ = a |0i + b |1i,
a, b ∈ C, |a|2 +|b|2 = 1, one would desire a quantum translation of this statement. For doing this, one may use the picture of the Riemann sphere C ∪ ∞
in parallel to that of the Bloch sphere and follow F. Klein lectures on the
icosahedron to perceive the platonic solids within the landscape [20]. This
picture fits well the Hopf fibrations [21], their entanglements described in
[22, 23] and quasicrystals [25, 24]. But we can be more ambitious and dress
S 3 in an alternative way that reproduces the historic thread of the proof
of Poincaré conjecture. Thurston’s geometrization conjecture, from which
Poincaré conjecture follows, dresses S 3 as a 3-manifold not homeomorphic
to S 3 . The wardrobe of 3-manifolds M 3 is huge but almost every dress is
hyperbolic and W. Thurston found the recipes for them [1]. Every dress is
identified thanks to a signature in terms of invariants. For our purpose, the
fundamental group π1 of M 3 does the job.
The three-dimensional space surrounding a knot K -the knot complement
S 3 \ K- is an example of a three-manifold [1, 26]. We will be especially
interested by the trefoil knot that underlies work of the first author [2]
as well as the figure-of-eight knot, the Whitehead link and the Borromean
rings because they are universal (in a sense described below), hyperbolic and
allows to build 3-manifolds from platonic manifolds [4]. Such manifolds carry
a quantum geometry corresponding to quantum computing and (possibly
informationally complete: IC) POVMs identified in our earlier work [2, 11,
12].
According to [3], the knot K and the fundamental group G = π1 (S 3 \ K)
are universal if every closed and oriented 3-manifold M 3 is homeomorphic
to a quotient H/G of the hyperbolic 3-space H by a subgroup H of finite
index d of G. The figure-of-eight knot and the Whitehead link are universal.
The catalog of the finite index subgroups of their fundamental group G and
of the corresponding 3-manifolds defined from the d-fold coverings [27] can
easily been established up to degree 8, using the software SnapPy [28].
In paper [2] of the first author, it has been found that d-dimensional
IC-POVMs may be built from finite index subgroups of the modular group
Γ = P SL(2, Z). To an IC is associated a subgroup of index d of Γ, a fundamental domain in the Poincaré upper-half plane and a signature in terms of
genus, elliptic points and cusps as summarized in [2, Fig. 1]. There exists a
relationship between the modular group Γ and the trefoil knot T1 since the
fundamental group π1 (S 3 \ T1 ) of the knot complement is the braid group
B3 , the central extension of Γ. But the trefoil knot and the corresponding
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
3
Figure
1. (a)
The
figure-of-eight
knot:
K4a1=otet0200001 = m004, (b) the Whitehead link
L5a1 =ooct0100001 = m129, (c) Borromean rings
L6a4 =ooct0200005 = t12067.
braid group B3 are not universal [29] which forbids the relation of the finite
index subgroups of B3 to all three-manifolds.
It is known that two coverings of a manifold M with fundamental group
G = π1 (M ) are equivalent if there exists a homeomorphism between them.
Besides, a d-fold covering is uniquely determined by a subgroup of index d
of the group G and the inequivalent d-fold coverings of M correspond to
conjugacy classes of subgroups of G [27]. In this paper we will fuse the
concepts of a three-manifold M 3 attached to a subgroup H of index d and
the POVM, possibly informationally complete (IC), found from H (thanks
to the appropriate magic state and related Pauli group factory).
Figure 2. (a) The trefoil knot T1 = K3a1 = 31 , (b) the
link L7n1 associated to the Hesse SIC, (c) the link L6a3
associated to the two-qubit IC.
1.2. Minimal informationally complete POVMs and uqc. In our approach [12, 2], minimal informationally complete (IC) POVMs are derived
MICHEL PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
4
from appropriate fiducial states under the action of the (generalized) Pauli
group. The fiducial states also allow to perform universal quantum computation [11].
A POVM is a collection of positive semi-definite operators {E1 , . . . , Em }
that sum to the identity. In the measurement of a state ρ, the i-th outcome
is obtained with a probability given by the Born rule p(i) = tr(ρEi ). For a
minimal IC-POVM, one needs d2 one-dimensional projectors Πi = |ψi i hψi |,
with Πi = dEi , such that the rank of the Gram matrix with elements
tr(Πi Πj ), is precisely d2 . A SIC-POVM (the S means symmetric) obeys
dδij +1
, that allows the explicit recovery
the relation|hψi |ψj i|2 = tr(Πi Πj ) = d+1
of the density matrix as in [30, eq. (29)].
New minimal IC-POVMs (i.e. whose rank of the Gram matrix is d2 ) and
with Hermitian angles |hψi |ψj i|i6=j ∈ A = {a1 , . . . , al } have been discovered
1
. The states en[2]. A SIC is equiangular with |A| = 1 and a1 = √d+1
countered are considered to live in a cyclotomic field F = Q[exp( 2iπ
n )], with
n = GCD(d, r), the greatest common divisor of d and r, for some r. The
1
Hermitian angle is defined as |hψi |ψj i|i6=j = k(ψi , ψj )k deg , where k.k means
the field norm of the pair (ψi , ψj ) in F and deg is the degree of the extension
F over the rational field Q [12].
The fiducial states for SIC-POVMs are quite difficult to derive and seem
to follow from algebraic number theory [31]. Except for d = 3, the ICPOVMs derived from permutation groups are not symmetric and most of
them can be recovered thanks to subgroups of index d of the modular group
Γ [2, Table 2]. For instance, for d = 4, the action of the two-qubit Pauli
group on the magic/fiducial state of type (0, 1, −ω, ω − 1) with ω = exp( iπ
3)
results into an IC whose geometry of triple products of projectors Πi , arising
from the the congruence subgroup Γ0 (3) of Γ, turns out to correspond to
the commutation graph of Pauli operators [2, Fig.2]. For d = 5, all five
congruence subgroups Γ′ , Γ(2), 3C 0 , Γ0 (4) or Γ0 (5) point out the geometry
of Borromean rings (see [2, Fig.2] and Table 1 below).
While Γ serves as a motivation for investigating the trefoil knot manifold in relation to uqc and the corresponding ICs, it is important to put
the uqc problem in the wider frame of Poincaré conjecture, the Thurston’s
geometrization conjecture and the related 3-manifolds. E.g. ICs may also
follow from hyperbolic or Seifert 3-manifolds as shown in tables 2 to 5 of
this paper.
1.3. Organization of the paper. The paper runs as follows. Sec. 2 deals
about the relationship between quantum information seen from the modular group Γ and from the trefoil knot 3-manifold. Sec. 3 deals about
the (platonic) 3-manifolds related to coverings over the figure-of-eight knot,
Whitehead link and Borromean rings, and how they relate to a few known
IC-POVMs. Sec. 4 describes the important role played by Dehn fillings
for describing the many types of 3-manifolds that may relate to topological
quantum computing.
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
5
2. Quantum information from the modular group Γ and the
related trefoil knot T1
In this section, we describe the results established in [2] in terms of the 3manifolds corresponding to coverings of the trefoil knot complement S 3 \ T1 .
d
2
3
.
4
.
.
5
.
6
.
.
.
.
.
.
.
7
.
.
8
.
.
.
ty
hom
cp gens CS link
type in [2]
1
cyc
1
2
-1/6
+
1
3
irr
1+1
2 2
1/4 L7n1 Γ0 (2), Hesse SIC
1
1
+
+
1
cyc
1
3
.
A4
2
2
irr
1+1
2 2
1/6 L6a3
Γ0 (3), 2QB IC
1
irr
1 3
.
4A0 , 2QB-IC
2 +1
1
cyc
1 2
.
S4
3 +1
cyc
1
1 2
5/6
A5
1
irr
1 3
.
5A0 , 5-dit IC
3 +1
reg
1+1+1
3 3
0 L6n1
Γ(2), 6-dit IC
cyc
1+1+1
1 3
. L6n1
Γ′ , 6-dit IC
irr
1+1+1
3 3
.
1
irr
2
3
.
3C 0 , 6-dit IC
+
1
+
1
2
1
irr
2 3
.
Γ0 (4), 6-dit IC
2 +1+1
1
irr
2 3
.
Γ0 (5), 6-dit IC
2 +1+1
irr 12 + 21 + 12 + 1 1 4
.
1
1
irr
1 3
.
3 + 3 +1
cyc
1
1 2
-5/6
irr
1+1
2 3
.
NC 7-dit IC
1
1
irr
1 4
.
7A0 7-dit IC
2 + 2 +1
irr
1+1
2 2
-1/6
1
cyc
2
2
.
+
1
3
1
cyc
2 3
.
3 +1+1
1
cyc
1
4
.
8A0 , ∼ 8-dit IC
+
1
6
Table 1. Coverings of degree d over the trefoil knot found
from SnapPy [28]. The related subgroup of modular group Γ
and the corresponding IC-POVM [2] (when applicable) is in
the right column. The covering is characterized by its type
ty, homology group hom (where 1 means Z), the number of
cusps cp, the number of generators gens of the fundamental
group, the Chern-Simons invariant CS and the type of link
it represents (as identified in SnapPy). The case of cyclic
coverings corresponds to Brieskorn 3-manifolds as explained
in the text: the spherical groups for these manifolds is given
at the right hand side column.
Let us introduce to the group representation of a knot complement π1 (S 3 \
K). A Wirtinger representation is a finite representation of π1 where the
relations are the form wgi w−1 = gj where w is a word in the k generators
{g1 , · · · , gk }. For the trefoil knot T1 = K3a1 = 31 sown in Fig. 2a, a
MICHEL PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
6
Wirtinger representation is [32]
π1 (S 3 \ T1 ) = hx, y|yxy = xyxi or equivalently π1 = x, y|y 2 = x3 .
In the rest of the paper, the number of d-fold coverings of the manifold
corresponding to the knot T will be displayed as the ordered list ηd (T ),
d ∈ {1..10 . . .}. For T1 it is
M3
ηd (T1 ) = {1, 1, 2, 3, 2, 8, 7, 10, 18, 28, . . .}.
Details about the corresponding d-fold coverings are in Table 1. As expected, the coverings correspond to subgroups of index d of the fundamental
group associated to the trefoil knot T1 .
Cyclic branched coverings over the trefoil knot. Let p, q, r be three positive
integers (with p ≤ q ≤ r), the Brieskorn 3-manifold Σ(p, q, r) is the intersection in C3 of the 5-sphere S 5 with the surface of equation z1p + z2q + z3r = 1.
In [34], it is shown that a r-fold cyclic covering over S 3 branched along
a torus knot or link of type (p, q) is a Brieskorn 3-manifold Σ(p, q, r) (see
also Sec 4.1). For the spherical case p−1 + q −1 + r −1 > 1, the group associated to a Brieskorn manifold is either dihedral [that is the group Dr
for the triples (2, 2, r)], tetrahedral [that is A4 for (2, 3, 3)], octahedral [that
is S4 for (2, 3, 4)] or icosahedral [that is A5 for (2, 3, 5)]. The Euclidean
case p−1 + q −1 + r −1 = 1 corresponds to (2, 3, 6), (2, 4, 4) or (3, 3, 3). The
remaining cases are hyperbolic.
The cyclic branched coverings with spherical groups for the trefoil knot
[which is of type (2, 3)] are identified in right hand side column of Table 1.
Irregular branched coverings over the trefoil knot. The right hand side column of Table 1 shows the subgroups of Γ identified in [2, Table 1] as corresponding to an IC-POVM. In particular, the Hesse SIC [already found
associated to the congruence subgroup Γ0 (2)] corresponds to the 7-crossing
link L7n1 while the two-qubit IC [already found associated to the congruence subgroup Γ0 (3)] corresponds to the 6- crossing link L6a3. The trefoil
knot and the former two links are pictured in Fig. 2.
Five coverings of degree 6 allow the construction of the (two-valued) 6-dit
IC-POVM whose geometry contain the picture of Borromean rings [2, Fig.
2c]. The corresponding congruence subgroups of Γ are identified in table 1.
The first two, viz Γ(2) and Γ′ , define 3-manifolds whose fundamental group
is the same than the one of the link L6n1, alias the Borromean rings with
surgeries of slope (0, 1) on two cusps (see Sect. 4 for more on this topic). The
other three coverings leading to the 6-dit IC are the congruence subgroups
3C 0 , Γ0 (4) and Γ0 (5).
3. Quantum information from universal knots and links
3.1. Three-manifolds pertaining to the figure-of-eight knot. The
fundamental group for the figure-of-eight knot K0 is
π1 (S 3 \ K0 ) = x, y|y ∗ x ∗ y −1 xy = xyx−1 yx .
and the number of d-fold coverings is in the list
ηd (K0 ) = {1, 1, 1, 2, 4, 11, 9, 10, 11, 38, . . .}.
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
7
Table 2 establishes the list of 3-manifolds corresponding to subgroups
of index d ≤ 7 of the universal group G = π1 (S 3 \ K0 ). The manifolds
are labeled otetNn in [4] because they are oriented and built from N = 2d
tetrahedra, with n an index in the table. The identification of 3-manifolds of
finite index subgroups of G was first obtained by comparing the cardinality
list ηd (H) of the corresponding subgroup H to that of a fundamental group
of a tetrahedral manifold in SnapPy table [28]. But of course there is more
straightforward way to perform this task by identifying a subgroup H to a
degree d covering of K0 [27]. The full list of d-branched coverings over the
figure eight knot up to degree 8 is available in SnapPy. Extra invariants
of the corresponding M 3 may be found there. In addition, the lattice of
branched coverings over K0 was investigated in [35].
ty
M3
cp
cyc
otet0400002 , m206
1
cyc
otet0600003 , s961
1
irr otet0800002 , L10n46, t12840 2
cyc
otet0800007 , t123839
1
5 cyc
otet1000019
1
irr
otet1000006 , L8a20
3
irr
otet1000026
2
6 cyc
otet1200013
1
irr
otet1200041
2
irr
otet1200039 , otet1200038
1
irr
otet1200017
2
irr
otet1200000
2
7 cyc
otet1400019
1
irr
otet1400002 , L14n55217
3
irr
otet1400035
1
d
2
3
4
rk
2
3
4
16
21
15,21
25
28
36
31
33
36
43
49
49
pp
comment
1
Mom-4s [36]
2-qubit IC
1
5-dit IC
2
6-dit IC
2
6-dit IC
2
2
7-dit IC
7-dit IC
Table 2. Table of 3-manifolds M 3 found from subgroups of
finite index d of the fundamental group π1 (S 3 \K0 ) (alias the
d-fold coverings of K0 ). The terminology in column 3 is that
of Snappy [28]. The identified M 3 is made of 2d tetrahedra
and has cp cusps. When the rank rk of the POVM Gram
matrix is d2 the corresponding IC-POVM shows pp distinct
values of pairwise products as shown.
Let us give more details about the results summarized in Table 2. Using Magma, the conjugacy class of subgroups of index 2 in the fundamental group G is represented by the subgroup on three generators and two
relations as follows H = x, y, z|y −1 zx−1 zy −1 x−2 , z −1 yxz −1 yz −1 xy , from
which the sequence of subgroups of finite index can be found as ηd (M 3 ) =
{1, 1, 5, 6, 8, 33, 21, 32, · · · }. The manifold M 3 corresponding to this sequence
is found in Snappy as otet0400002 , alias m206.
The conjugacy class of subgroups of index 3 in G is represented as
H = x, y, z|x−2 zx−1 yz 2 x−1 zy −1 , z −1 xz −2 xz −2 y −1 x−2 zy ,
MICHEL PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
8
with ηd (M 3 ) = {1, 7, 4, 47, 19, 66, 42, 484, · · · } corresponding to the manifold
otet0600003 , alias s961.
As shown in Table 2, there are two conjugacy classes of subgroups of
index 4 in G corresponding to tetrahedral manifolds otet0800002 (the permutation group P organizing the cosets is Z4 ) and otet0800007 (the permutation group organizing the cosets is the alternating group A4 ). The latter
group/manifold has fundamental group
Figure 3. Two platonic three-manifolds leading to the construction of the two-qubit IC-POVM. Details are given in
Tables 2 and 3.
H = x, y, z|yx−1 y −1 z −1 xy −2 xyzx−1 y, zx−1 yx−1 yx−1 zyx−1 y −1 z −1 xy −1 ,
with cardinality sequences of subgroups as ηd (M 3 ) = {1, 3, 8, 25, 36, 229, 435 · · · }.
To H is associated an IC-POVM [12, 2] which follows from the action of the
two-qubit Pauli group on a magic/fiducial state of type (0, 1, −ω6 , ω6 − 1),
with ω6 = exp(2iπ/6) a six-root of unity.
For index 5, there are three types of 3-manifolds corresponding to the
subgroups H. The tetrahedral manifold otet1000026 of sequence ηd (M 3 ) =
{1, 7, 15, 88, 123, 802, 1328 · · · }, is associated to a 5-dit equianguler IC-POVM,
as in [12, Table 5].
For index 6, the 11 coverings define six classes of 3-manifolds and two of
them: otet1200041 and otet1200000 are related to the construction of ICs. For
index 7, one finds three classes of 3-manifolds with two of them: otet1400002
(alias L14n55217) and otet1400035 are related to ICs. Finally, for index 7,
3 types of 3-manifolds exist, two of them relying on the construction of the
7-dit (two-valued) IC. For index 8, there exists 6 distinct 3-manifolds (not
shown) none of them leading to an IC.
A two-qubit tetrahedral manifold. The tetrahedral three-manifold otet0800007
is remarkable in the sense that it corresponds to the subgroup of index 4
of G that allows the construction of the two-qubit IC-POVM. The corresponding hyperbolic polyhedron taken from SnapPy is shown in Fig. 3a. Of
UNIVERSAL QUANTUM COMPUTING
d
2
3
4
5
6
ty
M3
cp
cyc ooct0200003 , t12066, L8n5 3
cyc
ooct0200018 , t12048
2
cyc
ooct0300011 , L10n100
4
cyc
ooct0300018
2
irr
ooct0300014 , L12n1741
3
irr
ooct0400058
4
irr
ooct0400061
3
irr
ooct0500092
3
irr
ooct0500285
2
irr
ooct0500098 , L13n11257
4
cyc
ooct0606328
5
irr
ooct0601972
3
irr
ooct0600471
4
AND THREE-MANIFOLDS
rk
2
2
3
3
9
16
16
25
25
25
36
36
36
pp
1
2
2
1
1
1
2
2
2
9
comment
Mom-4s[36]
Mom-4s[36]
qutrit Hesse SIC
2-qubit IC
2-qubit IC
5-dit IC
5-dit IC
5-dit IC
6-dit IC
6-dit IC
6-dit IC
Table 3. A few 3-manifolds M 3 found from subgroups of
the fundamental group associated to the Whitehead link. For
d ≥ 4, only the M 3 ’s leading to an IC are listed.
Figure 4. (a) The link L12n1741 associated to the qutrit
Hesse SIC, (b) The octahedral manifold ooct0300014 associated to the 2-qubit IC.
the 29 orientable tetrahedral manifolds with at most 8 tetrahedra, 20 are
two-colorable and each of those has at most 2 cusps. The 4 three-manifolds
(with at most 8 tetrahedra) identified in Table 2 belong to the 20’s and
the two-qubit tetrahedral manifold otet0800007 is one with just one cusp [37,
Table 1].
3.2. Three-manifolds pertaining to the Whitehead link. One could
also identify the 3-manifold substructure of another universal object, viz the
Whitehead link L0 [33].
MICHEL
10
PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
The cardinality list corresponding to the Whitehead link group π1 (L0 ) is
ηd (L0 ) = {1, 3, 6, 17, 22, 79, 94, 412, 616, 1659 . . .},
Table 3 shows that the identified 3-manifolds for index d subgroups of π1 (L0 )
are aggregates of d octahedra. In particular, one finds that the qutrit Hesse
SIC can be built from ooct0300014 and that the two-qubit IC-POVM may be
buid from ooct0400058 . The hyperbolic polyhedron for the latter octahedral
manifold taken from SnapPy is shown on Fig. 3b. The former octahedral
manifold follows from the link L12n1741 shown in Fig. 4a and the corresponding polyhedron taken from SnapPy is shown in Fig. 4b.
d
2
.
.
3
.
.
.
.
.
ty
hom
1
1
cyc
+
+
1+1+1
2
2
1
.
+
1
+
1
+1+1
2
.
1+1+1+1+1
1
1
cyc
3 + 3 +1+1+1
1
.
3 +1+1+1+1+1
.
1+1+1+1+1+1+1+1
irr
1+1+1+1
.
1+1+1+1+1+1
.
1+1+1+1+1+1
cp
3
4
5
3
5
7
4
4
5
M3
comment
ooct0400259
ooct0400055
ooct0400048 , L12n2226
ooct0607427 Hesse SIC
ooct0600463 Hesse SIC
ooct0600411
ooct0600466 Hesse SIC
ooct0600398
ooct0600407 , L14n63856
Table 4. Coverings of degrees 2 and 3 over the M 3 branched
along the Borromean rings. The identification of the corresponding hyperbolic 3-manifold is at the 5th column. It is
seen at the right hand side column that only three types of
3-manifolds allow to build the Hesse SIC.
3.3. A few three-manifolds pertaining to Borromean rings. Three-manifolds
corresponding to coverings of degree 2 and 3 of the 3-manifold branched along the
Borromean rings L6a4 [that is a not a (3,3)-torus link but an hyperbolic link] (see
Fig. 1c) are given in Table 4. The identified manifolds are hyperbolic octahedral
manifolds of volume 14.655 (for the degree 2) and 21.983 (for the degree 3).
4. A few Dehn fillings and their POVMs
To summarize our findings of the previous section, we started from a building
block, a knot (viz the trefoil knot T1 ) or a link (viz the figure-of-eight knot K0 )
whose complement in S 3 is a 3-manifold M 3 . Then a d-fold covering of M 3 was
used to build a d-dimensional POVM, possibly an IC. Now we apply a kind of
‘phase surgery’ on the knot or link that transforms M 3 and the related coverings
while preserving some of the POVMs in a way to be determined. We will start
with our friend T1 and arrive at a few standard 3-manifolds of historic importance,
the Poincaré homology sphere [alias the Brieskorn sphere Σ(2, 3, 5)], the Brieskorn
sphere Σ(2, 3, 7) and a Seifert fibered toroidal manifold Σ′ . Then we introduce
the 3-manifold ΣY resulting from 0-surgery on the figure-of-eight knot [42]. Later
in this section, we will show how to use the {3, 5, 3} Coxeter lattice and surgery
to arrive at a hyperbolic 3-manifold Σ120e of maximal symmetry whose several
coverings (and related POVMs) are close to the ones of the trefoil knot [38].
Let us start with a Lens space L(p, q) that is 3-manifold obtained by gluing the
boundaries of two solid tori together, so that the meridian of the first solid torus
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
11
T
Name
ηd (T )
T1
trefoil
{1,1,2,3,2, 8,7,10,10,28, 27,88,134,171,354}
T1 (−1, 1)
Σ(2, 3, 5)
{1,0,0,0,1, 1,0,0,0,1, 0,1,0,0,1}
T1 (1, 1)
Σ(2, 3, 7)
{1,0,0,0,0, 0,2,1,1,0, 0,0,0,9,3}
T1 (0, 1)
Σ′
{1,1,2,2,1, 5,3,2,4,1, 1,12,3,3,4}
K0 (0, 1)
ΣY
{1,1,1,2,2, 5,1,2,2,4, 3,17,1,1,2}
v2413 (−3, 2)
Σ120e
{1,1,1,4,1, 7,2,25,3,10 10,62,1,30,23}
Table 5. A few surgeries (column 1), their name (column 2)
and the cardinality list of d-coverings (alias conjugacy classes
of subgroups). Plain characters are used to point out the
possible construction of an IC-POVM in at least one the
corresponding three-manifolds (see [2] and Sec. 2 for the ICs
corresponding to T1 ).
goes to a (p, q)-curve on the second solid torus [where a (p, q)-curve wraps around
the longitude p times and around the meridian q times]. Then we generalize this
concept to a knot exterior, i.e. the complement of an open solid torus knotted like
the knot. One glues a solid torus so that its meridian curve goes to a (p, q)-curve
on the torus boundary of the knot exterior, an operation called Dehn surgery [1,
p. 275],[26, p 259],[39]. According to Lickorish’s theorem, every closed, orientable,
connected 3-manifold is obtained by performing Dehn surgery on a link in the
3-sphere.
4.1. A few surgeries on the trefoil knot.
The Poincaré homology sphere. The Poincaré dodecahedral space (alias the Poincaré
homology sphere) was the first example of a 3-manifold not the 3-sphere. It can be
obtained from (−1, 1) surgery on the trefoil knot T1 [40].
Let p, q, r be three positive integers and mutually coprime, the Brieskorn sphere
Σ(p, q, r) is the intersection in C3 of the 5-sphere S 5 with the surface of equation
z1p + z2q + z3r = 1. The homology of a Brieskorn sphere is that of the sphere S 3 .
A Brieskorn sphere is homeomorphic but not diffeomorphic to S 3 . The sphere
Σ(2, 3, 5) may be identified to the Poincaré homology sphere. The sphere Σ(2, 3, 7)
[42] may be obtained from (1, 1) surgery on T1 . Table 5 provides the sequences ηd
for the corresponding surgeries (±1, 1) on T1 . Plain digits in these sequences point
out the possibility of building ICs of the corresponding degree. This corresponds
to a considerable filtering of the ICs coming from T1 .
For instance, the smallest IC from Σ(2, 3, 5) has dimension five and is precisely
the one coming from the congruence subgroup 5A0 in Table 1. But it is built from a
non modular (fundamental) group whose permutation representation of the cosets
is the alternating group A5 ∼
= h(1, 2, 3, 4, 5), (2, 4, 3)i (compare [12, Sec. 3.3]).
The smallest dimensional IC derived from Σ(2, 3, 7) is 7-dimensional and twovalued, the same than the one arising from the congruence subgroup 7A0 given in
Table 1. But it arises from a non modular (fundamental) group with the permutation representation of cosets as P SL(2, 7) ∼
= h(1, 2, 4, 6, 7, 5, 3), (2, 5, 3)(4, 6, 7)i.
4.2. The Seifert fibered toroidal manifold Σ′ . An hyperbolic knot (or link)
in S 3 is one whose complement is 3-manifold M 3 endowed with a complete Riemannian metric of constant negative curvature, i.e. it has a hyperbolic geometry
and finite volume. A Dehn surgery on a hyperbolic knot is exceptional if it is reducible, toroidal or Seifert fibered (comprising a closed 3-manifold together with a
MICHEL
12
PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
decomposition into a disjoint union of circles called fibers). All other surgeries are
hyperbolic. These categories are exclusive for a hyperbolic knot. In contrast, a non
hyperbolic knot such as the trefoil knot admits a toroidal Seifert fiber surgery Σ′
obtained by (0, 1) Dehn filling on T1 [41].
The smallest dimensional ICs built from Σ′ are the Hesse SIC that is obtained
from the congruence subgroup Γ0 (2) (as for the trefoil knot) and the two-qubit IC
that comes from a non modular fundamental group [with cosets organized as the
alternating group A4 ∼
= h(2, 4, 3), (1, 2, 3)i].
4.3. Akbulut’s manifold ΣY . Exceptional Dehn surgery at slope (0, 1) on the
figure-of-eight knot K0 leads to a remarkable manifold ΣY found in [42] in the
context of 3-dimensional integral homology spheres smoothly bounding integral
homology balls. Apart from its topological importance, we find that some of its
coverings are associated to already discovered ICs and those coverings have the
same fundamental group π1 (ΣY ).
The smallest IC-related covering (of degree 4) occurs with integral homology Z
and the congruence subgroup Γ0 (3) also found from the trefoil knot (see Table 1).
Next, the covering of degree 6 and homology Z5 + Z leads to the 6-dit IC of type 3C 0
(also found from the trefoil knot). The next case corresponds to the (non-modular)
11-dimensional (3-valued) IC.
4.4. The hyperbolic manifold Σ120e . The hyperbolic manifold closest to the
trefoil knot manifold known to us was found in [38]. The goal in [38] is the search
of -maximally symmetric- fundamental groups of 3-manifolds. In two dimensions,
maximal symmetry groups are called Hurwitz groups and arise as quotients of
the (2, 3, 7)-triangle groups. In three dimensions, the quotients of the minimal
co-volume lattice Γmin of hyperbolic isometries, and of its orientation preserving
subgroup Γ+
min , play the role of Hurwitz groups. Let C be the {3, 5, 3} Coxeter
group, Γmin the split extension C ⋊ Z2 and Γ+
min one of the index two subgroups
of Γmin of presentation
3 2 2
2
2
5
.
Γ+
min = x, y, z|x , y , z , (xyz) , (xzyz) , (xy)
According to [38, Corollary 5], all torsion-free subgroups of finite index in Γ+
min
have index divisible by 60. There are two of them of index 60, called Σ60a and
Σ60b , obtained as fundamental groups of surgeries m017(−3, 2) and m016(−3, 2).
Torsion-free subgroups of index 120 in Γ+
min are given in Table 6. It is remarkable
that these groups are fundamental groups of oriented three-manifolds built with a
single icosahedron except for Σ120e .
manifold T
oicocld0100001 = s897(−3, 2)
oicocld0100000 = s900(−3, 2)
oicocld0100003 = v2051(−3, 2)
oicocld0100002 = s890(3, 2)
v2413(−3, 2) 6= oicocld0100004
oicocld0500005 = v3215(1, 2)
subgroup
ηd (T )
Σ120a
{1,0,0,0,0, 8,2,1,1,8}
Σ120b
{1,0,0,0,5, 8,10,15,5,24}
Σ120c
{1,0,0,0,0, 4,8,12,6,6}
Σ120d
{1,0,1,5,0, 9,0,35,9,2}
Σ120e
{1,1,1,4,1, 7,2,25,3,10}
Σ120f
{1,0,0,0,0, 14,10,5,10,17}
Table 6. The index 120 torsion free subgroups of Γ+
min and
their relation to the single isosahedron 3-manifolds [38]. The
icosahedral symmetry is broken for Σ120e (see the text for
details).
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
13
Σ120e is also special in the sense that many small dimensional ICs may be built
from it in contrast to the other groups in Table 6. The smallest ICs that may be
build from Σ120e are the Hesse SIC coming from the congruence subgroup Γ0 (2), the
two-qubit IC coming the congruence subgroup 4A0 and the 6-dit ICs coming from
the congruence subgroups Γ(2), 3C 0 or Γ0 (4) (see [2, Sec. 3 and Table 1]). Higher
dimensional ICs found from Σ120e does not come from congruence subgroups.
5. Conclusion
The relationship between 3-manifolds and universality in quantum computing
has been explored in this work. Earlier work of the first author already pointed out
the importance of hyperbolic geometry and the modular group Γ for deriving the
basic small dimensional IC-POVMs. In Sec. 2, the move from Γ to the trefoil knot
T1 (and the braid group B3 ) to non hyperbolic 3-manifolds could be investigated by
making use of the d-fold coverings of T1 that correspond to d-dimensional POVMs
(some of them being IC). Then, in Sec. 3, we passed to universal links (such as
the figure-of-eight knot, Whitehead link and Borromean rings) and the related hyperbolic platonic manifolds as new models for quantum computing based POVMs.
Finally, in Sec. 4, Dehn fillings on T1 were used to explore the connection of quantum computing to important exotic 3-manifolds [i.e. Σ(2, 3, 5) and Σ(2, 3, 7)], to
the toroidal Seifert fibered Σ′ , to Akbulut’s manifold ΣY and to a maximum symmetry hyperbolic manifold Σ120e slightly breaking the icosahedral symmetry. It is
expected that our work will have importance for new ways of implementing quantum computing and for the understanding of the link between quantum information
theory and cosmology [43, 44, 45].
Funding
The first author acknowledges the support by the French “Investissements d’Avenir”
program, project ISITE-BFC (contract ANR-15-IDEX-03). The other ressources
came from Quantum Gravity Research.
References
[1] W. P. Thurston, Three-dimensional geometry and topology (vol. 1), (Princeton University Press, Princeton, 1997).
[2] M. Planat, The Poincaré half-plane for informationally complete POVMs, Entropy
20 16 (2018).
[3] H. M. Hilden, M. T. Lozano, J. M. Montesinos and W. C. Whitten, On universal
groups and three-manifolds, Inventiones Mathematicae 87 441-45 (1987).
[4] E. Fominikh, S. Garoufalidis, M. Goerner, V. Tarkaev and A. Vesnin, A census of
tethahedral hyperbolic manifolds, Exp. Math. 25 466-481 (2016).
[5] A. Yu Kitaev, Fault-tolerant quantum computation by anyons, Annals Phys. 303
2-30 (2003).
[6] C. Nayak, S. H. Simon, A. Stern, M. Freedman and S. Das Sarma, Non-Abelian
anyons and topological quantum computation, Rev. Mod. Phys. 80 1083 (2008).
[7] Z. Wang, Topological Quantum Computation, (American Mathematical Soc., Rhode
Island, Number 112, 2010).
[8] J K. Pachos, Introduction to Topological Quantum Computation (Cambridge University Press, Cambridge, 2012).
[9] S. Vijay and L. Fu, A generalization of non-abelian anyons in three dimensions, arXiv
1706.07070 (2017).
[10] S. Bravyi and A. Kitaev, Universal quantum computation with ideal Clifford gates
and noisy ancillas, Phys. Rev. A71 022316 (2005).
[11] M. Planat and R. Ul Haq, The magic of universal quantum computing with permutations, Advances in mathematical physics 217, ID 5287862 (2017) 9 pp.
MICHEL
14
PLANAT†, RAYMOND ASCHHEIM‡, MARCELO M. AMARAL‡ AND KLEE IRWIN‡
[12] M. Planat and Z. Gedik, Magic informationally complete POVMs with permutations,
R. Soc. open sci. 4 170387 (2017).
[13] L. H. Kauffman and R. L. Baadhio, Quantum topology (Series on knots and everything), (World Scientific, 1993).
[14] L. H. Kauffman, Knot logic and topological quantum computing with Majorana
fermions, in Linear and algebraic structures in quantum computing, (J. Chubb, A.
Eskandarian and V. Harizanov eds, Lecture notes in logic 45, Cambridge Univ. Press,
2016).
[15] N. Seiberg, T. Senthil, C. Wang and E. Witten, A duality web in 2 + 1 dimensions
and condensed matter physics, Ann. Phys. 374 395-433 (2016).
[16] D. Gang, Y. Tachikawa and K. Yonekura, Smallest 3d hyperbolic manifolds via simple
3d theories, Phys. Rev. D 96 061701(R) (2017).
[17] N. C. Lim and S. E. Jackson, Molecular knots in biology and chemistry, J. Phys.
Condens. Matter 27 354101 (2015).
[18] K. Irwin, Toward a unification of physics and number theory,
https://www.researchgate.net/publication/314209738
Toward the Unification of Physics and Number Theory
[19] J. Milnor, The Poincaré conjecture 99 years later: a progress report (The Clay
Mathematics Institute 2002 Annual Report, 2003), http://www.math.sunysb.edu/∼
jack/PREPRINTS/poiproof.pdf, Retrieved 2018-01-01.
[20] M. Planat, On the geometry and invariants of qubits, quartits and octits, Int. J.
Geom. Methods in Mod. Phys. 8 303-313 (2011).
[21] N. S. Manton, Connections on discrete fiber bundles, Commun. Math. Phys. 113
341-351 (1987).
[22] R. Mosseri and R. Dandoloff, Geometry of entangled states, Bloch spheres and Hopf
fibrations, it J. Phys. A: Math. Gen. 34 10243 (2001).
[23] J. A. Nieto, Division-Algebras/Poincare-Conjecture Correspondence, J. Mod. Phys.
(Scientific Research) 4 32-36 (2013).
[24] A. Sen, R. Aschheim and K. Irwin, Emergence of an aperiodic Dirichlet space from
the tetrahedral units of an icosahedral internal space Mathematics 5 29 (2017).
[25] F. Fang, D. Hammock and K. Irwin, Methods for calculating empires in quasicrystals,
Crystals 7 304 (1997).
[26] C. C. Adams, The knot book, An elementary introduction to the mathematical theory
of knots (W. H. Freeman and Co, New York, 1994).
[27] A. D. Mednykh, A new method for counting coverings over manifold with fintely
generated fundamental group, Dokl. Math. 74 498-502 (2006).
[28] M. Culler, N. M. Dunfield, M. Goerner, and J. R. Weeks, SnapPy, a
computer program for studying the geometry and topology of 3-manifolds,
http://snappy.computop.org.
[29] H. M. Hilden, M. T. Lozano and J. M. Montesinoos, On knots that are universal,
Topology 24 499-504 (1985).
[30] Chris A. Fuchs, On the quantumness of a Hibert space, Quant. Inf. Comp. 4 467-478
(2004).
[31] M. Appleby, T. Y. Chien, S. Flammia and S. Waldron, Constructing exact symmetric informationally complete measurements from numerical solutions, Preprint
1703.05981 [quant-ph].
[32] D. Rolfsen, Knots and links (Mathematics Lecture Series 7, Houston, 1990).
[33] D. Gabai, The Whitehead manifold is a union of two Euclidean spaces, J. Topol. 4
529-534 (2011).
[34] J. Milnor, On the 3-dimensional Brieskorn manifolds M (p, q, r), in Knots, groups and
3-manifolds (ed. L. P. Neuwirth), Annals of Math. Study 84, Princeton Univ. Press,
Princeton, 175-225 (1975).
[35] J. Hempel, The lattice of branched covers over the Figure-eight knot, Topol. Appl.
34 183-201 (1990).
[36] R. C. Haraway, Determining hyperbolicity of compact orientable 3-manifolds with
torus boundary, Arxiv 1410.7115 [math.GT].
UNIVERSAL QUANTUM COMPUTING
AND THREE-MANIFOLDS
15
[37] S. A. Ballas, J. Danciger and G. S. Lee, Convex projective structures on nonhyperbolic three-manifolds, Arxiv 1508.04794 [math.GT].
[38] M. Conder, G. Martin and A. Torstensson, Maximal symmetry groups of hyperbolic
3-manifolds New Zealand J. Math. 35 3762 (2006).
[39] C. M. Gordon, Dehn filling: a survey, Knot Theory, Banach Center Publ. 42 129-144,
Polish Acad. Sci., Warsaw (1998).
[40] R. C. Kirby and M. G. Scharlemann, Eight faces of the Poincaré homology 3-sphere,
in Geometric Topology (Acad. Press, New York, 1979) pp 113-146.
[41] Ying-Qing Wu, Seifert fibered surgery on Montesinos knots, Arxiv 1207.0154
[math.GT], to appear in Comm. Anal. Geom.
[42] S. Akbulut and K. Larson, Brieskorn spheres bounding rational balls, ArXiv
1704.07739 [math.GT].
[43] K. T. Chan, H. Zainuddin, K. A. M. Atan and A. A. Siddig, Computing Quantum
Bound States on Triply Punctured Two-Sphere Surface, Chin. Phys. Lett. 33 090301
(2016).
[44] R. Aurich, F. Steiner and H. Then, Numerical computation of Maass waveforms and
an application to cosmology, in Hyperbolic Geometry and Applications in Quantum
Chaos and Cosmology (Jens Bolte and Frank Steiner eds, Cambridge Univ. Press,
2012), Preprint gr-qc/0404020.
[45] T. Asselmeyer-Maluga, Smooth quantum gravity: Exotic smoothness and Quantum
gravity, in At the Frontier of Spacetime Scalar-Tensor Theory, Bells Inequality, Machs
Principle, Exotic Smoothness (Fundamental Theories of Physics book series (FTP),
T. Asselmeyer-Maluga ed. (2016), pp 247-308).
† Université de Bourgogne/Franche-Comté, Institut FEMTO-ST CNRS UMR
6174, 15 B Avenue des Montboucons, F-25044 Besançon, France.
E-mail address: [email protected]
‡ Quantum Gravity Research, Los Angeles, CA 90290, USA
E-mail address: [email protected]
E-mail address: [email protected]
E-mail address: [email protected]
| 4math.GR
|
arXiv:1405.5245v1 [math.NA] 20 May 2014
An accurate and efficient numerical framework
for adaptive numerical weather prediction
Giovanni Tumolo(1), Luca Bonaventura(2)
January 16, 2018
(1)
Earth System Physics Section
The Abdus Salam International Center for Theoretical Physics
Strada Costiera 11, 34151 Trieste, Italy
[email protected]
(2)
MOX – Modelling and Scientific Computing,
Dipartimento di Matematica “F. Brioschi”, Politecnico di Milano
Via Bonardi 9, 20133 Milano, Italy
[email protected]
Keywords: Discontinuous Galerkin methods, adaptive finite elements, semiimplicit discretizations, semi-Lagrangian discretizations, shallow water equations, Euler equations.
AMS Subject Classification: 35L02, 65M60, 65M25, 76U05, 86A10
1
Abstract
We present an accurate and efficient discretization approach for
the adaptive discretization of typical model equations employed in
numerical weather prediction. A semi-Lagrangian approach is combined with the TR-BDF2 semi-implicit time discretization method
and with a spatial discretization based on adaptive discontinuous finite elements. The resulting method has full second order accuracy
in time and can employ polynomial bases of arbitrarily high degree in
space, is unconditionally stable and can effectively adapt the number
of degrees of freedom employed in each element, in order to balance accuracy and computational cost. The p−adaptivity approach employed
does not require remeshing, therefore it is especially suitable for applications, such as numerical weather prediction, in which a large number
of physical quantities are associated with a given mesh. Furthermore,
although the proposed method can be implemented on arbitrary unstructured and nonconforming meshes, even its application on simple
Cartesian meshes in spherical coordinates can cure effectively the pole
problem by reducing the polynomial degree used in the polar elements.
Numerical simulations of classical benchmarks for the shallow water
and for the fully compressible Euler equations validate the method
and demonstrate its capability to achieve accurate results also at large
Courant numbers, with time steps up to 100 times larger than those
of typical explicit discretizations of the same problems, while reducing
the computational cost thanks to the adaptivity algorithm.
2
1
Introduction
The Discontinuous Galerkin (DG) spatial discretization approach
is currently being employed by an increasing number of environmental
fluid dynamics models see e.g. [12], [17], [28],[39],[19],[25] and a more
complete overview in [5]. This is motivated by the many attractive
features of DG discretizations, such as high order accuracy, local mass
conservation and ease of massively parallel implementation.
On the other hand, DG methods imply severe stability restrictions when coupled with explicit time discretizations. One traditional
approach to overcome stability restrictions in low Mach number problems is the combination of semi - implicit (SI) and semi - Lagrangian (SL) techniques. In a series of papers [41], [42], [19], [14], [52]
it has been shown that most of the computational gains traditionally achieved in finite difference models by the application of SI, SL
and SISL discretization methods are also attainable in the framework
of DG approaches. In particular, in [52] we have introduced a dynamically p−adaptive SISL-DG discretization approach for low Mach
number problems, that is quite effective in achieving high order spatial
accuracy, while reducing substantially the computational cost.
In this paper, we apply the technique of [52] to the shallow water equations in spherical geometry and to the the fully compressible
Euler equations, in order to show its effectiveness for model problems typical of global and regional weather forecasting. The advective
form of the equations of motion is employed and the semi-implicit
time discretization is based on the TR-BDF2 method, see e.g. [21],
[32]. This combination of two robust ODE solvers yields a second
order accurate, A-stable and L-stable method (see e.g. [27]), that is
effective in damping selectively high frequency modes. At the same
time, it achieves full second order accuracy, while the off-centering in
the trapezoidal rule, typically necessary for realistic applications to
nonlinear problems (see e.g. [7], [11], [52]), limits the accuracy in time
to first order. Numerical results presented in this paper show that the
total computational cost of one TR-BDF2 step is analogous to that of
one step of the off-centered trapezoidal rule, as well as the structure
of the linear problems to be solved at each time step, thus allowing to
extend naturally to this more accurate method any implementation
based on the off-centered trapezoidal rule. Numerical simulations of
the shallow water benchmarks proposed in [55], [29], [22] and of the
non-hydrostatic benchmarks proposed in [47], [23] have been employed
to validate the method and to demonstrate its capabilities. In particular, it will be shown that the present approach enables the use of
time steps even 100 times larger than those allowed for DG models by
3
standard explicit schemes, see e.g. the results in [38].
The method presented in this paper, just as its previous version in
[52], can be applied in principle on arbitrarily unstructured and even
nonconforming meshes. For example, a model based on this method
could run on a non conforming mesh of rectangular elements built
around the nodes of a reduced Gaussian grid [20]. For simplicity,
however, no such implementation has been developed so far. Here,
only a simple Cartesian mesh has been used. If no degree adaptivity
is employed, this results in very high Courant numbers in the polar regions. These do not result in any special stability problems for
the present SISL discretization approach, as it will be shown by the
numerical results reported below. On the other hand, even with an
implementation based on a simple Cartesian mesh in spherical coordinates, the flexibility of the DG space discretization allows to reduce
the degree of the basis and test functions employed close to the poles,
thus making the effective model resolution more uniform and solving
the efficiency issues related to the pole problem by static p−adaptivity.
This is especially advantageous because the conditioning of the linear
system to be solved at each time step is greatly improved and, as a
consequence, the number of iterations necessary for the linear solver
is reduced by approximately 80%, while at the same time no spurious
reflections nor artificial error increases are observed.
Beyond these computational advantages, we believe that the present
approach based on p−adaptivity is especially suitable for applications to numerical weather prediction, in contrast to h−adaptivity
approaches (that is, local mesh coarsening or refinement in which the
size of some elements changes in time). Indeed, in numerical weather
prediction, information that is necessary to carry out realistic simulations (such as orography profiles, data on land use and soil type,
land-sea masks) needs to be reconstructed on the computational mesh
and has to be re-interpolated each time that the mesh is changed. Furthermore, many physical parameterizations are highly sensitive to the
mesh size. Although devising better parameterizations that require
less mesh-dependent tuning is an important research goal, more conventional parameterizations will still be in use for quite some time.
As a consequence, it is useful to improve the accuracy locally by
adding supplementary degrees of freedom where necessary, as done
in a p−adaptive framework, without having to change the underlying
computational mesh. In conclusion, the resulting modeling framework
seems to be able to combine the efficiency and high order accuracy of
traditional SISL pseudo-spectral methods with the locality and flexibility of more standard DG approaches.
In section 2, two examples of governing equations are introduced.
4
In section 3, the TR-BDF2 method is reviewed. In section 4 the
approach employed for the advection of vector fields in spherical geometry is described in detail. In section 5, we introduce the SISL-DG
discretization approach for the shallow water equations in spherical geometry. In section 6, we outline its extension to the fully compressible
Euler equations in a vertical plane. Numerical results are presented
in section 7, while in section 8 we try to draw some conclusions and
outline the path towards application of the concepts introduced here
in the context of a non hydrostatic dynamical core.
2
Governing equations
We consider as a basic model problem the two-dimensional shallow
water equations on a rotating sphere (see e.g. [15]). These equations
are a standard test bed for numerical methods to be applied to the
full equations of motion of atmospheric or oceanic circulation models,
see e.g. [55]. Among their possible solutions, they admit Rossby
and inertial gravity waves, as well as the response of such waves to
orographic forcing. We will use the advective, vector form of the
shallow water equations:
Dh
= −h∇ · u,
Dt
Du
= −g∇h − f k̂ × u − g∇b.
Dt
(1)
(2)
Here h represents the fluid depth, b the bathymetry elevation, f the
Coriolis parameter, k̂ the unit vector locally normal to the Earth’s surface and g the gravity force per unit mass on the Earth’s surface. Assuming that x, y are orthogonal curvilinear coordinates on the sphere
(or on a portion of it), we denote by mx and my the components
of the (diagonal) metric tensor. Furthermore, we set u = (u, v)T ,
where u and v are the contravariant components of the velocity vector in the coordinate direction x and y respectively, multiplied by the
D
the
corresponding metric tensor components. We also denote by Dt
Lagrangian derivative
D
∂
u ∂
v ∂
=
+
+
,
Dt
∂t mx ∂x my ∂y
Dy
so that u = mx Dx
Dt , v = my Dt . In particular, in this paper standard
spherical coordinates will be employed.
As an example of a more complete model, we will also consider the
fully compressible, non hydrostatic equations of motion. Following
5
e.g. [10], [4], [11], they can be written as
cp
DΠ
=−
− 1 Π∇ · u,
Dt
cv
Du
= −cp Θ∇Π − gk̂,
Dt
DΘ
= 0.
Dt
−R/cp
is the powhere, being p0 a reference pressure value, Θ = T pp0
p R/cp
tential temperature, Π = p0
is the Exner pressure, while cp , cv , R
are the constant pressure and constant volume specific heats and the
gas constant of dry air respectively. Here the Coriolis force is omitted
for simplicity. Notice also that, by a slight abuse of notation, in the
three-dimensional case u = (u, v, w)T denotes the three dimensional
D
, ∇ operators are also three-dimensional, while
velocity field and the Dt
we will assume u = (u, w)T in the description of (x, z) two dimensional, vertical slice models. It is customary to rewrite such equations
in terms of perturbations with respect to a steady hydrostatic reference profile, so that assuming Π(x, y, z, t) = π ∗ (z) + π(x, y, z, t),
∗
Θ(x, y, z, t) = θ ∗ (z) + θ(x, y, z, t) with cp θ ∗ dπ
dz = −g, one obtains for
a vertical plane
cp
DΠ
=−
− 1 Π∇ · u,
(3)
Dt
cv
Du
∂π
= −cp Θ ,
(4)
Dt
∂x
∂π
θ
Dw
= −cp Θ
+ g ∗,
(5)
Dt
∂z
θ
dθ ∗
Dθ
=−
w.
(6)
Dt
dz
It can be observed that equations (3)-(6) are isomorphic to equations
(1)-(2), which will allow to extend almost automatically the discretization approach proposed for the former to the more general model.
3
Review of the TR-BDF2 method
We review here some properties of the so called TR-BDF2 method,
which was first introduced in [2]. Given a Cauchy problem
y′ = f (y, t)
y(0) = y0
6
(7)
and considering a time discretization employing a constant time step
∆t, the TR-BDF2 method is defined by the two following implicit
stages:
un+2γ − γ∆tf (un+2γ , tn + 2γ∆t) = un + γ∆tf (un , tn ),
un+1 − γ2 ∆tf (un+1 , tn+1 ) = (1 − γ3 )un + γ3 un+2γ . (8)
Here γ ∈ [0, 1/2] is an implicitness parameter and
γ2 =
1 − 2γ
,
2(1 − γ)
γ3 =
1 − γ2
.
2γ
It is immediate that the first stage of (8) is simply the application
of the trapezoidal rule (or Crank-Nicolson method) over the interval
[tn , tn + 2γ∆t]. It could also be substituted by an off centered CrankNicolson step without reducing the overall accuracy of the method.
The outcome of this stage is then used to turn the two step BDF2
method into a single step, two stages method. This combination of two
robust stiff solvers yields a method with several interesting accuracy
and stability properties, that were analyzed in detail in [21]. As shown
in this paper, this analysis is most easily carried out by rewriting the
method as
k1 = f (un , tn )
k2 = f (un + γ∆tk1 + γ∆tk2 , tn + γ∆t)
1−γ
1−γ
∆tk1 +
∆tk2 + γ∆tk3 , tn+1
k3 = f un +
2
2
1−γ
1−γ
n+1
n
u
= u + ∆t
(9)
k1 +
k2 + γk3 .
2
2
In this formulation, the TR-BDF2 method is clearly a Singly Diagonal
Implicit Runge Kutta (SDIRK) method, so that one can rely on the
theory for this class of methods to derive stability and accuracy results
(see e.g. [27]). Notice that the same method has been rediscovered in
[6] and has been analyzed and applied also in [18], to treat the implicit
terms in the framework of an Additive Runge Kutta approach (see
e.g. [26]). As shown in [21], the TR-BDF2 method is second order
accurate and A-stable for any value of γ. Written as in (9), the method
can also be proven to constitute a (2,3) embedded Runge-Kutta pair,
with companion coefficients given by
√
√
√
2− 2
2
2
)/3, (1 + 3
)/3,
,
(1 −
4
4
6
7
provided that no off centering is employed in the first stage of (8).
This equips the method with an extremely efficient√estimator of the
time discretization error. Furthermore, for γ = 1 − 2/2 it is also Lstable. Therefore, with this coefficient value it can be safely applied to
problems with eigenvalues whose imaginary part is large, such as typically arise from the discretization of hyperbolic problems. This is not
the case for the standard trapezoidal rule (or Crank-Nicolson) implicit
method, whose linear stability region is exactly bounded by the imaginary axis. As a consequence, it is common to apply the trapezoidal
rule with off centering, see e.g. [7], [11] as well as [52], which results
in a first order time discretization. TR-BDF2 appears therefore to be
an interesting one step alternative to maintain full second order accuracy, especially considering that, if formulated as (8), it is equivalent
to performing two Crank-Nicolson steps with slightly modified coefficients. In order to highlight the advantages of the proposed method in
terms of accuracy with respect to other common robust stiff solvers,
we plot in figure 1 the contour levels of the absolute value of the linear stability function of the TR-BDF2 method without off centering in
the first stage, compared to the analogous contours of the off centered
Crank-Nicolson method with averaging parameter θ = 0.6, θ = 0.7 in
figures 2, 3, respectively, and to those of the BDF2 method in figure 4.
It is immediate to see that TR-BDF2 introduces less damping around
the imaginary axis for moderate values of the time step. On the other
hand, TR-BDF2 is more selective in damping very large eigenvalues,
as clearly displayed in figure 5, where the absolute values of the linear
stability functions of the same methods (with the exception of BDF2,
for which an explicit representation of the stability function is not
available) are plotted along the imaginary axis.
8
5
4
3
2
1
0
−1
−2
−3
−4
−5
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
Figure 1: Contour levels of the absolute value of the stability function of the
TR-BDF2 method without off centering in the first stage. Contour spacing
is 0.1 from 0.5 to 1.
5
4
3
2
1
0
−1
−2
−3
−4
−5
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
Figure 2: Contour levels of the absolute value of the stability function of
the off centered Crank-Nicolson method with averaging parameter θ = 0.6
(equivalent to an off centering parameter valued 0.05). Contour spacing is
0.1 from 0.5 to 1.
9
5
4
3
2
1
0
−1
−2
−3
−4
−5
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
Figure 3: Contour levels of the absolute value of the stability function of
the off centered Crank-Nicolson method with averaging parameter θ = 0.7
(equivalent to an off centering parameter valued 0.1). Contour spacing is 0.1
from 0.5 to 1.
5
4
3
2
1
0
−1
−2
−3
−4
−5
−5
−4.5
−4
−3.5
−3
−2.5
−2
−1.5
−1
−0.5
0
Figure 4: Contour levels of the absolute value of the stability function of the
BDF2 method. Contour spacing is 0.1 from 0.5 to 1.
10
1
CN off=0.05
CN off=0.01
TR−BDF2 off=0.0
TR−BDF2 off=0.02
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
−15
−10
−5
0
5
10
15
Figure 5: Graph of the absolute value of the stability functions of several
L-stable methods along the imaginary axis.
11
4 Review of Semi-Lagrangian evolution operators for vector fields on the
sphere
The semi-Lagrangian method can be described introducing the concept of evolution operator, along the lines of [37, 35]. Indeed, let
G = G(x, t) denote a generic function of space and time that is the
solution of
DG
∂G
u ∂G
v ∂G
=
+
+
= 0.
Dt
∂t
mx ∂x
my ∂y
To approximate this solution on the time interval [tn , tn+1 ], a numerical evolution operator E is introduced, that approximates the
exact evolution operator associated to the frozen velocity field u∗ =
(u∗ , v ∗ )T , that may coincide with the velocity field at time level tn or
with an extrapolation derived from more previous time levels. More
precisely, if X(t; tn+1 , x) denotes the solution of
dX(t; tn+1 , x)
= u∗ (X(t; tn+1 , x))
dt
(10)
with initial datum X(tn+1 ; tn+1 , x) = x at time t = tn+1 , then the expression [E(tn , ∆t)G](x) denotes a numerical approximation of Gn (xD )
where xD = X(tn ; tn+1 , x) and the notation Gn (x) = G(x, tn ) is used.
Since xD is nothing but the position at time tn of the fluid parcel
reaching location x at time tn+1 , according to standard terminology,
it is called the departure point associated with the arrival point x.
Different methods can be employed to approximate xD ; in this paper, for simplicity, the method proposed in [34] has been employed in
spherical geometry. Furthermore, to guarantee an accuracy compatible with that of the semi-implicit time-discretization, an extrapolation
1
un+ 2 of the velocity field at the intermediate time level tn + ∆t/2 was
used as u∗ in (10). On the other hand, in the application to Cartesian geometry (for the vertical slice discretization), a simple first order
Euler method with sub-stepping was employed, see e.g. [16], [45].
In case of the advection of a vector field
DG
∂G
u ∂G
v ∂G
=
+
+
= 0,
Dt
∂t
mx ∂x
my ∂y
as in the momentum equation (2), the extension of this approach has
to take into account the curvature of the spherical manifold. More
specifically, unit basis vectors at the departure point are not in general
aligned with those at the arrival point, i.e., if î, ĵ, k̂ represent a unit
vector triad, in general î(x) 6= î(xD ), ĵ(x) 6= ĵ(xD ), k̂(x) 6= k̂(xD ).
12
To deal with this issue two approaches are available. The first,
intrinsically Eulerian, consists in the introduction of the Christoffel symbols in the covariant derivatives definition, giving rise to the
well known metric terms, before the SISL discretization, and then
in the approximation along the trajectories of those metric terms.
This approach has been shown to be source of instabilities in a semiLagrangian frame, see e.g. [44, 8, 9, 13] and therefore is not adopted
in this work. The second approach, more suitable for semi-Lagrangian
discretizations, takes into account the curvature of the manifold only
at discrete level, i.e. after the SISL discretization has been performed. Many variations of this idea have been proposed, see e.g.
[44, 8, 9, 3, 49]. In [48], they have all been derived in a unified way
by the introduction of a proper rotation matrix that transforms vector components in the departure-point unit vector triad îD = î(xD ),
ĵ D = ĵ(xD ), k̂D = k̂(xD ) into vector components in the arrival-point
unit vector triad î = î(x), ĵ = ĵ(x), k̂ = k̂(x). To see how this rotation matrix comes into play, it is sufficient to consider the action of
the evolution operator E on a given vector valued function of space
and time G, defined as an approximation of
[E(tn , ∆t)G] (x) = Gn (xD ),
(11)
and to write this equation componentwise. Gn (xD ) is known through
its components in the departure point unit vector triad:
Gn (xD ) = Gxn (xD )îD + Gyn (xD )ĵ D + Gzn (xD )k̂D .
(12)
Therefore, via (11), the components of [E(tn , ∆t)G] (x) in the unit
vector triad at the same point are given by projection of (12) along î,
ĵ, k̂:
î · Gn (xD ) = Gxn (xD ) î · îD + Gyn (xD ) î · ĵ D + Gzn (xD ) î · k̂D ,
ĵ · Gn (xD ) = Gxn (xD ) ĵ · îD + Gyn (xD ) ĵ · ĵ D + Gzn (xD ) ĵ · k̂D ,
k̂ · Gn (xD ) = Gxn (xD )k̂ · îD + Gyn (xD )k̂ · ĵ D + Gzn (xD )k̂ · k̂D ,
i.e., in matrix notation:
n
î · [E(tn , ∆t)G] (x)
Gx
ĵ · [E(tn , ∆t)G] (x) = R Gyn
Gzn
k̂ · [E(tn , ∆t)G] (x)
î · îD î · ĵ D î · k̂D
R = ĵ · îD ĵ · ĵ D ĵ · k̂D .
k̂ · îD k̂ · ĵ D k̂ · k̂D
13
where
Under the shallow atmosphere approximation [51], R can be reduced
to the 2 × 2 rotation matrix
Λ
Λ12
Λ = Λ(x, xD ) = 11
,
(13)
Λ21 Λ22
where, as shown in [48], Λ11 = Λ22 = (R11 + R22 )/(1 + R33 ), Λ12 =
−Λ21 = (R12 −R21 )/(1+R33 ). Therefore, in the following the evolution
operator for vector fields will be defined componentwise as
î · [E(tn , ∆t)G] (x)
ĵ · [E(tn , ∆t)G] (x)
Gxn (xD )
=Λ
.
Gyn (xD )
(14)
5 A novel SISL time integration approach for the shallow water equations
on the sphere
The SISL discretization of equations.(1)-(2) based on (8) is then obtained by performing the two stages in (8) after reinterpretation of the
intermediate values in a semi-Lagrangian fashion. Furthermore, in order to avoid the solution of a nonlinear system, the dependency on h in
h∇ · u is linearized in time, as common in semi-implicit discretizations
based on the trapezoidal rule, see e.g. [7],[52]. Numerical experiments
reported in the following show that this does not prevent to achieve
second order accuracy in the regimes of interest for numerical weather
prediction. The TR stage of the SISL time semi-discretization of the
equations in vector form (1)-(2) is given by
hn+2γ
+ γ∆t hn ∇ · un+2γ
= E (tn , 2γ∆t) [h − γ∆t h ∇ · u] ,
i
h
un+2γ + γ∆t g∇hn+2γ + f k̂ × un+2γ = −γ∆t g∇b
h
io
n
+E tn , 2γ∆t u − γ∆t g(∇h + ∇b) + f k̂ × u .
(15)
(16)
The TR stage is then followed by the BDF2 stage:
hn+1 + γ2 ∆t hn+2γ ∇ · un+1
= 1 − γ3 E tn , ∆t h
+ γ3 E tn + 2γ∆t, (1 − 2γ)∆t h,
14
(17)
i
h
un+1 + γ2 ∆t g∇hn+1 + f k̂ × un+1 = −γ2 ∆t g∇b
+ 1 − γ3 E tn , ∆t u + γ3 E tn + 2γ∆t, (1 − 2γ)∆t u. (18)
For each of the two stages, the spatial discretization can be performed
along the lines described in [52], allowing for variable polynomial order to locally represent the solution in each element. The spatial
discretization approach considered is independent of the nature of the
mesh and could also be implemented for fully unstructured and even
non conforming meshes. For simplicity, however, in this paper only
an implementation on a structured mesh in longitude-latitude coordinates has been developed. In principle, either Lagrangian or hierarchical Legendre bases could be employed. We will work almost
exclusively with hierarchical bases, because they provide a natural environment for the implementation of a p−adaptation algorithm, see
for example [56]. A central issue in finite element formulations for
fluid problems is the choice of appropriate approximation spaces for
the velocity and pressure variables (in the context of SWE, the role of
the pressure is played by the free surface elevation). An inconsistent
choice of the two approximation spaces indeed may result in a solution
that is polluted by spurious modes, for the specific case of SWE see
for example [31, 53, 54] as well as the more recent and comprehensive
analysis in [30]. Here, we have not investigated this issue in depth, but
the model implementation allows for approximations of higher polynomial degree pu for the velocity fields than ph for the height field. Even
though no systematic study was performed, no significant differences
were noticed between results obtained with equal or unequal degrees.
In the following, only results with unequal degrees pu = ph + 1 are
reported, with the exception of an empirical convergence test for a
steady geostrophic flow.
All the integrals appearing in the elemental equations are evaluated
by means of Gaussian numerical quadrature formulae, with a number
of quadrature nodes consistent with the local polynomial degree being
used. In particular, notice that integrals of terms in the image of
the evolution operator E, i.e. of functions evaluated at the departure
points of the trajectories arriving at the quadrature nodes, cannot
be computed exactly (see e.g. [36, 40]), since such functions are not
polynomials. Therefore a sufficiently accurate approximation of these
integrals is needed, which may entail the need to employ numerical
quadrature formulae with more nodes than the minimal requirement
implied by the local polynomial degree. This overhead is actually
compensated by the fact that, for each Gauss node, the computation
15
of the departure point is only to be executed once for all the quantities
to be interpolated.
After spatial discretization has been performed, the discrete degrees of freedom representing velocity unknowns can be replaced in
the respective discrete height equations, yielding in each case a linear system whose structure is entirely analogous to that obtained in
[52]. The non-symmetric linear systems obtained from the TR-BDF2
stages are solved in our implementation by the GMRES method [46].
A classical stopping criterion based on a relative error tolerance of
10−10 was employed (see e.g. [24]). For the GMRES solver, so far,
only a block diagonal preconditioning was employed. As it will be
shown in section 7, the condition number of the systems to be solved
can be greatly reduced if lower degree elements are employed close
to the poles. In any case, the total computational cost of one TRBDF2 step is entirely analogous to that of one step of the standard
off centered trapezoidal rule employed in [52], since the structure of
the systems is the same but for each stage only a fraction of the time
step is being computed. Once hn+1 has been computed by solving this
linear system, then un+1 can be recovered by back substituting into
the momentum equation.
6 Extension of the time integration approach to the Euler equations
In this section, we show that the previously proposed method can
be extended seamlessly to the fully compressible Euler equations as
formulated in equations (3) - (6). For simplicity, only the application
to the (x, z) two dimensional vertical slice case is presented, but the
extension to three dimensions is straightforward. Again, in order to
avoid the solution of a nonlinear system, the dependency on Π in
Π∇ · u and the dependency on Θ in Θ∇π are linearized in time, as
common in semi-implicit discretizations based on the trapezoidal rule,
see e.g. [10], [4].
The semi-Lagrangian counterpart of the TR substep of (8) is first
applied to to (3) - (6), so as to obtain:
π n+2γ + γ∆t (cp /cv − 1) Πn ∇ · un+2γ = −π ∗
+E (tn , 2γ∆t) [Π − γ∆t (cp /cv − 1) Π ∇ · u] ,
16
(19)
∂π n+2γ
un+2γ + γ∆t cp Θn
=
∂x
∂π
n
,
E(t , 2γ∆t) u − γ∆t cp Θ
∂x
n+2γ
θ n+2γ
n ∂π
w
+ γ∆t cp Θ
=
−g ∗
∂z
θ
θ
∂π
n
−g ∗
,
E(t , 2γ∆t) w − γ∆t cp Θ
∂z
θ
(20)
n+2γ
θ
n+2γ
dθ ∗
dθ ∗ n+2γ
n
w
= E(t , 2γ∆t) θ − γ∆t
w .
+ γ∆t
dz
dz
(21)
(22)
Following [10] the time semi-discrete energy equation (22) can be
inserted into the time semi-discrete vertical momentum equation (21),
in order to decouple the momentum and the energy equations as follows
g dθ ∗
∂π n+2γ
1 + (γ∆t) ∗
wn+2γ + γ∆tcp Θn
=
θ dz
∂z
∂π
θ
n
E(t , 2γ∆t) w − γ∆t cp Θ
−g ∗
∂z
θ
∗
dθ
g
n
w .
+γ∆t ∗ E(t , 2γ∆t) θ − γ∆t
θ
dz
2
(23)
Equations (19), (20) and (23) are a set of three equations in three unknowns only, namely π, u, and w that can be compared with equations
(15), (16) with f = 0 and mx = my = 1 (Cartesian geometry). From
the comparison it is clear that the two formulations are isomorphic
under correspondence π ←→ h, u ←→ u, w ←→ v.
We can then consider the semi-Lagrangian counterpart of the BDF2
substep of (8) applied to (3) - (6) to obtain:
π n+1 + γ2 ∆t (cp /cv − 1) Πn+2γ ∇ · un+1
= −π ∗ + (1 − γ3 )[E (tn , ∆t) Π]
+ γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) Π],
17
(24)
∂π n+1
∂x
= (1 − γ3 )[E (tn , ∆t) u]
un+1 + γ2 ∆t cp Θn+2γ
+ γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) u],
w
n+1
n+1
θ n+1
n+2γ ∂π
+ γ2 ∆t cp Θ
−g ∗
∂z
θ
n
= (1 − γ3 )[E (t , ∆t) w]
+ γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) w],
dθ ∗ n+1
w
dz
= (1 − γ3 )[E (tn , ∆t) θ]
(25)
(26)
θ n+1 + γ2 ∆t
+ γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) θ].
(27)
Again, following [10], the time semi-discrete energy equation (27) can
be inserted into the time semi-discrete vertical momentum equation
(26), in order to decouple the momentum and the energy equations:
g dθ ∗
∂π n+1
1 + (γ2 ∆t) ∗
wn+1 + γ2 ∆t cp Θn+2γ
=
θ dz
∂z
(1 − γ3 )[E (tn , ∆t) w] + γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) w] +
(28)
g
γ2 ∆t ∗ {(1 − γ3 )[E (tn , ∆t) θ] + γ3 [E (tn + 2γ∆t, (1 − 2γ)∆t) θ]} .
θ
2
Now equations (24), (25) and (28) are a set of three equations
in three unknowns only, namely π, u, and w, that can be compared
with equations (17), (18) with f = 0 and mx = my = 1 (Cartesian
geometry). Again, it is easy to see that also in this case exactly the
same structure results as in equations (17)-(18) with the correspondence π ←→ h, u ←→ u, w ←→ v, so that the approach (and code)
proposed for the shallow water equations can be extended to the fully
compressible Euler equation in a straightforward way.
18
7
Numerical experiments
The numerical method introduced in section 5 has been implemented
and tested on a number of relevant test cases using different initial
conditions and bathymetry profiles, in order to assess its accuracy and
stability properties and to analyze the impact of the p−adaptivity
strategy. Whenever a reference solution was available, the relative
errors were computed in the L1 , L2 and L∞ norms at the final time tf
of the simulation according to [55] as:
I [ |h(·, tf ) − href (·, tf )| ]
,
I [ |href (·, tf )| ]
n h
2 io1/2
I
h(·, tf ) − href (·, tf )
l2 (h) =
,
1/2
I href (·, tf )2
max |h(·, tf ) − href (·, tf )|
l∞ (h) =
,
max |href (·, tf )|
l1 (h) =
(29)
(30)
(31)
where href denotes the reference solution for a model variable h and
I is a discrete approximation of the global integral
R
h mx my dx
,
(32)
I(h) = RΩ
Ω mx my dx
computed by an appropriate numerical quadrature rule, consistent
with the numerical approximation being tested, and the maximum is
computed over all nodal values.
The test cases considered for the shallow water equations in spherical geometry are
• a steady-state geostrophic flow: in particular, we have analyzed
results in test case 2 of [55] in the configuration least favorable
for methods employing longitude-latitude meshes;
• the unsteady flow with exact analytical solution described in [29];
• the polar rotating low-high, introduced in [33], aimed at showing
that no problems arise even in the case of strong cross polar flows;
• zonal flow over an isolated mountain and Rossby-Haurwitz wave
of wavenumber 4, corresponding respectively to test cases 5 and
6 in [55].
For the first two tests, analytic solutions are available and empirical
convergence tests can be performed. The test cases considered for the
discretization of equations (3)-(6) are
19
• inertia gravity waves involving the evolution of a potential temperature perturbation in a channel with periodic boundary conditions and uniformly stratified environment with constant BruntWäisälä frequency, as described in [47];
• a rising thermal bubble given by the evolution of a warm bubble
in a constant potential temperature environment, as described
in [23].
In all the numerical experiments performed for this paper, neither
spectral filtering nor explicit diffusion of any kind were employed,
the only numerical diffusion being implicit in the time discretization
approach. We have not yet investigated to which extent the quality
of the solutions is affected by this choice, but this should be taken
into account when comparing quantitatively the results of the present
method to those of reference models, such as the one described in
[22], in which explicit numerical diffusion is added. Sensitivity of
the comparison results to the amount of numerical diffusion has been
highlighted in several model validation exercises, see e.g. [43].
Since semi-implicit, semi-Lagrangian methods are most efficient for
low Froude number flows, where the typical velocity is much smaller
than that of the fastest propagating waves, all the tests considered
fall in this hydrodynamical regime. Therefore, in order to assess the
method efficiency, a distinction has been made between the maximum
Courant number based on the velocity, on one hand, and, on the
other hand, the maximum Courant number based on the celerity, or
the maximum Courant number based on the sound speed, defined
respectively as
kuk∞ ∆t
Cvel = max
∆x/p
p
√
(cp /cv )RΘΠ∆t
gh∆t
, Csnd = max
,
Ccel = max
∆x/p
∆x/p
where ∆x is to be interpreted as generic value of the meshsize in
either coordinate direction. For the tests in which p−adaptivity was
employed, if pnI denotes the local polynomial degree used at timestep
tn to represent a model variable inside the I − th element of the mesh,
while pmax is the maximum local polynomial degree considered, the
efficiency of the method in reducing the computational effort has been
measured by monitoring the evolution of the quantities
PN
ITNnadapt
(pn + 1)2
n
n
,
∆
=
,
∆dof = I=1 I
iter
N (pmax + 1)2
ITNnmax
where N is the total number of elements, ITNnadapt denotes the total
number of GMRES iterations at time step n for the adapted local
20
degrees configuration and ITNnmax the total number of GMRES iterations at time step n for the configuration with maximum degree
in all elements, respectively. Average values of these indicators over
the simulations performed are reported in the following, denoted by
respectively. The error between the adaptive
∆average
and ∆average
iter
dof
solution and the corresponding one obtained with uniform maximum
polynomial degree everywhere has been measured in terms of (29).
Finally, in some cases conservation of global invariants has been monitored by evaluating at each time step the following global integral
quantities:
J(q n ) =
I(q(·, tn )) − I(q(·, t0 ))
,
I(q(·, t0 ))
(33)
where I(q) has been defined in (32) and q n = q(·, tn ) is the density
associated to each global invariant. According to the choice of q,
following invariants are considered: mass, i.e. q = qmass = h, total
energy, i.e. q = qenerg = 12 (hu·u+g(h2 −b2 )), and potential enstrophy,
1
(k̂ · ∇ × u + f )2 .
i.e. q = qenstr = 2h
7.1
Steady-state geostrophic flow
We first consider the test case 2 of [55], where the solution is a steady
state flow with velocity field corresponding to a zonal solid body rotation and h field obtained from the velocity ones through geostrophic
balance. All the parameter values are taken as in [55]. The flow orientation parameter has been chosen here as α = π/2 − 0.05, making
the test more challenging on a longitute-latitude mesh. Error norms
associated to the solution obtained on a mesh of 10 × 5 elements for
different polynomial degrees are shown in tables 1, 2 and 3 for h, u and
v, respectively. All the results have been computed at tf = 10 days
at fixed maximum Courant numbers Ccel = 8, Cvel = 2, so that different values of ∆t have been employed for different polynomial order.
We remark that the resulting time steps are significantly larger than
those allowed by typical explicit time discretizations for analogous DG
space discretizations, see e.g. the results in [38]. The spectral decay in
the error norms can be clearly observed, until the time error becomes
dominant. For better comparison with the results in [38], we consider
again the configuration with ph = 6, pu = 7 on 10 × 5 elements, which
corresponds to the same resolution in space as for the 150 × 8 × 8 grid
used in [38]. While ∆t = 36 s is used in [38] giving a l∞ (h) ≈ 8× 10−6 ,
the proposed SISLDG formulation can be run with ∆t = 3600 s, in
which case l∞ (h) ≈ 3 × 10−7 , and the average number of iterations
21
required by the linear solver is 1 for the TR substep and 4 for the
BDF2 substep.
ph pu ∆t [s]
2
3
4
5
6
7
8
3
4
5
6
7
8
9
4800
3600
2880
2400
2057
1800
1600
l1 (h)
l2 (h)
l∞ (h)
5.558 × 10−3
6.017 × 10−4
1.743 × 10−5
1.586 × 10−6
8.829 × 10−8
1.246 × 10−8
5.641 × 10−9
6.805 × 10−3
8.176 × 10−4
2.405 × 10−5
2.281 × 10−6
1.206 × 10−7
1.590 × 10−8
5.952 × 10−9
1.914 × 10−2
2.569 × 10−3
9.024 × 10−5
1.058 × 10−5
4.926 × 10−7
4.158 × 10−8
6.320 × 10−9
Table 1: Relative errors on h for different polynomial degrees, SWE test case
2 with α = π/2 − 0.05 at time tf = 10 days.
ph pu ∆t [s]
2
3
4
5
6
7
8
3
4
5
6
7
8
9
4800
3600
2880
2400
2057
1800
1600
l1 (u)
l2 (u)
l∞ (u)
6.351 × 10−2
9.505 × 10−3
4.288 × 10−4
4.598 × 10−5
2.057 × 10−6
2.162 × 10−7
2.013 × 10−8
6.432 × 10−2
1.037 × 10−2
4.887 × 10−4
4.830 × 10−5
2.262 × 10−6
2.358 × 10−7
2.276 × 10−8
1.143 × 10−1
2.106 × 10−2
2.393 × 10−3
1.706 × 10−4
5.879 × 10−6
6.428 × 10−7
3.268 × 10−8
Table 2: Relative errors on u for different polynomial degrees, SWE test case
2 with α = π/2 − 0.05 at time tf = 10 days.
Another convergence test was performed for ph = pu = 3, increasing the number of elements and correspondingly decreasing the
value of the time step. In this case, the maximum Courant numbers vary because of the mesh inhomogeneity, so that 2 < Ccel < 18,
0.5 < Cvel < 4. The results are reported in tables 4, 5 and 6 for h,
u and v, respectively. The empirical convergence order q2emp based
on the l2 norm errors has also been estimated, showing that in this
stationary test convergence rates above the second order of the time
discretization can be achieved.
22
ph pu ∆t [s]
2
3
4
5
6
7
8
3
4
5
6
7
8
9
4800
3600
2880
2400
2057
1800
1600
l1 (v)
l2 (v)
l∞ (v)
1.001 × 10−1
1.859 × 10−2
7.376 × 10−4
8.185 × 10−5
3.074 × 10−6
3.370 × 10−7
2.175 × 10−8
1.016 × 10−1
1.823 × 10−2
7.428 × 10−4
8.307 × 10−5
3.173 × 10−6
3.432 × 10−7
2.317 × 10−8
2.698 × 10−1
6.848 × 10−2
2.884 × 10−3
2.574 × 10−4
1.123 × 10−5
1.323 × 10−6
5.124 × 10−8
Table 3: Relative errors on v for different polynomial degrees, SWE test case
2 with α = π/2 − 0.05 at time tf = 10 days.
Nx × Ny ∆t [s]
10 × 5
20 × 10
40 × 20
80 × 40
3600
1800
900
450
l1 (h)
l2 (h)
l∞ (h)
q2emp
2.557 × 10−4
2.187 × 10−5
2.530 × 10−6
3.996 × 10−7
3.495 × 10−4
2.889 × 10−5
3.353 × 10−6
5.534 × 10−7
1.403 × 10−3
1.566 × 10−4
1.430 × 10−5
3.134 × 10−6
−
3.6
3.1
2.6
Table 4: Relative errors on h for different number of elements, ph = pu = 3,
SWE test case 2 with α = π/2 − 0.05 at time tf = 10 days.
Nx × Ny ∆t [s]
10 × 5
20 × 10
40 × 20
80 × 40
3600
1800
900
450
l1 (u)
l2 (u)
l∞ (u)
q2emp
2.769 × 10−3
2.896 × 10−4
3.647 × 10−5
6.826 × 10−6
3.358 × 10−3
3.720 × 10−4
4.563 × 10−5
1.035 × 10−5
8.948 × 10−3
2.414 × 10−3
2.473 × 10−4
9.525 × 10−5
−
3.2
3.0
2.1
Table 5: Relative errors on u for different number of elements, ph = pu = 3,
SWE test case 2 with α = π/2 − 0.05 at time tf = 10 days.
23
Nx × Ny ∆t [s]
10 × 5
20 × 10
40 × 20
80 × 40
3600
1800
900
450
l1 (v)
l2 (v)
l∞ (v)
q2emp
3.309 × 10−3
4.016 × 10−4
5.180 × 10−5
9.405 × 10−6
3.346 × 10−3
4.233 × 10−4
5.578 × 10−5
1.214 × 10−5
8.250 × 10−3
1.255 × 10−3
2.329 × 10−4
7.763 × 10−5
−
3.0
2.9
2.2
Table 6: Relative errors on v for different number of elements, ph = pu = 3,
SWE test case 2 with α = π/2 − 0.05 at time tf = 10 days.
24
7.2
Unsteady flow with analytic solution
In a second, time dependent test, the analytic solution of (1)-(2) derived in [29] has been employed to assess the performance of the proposed discretization. More specifically, the analytic solution defined in
formula (23) of [29] was used. Since the exact solution is periodic, the
initial profiles also correspond to the exact solution an integer number
of days later. The proposed SISLDG scheme has been integrated up
to tf = 5 days with ph = 4 and pu = 5 on meshes with increasing
number of elements, while the time step has been decreased accordingly. In this case, the maximum Courant numbers vary because of
the mesh dishomogeneity, so that 4 < Ccel < 26, 1.25 < Cvel < 8.
Error norms for h, u, v of the above-mentioned integrations have been
computed at tf = 5 days and displayed in tables 7 - 9. An empirical order estimation shows that full second order accuracy in time is
attained.
Nx × Ny
10 × 5
20 × 10
40 × 20
80 × 40
∆t [s]
l1 (h)
l2 (h)
l∞ (h)
q2emp
3600
1800
900
450
5.456 × 10−3
1.246 × 10−3
3.039 × 10−4
7.548 × 10−5
6.120 × 10−3
1.397 × 10−3
3.410 × 10−4
8.475 × 10−5
9.537 × 10−3
2.143 × 10−3
5.207 × 10−4
1.292 × 10−4
2.1
2.0
2.0
Table 7: Relative errors on h at different resolutions, Läuter test case.
Nx × Ny
10 × 5
20 × 10
40 × 20
80 × 40
∆t [s]
l1 (u)
l2 (u)
l∞ (u)
q2emp
3600
1800
900
450
6.567 × 10−2
1.665 × 10−2
4.210 × 10−3
1.057 × 10−3
7.848 × 10−2
1.994 × 10−2
5.032 × 10−3
1.261 × 10−3
1.670 × 10−1
3.931 × 10−2
9.811 × 10−3
2.452 × 10−3
2.0
2.0
2.0
Table 8: Relative errors on u at different resolutions, Läuter test case.
For comparison, analogous errors have been computed with the
same discretization parameters but employing the off centered Crank
Nicolson method of [52] with θ = 0.6. The resulting improvement in
the errors between the TRBDF2 scheme and the off-centered Crank
Nicolson is achieved at an essentially equivalent computational cost in
terms of total CPU time employed.
25
Nx × Ny
10 × 5
20 × 10
40 × 20
80 × 40
∆t [s]
l1 (v)
l2 (v)
l∞ (v)
q2emp
3600
1800
900
450
1.174 × 10−1
2.939 × 10−2
7.336 × 10−3
1.833 × 10−3
1.198 × 10−1
3.002 × 10−2
7.497 × 10−3
1.874 × 10−3
2.316 × 10−1
5.561 × 10−2
1.390 × 10−2
3.464 × 10−3
2.0
2.0
2.0
Table 9: Relative errors on v at different resolutions, Läuter test case.
Nx × Ny
10 × 5
20 × 10
40 × 20
80 × 40
∆t [s]
l1 (h)
l2 (h)
l∞ (h)
q2emp
3600
1800
900
450
1.444 × 10−2
8.742 × 10−3
4.814 × 10−3
2.526 × 10−3
1.633 × 10−2
9.894 × 10−3
5.451 × 10−3
2.861 × 10−3
2.398 × 10−2
1.445 × 10−2
7.956 × 10−3
4.177 × 10−3
0.7
0.9
0.9
Table 10: Relative errors on h at different resolutions, Läuter test case with
off centered Crank Nicolson, θ = 0.6.
Nx × Ny
∆t [s]
l1 (u)
l2 (u)
l∞ (u)
q2emp
10 × 5
20 × 10
40 × 20
80 × 40
3600
1800
900
450
1.800 × 10−1
1.077 × 10−1
5.895 × 10−2
3.084 × 10−2
2.092 × 10−1
1.255 × 10−1
6.880 × 10−2
3.603 × 10−2
3.810 × 10−1
2.155 × 10−1
1.186 × 10−1
6.234 × 10−2
0.7
0.9
0.9
Table 11: Relative errors on u at different resolutions, Läuter test case with
off centered Crank Nicolson, θ = 0.6.
Nx × Ny
10 × 5
20 × 10
40 × 20
80 × 40
∆t [s]
l1 (v)
l2 (v)
l∞ (v)
q2emp
3600
1800
900
450
3.608 × 10−1
2.164 × 10−1
1.185 × 10−1
6.195 × 10−2
3.665 × 10−1
2.198 × 10−1
1.203 × 10−1
6.291 × 10−2
5.166 × 10−1
3.041 × 10−1
1.671 × 10−2
8.809 × 10−2
0.7
0.9
0.9
Table 12: Relative errors on v at different resolutions, Läuter test case with
off centered Crank Nicolson, θ = 0.6.
26
7.3
Zonal flow over an isolated mountain
We have then performed numerical simulations reproducing the test
case 5 of [55], given by a zonal flow impinging on an isolated mountain
of conical shape. The geostrophic balance here is broken by orographic
forcing, which results in the development of a planetary wave propagating all around the globe.
Plots of the fluid depth h as well as of the velocity components
u and v at 15 days are shown in figures 6-8. The resolution used
corresponds to a mesh of 60 × 30 elements with ph = 4, pu = 5, and
∆t = 900 s, giving a Courant number Ccel ≈ 58 in elements close to
the poles. It can be observed that all the main features of the flow are
correctly reproduced. In particular, no significant Gibbs phenomena
are detected in the vicinity of the mountain, even in the initial stages
of the simulation.
1.5
5250
52
5500
5750
55
5
5250
50
00
0
0
75
0.5
5500
525
5250
5500
1
5750
750
5750
0
0
550
575
5500
0
5250
50
−1.5
−3
52
−1
5250
−2
50
50
550
57
57
−0.5
55
0
5250 0
y
5
0
−1
0
x
1
2
3
Figure 6: h field after 15 days, isolated mountain wave test case, Ccel ≈ 58.
Contour lines spacing is 50 m.
The evolution in time of global invariants during this simulation is
shown in figures 9(a), 9(b), 9(c), respectively. Error norms for h and
u at different resolutions corresponding to a Ccel ≈ 6 and ph = pu = 3,
have been computed at tf = 5 days and are displayed in tables 13 14, with respect to a reference solution given by the National Center
for Atmospheric Research (NCAR) spectral model [22] at resolution
T511. It is apparent the second order of the proposed SISLDG scheme
in time. Since, as observed in [22], the National Center for Atmospheric Research (NCAR) spectral model incorporates diffusion terms
27
22
14
6
22
22
6
14
6
6
−1.5
−3
6
14
14
6
14
6
14
22
6
14
22
−2
14
−1
14
22
22
14
30
14
14
y
22
14
22
14
22
14
6
−0.5
6
22
30
14
22
14
14
0
14
14
6
14
14
−2
6
−2
−10
1
6
14
−2
6
0.5
−2
6
−2
30
1.5
6
−2
−1
0
x
1
2
3
Figure 7: u field after 15 days, isolated mountain wave test case, Ccel ≈ 58.
Contour lines spacing is 4 m s−1 .
in the governing equations, while the proposed SISLDG scheme does
not employ any diffusion terms nor filtering, nor smoothing of the topography, for this test it seemed more appropriate to compute relative
errors with respect to NCAR spectral model [22] solution at an earlier
time, tf = 5 days, when it can be assumed that the effects of diffusion
have less impact. Error norms for h and u have been computed at
tf = 15 days at different resolutions (corresponding to a Ccel ≈ 7),
ph = pu = 3, and displayed in tables 15 - 16.
Nx × Ny
∆t[min]
l1 (h)
l2 (h)
l∞ (h)
q2emp
12 × 6
24 × 12
48 × 24
20
10
5
8.19 × 10−4
1.49 × 10−4
2.88 × 10−5
1.08 × 10−3
2.08 × 10−4
4.25 × 10−5
5.90 × 10−3
1.92 × 10−3
8.40 × 10−4
2.4
2.3
Table 13: Relative errors on h at different resolutions, isolated mountain
wave test case, tf = 5 days.
Finally the mountain wave test case has been run on the same
mesh of 60 × 30 elements, ∆t = 900 s, with either static or static
plus dynamic adaptivity. The tolerance ǫ for the dynamic adaptivity
[52] has been set to ǫ = 10−2 . Results are reported in terms of error
norms with respect to a nonadaptive solution at the maximum uniform
28
−8
8
0
0
0
0
0
0
8
−8
0
−8
−8
8
−8
0
16
8
0
x
0
−8
−1
0
−2
8
−1.5
−3
−8
0
8
0
8
8
−16
0
−1
0
8
16
−8
0
−8
8
0
8
−16 −8
−8
8
−8
−16
−8
0
−0.5
0
0
0
8
0
y
0
0 0
8
0
0.5
16
0
0
8
1
−8
−16
8
−8
8
1.5
1
2
3
Figure 8: v field after 15 days, isolated mountain wave test case, Ccel ≈ 58.
Contour lines spacing is 4 m s−1 .
Nx × Ny
12 × 6
24 × 12
48 × 24
∆t[min]
l1 (u)
l2 (u)
l∞ (u)
q2emp
20
10
5
4.33 × 10−2
5.70 × 10−3
1.11 × 10−3
5.81 × 10−2
7.33 × 10−3
1.72 × 10−3
1.39 × 10−1
1.06 × 10−1
1.56 × 10−2
2.9
2.2
Table 14: Relative errors on u at different resolutions, isolated mountain
wave test case, tf = 5 days.
resolution and in terms of efficiency gain, measured through the saving
of number of linear solver iterations per time-step ∆average
as well as
iter
through the saving of number of degrees of freedom actually used
per timestep ∆average
; these results are summarized in tables 17 - 19:
dof
≈ 10.7% and
the use of static adaptivity only resulted in ∆average
iter
average
∆dof
≈ 88%, while the use of both static and dynamic adaptivity
led to ∆average
≈ 13% and ∆average
≈ 45%. The distribution of the
iter
dof
statically and dynamically adapted local polynomial degree used to
represent the solution after 15 days is shown in figure 10. It can be
noticed how, even after 15 days, higher polynomial degrees are still
automatically concentrated around the location of the mountain.
29
Nx × Ny
12 × 6
24 × 12
48 × 24
∆t[s]
l1 (h)
l2 (h)
l∞ (h)
q2emp
1500
750
375
2.34 × 10−3
5.99 × 10−4
2.00 × 10−4
2.92 × 10−3
7.72 × 10−4
2.74 × 10−4
1.49 × 10−2
3.87 × 10−3
1.87 × 10−3
1.9
1.5
Table 15: Relative errors on h at different resolutions, isolated mountain
wave test case, tf = 15 days.
Nx × Ny
12 × 6
24 × 12
48 × 24
∆t[s]
l1 (u)
l2 (u)
l∞ (u)
q2emp
1500
750
375
1.12 × 10−1
2.09 × 10−2
6.37 × 10−3
1.29 × 10−1
2.37 × 10−2
7.92 × 10−3
2.97 × 10−1
5.73 × 10−2
3.39 × 10−2
2.4
1.6
Table 16: Relative errors on u at different resolutions, isolated mountain
wave test case, tf = 15 days.
adaptivity
l1 (h)
l2 (h)
l∞ (h)
static
static + dynamic
1.415 × 10−4
1.660 × 10−4
3.314 × 10−4
3.419 × 10−4
2.117 × 10−3
2.038 × 10−3
Table 17: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for isolated mountain wave test case, h
field.
adaptivity
l1 (u)
l2 (u)
l∞ (u)
static
static + dynamic
1.289 × 10−2
1.509 × 10−2
3.275 × 10−2
3.309 × 10−2
1.524 × 10−1
1.475 × 10−1
Table 18: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for isolated mountain wave test case, u
field.
30
−5
−4
1
10
x 10
0
−1
−6
J (qmass )
J (qenerg )
10
−2
−3
−8
10
−4
−5
−10
10
0
5
days
10
15
−6
0
(a)
5
days
10
(b)
−5
4
x 10
I(qentsr )
3
2
1
0
−1
0
5
days
10
15
(c)
Figure 9: Integral invariants evolution, mass (a), energy (b), potential enstrophy (c), isolated mountain wave test case, Ccel ≈ 58.
adaptivity
l1 (v)
l2 (v)
l∞ (v)
static
static + dynamic
2.501 × 10−2
2.833 × 10−2
6.824 × 10−2
7.019 × 10−2
6.597 × 10−1
6.975 × 10−1
Table 19: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for isolated mountain wave test case, v
field.
31
15
4
1
y
3
0
2
−1
−3
1
−2
−1
0
x
1
2
3
Figure 10: Statically and dynamically adapted local ph distribution at 15
days, isolated mountain wave test case.
32
7.4
Rossby-Haurwitz wave
We have then considered test case 6 of [55], where the initial datum
consists of a Rossby-Haurwitz wave of wave number 4. This case
actually concerns a solution of the nondivergent barotropic vorticity
equation, that is not an exact solution of the system (1) - (2). For a
discussion about the stability of this profile as a solution of (1) - (2) see
[50]. Plots of the fluid depth h as well as of the velocity components
u and v at 15 days are shown in figures 11-13. The resolution used
corresponds to a mesh of 64 × 32 elements with ph = 4, pu = 5, and
∆t = 900 s, giving a Courant number Ccel ≈ 83 in elements close to
poles. It can be observed that all the main features of the flow are
correctly reproduced.
1.5
10500
00
90
8500
0
00
10
900
950 0
10 0
00
0
00
90
10500
10500
95 1
00 000
0
0
1
00
00
0
x
10
9 10
900500 000
0
y
00
00
85
−1
0
50
10
00
−2
90
00
95 100
00
90
−1.5
−3
00
8500
8500
−1
00
95
95
0
00
10 0
950 0
900
−0.5
00
951
0000
90
0
0
00 000
95 1
0
900 500 000
9 10
0.5
00
8500
00
8500
1
85
90
8500
8500
8500
2
3
Figure 11: h field after 15 days, Rossby-Haurwitz wave test case, Ccel ≈ 83.
Contour lines spacing is 100 m.
The evolution in time of global invariants during this simulation
is shown in figures 14(a), 14(b), 14(c), respectively. Error norms for
h and u at different resolutions, corresponding to a Ccel ≈ 32 and
ph = 4, pu = 5 have been computed at tf = 15 days and are displayed in tables 20 - 21, with respect to a reference solution given
by the National Center for Atmospheric Research (NCAR) spectral
model [22] at resolution T511. It is apparent the second order of the
proposed SISLDG scheme in time. Unlike the NCAR spectral model,
the proposed SISLDG scheme does not employ any explicit numerical
diffusion.
Finally, the Rossby-Haurwitz wave test case has been run on the
33
1.5
24
24
24
48
48
48
48
24
48
48
48
48
24
48
48
72
48
24
24
24
72
96
48
48
72
48
24
72
96
24
y
48
24
48
72
72
48
48
−1.5
−3
96
48
−1
24
72
96
72
−0.5
48
24
24
0
48
48
0.5
24
1
24
−2
24
−1
0
x
1
2
3
Figure 12: u field after 15 days, Rossby-Haurwitz wave test case, Ccel ≈ 83.
Contour lines spacing is 8 m s−1 .
Nx × Ny
10 × 5
20 × 10
40 × 20
∆t[min]
l1 (h)
l2 (h)
l∞ (h)
q2emp
60
30
15
2.92 × 10−2
5.50 × 10−3
1.40 × 10−3
3.82 × 10−2
6.80 × 10−3
1.80 × 10−3
6.75 × 10−2
1.11 × 10−2
3.20 × 10−3
2.4
2.0
Table 20: Relative errors on h at different resolutions, Rossby-Haurwitz wave
test case.
same mesh of 64 × 32 elements, ∆t = 900 s, with either static or static
plus dynamic adaptivity. The tolerance ǫ for the dynamic adaptivity
[52] has been set to ǫ = 5 × 10−2 . Results are reported in terms of
error norms with respect to a nonadaptive solution at the maximum
uniform resolution and in terms of efficiency gain, measured through
the saving of number of linear solver iterations per time-step ∆average
iter
as well as through the saving of number of degrees of freedom actually
used per timestep ∆average
; these results are summarized in tables 22 dof
24: the use of static adaptivity only resulted in ∆average
≈ 10.7% and
iter
∆average
≈
88%,
while
the
use
of
both
static
and
dynamic
adaptivity
dof
average
average
≈ 13% and ∆dof
≈ 45%. The distribution of the
led to ∆iter
statically and dynamically adapted local polynomial degree used to
represent the solution after 15 days is shown in figure 15. It can be
noticed how, even after 15 days, and even if the maximum allowed ph
34
−12
−36
60
−12
−3
−60 6
12
36
−36
−12
− 12
−12
−36
−12
36
− 12
−12
36
−36
−12
36
60
60
−12
− 12
36
36
−12
−3
−60 6
36
−60
36
−60
60
−60
36
12
−36
−36
−36
−12
12
12
36
12
−36
60
−60
− 12
12
12
y
12
36
−36
60
12
36
−60
−36
12
36
12
36
12
−1
−36
60
−0.5
−12
12
36
12
0
−36
60
12
−60
0.5
−12
−3
6
12
36
1
−12
−3
6
1.5
12
−1.5
−3
−2
−1
0
x
1
2
3
Figure 13: v field after 15 days, Rossby-Haurwitz wave test case, Ccel ≈ 83.
Contour lines spacing is 8 m s−1 .
Nx × Ny
10 × 5
20 × 10
40 × 20
∆t[min]
l1 (u)
l2 (u)
l∞ (u)
q2emp
60
30
15
4.065 × 10−1
7.79 × 10−2
2.04 × 10−2
3.775 × 10−1
7.33 × 10−2
1.95 × 10−2
2.305 × 10−1
5.67 × 10−2
1.76 × 10−2
2.4
1.9
Table 21: Relative errors on u at different resolutions, Rossby-Haurwitz wave
test case.
is 4, the use of the adaptivity criterion with ǫ = 5 × 10−2 leads to the
use of at most cubic polynomials for the local representation of h.
35
adaptivity
l1 (h)
l2 (h)
l∞ (h)
static
static + dynamic
2.182 × 10−4
2.358 × 10−3
3.434 × 10−4
2.963 × 10−3
2.856 × 10−4
5.157 × 10−3
Table 22: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for Rossby-Haurwitz wave test case, h
field.
adaptivity
l1 (u)
l2 (u)
l∞ (u)
static
static + dynamic
7.041 × 10−3
3.639 × 10−2
1.236 × 10−2
3.387 × 10−2
2.834 × 10−2
2.678 × 10−2
Table 23: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for Rossby-Haurwitz wave test case, u
field.
adaptivity
l1 (v)
l2 (v)
l∞ (v)
static
static + dynamic
3.158 × 10−3
2.723 × 10−2
3.250 × 10−3
2.432 × 10−2
1.148 × 10−2
2.646 × 10−2
Table 24: Relative errors between (statically and statically plus dynamically)
adaptive and nonadaptive solution for Rossby-Haurwitz wave test case, v
field.
36
−9
−4
−10
10
−8
−10
−6
10
J (qmass )
J (qenerg )
−7
−8
−10
−6
−10
10
−5
−10
−4
−10
10
0
5
days
10
15
(a)
−10
0
5
days
10
(b)
−4
2
x 10
1
I(qentsr )
0
−1
−2
−3
−4
−5
−6
0
5
days
10
15
(c)
Figure 14: Integral invariants evolution, mass (a), energy (b), potential enstrophy (c), Rossby-Haurwitz wave test case, Ccel ≈ 83.
37
15
Figure 15: Statically and dynamically adapted local ph distribution at 15
days, Rossby-Haurwitz test case.
38
7.5
Nonhydrostatic inertia gravity waves
In this section we consider the test case proposed in [47]. It consists in a set of inertia-gravity waves propagating in a channel with a
uniformly stratified reference atmosphere characterized by a constant
Brunt-Wäisälä frequency N 2 = 0.01. The domain and the initial and
boundary conditions are identical to those of [47]. The initial perturbation in potential temperature radiates symmetrically to the left and
to the right, but because of the superimposed mean horizontal flow
(u = 20m/s), does not remain centered around the initial position.
Contours of potential temperature perturbation, horizontal velocity,
and vertical velocity time tf = 3000 s are shown in figures 16, 17, 18,
respectively. The computed results compare well with the structure
displayed by the analytical solution of the linearized equations proposed in [1] and with numerical results obtained with other numerical
methods, see e.g. [4]. It is to be remarked that for this experiment
300 × 10 elements, pπ = 4, pu = 5 and a timestep ∆t = 15 s were used,
corresponding to a Courant number Csnd ≈ 25.
0.000977
0.000414
54
01
0.0
−0.000149
−0.000149
0.000414
0.000977
0.00154
−0.000149
−0.000149
−0.000149
0.000414
−0.000149
−0.000149
0.000414
0.000414
0.000414
−0.000149
0.000414
−0.000149
−0.000149
−0.000149
−0.000149
0.000414
54
0.00097
0.000414 7
z
−0.000149
0.000414
977
0.000
0.000414
0.00154
01
0.000414
0.0
0.000977
14
−0.000149
0.000414
0.000977
0.000977 0.000414
0.000414
−0.000149
−0.000149
−0.000149
0.000414
0.0004
2000
−0.000149
3000
0.000977
−0.000149
0.000977
4000
−0.000149
0.000414
0.000977
5000
0.000414
6000
0.00154
7000
0.000414
0.000414
0.00154
8000
0.000414
−0.000149
9000
−0.000149
10000
1000
0
0
0.5
1
1.5
x
2
2.5
3
5
x 10
Figure 16: Contours of perturbation potential temperature in the internal
gravity wave test.
39
20
20
20.004
20
20
20
19.996
20
20
20
20
20
20
5000
20
20
20
6000
z
20.008
20.004
19.996
7000
20
20
2020
20
8000
20.004
19.992
19.996
20
9000
20
10000
20
4000
20
19.996
20
20
20
20
20
1
20
0.5
20
0
0
20
20
20
1000
20.004
2000
20
20
20
20
3000
1.5
x
2
2.5
3
5
x 10
Figure 17: Contours of horizontal velocity in the internal gravity wave test.
−0.001
0.001
0.003
0.001
0.001
0.001
−0.001
0.001
0.001
0.001
0.001
0.001
−0.001
0.001
0.001
0.001
0.001
−0.001
0.003
0.001
1
−0.00
−0.001
01
−0.0
0.001
0.003 0.001
0.001
0.001
0.001
0.001
−0.001
0.001
z
01
.0
−0
−0.001
0.003
0.003
0.003
0.003
2000
0.001
0.001
0.001
0.003
0.003 0.001
3000
0.001
4000
−0.001
5000
0.001
6000
0.003
7000
0.001
0.001
1
−0.00
01
8000
0.0
9000
0.001
10000
1000
0
0
0.5
1
1.5
x
2
2.5
3
5
x 10
Figure 18: Contours of vertical velocity in the internal gravity wave test.
40
7.6
Rising thermal bubble
4000
3500
3500
3000
3000
2500
2500
2000
2000
1500
1500
1000
1000
500
500
0
0
z
z
4000
500
1000
1500
x
2000
2500
0
0
3000
4000
4000
3500
3500
3000
3000
2500
2500
2000
2000
z
z
As nonlinear nonhydrostatic time-dependent experiment, we consider
in this section the test case proposed in [23]. It consists in the evolution
of a warm bubble placed in an isentropic atmosphere at rest. All data
are as in [23]. Contours of potential temperature perturbation at
different times are shown in figure 19. These results were obtained
using 64 × 80 elements, pπ = 4, pu = 5 and a timestep ∆t = 0.5 s,
corresponding to a Courant number Csnd ≈ 17.
1500
1500
1000
1000
500
500
0
0
500
1000
1500
x
2000
2500
3000
0
0
500
1000
1500
x
2000
2500
3000
500
1000
1500
x
2000
2500
3000
Figure 19: Contours (every 0.2 K and the zero contour is omitted) of perturbation potential temperature in the rising thermal bubble test at time 10
min, 14 min, 15 min and 16 min respectively in clockwise sense.
41
8
Conclusions and future perspectives
We have introduced an accurate and efficient discretization approach
for typical model equations of atmospheric flows. We have extended
to spherical geometry the techniques proposed in [52], combining a
semi-Lagrangian approach with the TR-BDF2 semi-implicit time discretization method and with a spatial discretization based on adaptive
discontinuous finite elements. The resulting method is unconditionally
stable and has full second order accuracy in time, thus improving standard off-centered trapezoidal rule discretizations without any major
increase in the computational cost nor loss in stability, while allowing
the use of time steps up to 100 times larger than those required by
stability for explicit methods applied to corresponding DG discretizations. The method also has arbitrarily high order accuracy in space
and can effectively adapt the number of degrees of freedom employed
in each element in order to balance accuracy and computational cost.
The p−adaptivity approach employed does not require remeshing and
is especially suitable for applications, such as numerical weather prediction, in which a large number of physical quantities is associated
to a given the mesh.
Furthermore, although the proposed method can be implemented
on arbitrary unstructured and nonconforming meshes, like reduced
Gaussian grids employed by spectral transform models, even in applications on simple Cartesian meshes in spherical coordinates the
p−adaptivity approach can cure effectively the pole problem by reducing the polynomial degree in the polar elements, yielding a reduction
in the computational cost that is comparable to that achieved with reduced grids. Numerical simulations of classical shallow water and nonhydrostatic benchmarks have been employed to validate the method
and to demonstrate its capability to achieve accurate results even at
large Courant numbers, while reducing the computational cost thanks
to the adaptivity approach. The proposed numerical framework can
thus provide the basis of for an accurate and efficient adaptive weather
prediction system.
42
Acknowledgements
This research work has been supported financially by the The Abdus Salam International Center for Theoretical Physics, Earth System Physics Section. We are extremely grateful to Filippo Giorgi of
ICTP for his strong interest in our work and his continuous support.
Financial support has also been provided by the INDAM-GNCS 2012
project Sviluppi teorici ed applicativi dei metodi Semi-Lagrangiani and
by Politecnico di Milano. We would also like to acknowledge useful
conversations on the topics of this paper with C. Erath, F. X. Giraldo,
M. Restelli, N. Wood.
References
[1] M. Baldauf and S. Brdar. An analytic solution for linear gravity
waves in a channel as a test for numerical models using the nonhydrostatic, compressible Euler equations. Quarterly Journal of
the Royal Meteorological Society, 139:1977–1989, 2013.
[2] R.E. Bank, W.M. Coughran, W. Fichtner, E.H. Grosse, D.J.
Rose, and R.K. Smith. Transient Simulation of Silicon Devices
and Circuits. IEEE Transactions on Electron Devices., 32:1992–
2007, 1985.
[3] J.R. Bates, F.H.M. Semazzi, R.W. Higgins, and S.R.M. Barros.
Integration of the Shallow Water Equations on the Sphere Using a Vector Semi-Lagrangian Scheme with a Multigrid Solver.
Monthly Weather Review, 118:1615–1627, 1990.
[4] L. Bonaventura. A semi-implicit, semi-Lagrangian scheme using
the height coordinate for a nonhydrostatic and fully elastic model
of atmospheric flows. Journal of Computational Physics, 158:186–
213, 2000.
[5] L. Bonaventura, R. Redler, and R. Budich. Earth System
Modelling 2: Algorithms, Code Infrastructure and Optimisation.
Springer Verlag, New York, 2012.
[6] J. Butcher and D. Chen. A new type of singly-implicit RungeKutta method. Applied numerical mathematics, 34:179–188,
2000.
[7] V. Casulli and E. Cattani. Stability, accuracy and efficiency of
a semi-implicit method for three-dimensional shallow water flow.
Computational Mathematics and Applications, 27:99–112, 1994.
43
[8] J. Coté. A Lagrange multiplier approach for the metric terms of
semi-Lagrangian models on the sphere. Quarterly Journal of the
Royal Meteorological Society, 114:1347–1352, 1988.
[9] J. Coté and A. Staniforth. A Two-Time-Level Semi-Lagrangian
Semi-implicit Scheme for Spectral Models. Monthly Weather Review, 116:2003–2012, 1988.
[10] M.J.P. Cullen. A test of a semi-implicit integration technique for
a fully compressible non-hydrostatic model. Quarterly Journal of
the Royal Meteorological Society, 116:1253–1258, 1990.
[11] T. Davies, M.J.P. Cullen, A.J. Malcolm, M.H. Mawson, A. Staniforth, A.A. White, and N. Wood. A new dynamical core for the
Met Office’s global and regional modelling of the atmosphere.
Quarterly Journal of the Royal Meteorological Society, 131:1759–
1782, 2005.
[12] C.N. Dawson, J.J. Westerink, J.C. Feyen, and D. Pothina. Continuous, Discontinuous and coupled Discontinuous-Continuous
Galerkin finite element methods for the shallow water equations.
International Journal of Numerical Methods in Fluids, 52:63–88,
2006.
[13] F. Desharnais and A. Robert. Errors near the poles generated by
a semi-Lagrangian integration scheme in a global spectral model.
Atmosphere-Ocean, 28:162–176, 1990.
[14] M. Dumbser and V. Casulli. A staggered semi-implicit spectral discontinuous Galerkin scheme for the shallow water equations. Applied Mathematics and Computation, 219(15):8057 –
8077, 2013.
[15] A.E. Gill. Atmospheric-Ocean Dynamics. Academic Press, 1987.
[16] F.X. Giraldo. Trajectory computations for spherical geodesic
grids in cartesian space. Monthly Weather Review, 127:1651–
1662, 1999.
[17] F.X. Giraldo, J.S. Hesthaven, and T. Warburton. High-Order
Discontinuous Galerkin methods for the spherical shallow water
equations. Journal of Computational Physics, 181:499–525, 2002.
[18] F.X. Giraldo, J.F. Kelly, and E.M. Constantinescu. Implicitexplicit formulations of a three-dimensional nonhydrostatic unified model of the atmosphere (NUMA). SIAM Journal of Scientific Computing, 35(5):1162–1194, 2013.
[19] F.X. Giraldo and M. Restelli. High-order semi-implicit timeintegrators for a triangular discontinuous Galerkin oceanic shallow water model. International Journal of Numerical Methods in
Fluids, 63(9):1077–1102, 2010.
44
[20] M. Hortal and A. Simmons. Use of reduced Gaussian grids in
spectral models. Monthly Weather Review, 119:1057–1074, 1991.
[21] M.E. Hosea and L.F. Shampine. Analysis and implementation of
TR-BDF2. Applied Numerical Mathematics, 20:21–37, 1996.
[22] R. Jakob-Chien, J. Hack, and D. Williamson. Spectral transform
solutions to the shallow water test set. Journal of Computational
Physics, 119:164187, 1995.
[23] R. L. Carpenter Jr., K. K. Droegemeier, P. R. Woodward, and
C. E. Hane. Application of the Piecewise Parabolic Method
(PPM) to Meteorological Modeling. Monthly Weather Review,
118:586–612, 1990.
[24] C. T. Kelley. Iterative Methods for Linear and Nonlinear Equations. SIAM, Philadelphia, 1995.
[25] J.F. Kelly and F.X. Giraldo. Continuous and discontinuous Galerkin methods for a scalable three-dimensional nonhydrostatic
atmospheric model: Limited-area mode. Journal of Computational Physics, 231(24):7988–8008, 2012.
[26] C. Kennedy and M. Carpenter. Additive Runge-Kutta schemes
for convection-diffusion-reaction equations. Applied Numerical
Mathematics, 44:139–181, 2003.
[27] J.D. Lambert. Numerical methods for ordinary differential systems. Wiley, 1991.
[28] M. Läuter, F.X. Giraldo, D. Handorf, and K. Dethloff. A discontinuous Galerkin method for the shallow water equations in spherical triangular coordinates. Journal of Computational Physics,
227(24):10226–10242, 2008.
[29] M. Läuter, D. Handorf, and K. Dethloff. Unsteady analytical
solutions of the spherical shallow water equations. Journal of
Computational Physics, 210:535,553, 2005.
[30] D.Y. Le Roux. Spurious inertial oscillations in shallow-water
models. Journal of Computational Physics, 231:7959–7987, 2013.
[31] D.Y. Le Roux and G.F. Carey. Stability-dispersion analysis of
the discontinuous Galerkin linearized shallow-water system. International Journal of Numerical Methods in Fluids, 48:325–347,
2005.
[32] R. LeVeque. Finite difference methods for ordinary and partial
differential equations: steady-state and time-dependent problems.
Society for Industrial and Applied Mathematics, 2007.
45
[33] A. McDonald and J.R. Bates. Semi-Lagrangian Integration of a
Gridpoint Shallow Water Model on the Sphere. Monthly Weather
Review, 117:130–137, 1989.
[34] J.L. McGregor. Economical determination of departure points for
semi-Lagrangian models. Monthly Weather Review, 121:221–330,
1993.
[35] K. W. Morton. On the analysis of finite volume methods for
evolutionary problems. SIAM Journal of Numerical Analysis,
35:2195–2222, 1998.
[36] K. W. Morton, A. Priestley, and E. Süli. Stability of the
Lagrange-Galerkin scheme with inexact integration. RAIRO
Modellisation Matemathique et Analyse Numerique, 22:625–653,
1988.
[37] K. W. Morton and E. Süli. Evolution-Galerkin methods and their
supraconvergence. Numerische Mathematik, 71:331–355, 1995.
[38] R. D. Nair, S.J. Thomas, and R.D. Loft. A Discontinuous Galerkin Global Shallow Water Model. Monthly Weather Review,
133:876–888, 2005.
[39] R. D. Nair, S.J. Thomas, and R.D. Loft. A Discontinuous Galerkin transport scheme on the cubed sphere. Monthly Weather
Review, 133:814–828, 2005.
[40] A. Priestley. Exact Projections and the Lagrange-Galerkin Method: A Realistic Alternative to Quadrature. Journal of Computational Physics, 112:316–333, 1994.
[41] M. Restelli, L. Bonaventura, and R. Sacco. A semi-Lagrangian
Discontinuous Galerkin method for scalar advection by incompressible flows. Journal of Computational Physics, 216:195–215,
2006.
[42] M. Restelli and F.X. Giraldo. A conservative Discontinuous Galerkin semi-implicit formulation for the Navier-Stokes equations
in nonhydrostatic mesoscale modeling. SIAM Journal of Scientific Computing, 31:2231–2257, 2009.
[43] P. Ripodas, A. Gassmann, J. Förstner, D. Majewski, M. Giorgetta, P. Korn, L. Kornblueh, H. Wan, G. Zängl, L. Bonaventura,
and T. Heinze. Icosahedral Shallow Water Model (ICOSWM):
results of shallow water test cases and sensitivity to model parameters. Geoscientific Model Development, 2:231–251, 2009.
[44] H. Ritchie. Application of the Semi-Lagrangian Method to
a Spectral Model of the Shallow Water Equations. Monthly
Weather Review, 116:1587–1598, 1988.
46
[45] G. Rosatti, L. Bonaventura, and D. Cesari. Semi-implicit, semiLagrangian environmental modelling on cartesian grids with cut
cells. Journal of Computational Physics, 204:353–377, 2005.
[46] Y. Saad and M.H. Schultz. GMRES: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM
Journal on Scientific and Statistical Computing, 7:856–869, 1986.
[47] W.C. Skamarock and J.B. Klemp. Efficiency and accuracy of the
Klemp-Wilhelmson time-splitting technique. Monthly Weather
Review, 122:2623–2630, 1994.
[48] A. Staniforth, A.A. White, and N. Wood. Treatment of vector equations in deep-atmosphere semi-Lagrangian models.I: Momentum equation. Quarterly Journal of the Royal Meteorological
Society, 136:497–506, 2010.
[49] C. Temperton, M. Hortal, and A. Simmons. A two-time-level
semi-Lagrangian global spectral model. Quarterly Journal of the
Royal Meteorological Society, 127:111–127, 2001.
[50] J. Thuburn and Y. Li. Numerical simulations of Rossby-Haurwitz
waves. Tellus A, 52:181189, 2000.
[51] J. Thuburn and A.A. White. A geometrical view of the
shallow-atmosphere approximation, with application to the semilagrangian departure point calculation. Quarterly Journal of the
Royal Meteorological Society, 139:261268, 2013.
[52] G. Tumolo, L. Bonaventura, and M. Restelli. A semi-implicit,
semi-Lagrangian, p-adaptive discontinuous Galerkin method for
the shallow water equations. Journal of Computational Physics,
232:46–67, January 2013.
[53] R.A. Walters. Numerically induced oscillations in finite-element
approximations to the shallow-water equations. International
Journal of Numerical Methods in Fluids, 3:591–604, 1983.
[54] R.A. Walters and G.F. Carey. Analysis of spurious oscillation
modes for the shallow-water and Navier-Stokes equations. Computers and Fluids, 11:51–68, 1983.
[55] D.L. Williamson, J.B. Drake, J.J. Hack, R. Jacob, and P.N.
Swarztrauber. A Standard Test Set for the Numerical Approximations to the Shallow Water Equations in Spherical Geometry.
Journal of Computational Physics, 102:211–224, 1992.
[56] O.C. Zienkiewicz, J.P.Gago, and D.W. Kelly. The hierarchical
concept in finite element analysis. Computers and Structures,
16:53–65, 1983.
47
| 5cs.CE
|
Fundamental Diagram of Rail Transit
and Its Application to Dynamic Assignment
arXiv:1708.02147v1 [cs.SY] 7 Aug 2017
Toru Seo∗† Kentaro Wada‡ Daisuke Fukuda§
August 8, 2017
Abstract
Urban rail transit often operates with high service frequencies to serve heavy passenger demand
during rush hours. Such operations can be delayed by train congestion, passenger congestion, and
the interaction of the two. Delays are problematic for many transit systems, as they become
amplified by this interactive feedback. However, there are no tractable models to describe
transit systems with dynamical delays, making it difficult to analyze the management strategies of
congested transit systems in general, solvable ways. To fill this gap, this article proposes simple yet
physical and dynamic models of urban rail transit. First, a fundamental diagram of a transit system
(3-dimensional relation among train-flow, train-density, and passenger-flow) is analytically derived
by considering the physical interactions in delays and congestion based on microscopic operation
principles. Then, a macroscopic model of a transit system with time-varying demand and supply is
developed as a continuous approximation based on the fundamental diagram. Finally, the accuracy
of the macroscopic model is investigated using a microscopic simulation, and applicable range of
the model is confirmed.
Keywords: public transport; rush hour; fundamental diagram; kinematic wave theory; MFD; dynamic
traffic assignment
1
Introduction
Urban rail transit, such as metro systems, plays a significant role in handling the transportation needs
of metropolitan areas (Vuchic 2005). Its most notable usage is the morning commute, in which
heavy passenger demand is focused into a short time period. To obtain general policy implications
for management strategies of transit systems (e.g., pricing, gating, scheduling), many studies have
theoretically analyzed such situations under certain simplifications, such as the static travel time of
transit operations (de Cea and Fernández 1993; Tabuchi 1993; Kraus and Yoshida 2002; Tian et al.
2007; Gonzales and Daganzo 2012; Trozzi et al. 2013; de Palma et al. 2015a,b).
It is known that urban mass transit often suffers from delays caused by congestion, even if no
serious incidents or accidents occur (Kato et al. 2012; Tirachini et al. 2013; Kariyazaki et al. 2015).
This means that the dynamical aspect of transit systems is important during periods of congestion.
∗
Corresponding author. [email protected]
Tokyo Institute of Technology, 2-12-1-M1-20, O-okayama, Meguro, Tokyo 152-8552, Japan
‡
Institute of Industrial Science, The University of Tokyo, Cw504, 4-6-1 Komaba, Meguro, Tokyo 153-8505, Japan
§
Tokyo Institute of Technology, 2-12-1-M1-11, O-okayama, Meguro, Tokyo 152-8552, Japan
†
1
For instance, in the Tokyo metropolitan area (TMA), which is one of the most populated regions in
the world, rail transit systems are essential and operated with high service frequency (up to 30 trains
per hour per line; headway of two minutes) to serve the heavy passenger demand during peak hours
(Kariyazaki 2016). Unfortunately, even if there are no accidents, chronic delays occur almost daily and
passengers experience longer and unreliable travel times due to congestion. For example, the mean
delay of one of the major transit lines in Tokyo during the rush hour is about eight minutes, whereas
the standard deviation of the delay is about two minutes (Iwakura et al. 2013). Kariyazaki (2016)
estimated that, on a typical weekday, three million commuters across the entire TMA experience such
delays, and the social cost caused by the delay corresponds to 230 billion Japanese yen (approximately
2 billion USD) per year. Appropriate management strategies to solve this issue are therefore desirable.
In general, the following types of congestion are observed in urban rail transit:
• train-congestion: congestion involving consecutive trains using the same tracks, also known as
‘knock-on delay’ (Carey and Kwieciński 1994).
• passenger-congestion: congestion of passengers at station platforms, namely, bottleneck
congestion at the doors of a train while it is stopped at a station (Wada et al. 2012; Kariyazaki
et al. 2015).1
These two types of congestion interact with each other and cause delay (Newell and Potts 1964; Wada
et al. 2012; Kato et al. 2012; Tirachini et al. 2013; Kariyazaki et al. 2015; Cuniasse et al. 2015). For
example, passenger-congestion can prolong the time that a train spends at a station. This extended
dwell time interrupts the operation of subsequent trains, and causes train-congestion at times of high
service frequency. As passenger throughput deteriorates when train-congestion occurs, the passenger
congestion at stations intensifies—this is a short-term vicious cycle. The extreme case is known as
‘bunching’ (Newell and Potts 1964). Cuniasse et al. (2015) reported macroscopic-scale production
loss phenomena which occur almost daily in a railway system due to congestion. Moreover, in the long
term, such chronic delays could affect passengers’ departure time choice; Kato et al. (2012) observed
this phenomenon in TMA. Kim et al. (2015) also reported that route choice of metro passengers is
affected by congestion, crowding, and delay. Therefore, these congestion dynamics affect both the
within-day and day-to-day dynamics of transit systems.
For this reason, it would be preferable to consider these dynamical aspects of transit systems in
order to obtain general policy implications for transit management under heavy demand during rush
hours, similar to road traffic congestion problems (c.f., dynamic traffic assignment, Szeto and Lo 2006;
Iryo 2013). However, to the authors’ knowledge, no study has investigated such problems in transit
systems. In the aforementioned theoretical studies on transit commuting (i.e., de Cea and Fernández
1993, and others), the travel time of a transit system is assumed to be constant and/or determined by
static models, meaning that the dynamical aspect is neglected. One reason for this might be that we do
not have tractable models of transit systems that can consider the dynamics of delay and congestion.2
To fill this gap, this article proposes tractable models of the dynamics of urban rail transit
considering the physical interaction between train-congestion and passenger-congestion. The
1Note that passenger-congestion differs from in-vehicle passenger-crowding, which results in discomfort due to standing
and crowding, but is not necessarily cause any delay directly.
2Several operation models have been proposed by considering the detailed mechanism of such delay and congestion
(see Vuchic 2005; Koutsopoulos and Wang 2007; Parbo et al. 2016; Cats et al. 2016; Li et al. 2017; Alonso et al. 2017,
and references therein), and these have been used to develop efficient operation schemes. However, their purposes are
case-specific optimization and evaluation such as time-tabling; it would be difficult to use them to obtain general policy
implications for management strategies, as they are essentially complex and intractable.
2
remainder of this article is organized as follows. In Section 2, a simple and tractable operation
model of rail transit is formulated by considering train-congestion, passenger-congestion, and the
interaction between them. The model describes the theoretical relation among train-flow, train-density,
and passenger-flow under ideal conditions—that is, a fundamental diagram (FD). In Section 3, a
macroscopic loading model of a transit system in which demand and supply change dynamically is
developed based on the proposed FD. The model is based on a continuous approximation approach
with the FD, which is widely used for automobile traffic flows—that is, an exit-flow model. The model
is called macroscopic because it describes the aggregated behavior of trains and passengers in a certain
(e.g., urban-scale) spatial domain. In Section 4, the approximation accuracy and other properties of the
proposed macroscopic model are investigated through a comparison with the results of a microscopic
simulation. Section 5 concludes this article.
2
Fundamental Diagram of Rail Transit System
In this section, we analytically derive an FD of a rail transit system based on microscopic operation
principles. The FD is defined as the relation among train-flow, train-density, and passenger-flow.
2.1
Assumptions of the Rail Transit System
We assume two principles of rail transit operation, namely, the train’s dwell behavior at a station for
passenger boarding and the cruising behavior on the railroad. Note that they are equivalent to those
employed by Wada et al. (2012).
The passenger boarding time is modeled using a point-queue bottleneck model. That is, the
flow-rate of passenger boarding to a dwelling train is assumed to be constant, µp ; and there is a buffer
time (e.g., time required for door opening/closing), gb , for the dwell time. The dwell time of a train at
a station, tb , is expressed as
tb =
np
+ gb ,
µp
(1)
where np is the number of passengers waiting to board the train at the station.3 Passengers waiting
for a train at a station are assumed to board the first train that arrives. This means that the passenger
storage capacity of a train is assumed to be unlimited.
The cruising behavior of a train is modeled using Newell’s simplified car-following model (Newell
2002), which is a special case of the well-known road traffic flow model, the Lighthill–Whitham–
Richards model (Lighthill and Whitham 1955; Richards 1956). In this model, a vehicle travels as fast
as possible while maintaining the minimum safety clearance. Specifically, let xm (t) be the position of
train m at time t. It is described as
xm (t) = min {xm (t − τ ) + vf τ, xm−1 (t − τ ) − δ} ,
(2)
where m−1 indicates the preceding train, τ is the physical minimum headway time (similar to reaction
time of the vehicle), vf is the free-flow speed (i.e., maximum speed), and δ is the minimum spacing.
3 The parameter np can also be interpreted as the total number of passengers who are getting on and off the train. In
fact, this more general definition is preferable in some sense. However, it will complicate the following discussions; thus,
we neglect passengers getting off the train. Note that, by carefully distinguishing the two types of passengers, the following
discussions are valid and do not affect the final results.
3
The first term in the min operation indicates that the traffic is in the free-flowing regime, where the
train can travel at its maximum speed. The second term indicates the traffic is in the congested regime,
where the train catches the preceding one and is required to decrease its speed to maintain the safety
headway τ and distance δ simultaneously. In a critical regime, a train’s speed is vf and the train catches
the preceding one. Without loss of generality, we introduce a variable buffer headway time hf ≥ 0 to
describe traffic in the free-flow regime.
2.2
Steady State of Rail Transit System
Here, we consider the steady state of rail transit operation under the assumptions stated in Section 2.1.
The steady state is an idealized traffic state that does not change over time, and its traffic state variables
(typically combination of flow, density, and speed) are characterized by certain relation (called an FD)
of the traffic flow (e.g., Daganzo 1997). In the case of rail transit operation, the steady state can be
defined as a state that satisfies all of the following conditions:
•
•
•
•
•
The model parameters, namely, µp , gb , vf , τ, δ, are constant.
The distance between adjacent stations, l, is constant.
The headway time between successive trains, H, is constant.
The cruising speed v of all trains is the same.
The passenger-flow arriving to each station platform, qp , is the same.
Additionally, we assume that all trains stop at every station. In order to operate the transit system,
qp < µp is assumed; otherwise, passenger boarding will never end. Under the steady state, the dwell
time of a train at a station, represented by Eq. (1), can be transformed to qp H/µp + gb , because
np is equal to qp H. Note that the control strategy of the transit operation (e.g., schedule-based,
headway-based) need not be specified here, because reasonable control strategies should follow such
steady operation if there is no disturbance; otherwise, train bunching will occur (Wada et al. 2012).
Transit systems under different steady states are illustrated as time–space diagrams in Fig. 1, where
the horizontal axis indicates the time of day, the vertical axis indicates space, and the curves indicate
train trajectories. In each sub-figure, train m arrives at and departs from station i, then travels to station
i + 1 at cruising speed v, and finally arrives at station i + 1 under different conditions. In Fig. 1a, the
speed v is equal to the free-flow speed vf and hf is greater than zero; therefore, the state is classified
into the free-flowing regime. In Fig. 1b, the speed is equal to vf and hf is equal to zero; therefore, the
state is classified into the critical regime. In Fig. 1c, the speed is less than vf ; therefore, the state is
classified into the congested regime.
2.3
Fundamental Diagram
In general, the following can be considered as the traffic state variables of a rail transit system:
•
•
•
•
•
•
train-flow q,
train-density k,
train-mean-speed v̄,
passenger-flow qp ,
passenger-density kp ,
passenger-mean-speed v̄p .
4
Space x
Train m − 1
Train m
Station i + 1
δ
hf
τ
v
l
qp H/µp + gb
Station i
H
Time t
(a) Free-flowing regime: v = vf , hf > 0.
Space x
Train m − 1
Train m
Station i + 1
δ
τ
v
l
qp H/µp + gb
Station i
H
Time t
(b) Critical regime: v = vf , hf = 0.
Space x
Train m − 1
Station i + 1
Train m
δ
τ
v
l
qp H/µp + gb
Station i
H
Time t
(c) Congested regime: v < vf , hf = 0.
Figure 1: Time–space diagrams of rail transit system under steady states.
5
Among these, there are three independent variables: for example, the combination of q, k, and qp .
This is because of the identities in continuum flow, namely, q = kv̄ and qp = kp v̄p , and the identity
v̄ = v̄p .4
Now suppose that the relation among the independent variables of the traffic state under every
steady state can be expressed using a function Q as
q = Q(k, qp ).
(3)
The function Q can be regarded as an FD of the rail transit system. If the rail transit operation principle
follows Eqs. (1) and (2), the FD function can be specified as
lk − qp /µp
,
if k < k ∗ (qp ),
gb + l/vf
Q(k, qp ) =
(4)
lδ
−
(k − k ∗ (qp )) + q ∗ (qp ), if k ≥ k ∗ (qp ),
(l − δ)gb + τ l
with
1 − qp /µp
,
gb + δ/vf + τ
(l − δ)/vf − τ
gb + l/vf
k ∗ (qp ) = −
qp +
,
(gb + δ/vf + τ )µp l
(gb + δ/vf + τ )l
q ∗ (qp ) =
(5)
(6)
where q ∗ (qp ) and k ∗ (qp ) represent train-flow and train-density, respectively, at a critical state with
passenger-flow qp . For the derivation of Eqs. (4)–(6), see Appendix A.
2.4
2.4.1
Discussions
Features of Fundamental Diagram
The FD has the following features that can be derived analytically from Eq. (4). Note that they can
easily be found in the numerical example in Fig. 2.
The FD can be interpreted as a function that determines train-flow q and mean-speed v̄ = q/k
under a given train-density k (i.e., supply) and passenger-flow qp (i.e., demand) for the given
technical parameters of the transit system (i.e., µp , gb , vf , τ, δ, l). Although the FD equation (4)
looks complicated, it represents a simple relation: namely, a piecewise linear (i.e., triangular) relation
between q and k under fixed qp .
As mentioned, the traffic state of a transit system can be categorized into three regimes (free-flowing,
critical, and congested), as in the standard traffic flow theory. Therefore, there is a critical train-density
k ∗ (qp ) for any given qp . Train traffic is in the free-flowing regime if k < k ∗ (qp ), in the critical
regime if k = k ∗ (qp ), or in the congested regime otherwise. The congested regime can be considered
as inefficient compared with the free-flowing regime, because the congested regime takes more time
to transport the same amount of passengers. The critical regime is the most efficient in terms of
travel time (i.e., 1/v̄, 1/v̄p ) as well as passenger-crowding (i.e., number of passengers per train, qp /q).
However, the critical regime requires more trains (i.e., higher train- density) than the free-flowing
regime; therefore, it may not be the most efficient if the operation cost is taken into account.
4Note that the mean speed v̄ differs from the cruising speed v; the former takes the dwelling time at a station and cruising
between stations into account, whereas the latter only considers the cruising time.
6
Even in the critical regime, the mean speed v̄ is inversely proportional to passenger demand qp .
This means that travel time increases as passenger demand increases. In addition, the size of the
feasible area of (q, k) narrows as qp increases. Thus, the operational flexibility of the transit system
declines as the passenger demand increases.
Flow and density in the critical regime satisfy the following relations:
q ∗ (qp ) =
l
1
k ∗ (qp ) −
.
(l − δ)/vf − τ
(l − δ)/vf − τ
(7)
(Here, we have assumed (l − δ)/vf − τ 6= 0.) Therefore, the critical regime can be represented as a
straight line whose slope (l/[(l−δ)/vf −τ ]) is either positive or negative in the q–k plane. This implies
a qualitative difference between transit systems. Specifically, if the slope is positive, a transit operation
with constant train-density would transition from the free-flowing regime to the congested regime as
passenger demand increases (Fig. 2). On the contrary, if the slope is negative, such an operation would
transition from free-flowing to congested as passenger demand decreases. This seems paradoxical, but
it is actually reasonable because the operational efficiency can be degraded if the number of trains is
excessive compared to passenger demand.
Note that Eqs. (3) and (4) are consistent with Edie’s generalized definition (Edie 1963) of traffic
states. Therefore, the FD is consistent with the fundamental definition of traffic. For steady-state
transit operation, Edie’s traffic state is derived as5
1
,
H
qp H/µp + gb + l/v
k=
,
lH
l
v̄ =
.
qp H/µp + gb + l/v
q=
(8)
(9)
(10)
One can easily confirm that Eqs. (8)–(10) satisfy the FD equation.
2.4.2
Numerical Example
For ease of understanding, a numerical example is shown in Fig. 2; the parameter values are presented
in Table 1. In the figure, the horizontal axis represents train-density k, the vertical axis represents
train-flow q, and the plot color represents passenger-flow qp . The slope of the straight line from a
traffic state to the origin represents the mean speed v̄ of the state.
The features of the FD described in Section 2.4.1 can be easily confirmed. For example, the figure
can be read as follows. Suppose that the passenger demand per station is qp = 16000 (pax/h). If
the number of trains in the transit system is given by the train-density k = 0.3 (train/km), then the
resulting train traffic has a train-flow of q ' 15 (train/h) and a mean speed of v̄ ' 50 (km/h). This is
the traffic state in the free-flowing regime. There is a congested state corresponding to a free-flowing
state: for the aforementioned state with (q, k, v̄) ' (15 veh/h, 0.3 veh/km, 50 km/h), the corresponding
congested state is (15 veh/h, 0.55 veh/km, 27 km/h). The critical state under qp = 16000 (pax/h) is
(22 veh/h, 0.42 veh/km, 52 km/h). Notice that this state has the fastest average speed under the given
5These relations are derived by applying Edie’s definition to the ‘minimum component of the time–space diagram’ of
the steady state, which is a parallelogram-shaped area in Fig. 1 whose vertexes are time–space points of (i) train m departs
from station i, (ii) train m arrives at station i + 1, (iii) train m − 1 arrives at station i + 1, and (iv) train m − 1 departs from
station i.
7
Figure 2: Numerical example of the FD.
Table 1: Parameters of the numerical example.
parameter
u
τ
δ
µp
gb
l
value
70 km/h
1/70 h
1 km
36000 pax/h
10/3600 h
3 km
passenger demand. The triangular q–k relation mentioned before is clearly shown in the figure; the
‘left edge’ of the triangle corresponds to the free-flowing regime, the ‘top vertex’ corresponds to the
critical regime, and the ‘right edge’ corresponds to the congested regime.
2.4.3
Validity of the Assumptions
Here, we discuss the relation between the operation principles, the proposed FD, and an actual transit
system. First of all, it is worth mentioning that all parameters in the proposed model have an explicit
physical meaning. Therefore, the parameter calibration required to approximate an actual transit system
is relatively simple.
In the train cruising model (2), each train is assumed to maintain a headway that is greater than the
given minimum headway. This is reasonable; similar models with minimum headways have been used
in existing studies (Carey and Kwieciński 1994; Higgins and Kozan 1998; Huisman et al. 2005) to
analyze the effect of train congestion (i.e., knock-on delay). Additionally, the model can be considered
as the so-called ‘moving block control,’ which is one of the standard operation schemes for trains
(Wada et al. 2012).
Under the presence of well-designed adaptive control strategies, such as schedule-based and
headway-based control (Daganzo 2009; Wada et al. 2012), the steady state is likely to be realized. This
is because the aim of such adaptive control is usually to eliminate bunching—in other words, such
control makes the operation steady.
8
The passenger boarding model (1), namely, the bottleneck model, is a coarse approximation of
actual phenomena that would be fairly reasonable. This is because the capacity of a fixed-width
bottleneck for ordinary pedestrian flow is often considered to be constant (Lam et al. 1998;
Hoogendoorn and Daamen 2005). However, some observational studies have reported that, in heavily
crowded conditions, the boarding time could increase nonlinearly as passenger numbers increase,
probably due to passenger-crowding (interference between passengers and a lack of space in carriages)
(Harris 2006; Tirachini et al. 2013). Moreover, there is no stock capacity for passengers in the proposed
model. Therefore, states with excessively large qp and small q in the FD might correspond to unrealistic
situations. This is a limitation of the current model. Nevertheless, the scale of passenger-crowding
can be derived by the model: that is, qp /q represents the number of passengers per train.
2.4.4
Relation to the Macroscopic Fundamental Diagram
The proposed FD resembles the macroscopic fundamental diagram (MFD) (Geroliminis and Daganzo
2007; Daganzo 2007) and its extensions (e.g., Geroliminis et al. 2014; Chiabaut 2015). They are
similar in the following sense. First, they both consider dynamic traffic. Second, they both describe
the relations among macroscopic traffic state variables in which the traffic is not necessarily steady
or homogeneous at the local scale (i.e., they use area-wide aggregations based on Edie’s definition).
Third, they both have unimodal relations, meaning that there are free-flowing and congested regimes,
where the former has higher performance than the latter; in addition, there is a critical regime where
the throughput is maximized. Therefore, it is expected that existing approaches for MFD applications,
such as modeling, control and the optimization of transport systems (e.g., Daganzo 2007; Geroliminis
and Levinson 2009; Geroliminis et al. 2013; Fosgerau 2015), are also suitable for the proposed transit
FD.
However, there are substantial differences between the proposed FD and the existing MFD-like
concepts. In comparison with the original MFD (Geroliminis and Daganzo 2007; Daganzo 2007)
and its railway variant (Cuniasse et al. 2015), the proposed FD has an additional dimension, that
is, passenger-flow. In comparison with the three-dimensional MFD of Geroliminis et al. (2014),
which describes the relations among total traffic flow, car density, and bus density in a multi-modal
traffic network, the proposed FD explicitly models the physical interaction among the three variables.
In comparison with the passenger MFD of Chiabaut (2015), which describes the relation between
passenger flow and passenger density when passengers can choose to travel by car or bus, in the
proposed FD, passenger demand can degrade the performance (i.e., speed) of the vehicles because of
the inclusion of the boarding time.
3
Dynamic Model Based on Fundamental Diagram
Recall that the proposed FD describes the relationship among traffic variables under the steady state;
therefore, the behavior of a dynamical system in which demand and supply change over time is not
described by the FD itself. This feature is the same as in the road traffic FD and MFDs. In this section,
we formulate a model of urban rail transit operation where the demand (i.e., passenger-flow) and
supply (i.e., train-density) change dynamically. In the proposed model, individual train and passenger
trajectories are not explicitly described; therefore, the model is called macroscopic.
The proposed model is based on an exit-flow model (Merchant and Nemhauser 1978; Carey and
McCartney 2004) in which the proposed FD is employed as the exit-flow function. In other words,
the transit system is considered as an input–output system, as illustrated in Fig. 3. The exit-flow
9
modeling approach is often employed for area-wide traffic approximations and analysis using MFDs,
such as optimal control to avoid congestion (Daganzo 2007) and analyses of user equilibrium and
social optimum in morning commute problems (Geroliminis and Levinson 2009; Fosgerau 2015). The
advantage of this approach is that it may be possible to conduct mathematically tractable analysis
of dynamic, large-scale, and complex transportation systems, where the detailed traffic dynamics are
difficult to model in a tractable manner—this is the case for transit operations.
train in-flow: a(t)
its cumulative: A(t)
passenger in-flow: ap (t)
its cumulative: Ap (t)
Railway system
internal average train-flow: Q(k(t) ap (t))
dynamics of internal average train-density:
dk(t)
= a(t) − Q(k(t), ap (t))
dt
travel time: T (t)
train out-flow: Q(k(t), ap (t))
its cumulative: D(t)
passenger out-flow: dp (t)
(determined by the model)
its cumulative: Dp (t)
Figure 3: Railway system as an input–output system.
3.1
Formulation
At time t, let a(t) be the inflow of trains to the transit system, ap (t) be the inflow of passengers, d(t)
be the outflow of trains from the transit system, and dp (t) be the outflow of passengers. We set the
initial time to be 0, and therefore t ≥ 0. Let A(t), Ap (t), D(t), and Dp (t) be the cumulative values
Rt
of a(t), ap (t), d(t), and dp (t), respectively (e.g., A(t) = 0 a(s)ds). Let T (t) be the travel time of
a train that entered the system at time t, and let its initial value T (0) be given by the free-flow travel
time under q = a(0) and qp = ap (0). To simplify the formulation, the trip length of the passengers is
assumed to be equal to that of the trains.6 This means that T (·) is the travel time of both the trains and
the passengers. These functions can be interpreted as follows:
a(t): trains’ departure rate from their origin station at time t.
ap (t): passengers’ arrival rate at the platform of their origin station at time t.
d(t): trains’ arrival rate at their final destination station at time t.
dp (t): passengers’ arrival rate at their destination station at time t.
T (t): travel time of a train and passengers from origin (departs at time t) to destination. Note that the
arrival time at the destination is t + T (t).
Therefore, in reality, a(·) and ap (·) will be determined by the transit operation plan and passenger
departure time choice, respectively. d(·), dp (·), and T (·) are endogenously determined through the
operational dynamics.
In accordance with exit-flow modeling, the train traffic is modeled as follows. First, the exit flow
d(t) is assumed to be
d(t) = Q(k(t), ap (t))
(11)
6This assumption is reasonable if the average trip length is shared by trains and passengers. If they are different, a
modification such as Tp (t) = T (t)/λ, where λ is the ratio of average trip length of the passengers to that of the trains, would
be useful.
10
where the FD function Q(·) is considered to be an exit-flow function.7 This means that the dynamics
of the transit system are modeled by taking the conservation of trains into account as follows:
dLk(t)
= a(t) − Q(k(t), ap (t)),
dt
(12)
where L represents the length of the transit route. This exit-flow model has been employed in several
studies to represent the macroscopic behavior of a transportation system (e.g., Merchant and Nemhauser
1978; Carey and McCartney 2004; Daganzo 2007). Note that the average train-density k(t) can be
defined as
k(t) =
A(t) − D(t)
,
L
(13)
which is consistent with Eq. (12). Based on above functions and equations (i.e., T (0), a(t), ap (t), and
Eqs. (11) and (13)), d(t) and D(t) can be sequentially computed—in other words, the train traffic can
be computed using the initial and boundary conditions and the exit-flow model based on the FD.
The passenger traffic can be derived as follows. By the definition of the travel time of trains,
A(t) = D(t + T (t))
(14)
holds. As A(t) and D(t) have already been obtained, the travel time T (t) such that Eq. (14) holds
can be computed. Then, Dp (t) and dp (t) can be computed from the definition of the travel time of
passengers, which is also T (t):
Ap (t) = Dp (t + T (t)).
3.2
(15)
Discussion
The proposed macroscopic model computes train out-flow d(t) and passenger out-flow d(t) based on
the FD function Q(·), the initial and boundary conditions a(t), ap (t), and T (0). The notable feature
of the model is that it is highly tractable, as it is based on an exit-flow model. Therefore, we expect
the proposed model to be useful for analyzing various management strategies for transit systems (e.g.,
dynamic pricing during the morning commute).
The proposed model can accurately approximate the macroscopic behavior of a transit operation
with high-frequency operation (i.e., small headway time) under moderate changes in demand and/or
supply. This is because exit-flow models are reasonable when the changes in inflow are moderate
compared with the relaxation time of the dynamical system. Because such situations often occur in
busy metropolitan subway systems, which suffer from congestion and delay during rush hours because
of heavy demand, the model may be useful for investigating such congestion problems. However, the
accuracy of the model is expected to decrease if the operation has low steadiness, such as in the event of
train bunching. In the next section, the quantitative accuracy of the model is verified using numerical
experiments.
The model can also derive the social cost and benefit of a transit system. For example, the
generalized travel cost of passengers (e.g., travel time, schedule delay, crowding disutility) can be
calculated from Ap (t) and Dp (t). In addition, the operation cost of the transit system can be calculated
from A(t), D(t), and the FD parameters.
7If np is considered as the sum of the number of passengers who are boarding and alighting (as mentioned in note 3), we
can simply define d(t) to be equal to Q(k(t), ap (t) + dp (t)). Such a model is also computable using a similar procedure.
11
4
Verification of the Macroscopic Model
In this section, we verify the quantitative accuracy of the macroscopic model by comparing its results
with that of the microscopic model (i.e., Eqs. (1) and (2)). The validity of the macroscopic model can
be investigated by comparing its solutions with those of microscopic models using the same initial and
boundary conditions and model parameters.
4.1
Simulation Setting
The parameter values of the transit operation are listed in Table 1 for both the microscopic and
macroscopic models. The railroad is considered to be a one-way corridor. The stations are equally
spaced at intervals of l, and there are a total of 10 stations. Trains enter the railroad with flow a(t);
in the microscopic model, a discrete train enters the railroad from the upstream boundary station if
bA(t)c (i.e., integer part of A(t)) is incremented. In the microscopic model, trains leave the railroad
from the downstream boundary station without any restrictions, other than the passenger boarding and
minimum headway clearance. Passengers arrive at each station with flow ap (t).
The functions a(t) and ap (t) are exogenously determined to mimic morning rush hours, with each
having a peak at t = 2. The flow before the peak time increases monotonically, whereas the flow after
the peak time decreases monotonically—in other words, the so-called S-shaped A(t) and Ap (t) are
considered. These functions are specified as
if t < 2,
a(0) + (a(2) − a(0)) 2t ,
4−t
a(0) + (a(2) − a(0)) 2 , if 2 ≤ t < 4,
a(t) =
(16)
a(0),
if 4 ≤ t,
if t < 2,
ap (0) + (ap (2) − ap (0)) 2t ,
4−t
ap (0) + (ap (2) − ap (0)) 2 , if 2 ≤ t < 4,
(17)
ap (t) =
ap (0),
if 4 ≤ t,
where the values of a(0), a(2), ap (0), and ap (2) are given as scenario parameters. The simulation
duration is set to 4 h for the baseline scenario in Section 4.2.1 and to 8 h for the sensitivity analysis in
Section 4.2.2 (the reason will be explained later).
The microscopic model without any control is asymptotically unstable, as proven by Wada et al.
(2012); this means that time-varying demand and supply always cause train bunching, making the
experiment unrealistic and useless. Therefore, the headway-based control scheme proposed by Wada
et al. (2012) is implemented in the microscopic model to prevent bunching and stabilize the operation.
This scheme has two control measures: holding (i.e., extending the dwell time) and an increase in
free-flow speed, similar to Daganzo (2009). The former is activated by a train if its following train is
delayed, and can be represented as an increase in gb in the microscopic model. The latter is activated
by a train if it is delayed, and can be represented as an increase in vf up to a maximum allowable
speed vmax . In this experiment, vmax is set to 80 km/h and vf is 70 km/h. This control scheme can
be considered realistic and reasonable, as similar operations are executed in practice. See Appendix B
for further details of the control scheme.
4.2
Results
First, to examine how well the proposed model reproduces the behavior of the transit system under
time-varying conditions, the results for the baseline scenario are presented in Section 4.2.1. A
12
Figure 4: Result of the microscopic model in the baseline scenario.
(a) Train
(b) Passenger
Figure 5: Result of the macroscopic model in the baseline scenario.
sensitivity analysis of the demand/supply conditions is then conduced and applicable ranges of the
proposed model are investigated in Section 4.2.2.
4.2.1
Baseline scenario
The baseline scenario with parameter values a(0) = 10 (train/h), a(2) = 15 (train/h), ap (0) = 0.1µp
(pax/h), and ap (2) = 0.5µp (pax/h) is investigated first. A solution of the microscopic model is shown
in Fig. 4 as a time–space diagram. The colored curves represent the trajectories of each train traveling
in the upward direction while stopping at every station. Around the peak time period (t = 2), train
congestion occurs; namely, some of the trains stop occasionally between stations in order to maintain
the safety interval. The congestion is caused by heavy passenger demand; therefore, the situation
during rush hour is reproduced.
The result given by the macroscopic model is shown in Fig. 5 as cumulative plots. Fig. 5a shows
the cumulative curves for the trains, where the blue curve represents the inflow A and the red curve
represents the outflow D. Fig. 5b shows those of passengers in the same manner. Congestion and
delay can be observed around the peak period (it is more remarkable in the passenger traffic). For
example, during the peak time period, dp (t) is less than ap (t) and ap (t0 ), where t0 is time such that
t = t0 + T (t0 ). This means that the throughput of the transit system is reduced by the heavy passenger
demand. Consequently, T (t) is greater during peak hours than in off-peak periods such as T (0),
meaning that delays occur due to the congestion.
The macroscopic and microscopic models are compared in terms of the cumulative number of
13
Figure 6: Comparison between the macroscopic and microscopic models in the baseline scenario.
trains in Fig. 6. In the figure, the solid curves denote the macroscopic model and the dots denote the
microscopic model. It is clear that D in the macroscopic model follows that of the microscopic model
fairly precisely. For example, the congestion and delay during the peak time period are captured very
well. However, there is a slight bias: the macroscopic model gives a slightly shorter travel time. This is
mainly due to the large-scale unsteady state (i.e., train bunching) generated in the microscopic model;
the delay caused by such large-scale bunching cannot be recovered by the microscopic model under the
implemented headway-based control scheme (for details, see Appendix B). It means that if the control
is schedule-based, the bias could be reduced.
4.2.2
Sensitivity analysis of the demand/supply conditions
The accuracy of the macroscopic model regarding the dynamic patterns of demand/supply is now
examined. This is worth investigating quantitatively because it is qualitatively clear that the exit-flow
model is valid if the speed of demand/supply changes is ‘sufficiently’ small as discussed in Section 3.2.
Specifically, the sensitivity of the peak passenger demand ap (2) and train supply a(2) is evaluated by
assigning various values to these parameters. The simulation duration is set to 8 h to take the residual
delay after t = 4 (h) in some scenarios into account. The other parameters are the same as in the
baseline scenario.
The results are summarized in Fig. 7. Fig. 7a shows the relative difference in total travel time
(TTT) of trains between the microscopic and macroscopic models for various peak passenger flows
(negative values indicate that TTT of the macroscopic model is smaller). The relative difference can
be considered as an error index of the macroscopic model. Fig. 7b compares the absolute value of TTT
in each model. Note that there are some missing values, such as the relative error with a(2) = 18 and
ap (2)/µp = 0.60; this is due to that the macroscopic model does not derive a solution under the given
conditions (e.g., k(t) exceeds the jam density); this corresponds to ‘gridlock’ in the transportation
system.
According to the results in Fig. 7, the accuracy of the macroscopic model is high when the peak
passenger demand is low. These are the expected results, as the speed of demand change is slow in
these cases. TTT given by the macroscopic model is almost always less than that of the microscopic
model; this might be due to the aforementioned inconsistency between the steady state assumption of
the macroscopic model and headway-based control of the microscopic model.
As the peak passenger demand increases, the relative error increases gradually when the demand
is low, and increases suddenly when the demand exceeds a certain value. This sudden increase in the
14
(a) Relative error regarding passenger demand
(b) Absolute values of TTT
Figure 7: Comparison between the microscopic and macroscopic models under different
demand/supply conditions.
error is because of extraordinary large-scale train bunching in the microscopic model. As confirmed
by Fig. 7b, the absolute value of TTT in the microscopic model also exhibits a sudden increase when
the demand exceeds the certain value. This bunching often occurs in cases with excessive passenger
demand, such as ap (2) > µp /2. Such demand can be considered as unrealistically excessive, as the
dwell time of a train at a station is longer than the cruising time between adjacent stations in such
situations; this usually does not occur even in rush hours.
As for the sensitivity of the train supply a(·), there is a weak tendency for faster variations in supply
to cause larger errors. This is also the expected result.
From these results, we conclude that the proposed model is fairly accurate under ordinary passenger
demand, although it is not able to reproduce extraordinary and unrealistic situations (for daily travel)
with excessive train bunching. This might be acceptable for representing transit systems during usual
rush hours.
5
Conclusion
In this paper, the following three models of an urban rail transit system have been analyzed:
Microscopic model: A model describing the trajectories of individual trains and passengers based on
Newell’s car-following model and passenger boarding model. This is represented in Eqs. (1)
and (2), and can be solved using simulations.
Fundamental diagram: An exact relationship among train-flow, train-density, and passenger-flow in
the microscopic model under a steady state. This is represented in Eqs. (3)–(6). It is a closed-form
equation.
Macroscopic model: A model describing train and passenger traffic using an exit-flow model whose
exit-flow function is the FD. This is represented in Eqs. (11), (13), (14), and (15), and can be
solved using simple simulations.
The FD and macroscopic model are the original contributions of this study, whereas the microscopic
model was proposed by Wada et al. (2012). The microscopic model can be considered as a lower-order
15
approximation of an actual transit system. The FD represents the exact relation among steady state
traffic variables in the microscopic model. The macroscopic model can be considered as an FD-based
macroscopic approximation of the area-wide behavior of the microscopic model.
The FD itself implies several insights on transit system. In addition, according to the results
of the numerical experiment, the macroscopic model can reproduce the behavior of the microscopic
model accurately, except for cases with unrealistically excessive demands. Because of the simplicity,
mathematical tractability, and good approximation accuracy of the proposed FD and macroscopic
model in ordinary situations, we expect that they will contribute for obtaining general policy
implications on management strategies of rail transit systems, such as pricing and control for morning
commute problems.
Some improvements to the proposed model can be considerable. First, the model ignores in-vehicle
passenger-crowding. This could be solved using a nonlinear passenger boarding model instead of
Eq. (1), or by introducing a disutility crowding term into the departure time choice problem in the
macroscopic model, as in Tian et al. (2007); de Palma et al. (2015a). Second, the variability and
reliability of a transit system (e.g., robustness against unpredictable disturbances, travel time reliability
issues) are not considered by the proposed model. A stochastic extension of the model might be useful
for this problem. Third, the extension to cases with heterogeneity, such as spatially heterogeneous
station distributions and passenger demand, would make the model considerably more realistic.
As an application of the proposed model, the following morning commute problems are being
investigated by the authors: user equilibrium departure time choice problem (find the equilibrium
Ap (t) for a given A(t) and desired arrival time of passengers, Zp (t)), optimal demand control problem
(find Ap (t) such that the total travel cost is minimized for a given A(t) and Zp (t)), and optimal demand
and supply control problem (find Ap (t) and A(t) such that total travel cost is minimized for a given
Zp (t)). The solutions to these problems would provide general insights into both demand and supply
management strategies for transit systems (e.g., dynamic pricing, operation planning). Furthermore,
multimodal commuting problems combined with travel mode choices (e.g., trains modeled by the
proposed FD and cars and buses modeled by MFDs) are also considerable.
Acknowledgements
Part of this research was financially supported by a KAKENHI Grant-in-Aid for Scientific Research
(B) #17H03320.
A
Derivation of FD
This appendix describes derivation of the FD expressed in Eqs. (4)–(6). Consider a looped rail transit
system under steady state operation. Let L be the length of the railroad, S be the number of the
stations, M be the number of trains, H be the headway time of the operation, tb be the dwelling time of
a train at a station, tc be the cruising time of a train between adjacent stations, and qp be the passenger
demand flow rate per station. Note that the distance between adjacent stations l is L/S and the number
of passengers boarding a train at each station is qp H.
The headway time of the operation is derived as follows. The round trip time of a train in the
looped railroad is S(tb + tc ), and M trains pass the station during that time. Then, the identities
N H = S(tb + tc ) and
16
H=
gb + tc
M/S − qp /µp
(18)
hold. Moreover, by the definition of headway and Newell’s car-following rule, the headway time H
must satisfy
δ + vτ
+ hf .
v
(19)
gb + δ/v + τ
+ hf .
1 − qp /µp
(20)
H = tb +
This reduces to
H=
The q–k relation in a free-flowing regime is derived as follows. As the train-flow is 1/H and
train-density is M/L by definition, Eq. (18) can be transformed to
q(k) =
kl − qp /µp
.
gb + l/vf
The train-flow and train-density under a critical state, (q ∗ , k ∗ ), are derived as follows.
substituting v = vf and hf = 0 into Eq. (20) and using the identity q = kv̄, we obtain
1 − qp /µp
,
gb + δ/vf + τ
(1 − qp /µp )(gb + l/vf )
,
k ∗ = k0 +
(gb + δ/vf + τ )l
q∗ =
(21)
By
(22)
(23)
where k0 is the minimum train-density where the train-flow is zero, namely, k0 = qp /(µp l).
The q–k relation in a congested regime is derived as follows. First, the k–v relation in a congested
regime can be easily derived from the q–v relation (20) with hf = 0 and the identity q = kv̄:
k(v) = k0 +
(1 − qp /µp )(gb + l/v)
.
(gb + δ/v + τ )l
(24)
Now, consider dq/dk, which is identical to (dq/dv) · (dv/dk). This can be derived as
dq
lδ
=
,
dk (δ − l)gb − τ l
(25)
which is constant and negative; therefore, the q–k relation is linear in a congested regime. Then,
recalling that the linear q–k curve passes the point (q ∗ , k ∗ ) with a slope of dq/dk, the q–k relation in
a congested regime can be derived as
q(k) =
with
lδ
k + q0
(δ − l)gb − τ l
dq ∗
·k .
dk
Eqs. (4)–(6) are constructed based on Eqs. (21), (22), (23), (26), and (27).
q0 = q ∗ −
17
(26)
(27)
B
Adaptive control scheme in the microscopic model
This appendix briefly explains the adaptive control scheme for preventing train bunching, proposed by
Wada et al. (2012). This scheme consists of two control measures: holding at a station and increasing
the maximum speed during cruising.
First, the scheme modifies the buffer time for dwelling (originally defined as gb in Eq. (1)) of train
m at station i to
gb := max{0, gb − Em (i)}
(28)
Em (i) = (1 − α)εm (i) + αµp (εm (i) − εm−1 (i)) ,
(29)
with
where εm (i) ≡ tm (i) − Tm,i represents the delay, tm (i) represents the time at which train m arrives
at station i, Tm,i represents the scheduled time (i.e., without delay) at which train m should arrive at
station i, and α ∈ [0, 1] is a weighting parameter. This scheme represents a typical holding control
strategy, similar to the bunching prevention method of Daganzo (2009), which extends the dwelling
time of a vehicle if the headway to the preceding vehicle is too small and vice versa.
Second, the scheme modifies the free-flow cruising speed vf such that the interstation travel time
is reduced by
min l/vf − l/vmax , max{0, Em (i) − gb } .
(30)
This means that, in the event of a delay, the train tries to catch up by increasing its cruising speed
up to the maximum allowable speed vmax (which implies that the free-flow speed vf is a ‘buffered’
maximum speed).
Meanwhile, the proposed train operation model in this study does not have a schedule—it is a
frequency-based operation. Therefore, in this study, the scheduled headway in the scheme (Tm,i −
Tm−1,i ) is approximated by the planned frequency (1/a(tm (i))). Thus, we set α = 1 and substitute
Em (i) with
µp tm (i) − tm−1 (i) − 1/a(tm (i)) .
(31)
The stationary state of the operational dynamics under the original scheme is identical to the steady
state defined in Section 2.2. In the case of α < 1, the scheme makes the train operation asymptotically
stable, meaning that the operation schedule is robust to small disturbances. In the case of α = 1,
the scheme prevents the propagation and amplification of delay, but does not recover the original
schedule (the small ‘shift’ found in Fig. 6 is due to α = 1). Note that these control measures do not
interrupt passenger boarding or violate the safety clearance between trains, meaning that most of the
fundamental assumptions of the proposed FD are satisfied.
References
B. Alonso, J. C. Munoz, A. Ibeas, and J. L. Moura. A congested and dwell time dependent transit corridor
assignment model. Journal of Advanced Transportation, 2017.
18
M. Carey and A. Kwieciński. Stochastic approximation to the effects of headways on knock-on delays of trains.
Transportation Research Part B: Methodological, 28(4):251–267, 1994.
M. Carey and M. McCartney. An exit-flow model used in dynamic traffic assignment. Computers & Operations
Research, 31(10):1583–1602, 2004.
O. Cats, J. West, and J. Eliasson. A dynamic stochastic model for evaluating congestion and crowding effects in
transit systems. Transportation Research Part B: Methodological, 89:43–57, 2016.
N. Chiabaut. Evaluation of a multimodal urban arterial: The passenger macroscopic fundamental diagram.
Transportation Research Part B: Methodological, 81:410–420, 2015.
P.-A. Cuniasse, C. Buisson, J. Rodriguez, E. Teboul, and D. de Almeida. Analyzing railroad congestion in a
dense urban network through the use of a road traffic network fundamental diagram concept. Public Transport,
7(3):355–367, 2015.
C. F. Daganzo. Fundamentals of Transportation and Traffic Operations. Pergamon Oxford, 1997.
C. F. Daganzo. Urban gridlock: Macroscopic modeling and mitigation approaches. Transportation Research
Part B: Methodological, 41(1):49–62, 2007.
C. F. Daganzo. A headway-based approach to eliminate bus bunching: Systematic analysis and comparisons.
Transportation Research Part B: Methodological, 43(10):913–921, 2009.
J. de Cea and E. Fernández. Transit assignment for congested public transport systems: an equilibrium model.
Transportation Science, 27(2):133–147, 1993.
A. de Palma, M. Kilani, and S. Proost. Discomfort in mass transit and its implication for scheduling and pricing.
Transportation Research Part B: Methodological, 71:1–18, 2015a.
A. de Palma, R. Lindsey, and G. Monchambert. The economics of crowding in public transport. Working Paper
(hal-01203310), 2015b.
L. C. Edie. Discussion of traffic stream measurements and definitions. In J. Almond, editor, Proceedings of the
2nd International Symposium on the Theory of Traffic Flow, pages 139–154, 1963.
M. Fosgerau. Congestion in the bathtub. Economics of Transportation, 4:241–255, 2015.
N. Geroliminis and C. F. Daganzo. Macroscopic modeling of traffic in cities. In Transportation Research Board
86th Annual Meeting, 2007.
N. Geroliminis and D. M. Levinson. Cordon pricing consistent with the physics of overcrowding. In W. H. K.
Lam, S. C. Wong, and H. K. Lo, editors, Transportation and Traffic Theory 2009, pages 219–240. Springer,
2009.
N. Geroliminis, J. Haddad, and M. Ramezani. Optimal perimeter control for two urban regions with macroscopic
fundamental diagrams: A model predictive approach. IEEE Transactions on Intelligent Transportation
Systems, 14(1):348–359, 2013.
N. Geroliminis, N. Zheng, and K. Ampountolas. A three-dimensional macroscopic fundamental diagram for
mixed bi-modal urban networks. Transportation Research Part C: Emerging Technologies, 42:168–181,
2014.
E. J. Gonzales and C. F. Daganzo. Morning commute with competing modes and distributed demand: User
equilibrium, system optimum, and pricing. Transportation Research Part B: Methodological, 46(10):1519–
1534, 2012.
19
N. G. Harris. Train boarding and alighting rates at high passenger loads. Journal of Advanced Transportation,
40(3):249–263, 2006.
A. Higgins and E. Kozan. Modeling train delays in urban networks. Transportation Science, 32(4):346–357,
1998.
S. P. Hoogendoorn and W. Daamen. Pedestrian behavior at bottlenecks. Transportation Science, 39(2):147–159,
2005.
D. Huisman, L. G. Kroon, R. M. Lentink, and M. J. C. M. Vromans. Operations research in passenger railway
transportation. Statistica Neerlandica, 59(4):467–497, 2005.
T. Iryo. Properties of dynamic user equilibrium solution: existence, uniqueness, stability, and robust solution
methodology. Transportmetrica B: Transport Dynamics, 1(1):52–67, 2013.
S. Iwakura, I. Takahashi, and S. Morichi. A multi agent simulation model for estimating knock-on train delays
under high-frequency urban rail operation. In Transport Policy Studies’s Review, volume 15. Institute For
Transport Policy Studies, 2013. (in Japanese).
K. Kariyazaki. Investigation of train delay recovery mechanism and delay prevention schemes in urban railway.
PhD thesis, National Graduate Institute for Policy Studies, 2016. (in Japanese).
K. Kariyazaki, N. Hibino, and S. Morichi. Simulation analysis of train operation to recover knock-on delay
under high-frequency intervals. Case Studies on Transport Policy, 3(1):92–98, 2015.
H. Kato, Y. Kaneko, and Y. Soyama. Departure-time choices of urban rail passengers facing unreliable service:
Evidence from Tokyo. In Proceedings of the International Conference on Advanced Systems for Public
Transport 2012, 2012.
K. M. Kim, S.-P. Hong, S.-J. Ko, and D. Kim. Does crowding affect the path choice of metro passengers?
Transportation Research Part A: Policy and Practice, 77:292–304, 2015.
H. Koutsopoulos and Z. Wang. Simulation of urban rail operations: Application framework. Transportation
Research Record: Journal of the Transportation Research Board, 2006:84–91, 2007.
M. Kraus and Y. Yoshida. The commuter’s time-of-use decision and optimal pricing and service in urban mass
transit. Journal of Urban Economics, 51(1):170–195, 2002.
W. H. K. Lam, C. Y. Cheung, and Y. F. Poon. A study of train dwelling time at the Hong Kong mass transit
railway system. Journal of Advanced Transportation, 32(3):285–295, 1998.
S. Li, M. M. Dessouky, L. Yang, and Z. Gao. Joint optimal train regulation and passenger flow control strategy
for high-frequency metro lines. Transportation Research Part B: Methodological, 99:113–137, 2017.
M. J. Lighthill and G. B. Whitham. On kinematic waves. II. a theory of traffic flow on long crowded roads.
Proceedings of the Royal Society of London. Series A. Mathematical and Physical Sciences, 229(1178):
317–345, 1955.
D. K. Merchant and G. L. Nemhauser. A model and an algorithm for the dynamic traffic assignment problems.
Transportation Science, 12(3):183–199, 1978.
G. F. Newell. A simplified car-following theory: a lower order model. Transportation Research Part B:
Methodological, 36(3):195–205, 2002.
G. F. Newell and R. B. Potts. Maintaining a bus schedule. In Proceedings of the 2nd Australian Road Research
Board, volume 2, 1964.
20
J. Parbo, O. A. Nielsen, and C. G. Prato. Passenger perspectives in railway timetabling: A literature review.
Transport Reviews, 36(4):500–526, 2016.
P. I. Richards. Shock waves on the highway. Operations Research, 4(1):42–51, 1956.
W. Y. Szeto and H. K. Lo. Dynamic traffic assignment: properties and extensions. Transportmetrica, 2(1):
31–52, 2006.
T. Tabuchi. Bottleneck congestion and modal split. Journal of Urban Economics, 34(3):414–431, 1993.
Q. Tian, H.-J. Huang, and H. Yang. Equilibrium properties of the morning peak-period commuting in a
many-to-one mass transit system. Transportation Research Part B: Methodological, 41(6):616–631, 2007.
A. Tirachini, D. A. Hensher, and J. M. Rose. Crowding in public transport systems: effects on users, operation
and implications for the estimation of demand. Transportation Research Part A: Policy and Practice, 53:
36–52, 2013.
V. Trozzi, G. Gentile, M. G. H. Bell, and I. Kaparias. Dynamic user equilibrium in public transport networks
with passenger congestion and hyperpaths. Transportation Research Part B: Methodological, 57:266–285,
2013.
V. R. Vuchic. Urban Transit: Operations, Planning, and Economics. John Wiley & Sons, 2005.
K. Wada, S. Kil, T. Akamatsu, and M. Osawa. A control strategy to prevent delay propagation in high-frequency
railway systems. Journal of Japan Society of Civil Engineers, Ser. D3 (Infrastructure Planning and
Management), 68(5):I_1025–I_1034, 2012. (in Japanese; extended abstract in English was presented
at the 1st European Symposium on Quantitative Methods in Transportation Systems and available at
https://www.researchgate.net/publication/281823577).
21
| 3cs.SY
|
arXiv:1707.07726v1 [math.NT] 24 Jul 2017
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC
GROUPS
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
Abstract. In this paper, we formulate an analogue of Waring’s problem for an algebraic group G. At the field level we consider a morphism
of varieties f : A1 → G and ask whether every element of G(K) is the
product of a bounded number of elements f (A1 (K)) = f (K). We give
an affirmative answer when G is unipotent and K is a characteristic zero
field which is not formally real.
The idea is the same at the integral level, except one must work with
schemes, and the question is whether every element in a finite index
subgroup of G(O) can be written as a product of a bounded number of
elements of f (O). We prove this is the case when G is unipotent and O
is the ring of integers of a totally imaginary number field.
0. Introduction
The original version of Waring’s problem asks whether, for every positive
integer n there exists M := Mn such that every non-negative integer is of
the form an1 + · · · + anM , ai ∈ N, and, if so, what is the minimum value for
Mn . Since 1909, when Hilbert proved that such a bound exists, an enormous
literature has developed, largely devoted to determining Mn . There is also
a substantial literature devoted to variants of Waring’s problem. Kamke
proved [Ka] a generalization of the theorem in which nth powers are replaced
by general polynomials. In a series of papers [W1, W2, W3], Wooley solved
Waring’s problem for vector-valued polynomials. Siegel [Si, Si1] treated the
case of rings of integers in number fields, and since then, many papers have
analyzed Waring’s problem for a wide variety of rings, for instance [Bi, Ca,
Vo, GV, LW, Ch, El]. Also, there has been a flurry of recent activity on
“Waring’s problem for groups”; the typical problem here is to prove that
every element in G is a product of a small number of nth powers of elements
of G (see, for instance, [Sh, LST, AGKS, GT] and the references therein.)
This paper explores the view that algebraic groups are the natural setting
for Waring’s problem. To this extent, it resembles the work on Waring’s
problem for groups of Lie type. The work on the polynomial-valued and
vector-valued variants of Waring’s problem also fit naturally in this framework. We will consider morphisms of varieties (resp. schemes) f : A1 → G
defined over a field (resp. a number ring) and look at bounded generation of
the groups generated by the images.
ML was partially supported by NSF Grant DMS-1401419.
1
2
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
The strategy is developed in §2 for unipotent algebraic groups over fields of
characteristic 0 which are not formally real. (Some justification for concentrating on the unipotent case is given in Lemma 1.4 below and the following
remarks.) In §3, we solve the unipotent version of Waring’s problem for totally imaginary number rings. In §4, we work over general characteristic 0
fields and general number rings, but consider only the “easier Waring’s problem”, in which one is allowed to use inverses. Our methods throughout are
elementary. The only input from analytic number theory is Siegel’s solution
of Waring’s problem over number rings.
Unfortunately, in the original situation of Waring’s problem, namely the
ring Z, the additive group Ga , and the morphism f : A1 → Ga given by
f (x) = xn , our results fall short of Hilbert’s theorem; we can prove only the
easier Waring’s problem in this case, rather than the statement that every
positive integer can be represented as a bounded sum of non-negative nth
powers. The difficulty, of course, is the ordering on Z. It seems natural
to ask whether, for unipotent groups over general number rings, one can
characterize the set which ought to be expressible as a bounded product of
images. In proving the easier Waring’s problem, we simply avoid this issue.
1. Generating subvarieties
Throughout this paper, K will always be a field of characteristic 0, and
G will be an algebraic group over K. A variety over K will be a reduced
separated scheme of finite type and, in particular, need not be connected.
A subvariety, closed subgroup, etc., will always be understood to be defined
over K.
Definition 1.1. Let G be an algebraic group over a field K. A subvariety
X of G is generating if there exists n such that every generic point of G lies
in the image of the product map from X ×n := X × · · · × X to G. A finite
collectionSfi : Xi → G of morphisms is generating if the union of Zariski
closures i f (Xi ) is generating.
We have the following necessary and sufficient condition for a subvariety
to be generating.
Proposition 1.2. Let G be an algebraic group over K and Z ⊆ G a closed
subvariety. Then Z is generating if and only if it satisfies the following two
properties:
I. Z is not contained in any proper closed subgroup of G.
II. For every proper closed normal subgroup H of G, the image of Z →
G/H has positive dimension.
We first prove the following technical lemma:
Lemma 1.3. Let K be algebraically closed. Let X, Y be irreducible closed
subvarieties of G. Assume:
(1) dim(XX) = dim X;
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
3
(2) dim(XY X) = dim X.
Then there exists a closed subgroup H of G such that the following statements
are true:
(i) X = xH = Hx for all x ∈ X(K);
(ii) Y ⊂ yH for some y ∈ NG (H)(K).
Proof. As X is irreducible, X ×2 is irreducible, with generic point η. The
closure X 2 = XX of its image in G is therefore the closure of the image
of η in G and thus irreducible. If x ∈ X(K), then xX and Xx are closed
subvarieties of X 2 of dimension dim X = dim X 2 , so xX = X 2 = Xx. Thus,
Xx−1 = x−1 X.
It follows that for x1 , x2 ∈ X(K),
−1 −1 2
−1
−1
−1 2 −1
−1
−1
x−1
1 X = x1 (x2 X ) = (x1 X)(x2 X) = (x1 X )x2 = Xx2 = x2 X.
Defining H := x−1 X for x ∈ X(K), we see that H does not depend on the
choice of x and, moreover, that H 2 = H. As every h ∈ H(K) can be written
−1 = x−1 x ∈ H(K). Thus H(K) is
x−1
1
1 x2 , x1 , x2 ∈ X(K), it follows that h
2
a subgroup, which since K is algebraically closed implies that H is a closed
subgroup of G, which implies (i).
For y ∈ Y (K), XY X is connected and contains XyX, which has dimension ≥ dim X = dim XY X. Thus,
XY X = XyX = HxyxH
is connected and has dimension dim H. It follows that the double coset
HxyxH consists of a single left coset, so xyx ∈ NG (H)(K). By (i), x also
normalizes H, and it follows that y normalizes H. Finally,
Y ⊆ x−1 XY Xx−1 = (x−1 Hx)y(x−1 Hx) = HyH = yH.
Using this, we can prove Proposition 1.2.
Proof. Clearly, if Z ⊆ H ( G, then the same is true for Z n , and if the
image of Z in G/H is finite, the same is true for Z n . This proves necessity
of conditions I and II.
For the sufficiency, we may assume without loss of generality that K is
algebraically closed. For all z ∈ Z(K), ZZ n ⊆ Z n+1 , so dim Z n is a bounded
non-decreasing sequence of integers. It therefore stabilizes for some n. Let
X denote an irreducible component of Z n of dimension dim Z n and Y any
irreducible component of Z. Then dim X ≤ dim X 2 ≤ dim Z n and dim X ≤
dim XY X ≤ dim Z n , so conditions (1) and (2) of Lemma 1.3 are satisfied.
Let H be the closed subgroup of G satisfying (i) and (ii). As H is a translate
of X, it is irreducible.
If H = G◦ , then X is a connected component of G, which means Z m
is a union of components of G whenever m ≥ n. Applying condition I to
subgroups of G containing G◦ , it follows that every generic point of G lies
4
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
in Z m for some m ≥ n, so Z is generated, as claimed. If H 6= G◦ < then
dim H < dim G. If H is normal in G, then the image of Z in G/H is finite,
contrary to condition II. If H has normalizer N ( G, then Y is contained
in N . Since N does not depend on the choice of component Y , Z ⊆ N ,
contrary to condition I.
Henceforth, we assume G is connected. We are interested in generating
collections of morphisms A1 → G. By a theorem of Chevalley, Barsotti,
and Rosenlicht [Ro], every connected algebraic group G has a closed normal
subgroup H which is a linear algebraic group and such that G/H is an
abelian variety. Every map from a rational curve to an abelian variety is
trivial. Thus, unless G is a linear algebraic group, it is impossible for any
collection of morphisms A1 → G to be generating.
Let R and U denote respectively the radical and the unipotent radical of
G.
Lemma 1.4. If U ( R, then there does not exist a generating set of morphisms A1 → G.
Proof. It suffices to prove that there is no generating set of morphisms from
A1 to the connected reductive group G/U . Thus, we may assume without
loss of generality that G is connected reductive. If the radical R is non-trivial,
then the inclusion map R → G induces an isogeny of tori R → G/[G, G], so
it suffices to prove that there no generating set of morphisms from A1 to a
non-trivial torus T . Without loss of generality, we may assume that K = K̄.
Thus we may replace T by a quotient isomorphic to the multiplicative group,
and it suffices to prove there is no non-constant morphism of curves from A1
to A1 \ {0}. At the level of coordinate rings, this is the obvious statement
that every K-homomorphism from K[t, t−1 ] to K[x] maps t to an element
of K ∗ , or, equivalently, the fact that K[x]∗ = K ∗ .
We need only consider, then, the case that G is the extension of a semisimple group by a unipotent group, both connected. The semisimple case is perhaps even more interesting, but we know that, at least for SL2 , we cannot
always expect bounded generation, since, for example, SL2 (Z) and SL2 (Z[i])
do not have bounded generation by elementary matrices [GS, Ta].
Since the characteristic of K is 0, if G is unipotent, it is necessarily connected [DG, IV, §2, Prop. 4.1]. The derived group G′ is then likewise unipotent [DG, IV, §2, Prop. 2.3], and therefore connected. The quotient G/G′
is unipotent [DG, IV, §2, Prop. 2.3], and commutative and is therefore a
vector group [DG, IV, §2, Prop. 4.2]. The non-abelian Galois cohomology
group H 1 (K, G′ ) vanishes [Se, III Prop. 6], so the cohomology sequence for
the short exact sequence 1 → G′ → G → G/G′ → 1 [Se, I Prop. 38] implies (G/G′ )(K) = G(K)/G′ (K). We identify these groups. We do not
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
5
distinguish between closed (vector) subgroups of G/G′ at the level of algebraic groups over K and the corresponding K-subspaces of the vector space
(G/G′ )(K). If V is a subspace of G/G′ , we denote by V G′ the inverse image
of V in G, regarded as an algebraic group.
Lemma 1.5. Let G be a connected unipotent algebraic group, and let H be
a proper closed subgroup of G. Then the normalizer of H in G is strictly
larger than H.
Proof. We use induction on dim G. The case dim G = 1 is trivial since this
implies G is commutative, so the normalizer of every subgroup is all of G.
For general unipotent G, the fact that the lower central series goes to {1}
implies that the center Z of G is of positive dimension. If Z is not contained
in H, then ZH ⊆ NG (H) is strictly larger than G. Otherwise, replacing G
and H by G/Z and H/Z respectively, we see that H/Z is normal in N/Z
for some N ⊆ G strictly larger than H, so H is normal in N .
Proposition 1.6. If G is a unipotent group over K, then every proper closed
subgroup H of G is contained in a normal subgroup N of codimension 1 in
G which contains the derived group G′ of G.
Proof. As K is of characteristic zero, G is connected. If H = {e}, the proposition asserts that G contains a codimension 1 normal subgroup containing
G′ . As H is a proper subgroup, G is non-trivial, so G/G′ is non-trivial,
and the proposition amounts to the obvious statement that every non-trivial
vector group contains a normal subgroup of codimension 1.
For the general case, applying the previous lemma, we can replace H by
a strictly larger group NG (H) unless H is normal in G. This operation
can be repeated only finitely many times, since NG (H), being strictly larger
than H, must be of strictly higher dimension (since every closed subgroup
of a unipotent group is unipotent and therefore connected). Thus, we may
assume H is normal in G. Then G/H is unipotent. Replacing G and H by
G/H and {e} respectively, we are done.
From Proposition 1.2, we deduce that for unipotent groups, we have the
following simple criterion.
Lemma 1.7. Let G be a unipotent group over K. A subvariety X of G
(resp. a set {f1 , . . . , fn } of morphisms A1 → G) is generating if and only if
for each proper subspace V ( G/G′ , such that the projection of X to G/V G′
is of positive dimension (resp. the composition of some fi with the projection
G → G/V G′ is non-constant.)
Note that the question of whether a set of morphisms fi is generating
depends only on the set of compositions f¯i of fi with the quotient map
G → G/G′ . It is also invariant under left or right translation of the fi by
any element of G(K).
6
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
Lemma 1.8. If {f1 , . . . , fn } is not generating, then for all positive integers
N,
(f1 (K) ∪ · · · ∪ fn (K))N ( G(K).
Proof. The image of (f1 (K) ∪ · · · ∪ fn (K))N in (G/V G′ )(K) is the same
as the image of {f1 (0), . . . , fn (0)}N and is therefore a finite subgroup of an
infinite group.
We record the following lemma, which will be needed later.
Lemma 1.9. Let G be a unipotent group over K, G′ its derived group and
G′′ the derived group of G′ . If V is a proper subspace of G′ /G′′ , then there
exists a dense open subvariety U1 ⊂ G and for all γ1 ∈ U1 (K), a dense open
subvariety U2 of the form G \ W G′ , such that for all γ2 ∈ U2 (K), [γ1 , γ2 ]
does not lie in V G′′ (K).
Proof. Without loss of generality, we assume K is algebraically closed. As
the characteristic of K is 0, G and G′ are connected, so G′ /G′′ is connected.
The composition G × G → G′ /G′′ of the commutator map and the quotient
map has the property that its image generates G′ /G′′ and is therefore not
contained in V . It follows that the inverse image U of the complement of
V is dense and open in G × G. By Chevalley’s theorem, the projection of
U ⊆ G × G onto the first factor G is a constructible set containing the
generic point; it therefore contains an open dense U1 . The fiber over any
point γ1 ∈ U1 (K) is non-empty. The condition on γ2 that [γ1 , γ2 ] 6∈ V G′ is
linear on the image of γ2 in G/G′ and is satisfied for at least one γ2 , so U2 ,
defined by the condition [γ1 , γ2 ] 6∈ V G′ satisfies the properties claimed.
2. The unipotent Waring Problem over nonreal fields
Definition 2.1. We say a field K is nonreal if it is of characteristic zero
but not formally real (i.e., −1 is a sum of squares in K).
The main theorem of the section is the following:
Theorem 2.2. If G is a unipotent algebraic group over a nonreal field K
and {f1 , . . . , fn } is a generating set of K-morphisms A1 → G, then for some
positive integer M ,
(f1 (K) ∪ · · · ∪ fn (K))M = G(K).
The proof occupies the rest of this section. It depends on the following
two propositions:
Proposition 2.3. Theorem 2.2 holds when G is a vector group.
Proposition 2.4. Under the hypotheses of Theorem 2.2, there exists an
integer m, a sequence of elements g1 , . . . , gm ∈ G(K), a sequence of positive integers k1 , . . . , km , for each i ∈ {1, . . . , m}, a sequence of integers
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
7
ℓi,1 , . . . , ℓi,ki ∈ [1, n] and of ai,j , bi,j ∈ K, such that for each i ∈ {1, . . . , m},
the K-morphisms h1 , . . . , hm : A1 → G defined by
hi (x) := gi fℓi,1 (ai,1 x + bi,1 ) · · · fℓi,ki (ai,ki x + bi,ki )
map A1 → G′ and as morphisms to G′ are generating.
Assuming both propositions hold, we can prove Theorem 2.2 by induction
on dimension. If G is commutative, then Proposition 2.3 applies. Otherwise, we apply Proposition 2.4 to construct h1 , . . . , hm . Letting f¯i denote
the composition of fi with G → G/G′ , Proposition 2.3 asserts that every
element of G(K)/G′ (K) = (G/G′ )(K) is represented by a bounded product
of elements of f¯1 (K) ∪ · · · ∪ f¯n (K). For each gi , there exists gi′ , which is a
bounded product of elements of f1 (K) ∪ · · · ∪ fn (K) and lies in the same
G′ (K)-coset of G(K). Defining
h′i (x) := gi′ fℓi,1 (ai,1 x + bi,1 ) · · · fℓi,ki (ai,ki x + bi,ki ),
it suffices to prove that every element of G′ (K) is a bounded product of
elements of h′1 (K) ∪ · · · ∪ h′m (K). As the hi are generating for G′ , the same
is true for h′i , and the theorem follows by induction. Thus, we need only
prove Propositions 2.3 and 2.4.
To prove Proposition 2.3, we begin with a special case.
Proposition 2.5. If K is a characteristic zero field which is not formally
real and d is a positive integer, there exists an integer N > 0 such that every
vector in K d is a sum of elements of
X1d := {(x, x2 , . . . , xd ) | x ∈ K}.
Proof. For each integer k > 0, let
Xkd := X1d + · · · + X1d .
{z
}
|
k
S
Thus
Xid +
pd+1 :
π d+1 : X d+1 → K the projection map onto the last coordinate. In particular,
pd+1 (Xkd+1 ) = Xkd for all positive integers k.
It is a theorem [El, Theorem 2] that for each positive integer d there exists
M such that every element in K is the sum of M dth powers of elements of
K.
We proceed by induction on d. The theorem is trivial for d = 1. Assume
d = X d = K d and every
it holds for d and choose M large enough that XM
element of K is the sum of M (d + 1)st powers. In particular,
X1d ⊆ X2d ⊆ · · · , and we denote by X d the limit i Xid . Clearly
d
d . Taking unions, X d is a semiring. Let
Xjd = Xi+j
and Xid Xjd ⊆ Xij
X d+1 → X d denote the projection map onto the first d coordinates and
(1)
d+1
) = K.
π d+1 (XM
d+1
d+1
d+1
d+1
( XM
If XM
+1 , then choosing w ∈ XM +1 \ XM , there exists an eled+1
d+1
is chosen with
such that pd+1 (v) = pd+1 (w). If u ∈ XM
ment v ∈ XM
8
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
pd+1 (u) = −pd+1 (v), then either u + v or u + w is non-zero, and either way
d+1
d+1 (t) = 0. By (1),
there exists an element t ∈ X2M
+1 \ {0} with p
d+1
{(0, . . . , 0, x) | x ∈ K} ⊂ XM
(2M +1) ,
d+1
d+1 , and we are done.
so by the induction hypothesis, XM
+M (2M +1) = K
d+1
d+1
d+1
= X d+1 .
= XM
We may therefore assume XM
+1 , which implies XM
Moreover, if pd+1 fails to be injective, the same argument applies, so we
may assume that pd+1 is an isomorphism of semirings, and therefore an
isomorphism of rings (since the target K d is a ring).
Thus, we can regard π d+1 ◦(pd+1 )−1 as a ring homomorphism φ : K d → K.
If ei = (0, . . . , 0, 1, 0, . . . , 0), then φ(ei ) maps to an idempotent of K, which
can only be 0 or 1. Since φ(e1 + · · · + ed ) = 1, there exists i such that
φ(ei ) = 1, and it follows that φ factors through projection onto the ith
coordinate. Thus, there exists a ring endomorphism ψ : K → K such that
for all (x1 , . . . , xd+1 ) ∈ X d+1 , ψ(xi ) = xd+1 . As (2, 4, 8, . . . , 2d+1 ) ∈ X d+1 ,
ψ(2i ) = 2d+1 , which is absurd.
We now prove Proposition 2.3.
Proof. Let
(2)
fj (x) = (P1j (x), . . . , Pmj (x)),
where d is the maximum of the degrees of the Pij for 1 ≤ i ≤ m, 1 ≤ j ≤ n.
Let N be chosen as in Proposition 2.5. We write
(3)
Pij (x) =
d
X
aijk xk
k=0
for 1 ≤ i ≤ m, 1 ≤ j ≤ n. Given c1 , . . . , cn , our goal is to find xjℓ ∈ K,
1 ≤ j ≤ n, 1 ≤ ℓ ≤ N that satisfy the system of equations
X
(4)
aijk xkjℓ = ci , i = 1, 2, . . . , m.
j,k,ℓ
By Proposition 2.5, by choosing xjℓ suitably, we can choose the values
yjk =
N
X
xkjℓ
ℓ=1
independently for 1 ≤ j ≤ n and 1 ≤ k ≤ d, while yj0 = N by definition.
Thus, we can rewrite the system of equations (4) as
(5)
d X
n
X
k=1 j=1
aijk yjk = ci − N
n
X
aij0 , i = 1, 2, . . . , m.
j=1
This is always solvable unless there is a non-trivial relation among the linear
forms on the left hand side in this system, i.e., a non-zero sequence b1 , . . . , bm
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
such that
m
X
9
bi aijk = 0
i=1
for all j and k ≥ 1. If this is true, then
m
X
X
bi Pij =
bi aij0 , j = 1, . . . , n.
i=1
i
In other words, defining π(t1 , . . . , tn ) := b1 t1 + · · · + bm tm , π ◦ fj is constant
for all j, contrary to assumption.
Finally, we prove Proposition 2.4.
Proof. Suppose we have already constructed h1 , . . . , hr . Let h̄i denote the
composition of hi with the projection G′ → G′ /G′′ . Let W denote the
vector space spanned by the set {h̄i (t) − h̄i (0) | 1 ≤ i ≤ r, t ∈ K}. Suppose
W = G′ /G′′ . Then for all proper subspaces W ′ ( W there exist i and t such
that h̄i (0) and h̄i (t) represent different classes in W/W ′ . It follows that the
composition of hi with the projection G′ → G′ /W ′ G′′ is non-constant, and
therefore, {h1 , . . . , hr } is a generating set of morphisms to G′ .
Thus, we may assume that W is a proper subspace of G′ /G′′ . We apply
Lemma 1.9 to deduce the existence of γ1 ∈ G(K) and a proper closed subspace V of G/G′ such that for all γ2 ∈ G(K) \ V G′ (K), the commutator of
γ1 and γ2 is not in W G′′ (K). Let f¯i (x) denote the composition of fi (x) with
the quotient map G → G/G′ . As the fi are generating, there exists i such
that for all but finitely many values x ∈ K, f¯i (x) is not in V . Without loss
of generality, we assume i = 1.
By Proposition 2.3, there exists a bounded product
g = fs1 (b1 ) · · · fsM (bM ), si ∈ {1, 2, . . . , n}, bi ∈ K
such that
f¯s1 (b1 ) + · · · + f¯sM (bM ) = γ̄1 .
We write
f¯1 (x) = v0 + xv1 + · · · + xd vd ,
with vi ∈ K m . By Proposition 2.5, there exist a1 , . . . , aN ∈ K, not all zero,
such that
n
X
aki = 0, k = 1, 2, . . . , d.
(6)
i=1
Thus,
f¯1 (a1 x) + f¯2 (a2 x) + f¯3 (a3 x) + · · · + f¯N (aN x) = N v0 ,
which means that x 7→ f1 (a1 x)f2 (a2 x) · · · fN (an x) goes to a constant G′ coset. Without loss of generality, we may assume a1 6= 0.
Let gr+1 be an element of G(K) which can be realized as a product of at
most N values of fi at elements of K and such that
(7)
gr+1 fs1 (b1 ) · · · fbM (bM )f1 (a1 x)f2 (a2 x) · · · fN (an x)
10
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
belongs to G′ (K) for x = 0. By Proposition 2.3, such a gr+1 exists. We
choose hr+1 (x) to be either (7) or
(8)
gr+1 f1 (a1 x)fs1 (b1 ) · · · fsM (bM )f2 (a2 x) · · · fN (an x),
Either way, hr+1 (0) ∈ G′ (K). By (6), hr+1 (x) ∈ G′ (K) for all x ∈ K.
Because the commutator [f1 (a1 x), fs1 (b1 ) · · · fsM (bM )] lies in W G′′ for at
most finitely many x-values, at least one of (7) and (8) is non-constant (mod
W G′′ ). By induction on the codimension of W , the proposition follows.
3. The unipotent Waring Problem over totally imaginary
number rings
In this section K denotes a totally imaginary number field, O its ring of
integers, and G a closed O-subscheme of the group scheme Uk of unitriangular
k × k matrices. Thus the generic fiber GK will be a closed subgroup of Uk
over K and therefore unipotent. Moreover, there is a filtration of G(O) by
normal subgroups such that the successive quotients are finitely generated
free abelian groups. In particular, it is (by definition) a finitely generated
torsion-free nilpotent group, whose Hirsch number is the sum of the ranks
of these successive quotients.
A set {f1 , . . . , fn } of O-morphisms A1 → G is said to be generating if it is
so over K. The main theorem in this section is the following integral version
of Theorem 2.2:
Theorem 3.1. If {f1 , . . . , fn } is a generating set of O-morphisms A1 → G,
then for some positive integer M ,
(f1 (O) ∪ · · · ∪ fn (O))M
is a subgroup of finite index in G(O).
We begin by proving results that allow us to establish that some power of
a subset of a group Γ gives a finite index subgroup of Γ.
Lemma 3.2. Let Γ be a group, ∆ a finite index subgroup of Γ, Ξ a subset
of Γ, and M a positive integer. If ∆ ⊆ ΞM , then there exists N such that
ΞN is a finite index subgroup of Γ.
Proof. Without loss of generality, we assume that ∆ is normal in Γ. Consider
the (finite set)
{(m̄, γ̄) ∈ Z/M Z × Γ/∆ | m ∈ N, γ ∈ Ξm }.
For each element, choose a pair (m, γ), m ∈ N, γ ∈ Ξm representing it.
Choose A to be greater than all values m appearing in such pairs. Let N be
a multiple of M which is greater than A+M . For all positive integers k, ΞkN
is a union of cosets of ∆ in Γ and does not depend on k. The image of ΞN in
Γ/∆ is therefore a subset of a finite group and closed under multiplication.
It is therefore a subgroup, and the lemma follows.
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
11
Lemma 3.3. Let Γ be a finitely generated nilpotent group and ∆ a normal
subgroup of Γ. Then every finite index subgroup ∆0 of ∆ contains a finite
index subgroup ∆◦0 which is normal in Γ.
Proof. We prove there exists a function f : N → N depending only on Γ such
that for any normal subgroup ∆ of Γ, every subgroup ∆0 of ∆ of index n
contains a normal subgroup of Γ of index ≤ f (n) in ∆. Replacing ∆0 with
the kernel of the left action of ∆ on ∆/∆0 , we may assume without loss of
generality that ∆0 is normal in ∆. We prove the claim by induction on the
total number of prime factors of n.
If n = p is prime, it suffices to prove that there is an upper bound,
independent of ∆, on the number of normal subgroups of ∆ of index p. This
is true because intersecting a fixed central series of Γ = Γ0 ⊲ Γ1 ⊲ · · · with ∆
gives a central series of ∆, and every index p normal subgroup of ∆ is the
inverse image in ∆ of an index p subgroup of the finitely generated abelian
group ∆/∆ ∩ Γ1 ⊆ Γ/Γ1 .
If n has ≥ 2 prime factors, then for some prime factor p of n, ∆0 is a
normal subgroup of index n/p of a normal subgroup ∆p of index p in ∆. By
the induction hypothesis, ∆p contains a normal subgroup ∆◦p of Γ, of index
≤ f (p) in ∆. The index of ∆0 ∩ ∆◦p in ∆◦p divides n/p, and applying the
induction hypothesis, we deduce that the existence of a normal subgroup ∆◦0
of Γ of index ≤ maxi≤n/p f (i) in ∆◦p .
Proposition 3.4. Let Γ be a finitely generated nilpotent group, ∆ a normal
subgroup of Γ, Ξ a subset of Γ and M1 , M2 positive integers such that ΞM1
contains a finite index subgroup of ∆ and the image of ΞM2 in Γ/∆ contains
a finite index subgroup of Γ/∆. Then there exists M3 such that ΞM3 is a
finite index subgroup of Γ.
Proof. Let hΞi denote the subgroup of Γ generated by Ξ. The intersection
hΞi ∩ ∆ is of finite index in ∆, and the image hΞi∆/∆ is of finite index in
Γ/∆, so hΞi is of finite index in Γ. As a subgroup of a finitely generated
nilpotent group, it is also finitely generated and nilpotent. Replacing Γ and
∆ by hΞi and hΞi ∩ ∆ respectively, we assume without loss of generality that
Ξ generates Γ.
Replacing Ξ with ΞM1 , we may assume Ξ contains a finite index subgroup
∆0 of ∆. By Lemma 3.3, we may assume that ∆0 is a normal subgroup of
Γ. Let Ξ̄ denote the image of Ξ in Γ/∆0 . If Ξ̄N is a finite index subgroup of
Γ̄ := Γ/∆0 , then ΞN +1 is the inverse image of this subgroup in Γ, and the
¯ := ∆/∆0 , and Ξ̄ respectively,
proposition holds. Replacing Γ, ∆, Ξ by Γ̄, ∆
we reduce to the case that ∆ is finite. We need only show that if ΞM2 ∆
contains a finite index subgroup Γ0 of Γ, then ΞN is a finite index subgroup
of Γ for some N .
Replacing Ξ by ΞM2 ∩ Γ0 ∆, we may assume that Ξ ⊆ Γ0 ∆ and Ξ meets
every fiber of π : Γ0 ∆ → Γ0 ∆/∆. In particular, Ξ contains an element of ∆,
12
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
so replacing Ξ with Ξ|∆| , we may assume Ξ contains the identity, so
Ξ ⊆ Ξ2 ⊆ Ξ3 ⊆ · · ·
(9)
For i a positive integer, let mi denote the maximum over all fibers of π of
the cardinality of the intersection of the fiber with Ξi . Thus, the intersection
of every fiber of π with Ξi+1 is at least mi . Since fiber size is bounded
above by ∆, the sequence (9) must eventually stabilize. Replacing Ξ with
a suitable power, we have Ξ2 = Ξ. Thus Ξ is closed under multiplication.
As Ξ meets every fiber of π in the same number of points, γ ∈ Ξ implies
γ(Ξ ∩ π −1 (π(γ)−1 )) = Ξ ∩ ∆0, which implies γ −1 ∈ Ξ. Thus, Ξ is a subgroup
of Γ of bounded index.
Next, we prove a criterion for a subgroup of G(O) to be of finite index.
Proposition 3.5. Let Γ ⊆ G(O) ⊂ G(K). Then the Hirsch number hΓ of Γ
satisfies
(10)
hΓ ≤ [K : Q] dim GK .
If equality holds in (10), then Γ is of finite index in G(O).
Proof. Hirsch number is additive in short exact sequences. Let G0 := GK ,
and let G0 ) G1 ) · · · ) Gk = {1} be a central series. Then we have a
decreasing filtration of Γ by Γi := Γ ∩ Gi (K), and each quotient Γi /Γi+1
is a free abelian subgroup of Gi (K)/Gi+1 (K) ∼
= K dim Gi /Gi+1 . Every free
r
abelian subgroup of K has rank ≤ r[K : Q] with equality if and only if it is
commensurable with Or . This implies (10).
Applying the same argument to G(O), we get
hΓ ≤ hG(O) ≤ [K : Q] dim GK .
If equality holds in (10), then G(O)i /G(O)i+1 and its subgroup Γi /Γi+1 are
commensurable, and this implies that Γ is of index
k−1
Y
[G(O)i /G(O)i+1 : Γi /Γi+1 ] < ∞
i=0
in Γ.
We prove Theorem 3.1 by showing that h(f1 (O) ∪ · · · ∪ fn (O))M contains
a subset which is a group of Hirsch number [K : Q] dim GK . We first treat
the commutative case.
Proposition 3.6. Theorem 3.1 holds if G is commutative.
Proof. First we claim that for all d > 0 there exist integers L, M > 0 such
that
LOd ⊆ {(x1 + · · · + xM , x21 + · · · + x2M , . . . , xd1 + · · · + xdM ) | xi ∈ O}.
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
13
Since this is of finite index in Od , replacing M by a larger integer (also
denoted M ), we can guarantee that every element in the group generated by
{(x, x2 , . . . , xd ) | x ∈ O} can be written as a sum of M elements.
To prove the claim, we use Proposition 2.5 to show that each basis vector
ei is a sum of M elements of {(x, x2 , . . . , xd ) | x ∈ K}. Replacing each
x in the representation of ei by Dx for some sufficiently divisible positive
integer D, it follows that each ki ei can be written as a sum of M elements
of {(x, x2 , . . . , xd ) | x ∈ O} for suitable positive integers ki .
For each α ∈ O, we see from the (i − 1)th difference of αi (see [Wr,
Theorem 1, p.267]) that
i−1
X
1
m i−1
(−1)
i!α =
(α + m)i − (i − 1)i!.
(11)
m
2
m=0
Thus every element of i!O is in the subring O(i) of O generated by ith powers
of elements of O. A theorem of Siegel (see [Si1, Theorem VI]) implies that
there exist N1 , N2 , . . . such that every element of O(i) is a sum of Ni ith
powers of elements of O. Thus every element of i!O is a sum of Ni ith
powers of elements of O, and therefore every element of i!ki Oei is a sum of
M Ni elements of {(x, x2 , . . . , xd ) | x ∈ O}.
Letting L denote a positive integer divisible by 1!k1 , 2!k2 . . . , d!kd , and
replacing M by M (N1 + · · · + Nd ), we can write every element of LOd as a
sum of M elements of {(x, x2 , . . . , xd ) | x ∈ O}.
Restricting fj to the generic fiber, we can write it as a vector of polynomials (2) with the Pij given by (3). We can solve the system (4) of equations in
O whenever we can solve (5) in yjk ∈ LO. This
P system is always solvable in
K, so it is solvable in LO whenever the ci −N j aij0 is sufficiently divisible.
Thus, there exists an integer D such that if N and the ci are divisible by
D and N is sufficiently large, then (c1 , . . . , cm ) is a sum of N terms each of
which belongs to (f1 (O) ∪ · · · ∪ fn (O)). Let Λ := DOm .
Now, Λ ⊆ G(O) ⊂ G(K) ∼
= K m . As G(O) has a finite filtration whose
quotients are finitely generated free abelian groups, it must contain Λ as a
subgroup of finite index. Defining
Xi := f1 (O) ∪ · · · ∪ fn (O) + · · · + f1 (O) ∪ · · · ∪ fn (O) ,
|
{z
}
iN
we have Λ ⊆ Xi ⊆ G(O) for all i ≥ 1, and Xi+1 = X1 + Xi . It follows that
Xi+1 contains every Λ-coset in G(O) represented by any element of Xi , and
therefore the sequence X1 , X2 , . . . stabilizes to a subgroup of G(O) of rank
m[K : Q] and of finite index in G(O).
Now we prove Theorem 3.1
Proof. We first observe that Proposition 2.4 remains true over O; more precisely, assuming that the morphisms fi are defined over O, the elements gi
can be taken to be in G(O) and ai,j , bi,j ∈ O, so the morphisms hi are defined
14
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
over O. Instead of using Proposition 2.3, we use Proposition 3.6. The image
γ̄1 of the element γ1 guaranteed by Lemma 1.9 may not lie in the lattice
Λ ⊂ G(K)/G′ (K) ∼
= K m , but some positive integer multiple of γ̄1 will do so,
and the property of γ1 with respect to V is unchanged when it is replaced by
a non-trivial power. The elements ai guaranteed by Proposition 2.5 may not
lie in O, but again we can clear denominators by multiplying by a suitable
positive integer. The element gr+1 will exist as long as N v0 ∈ Λ. This can
be guaranteed by replacing N with a suitable positive integral multiple.
induction on dim G.
Now we proceed as in the proof of Theorem 2.2, usingS
By the induction hypothesis, there exists N such that ( i hi (O))N contains
a subgroup of G′ (K) of Hirsch number [K : Q] dim G′ . On the other hand,
by Proposition 3.6, there exists a bounded power of (f¯1 (O) ∪ · · · ∪ f¯n (O))
which contains a subgroup of (G/G′ )(K) of Hirsch number [K : Q] dim G/G′ .
Here for each 1 ≤ i ≤ n, f¯i (x) denotes the composition of fi (x) with the
quotient map G → G/G′ . The theorem follows from Proposition 3.4 and the
additivity of Hirsch numbers.
4. The easier unipotent Waring problem
We recall that the classical “easier Waring problem” [Wr] is to prove that
for every positive integer n there exists m such that every integer can be
written in the form ±an1 ± · · · ± anm , ai ∈ Z, and to determine the minimum
value of m for each n.
In this section, we prove unipotent analogues of the easier Waring problem
for arbitrary fields of characteristic zero and rings of integers of arbitrary
number fields:
Theorem 4.1. If G is a unipotent algebraic group over a field K of characteristic zero and {f1 , . . . , fn } is a generating set of K-morphisms A1 → G,
then for some positive integer M ,
M
[
(f1 (K)e1 ∪ · · · ∪ fn (K)em ) = G(K).
e1 ,...,en ∈{±1}
Theorem 4.2. Let K be a number field, O its ring of integers, and G a
closed O-subscheme of the group scheme Uk of unitriangular k × k matrices.
If {f1 , . . . , fn } is a generating set of O-morphisms A1 → G, then for some
positive integer M ,
M
[
f1 (O)e1 ∪ · · · ∪ fn (O)en
e1 ,...,en ∈{±1}
is a subgroup of bounded index in G(O).
The proof of Theorem 4.1 depends on variants of Propositions 2.3 and 2.4.
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
15
Proposition 4.3. If K is a field of characteristic zero and d is a positive
integer, there exists an integer N > 0 such that K d can be represented as
K d = (X1d + · · · + X1d ) − (X1d + · · · + X1d ),
{z
} |
{z
}
|
N
where
N
X1d := {(x, x2 , . . . , xd ) | x ∈ K}.
Proof. This is [IL, Theorem 3.2].
Proposition 4.4. Theorem 4.1 holds when G is a vector group.
Proof. Let
(12)
fj (x) = (P1j (x), . . . , Pmj (x)),
where d is the maximum of the degrees of the Pij for 1 ≤ i ≤ m and
1 ≤ j ≤ n. Let N be chosen as in Proposition 4.3. Writing
(13)
Pij (x) =
d
X
aijk xk
k=0
for 1 ≤ i ≤ m and 1 ≤ j ≤ n. Given (c1 , . . . , cm ), our goal is to find suitable
ǫjℓ ∈ {±1} and xjℓ ∈ K such that
(c1 , . . . , cm ) =
2N X
n
X
ǫjℓ fj (xjℓ ).
ℓ=1 j=1
In light of Proposition 4.3, for each 1 ≤ j ≤ n, one can let ǫjℓ = 1 if
1 ≤ ℓ ≤ N , and let ǫjℓ = −1 if N + 1 ≤ ℓ ≤ 2N . Thus the above system is
equivalent to the system of equations
!
2N
N
d
n X
X
X
X
(14)
xkjℓ , i = 1, . . . , m.
xkjℓ −
aijk
ci =
j=1 k=0
ℓ=N +1
ℓ=1
By Proposition 4.3, by choosing xjℓ ∈ K suitably, we can choose the values
yjk =
N
X
ℓ=1
xkjℓ
−
2N
X
xkjℓ
ℓ=N +1
independently for 1 ≤ j ≤ n and 1 ≤ k ≤ d, while yj0 = 0 by definition.
Thus we can rewrite the system of equations (14) as
(15)
d
n X
X
aijk yjk = ci , i = 1, . . . , m.
j=1 k=1
Arguing as in the proof of Proposition 2.3, we see that the above system
of equations is always solvable unless fj is constant modulo some proper
subspace V of Am for all 1 ≤ j ≤ n, i.e., each π ◦ fj is constant, where
16
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
π : Am → Am /V is the canonical projection. This is impossible since the set
of morphisms {f1 , . . . , fn } is generating.
Proposition 4.5. Under the hypotheses of Theorem 4.1, there exists an integer m, a sequence of elements g1 , . . . , gm ∈ G(K), a sequence of positive integers k1 , . . . , km , for each i ∈ {1, . . . , m}, a sequence of integers ℓi,1 , . . . , ℓi,ki ∈
[1, n], a sequence of integers ei,1 , . . . , ei,ki ∈ {±1}, and sequences of elements ai,1 , bi,1 , . . . , ai,ki , bi,ki ∈ K, such that for each i ∈ {1, . . . , m}, the
K-morphisms h1 , . . . , hm : A1 → G defined by
hi (x) := gi fℓi,1 (ai,1 x + bi,1 )ei,1 · · · fℓi,ki (ai,ki x + bi,ki )ei,ki
map A1 → G′ and as morphisms to G′ are generating.
Proof. Using Proposition 4.4, and the same arguments as in Proposition 2.4,
Proposition 4.5 follows immediately.
Proof of Theorem 4.1. The proof of Theorem 4.1 is the same as that of Theorem 2.2. Using Propositions 4.4 and 4.5, we proceed as in the proof of
Theorem 2.2, using induction on dim(G). Theorem 4.1 follows immediately.
Next, we prove an integral variant of Proposition 4.3, in greater generality
than we need for Theorem 4.2:
Proposition 4.6. Let O be any integral domain whose quotient field K is
of characteristic zero. For all positive integers d, there exist λ ∈ O \ {0} and
N ∈ Z>0 such that
λOd ⊆ (Y1d + · · · + Y1d ) − (Y1d + · · · + Y1d ),
{z
} |
{z
}
|
N
where
N
Y1d = {(x, x2 , . . . , xd ) | x ∈ O}.
Proof. For each integer k > 0, set
d
= (Y1d + · · · + Y1d ) − (Y1d + · · · + Y1d ) .
Yk,k
|
{z
} |
{z
}
k
k
Choose N > 0 as in Proposition 4.3. For each 1 ≤ m ≤ d, the basis vector
em can be written in the form
em
N
N
X
X
2
d
(yj , yj2 , . . . , yjd )
(xj , xj , . . . , xj ) −
=
j=1
j=1
for some xj , yj ∈ K. Replacing each xj , yj in the above representation by
δxj , δyj for some non-zero δ ∈ O, it follows that there exists a non-zero
κm ∈ O such that
(16)
d
κm em ∈ YN,N
.
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
17
For each α ∈ O, we apply (11) to prove that
d
m!κm Oem ⊆ Y2N
Nm ,2N Nm .
Let λ := d!
Qd
i=1 κi .
Replacing N by 2N (N 1 + · · · + Nd ), we deduce that
d
λOd ⊆ YN,N
.
The next result is a variant of Proposition 3.6.
Proposition 4.7. Theorem 4.2 holds if G is commutative.
Proof. Let λ, N be chosen as in Proposition 4.6. Restricting fj to the generic
fiber, we can write it as a vector of polynomials (12) with the Pij given by
(13). We can solve the system (14) of equations whenever we can solve
the system (15) in yjk ∈ λO. This system is always solvable in K; so
it is solvable in λO whenever the ci are sufficiently divisible. Thus there
exists an integer D such that if the ci are divisible by D and N is sufficiently
S large, then (c1 , . . . , cm ) is a sum of N terms, each of which belongs
to e1 ,...,en∈{±1} (e1 f1 (O) ∪ · · · ∪ en fn (O)). Let Λ := DOm .
Set
[
U=
(e1 f1 (O) ∪ · · · ∪ en fn (O)).
e1 ,...,en ∈{±1}
For each i ≥ 1, define
Xi = |U + ·{z
· · + U} .
iN copies of U
We have Λ ⊆ Xi ⊆ G(O) ⊂ G(K) ∼
= K m for all i ≥ 1, and Xi+1 = X1 +Xi .
Using the same arguments as in the proof of Proposition 3.6, Λ is a subgroup
of finite index in G(O), and therefore the sequence X1 , X2 , . . . stabilizes to a
subgroup of G(O) of rank m[K : Q], and of finite index in G(O).
We now prove Theorem 4.2.
Proof of Theorem 4.2. We first observe that Proposition 4.5 remains true
over O; more precisely, assuming that the morphisms fi are defined over
O, the elements gi can be taken to be in G(O) and ai,j , bi,j ∈ O, so the
morphisms hi are defined over O.
Now we proceed as in the proof of Theorem 4.1, using induction on
dim G.S By the induction hypothesis, there exists an integer N > 0 such
that ( e1 ,...,en∈{±1} (h1 (O)e1 ∪ · · · ∪ hn (O)en ))N is a subgroup of G′ (O) of
Hirsch number [K : Q] dimS
G′ . On the other hand, by Proposition 4.7, there
exists a bounded power of e1 ,...,en ∈{±1} (f¯1 (O)e1 ∪ · · · ∪ f¯n (O)en ) which is a
subgroup of (G/G′ )(O) of Hirsch number [K : Q] dim G/G′ . Here for each
1 ≤ i ≤ n, f¯i (x) denotes the composition of fi (x) with the quotient map
18
MICHAEL LARSEN AND DONG QUAN NGOC NGUYEN
G → G/G′ . The theorem follows by Proposition 3.4 and the additivity of
Hirsch numbers.
References
[AGKS] Avni, Nir; Gelander, Tsachik; Kassabov, Martin; Shalev, Aner: Word values in
p-adic and adelic groups. Bull. Lond. Math. Soc. 45 (2013), no. 6, 1323–1330.
[Bi]
Birch, B. J.: Waring’s problem for p-adic number fields. Acta Arith. 9 (1964),
169–176.
[Ca]
Car, Mireille: Le problème de Waring pour les corps de fonctions. Journées Arithmétiques, 1989 (Luminy, 1989). Astérisque No. 198–200 (1991), 77–82.
[CK1] Carter, David; Keller, Gordon: Bounded elementary generation of SLn (O).
Amer. J. Math. 105 (1983), no.3, 673–687.
[CK2] Carter, David; Keller, Gordon: Elementary expressions for unimodular matrices.
Comm. Algebra 12 (1984), no. 3-4, 379–389.
[Ch]
Chinburg, Ted: Infinite easier Waring constants for commutative rings. Topology
Appl. 158 (2011), no. 14, 1844–1847.
[DG]
Demazure, Michel; Gabriel, Pierre: Groupes algébriques. Tome I: Géométrie algébrique, généralités, groupes commutatifs. Avec un appendice Corps de classes
local par Michiel Hazewinkel. Masson & Cie, Éditeur, Paris; North-Holland Publishing Co., Amsterdam, 1970.
[El]
Ellison, William: Waring’s problem for fields. Acta Arith. 159 (2013), no. 4,
315–330.
[GV]
Gallardo, Luis H.; Vaserstein, Leonid N.: The strict Waring problem for polynomial rings. J. Number Theory 128 (2008), no. 12, 2963–2972.
[GS]
Grunewald, Fritz J.; Schwermer, Joachim: Free nonabelian quotients of SL2 over
orders of imaginary quadratic numberfields. J. Algebra 69 (1981), no. 2, 298–304.
[GT]
Guralnick, Robert M.; Tiep, Pham Huu: Effective results on the Waring problem
for finite simple groups. Amer. J. Math. 137 (2015), no. 5, 1401–1430.
[IL]
Im, Bo-Hae; Larsen, Michael: Waring’s Problem for rational functions in one
variable, preprint.
[Ka]
Kamke, E.: Verallgemeinerungen des Waring-Hilbertschen Satzes. Math. Ann.
83 (1921), no. 1–2, 85–112.
[LST] Larsen, Michael; Shalev, Aner; Tiep, Pham Huu: The Waring problem for finite
simple groups. Annals of Math. (2) 174 (2011), no. 3, 1885–1950.
[LW]
Liu, Yu-Ru; Wooley, Trevor D.: Waring’s problem in function fields. J. Reine
Angew. Math. 638 (2010), 1–67.
[Ro]
Rosenlicht, Maxwell: Some basic theorems on algebraic groups. Amer. J. Math.
78 (1956), 401–443.
[Se]
Serre, Jean-Pierre: Cohomologie galoisienne. With a contribution by Jean-Louis
Verdier. Lecture Notes in Mathematics, No. 5. Troisim̀e édition, 1965 SpringerVerlag, Berlin-New York, 1965.
[Sh]
Shalev, Aner: Word maps, conjugacy classes, and a noncommutative Waringtype theorem. Annals of Math. (2) 170 (2009), no. 3, 1383–1416.
[Si]
Siegel, Carl Ludwig: Generalization of Waring’s problem to algebraic number
fields. Amer. J. Math. 66, (1944), 122–136.
[Si1]
Siegel, Carl Ludwig: Sums of m-th powers of algebraic integers. Ann. of Math.
(2) 46, (1945). 313–339.
[Ta]
Tavgen’, O. I.: Bounded generation of normal and twisted Chevalley groups over
the rings of S-integers. Proceedings of the International Conference on Algebra,
Part 1 (Novosibirsk, 1989), 409–421, Contemp. Math., 131, Part 1, Amer. Math.
Soc., Providence, RI, 1992.
WARING’S PROBLEM FOR UNIPOTENT ALGEBRAIC GROUPS
[Vo]
[W1]
[W2]
[W3]
[Wr]
19
Voloch, José Felipe On the p-adic Waring’s problem. Acta Arith. 90 (1999), no.
1, 91–95.
Wooley, Trevor D. On simultaneous additive equations. I. Proc. London Math.
Soc. (3) 63 (1991), no. 1, 1–34.
Wooley, Trevor D. On simultaneous additive equations. II. J. Reine Angew. Math.
419 (1991), 141–198.
Wooley, Trevor D. On simultaneous additive equations. III. Mathematika 37
(1990), no. 1, 85–96.
Wright, Edward Maitland: An easier Waring problem J. London Math. Soc.
S1–9, (1934), no.4 267–272.
Department of Mathematics, Indiana University, Bloomington, Indiana
47405, USA
E-mail address: [email protected]
Department of Applied and Computational Mathematics and Statistics,
University of Notre Dame, Notre Dame, Indiana 46556, USA
E-mail address: [email protected]
| 4math.GR
|
arXiv:1711.04520v1 [cs.GT] 13 Nov 2017
Fair Knapsack∗
Till Fluschnik† 1 , Piotr Skowron1 , Mervin Triphaus1 , and Kai Wilker1
1
Institut für Softwaretechnik und Theoretische Informatik, TU Berlin, Germany,
[email protected], [email protected],
{mervin.triphaus,wilker}@campus.tu-berlin.de
Abstract
We study the following multiagent variant of the knapsack problem. We are given a
set of items, a set of voters, and a value of the budget; each item is endowed with a cost
and each voter assigns to each item a certain value. The goal is to select a subset of items
with the total cost not exceeding the budget, in a way that is consistent with the voters’
preferences. Since the preferences of the voters over the items can vary significantly,
we need a way of aggregating these preferences, in order to select the socially most
preferred valid knapsack. We study three approaches to aggregating voters preferences,
which are motivated by the literature on multiwinner elections and fair allocation. This
way we introduce the concepts of individually best, diverse, and fair knapsack. We study
computational complexity (including parameterized complexity, and complexity under
restricted domains) of computing the aforementioned concepts of multiagent knapsacks.
1
Introduction
In the classic knapsack problem we are given a set of items, each having a cost and a
value, and a budget. The goal is to find a subset of items with the maximal sum of the
values subject to the constraint that the total cost of the selected items must not exceed
the budget. In this paper we are studying the following variant of the knapsack problem:
instead of having a single objective value for each item we assume that there is a set of agents
(also referred to as voters) who have potentially different valuations of the items. When
choosing a subset of items we want to take into account possibly conflicting preferences of
the voters with respect to which items should be selected: in this paper we discuss three
different approaches to how the voters’ valuations can be aggregated.
Multiagent knapsack forms an abstract model for a number of real-life scenarios. First,
observe that it is a natural generalization of the model for multiwinner elections [19] to the
∗
This research was initiated within the student project “Research in Teams” organized by the research
group Algorithmics and Computational Complexity of TU Berlin, Berlin, Germany.
†
Supported by the DFG, project DAMM (NI 369/13).
1
case where the items come with different costs (in the literature on multiwinner elections,
items are often called candidates). Multiwinner voting rules are applicable in a broad class of
scenarios, ranging from selecting a representative committee of experts, through recommendation systems [34]1 , to resource allocation [37, 50, 49] and facility location problems [21]. In
each of these settings it is quite natural to consider that different items/candidates can incur
different costs. Further, algorithms for multiagent knapsack can be viewed as tools for the
participatory budgeting problem [9], where the authorities aggregate citizens’ preferences
in order to decide which of the potential local projects should obtain funding.
Perhaps the most straightforward way to aggregate voters’ preferences is to select a
subset (a knapsack) that maximizes the sum of the utilities of all the voters over all the
selected items. This approach—which we call selecting an individually best knapsack —
subject to differences in methods used for elicitating voters’ preferences, has been taken by
Benabbou and Perny [3], and in the context of participatory budgeting by Goel et al. [28] and
Benade et al. [4]. However, by selecting an individually best knapsack we can discriminate
even large minorities of voters, which is illustrated by the following simple example: assume
that the set of items can be divided into two subsets, A1 and A2 , that all items have the
same unit cost, and that 51% of the voters like items from A1 (assigning the utility of 1
to them, and the utility of 0 to the other items) and the remaining 49% of voters like only
items from A2 . An individually best knapsack would contain only items from A1 , that is
49% of the voters would be effectively disregarded.
In this paper we introduce two other approaches to aggregating voters’ preferences for
selecting a collective knapsack. One such approach—which we call selecting a diverse knapsack —is inspired by the Chamberlin–Courant rule [11] from the literature on multiwinner
voting. Informally speaking, in this approach we aim at maximizing the number of voters
who have at least one preferred item in the selected knapsack. For the second approach—
which is the main focus of the paper and which we call selecting a fair knapsack —we use
the concept of Nash welfare [39] from the literature on fair allocation. Nash welfare is a
solution concept that implements a tradeoff between having an objectively efficient resource
allocation (knapsack, in our case), and having an allocation which is acceptable for a large
population of agents. Indeed, the properties of Nash welfare have been recently extensively
studied in the literature on fair allocation [10, 38, 14, 46] and this solution concept has
been considered e.g., in the context of public decision making [12], online resource allocation [25] or transmission congestion control [29] (where it is referred to as proportional
fairness). Thus, our work introduces a new application domain—where the goal is to select
a set of shared items—for the concept of Nash welfare. In particular, as a side note, we will
explain that our approach leads to a new class of multiwinner rules, which can be viewed as
generalizations of the Proportional Approval Voting rule [54] beyond the approval setting.
Apart from introducing the new class of multiagent knapsack problems, our contribution
is the following:
1
An example often described in the literature is when an enterprise considers which set of products should
be pushed to production—it is natural to view such a problem as an instance of multiwinner elections with
products corresponding to the items/candidates and potential customers corresponding to the voters.
2
Table 1: Overview of our results: Herein, SP and SC abbreviate single-peaked and single-crossing
preferences, respectively, and “# voters” refers to “when parameterized by the number of voters”.
unary
general
IB Knapsack
Diverse Knapsack
Fair Knapsack
SP
SC
# voters
←
P (Thm. 2)
NP-hard
P
P
FPT
([44])
(Thm. 4)
(Prop. 1)
(Thm. 5)
NP-hard
NP-hard
W[1]-hard
(Thm. 6)
(Thm. 10)
(Thm. 7)
(1) We study the complexity of computing an optimal individually best, diverse, and fair
knapsack. This problem is in general hard, except for the case of individually best
knapsack with the utilities of the voters represented in unary encoding.
(2) We study the parameterized complexity of the problem, focusing on the number of
voters. Considering this parameter is relevant, e.g., for the case when the set of voters
is in fact a relatively small group of experts acting on behalf of a larger population
of agents. Redelegating the task of evaluating the items to the committee of experts
is reasonable for several reasons. For instance, coming up with accurate valuations
of items may require a specialized knowledge and a significant cognitive effort, and
so it would often be impossible to evaluate items efficiently and accurately among a
large group of common people. We show that for unary-encoded utilities of the voters
computing a diverse knapsack is FPT when parameterized by the number of voters.
On the other hand, the problem of computing a fair knapsack is W[1]-hard for the same
parameter.
(3) We study the complexity of the considered problems for single-peaked and singlecrossing preferences. We show that (under unary encoding of voters’ utilities) a diverse knapsack can be computed efficiently when the preferences are single-peaked or
single-crossing. Interestingly, computing fair knapsack stays NP-hard even when the
preferences are both single-peaked and single-crossing.
Our results are summarized in Table 1. We additionally show that all three problems are
NP-hard in the non-unary case (Theorems 3 and 6) and prove fixed-parameter intractability
for the parameter being the budget (Proposition 2 and Corollary 1).
2
The Model
For any pair of natural numbers i, j ∈ N, i ≤ j, by [i, j] we denote the set {i, i + 1, . . . , j}.
Further, by [j] we denote the set [1, j].
3
Let V = {v1 , . . . , vn } be the set of n voters and A = {a1 , . . . , am } be the set of m items.
The voters have preferences over the items, which are represented as a utility profile u =
(ui (a) | i ∈ [n], a ∈ A): for each i ∈ [n] and a ∈ A we use ui (a) to denote the utility that
vi assigns to a; this utility quantifies the extent to which vi enjoys a. We assume that all
utilities are nonnegative integers.
Each item a ∈ A comes with a cost c(a) ∈ N, and we are given a global budget B. We
call a knapsack a subset S of items whose total cost does not exceed B, that is c(S) =
P
a∈S c(a) ≤ B. Our goal is to select a knapsack that would be, in some sense, most
preferred by the voters. Below, we describe three representative rules which extend the
preferences of the individual voters over individual items to their aggregated preferences
over all knapsacks. Each such a rule induces a corresponding method for selecting the best
knapsack. Our rules are motivated with concepts from the literature on fair division and
on multiwinner elections:
Individually best knapsack: this is the knapsack
which maximizes the total utility
P SP
of the voters from the selected items uIB (S) = a∈S vi ∈V ui (a). This defines perhaps
the most straightforward way to select the knapsack: we call it individually best, because
the formula uIB (S) treats the items separately and does not take into account fairnessrelated issues. Indeed, such a knapsack can be very unfair, which is illustrated by the
following example:
Example 1. Let B be an integer, and consider a set of n = B voters and m = B 2
items, all having a unit cost (c(a) = 1 for each a ∈ A). Let us rename the items so that
A = {ax,y | x, y ∈ [B]} and consider the following utility profile:
if i = x = 1
L + 1
ui (ax,y ) = L
if i = x 6= 1
0
otherwise,
for some large L ∈ N. In this case, the individually best knapsack is SIB = {a1,y | y ∈ B},
that is it consists only of the items liked by a single voter v1 . At the same time, there
exists a much more fair knapsack SFair = {ax,1 | x ∈ B} that for each voter v ∈ V
contains an item liked by v.
Diverse knapsack:
this is the knapsack S that maximizes the utility uDiv (S) defined as
P
uDiv (S) = vi ∈V maxa∈S ui (a). In words, in the definition of uDiv we assume that each
voter cares only about his or her most preferred item in the knapsack. This approach
is inspired by the Chamberlin–Courant rule [11] from the literature on multiwinner
elections and by classic models from the literature on facility location [21]. We call
such a knapsack diverse following the convention from the multiwinner literature [19].
Intuitively, such a knapsack represents the diversity of the opinions among the population
of voters; in particular, if the preferences of the voters are very diverse, such a knapsack
tries to incorporate the preferences of as many groups of voters as possible at the cost
of containing only one representative item for each “similar” group.
4
Fair knapsack: we use Nash welfare [39] as a solution concept Q
for fairness.P
Formally, we
call a knapsack S fair if it maximizes the product uFair (S) = vi ∈V 1 + a∈S ui (a) .2
Alternatively,
logarithm of uFair we can represent fair knapsack as the one
Pby taking theP
maximizing vi ∈V log(1 + a∈S ui (a)).
In Section 1 we referred the reader to the literature supporting the use of Nash welfare
in various settings. Let us complement these arguments with one additional observation.
When the utilities of the voters come from the binary set {0, 1} and costs of all items
are equal to one, then our multiagent knapsack framework boils down to the standard
multiwinner elections model with approval preferences. In such a case, a very appealing rule,
Proportional
P
P Approval Voting [19, 54], can be expressed as finding a knapsack maximizing
H(
vi ∈V
a∈S ui (a)), where H(i) is the i-th harmonic number. This is almost equivalent
to finding a fair knapsack (maximizing the Nash welfare) since the harmonic function can
be viewed as a discrete version of the logarithm. Thus, fair knapsack can be considered a
generalization of PAV to the model with cardinal utilities and costs. In particular (as a
side note), observe that the notion of fair knapsack combined with positional scoring rules
induces rules that can be viewed as an adaptations of PAV to the ordinal model.
3
Related Work
Our work extends the literature on the multi-objective (MO) knapsack problem, that is
on the variant of the classic knapsack problem with multiple independent functions valuating the items. Typically, in the MO knapsack problem the goal is to find a (the set of)
Pareto optimal solution(s) according to multiple objectives defined through given functions
valuating items. Our approach is different since we consider specific forms of aggregating
the objectives; in particular each of the concepts we study—the individually best, diverse
and fair knapsack—is a Pareto optimal solution. For an overview of the literature on the
MO knapsack problem (with the focus on the analysis of heuristic algorithms) we refer the
reader to the survey by Lust and Teghem [35].
Multidimensional (MD) knapsack is yet another generalization of the original knapsack
problem; in the MD knapsack we have multiple cost constraints (each item comes with
different costs for different constraints), and the goal is to maximize a single objective while
respecting all the constraints. Approximation algorithms for the problem with submodular
objective functions have been considered, e.g., by Kulik et al. [30], Sviridenko [53], and
Lee et al. [32]. Further, Fréville [26] and Puchinger et al. [45] provide an overview of
heuristic algorithms for the problem. Finally, Florios et al. [23] consider algorithms for the
multidimensional multi-objective variant of the knapsack problem.
Lu and Boutilier [34] studied a variant of the Chamberlin–Courant rule which includes
knapsack constraints, and so, which is very similar to our diverse knapsack problem. The
2
P
Q
Typically,
Nash welfare would be defined as vi ∈V
a∈S ui (a) . In our definition, we add one to the
P
sum a∈S ui (a) in order to avoid pathological situations when the sum is equal to zero for some voters.
This also allows us to represent the expression we optimize as a sum of logarithms, and thus, to expose the
close relation between the fair knapsack and the Proportional Approval Voting rule.
5
difference is that (i) they consider utilities which are extracted from the voters’ preference
rankings, thus these utilities have a specific structure, and (ii) in their model the items are
not shared; instead, the selected items can be copied and distributed among the voters. Lu
and Boutilier consider a model with additional costs related to copying a selected item and
sending it to a voter. Consequently, their general model is more complex than our diverse
knapsack; they also considered a more specific variant of this model, equivalent to winner
determination under the Chamberlin–Courant rule.
The computational complexity of winner determination under the Chamberlin–Courant
rule (i.e., the variant of the diverse knapsack, where the costs of all items are equal to
one) has been extensively studied in the computational social choice (ComSoc) literature.
Procaccia et al. [44] showed that the problem is NP-hard. The parameterized complexity of the problem was investigated by Betzler et al. [5], and its computational complexity under restricted domains by Betzler et al. [5], Yu et al. [55], Elkind and Lackner [16],
Skowron et al. [51], and Peters and Lackner [43]. Lu and Boutilier [34] and Skowron et al. [50]
investigated approximation algorithms for the problem; superpolynomial (FPT) approximation algorithms have been considered by Skowron and Faliszewski [48].
A variant of the diverse knapsack problem with the utilities satisfying a form of the
triangle inequality is known under the name of the knapsack median problem; see the work
of Byrka et al. [8] for a discussion on the approximability of the problem.
The Chamberlin–Courant method is a multiwinner election rule (in short, the multiagent
knapsack model extends the multiwinner model by allowing the items to have different
costs). There is a broad class of multiwinner rules aggregating voter preferences in various
ways. In particular, there exists a number of spectra of rules “between” the individually
best and the Chamberlin–Courant objectives [20, 31]. For an overview of other multiwinner
rules which can be adapted to our setting, see [52, 19].
As we discussed in the introduction, the multiagent variant of the knapsack problem
has been often considered in the context of participatory budgeting, yet to the best of
our knowledge this literature focused on the simplest aggregation rule corresponding to
our individually best knapsack approach [9, 28, 4]. Another avenue has been explored by
Fain et al. [18], who studied rules that determine the level of funding provided to different
projects (items, in our nomenclature) rather than rules selecting subsets of projects with
predefined funding requirements.
As we mentioned before, Nash welfare is an established solution concept used in the
literature on fair allocation [10, 38, 14, 46]. Nguyen et al. [40] provided a thorough survey
on the complexity of computing Nash welfare in the context of allocating indivisible goods
in the multiagent setting. To the best of our knowledge, our paper is the first work studying
fairness solution concepts for the problem of selecting a collective knapsack.
4
Computing Collective Knapsacks
In this section we investigate the computational complexity of finding individually best,
diverse, and fair knapsack. Formally, we define the computational problem for individually
6
best knapsack as:
Individually Best Knapsack
Input: An instance (V, A, u, c) and a budget B.
Task: P
Compute
P a knapsack S ⊆ A such that c(S)
vi ∈V
a∈S ui (a) is maximum.
≤
B and uIB (S) =
We define the computational problems Diverse Knapsack and Fair Knapsack
analogously—the difference is only in the expression to maximize, which for the two problems is uDiv and uFair , respectively. We will use the same names when referring to the
decision variants of these problems; in such cases we will assume that one additional integer
x is given in the input, and that the decision question is whether there exists S with value
uIB (S) (respectively, uDiv (S) or uFair (S)) greater or equal to x, and c(S) ≤ B.
We observe that the functions uIB , uDiv , and uFair (when represented as a sum of logarithms) are submodular. Thus, we can use the algorithm of Sviridenko [53] with the
following guarantees.
Theorem 1. There exists a polynomial-time (1 − 1/e)-approximation algorithm for Individually Best Knapsack, Diverse Knapsack, and Fair Knapsack.
In the remaining part of the paper we will focus on computing an exact solution to the
three problems. In particular, we study the complexity under the following two restricted
domains:
Single-peaked preferences [6, 1]. Let topi denote vi ’s most preferred item, and let ⊳
be an order of the items. We say that a utility profile u is single-peaked with respect to
⊳ if for each a, b ∈ A and each vi ∈ V such that a ⊳ b ⊳ topi or topi ⊳ b ⊳ a we have that
ui (b) ≥ ui (a).
Single-crossing preferences [36, 47]. Let ⊳ be an order of the voters. We say that a
utility profile u is single-crossing with respect to ⊳ if for each two items a, b ∈ A the set
{vi ∈ V | ui (b) ≥ ui (a)} forms a consecutive block according to ⊳.
We say that a profile u is single-peaked (resp., single-crossing) if there exists an order ⊳
of the items (resp., of the voters) such that u is single-peaked (resp., single-crossing) with
respect to ⊳. Note that an order witnessing single-peakedness or single-crossingness can be
computed in polynomial time (see, e.g., [17]).
We will also study the parameterized complexity of the three problems. For a given
parameter p, we say that an algorithm for a problem is FPT with respect to p if it solves
each instance I of the problem in O f (p) · poly(|I|) time, where f is some computable
function. In the parameterized complexity theory, FPT algorithms are considered efficient.
There is a whole hierarchy of complexity classes, but informally speaking, a problem that
is W[1]- or W[2]-hard is assumed not to be FPT and, hence, hard from the parameterized
point of view (see [13, 41, 24, 15] for more details on parameterized complexity).
7
4.1
Individually Best Knapsack
We first look at the simplest case of individually best knapsack.
Theorem 2. Individually Best Knapsack is solvable in polynomial time when the
utilities of voters are unary-encoded.
P
P
Proof. Consider an instance (V, A, u, c, B), and let û := vi ∈V a∈A ui (a). We apply dynamic programming with table T , where T [i, x] denotes the minimal cost of S ⊆ {a1 , . . . , ai }
with value uIB (S) at least equal to x. We initialize T [i, 0] = 0 for i ∈ [m] and T [0, x] = ∞
for each x ∈ [û]. For i ∈ [m], we have
T [i − 1, x],
h
i .
T [i, x] = min
P
c(ai ) + T i − 1, max(0, x − vj ∈V uj (ai ))
By precomputing
P
vj ∈V
uj (a) for each a ∈ A, we get a running time of O(nm + mû).
Note that if the utilities are not encoded in unary, then the problem is NP-hard even
for one voter (see Theorem 6).
4.2
Diverse Knapsack
We now turn our attention to the problem of computing a diverse knapsack. Through a
straightforward reduction from the standard knapsack problem, we get that the problem
is computationally hard even for profiles which are both single-peaked and single-crossing,
unless the utilities are provided in unary encoding.
Theorem 3. Diverse Knapsack is NP-hard even for single-peaked and single-crossing
utility profiles.
Proof. We present a many-one reduction from Knapsack. Let (X = {x1 , . . . , xn }, x, y) be
an instance of Knapsack where each xi comes with
P value ν(xi ) ≥ 1 and
P weight ω(xi );
the question is whether there exists S ⊆ X with xi ∈S ν(xi ) ≥ x and xi ∈S ω(xi ) ≤ y.
We set our set of items A = {a1 , . . . , an } with c(ai ) := ω(xi ) for each i ∈ [n]. We add n
voters v1 , . . . , vn with
2
3n · ν(aj ), i = j,
ui (aj ) := j,
i > j,
2n − j + 1, i < j.
P
P
It
is
immediate
that
for
each
S
we
have
that
c(a
)
=
Further,
i
a
∈S
xi ∈S ω(xi ).
i
P
P
2 x if and only if
max
u
(a
)
≥
3n
ν(x
)
≥
x,
which
proves
the
correctness.
a
∈S
i
j
j
j
vi ∈V
xj ∈S
It is immediate to check that the utility profile is single-peaked and single-crossing.
8
Note that computing a diverse knapsack is also NP-hard for unary encoding, as it
generalizes the Chamberlin–Courant rule, which is computationally hard [44]. For singlepeaked and single-crossing profiles the Chamberlin-Courant rule is computable in polynomial time [5, 51]. These known algorithms can be extended (by considering dynamic
programs with induction running over other dimensions) to the case of the diverse knapsack.
Theorem 4. Diverse Knapsack is solvable in polynomial time when the utility profile is
single-peaked and encoded in unary.
Proof. Consider an input instance (V = {v1 , . . . , vn }, A = {a1 , . . . , am }, u, c, B), where A
is enumerated such that the order is single-peaked
(note that such an ordering can be
P
P
computed in polynomial time [17]). Let û := vi ∈V a∈A ui (a). We apply dynamic programming with table T , where T [i, x] denotes the minimal cost of a subset S ⊆ {a1 , . . . , ai }
containing ai (ai ∈ S) with value at least equal to x (uDiv (S) ≥ x). We define the helper
function
(
P
c(ai ),
vj ∈V uj (ai ) ≥ x,
f (i, x) :=
∞,
otherwise.
We initialize T [1, x] = f (1, x) for all x ∈ [0, û]. Then we set
T [i, x] = min
f [i, x],
c(ai ) + min1≤j<i T [j, x − d(i, j)]
!
,
P
where d(i, j) := vℓ ∈V max(0, uℓ (ai ) − uℓ (aj )). Let M := {(i, x) | T [i, x] ≤ B}. Then we
can derive the value of the best diverse knapsack from max{x | ∃i : (i, x) ∈ M }.
We inductively argue over 1 ≤ i ≤ n. Clearly, the best diverse knapsack over item
set {a1 } has cost c(a1 ). Consider T [i, x] with i > 1. Let Ai ⊆ {a1 , . . . , ai } be a set of
items with value at least x of minimal cost containing ai . Then, either Ai = {ai }, or
Ai = Aj ∪ {ai }, where j ∈ [i − 1] with aj ∈ Ai and there is no j ′ ∈ [i − 1] such that aj ′ ∈ Ai
and j < j ′ < i. In the first case, c(Ai ) = c(ai ) = f [i, x]. Consider the second case. Clearly,
c(Aj ) = c(Ai ) − c(ai ). Let V1 = vℓ ∈ V | uℓ (aj ) ≤ uℓ (ai ) and let V2 := V \ V1 . From the
single-peakedness, we have that uℓ (ai ) − max1≤j ′ ≤j uℓ (aj ′ ) = uℓ (ai ) − uℓ (aj ), if vℓ ∈ V1 , and
clearly uℓ (ai ) −P
max1≤j ′ ≤j uℓ (aj ′ ) < 0, if vℓ ∈ V2 . Hence, the value of Ai is greater than the
value of Aj by vℓ ∈V max(0, uℓ (ai ) − uℓ (aj )) = d(i, j).
Before we provide an analogous result for single-crossingness, let us define a set of useful
tools. We will also use these tools later on, when analyzing the parameterized complexity
of the problem.
~ = (v1 , . . . , vn ) and a subset S ⊆ A of items, we define
Given a tuple of voters V
an assignment πS,V~ as a surjection [n] → S. An assignment is called connected, if for
every s ∈ S it holds that π −1~ (s) := {i ∈ [n] | s = πS,V~ (i)} = [x, y] for some x, y ∈ [n]. For
S,V
our first tool we introduce the following auxiliary problem.
9
Ordered Diverse Knapsack
~ , A, u, c) where V
~ = (v1 , . . . , vn ) is ordered and a budget B.
Input: An instance (V
Task: Compute a knapsack
S ⊆ A such that c(S) ≤ B, and uOrd =
Pn
maxconnected πS,V~ i=1 ui (πS,V~ (i)) is maximum.
If S = {s1 , . . . , sℓ } ⊆ A is a cost-minimal solution to Diverse Knapsack on (V, A, u, c),
~ = (V1 , . . . , Vℓ ),
then let Vi := {vj ∈ V | si = arg maxa∈S uj (a)}. Consider an ordering V
where for each i ∈ [ℓ], the voters in Vi are arbitrarily ordered. Then it is not difficult to
see that the assignment πS,V~ (i) = arg maxa∈S ui (a) is connected. Hence, we obtain the
following connection between Diverse Knapsack and Ordered Diverse Knapsack.
~ on the voters V such that there is an S ⊆ A
Observation 1. There is an ordering V
that forms a cost-minimal solution for Ordered Diverse Knapsack and for Diverse
Knapsack.
Next, we give a dynamic program for computing knapsacks that qualitatively lie “between” optimal knapsacks for Ordered Diverse Knapsack and Diverse Knapsack
(we will specify what we mean by “lying in between” later on).
~ = (v1 , . . . , vn ) of the voters. We
Let us
an input (V, A, u, c, B) and an ordering V
Pfix
n P
set û := i=1 a∈A ui (a). We give a dynamic program with table T , where T [i, x] denotes
“some” cost of a knapsack with a value assigned by voters from (v1 , . . . , vi ) at least equal
to x. We set T [1, x] = min{c(a) | a ∈ A, u1 (a) ≥ x}, if there is an a ∈ A such that u1 (a) ≥ x,
and T [1, x] = ∞ otherwise. We define the helper function
(
Pi
c(a),
j=1 uj (a) ≥ x,
f (i, a, x) =
∞,
otherwise.
We set
T [i, x] = min
a∈A
f (i, a, x),
c(a) + min T [j, max(0, x −
j∈[i−1]
Pi
ℓ=j+1 uℓ (a))]
.
(1)
Observation 2. When the utilities are unary-encoded, we can compute all entries in T in
polynomial time.
Lemma 1. Let S be a cost-minimal solution to Diverse Knapsack on (V, A, u, c) and
let x = uDiv (S). Then T [n, x] ≥ c(S).
Proof. Suppose that this is not the case, that is, T [n, x] < c(S). Then we construct a
knapsack S ′ from T [n, x] as follows. Let a ∈ A be an item that minimizes (1) for T [n, x],
then make a ∈ S ′ . If T [n, x] = f (n, a, x), then T [n, x] = c(a) < c(S), contradicting the fact
that S is cost-minimal. Otherwise,
T [n, x] = c(a) + T [j, x′ := max(0, x −
n
X
ℓ=j+1
10
uℓ (a))].
Then we proceed towards a contradiction as before: Let a′ ∈ A be an item that minimizes (1)
for T [j, x′ ], then make a′ ∈ S ′ , and continue the same reasoning.
We next give the relation to Ordered Diverse Knapsack.
Lemma 2. Let S be a cost-minimal solution to Ordered Diverse Knapsack
~ , A, u, c) where V
~ = (v1 , . . . , vn ) is ordered and let x = uOrd (S). Then T [n, x] ≤ c(S).
on (V
Proof. Assume S = {s1 , . . . , sℓ } being enumerated. Let πS,V~ be an connected assignment
P
such that vi ∈V ui (πS,V~ (i)) = x. Let 0 = i0 < i1 < . . . < iℓ = n be such that [ij + 1, ij+1 ] =
P ij
ui (πS,V~ (i)), for j ∈ [ℓ. By our definition of T , we have
π −1~ (sj+1 ). Moreover, let xj := i=1
S,V
that T [i1 , x1 ] ≤ c(s1 ). Moreover, we have T [i2 , x2 ] ≤ T [i1 , x1 ] + c(s2 ). It follows inductively
that T [iℓ , xℓ ] ≤ c(S).
We have all ingredients at hand to prove our main results.
Proposition 1. Diverse Knapsack is solvable in polynomial time when utility profiles
are single-crossing and encoded in unary.
~ is a single-crossing ordering on the voters V , then there is an S ⊆ A that forms
Proof. If V
a cost-minimal solution for Ordered Diverse Knapsack and for Diverse Knapsack.
By Lemmas 1 and 2 we are guaranteed that our algorithm will find it.
Further, we can use our tools to obtain an FPT algorithm (with respect to the number
of voters) for unrestricted domains.
Theorem 5. Diverse Knapsack is in FPT when parameterized by the number of voters
when the utilities are unary-encoded.
~ on the voters such that there
Proof. By Observation 1, we know that there is an ordering V
is an S ⊆ A that forms a cost-minimal solution for Ordered Diverse Knapsack and
~ our dynamic
for Diverse Knapsack. Together with Lemmas 1 and 2 we obtain that for V
program will find such S. Hence, for each ordering on the voters in V , we compute T [n, x].
Then, we take the minimum over all observed values. Note that x is the largest value such
that T [n, x] ≤ B for some ordering on the voters in V . Altogether, this yields a running
time of O(n! poly(û + n + m)) ⊆ O(2n log n poly(û + n + m)).
Finally, we complement Theorem 5 by proving a lower bound on the running time,
assuming ETH.
Proposition 2. Diverse Knapsack with binary utilities and unary costs is W[2]-hard
when parameterized by the budget B and, unless the ETH breaks, there is no 2o(n+m) ·
poly(n + m) algorithm.
11
Proof. We give a many-one reduction from Dominating Set. An instance of dominating
set consists of a graph G = (W, E) and an integer k; the question is whether there exists
a subset S of at most k vertices such that for each vertex w ∈ W there is an s ∈ S such
that w ∈ NG [s], where NG [s] = {v ∈ W | {v, s} ∈ E} ∪ {s} denotes the closed neighborhood
of s in G. For each vertex w ∈ W , we introduce a voter vw to V and an item aw to A of cost
one. We set uvw (aw′ ) = 1, if w′ ∈ NG [w], and uvw (aw′ ) = 0, otherwise. Furthermore, we set
the budget B = k. It is not difficult to see that there is a diverse knapsack S with c(S) ≤ B
and uDiv (S) ≥ n if and only if (G, k) is a yes-instance. As B = k and |V | = |A| = n, the
lower bounds follow.
4.3
Fair Knapsack
Let us now turn to the problem of computing a fair knapsack. We first prove that the problem is NP-hard, even for restricted cases, and then we study its parameterized complexity.
Theorem 6. Fair Knapsack is NP-hard, even
(1) for one voter;
(2) for two voters and when all costs are equal to one;
(3) if all utilities are in {0, 1} and all costs are equal to one.
Proof. (1): We provide a many-one reduction from the Partition problem [27]: Given a
set S = {s1 , . . . , sn } of nPpositive integers,
the question is to decide whether there exists a
P
subset S ′ ⊆ S such that s∈S ′ s = 21 s∈S s. Given an instance (S) of Partition where all
integers
P are divisible by two, we construct an instance of Fair Knapsack as follows. Let
T := s∈S s. For each si ∈ S, we introduce an item ai with cost si . Further, we introduce
one voter v1 with utility u1 (ai ) = si for each i ∈ [n]. We set the budget B = T /2, and we
ask if there exists a knapsack with a Nash welfare W of at least T /2 + 1.
Let (S) be a yes-instance and let S ′ ⊆ S be P
a solution. Then,
P the subset of items
′
′
A := {ai ∈ A | si ∈ S } forms aPfair knapsack, as P
s∈S ′ s = T /2 ≤ B, and
a∈A′ c(a) =
the Nash welfare is at least 1 + a∈A′ (u1 (a)) = 1 + s∈S ′ s = T /2 + 1
Conversely, let the constructed instance of Fair Knapsack be a yes-instance, and
′
let A′ ⊆ A be a fair knapsack. Denote
to
P
P subset of integers in S corresponding
P by S the
′
s
=
Moreover,
1
+
the P
items in A . Then it holds that s∈S ′ s = a∈A′ c(a) ≤ T /2.
′
s∈S
P
1 + a∈A′ (u1 (a)) ≥ T /2 + 1. Together, both inequalities yield s∈S ′ = T /2, and hence S ′
forms a solution to (S).
(2): We provide a many-one reduction from the Exact Partition problem: Given a
set S = {s1 , . . . , sn } of n positive integers
k ≥ 1, decide whether there is
Pand an integer
1P
′
′
a subset S ⊆ S with |S | = k such that s∈S ′ s = 2 s∈S s. Given an instance (S, k) of
Exact Partition where all integers are divisible by two and byP
k, we construct an instance
of Fair Knapsack as follows. Similarly as before we set T := s∈S s. For each si ∈ S, we
introduce an item ai with cost 1. Further, we introduce two voters, v1 and v2 , with utility
12
functions u1 (ai ) = T + si and u2 (ai ) = T + Tk − si , respectively. We set the budget B = k
and ask for a knapsack with a Nash welfare W at least equal to (1 + kT + T /2)2 .
Let (S, k) be a yes-instance and let S ′ ⊆ S be a solution. We claim that the subset ofPitems A′ := {ai ∈ A | si ∈ S ′ } forms an appropriate fair knapsack. It holds
that a∈A′ c(a) = |S ′ | = k ≤ B. The Nash welfare is at least
!
!
X
X
1+
u1 (a) · 1 +
u2 (a) =
a∈A′
1 + kT +
a∈A′
X
s∈S ′
s
!
·
1 + kT +
X T
s∈S ′
(1 + kT + T /2) 1 + kT + T −
X
s∈S ′
s
k
!
−s
!
=
= (1 + kT + T /2)2 .
Conversely, let the constructed instance of Fair Knapsack be a yes-instance, and
let A′ ⊆ A be a corresponding fair knapsack. Let S ′ denote P
the subset of integers in S
corresponding to the items in A′ . Then it holds that |S ′ | = a∈A′ c(a) ≤Pk. Moreover,
for each item a ∈ A it holds that u1 (a) + u2P
(a) = 2T + T /k, and
P hence a∈A′ (u1 (a) +
u2P
(a)) = |A′ |(2T P
+ T /k). The product (1P+ a∈A′ u1 (a)) · (1 + a∈A
′ u2 (a)) is maximal
T
. Together, it follows
if a∈A′ u1 (a) = a∈A′ u2 (a), leading to a∈A′ u1 (a) = |A′ | T + 2k
′
′
that |A | = k, and hence S forms a solution for (S, k).
(3): We provide a many-one reduction from the Exact Regular Set Packing (ERSP)
problem (there is a straight-forward parameterized reduction from Exact Regular Independent Set [2]): Given a set X, set F = {F1 , . . . , Fm } of subsets of X with |Fi | = d
for all i ∈ [m], and an integer k ≥ 1, decide whether there exists a subset F ′ ⊆ F
with |F ′ | = k such that for each distinct F, F ′ ∈ F it holds F ∩ F ′ = ∅. Let
(X = {x1 , . . . , xn }, F = {F1 , . . . , Fm }, k) be an instance of ERSP where |Fi | = d for all
i ∈ [m]. We construct an instance of Fair Knapsack as follows. Let A := {ai | Fi ∈ F} be
the set of items each with cost equal to one. Further, we introduce n voters with ui (aj ) = 1
if xi ∈ Fj , and ui (aj ) = 0 otherwise, for all i ∈ [n], j ∈ [m]. We set B = k and the desired
Nash welfare to W = 2dk . This finishes the construction.
Assume that (X, F, k) admits a solution F ′ . We claim that A′ := {ai ∈ A | Fi ∈ F ′ }
is a fair knapsack with the desired value of the Nash welfare (note that |A′ | = k). By the
construction, each item a ∈ A contributes one to exactly d voters. Moreover, each distinct
a, a′ ∈ A′ contribute to disjoint sets of voters. Hence,
!
Y
X
1+
ui (a) = 2dk .
1≤i≤n
a∈A′
Conversely, let A′ ⊆ A be a fair knapsack, and let F ′ P
= {Fi ∈P
F | ai ∈ A′ }. We claim
that F ′ forms a solution to (X, F, k). First, observe that 1≤i≤n a∈A′ ui (a) = |A′ |d. Let
13
M := {i ∈ [n] | xi ∈
Then
S
Y
i∈M
F ′ } be the set of elements covered by F ′ . Note that 1 ≤ |M | ≤ |A′ |·d.
1+
X
a∈A′
!
ui (a)
≤
|A′ |d
1+
|M |
|M |
′
≤ 2|A |d ≤ 2kd .
For the second inequality, observe that the function (1 + y/x)x is increasing on the interval
(0, y] for every y > 0. Hence, we have that |A′ | = k and |M | = k · d. Thus, F ′ is a set of
exactly k pairwise disjoint sets.
Given that Exact Regular Set Packing (ERSP) is W[1]-hard with respect to the
size of the solution [2], the proof of Theorem 6 (3) implies the following.
Corollary 1. Fair Knapsack is W[1]-hard when parameterized by the budget, even if all
utilities are in {0, 1} and all costs are equal to one.
Using a more clever construction, we can show that for the combination of the two
parameters—the number of voters and the budget—we still get fixed-parameter intractability.
Theorem 7. Fair Knapsack is W[1]-hard when parameterized by the number of voters
and the budget, even if the utilities and the budget are represented in unary encoding and
the costs of all items are equal to one.
Proof. We provide a parameterized reduction from the k-Multicolored Clique problem, which is known to be W[1]-hard with respect to the number of colors [22]. Let I be
an instance of k-Multicolored Clique. In I we are given a graph G with the set of
vertices V (G) and the set of edges E(G), a natural number k ∈ N, and a coloring function
f : V (G) → [k] that assigns one of k colors to each vertex. We ask if G contains k pairwise
connected vertices, each having a different color. Without loss of generality we assume
that k ≥ 2.
From I we construct an instance IF of Fair Knapsack as follows (we refer to Figure 1
for an illustration). Let T = |V (G)|. We set the set of items to V (G) ∪ E(G), that is we
associate one item with each vertex and with each edge. We construct the set of voters as
follows (unless specified otherwise, by default we assume that a voter assigns utility of zero
to an item):
(1) For each color we introduce one voter who assigns utility of T to each vertex with this
color. Clearly, there are k such voters.
(2) For each pair of two different colors we introduce k − 2 voters, each assigning utility of
T to each edge that connects two vertices with these two colors. There are (k − 2) k2
such voters.
14
, V (G)
n11 · · ·
v1
(1)
T
···
n1i · · ·
n1ℓ
···
T
T
v2
n2j · · ·
n2ℓ · · · · · · {n1i , n2j } · · ·
0
···
T
0
..
.
n21 · · ·
, E(G)
···
T
T
0
···
0
1
v{1,2}
T
..
.
T
..
.
(2)
k−2
v{1,2}
0
..
.
..
.
a
v(1,2)
b
v(1,2)
(3)
0
0
1
···
T − 1 ···
i
···
T − i ···
a
v(2,1)
0
b
v(2,1)
0
ℓ
0
T −i
T −ℓ
0
i
···
1
T − 1 ···
..
.
j
···
T − j ···
ℓ
0
T −j
T −ℓ
0
j
0
Figure 1: Illustration of the instance obtained in the proof of Theorem 7. Herein, ncb denotes
vertex b in color class c, where each color class contains ℓ vertices. In the presented example, the
vertices n1i and n2j are adjacent. Blocks containing a zero indicate that the corresponding entries
are zero.
(3) For each ordered pair of colors, c1 and c2 , with c1 6= c2 we introduce two vertices, call
them a and b, with the following utilities. Consider the set of vertices with color c1 and
rename them in an arbitrary way so that they can be put in a sequence n1 , n2 , . . . , nℓ .
For each i ∈ [ℓ] voter a assigns utility i to vertex ni and utility (T − i) to each edge
that connects ni with a vertex with color c2 . Voter b assigns utility (T − i) to ni and
utility i to each edge that connects ni with a vertex with color c2 . There are 2k(k − 1)
such voters.
We set the cost of each item to one, and the total budget to B = k + k2 . By a simple
calculation one can check that the total number of voters is equal to k + (k − 2) · k2 +
2k(k − 1) = kB. This completes our construction.
First, observe that in total each item is assigned utility of kT from all the voters. Indeed,
each item corresponding to a vertex gets utility of T from exactly one voter from the first
group, and total utility of (k − 1) · T from 2(k − 1) voters from the third group. Similarly,
each item corresponding to an edge gets utility of T from k − 2 voters from the second
15
group, and total utility of 2 · T from four voters from the third group. Thus, independently
of how we select B items, the sum of the utilities they are assigned from the voters will
always be the same, that is BkT . Thus, clearly the Nash welfare would be maximized if
the total utility assigned to the selected items by each voter is the same, and equal to T .
Only in such case the Nash welfare would be equal to (T + 1)kB . We will show, however,
that each voter assigns to the set of B items utility
T if and only if k out of such items are
k
vertices with k different colors, the remaining 2 of such items are edges, and each selected
edge connects two selected vertices.
Indeed, it is easy to see that if the selected set of items has the structure as described
above, then each voter assigns to this set the utility of T . We will now prove the other
implication. Assume that for the set of B items S each voter assigns total utility of T . By
looking at the first group of voters, we infer that k items from S correspond to the vertices,
and that these k vertices have different colors. By looking at the second group of voters,
we infer that for each pair of two different colors, S contains exactly one edge connecting
vertices with such colors. Finally, by looking at the third group of voters we infer that each
edge from S that connects colors c1 and c2 is adjacent to the vertices from S with colors c1
and c2 . This completes the proof.
On the other hand, each instance I of Fair Knapsack with utilities represented in
unary encoding is solvable in O(|I|f (n) ) time (it is in XP when parameterized by n), where n
is the number of voters and f is some computable function only depending on n.
Theorem 8. For utilities represented in unary encoding, Fair Knapsack is in XP when
parameterized by the number of voters.
Proof. We provide an algorithm based on dynamic programing. We construct a table T
where for each sequence of n + 1 integers, z1 , z2 , . . . , zn , and i, entry T [z1 , z2 , . . . , zn , i]
represents the lowest possible value of the budget x such that there exists a knapsack S
with P
the following properties: (i) the total cost of all items in the knapsack is equal to x (i.e.,
x = a∈S c(a)), (ii) the last index of an P
item in the knapsack S is i (i.e., i = maxaj ∈S j),
and (iii) for each voter vj we have that a∈S uj (a) = zj . This table can be constructed
recursively:
T [z1 , z2 , . . . , zn , i] = c(ai ) + min T [z1 − u1 (ai ), z2 − u2 (ai ), . . . , zn − un (ai ), j].
j≤i
We handle the corner cases by setting T [0, 0, . . . , 0, i] = 0 for each i, and T [z1 , z2 , . . . , zn , i] =
∞ whenever zi < 0 for some i ∈ [n].
Clearly, if n is fixed and if the utilities are represented in unary encoding, then the table
can be filled in polynomial time. Now, it is sufficient
to traverse the table and to find the
Q
entry T [z1 , z2 , . . . , zn , i] ≤ B which maximizes nj=1 (zj + 1).
On the positive side, with stronger requirements on the voters’ utilities, that is, if the
number of different values over the utility functions is small, we can strengthen Theorem 8
and prove the membership in FPT.
16
Theorem 9. Fair Knapsack is FPT when parameterized by the combination of the number of voters and the number of different values that a utility function can take.
Proof. We will use the classic result of Lenstra [33] which says that an integer linear program
(ILP) can be solved in FPT time with respect to the number of integer variables. We will
also use a recent result of Bredereck et al. [7] who proved that one can apply concave/convex
transformations of certain variables in an ILP, and that such a modified program can be
still solved in an FPT time. We construct an ILP as follows. Let U be the set of values
that a utility function can take. For each vector z = (z1 , . . . , zn ) with zi ∈ U for each i, we
define Az as the set of items a such that for each voter vi we have ui (a) = zi . Intuitively, Az
describes a subcollection of the items with the same “type”: such items are indistinguishable
when we look only at the utilities assigned by the voters; they may vary only with their
costs. For each such a set Az we introduce an integer variable xz which intuitively denotes
the number of items from the optimal solution that belong to Az . Further, we construct a
function fz such that fz (x) is the cost of the x cheapest items from Az ; clearly fz is convex.
We formulate the following program:
maximize:
X
log
X
zi · xz
z∈U n
vi ∈V
subject to:
X
!
fz (xz ) ≤ B
z∈U n
xz ∈ Z,
z ∈ Un
The above program uses concave transformations (logarithms) for the maximized expression,
and convex transformations (functions fz ) in the left-hand sides of the constraints, so we
can use the result of Bredereck et al. [7] and claim that this program can be solved in an
FPT time with respect to the number of integer variables. This completes the proof.
4.4
Fair Knapsack under Restricted Domains
In contrast to Individually Best Knapsack and Diverse Knapsack, both being solvable in polynomial time on restricted domains, computing a fair knapsack remains NP-hard
on utility profiles that are even both, single-peaked and single-crossing.
Theorem 10. Fair Knapsack is NP-hard even on single-peaked single-crossing domains,
when the costs of all items are equal to one, and the utilities of each voter come from the
set {0, . . . 6}.
Proof. We give a many-one reduction from the NP-hard Exact-Set-Cover (X3C) problem: Given a universe U with 3k elements and a set F of 3-size subsets of U , the question
is to decide whether there exist exactly k subsets in F that cover U . Without loss of
generality, we can additionally assume that each element in U appears in exactly three
sets from F [27]. Given an instance (U = {e1 , . . . , en }, F = {F1 , . . . , Fm }) of X3C (note
17
a1 a2 · · · ai
x1
am a′m
···
a′i · · ·
6
···
a′2 a′1
0
6
x2
0
(i)
y1
6
(i)
y2
(ℓ)
z1
(ℓ)
z2
3
6
5
0
4
3
6
5
4
3
2
1
0
6
3
0
2
1
0
Figure 2: Visualization of the utilities of the voters used in the proof of Theorem 10. The solid lines
can be interpreted as plots depicting the utilities of the voters from different items. For instance,
agent x1 assigns utility of 0 to the items a1 , . . . am , and utility of 6 to the items am+1 , . . . a2m (note
(ℓ)
(ℓ)
that a′j = a2m−j+1 ). Agents z1 and z2 depicted in the figure correspond to the element eℓ such
that eℓ ∈ F2 ∩ Fi ∩ Fm .
that n = m = 3k), we compute an instance of the problem of computing a fair knapsack as
follows (the utilities of the voters are depicted in Figure 2).
First, for each i ∈ [m], we introduce two items, ai and a2m−i , that correspond to set Fi ,
each with the cost of one. Further, we introduce three different types of voters:
(1) We add two voters x1 and x2 with ux1 (ai ) = ux1 (am+i ) − 6 = 0 and ux2 (am+i ) =
ux2 (ai ) − 6 = 0 for all i ∈ [m].
(i)
(i)
(2) For each i ∈ [m], we add two voters y1 and y2 with
0,
uy(i) (aj ) = 3,
1
6,
1 ≤ j ≤ i,
i < j < 2m − i + 1,
2m − i + 1 ≤ j ≤ 2m,
and uy(i) (aj ) = uy(i) (a2m−j+1 ).
2
1
(i)
(i)
(3) For each i ∈ [n], we add two voters z1 and z2 with uz (i) (aj ) = fi (aj ) and uz (i) (aj ) =
1
uz (i) (a2m−j+1 ), where
2
1
fi (aj ) = |{ℓ | 1 ≤ ℓ ≤ j, ei ∈ Fℓ }|,
for j ∈ [m],
fi (a2m−j+1 ) = 3 + |{ℓ | j ≤ ℓ ≤ m, ei ∈ Fℓ }|,
for j ∈ [m].
18
We set the budget to B = 2k and the required Nash welfare to W = (6k + 1)2+2m (6k + 2)2n .
It is apparent that with the order (a1 , . . . , a2m ), this is profile is single-peaked. For single(i)
(i)
crossingness, note that the utilities of agents xi , yj , zj ′ are increasing over (a1 , . . . , a2m )
if i = 1, and decreasing, if i = 2. Hence, the order of voters
(1)
(1)
(2)
(2)
(1)
(1)
(2)
(2)
(x1 , y1 , . . . , ym
, z1 , . . . , zm
, x2 , y 1 , . . . , y m
, z1 , . . . , zm
)
witnesses single-crossingness.
We will prove that (U = {e1 , . . . , e3k }, F = {F1 , . . . , Fm }) is a yes-instance for X3C if
and only if the constructed instance of Fair Knapsack is a yes-instance.
(⇒) Let F ′ = {Fb1 , . . . , Fbk } ⊆ F be an exact cover of U . We claim that S =
{abi , a2m−bi +1 | i ∈ [k]} is a fair knapsack. First observe that c(S) = 2k ≤ B. We
consider the welfare for each of P
the three typesP
(1)-(3) of voters separately.
For x1 and x2 (1), we have a∈S ux1 (a) = a∈S ux2 (a) = 6k.
(i)
Next, consider the voters of type (2). Consider y1 , i ∈ [m]:
k
X
[uy(i) (abj ) + uy(i) (a2m−bj +1 )] =
j=1
1
1
=
X
[uy(i) (abj ) + uy(i) (a2m−bj +1 )] +
1
1
1≤bj ≤i
X
[uy(i) (abj ) + uy(i) (a2m−bj +1 )]
i<bj ≤m
1
1
= {j | 1 ≤ bj ≤ i} · (0 + 6) + {j | i < bj ≤ m} · (3 + 3) = 6k.
By symmetry,
Pk
j=1 [uy (i) (abj )
2
+ uy(i) (a2m−bj +1 )] = 6k.
2
(i)
Finally, consider the voters of type (3). Consider a voter z1 , i ∈ [m]. Let j ∗ be the
index such that ei ∈ Fbj∗ (recall exact cover). We have
X
a∈S
uz (i) (a) =
1
k
X
[fi (abj ) + fi (a2m−bj +1 )]
j=1
= fi (abj∗ ) + fi (a2m−bj∗ +1 ) +
X
[fi (abj ) + fi (a2m−bj +1 )]
j∈[k]\{j ∗ }
X
=6+1+
j∈[k]\{j ∗ }
X
=6+1+
|{ℓ | 1 ≤ ℓ ≤ bj , ei ∈ Fℓ }| + 3 + |{ℓ | bj ≤ ℓ ≤ m, ei ∈ Fℓ }|
(|{ℓ | 1 ≤ ℓ ≤ m, ei ∈ Fℓ }| + 3) +
j∈[k]\{j ∗ }
X
=6+1+
j∈[k]\{j ∗ }
By symmetry,
P
a∈S
X
j∈[k]\{j ∗ }
6+
X
0 = 6k + 1
j∈[k]\{j ∗ }
uz (i) (a) = 6k + 1.
2
19
{ei } ∩ Fbj
Hence, we get in total that the Nash welfare is equal to
(1 + 6k)2 (1 + 6k)2m (2 + 6k)2n = W .
(⇐) Let S ⊆ {a1 , . . . , a2m } be a fair knapsack with c(S) ≤ 2k and with the Nash welfare
at least equal to W . We will now show that the total utility that all voters assign to each
item aj ∈ A is equal to 6 + 6m + 6n + 3. Indeed, the two voters from (1) assign to aj
(i)
(i)
the total utility of 6. Similarly, any pair of voters, y1 and y2 , assigns utility of 6 to aj .
(i)
(i)
Finally, observe that, whenever ei ∈ Fj , then voters z1 and z2 assign utility of 7 to aj ;
otherwise they assign utility of 6 to aj . Since each set Fj contains exactly 3 elements, we
get that aj gets total utility of (n − 3)6 + 3 · 7 from the voters from (3).
Hence, 2k items contribute 2k(6 + 6m + 6n + 3) to the total utility, and so, for the Nash
welfare to be equal to W , this total utility must be distributed as equally as possible among
the voters. Specifically, 2m + 2 voters need to get the total utility of 6k, and 2n voters must
get the total utility of 6k + 1.
Now, we claim that for each i ∈ [m], ai ∈ S ⇐⇒ a2m−i+1 ∈ S. Suppose this is not
the case, and let i ∈ [m] be the smallest index such that either (i) ai ∈ S ∧ a2m−i+1 6∈ S
or (ii) ai 6∈ S ∧ a2m−i+1 ∈ S. Consider the first case (i). Let k1 = |{1 ≤ j ≤ i | aj ∈ S}|
and k2 = |{2m − i + 1 ≤ j ≤ 2m | aj ∈ S}|. It holds that k1 ≥ k2 + 1, and it follows for
(i)
voter y2 that
X
uy(i) (aj ) = k1 6 + (2k − (k1 + k2 )) · 3 + k2 · 0
aj ∈S
2
= 3(2k + (k1 − k2 )) ≥ 6k + 3 ∈
/ {6k, 6k + 1}.
Case P(ii) works analogously, and hence, our claim follows. P
From this we infer
that a∈S uy(ℓ) (a) = 6k for each i ∈ {1, 2} and ℓ ∈ [m], and that a∈S uxi (a) = 6k for
i
P
each i ∈ {1, 2}. Thus, for each voter z from (3) it must be the case that a∈S uz (a) = 6k+1.
Finally, we will prove that F ′ = {Fb1 , . . . , Fbk } = {Fi ∈ F | ai ∈ S, i ∈ [m]} forms a
cover of U . Towards a contradiction suppose that there is an element ei ∈ U such that ei
(i)
is not covered by F ′ . We consider voter z1 . Observe that since ei 6∈ Fbj for each j ∈ [k],
we have
X
X
6 = 6k < 6k + 1.
uz (i) (abj ) + uz (i) (a2m−bj +1 ) =
j∈[k]
1
1
j∈[k]
Thus, we reached a contradiction, and consequently we get that every element in U is
covered by F. This completes the proof.
As we discussed in Section 2, if the voters’ utilities come from the binary set {0, 1} and
if the costs of the items are equal to one, then the problem of computing a fair knapsack
is equivalent to computing winners according to Proportional Approval Voting. For this
case with single-peaked preferences, Peters [42] showed that the problem can be formulated
20
as an integer linear program with total unimodular constraints, and thus it is solvable in
polynomial time. This makes our result interesting, as it shows that by allowing slightly
more general utilities (coming from the set {0, . . . , 6}, instead of {0, 1}) the problem becomes
already NP-hard (even if we additionally assume single-crossingness of the preferences).
This draws quite an accurate line separating instances which are computationally easy
from those which are intractable.
5
Conclusion
In this paper we study three variants of the knapsack problem in multiagent settings. One of
these variants, selecting an individually best knapsack, has been considered in the literature
before, and this work introduces the other two concepts: diverse and fair knapsack. Our
paper establishes a relation between the multi-objective knapsack problem and a broad
literature including a literature on multiwinner voting and on fair allocation. This way,
we expose a variety of ways in which the preferences of the voters can be aggregated in
a number of real-life scenarios that are captured by the abstract model of the multiagent
knapsack.
Our computational results are outlined in Table 1. In summary, our results show that
the problem of computing an individually-best or a diverse knapsack can be handled efficiently under some simplifying assumptions. On the other hand, we give multiple evidences
that computing a fair knapsack is a hard problem. Thus, this research provides theoretical
foundations motivating, and calls for, studying approximation and heuristic algorithms for
the problem of computing a fair knapsack.
References
[1] K. Arrow. Social Choice and Individual Values. John Wiley and Sons, 1951 (revised
editon, 1963).
[2] G. Ausiello, A. D’Atri, and M. Protasi. Structure preserving reductions among convex
optimization problems. Journal of Computer and System Sciences, 21:136–153, 1980.
[3] N. Benabbou and P. Perny. Solving multi-agent knapsack problems using incremental
approval voting. In Proceedings of the 22nd European Conference on Artificial Intelligence, pages 1318–1326, 2016.
[4] G. Benade, S. Nath, A. Procaccia, and N. Shah. Preference elicitation for participatory
budgeting. In Proceedings of the 31st AAAI Conference on Artificial Intelligence, pages
376–382, 2017.
[5] N. Betzler, A. Slinko, and J. Uhlmann. On the computation of fully proportional
representation. Journal of Artificial Intelligence Research, 47:475–519, 2013.
21
[6] D. Black. On the rationale of group decision-making. Journal of Political Economy,
56(1):23–34, 1948.
[7] R. Bredereck, P. Faliszewski, R. Niedermeier, P. Skowron, and N. Talmon. Mixed integer programming with convex/concave constraints: Fixed-parameter tractability and
applications to multicovering and voting. Technical Report arXiv:1709.02850 [cs.DS],
arXiv.org, 2017.
[8] J. Byrka, T. Pensyl, B. Rybicki, J. Spoerhase, A. Srinivasan, and K. Trinh. An improved approximation algorithm for knapsack median using sparsification. In Proceedings of 23rd Annual European Symposium on Algorithms (ESA-2015), pages 275–287,
2015.
[9] Y. Cabannes. Participatory budgeting: a significant contribution to participatory
democracy. Environment and Urbanization, 16(1):27–46, 2004.
[10] I. Caragiannis, D. Kurokawa, H. Moulin, A. Procaccia, N. Shah, and J. Wang. The
unreasonable fairness of maximum nash welfare. In Proceedings of the 2016 ACM
Conference on Economics and Computation (EC-16), pages 305–322.
[11] B. Chamberlin and P. Courant. Representative deliberations and representative decisions: Proportional representation and the Borda rule. American Political Science
Review, 77(3):718–733, 1983.
[12] V. Conitzer, R. Freeman, and N. Shah. Fair public decision making. In Proceedings of
the 2017 ACM Conference on Economics and Computation (EC-17), pages 629–646,
2017.
[13] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015.
[14] A. Darmann and J. Schauer. Maximizing nash product social welfare in allocating
indivisible goods. European Journal of Operational Research, 247(2):548–559, 2015.
[15] R. Downey and M. Fellows. Fundamentals of Parameterized Complexity. Texts in
Computer Science. Springer, 2013.
[16] E. Elkind and M. Lackner. Structure in dichotomous preferences. In Proceedings of the
24th International Joint Conference on Artificial Intelligence, pages 2019–2025, 2015.
[17] E. Elkind, M. Lackner, and D. Peters. Structured preferences. In U. Endriss, editor,
Trends in Computational Social Choice. AI Access, 2017.
[18] B. Fain, A. Goel, and K. Munagala. The core of the participatory budgeting problem.
In Proceedings of the 12th Conference on Web and Internet Economics, pages 384–399,
2016.
22
[19] P. Faliszewski, P. Skowron, A. Slinko, and N. Talmon. Multiwinner voting: A new
challenge for social choice theory. In U. Endriss, editor, Trends in Computational
Social Choice. AI Access, 2017.
[20] P. Faliszewski, P. Skowron, A. Slinko, and N. Talmon. Multiwinner rules on paths from
k-Borda to Chamberlin–Courant. pages 192–198, 2017.
[21] F. Zanjirani Farahani and M. Hekmatfar, editors. Facility Location: Concepts, Models,
and Case Studies. Springer, 2009.
[22] M. Fellows, D. Hermelin, F. Rosamond, and S. Vialette. On the parameterized complexity of multiple-interval graph problems. Theoretical Computer Science, 410(1):
53–61, 2009.
[23] K. Florios, G. Mavrotas, and D. Diakoulaki. Solving multiobjective, multiconstraint
knapsack problems using mathematical programming and evolutionary algorithms. European Journal of Operational Research, 203(1):14–21, 2010.
[24] J. Flum and M. Grohe. Parameterized Complexity Theory. Springer-Verlag, 2006.
[25] R. Freeman, S. Zahedi, and V. Conitzer. Fair and efficient social choice in dynamic
settings. pages 4580–4587, 2017.
[26] A. Fréville. The multidimensional 0–1 knapsack problem: An overview. European
Journal of Operational Research, 155(1):1–21, 2004.
[27] M. Garey and D. Johnson. Computers and Intractability: A Guide to the Theory of
NP-Completeness. W. H. Freeman and Company, 1979.
[28] A. Goel, A. Krishnaswamy, S. Sakshuwong, and T. Aitamurto. Knapsack voting:
Voting mechanisms for participatory budgeting. Manuscript, 2016.
[29] Frank Kelly. Charging and rate control for elastic traffic. European Transactions on
Telecommunications, 8:33–37, 1997.
[30] A. Kulik, H. Shachnai, and T. Tamir. Approximations for monotone and nonmonotone submodular maximization with knapsack constraints. Mathematics of Operations
Research, 38(4):729–739, 2013.
[31] M. Lackner and P. Skowron. Consistent approval-based multi-winner rules. Technical
Report arXiv:1704.02453v1 [cs.GT], arXiv.org, April 2017.
[32] J. Lee, V. Mirrokni, V. Nagarajan, and M. Sviridenko. Non-monotone submodular
maximization under matroid and knapsack constraints. In Proceedings of the Fortyfirst Annual ACM Symposium on Theory of Computing (STOC-2009), pages 323–332,
2009.
23
[33] H. Lenstra, Jr. Integer programming with a fixed number of variables. Mathematics
of Operations Research, 8(4):538–548, 1983.
[34] T. Lu and C. Boutilier. Budgeted social choice: From consensus to personalized decision making. In Proceedings of the 22nd International Joint Conference on Artificial
Intelligence, pages 280–286, 2011.
[35] T. Lust and J. Teghem. The multiobjective multidimensional knapsack problem: a
survey and a new approach. International Transactions in Operational Research, 19
(4):495–520, 2012.
[36] J. Mirrlees. An exploration in the theory of optimal income taxation. Review of
Economic Studies, 38:175–208, 1971.
[37] B. Monroe. Fully proportional representation. American Political Science Review, 89
(4):925–940, 1995.
[38] H. Moulin. Fair division and collective welfare. The MIT Press, 2003.
[39] J. Nash. The bargaining problem. Econometrica, 18(2):155–162, 1950.
[40] T. Nguyen, M. Roos, and J. Rothe. A survey of approximability and inapproximability
results for social welfare optimization in multiagent resource allocation. Annals of
Mathematics and Artificial Intelligence, 68(1-3):65–90, 2013.
[41] R. Niedermeier. Invitation to Fixed-Parameter Algorithms. Oxford University Press,
2006.
[42] D. Peters. Single-peakedness and total unimodularity: Efficiently solve voting problems
without even trying. Technical Report 1609.03537, arXiv.org, 2017.
[43] D. Peters and M. Lackner. Preferences single-peaked on a circle. In Proceedings of the
31st AAAI Conference on Artificial Intelligence, pages 649–655, 2017.
[44] A. Procaccia, J. Rosenschein, and A. Zohar. On the complexity of achieving proportional representation. Social Choice and Welfare, 30(3):353–362, 2008.
[45] J. Puchinger, G. Raidl, and U. Pferschy. The multidimensional knapsack problem:
Structure and algorithms. INFORMS Journal on Computing, 22(2):250–265, 2010.
[46] S. Ramezani and U. Endriss. Nash Social Welfare in Multiagent Resource Allocation,
pages 117–131. Springer Berlin Heidelberg, 2010.
[47] K. W. S. Roberts. Voting over income tax schedules. Journal of Public Economics, 8
(3):329–340, 1977.
24
[48] P. Skowron and P. Faliszewski. Fully proportional representation with approval ballots:
Approximating the MaxCover problem with bounded frequencies in FPT time. In
Proceedings of the 29th AAAI Conference on Artificial Intelligence, pages 2124–2130,
2015.
[49] P. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From
proportional multirepresentation to group recommendation. In Proceedings of the 29th
AAAI Conference on Artificial Intelligence, 2015.
[50] P. Skowron, P. Faliszewski, and A. Slinko. Achieving fully proportional representation:
Approximability result. Artificial Intelligence, 222:67–103, 2015.
[51] P. Skowron, L. Yu, P. Faliszewski, and E. Elkind. The complexity of fully proportional
representation for single-crossing electorates. Theoretical Computer Science, 569:43–57,
2015.
[52] P. Skowron, P. Faliszewski, and J. Lang. Finding a collective set of items: From
proportional multirepresentation to group recommendation. Artificial Intelligence, 241:
191–216, 2016.
[53] M. Sviridenko. A note on maximizing a submodular set function subject to a knapsack
constraint. Operations Research Letters, 32(1):41–43, 2004.
[54] T. N. Thiele. Om flerfoldsvalg. In Oversigt over det Kongelige Danske Videnskabernes
Selskabs Forhandlinger, pages 415–441. 1895.
[55] L. Yu, H. Chan, and E. Elkind. Multiwinner elections under preferences that are
single-peaked on a tree. In Proceedings of the 23rd International Joint Conference on
Artificial Intelligence, pages 425–431, 2013.
25
| 8cs.DS
|
AN INFORMAL OVERVIEW OF TRIPLES AND SYSTEMS
arXiv:1709.03174v3 [math.AC] 14 Nov 2017
LOUIS ROWEN
Abstract. We describe triples and systems, expounded as an axiomatic algebraic umbrella theory for
classical algebra, tropical algebra, hyperfields, and fuzzy rings.
1. Introduction
The goal of this overview is to present an axiomatic algebraic theory which unifies, simplifies, and
“explains” aspects of tropical algebra [28, 34, 35, 32], hyperfields [5, 21, 41, 61], and fuzzy rings [16, 17, 21]
in terms of familiar algebraic concepts. It was motivated by an attempt to understand whether or not it
is coincidental that basic algebraic theorems are mirrored in supertropical algebra, and was spurred by
the realization that some of the same results are obtained in parallel research on hyperfields and fuzzy
rings. Our objective is to hone in on the precise axioms that include these various examples, formulate
the axiomatic structure, describe its uses, and review five papers [58, 4, 45, 46, 44] in which the theory
is developed. The bulk of this survey concerns [58], in which the axiomatic framework is laid out, since
the other papers build on it.
Other treatments can be found in [13, 23, 49]. Although we deal with general categorical issues,
ours is largely a “hands on” approach, emphasizing a “negation map” which exists in all of the abovementioned examples, and which often is obtained by means of a “symmetrization” functor. Although
the investigation has centered on semirings, having grown out of tropical considerations, it also could be
used to develop a parallel Lie semi-algebra theory (and more generally Hopf semi-algebra theory).
1.1. Acquaintance with basic notions.
One starts with a set T that we want to study, called the set of tangible elements, endowed with a
partial additive algebraic structure which however is not defined on all of T ; this is resolved by embedding
T in a larger set A with a fuller algebraic structure. Often T is a multiplicative monoid 1, a situation
developed by Lorscheid [49, 50] when A is a semiring. However, there also are examples (such as Lie
algebras) lacking associative multiplication. We usually denote a typical element of T as a, and a typical
element of A as b.
Definition 1.1. A T -module over a set T 2 is an additive monoid (A, +, 0A ) together with scalar
multiplication T × A → A satisfying distributivity over T 3 in the sense that
a(b1 + b2 ) = ab1 + ab2
for a ∈ T , bi ∈ A, also stipulating that a0A = 0A .
A T -monoid module over a multiplicative monoid T is a T -module Asatisfying the extra conditions
1T b = b,
(a1 a2 )b = a1 ,
∀ai ∈ T , b ∈ A.
Date: November 15, 2017.
2010 Mathematics Subject Classification. Primary 08A05, 16Y60, 12K10, 13C60, 20N20; Secondary 03C05, 06F05,
13C10, 14T05.
Key words and phrases. bipotent, category, congruence, dual basis, homology, hyperfield, linear algebra, matrix, metatangible, morphism, negation map, module, polynomial, prime, projective, tensor product, semifield, semigroup, semiring,
split, supertropical algebra, superalgebra, surpassing relation, symmetrization, system, triple, tropical.
File name: InformalOverview29.
1Geometry over monoids has been a subject of recent interest, cf. [14].
2More generally, when T is not a monoid, Hopf theory can play an interesting role examined in [46].
3It turns out that distributivity over elements of T is enough to run the theory, since one can (re)define multiplication
on A to make it distributive, as seen in [58, Theorem 2.9]. This rather easy result applies for instance to hyperfields such
as the phase hyperfield.
1
2
L. ROWEN
For the sake of this exposition, we assume that A is a T -module and T ⊆ A. Next, we introduce a
formal negation map (−), which we describe in §1.3 after some introductory examples, such that T ∪(−)T
generates A additively, creating a T -triple (A, T , (−)).
When a formal negation map is not available at the outset, we can introduce it in two ways, to be
elaborated shortly:
• Declare the negation map to be the identity, as in the supertropical case, cf. §1.2.2.
• Apply symmetrization, to get the switch map, of second kind, cf. §1.2.4. Often [3] is applicable,
where T could take the role of the “thin elements.”
The element b◦ := b(−)b is called a quasi-zero. We write A◦ for {b◦ : b ∈ A}, and usually require
that T ∩ A◦ = ∅, i.e., a quasi-zero cannot be tangible.
In classical algebra, the only quasi-zero is 0 itself, and a1 − a1 = 0 = a2 − a2 for all a1 , a2 . Accordingly,
we call a triple “T -classical” [58, Definition 2.45], when a1 (−)a1 = a2 (−)a2 for some a1 6= (±)a2 in T .
Examples from classical mathematics might provide some general intuition about employing A to
study T . A rather trivial example: T is the multiplicative subgroup of a field A. Or A could be a graded
associative algebra, with T its multiplicative submonoid of homogeneous elements. A deeper example
tied to Gröbner-Shirshov bases is given in [45, Example 2.2].
But we are more interested in the non-classical situation, involving semirings which are not rings.
Some motivating examples: The supertropical semiring, where T is the set of tangible elements, the
symmetrized semiring, and the power set of a hyperfield, where T is the hyperfield itself. Since hyperfields
are so varied, they provide a good test for this theory. Semirings in general, without negation maps, are
too broad to yield as decisive results as we would like, which is the reason that negation maps and triples
are introduced in the first place.
Since we need to correlate two structures (T and A), as well as the negation map (which could be
viewed as a unary operator), it is convenient to work in the context of universal algebra, which was
designed precisely for the purpose of discussing diverse structures together. (More recently these have
been generalized to Lawvere theories and operads, but we do not delve into these aspects.)
To round things out, given a triple, we introduce a surpassing relation , to replace equality in
our theorems. (In classical mathematics, is just equality.) The quadruple (A, T , (−), ) is called a
T -system.
1.2. Motivating examples.
A semiring† satisfies all the axioms of ring except the existence of a 0 element and of negatives. A
semiring is a semiring† with 0. We elaborate the main non-classical examples motivating this theory.
1.2.1. Idempotent semirings.
Tropical geometry has assumed a prominent position in mathematics because of its ability to simplify algebraic geometry while not changing certain invariants (often involving intersection numbers
of varieties), thereby simplifying difficult computations. Outstanding applications abound, including
[1, 26, 39, 53, 56].
The main original idea, as expounded in [27, 51], was to take the limit of the logarithm of the absolute
values of the coordinates of an affine variety as the base of the logarithm goes to ∞. The underlying
algebraic structure reverted from C to the max-plus algebra Rmax , an ordered multiplicative monoid in
which one defines a + b to be max{a, b}. This is a semiring† and is clearly additively bipotent in the
sense that a + b ∈ {a, b}. Such algebras have been studied extensively some time ago, cf. [12, 20].
Definition 1.2. A semigroup (A, +) has characteristic k > 0 if (k + 1)a = a for all a ∈ A, with k ≥ 1
minimal. A has characteristic 0 if A does not have characteristic k for any k ≥ 1.
Idempotent (in particular bipotent) semirings have characteristic 1, and their geometry has been
studied intensively as “F1 -geometry,” cf. [9, 13]. But logarithms cannot be taken over the complex
numbers, and the algebraic structure of bipotent semirings is often without direct interpretation in tropical
geometry, so attention of tropicalists passed to the field of Puisseux series, which in characteristic 0 also is
an algebraically closed field, but now with a natural valuation, thereby making available tools of valuation
theory, cf. [8]. The collection [7] presents such a valuation theoretic approach. Thus one looks for an
alternative to the max-plus algebra.
AN INFORMAL OVERVIEW OF TRIPLES AND SYSTEMS
3
Properties of the characteristic are described in [58, §4.4]. Most of our major examples have characteristic 0, but some interesting examples have characteristic 2 or more.
1.2.2. Supertropical semirings.
Izhakian [28] overcame many of the structural deficiencies of a max-plus algebra T by adjoining an
extra copy of T , called the ghost copy G in [34, Definition 3.4], and modifying addition. More generally,
a supertropical semiring is a semiring with ghosts (R, G, ν) := T ∪ G, together with a projection
ν : R → G satisfying the extra properties:
(a) (ν-Bipotence) a + b ∈ {a, b}, ∀a, b ∈ R such that ν(a) 6= ν(b);
(b) (Supertropicality) a + b = ν(a) if
ν(a) = ν(b).
The supertropical semiring is standard if ν|T is 1:1.
Mysteriously, although lacking negation, the supertropical semiring provides affine geometry and linear
algebra quite parallel to the classical theory, by taking the negation map (−) to be the identity, where
the ghost ideal G takes the place of the 0 element. In every instance, the classical theorem involving
equality f = g is replaced by an assertion that f = g + ghost, called ghost surpassing. In particular
when g = 0 this means that f itself is a ghost.
For example, an irreducible affine variety should be the set of points which when evaluated at a given
set of polynomials is ghost (not necessarily 0), leading to:
• a version of the Nullstellensatz in [34, Theorem 7.17],
• a link between decomposition of affine varieties and (non-unique) factorization of polynomials,
illustrated in one indeterminate in [34, Remark 8.42 and Theorem 8.46],
• a version of the resultant of polynomials that can be computed by the classical Sylvester matrix
and [37, Theorem 4.12 and Theorem 4.19].
Matrix theory also can be developed along supertropical lines. The supertropical Cayley-Hamilton
theorem [35, Theorem 5.2] says that the characteristic polynomial evaluated on a matrix is a ghost. A
matrix is called singular when its permanent (the tropical replacement of determinant) is a ghost; in [35,
Theorem 6.5] the row rank, column rank, and submatrix rank of a matrix (in this sense) are seen to be
equal. Solution of tropical equations is given in [2, 3, 19, 36]. Supertropical singularity also gives rise to
semigroup versions of the classical algebraic group SL, as illustrated in [33], also cf. [46].
Valuation theory is handled in a series of papers starting with [29], also cf. [43], generalized further
in [30] and [31].
Note that supertropical semirings “almost” are bipotent, in the sense that a1 + a2 ∈ {a1 , a2 } for any
a1 6= a2 in T . This turns out to be an important feature in triples.
1.2.3. Hyperfields and other related constructions.
Another algebraic construction is hyperfields [61], which are multiplicative groups in which sets replace
elements when one takes sums. Hyperfields have received considerable attention recently [5, 25, 41] in
part because of their diversity, and in fact Viro’s “tropical hyperfield” matches Izhakian’s construction.
But there are important nontropical hyperfields (such as the hyperfield of signs, the phase hyperfield, and
the “triangle” hyperfield) whose theories we also want to understand along similar lines. In hyperfield
theory, one can replace “zero” by the property that a given set contains 0.
An intriguing phenomenon is that linear algebra over some classes of hyperfields follows classical lines
as in the supertropical case, but the hyperfield of signs provides easy counterexamples to others, as
discussed in [4].
1.2.4. Symmetrization.
This construction uses Gaubert’s “symmetrized algebras” [19, 55, 2, 20] (which he designed for linear
algebra) as a prototype. We start with T , take A = Tb := T × T , and define the “switch map” (−)
by (−)(a0 , a1 ) = (a1 , a0 ). The reader might already recognize this as the first step in constructing the
integers from the natural numbers, where one identifies (a0 , a1 ) with (a′0 , a′1 ) if a0 + a′1 = a′0 + a1 , but
the trick here is to recognize the equivalence relation without modding it out, since everything could
degenerate in the nonclassical applications. Equality (a0 , a1 ) = (b0 , b1 ) often is replaced by the assertion
4
L. ROWEN
(a0 , a1 ) = (b0 , b1 ) + (c, c) for some c ∈ T . The “symmetrized” T -module also can be viewed as a Tb -supermodule (i.e., 2-graded), via the twist action
(a0 , a1 ) ·tw (b0 , b1 ) = (a0 b0 + a1 b1 , a0 b1 + a1 b0 ),
(1.1)
utilized in [40] to define and study the prime spectrum.
1.2.5. Fuzzy rings.
Dress [16] introduced “fuzzy rings” a while ago in connection with matroids, and these also have been
seen recently to be related to hypergroups in [21].
1.3. Negation maps, triples, and systems.
These varied examples and their theories, which often mimic classical algebra, lead one to wonder
whether the parallels among them are happenstance, or whether there is some straightforward axiomatic
framework within which they can all be gathered and simplified. Unfortunately semirings may lack
negation, so we also implement a formal negation map (−) to serve as a partial replacement for negation.
Definition 1.3. A negation map on a T -module (A, ·, +) is a map (−) : (T , +) → (T , +) together
with a semigroup isomorphism
(−) : (A, +) → (A, +),
both of order ≤ 2, written a 7→ (−)a, satisfying
(−)(ab) = ((−)a)b = a((−)b),
∀a ∈ T , b ∈ A.
(1.2)
Obvious examples of negation maps are the identity map, which might seem trivial but in fact is the
one used in supertropical algebra, the switch map ((−)(a0 , a1 ) = (a1 , a0 )) in the symmetrized algebra,
the usual negation map (−)a = −a in classical algebra, and the hypernegation in the definition of
hypergroups. Accordingly, we say that the negation map (−) is of the first kind if (−)a = a for all
a ∈ T , and of the second kind if (−)a 6= a for all a ∈ T .
As indicated earlier, the quasi-zeros take the role customarily assigned to the zero element. In the
supertropical theory the quasi-zeros are the “ghost” elements. In [3, Definition 2.6] the quasi-zeros are
called “balanced elements” and have the form (a, a).
When 1 ∈ T , the element (−)1 determines the negation map, since (−)b = (−)(1b) = ((−)1)b. When
T ⊆ A, several important elements of A then are:
e = 1◦ = 1(−)1,
e′ = e + 1,
e◦ = e(−)e = e + e = 2e.
(1.3)
The most important quasi-zero for us is e. (For fuzzy rings, e = 1 + ǫ.) But e need not absorb in
multiplication; rather, in any semiring† with negation, Definition 1.1 implies
ae = a(−)a = a◦ .
(1.4)
Definition 1.4. A pseudo-triple is a collection (A, T , (−)), where A is a T -module with T ⊂ A, and
(−) is a negation map.
A triple is a pseudo-triple (A, T , (−)), where T generates (A, +).
Our structure of choice, a system, is a quadruple (A, T , (−), ), where (A, T , (−)) is a triple and
is a T -surpassing relation ([58, Definition 2.65]). The main T -surpassing relations are:
• ◦ , defined by a ◦ b if b = a + c◦ for some c.
• (on sets) is ⊆.
The relation has an important theoretical role, replacing “=” and enabling us to define a broader
category than one would obtain directly from universal algebra, cf. [58, §6]. One major reason why
can formally replace equality in much of the theory is found in the “transfer principle” of [3], given in
the context of systems in [58, Theorem 6.17].
Example 1.5. The four main non-classical examples are:
• (The standard supertropical triple) (A, T , (−)) where A = T ∪ G as before and (−) is the identity
map. We get the system by taking to be ◦ .
AN INFORMAL OVERVIEW OF TRIPLES AND SYSTEMS
5
• (The symmetrized triple) (Â, Tb , (−)) where  = A × A with componentwise addition, and Tb =
(T × {0}) ∪ ({0} × T ) with multiplication Tb × Ab → Ab given by
(a0 , a1 )(b0 , b1 ) = (a0 b0 + a1 b1 , a0 b1 + a0 b1 ).
Here we take (−) to be the switch map ((−)(a0 , a1 ) = (a1 , a0 )), which is of second kind. Again
is ◦ .
• (The fuzzy triple [58, Appendix A]) For any T -monoid module A with an element 1′ ∈ T satisfying
(1′ )2 = 1, we can define a negation map (−) on T and A given by a 7→ 1′ a. In particular,
(−)1 = 1′ . This enables us to view fuzzy rings as systems. Again is ◦ . The same argument
shows that tracts [6] are systems.
• (The hyperfield pseudo-triple [58, §2.4.1]) (P(T ), T , (−)) where T is the original hyperfield, P(T )
is its power set (with componentwise operations), and (−) on the power set is induced from the
hypernegation. Here is ⊆.
Although we introduced pseudo-triples since T need not generate (A, +) (for example, taking A =
P(T ) for the phase hypergroup), we are more concerned with triples, and furthermore in a pseudo-triple
one can take the sub-triple generated by T . More triples and systems related to tropical algebra are
presented in [58, §3.2].
Structures other than monoids also are amenable to such an approach. This can all be formulated
axiomatically in the context of universal algebra, as treated for example in [38]. Once the natural
categorical setting is established, it provides the context in which tropicalization becomes a functor,
thereby providing guidance to understand tropical versions of an assortment of mathematical structures.
1.3.1. Ground triples versus module triples.
Classical structure theory involves the investigation of an algebraic structure as a small category (for
example, viewing a monoid as a category with a single object whose morphisms are its elements), and
homomorphisms then are functors between two of these small categories. On the other hand, one obtains
classical representation theory via an abelian category, such as the class of modules over a given ring.
Analogously, there are two aspects of triples. We call a triple (resp. system) a ground triple
(resp. ground system) when we study it as a small category with a single object in its own right,
usually a semidomain. Ground triples have the same flavor as Lorscheid’s blueprints (albeit slightly
more general, and with a negation map), whereas representation theory leads us to module systems,
described below in §4.1 and §6.
This situation leads to a fork in the road: The first path takes us to a structure theory based on
functors of ground systems, translating into homomorphic images of systems via congruences in [45, §7]
(especially prime systems, in which the product of non-trivial congruences is nontrivial [45, §7.2]).
Ground systems often are designated in terms of the structure of A or T , such as “semiring systems” or
“Hopf systems” or “hyperfield systems.”
The paper [4] has a different flavor, dealing with matrices and linear algebra over ground systems, and
focusing on subtleties concerning Cramer’s rule and the equality of row rank, column rank, and submatrix
rank.
The second path takes us to categories of module systems. In [45] we also bring in tensor products and
Hom functors. This is applied to geometry in [46]. In [44] we develop the homological theory, relying on
work done already by Grandis [23] under the name of N -category and homological category (without
the negation map), and there is a parallel approach of Connes and Consani in [13].
2. Contents of [58]: meta-tangible systems
The emphasis in [58] is on ground systems. One can apply the familiar constructions and concepts of
classical algebra (direct sums [58, Definition 2.10], matrices [58, §4.5], involutions [58, §4.6], polynomials
[58, §4.7], localization [58, §4.8], and tensor products [58, Remark 6.34]) to produce new triples and
systems. The simple tensors a ⊗ b where a, b ∈ T comprise the tangible elements of the tensor product.
The properties of tensors and Hom are treated in much greater depth in [45], and localization is analyzed
for module systems in [46].
6
L. ROWEN
2.0.1. Basic properties of triples and systems.
Let us turn to important properties which could hold in triples. One crucial axiom for this theory,
holding in all tropical situations and most related theories, is:
Definition 2.1. The T -triple (A, T , (−)) is uniquely negated if a + b ∈ A◦ for a, b ∈ T implies
b = (−)a.
By definition, hyperfield triples are uniquely negated. We can hone in further, to obtain two of the
principal concepts of [58]:
Definition 2.2. A uniquely negated T -triple (A, T , (−)) is meta-tangible, if the sum of two tangible
elements is tangible unless they are quasi-negatives of each other.
A special case: (A, T , (−)) is (−)-bipotent if a + b ∈ {a, b} whenever a, b ∈ T with b 6= (−)a. In other
words, a + b ∈ {a, b, a◦ } for all a, b ∈ T .
The stipulation in the definition that b 6= (−)a is of utmost importance, since otherwise all of our main
examples would fail. [58, Proposition 5.64] shows how to view a meta-tangible triple as a hypergroup,
thereby enhancing the motivation of transferring hyperfield notions to triples and systems in general.
Any meta-tangible triple satisfying T ∩ A◦ = ∅ is uniquely negated. The supertropical triple is (−)bipotent, as is the modification of the symmetrized triple described in [58, Example 2.53].
The Krasner hyperfield triple (which is just the supertropicalization of the Boolean semifield B) and
the triple arising from the hyperfield of signs (which is just the symmetrization of B) are (−)-bipotent,
but the phase hyperfield triple and the triangle hyperfield triple are not even metatangible (although
the latter is idempotent). But as seen in [58, Theorem 4.46], hyperfield triples satisfy another different
property of independent interest:
Definition 2.3. [58, Definition 4.13] A surpassing relation in a system is called T -reversible if
a b + c implies b a(−)c for a, b ∈ T .
The category of hyperfields as given in [41] can be embedded into the category of uniquely negated
T -reversible systems ([58, Theorem 6.7]). Reversibility enables one to apply systems to matroid theory,
although we have not yet embarked on that endeavor in earnest.
The height of an element c ∈ A (sometimes called “width” in the group-theoretic literature) is the
Pt
minimal t such that c = i=1 ai with each ai ∈ T . (We say that 0 has height 0.) By definition, every
element of a triple has finite height. The height of A is the maximal height of its elements, when these
heights are bounded. For example, the supertropical semiring has height 2, as does the symmetrized
semiring of an idempotent semifield T .
Some unexpected examples of meta-tangible systems sneak in when the triple has height ≥ 3, as
described in [58, Examples 2.73].
In [58, §5] the case is presented that meta-tangibility could be the major axiom in the theory of ground
systems over a group T , leading to a bevy of structure theorems on meta-tangible systems, starting with
the observation [58, Lemma 5.5] that for a, b ∈ T either a = (−)b, a + b = a (and thus a◦ + b = a◦ ), or
a◦ + b = b.
In [58, Proposition 5.17] the following assertions are seen to be equivalent for a triple (A, T , (−))
containing 1:
(i) T ∪ T ◦ = A,
(ii) A is meta-tangible of height ≤ 2,
(iii) A is meta-tangible with e′ ∈ {1, e}.
We obtain the following results for a meta-tangible system (A, T , (−), ):
• [58, Theorem 5.20] If A is not (−)-bipotent then e′ = 1, with A of characteristic 2 when (−) is
of the first kind.
• [58, Theorem 5.27] Every element has the form c◦ or mc for c ∈ T and m 6= 2. The extent to
which this presentation is unique is described in [58, Theorem 5.31].
• [58, Theorem 5.33] Distributivity follows from the other axioms.
• [58, Theorem 5.34] The surpassing relation must “almost” be ◦ .
• [58, Theorem 5.37] A key property of fuzzy rings holds.
• [58, Theorem 5.43] Reversibility holds, except in one pathological situation.
AN INFORMAL OVERVIEW OF TRIPLES AND SYSTEMS
7
• [58, Theorem 5.55] A criterion is given in terms of sums of squares for (A, T , (−)) to be isomorphic
to a symmetrized triple.
One would want a classification theorem of meta-tangible systems that reduces to classical algebras,
the standard supertropical semiring, the symmetrized semiring, layered semirings, power sets of various
hyperfields, or fuzzy rings, but there are several exceptional cases. Nonetheless, [58, Theorem 5.56] comes
close. Namely, if (−) is of the first kind then either A has characteristic 2 and height 2 with A◦ bipotent,
or (A, T , (−)) is isomorphic to a layered system. If (−) is of the second kind then either T is (−)-bipotent,
with A of height 2 (except for an exceptional case), A is isometric to a symmetrized semiring when A is
real, or A is classical. More information about the exceptions are given in [58, Remark 5.57].
[58, §7] continues with some rudiments of linear algebra over a ground triple, to be discussed shortly.
In [58, §8], tropicalization is cast in terms of a functor on systems (from the classical to the nonclassical).
This principle enables one in [58, §9] to define the “right” tropical versions of classical algebraic structures,
including exterior algebras, Lie algebras, Lie superalgebras, and Poisson algebras.
3. Contents of [4]: Linear algebra over systems
The paper [4] was written with the objective of understanding some of the diverse theorems in linear
algebra
over semiring systems. We define a set of vectors {vi ∈ A(n) : i ∈ I} to be T -dependent if
P
◦ (n)
for some nonempty subset I ′ ⊆ I and αi ∈ T , and the row rank of a matrix to be
i∈I ′ αi vi ∈ (A )
the maximal number of T -independent rows.
A tangible vector is a vector all of whose elements are in T0 . A tangible matrix is a matrix all of
whose rows are tangible vectors.
The (−)-determinant |A| of an n × n matrix A = ai,j is
X
(−)π aπ ,
π∈Sn
Qn
where aπ := i=1 ai,π(i) . A matrix A is nonsingular if |A| ∈ T .
The submatrix rank of A is the largest size of a nonsingular square submatrix of A.
In [58, Corollary 7.4] we see that the submatrix rank of a matrix over a T -cancellative meta-tangible
triple is less than or equal to both the row rank and the column rank. This is improved in [4]:
Theorem 3.1 ([4, Theorem 4.7(i)]). Let (A, T , (−), ◦ ) be a system. For any vector v, the vector
1
1
y = |A|
adj(A)v satisfies |A|v ◦ Ay. In particular, if |A| is invertible in T , then x := |A|
adj(A)v satisfies
v ◦ Ax.
The existence of a tangible such x is subtler. One considers valuations of systems ν : (A, +) → (G, +),
and their fibers {a ∈ T : ν(a) = g} for g ∈ G; we call the system T -Noetherian if any ascending chain
of fibers stabilizes.
Theorem 3.2 ([4, Corollary 4.26]). In a T -Noetherian G-valued system A, if |A| is invertible, then for
1
any vector v, there is a tangible vector x with |x| = |A|
| adj(A)|v, such that Ax + v ∈ A◦ .
One obtains uniqueness of x [4, Theorem 4.7(ii)] using a property called “strong balance elimination.”
After translating some more of the concepts of [3] into the language of systems, we turn to the question,
raised privately for hyperfields by Baker:
Question A. When does the submatrix rank equal the row rank?
Our initial hope was that this would always be the case, in analogy to the supertropical situation.
However, Gaubert observed that a (nonsquare) counterexample to Question A already can be found in [2],
and the underlying system even is meta-tangible. Here the kind of negation map is critical: A rather
general counterexample for triples of the second kind is given in [4, Proposition 3.3]; the essence of the
example already exists in the “sign hyperfield.” Although the counterexample as given is a nonsquare
(3 × 4) matrix, it can be modified to an n × n matrix for any n ≥ 4. This counterexample is minimal in
the sense that Question A has a positive answer for n ≤ 2 and for 3 × 3 matrices under a mild assumption,
cf. [4, Theorem 5.7].
Nevertheless, positive results are available. A positive answer for Question A along the lines of [3,
Theorems 5.11, 5.20, 6.9] is given in [4, Theorem 5.8] for systems satisfying certain technical conditions. In
8
L. ROWEN
[4, Theorem 5.11], we show that Question A has a positive answer for square matrices over meta-tangible
triples of first kind of height 2, and this seems to be the “correct” framework in which we can lift theorems
from classical algebra. A positive answer for all rectangular matrices is given in [4, Theorem 5.19], but
with restrictive hypotheses that essentially reduce to the supertropical situation.
4. Contents of [45]: Basic categorical considerations
The paper [45] elaborates on the categorical aspects of systems, with emphasis on important functors.
The functor triple C S [45, Definition 3.7] embraces important constructions including the symmetrized
triple and polynomial triples (via the convolution product given before [45, Definition 4.4]).
As in [49, 50], the emphasis on [45] is for T to be a cancellative multiplicative monoid (even a group),
which encompasses many major applications. This slights the Lie theory, and indeed one could consider
Hopf systems, as discussed briefly in [45, §2.1] and in more depth in [46]. Motivation can be found in [59].
An issue that must be confronted is the proper definition of morphism, cf. [45, Definitions 5.1, 5.3]. In
categories arising from universal algebra, one’s intuition would be to take the homomorphisms, i.e., those
maps which preserve equality in the operators. We call these morphisms “strict.” However, this approach
loses some major examples of hypergroups. Applications in tropical mathematics and hypergroups (cf. [41,
Definition 2.3]) tend to depend on the “surpassing relation” ([58, Definition 2.65]), so we are led to
a broader definition called -category in [45, Definition 4.8]. -morphisms ([58, Definition 6.3]) often
provide the correct venue for studying ground systems. On the other hand, [45, Proposition 5.4] gives a
way of verifying that some morphisms automatically are strict.
The situation is stricter for module systems [45, §5, 6] over ground triples. The sticky point here is
that the semigroups of morphisms Mor(A, B) in our T -module categories are not necessarily groups, so
the traditional notion of abelian category has to be replaced by “semi-abelian,” [45, Definition 3.16], and
these lack some of the fundamental properties of abelian categories. The tensor product is only functorial
when we restrict our attention to strict morphisms, [45, Proposition 6.8]!
4.1. Module systems.
In both cases, in the theory of semirings and their modules, homomorphisms are described in terms of
congruences, so congruences should be a focus of the theory. The null congruences contain the diagonal,
and not necessarily zero, and lead us to null morphisms, [45, Definition 3.2]. An alternate way of viewing
c is given in [45, §8.4]. In [45, §9], “Hom” is studied
congruences in terms of “transitive” modules of M
together with its dual, and again one only gets all the desired categorical properties (such as the adjoint
isomorphism [45, Lemma 9.8]) when considering strict morphisms. In this way, the categories comprised
of strict morphisms should be amenable to a categorical view, to be carried out in [46] for geometry and
and in [44] for homology, again at times with a Hopfian flavor.
The functors between the various categories arising in this theory are described in [45, §10], also with
an eye towards valuations of triples.
5. Contents of [46]: Geometry
The paper [46] focuses on module systems, leading to a geometric theory over ground systems, and is
comprised of the following parts:
• Group and Lie semialgebra systems.
• The symmetrization functor from Lorscheid’s blueprints to symmetrized triples.
• Localization theory for module triples over semiring triples.
• A geometrical category theory including a representation theorem for semi-abelian N -categories
with negation.
• Schemes of semiringed spaces.
• Sheaves of module systems.
• A Hopf semialgebra approach to ground systems and module systems.
As in classical algebra, the “prime” systems [58, Definitions 2.11, 2.25] play an important role in
affine geometry, via the Zariski topology [45, §5.3.1], so it is significant that we have a version of the
fundamental theorem of algebra in [45, Theorem 7.29], which implies that a polynomial system over a
prime system is prime [45, Corollary 7.30].
AN INFORMAL OVERVIEW OF TRIPLES AND SYSTEMS
9
6. Contents of [44]: Homology
This is work in progress. We start with a version of split epics (weaker than the classical definition):
Definition 6.1. An epic π : M → N -splits if there is an N -monic ν : N → M such that 1N πν.
In this case, we also say that (π, ν) -splits, and N is a -retract of M.
A module system M = (M, TM , (−), ) is the T --direct sum of subsystems (M1 , TM1 , (−), )
and (M2 , TM2 , (−), ) if TM1 ∩ TM2 = ∅ and every a ∈ M can be written a a1 + a2 for ai ∈ Mi .
This leads to projective module systems.
Definition 6.2. A T -module system P is projective if for any strict epic h : M → M′ of T -module
systems, every morphism f : P → M′ lifts to a morphism f˜ : P → M, in the sense that hf˜ = f.
P is -projective if for any (T -module system) strict epic h : M → M′ , every morphism f : P → M′
-lifts to a morphism f˜ : P → M, in the sense that f hf˜.
Their fundamental properties are then obtained, including the -Dual Basis Lemma [44, Proposition 5.14], leading to -projective resolutions and -projective dimension. Then one obtains a homology
theory in the context of homological categories [23] and derived functors, in connection to the recent work
of Connes and Consani [13].
References
[1] K. Adiprasito, J. Huh, and E. Katz, Hodge Theory of Matroids, Notices of the AMS 64 (2017), 26–30.
[2] M. Akian, S. Gaubert, and A. Guterman. Linear independence over tropical semirings and beyond. In Tropical and
Idempotent Mathematics, G.L. Litvinov and S.N. Sergeev, (eds.), Contemp. Math. 495 (2009), 1–38.
[3] M. Akian, S. Gaubert, and A. Guterman. Tropical Cramer Determinants revisited, Contemp. Math. 616, Amer. Math.
Soc. (2014), 1–45.
[4] M. Akian, S. Gaubert, and L. Rowen, Linear algebra over systems, preprint (2017).
[5] M. Baker and N. Bowler, Matroids over hyperfields, arXiv:1601.01204v4, 24 Aug 2016.
[6] M. Baker and N. Bowler, Matroids over partial hyperstructures, preprint, Sept. 2017.
[7] M. Baker and S. Payne (ed), Nonarchimedean and Tropical Geometry , Simons Symposia, Springer-Verlag, 2016.
[8] V. Berkovich, Non-archimedean analytic geometry: first steps in p-adic Geometry: Lectures from the 2007 Arizona
Winter School, University Lecture Series, vol. 45, American Mathematical Society, Providence, R.I., (2008), 1–7.
[9] V. Berkovich, Algebraic and analytic geometry over the field with one element (2014).
[10] A. Bertram and R. Easton, The tropical nullstellensatz for congruences, Advances in Mathematics 308 (2017), 36–82.
[11] N. Bourbaki. Commutative Algebra. Paris and Reading, 1972.
[12] P. Butkovic. Max-algebra: the linear algebra of combinatorics? Lin. Alg. and Appl. (2003), 313–335.
[13] A. Connes and C. Consani, Homological algebra in characteristic one, arXiv:1703.02325v1 [math.AG] 7 Mar 2017.
[14] G. Cortinas, C. Haesemeyer, M. Walker, and C. Weibel, Toric varieties, monoid schemes, and cdh-descent, J. Reine
Angew. Math. 698 (2015), 1-54.
[15] A.A. Costa. Sur la thêorie générale des demi-anneaux, Publ. Math. Decebren 10 (1963), 14–29.
[16] A. Dress, Duality theory for finite and infinite matroids with coefficients, Advances in Mathematics 93(2) (1986),
214–250.
[17] A. Dress and W. Wenzel, Algebraic, tropical, and fuzzy geometry, Beitrage zur Algebra und Geometrie/ Contributions
to Algebra und Geometry 52 (2) (2011), 431–461.
[18] P. Etingof, S. Gelaki, D. Nikshych,and V. Ostrik, Tensor Categories. Mathematical Surveys and Monographs Volume
205, American Mathematical Society.
[19] S. Gaubert. Théorie des systèmes linéaires dans les diodes. Thèse, École des Mines de Paris, 1992.
[20] S. Gaubert and M. Plus. Methods and applications of (max,+) linear algebra. In R. Reischuk and M. Morvan, editors,
STACS97, number 1200 in LNCS, Lubeck, March 1997. Springer.
[21] J. Giansiracusa, J. Jun, and O. Lorscheid, On the relation between hyperrings and fuzzy rings.
[22] J. Golan. The theory of semirings with applications in mathematics and theoretical computer science, volume 54.
Longman Sci & Tech., 1992.
[23] M. Grandis. Homological Algebra: In Strongly Non-abelian Settings, World scientific, 2013.
[24] A. Grothendieck, Produits tensoriels topologiques et espaces nucleaires, Memoirs of the Amer. Math. Soc., no. 16, 1955.
[25] S. Henry, Symmetrization of monoı̈ds as hypergroups, arXiv preprint arXiv:1309.1963 (2013).
[26] I. Itenberg, V. Kharlamov, and E. Shustin, Welschinger invariants of real Del Pezzo surfaces of degree ≥ 3, Math.
Annalen 355 (2013), no. 3, 849-878.
[27] I. Itenberg, G. Mikhalkin and E. Shustin. Tropical Algebraic Geometry, Oberwolfach Seminars, 35, Birkhäuser Verlag,
Basel, 2007.
[28] Z. Izhakian, Tropical arithmetic & algebra of tropical matrices, Preprint at arXiv:math.AG/0505458, 2005.
[29] Z. Izhakian, M. Knebusch, and L. Rowen. Supertropical semirings and supervaluations. J. Pure and Appl. Alg., 215(10)
(2011), 2431–2463.
10
L. ROWEN
[30] Z. Izhakian, M. Knebusch, and L. Rowen. Layered tropical mathematics, Journal of Algebra, 416 (2014), 200–273.
[31] Z. Izhakian, M. Knebusch, and L. Rowen. Categories of layered semirings, Commun. in Algebra 43(5) (2015), 1807–1836.
[32] Z. Izhakian, M. Knebusch, and L. Rowen. Algebraic structures of tropical mathematics, in Tropical and Idempotent
Mathematics, G. L. Litvinov and S. N. Sergeev, (eds.), Contemporary Mathematics (AMS) 616(28) (2014), 125–150.
(Preprint at arXiv:1305.3906.)
[33] Z. Izhakian, A. Niv, and L. Rowen. Supertropical SL, Linear and multilinear algebra, to appear.
[34] Z. Izhakian and L. Rowen. Supertropical algebra, Advances in Mathematics, 225(4) (2010), 2222–2286.
[35] Z. Izhakian and L. Rowen, Supertropical matrix algebra, Israel J. Math., 182(1) (2011), 383–424.
[36] Z. Izhakian and L. Rowen, Supertropical matrix algebra II: Solving tropical equations. Israel J. Math., 186(1) (2011),
69–97.
[37] Z. Izhakian and L. Rowen, Supertropical polynomials and resultants, J. of Algebra, 324 (2010), 1860–1886.
[38] N. Jacobson. Basic algebra II, Freeman, 1980.
[39] D. Jensen and S. Payne, Combinatorial and inductive methods for the tropical maximal rank conjecture, J. Combin.
Theory Ser. A. 152 (2017), 138-158.
[40] D. Joo and K. Mincheva, Prime congruences of idempotent semirings and a Nullstellensatz for tropical polynomials,
arXiv:1408.3817v1 (to appear in Selecta Mathematica).
[41] J. Jun, Algebraic geometry over hyperrings, arxiv:math.AG/1512.04837 (2015).
[42] J. Jun, Cech cohomology of semiring schemes, arXiv preprint arXiv:1503.01389 (2015).
[43] J. Jun, Valuations of semirings (2015), arXiv preprint arXiv:1503.01392 (to appear in Journal of Pure and Applied
Algebra).
[44] J. Jun, K. Mincheva, and L.Rowen, Homology of systems, in preparation.
[45] J. Jun and L.Rowen, Categories with negation, arXiv 1709.0318.
[46] J. Jun and L. Rowen, Geometry of systems (2017).
[47] Y. Katsov, Tensor products of functors, Siberian J. Math. 19 (1978), 222-229, trans. from Sirbiskii Mathematischekii
Zhurnal 19 no. 2 (1978), 318-327.
[48] Y. Katsov, Toward homological characterization of semirings: Serreś conjecture and Bass´ perfectness in a semiring
context, Algebra Universalis 52 (2004), 197-214.
[49] O. Lorscheid, The geometry of blueprints. Part I: algebraic background and scheme theory, Adv. Math. 229 (3) (2012),
pp. 1804–1846.
[50] O. Lorscheid, A blueprinted view on F1-geometry, Absolute Arithmetic and -Geometry, European Mathematical Society
(2016). arXiv:1301.0083v2.
[51] D. Maclagan and B. Sturmfels, Introduction to Tropical Geometry, American Mathematical Society Graduate Studies
in Mathematics 161, 2015.
[52] R. McKenzie, G. McNulty and W. Taylor, Algebras, Lattices and Varieties, Vol. 1 Wadsworth and Brooks, 1987.
[53] G. Mikhalkin. Enumerative tropical algebraic geometry in R2 , J. Amer. Math. Soc 18 (2005), 313–377.
[54] A. Patchkoria, On exactness of long sequences of homology semimodules, Journal of Homotopy and Related Structures,
vol. 1(1), (2006), 229-243.
[55] M. Plus (M. Akian, G. Cohen, S. Gaubert, R. Nikoukhah, and J.P. Quadrat), Max L’algèbre (max,+) et sa symétrisation
ou l’algèbre des équilibres. (French) [(max,+) algebra and its symmetrization or the algebra of balances] C. R. Acad.
Sci. Paris Sér. II Méc. Phys. Chim. Sci. Univers Sci. Terre 311 (1990), no. 4, 443-448.
[56] Q. Ren, K. Shaw, and B. Sturmfels, Tropicalization of Del Pezzo surfaces, Advances in Mathematics 300 (2016),
156–189.
[57] L.H. Rowen, Symmetries in tropical algebra, arXiv:1602.00353 [math.RA]
[58] L.H. Rowen, Algebras with a negation map, 78 pages, arXiv:1602.00353
[59] Y. Sommerhäuser, Hopf algebras and associated group actions, slides of lecture at ACC conference Combinatorics of
group actions, Saint John’s, Newfoundland, August 29, 2017.
[60] B. Toën et M. Vaquié, Au-dessous de Spec Z, Journal K-Theory, 3 (3) (2009), pp. 437–500.
[61] O.Y. Viro, Hyperfields for tropical geometry I. Hyperfields and dequantization. ArXiv AG/1006.3034.
Department of Mathematics, Bar-Ilan University, Ramat-Gan 52900, Israel
E-mail address: [email protected]
| 0math.AC
|
arXiv:1706.08106v1 [cs.AI] 25 Jun 2017
Random Forests for Industrial Device
Functioning Diagnostics Using Wireless Sensor
Networks
W. ElGhazel, C. Guyeux, A. Farhat, M. Hakem,
K. Medjaher, N. Zerhouni, and J.M. Bahi
Abstract
In this paper, random forests are proposed for operating devices diagnostics in the presence of a variable number of features. In various
contexts, like large or difficult-to-access monitored areas, wired sensor
networks providing features to achieve diagnostics are either very costly
to use or totally impossible to spread out. Using a wireless sensor network
can solve this problem, but this latter is more subjected to flaws. Furthermore, the networks’ topology often changes, leading to a variability
in quality of coverage in the targeted area. Diagnostics at the sink level
must take into consideration that both the number and the quality of the
provided features are not constant, and that some politics like scheduling
or data aggregation may be developed across the network. The aim of
this article is (1) to show that random forests are relevant in this context,
due to their flexibility and robustness, and (2) to provide first examples
of use of this method for diagnostics based on data provided by a wireless
sensor network.
1
Introduction
In machine learning, classification refers to identifying the class to which
a new observation belongs, on the basis of a training set and quantifiable
observations, known as properties.
In ensemble learning, the classifiers are combined to solve a particular
computational intelligence problem. Many research papers encourage
adapting this solution to improve the performance of a model, or reduce
the likelihood of selecting a weak classifier. For instance, Dietterich argued that averaging the classifiers’ outputs guarantees a better performance than the worst classifier [6]. This claim was theoretically proven
correct by Fumera and Roli [8]. In addition to this, and under particular
hypotheses, the fusion of multiple classifiers can improve the performance
of the best individual classifier [15].
Two of the early examples of ensemble classifiers are Boosting and
Bagging. In Boosting algorithm [12], the distribution of the training set
1
changes adaptively based on the errors generated by the previous classifiers. In fact, at each step, a higher degree of importance is accorded
to the misclassified instances. At the end of the training, a weight is accorded to each classifier, regarding its individual performance, indicating
its importance in the voting process. As for Bagging [2], the distribution
of the training set changes stochastically and equal votes are accorded to
the classifiers. For both classifiers, the error rate decreases when the size
of the committee increases.
In a comparison made by Tsymbal and Puuronen [14], it is shown that
Bagging is more consistent but unable to take into account the heterogeneity of the instance space. In the highlight of this conclusion, the authors
emphasize the importance of classifiers’ integration. Combining various
techniques can provide more accurate results as different classifiers will
not behave in the same manner faced to some particularities in the training set. Nevertheless, if the classifiers give different results, a confusion
may be induced [10]. It is not easy to ensure reasonable results while combining the classifiers. In this context, the use of random methods could
be beneficial. Instead of combining different classifiers, a random method
uses the same classifier over different distributions of the training set. A
majority vote is then employed to identify the class.
In this article, the use of random forests (RF) is proposed for industrial functioning diagnostics, particularly in the context of devices being
monitored using a wireless sensor network (WSN). A prerequisite in diagnostics is to consider that data provided by sensors are either flawless
or simply noisy. However, deploying a wired sensor network on the monitored device is costly in some well-defined situations, specifically in large
scale, moving, or hardly accessible areas to monitor. Such situations encompass nuclear power plants or any structure spread in deep water or in
the desert. Wireless sensors can be considered in these cases, due to their
low cost and easy deployment.
WSNs monitoring is somehow unique in the sense that sensors too
are subjected to failures or energy exhaustion, leading to a change in
the network topology. Thus, monitoring quality is variable too and it
depends on both time and location on the device. Various strategies can
be deployed on the network to achieve fault tolerance or to extend the
WSN’s lifetime, like nodes scheduling or data aggregation. However, the
diagnostic processes must be compatible with these strategies, and with a
device coverage of a changing quality. The objective of this research work
is to show that RF achieve a good compromise in that situation, being
compatible with a number of sensors which may be variable over time,
some of them being susceptible to errors. More precisely, we will explain
why random methods are relevant to achieve accurate diagnostics of an
industrial device being monitored using a WSN. The functioning of RF
will then be recalled and applied in the monitoring context. Algorithms
will be provided, and an illustration on a simulated WSN will finally be
detailed.
The remainder of this article is organized as follows. Section 2 summarizes the related work. In Section 3, we overview the research works in
industrial diagnostics. We present the random forest algorithm in Section
4 and give simulation results in Section 5. This research work ends with
2
a conclusion section, where the contribution is summarized and intended
future work is provided.
2
Related work
Many research works have contributed in improving the classification’s
accuracy. For instance, tree ensembles use majority voting to identify
the most popular class. They have the advantage of transforming weak
classifiers into strong ones by combining their knowledge to reduce the
error rate.
Usually, the growth of each tree is governed by random vectors sampled
from the training set, and bagging is one of the early examples of this. In
this method, each tree is grown by randomly selecting individuals from
the training set without replacing them [2]. The use of bagging can be
motivated by three main reasons: (1) it enhances accuracy with the use
of random features, (2) it gives ongoing estimates of the generalization
error, strength, and correlation of combined trees, and (3) it is also good
for unstable classifiers with large variance.
Meanwhile, Freund introduced the adaptive boosting algorithm Adaboost, which he defined as [7]: “a deterministic algorithm that selects
the weights on the training set for input to the next classifier based on
the wrong classifications in the previous classifiers”.
The fact that the classifier focuses on correcting the errors at each new
step remarkably improved the accuracy of classifications.
Shortly after, in [1] randomness was again used to grow the trees. The
split was defined at each node by searching for the best random selection
of features in the training set. Ho [9] introduced the random subspace,
in which he randomly selects a subset of vectors of features to grow each
tree. Diettrich introduced the random split selection where at each node,
a split is randomly selected among k best splits [6].
For these methods, and like bagging, a random vector sampled to
grow a tree is completely independent from the previous vectors, but is
generated with the same distribution.
Random split selection [6] and introducing random noise into the outputs [3] both gave better results than bagging. Nevertheless, the algorithms implementing ways of re-weighting the training set, such as Adaboost [7], outperform these two methods [4].
Therefore, Breiman combined the strengths of the methods detailed
above into the random forest algorithm. In this method, individuals are
randomly selected from the training set with replacement. At each node,
a split is selected by reducing the dispersion generated by the previous
step and consequently lowering the error rate. This algorithm is further
detailed in Section 4.
3
Overview of diagnostics
With their constantly growing complexity, current industrial systems witness costly downtime and failures. Therefore, an efficient health assess-
3
ment technique is mandatory. In fact, in order to avoid expensive shutdowns, maintenance activities are scheduled to prevent interruptions in
system operation. In early frameworks, maintenance takes place either
after a failure occurs (corrective maintenance), or according to predefined
time intervals (periodic maintenance). Nevertheless, this still generates
extra costs due to “too soon” or “too late” maintenances. Accordingly,
considering the actual health state of the operating devices is important in
the decision making process. Maintenance here becomes condition-based,
and is only performed after the system being diagnosed in a certain health
state.
Diagnostics is an understanding of the relationship between what we
observe in the present and what happened in the past, by relating the
cause to the effect. After a fault takes place, and once detected, an
anomaly is reported in the system behavior. The fault is then isolated
by determining and locating the cause (or source) of the problem. Doing
so, the component responsible for the failure is identified and the extent
of the current failure is measured. This activity should meet several requirements in order to be efficient [5]. these requirements are enumerated
in the following.
• Early detection: in order to improve industrial systems’ reliability, fault detection needs to be quick and accurate. Nevertheless,
diagnostic systems need to find a reasonable trade-off between quick
response and fault tolerance. In other words, an efficient diagnostic
system should differentiate between normal and erroneous performances in the presence of a fault.
• Isolability: fault isolation is a very important step in the diagnostic process. It refers to the ability of a diagnostic system to
determine the source of the fault and identify the responsible component. With the isolability attribute, the system should discriminate
between different failures. When an anomaly is detected, a set of
possible faults is generated. While the completeness aspect requires
the actual faults to be a subset of the proposed set, resolution optimization necessitates that the set is as small as possible. A tradeoff
then needs to be found while respecting the accuracy of diagnostics.
• Robustness and resources: it is highly desirable that the diagnostic system would degrade gracefully rather than fail suddenly. For
this finality, the system needs to be robust to noise and uncertainties. In addition to this, a trade-off between system performance and
computational complexity is to be considered. For example, on-line
diagnostics require low complexity and higher storage capacities.
• Faults identifiability: a diagnostics system is of no interest if it
cannot distinguish between normal and abnormal behaviors. It is
also crucial that not only the cause of every fault is identified, but
also that new observations of malfunctioning would not be misclassified as a known fault or as normal behavior. While it is very common
that a present fault leads to the generation of other faults, combining
the effects of these faults is not that easy to achieve due to a possible non-linearity. On the other hand, modeling the faults separately
may exhaust the resources in case of large processes.
4
• Clarity: when diagnostic models and human expertise are combined
together, the decision making support is more reliable. Therefore, it
is appreciated that the system explains how the fault was triggered
and how it propagated, and keeps track on the cause/effect relationship. This can help the operator use their experience to evaluate the
system and understand the decision making process.
• Adaptability: operating conditions, external inputs, and environmental conditions change all the time. Thus, to ensure relevant diagnostics at all levels,the system should adapt to changes and evolve
in the presence of new information.
Existent diagnostic models have several limitations. Some of which
are summarized in Table 1.
Diagnostic model
Markovian process
Bayesian networks
Neural networks
Fuzzy systems
Drawbacks
-Aging is not considered
-Different stages of degradation process
cannot be accounted for
-Large volume of data is required for the
training
-The assumptions are not always practical
-Prior transitions are not considered
-Complete reliance on accurate thresholds
-Many state transitions are needed for efficient results
-Unable to predict unanticipated states
-Significant amount of data for the training
-Retraining is necessary with every change
of conditions
-Pre-processing is needed to reduce inputs
-Increasing complexity with every new entry
-Domain experts are required
-Results are as good as the developers’ understanding
Table 1: Limitations of diagnostic models
The degradation process can be considered as a stochastic process.
The evolution of the degradation is a random variable that describes the
different levels of the system’s health state, from good condition to complete deterioration. The deterioration process is multistate and can be
divided into two main categories [11]:
1. Continuous-state space: the device is considered failed when the
predefined threshold is reached.
2. Discrete-state space: the degradation process is divided into a finite
number of discrete levels.
5
As condition-based maintenance relies on reliable scheduling of maintenance activities, an understanding of the degradation process is required.
For this finality, in this paper, we consider the discrete-state space deterioration process.
4
Random forests
The RF algorithm is mainly the combination of Bagging [2] and random
subspace [9] algorithms, and was defined by Leo Breiman as “a combination of tree predictors such that each tree depends on the values of
a random vector sampled independently and with the same distribution
for all trees in the forest” [4]. This method resulted from a number of
improvements in tree classifiers’ accuracy.
This classifier maximizes the variance by injecting randomness in variable selection, and minimizes the bias by growing the tree to a maximum
depth (no pruning). The steps of constructing the forest are detailed in
Algorithm 1.
Algorithm 1 Random forest algorithm
Input: Labeled training set S, Number of trees T , Number of features F .
Output: Learned random forest RF .
initialize RF as empty
for i in 1..T do
Si0 ← bootstrap (S)
initialize the root of tree i
repeat
if current node is terminal then
affect a class
go to the next unvisited node if any
else
select the best feature f ∗ among F
sub-tree ← split(Si0 , f ∗ )
add (leftChild, rightChild) to tree i
end if
until all nodes are visited
add tree i to the forest
end for
In a RF, the root of a tree i contains the instances from the training
subset Si0 , sorted by their corresponding classes. A node is terminal if
it contains instances of one single class, or if the number of instances
representing each class is equal. In the alternative case, it needs to be
further developed (no pruning). For this purpose, at each node, the feature
that guarantees the best split is selected as follows.
1. The information acquired by choosing a feature can be computed
through:
6
(a) The entropy of Shannon, which measures the quantity of information
Entropy(p) = −
c
X
P (k/p) × log(P (k/p))
(1)
k=1
where p is the number of examples associated to a position in the
tree, c is the total number of classes, k/p denotes the fraction of
examples associated to a position in the tree and labelled class
k, P (k/p) is the proportion of elements labelled class k at a
position p.
(b) The Gini index, which measures the dispersion in a population
Gini(x) = 1 −
c
X
P (k/p)2
(2)
k=1
where x is a random sample, c is the number of classes, k/p
denotes the fraction of examples associated to a position in the
tree and labelled class k, P (k/p) is the proportion of elements
labelled class k at a position p.
2. The best split is then chosen by computing the gain of information from growing the tree at given position, corresponding to each
feature as follows:
Gain(p, t) = f (p) −
n
X
Pj × f (pj )
(3)
j=1
where p corresponds to the position in the tree, t denotes the test at
branch n, Pj is the proportion of elements at position p and that go
to position pj , f (p) corresponds to either Entropy(p) or Gini(p).
The feature that provides the higher Gain is selected to split the
node.
The optimal training of a classification problem can be NP-hard. Tree
ensembles have the advantage of running the algorithm from different
starting points, and this can better approximate the near-optimal classifier.
In his paper, Leo Breiman discusses the accuracy of random Forests.
In particular, he gave proof that the generalized error, although different from one application to another, always has an upper bound and so
random forests converge [4].
The injected randomness can improve accuracy if it minimizes correlation while maintaining strength. The tree ensembles investigated by
Breiman use either randomly selected inputs or a combination of inputs at
each node to grow the tree. These methods have interesting characteristics
as:
- Their accuracy is at least as good as Adaboost
- They are relatively robust to outliers and noise
- They are faster than bagging or boosting
7
- They give internal estimates of error, strength, correlation, and variable importance
- They are simple and the trees can be grown in parallel
There are four different levels of diversity which were defined in [13],
level 1 being the best and level 4 the worst.
• Level 1: no more than one classifier is wrong for each pattern.
• Level 2: the majority voting is always correct.
• Level 3: at least one classifier is correct for each pattern.
• Level 4: all classifiers are wrong for some pattern.
RF can guarantee that at least level two is reached. In fact, a trained
tree is only selected to contribute in the voting if it does better than
random, i.e. the error rate generated by the corresponding tree has to be
less than 0.5, or the tree will be dropped from the forest [4].
In [16], Verikas et al. argue that the most popular classifiers (Support
Vector Machine SVM, MultiLayer Perceptron MLP, and Relevance Vector
Machine RVM) provide too little insight about the variable importance to
the derived algorithm. They compared each of these methodologies to the
random forest algorithm to find that in most cases RF outperform other
techniques by a large margin.
5
Experimental study
5.1
Data collection
In this paper, we consider two sets of experiments. The sensor network is
constituted by 110 nodes, sensing respectively the levels of temperature
(50 sensors), pressure (50), and humidity (10) on the industrial device
under consideration.
5.1.1
Set of experiment 1
In this set of experiments, we consider that no level of correlation is introduced betweent the different features. Moreover, we suppose that at
time t:
• Under normal conditions, temperature sensors follow a Gaussian law
of parameter (20×(1+0.005t), 1), while these parameters are mapped
to (35, 1) in case of a malfunction of the industrial device. Finally,
these sensors return the value 0 when they break down.
• The Gaussian parameters are (5 × (1 + 0.01t), 0.3) when both the
industrial device and the pressure sensors are in normal conditions.
The parameters are changed to (15, 1) in case of industrial failure,
while the pressure sensors return 1 when they are themselves broken
down.
8
• Finally, the 10 humidity sensors produce data following a Gaussian
law of parameter (52.5 × (1 + 0.001t), 12.5) when they are sensing a
well-functioning device. These parameters are set to (70, 10) in case
of device failure, while malfunctioning humidity sensors produce the
value 0.
5.1.2
Set of experiment 2
For this set, a linear correlation is injected between the studied features.
• Under normal conditions, temperature sensors follow a Gaussian law
of parameter (20×(1+0.005t), 1), while these parameters are mapped
to (35, 1) in case of a malfunction of the industrial device. Finally,
these sensors return the value 0 when they break down.
• When both the industrial device and the pressure sensors are in
normal conditions, the value of pressure is computed as (x ÷ 2 + 10),
where x is the value of temperature. The parameters are changed to
(15, 1) in case of industrial failure, while the pressure sensors return
1 when they are themselves broken down.
• For a well-functioning device, the 10 humidity sensors produce data
in the form of (x × 525 + 12). These parameters are set to (70, 10) in
case of device failure, while malfunctioning humidity sensors produce
the value 0.
For both data sets, the probability that a failure occurs at time t
follows a Bernoulli distribution of parameter t ÷ 35000.
Five levels of functioning are attributed to each category of sensors,
depending on the abnormality of the sensed data. These levels are defined
thanks to 4 thresholds, which are 22.9, 24.5, 26, and 28 degrees for the
temperature (a temperature lower than 22.9°C is normal, while a sensed
value larger than 28°C is highly related to a malfunctioning), 5.99, 6.4,
7.9, and 9 bars for the pressure parameter, and finally 68, 80, 92, and 95
percents for the humidity.
Data is generated as follows.
• For each time unit t = 1..100 during the industrial device monitoring,
– For each category c (temperature, pressure, humidity) of sensors:
∗ For each sensor s belonging to category c:
· If s has not yet detected a device failure:
1. s picks a new data, according to the Gaussian law corresponding to a well-functioning device, which depends
on both t and c,
2. a random draw from the exponential law detailed previously is realized, to determine if a breakdown occurs
on the location where s is placed.
· Else s picks a new datum according to the Bernoulli distribution of a category c sensor observing a malfunctioning device.
9
The global failure level F t of a set of 110 sensed data produced by
the wireless sensor network at a given time t is defined as follows. For
each sensed datum dti , i = 1..110, let fit ∈ {1, .., 5} be the functioning
level related to its category (pressure, temperature, or humidity). Then
F t = max fit | i = 1..110.
5.2
Random forest design
Figure 1: Example of a tree in the random forest
The random forest, constituted in this set of experiments by 100 trees,
is defined as follows. For each tree Ti , i = 1..100:
• A sample of 67% of dates τ1 , . . . , τ67 ∈ {1, ..., 100} is extracted
• The root of the tree Ti is the tuple (]{j | F τj = n, j = 1..67})n=1..5 ,
where ]X is the cardinality of the finite set X. Thus, its n-th coordinate corresponds to the number of times the device has been in
the global failure n in this sample of observation dates.
• The category c having the largest Gain for the dates in the root
node is selected. The dates are divided into five sets depending on
thresholds related to c. Then, 5 edges labeled by both c and failure
levels li0 = {1, .., 5} are added to Ti , as depicted in Figure 1. They
are directed to (at most) 5 new vertices containing the tuples
τi
]{j | F τj = n and di j has a c level equal to li }
.
n=1,..,5
In other words, we only consider in this node a sub-sample of dates
having their functioning level for category c equal to li0 , and we divide
the sub-sample into 5 subsets, depending on their global functioning
levels: the tuple is constituted by each cardinality of these subsets,
see Fig. 1.
• The process is continued, with: this vertex as a new root, the reduced
set of observed dates, and the categories minus c. It is stopped when
either all the categories have been regarded, or when tuple of the
node has at least 4 components equal to 0.
10
5.3 Providing a diagnostic on a new set of observations
Finally, given a new set of observations at a given time, the diagnostics
of the industrial device is obtained as follows.
Let T be a tree in the forest. T will be visited starting from its root
until reaching a leaf as described below.
1. All the edges connected to the root of T are labeled with the same
category c, but with various failure levels. The selected edge e is
the one whose labeled level of failure regarding c corresponds to the
c-level of failure of the observations.
2. If the obtained node n following edge e is a leaf, then the global
level of failure of the observations according to T is the coordinate
of the unique non zero component of the tuple. If not, the tree walk
is continued at item 1 with node n as new root.
The global diagnostics for the given observation is a majority consensus
of all the responses of all the trees in the forest.
5.4
Numerical simulations
The training set is obtained by simulating 100 observations for 10 successive times, which results in 1000 instances. The resulting data base is
then used to train 100 trees that will constitute the trained random forest.
Figure 2 presents the delay between the time the system enters a failure mode and the time of its detection. This is done in the absence of
correlations between the different features. The 0 time value of delay, the
negative values, and positive value refer to in time predictions, early predictions and late predictions of failures, respectively. The plotted values
are the average result per number of simulations which varies from 1 to
100. With time, sensor nodes start to fail in order to simulate missing
data packets. As a result, the RF algorithm was able to detect 54% of the
failures either in time or before their occurrence.
For each of the 100 performed simulations, we calculated the average
number of errors in fault detection, produced by the trees in the forest.
Figure 3 shows that this error rate remained below 15% through the simulation. This error rate includes both “too early” and “too late” detections.
When certain sensor nodes stop functioning, this leads to a lack on information, which has an impact on the quality of predictions; this explains a
sudden increase in the error rate with time. We can conclude from the low
error rate in the absence of some data packets that increasing the number
of trees in the RF helps improve the quality and accuracy of predictions.
As described in Section 5.1.2, a correlation was introduced between
the features. Figure 4 shows the number of successful diagnostics when
the number of tree estimators in the forest changes. As shown in this
figure, the RF method guarantees a 60% success rate when the number of
trees is limited to 5. As this number grows, the accuracy of the method
increases to reach 80% when the number of trees is around 100.
11
Figure 2: Delay in failure detection with respect to the number of simulations.
Figure 3: Error rate in diagnostics with respect to the number of simulations
6
Conclusion
Instead of using wired sensor networks for a diagnostics and health management method, it is possible to use wireless sensors. Such a use can be
motivated by cost reasons or due to specific particularities of the monitored device. In the context of a changing number and quality of provided
features, the use of random forests may be of interest. These random classifiers were recalled with details in this article, and the reason behind their
use in the context of a wireless sensors network monitoring was explained.
Finally, algorithms and first examples of use of these random forests for
diagnostics using a wireless sensor network were provided. The simulation
results showed that the algorithm guarantees a certain level of accuracy
12
Figure 4: Number of successful diagnostics with respect to the number of trees.
even when some data packets are missing.
In future work, the authors’ intention is to compare various tools for
diagnostics to the random forests, either when considering wireless sensor
networks or wired ones. Comparisons will be carried out both theoretical
and practical aspects. The algorithm of random forests, for its part, will
be extended to achieve prognostics and health management too. Finally,
the method for diagnosing an industrial device will be tested on a life size
model, to illustrate the effectiveness of the proposed approach.
References
[1] Yali Amit and Donald Geman. Shape quantization and recognition
with randomized trees. Neural Computation, 9:1545–1588, 1997.
[2] Leo Breiman. Bagging predictors. Machine Learning, 24:123–140,
1996.
[3] Leo Breiman. Using adaptive bagging to debias regressions. Technical
report, Statics Department UCB, 1999.
[4] Leo Breiman. Random forests. Machine Learning, 45:5–32, 2001.
[5] Sourabh Dash and Venkat Venkatasubramanian. Challenges in the
industrial applications of fault diagnostic systems. Computers and
Chemical Engineering, 24:785–791, 2000.
[6] Thomas G. Dietterich. An experimental comparison of three methods
for constructing ensembles of decision trees: Bagging, boosting, and
randomization. Machine Learning, 40:139–157, 2000.
13
[7] Y. Freund and R. Schapire. Experiments with a new boosting algorithm. In Proceedings of the Thirteenth International Conference on
Machine Learning, pages 148–156, 1996.
[8] Giorgio Fumera and Fabio Roli. A theoretical and experimental analysis of linear combiners for multiple classifier systems. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(6):942–956,
2005.
[9] Tin Kam Ho. The random subspace method for constructing decision forests. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 20(8):832–844, 1998.
[10] Shigeru Kanemoto, Norihiro Yokotsuka, Noritaka Yusa, and
Masahiko Kawabata. Diversity and integration of rotating machine
health monitoring methods. In Chemical Engineering Transactions,
number 33, pages 169–174, Milan, Italy, 2013.
[11] Ramin Moghaddass and Ming Zuo. An integrated framework for
online diagnostic and prognostic health monitoring using a multistate
deterioration process. Reliability Engieneering and System Safety,
124:92–104, 2014.
[12] Robert E. Schapire. A brief introduction to boosting. In Proceedings
of the sixteenth International Joint Conference on Artificial Intelligence, 1999.
[13] A. Sharkey and N. Sharkey. Combining diverse neural nets. The
Knowledge Egineering Review, 12(3):231–247, 1997.
[14] Alexey Tsymbal and Seppo Puuronen. Bagging and boosting with
dynamic integration of classifiers. In The 4th European Conference
on Principles and Practice of Knowledge Discovery in Data Bases
PKDD, pages 116–125, 2000.
[15] K. Tumer and J. Ghosh. Error correlation and error reduction in
ensemble classifiers. Connection Science, 8:385–404, 1996.
[16] A. Verikas, A. Gelzinis, and M. Bacauskiene. Mining data with random forests: A survey and results of new tests. Pattern Recognition,
44:330–349, 2011.
14
| 2cs.AI
|
IMPROVEMENTS TO DEEP CONVOLUTIONAL NEURAL NETWORKS FOR LVCSR
Tara N. Sainath1, Brian Kingsbury1 , Abdel-rahman Mohamed2 , George E. Dahl2 , George Saon1
Hagen Soltau1, Tomas Beran1, Aleksandr Y. Aravkin1, Bhuvana Ramabhadran1
arXiv:1309.1501v3 [cs.LG] 10 Dec 2013
1
IBM T. J. Watson Research Center, Yorktown Heights, NY 10598
2
Department of Computer Science, University of Toronto
1
{tsainath, bedk, gsaon, hsoltau, tberan, saravkin, bhuvana}@us.ibm.com,
2
{asamir, gdahl}@cs.toronto.edu
ABSTRACT
Deep Convolutional Neural Networks (CNNs) are more powerful
than Deep Neural Networks (DNN), as they are able to better reduce
spectral variation in the input signal. This has also been confirmed
experimentally, with CNNs showing improvements in word error
rate (WER) between 4-12% relative compared to DNNs across a variety of LVCSR tasks. In this paper, we describe different methods
to further improve CNN performance. First, we conduct a deep analysis comparing limited weight sharing and full weight sharing with
state-of-the-art features. Second, we apply various pooling strategies that have shown improvements in computer vision to an LVCSR
speech task. Third, we introduce a method to effectively incorporate
speaker adaptation, namely fMLLR, into log-mel features. Fourth,
we introduce an effective strategy to use dropout during Hessian-free
sequence training. We find that with these improvements, particularly with fMLLR and dropout, we are able to achieve an additional
2-3% relative improvement in WER on a 50-hour Broadcast News
task over our previous best CNN baseline. On a larger 400-hour
BN task, we find an additional 4-5% relative improvement over our
previous best CNN baseline.
1. INTRODUCTION
Deep Neural Networks (DNNs) are now the state-of-the-art in acoustic modeling for speech recognition, showing tremendous improvements on the order of 10-30% relative across a variety of small and
large vocabulary tasks [1]. Recently, deep convolutional neural networks (CNNs) [2, 3] have been explored as an alternative type of
neural network which can reduce translational variance in the input
signal. For example, in [4], deep CNNs were shown to offer a 4-12%
relative improvement over DNNs across different LVCSR tasks. The
CNN architecture proposed in [4] was a somewhat vanilla architecture that had been used in computer vision for many years. The goal
of this paper is to analyze and justify what is an appropriate CNN architecture for speech, and to investigate various strategies to improve
CNN results further.
First, the architecture proposed in [4] used multiple convolutional layers with full weight sharing (FWS), which was found to be
beneficial compared to a single FWS convolutional layer. Because
the locality of speech is known ahead of time, [3] proposed the use
of limited weight sharing (LWS) for CNNs in speech. While LWS
has the benefit that it allows each local weight to focus on parts of
the signal which are most confusable, previous work with LWS had
just focused on a single LWS layer [3], [5]. In this work, we do a
detailed analysis and compare multiple layers of FWS and LWS.
Second, there have been numerous improvements to CNNs in
computer vision, particularly for small tasks. For example, using
lp [6] or stochastic pooling [7] provides better generalization than
max pooling used in [4]. Second, using overlapping pooling [8] and
pooling in time [9] also improves generalization to test data. Furthermore, multi-scale CNNs [6], that is, combining outputs from
different layers of the neural network, has also been successful in
computer vision. We explore the effectiveness of these strategies for
larger scale speech tasks.
Third, we investigate using better features for CNNs. Features
for CNNs must exhibit locality in time and frequency. In [4] it
was found that VTLN-warped log-mel features were best for CNNs.
However, speaker adapted features, such as feature space maximum
likelihood linear regression (fMLLR) features [10], typically give
the best performance for DNNs. In [4], the fMLLR transformation
was applied directly to a correlated VTLN-warped log-mel space.
However, no improvement was observed as fMLLR transformations
typically assume uncorrelated features. In this paper, we propose a
methodology to effectively use fMLLR with log-mel features. This
involves transforming log-mel into an uncorrelated space, applying
fMLLR in this space, and then transforming the new features back
to a correlated space.
Finally, we investigate the role of rectified linear units (ReLU)
and dropout for Hessian-free (HF) sequence training [11] of CNNs.
In [12], ReLU+dropout was shown to give good performance for
cross-entropy (CE) trained DNNs but was not employed during HF
sequence-training. However, sequence-training is critical for speech
recognition performance, providing an additional relative gain of 1015% over a CE-trained DNN [11]. During CE training, the dropout
mask changes for each utterance. However, during HF training, we
are not guaranteed to get conjugate directions if the dropout mask
changes for each utterance. Therefore, in order to make dropout
usable during HF, we keep the dropout mask fixed per utterance for
all iterations of conjugate gradient (CG) within a single HF iteration.
Results with the proposed strategies are first explored on a 50-hr
English Broadcast News (BN) task. We find that there is no difference between LWS and FWS with multiple layers for an LVCSR
task. Second, we find that various pooling strategies that gave improvements in computer vision tasks, do not help much in speech.
Third, we observe that improving the CNN input features by including fMLLR gives improvements in WER. Finally, fixing the dropout
mask during the CG iterations of HF lets us use dropout during HF
sequence training and avoids destroying the gains from dropout accrued during CE training. Putting together improvements from fMLLR and dropout, we find that we are able to obtain a 2-3% relative
reduction in WER compared to the CNN system proposed in [4]. In
addition, on a larger 400-hr BN task, we can also achieve a 4-5%
relative improvement in WER.
The rest of this paper is organized as follows. Section 2 describes the basic CNN architecture in [4] that serves as a starting
point to the proposed modifications. In Section 3, we discuss experiments with LWS/FWS, pooling, fMLLR and ReLU+dropout for
HF. Section 4 presents results with the proposed improvements on a
50 and 400-hr BN task. Finally, Section 5 concludes the paper and
discusses future work.
2. BASIC CNN ARCHITECTURE
In this section, we describe the basic CNN architecture that was introduced in [4], as this will serve as the baseline system which we
improve upon. In [4], it was found that having two convolutional layers and four fully connected layers, was optimal for LVCSR tasks.
We found that a pooling size of 3 was appropriate for the first convolutional layer, while no pooling was used in the second layer. Furthermore, the convolutional layers had 128 and 256 feature maps respectively, while the fully connected layers had 1,024 hidden units.
The optimal feature set used was VTLN-warped log-mel filterbank
coefficients, including delta + double delta. Using this architecture
for CNNs, we were able to achieve between 4-12% relative improvement over DNNs across many different LVCSR tasks.
In this paper, we explore feature, architecture and optimization
strategies to improve the CNN results further. Preliminary experiments are performed on a 50-hr English Broadcast News task [11].
The acoustic models are trained on 50 hours from the 1996 and
1997 English Broadcast News Speech Corpora. Results are reported
on the EARS dev04f set. Unless otherwise noted, all CNNs are
trained with cross-entropy, and results are reported in a hybrid setup.
3. ANALYSIS OF VARIOUS STRATEGIES FOR LVCSR
3.1. Optimal Feature Set
Convolutional neural networks require features which are locally
correlated in time and frequency. This implies that Linear Discriminant Analysis (LDA) features, which are very commonly used in
speech, cannot be used with CNNs as they remove locality in frequency [3]. Mel filter-bank (FB) features are one type of speech
feature which exhibit this locality property [?]. We explore if any
additional transformations can be applied to these features to further
improve WER. Table 3 shows the WER as a function of input feature
for CNNs. The following can be observed:
• Using VTLN-warping to help map features into a canonical
space offers improvements.
• Using fMLLR to further speaker-adapt the input does not
help. One reason could be that fMLLR assumes the data is
well modeled by a diagonal model, which would work best
with decorrelated features. However, the mel FB features are
highly correlated.
• Using delta and double-delta (d + dd) to capture further timedynamic information in the feature helps.
• Using energy does not provide improvements.
In conclusion, it appears VTLN-warped mel FB + d+dd is the
optimal input feature set to use. This feature set is used for the remainder of the experiments, unless otherwise noted.
Feature
Mel FB
VTLN-warped mel FB
VTLN-warped mel FB + fMLLR
VTLN-warped mel FB + d + dd
VTLN-warped mel FB + d + dd + energy
WER
21.9
21.3
21.2
20.7
21.0
Table 1. WER as a function of input feature
3.2. Number of Convolutional vs. Fully Connected Layers
Most CNN work in image recognition makes use of a few convolutional layers before having fully connected layers. The convolutional layers are meant to reduce spectral variation and model spectral correlation, while the fully connected layers aggregate the local
information learned in the convolutional layers to do class discrimination. However, the CNN work done thus far in speech [3] introduced a novel framework for modeling spectral correlations, but this
framework only allowed for a single convolutional layer. We adopt
a spatial modeling approach similar to the image recognition work,
and explore the benefit of including multiple convolutional layers.
Table 2 shows the WER as a function of the number of convolutional and fully connected layers in the network. Note that for
each experiment, the number of parameters in the network is kept the
same. The table shows that increasing the number of convolutional
layers up to 2 helps, and then performance starts to deteriorate. Furthermore, we can see from the table that CNNs offer improvements
over DNNs for the same input feature set.
# of Convolutional vs.
Fully Connected Layers
No conv, 6 full (DNN)
1 conv, 5 full
2 conv, 4 full
3 conv, 3 full
WER
24.8
23.5
22.1
22.4
Table 2. WER as a Function of # of Convolutional Layers
3.3. Number of Hidden Units
CNNs explored for image recognition tasks perform weight sharing across all pixels. Unlike images, the local behavior of speech
features in low frequency is very different than features in high frequency regions. [3] addresses this issue by limiting weight sharing
to frequency components that are close to each other. In other words,
low and high frequency components have different weights (i.e. filters). However, this type of approach limits adding additional convolutional layers [3], as filter outputs in different pooling bands are not
related. We argue that we can apply weight sharing across all time
and frequency components, by using a large number of hidden units
compared to vision tasks in the convolutional layers to capture the
differences between low and high frequency components. This type
of approach allows for multiple convolutional layers, something that
has thus far not been explored before in speech.
Table 3 shows the WER as a function of number of hidden units
for the convolutional layers. Again the total number of parameters
in the network is kept constant for all experiments. We can observe
that as we increase the number of hidden units up to 220, the WER
steadily decreases. We do not increase the number of hidden units
past 220 as this would require us to reduce the number of hidden
units in the fully connected layers to be less than 1,024 in order to
keep the total number of network parameters constant. We have observed that reducing the number of hidden units from 1,024 results in
an increase in WER. We were able to obtain a slight improvement by
using 128 hidden units for the first convolutional layer, and 256 for
the second layer. This is more hidden units in the convolutional layers than are typically used for vision tasks [2], [9], as many hidden
units are needed to capture the locality differences between different
frequency regions in speech.
Number of Hidden Units
64
128
220
128/256
WER
24.1
23.0
22.1
21.9
Table 3. WER as a function of # of hidden units
3.4. Limited vs. Full Weight Sharing
In speech recognition tasks, the characteristics of the signal in lowfrequency regions are very different than in high frequency regions.
This allows a limited weight sharing (LWS) approach to be used
for convolutional layers [3], where weights only span a small local
region in frequency. LWS has the benefit that it allows each local
weight to focus on parts of the signal which are most confusable, and
perform discrimination within just that small local region. However,
one of the drawbacks is that it requires setting by hand the frequency
region each filter spans. Furthermore, when many LWS layers are
used, this limits adding additional full-weight sharing convolutional
layers, as filter outputs in different bands are not related and thus the
locality constraint required for convolutional layers is not preserved.
Thus, most work with LWS up to this point has looked at LWS with
one layer [3], [5].
Alternatively, in [4], a full weight sharing (FWS) idea in convolutional layers was explored, similar to what was done in the image
recognition community. With that approach, multiple convolutional
layers were allowed and it was shown that adding additional convolutional layers was beneficial. In addition, using a large number
of hidden units in the convolutional layers better captures the differences between low and high frequency components.
Since multiple convolutional layers are critical for good performance in WER, in this paper we explore doing LWS with multiple
layers. Specifically, the activations from one LWS layer have locality preserving information, and can be fed into another LWS layer.
Results comparing LWS and FWS are shown in Table 4. Note these
results are with stronger VTLN-warped log-mel+d+dd features, as
opposed to previous LWS work which used simpler log-mel+d+dd.
For both LWS and FWS, we used 2 convolutional layers, as this
was found in [4] to be optimal. First, notice that as we increase the
number of hidden units for FWS, there is an improvement in WER,
confirming our belief that having more hidden units with FWS is
important to help explain variations in frequency in the input signal.
Second, we find that if we use LWS but match the number of parameters to FWS, we get very slight improvements in WER (0.1%). It
seems that both LWS and FWS offer similar performance. Because
FWS is simpler to implement, as we do not have to choose filter locations for each limited weight ahead of time, we prefer to use FWS.
Because FWS with 5.6M parameters (256/256 hidden units per convolution layer) gives the best tradeoff between WER and number of
parameters, we use this setting for subsequent experiments.
3.5. Pooling Experiments
Pooling is an important concept in CNNs which helps to reduce
spectral variance in the input features. Similar to [3], we explore
Method
FWS
FWS
FWS
FWS
LWS
LWS
Hidden Units in Conv Layers
128/256
256/256
384/384
512/512
128/256
256/256
Params
5.1M
5.6M
7.6M
10.0M
5.4M
6.6M
WER
19.3
18.9
18.7
18.5
18.8
18.7
Table 4. Limited vs. Full Weight Sharing
pooling in frequency only and not time, as this was shown to be optimal for speech. Because pooling can be dependent on the input
sampling rate and speaking style, we compare the best pooling size
for two different 50 hr tasks with different characteristics, namely
8kHZ speech - Switchboard Telephone Conversations (SWB) and
16kHz speech, English Broadcast News (BN). Table 5 indicates that
not only is pooling essential for CNNs, for all tasks pooling=3 is the
optimal pooling size. Note that we did not run the experiment with
no pooling for BN, as it was already shown to not help for SWB.
No pooling
pool=2
pool=3
pool=4
WER-SWB
23.7
23.4
22.9
22.9
WER- BN
20.7
20.7
21.4
Table 5. WER vs. pooling
3.5.1. Type of Pooling
Pooling is an important concept in CNNs which helps to reduce
spectral variance in the input features. The work in [4] explored
using max pooling as the pooling strategy. Given a pooling region
Rj and a set of activations {a1 , . . . a|Rj | } ∈ Rj , the operation for
max-pooling is shown in Equation 1.
sj = max ai
i∈Rj
(1)
One of the problems with max-pooling is that it can overfit the
training data, and does not necessarily generalize to test data. Two
pooling alternatives have been proposed to address some of the problems with max-pooling, lp pooling [6] and stochastic pooling [7].
lp pooling looks to take a weighted average of activations ai in
pooling region Rj , as shown in Equation 2.
1
p
X p
sj =
ai
(2)
i∈Rj
p = 1 can be seen as a simple form of averaging while p = ∞ corresponds to max-pooling. One of the problems with average pooling
is that all elements in the pooling region are considered, so areas of
low-activations may downweight areas of high activation. lp pooling
for p > 1 is seen as a tradeoff between average and max-pooling. lp
pooling has shown to give large improvements in error rate in computer vision tasks compared to max pooling [6].
Stochastic pooling is another pooling strategy that addresses the
issues of max and average pooling. In stochastic pooling, first a set
of probabilities p for each region j is formed by normalizing the
activations across that region, as shown in Equation 3.
ai
pi = P
(3)
k∈Rj ak
sj = al where l ∼ P (p1 , p2 , . . . p|Rj | )
(4)
A multinomial distribution is created from the probabilities and
the distribution is sampled based on p to pick the location l and
corresponding pooled activation al . This is shown by Equation 4.
Stochastic pooling has the advantages of max-pooling but prevents
overfitting due to the stochastic component. Stochastic pooling has
also shown huge improvements in error rate in computer vision [7].
Given the success of lp and stochastic pooling, we compare both
of these strategies to max-pooling on an LVCSR task. Results for
the three pooling strategies are shown in Table 6. Stochastic pooling
seems to provide improvements over max and lp pooling, though the
gains are slight. Unlike vision tasks, in appears that in tasks such as
speech recognition which have a lot more data and thus better model
estimates, generalization methods such as lp and stochastic pooling
do not offer great improvements over max pooling.
Method
Max Pooling
Stochastic Pooling
lp pooing
WER
18.9
18.8
18.9
Table 6. Results with Different Pooling Types
3.5.2. Overlapping Pooling
The work presented in [4] did not explore overlapping pooling in
frequency. However, work in computer vision has shown that overlapping pooling can improve error rate by 0.3-0.5% compared to
non-overlapping pooling [8]. One of the motivations of overlapping
pooling is to prevent overfitting.
Table 7 compares overlapping and non-overlapping pooling on
an LVCSR speech task. One thing to point out is that because overlapping pooling has many more activations, in order to keep the experiment fair, the number of parameters between non-overlapping
and overlapping pooling was matched. The table shows that there
is no difference in WER between overlapping or non-overlapping
pooling. Again, on tasks with a lot of data such as speech, regularization mechanisms such as overlapping pooling, do not seem to
help compared to smaller computer vision tasks.
Method
Pooling No Overlap
Pooling with Overlap
WER
18.9
18.9
Table 7. Pooling With and Without Overlap
3.5.3. Pooling in Time
Most previous CNN work in speech explored pooling in frequency
only ([4], [3], [5]), though [13] did investigate CNNs with pooling
in time, but not frequency. However, most CNN work in vision performs pooling in both space and time [6], [8]. In this paper, we do a
deeper analysis of pooling in time for speech. One thing we must ensure with pooling in time in speech is that there is overlap between
the pooling windows. Otherwise, pooling in time without overlap
can be seen as subsampling the signal in time, which degrades performance. Pooling in time with overlap can thought of as a way to
smooth out the signal in time, another form of regularization.
Table 8 compares pooling in time for both max, stochastic and
lp pooling. We see that pooling in time helps slightly with stochastic and lp pooling. However, the gains are not large, and are likely
to be diminished after sequence training. It appears that for large
tasks with more data, regularizations such as pooling in time are not
helpful, similar to other regularization schemes such as lp /stochastic
pooling and pooling with overlap in frequency.
Method
Baseline
Pooling in Time, Max
Pooling in Time, Stochastic
Pooling in Time, lp
WER
18.9
18.9
18.8
18.8
Table 8. Pooling in Time
3.6. Incorporating Speaker-Adaptation into CNNs
In this section, we describe various techniques to incorporate speaker
adapted features into CNNs.
3.6.1. fMLLR Features
Since CNNs model correlation in time and frequency, they require
the input feature space to have this property. This implies that commonly used feature spaces, such as Linear Discriminant Analysis,
cannot be used with CNNs. In [4], it was shown that a good feature
set for CNNs was VTLN-warped log-mel filter bank coefficients.
Feature-space maximum likelihood linear regression (fMLLR)
[10] is a popular speaker-adaptation technique used to reduce variability of speech due to different speakers. The fMLLR transformation applied to features assumes that either features are uncorrelated
and can be modeled by diagonal covariance Gaussians, or features
are correlated and can be modeled by a full covariance Gaussians.
While correlated features are better modeled by full-covariance
Gaussians, full-covariance matrices dramatically increase the number of parameters per Gaussian component, oftentimes leading to parameter estimates which are not robust. Thus fMLLR is most commonly applied to a decorrelated space. When fMLLR was applied
to the correlated log-mel feature space with a diagonal covariance
assumption, little improvement in WER was observed [4].
Semi-tied covariance matrices (STCs) [14] have been used to
decorrelate the feature space so that it can be modeled by diagonal
Gaussians. STC offers the added benefit in that it allows a few full
covariance matrices to be shared over many distributions, while each
distribution has its own diagonal covariance matrix.
In this paper, we explore applying fMLLR to correlated features
(such as log-mel) by first decorrelating them such that we can appropriately use a diagonal Gaussian approximation with fMLLR. We
then transform the fMLLR features back to the correlated space so
that they can be used with CNNs.
The algorithm to do this is described as follows. First, starting
from correlated feature space f , we estimate an STC matrix S to
map the features into an uncorrelated space. This mapping is given
by transformation 5
Sf
(5)
Next, in the uncorrelated space, an fMLLR M matrix is estimated, and is applied to the STC transformed features. This is shown
by transformation 6
MSf
(6)
Thus far, transformations 5 and 6 demonstrate standard transformations in speech with STC and fMLLR matrices. However, in
speech recognition tasks, once features are decorrelated with STC,
further transformation (i.e. fMLLR, fBMMI) are applied in this
decorrelated space, as shown in transformation 6. The features are
never transformed back into the correlated space.
However for CNNs, using correlated features is critical. By multiplying the fMLLR transformed features by an inverse STC matrix,
we can map the decorrelated fMLLR features back to the correlated
space, so that they can be used with a CNN. The transformation we
propose is given in transformation 7
S−1 MSf
(7)
3.7. Multi-scale CNN/DNNs
The information captured in each layer of a neural network varies
from more general to more specific concepts. For example, in speech
lower layers focus more on speaker adaptation and higher layers focus more on discrimination. In this section, we look to combine
inputs from different layers of a neural network to explore if complementarity between different layers could potentially improve results further. This idea, known as multi-scale neural networks [6]
has been explored before for computer vision.
Specifically, we look at combining the output from 2 fullyconnected and 2 convolutional layers. This output is fed into 4 more
fully-connected layers, and the entire network is trained jointly. This
can be thought of as combining features generated from a DNN-style
and CNN-style network. Note for this experiment, the same input
feature, (i.e., log-mel features) were used for both DNN and CNN
streams. Results are shown in Table 5. A small gain is observed
by combining DNN and CNN features, again much smaller than
gains observed in computer vision. However, given that a small
improvement comes at the cost of such a large parameter increase,
and the same gains can be achieved by increasing feature maps in the
CNN alone (see Table 4), we do not see huge value in this idea. It
is possible however, that combining CNNs and DNNs with different
types of input features which are complimentary, could potentially
show more improvements.
3.8. I-vectors
order Hessian-free (HF) optimization method is critical for performance gains with sequence training compared to SGD-style optimization, though not as important for CE-training [11].
Rectified Linear Units (ReLU) and Dropout [16] have recently
been proposed as a way to regularize large neural networks. In fact,
ReLU+dropout was shown to provide a 5% relative reduction in
WER for cross-entropy-trained DNNs on a 50-hr English Broadcast
News LVCSR task [12]. However, subsequent HF sequence training
[11] that used no dropout erased some of these gains, and performance was similar to a DNN trained with a sigmoid non-linearity
and no dropout. Given the importance of sequence-training for neural networks, in this paper, we propose a strategy to make dropout
effective during HF sequence training. Results are presented in the
context of CNNs, though this algorithm can also be used with DNNs.
3.9.1. Hessian-Free Training
One popular 2nd order technique for DNNs is Hessian-free (HF) optimization [17]. Let θ denote the network parameters, L(θ) denote a
loss function, ∇L(θ) denote the gradient of the loss with respect to
the parameters, d denote a search direction, and B(θ) denote a Hessian approximation matrix characterizing the curvature of the loss
around θ. The central idea in HF optimization is to iteratively form
a quadratic approximation to the loss and to minimize this approximation using conjugate gradient (CG).
L(θ + d) ≈ L(θ) + ∇L(θ)T d +
1 T
d B(θ)d
2
(8)
During each iteration of the HF algorithm, first, the gradient is
computed using all training examples. Second, since the Hessian
cannot be computed exactly, the curvature matrix B is approximated
by a damped version of the Gauss-Netwon matrix G(θ) + λI, where
λ is set via Levenberg-Marquardt. Then, Conjugate gradient (CG)
is run for multiple-iterations until the relative per-iteration progress
made in minimizing the CG objective function falls below a certain
tolerance. During each CG iteration, Gauss-Newton matrix-vector
products are computed over a sample of the training data.
3.8.1. Results
3.9.2. Dropout
Results with the proposed fMLLR idea are shown in Table 9. Notice
that by applying fMLLR in a decorrelated space, we can achieve a
0.5% improvement over the baseline VTLN-warped log-mel system.
This gain was not possible in [4] when fMLLR was applied directly
to correlated log-mel features.
Dropout is a popular technique to prevent over-fitting during neural network training [16]. Specifically, during the feed-forward operation in neural network training, dropout omits each hidden unit
randomly with probability p. This prevents complex co-adaptations
between hidden units, forcing hidden units to not depend on other
units. Specifically, using dropout the activation yl at layer l is given
by Equation 9, where yl−1 is the input into layer l, Wl is the weight
for layer l, b is the bias, f is the non-linear activation function (i.e.
ReLU) and r is a binary mask, where each entry is drawn from a
Bernoulli(p) distribution with probability p of being 1. Since dropout
1
is not used during decoding, the factor 1−p
used during training ensures that at test time, when no units are dropped out, the correct
total input will reach each layer.
1
Wl (rl−1 ∗ yl−1 ) + bl
(9)
yl = f
1−p
Feature
VTLN-warped log-mel+d+dd
proposed fMLLR + VTLN-warped log-mel+d+dd
WER
18.8
18.3
Table 9. WER With Improved fMLLR Features
3.9. Rectified Linear Units and Dropout
At IBM, two stages of Neural Network training are performed. First,
DNNs are trained with a frame-discriminative stochastic gradient descent (SGD) cross-entropy (CE) criterion. Second, CE-trained DNN
weights are re-adjusted using a sequence-level objective function
[15]. Since speech is a sequence-level task, this objective is more
appropriate for the speech recognition problem. Numerous studies
have shown that sequence training provides an additional 10-15%
relative improvement over a CE trained DNN [11], [4]. Using a 2nd
3.9.3. Combining HF + Dropout
Conjugate gradient tries to minimize the quadratic objective function given in Equation 8. For each CG iteration, the damped GaussNetwon matrix, G(θ), is estimated using a subset of the training
data. This subset is fixed for all iterations of CG. This is because if
the data used to estimate G(θ) changes, we are no longer guaranteed
to have conjugate search directions from iteration to iteration.
Recall that dropout produces a random binary mask for each presentation of each training instance. However, in order to guarantee
good conjugate search directions, for a given utterance, the dropout
mask per layer cannot change during CG. The appropriate way to
incorporate dropout into HF is to allow the dropout mask to change
for different layers and different utterances, but to fix it for all CG
iterations while working with a specific layer and specific utterance
(although the masks can be refreshed between HF iterations).
As the number of network parameters is large, saving out the
dropout mask per utterance and layer is infeasible. Therefore, we
randomly choose a seed for each utterance and layer and save this
out. Using a randomize function with the same seed guarantees that
the same dropout mask is used per layer/per utterance.
3.9.4. Results
We experimentally confirm that using a dropout probability of p =
0.5 in the 3rd and 4th layers is reasonable, and the dropout in all
other layers is zero. For these experiments, we use 2K hidden units
for the fully connected layers, as this was found to be more beneficial
with dropout compared to 1K hidden units [12].
Results with different dropout techniques are shown in Table 10.
Notice that if no dropout is used, the WER is the same as sigmoid, a
result which was also found for DNNs in [12]. By using dropout but
fixing the dropout mask per utterance across all CG iterations, we
can achieve a 0.6% improvement in WER. Finally, if we compare
this to varying the dropout mask per CG training iteration, the WER
increases. Further investigation in Figure 1 shows that if we vary the
dropout mask, there is slow convergence of the loss during training,
particularly when the number of CG iterations increases during the
later part of HF training. This shows experimental evidence that if
the dropout mask is not fixed, we cannot guarantee that CG iterations
produce conjugate search directions for the loss function.
Non-Linearity
Sigmoid
ReLU, No Dropout
ReLU, Dropout Fixed for CG Iterations
ReLU, Dropout Per CG Iteration
WER
15.7
15.6
15.0
15.3
Table 10. WER of HF Sequence Training + Dropout
0.16
training is that it is more closely linked to the speech recognition
objective function compared to cross-entropy. Using this fact, we
explore how many iterations of CE are actually necessary before
moving to HF training. Table 11 shows the WER for different CE
iterations, and the corresponding WER after HF training. Note that
HF training is started and lattices are dumped using the CE weight
that is stopped at. Notice that just by annealing two times, we can
achieve the same WER after HF training, compared to having the
CE weights converge. This points to the fact that spending too much
time in CE is unnecessary. Once the weights are in a relatively decent space, it is better to just jump to HF sequence training which is
more closely matched to the speech objective function.
CE Iter
4
6
8
13
# Times Annealed
1
2
3
7
CE WER
20.8
19.8
19.4
18.8
HF WER
15.3
15.0
15.0
15.0
Table 11. HF Seq. Training WER Per CE Iteration
4. RESULTS
In this section, we analyze CNN performance with the additions proposed in Section 3, namely fMLLR and ReLU + dropout. Results are
shown on both a 50 and 400 hr English Broadcast News task.
4.1. 50-hour English Broadcast News
4.1.1. Experimental Setup
Following the setup in [4], the hybrid DNN is trained using speakeradapted, VTLN+fMLLR features as input, with a context of 9
frames. A 5-layer DNN with 1,024 hidden units per layer and a
sixth softmax layer with 2,220 output targets is used. All DNNs are
pre-trained, followed by CE training and then HF sequence-training
[11]. The DNN-based feature system is also trained with the same
architecture, but uses 512 output targets. A PCA is applied on top of
the DNN before softmax to reduce the dimensionality from 512 to
40. Using these DNN-based features, we apply maximum-likelihood
GMM training, followed by feature and model-space discriminative
training using the BMMI criterion. In order to fairly compare results
to the DNN hybrid system, no MLLR is applied to the DNN featurebased system. The old CNN systems are trained with VTLN-warped
log-mel+d+dd features, and a sigmoid non-linearity. The proposed
CNN-based systems are trained with the fMLLR features described
in Section 3.6, and ReLU+Dropout discussed in Section 3.9.
Dropout Fixed Per CG
Dropout Varied Per CG
0.15
4.1.2. Results
Loss
0.14
0.13
0.12
0.11
0.1
0
5
10
15
20
25
HF Iteration
Fig. 1. Held-out Loss With Dropout Techniques
Finally, we explore if we can reduce the number of CE iterations
before moving to sequence training. A main advantage of sequence
Table 12 shows the performance of proposed CNN-based feature and
hybrid systems, and compares this to DNN and old CNN systems.
The proposed CNN hybrid system offers between a 6-7% relative
improvement over the DNN hybrid, and a 2-3% relative improvement over the old CNN hybrid system. While the proposed CNNbased feature system offers a modest 1% improvement over the old
CNN-based feature system, this slight improvements with featurebased system is not surprising all. We have observed huge relative
improvements in WER (10-12%) on a hybrid sequence trained DNN
with 512 output targets, compared to a hybrid CE-trained DNN.
However, after features are extracted from both systems, the gains
diminish down to 1-2% relative [18]. Feature-based systems use the
neural network to learn a feature transformation, and seem to saturate in performance even when the hybrid system used to extract the
features improves. Thus, as the table shows, there is more potential
to improve a hybrid system as opposed to a feature-based system.
model
Hybrid DNN
Old Hybrid CNN [4]
Proposed Hybrid CNN
DNN-based Features
Old CNN-based Features [4]
Proposed CNN-based Features
dev04f
16.3
15.8
15.4
17.4
15.5
15.3
rt04
15.8
15.0
14.7
16.6
15.2
15.1
Table 12. WER on Broadcast News, 50 hours
4.2. 400 hr English Broadcast News
[1] G. Hinton, L. Deng, D. Yu, G. Dahl, A. Mohamed, N. Jaitly, A. Senior,
V. Vanhoucke, P. Nguyen, T. N. Sainath, and B. Kingsbury, “Deep
Neural Networks for Acoustic Modeling in Speech Recognition,” IEEE
Signal Processing Magazine, vol. 29, no. 6, pp. 82–97, 2012.
[2] Y. LeCun and Y. Bengio, “Convolutional Networks for Images, Speech,
and Time-series,” in The Handbook of Brain Theory and Neural Networks. MIT Press, 1995.
[3] O. Abdel-Hamid, A. Mohamed, H. Jiang, and G. Penn, “Applying
Convolutional Neural Network Concepts to Hybrid NN-HMM Model
for Speech Recognition,” in Proc. ICASSP, 2012.
[4] T.N. Sainath, A. Mohamed, B. Kingsbury, and B. Ramabhadran, “Deep
Convolutional Neural Networks for LVCSR,” in Proc. ICASSP, 2013.
[5] L. Deng, O. Abdel-Hamid, and D. Yu, “A Deep Convolutional Neural
Network using Heterogeneous Pooling for Trading Acoustic Invariance
with Phonetic Confusion,” in Proc. ICASSP, 2013.
[6] P. Sermanet, S. Chintala, and Y. LeCun, “Convolutional neural networks applied to house numbers digit classification,” in Pattern Recognition (ICPR), 2012 21st International Conference on, 2012.
4.2.1. Experimental Setup
We explore scalability of the proposed techniques on 400 hours of
English Broadcast News [15]. Development is done on the DARPA
EARS dev04f set. Testing is done on the DARPA EARS rt04
evaluation set. The DNN hybrid system uses fMLLR features, with
a 9-frame context, and use five hidden layers each containing 1,024
sigmoidal units. The DNN-based feature system is trained with 512
output targets, while the hybrid system has 5,999 output targets. Results are reported after HF sequence training. Again, the proposed
CNN-based systems are trained with the fMLLR features described
in Section 3.6, and ReLU+Dropout discussed in Section 3.9.
4.2.2. Results
Table 13 shows the performance of the proposed CNN system compared to DNNs and the old CNN system. While the proposed 512hybrid CNN-based feature system did improve (14.1 WER) over the
old CNN (14.8 WER), performance slightly deteriorates after CNNbased features are extracted from the network. However, the 5,999hybrid CNN offers between a 13-16% relative improvement over
the DNN hybrid system, and between a 4-5% relative improvement
over the old CNN-based features systems. This helps to strengthen
the hypothesis that hybrid CNNs have more potential for improvement, and the proposed fMLLR and ReLU+dropout techniques provide substantial improvements over DNNs and CNNs with a sigmoid
non-linearity and VTLN-warped log-mel features.
model
Hybrid DNN
DNN-based Features
Old CNN-based Features [4]
Proposed CNN-based Features
Proposed Hybrid CNN
6. REFERENCES
dev04f
15.1
15.3
13.4
13.6
12.7
rt04
13.4
13.5
12.2
12.5
11.7
Table 13. WER on Broadcast News, 400 hrs
5. CONCLUSIONS
In this paper, we explored various strategies to improve CNN performance. We incorporated fMLLR into CNN features, and also made
dropout effective after HF sequence training. We also explored various pooling and weight sharing techniques popular in computer vision, but found they did not offer improvements for LVCSR tasks.
Overall, with the proposed fMLLR+dropout ideas, we were able to
improve our previous best CNN results by 2-5% relative.
[7] M. Zeiler and R. Fergus, “Stochastic Pooling for Regularization of
Deep Convolutional Neural Networks,” in Proc. of the International
Conference on Representaiton Learning (ICLR), 2013.
[8] A. Krizhevsky, I. Sutskever, and G. Hinton, “Imagenet Classification
with Deep Convolutional Neural Networks,” in Advances in Neural
Information Processing Systems, 2012.
[9] Y. LeCun, F. Huang, and L. Bottou, “Learning Methods for Generic
Object Recognition with Invariance to Pose and Lighting,” in Proc.
CVPR, 2004.
[10] M.J.F. Gales, “Maximum likelihood linear transformations for HMMbased Speech Recognition,” Computer Speech and Language, vol. 12,
no. 2, pp. 75–98, 1998.
[11] B. Kingsbury, T. N. Sainath, and H. Soltau, “Scalable Minimum Bayes
Risk Training of Deep Neural Network Acoustic Models Using Distributed Hessian-free Optimization,” in Proc. Interspeech, 2012.
[12] G.E. Dahl, T.N. Sainath, and G.E. Hinton, “Improving Deep Neural
Networks for LVCSR Using Rectified Linear Units and Dropout,” in
Proc. ICASSP, 2013.
[13] A. Waibel, T. Hanazawa, G. Hinton, K. Shikano, and K.J Lang,
“Phoneme Recognition using Time-delay Neural Networks,” IEEE
Transactions on Acoustics, Speech and Signal Processing, vol. 37, no.
3, pp. 328–339, 1989.
[14] M.J.F. Gales, “Semi-tied Covariance Matrices for Hidden Markov
Models,” IEEE Transactions on Speech and Audio Processing, vol.
7, pp. 272–281, 1999.
[15] B. Kingsbury, “Lattice-Based Optimization of Sequence Classification
Criteria for Neural-Network Acoustic Modeling,” in Proc. ICASSP,
2009.
[16] G.E. Hinton, N. Srivastava, A. Krizhevsky, I. Sutskever, and
R. Salakhutdinov, “Improving Neural Networks by Preventing CoAdaptation of Feature Detectors,” The Computing Research Repository
(CoRR), vol. 1207.0580, 2012.
[17] J. Martens, “Deep learning via Hessian-free optimization,” in Proc.
Intl. Conf. on Machine Learning (ICML), 2010.
[18] T. N. Sainath, B. Kingsbury, and B. Ramabhadran, “Auto-Encoder
Bottleneck Features Using Deep Belief Networks,” in Proc. ICASSP,
2012.
| 9cs.NE
|
Model Identification via Physics Engines for Improved Policy Search
arXiv:1710.08893v1 [cs.RO] 24 Oct 2017
Shaojun Zhu, Andrew Kimmel, Kostas E. Bekris and Abdeslam Boularias
Abstract— This paper presents a practical approach for
identifying unknown mechanical parameters, such as mass
and friction models of manipulated rigid objects or actuated
robotic links, in a succinct manner that aims to improve the
performance of policy search algorithms. Key features of this
approach are the use of off-the-shelf physics engines and the
adaptation of a black-box Bayesian optimization framework
for this purpose. The physics engine is used to reproduce in
simulation experiments that are performed on a real robot,
and the mechanical parameters of the simulated system are
automatically fine-tuned so that the simulated trajectories
match with the real ones. The optimized model is then used for
learning a policy in simulation, before safely deploying it on the
real robot. Given the well-known limitations of physics engines
in modeling real-world objects, it is generally not possible to
find a mechanical model that reproduces in simulation the real
trajectories exactly. Moreover, there are many scenarios where
a near-optimal policy can be found without having a perfect
knowledge of the system. Therefore, searching for a perfect
model may not be worth the computational effort in practice.
The proposed approach aims then to identify a model that
is good enough to approximate the value of a locally optimal
policy with a certain confidence, instead of spending all the
computational resources on searching for the most accurate
model. Empirical evaluations, performed in simulation and on
a real robotic manipulation task, show that model identification
via physics engines can significantly boost the performance of
policy search algorithms that are popular in robotics, such as
TRPO, PoWER and PILCO, with no additional real-world data.
I. INTRODUCTION
This paper presents an approach for model identification
by exploiting the availability of off-the-shelf physics engines
that are used for simulating dynamics of robots and objects
they interact with. There are many examples of popular
physics engines that are becoming increasingly efficient [1]–
[6]. Physics engines take as inputs mechanical and mesh
models of objects in a particular environment, in addition to
forces and torques applied to them at different time-steps,
and return predictions of how the objects would move.
The accuracy of the predicted motions depends on several
factors. The first one is the limitation of the mathematical
model used by the engine (e.g., Coulomb’s law of friction).
The second factor is the accuracy of the numerical algorithm
used for solving the differential equations of motion. Finally,
the prediction depends on the accuracy of the mechanical
parameters of the robot and the objects’ models, such as
mass, friction and elasticity. In this work, we focus on this
The
puter
authors
Science,
are
with
the
Department
of
Rutgers
University,
New
Jersey,
{shaojun.zhu,ask139,kostas.bekris,ab1544}
@cs.rutgers.edu
ComUSA.
The Baxter robot needs to pick up the bottle but it cannot
reach it, while the Motoman robot can. The Motoman gently pushes
the object locally without risking to lose it. From observed motions,
mechanical properties of the object are identified via a physics
engine. The object is then pushed into Baxter’s workspace using a
policy learned in simulation with the identified property parameters.
Fig. 1.
last factor and propose a method for improving the accuracy
of mechanical parameters used in physical simulations.
For motivation, consider the setup illustrated in Figure 1,
where a static robot (Motoman) assists another one (Baxter)
to reach and pick up a desired object (a bottle). The object
is known and both robots have the capability to pick it up.
However, the object can be reached only by Motoman and not
by Baxter. And due to the considerable distance between the
two static robots, the intersection of their reachable workspace
is empty, which restricts the execution of a direct hand-off.
In this case, the Motoman robot must learn an action, such as
rolling or sliding, that would move the bottle to a distant target
zone. If the robot simply executes the maximum velocity push
on the object, the result causes the object to fall off the table.
Similarly, if the object is rolled too slowly, it could end up
stuck in the region between the two robot’s workspaces and
neither of them could reach it. Both outcomes are undesirable
as they would ruin the autonomy of the system and require a
human intervention to reset the scene or perform the action.
This example highlights the need for identifying an object’s
mechanical model to predict where the object would end up
on the table given different initial velocities. An optimal
velocity could be derived accordingly using simulations.
The technique presented in this paper aims at improving the
accuracy of the mechanical parameters used in physics engines
in order to perform a given robotic task. Given recorded
real trajectories of the object in question, we search for the
best model parameters so that the simulated trajectories are
as close as possible to the observed real trajectories. This
search is performed through an anytime black-box Bayesian
optimization, where a probability distribution (belief) on
the optimal model is repeatedly updated. When the time
consumed by the optimization exceeds a certain preallocated
time budget, the optimization is halted and the model with
the highest probability is returned. A policy search subroutine
takes over the returned model and finds a policy that aims
to perform the task. The policy search subroutine could be a
control method, such as LQR, or a reinforcement learning
(RL) algorithm that runs on the physics engine with the
identified model instead of the real world for the sack of
data-efficiency and also for safety. The obtained policy is
then deployed on the robot, run in the real world and the
new observed trajectories are handed back again to the model
identification module, to repeat the same process.
The question that arises here is: how accurate should the
identified model be in order to find the optimal policy? Instead
of spending a significant amount of time searching for the
most accurate model, it would be useful to stop the search
whenever a model that is sufficiently accurate for the task at
hand is found. Answering this question exactly is difficult,
because that would require knowing in advance the optimal
policy for each model, in which case the model identification
process can be stopped simply when there is a consensus
among the most likely models on which policy is optimal.
Our solution to this problem is motivated by a key quality
that is desired in robot RL algorithms. To ensure safety,
most robot RL algorithms constrain the changes in the
policy between two iterations to be minimal and gradual.
For instance, both Relative-Entropy Policy Search (REPS)
and Trust Region Policy Optimization (TRPO) algorithms
guarantee that the KL distance between an updated policy and
a previous one in a learning loop is bounded by a predefined
constant. Therefore, one can in practice use the previous best
policy as a proxy to verify if there is a consensus among the
most likely models on the best policy in the next iteration
of the policy search. This is justified by the fact that the
new policies are not too different from the previous one in
a policy search. The model identification process is stopped
whenever the most likely models predict almost the same
value for the previous policy. In other terms, if all models that
have reached a high probability in the anytime optimization
predict the same value for the previous policy, then any of
these models could be used for searching for the next policy.
While the current paper does not provide theoretical
guarantees of the proposed method, our empirical evaluations
show that it can indeed improve the performance of several
RL algorithms. The first part of the experiments is performed
on systems in the MuJoCo simulator [3]. The second part is
performed on the robotic task shown in Figure 1.
II. R ELATED W ORK
Two high-level approaches exist for learning to perform
tasks in systems with unknown parameters: model-free and
model-based ones. Model-free methods search for a policy
that best solves the task without explicitly learning the
system’s dynamics [7]–[9]. Model-free methods are accredited
with the recent success stories of RL, in video games for
example [10]. In robot learning, the REPS algorithm was
used to successfully train a robot to play table tennis [11].
The PoWER algorithm [12] is another model-free policy
search approach widely used for learning motor skills. The
Trust Region Policy Optimization (TRPO) algorithm [13] is
arguably the state-of-the-art model-free RL technique. Policy
search can also be achieved through Bayesian optimization
and has been used for gait optimization [14], where Central
Pattern Generators are a popular policy parameterization [15].
Model-free methods, however, tend to require a lot of
training data and can also jeopardize the safety of a robot.
Model-based approaches are alternatives that explicitly learn
the unknown parameters of the system, and search for
an optimal policy accordingly. There are many examples
of model-based approaches for robotic manipulation [16]–
[20], some of which have used physics-based simulation
to predict the effects of pushing flat objects on a smooth
surface [16]. A nonparametric approach was employed for
learning the outcome of pushing large objects [18]. For
general-purpose model-based RL, the PILCO algorithm has
been proven efficient in utilizing a small amount of data to
learn dynamical models and optimal policies [21]. Several
cognitive models that combine such Bayesian inference with
approximate knowledge of Newtonian physics have been
proposed recently [22]–[24]. A common characteristic of
many model-based methods is the fact that they learn a
transition function using a purely statistical approach without
taking advantage of the known equations of motion of
rigid-objects. NARMAX [25] is an example of popular
general-purpose model identification techniques that are not
specifically designed for rigid-body dynamics. In contrast
to these methods, we use a physics engine, and concentrate
on identifying only the mechanical properties of the objects
instead of learning laws of motion from scratch.
There is also work on identifying sliding models of objects
using white-box optimization [17], [20], [26]. It is not clear,
however, how these methods would perform, since they are
tailored to specific tasks, such as pushing planar objects [20],
unlike the proposed general approach.
An increasingly popular alternative addresses these challenges through end-to-end learning [27]–[38]. This involves
the demonstration of successful examples of physical interaction and learning a direct mapping of the sensing input to
controls. While a desirable result, these approaches usually
require many physical experiments to effectively learn.
Some recent works also proposed to use physics engines
in combination with real-world experiments to boost policy
search algorithms [39], [40], although these methods do not
explicitly identify mechanical models of objects.
A key contribution of the current work is linking the modelidentification process to the policy search process. Instead
of searching for the most accurate model, we search for a
model that is accurate enough to predict the value function
of a policy that is not too different from the searched policy.
Therefore, the proposed approach can be used in combination
with any policy search algorithm that guarantees smooth
changes in the learned policy.
Physical interaction
force µt
xt
xt+1
t
xt+1
simulation
error ek
position
x̂kt+1
x̂kt+1
= f (xt , µt , ✓k )
x
0.065
0 model distribution
simulate with model ✓ 2
simulation
error e2
x̂2t+1
xt+1
✓
0 model distribution
position x̂2t+1 = f (xt , µt , ✓2 )
✓
simulate with model ✓ 1
simulation
error e1
x̂1t+1
Simulation
✓
0.13
error(✓)
0.13
x
P r(✓)
x
x
simulate with model ✓ k
error(✓)
P r(✓)
x
x
x
x
0 model distribution
t+1
0.26
x
x
x
e1
n
ia or
ss err
au d
r G ve
ro er
er obs
e
th ith
te w
da ss
up oce
Pr
t
0.2
e2
n
ia or
ss err
au d
r G ve
ro er
er obs
e
th ith
te w
da ss
up oce
Pr
use the final model distribution
to find force µt in new state xt
0.4
actual observed pose at time t + 1
P r(✓)
error(✓)
pose at time
xt+1
position x̂1t+1 = f (xt , µt , ✓1 )
Learning a mechanical model of an object (bottle) through physical simulations. The key idea is to search for a model that closes
the gap between simulation with a physics engine and reality, by using an anytime Bayesian optimization. The search stops when the
models with the highest probabilities predict similar values for a given policy. This process is repeated in this figure after each time-step t,
i.e. after every real-world action. In practice, it is more efficient to do the model identification only after a certain number of time-steps.
Fig. 2.
III. P ROPOSED A PPROACH
We start with an overview of the model identification and
policy search system. We then present the main algorithm
and explain the model identification part in more details.
A. System Overview and Notations
Figure 2 shows an overview of the proposed approach.
The example is focused on the non-prehensile manipulation
application but the same approach is used to identify physical
properties of actuated robotic links. In object manipulation
problems (Figure 1), the targeted mechanical properties
correspond to the object’s mass and the static and kinetic
friction coefficients of different regions of the object. The
surface of an object is divided into a regular grid. This allows
to identify the friction parameters of each part of the grid.
These physical properties are all concatenated in a single
vector and represented as a D-dimensional vector θ ∈ Θ,
where Θ is the space of all possible values of the physical
properties. Θ is discretized with a regular grid resolution. The
proposed approach returns a distribution P on discretized Θ
instead of a single point θ ∈ Θ since model identification
is generally an ill-posed problem. In other terms, there are
multiple models that can explain an observed movement of
an object with equal accuracies. The objective is to preserve
all possible explanations and their probabilities.
The online model identification algorithm takes as input a
prior distribution Pt , for time-step t ≥ 0, on the discretized
space of physical properties Θ. Pt is calculated based on
the initial distribution P0 and a sequence of observations
(x0 , µ0 , x1 , µ1 , . . . , xt−1 , µt−1 , xt ). For instance, in the case
of object manipulation, xt is the 6D pose (position and
orientation) of the manipulated object at time t and µt is
a vector describing a force applied by the robot’s fingertip on
the object at time t. Applying a force µt results in changing
the object’s pose from xt to xt+1 . The algorithm returns a
distribution Pt+1 on the models Θ.
The robot’s task is specified by a reward function R that
maps state-actions (x, µ) into real numbers. A policy π returns
an action µ = π(x) for state x. The value V π (θ ) of policy π
H
given model θ is defined as V π (θ ) = ∑t=0
R(xt , µt ), where
H is a fixed horizon, x0 is a given starting state, and xt+1 =
f (xt , µt , θ ) is the predicted state at time t + 1 after simulating
force µt in state xt using physical parameters θ . For simplicity,
we focus here only on systems with deterministic dynamics.
B. Main Algorithm
Given a reward function R and a simulator with model
parameters θ , there are many techniques that can be used for
searching for a policy π that maximizes value V π (θ ). For
example, one can use Differential Dynamic Programming
(DDP), Monte Carlo (MC) methods, or simply run a modelfree RL algorithm on the simulator if the system is highly
nonlinear and a good policy cannot be found with former
methods. The choice of a particular policy search method is
open and depends on the task.
The main loop of our system is presented in Algorithm 1.
This meta-algorithm consists in repeating three main steps: (1)
data collection using the real robot, (2) model identification
using a simulator, and (3) policy search in simulation using
the best identified model.
C. Value-Guided Model Identification
The process, explained in Algorithm 2, consists of simulating the effects of forces µi on the object in states xi under
t ← 0;
Initialize distribution P over Θ to a uniform distribution;
Initialize policy π;
repeat
Execute policy π for H iterations on the real robot,
and collect new state-action-state data {(xi , µi , xi+1 )}
for i = t, . . . ,t + H − 1;
t ← t + H;
Run Algorithm 2 with collected state-action-state data
and reference policy π for updating distribution P;
Initialize a policy search algorithm (e.g. TRPO) with
π and run the algorithm in the simulator with the
model arg maxθ ∈Θ P(θ ) to find an improved policy
π 0; π ← π 0;
until Timeout;
Algorithm 1: Main Loop
various values of parameters θ and observing the resulting
states x̂i+1 , for i = 0, . . . ,t. The accompanying implementation
is using the Bullet and MuJoCo physics engines for this
purpose [2], [3]. The goal is to identify the model parameters
that make the outcomes of the simulation as close as possible
to the real observed outcomes. In other terms, the following
black-box optimization problem is solved:
de f
θ ∗ = arg min E(θ ) =
θ ∈Θ
t
∑ kxi+1 − f (xi , µi , θ )k2 ,
(1)
i=0
wherein xi and xi+1 are the observed states of the object at
times i and i + 1, µi is the force that moved the object from
xi to xi+1 , and f (xi , µi , θ ) = x̂i+1 , the predicted state at time
t + 1 after simulating force µi in state xi using θ .
High-fidelity simulations are computationally expensive. It
is therefore important to minimize the number of simulations,
i.e., evaluations of function E, while searching for the optimal
parameters that solve Equation 1. We solve this problem by
using the Entropy Search technique [41]. This method is wellsuited for our purpose because it explicitly maintains a belief
on the optimal parameters, unlike other Bayesian optimization
methods such as Expected Improvement that only maintain a
belief on the objective function. In the following, we explain
how this technique is adapted to our purpose, and show why
keeping a distribution on all models is needed for deciding
when to stop the optimization.
The error function E does not have an analytical form, it
is gradually learned from a sequence of simulations with
a small number of parameters θk ∈ Θ. To choose these
parameters efficiently in a way that quickly leads to accurate
parameter estimation, a belief about the actual error function is
maintained. This belief is a probability measure over the space
of all functions E : RD → R, and is represented by a Gaussian
Process (GP) [42] with mean vector m and covariance matrix
K. The mean m and covariance
K of the GP
are learned
from data points { θ0 , E(θ0 ) , . . . , θk , E(θk ) }, where θk is
a selected vector of physical properties of the object, and
E(θk ) is the accumulated distance between actual observed
states and states that are obtained from simulation using θk .
Input: state-action-state data {(xi , µi , xi+1 )} for i = 0, . . . ,t
a discretized space of possible values of physical
properties Θ,
a reference policy π,
minimum and maximum number of evaluated
models kmin , kmax ,
model confidence threshold η,
value error threshold ε ;
Output: probability distribution P over Θ;
Sample θ0 ∼ Uniform(Θ); L ← 0;
/ k ← 0; stop ← f alse;
repeat
/* Calculating the accuracy of model θk
*/
lk ← 0;
for i = 0 to t do
Simulate {(xi , µi )} using a physics engine with
physical parameters θk and get the predicted next
state x̂i+1 = f (xi , µi , θk ) ;
lk ← lk + kx̂i+1 − xi+1 k2 ;
end
L ← L ∪ {(θk , lk )};
Calculate GP(m, K) on error function E, where
E(θ ) = l, using data (θ , l) ∈ L;
/* Monte Carlo sampling
Sample E1 , E2 , . . . , En ∼ GP(m, K) in Θ;
foreach θ ∈ Θ do
1 n
P(θ ) ≈ ∑ 1θ =arg minθ 0 ∈Θ E j (θ 0 )
n j=0
end
*/
(2)
/* Selecting the next model to
evaluate
*/
/* Checking the stopping condition
*/
θk+1 = arg minθ ∈Θ P(θ ) log P(θ ) ;
k ← k + 1;
if k ≥ kmin then
θ ∗ ← arg maxθ ∈Θ P(θ );
Calculate the values V π (θ ) with all models θ that
have a probability P(θ ) ≥ η by using the physics
engine for simulating trajectories with models θ ;
if ∑θ ∈Θ 1P(θ )≥η |V π (θ ) −V π (θ ∗ )| ≤ ε then
stop ← true;
end
end
if k = kmax then
stop ← true;
end
until stop = true;
Algorithm 2: Value-Guided Model Identification
The probability distribution P on the identity of the best
physical model θ ∗ , returned by the algorithm, is computed
from the learned GP as
de f
P(θ ) = P θ = arg min
E(θ 0 )
0
θ ∈Θ
Z
(3)
=
pm,K (E)Πθ 0 ∈Θ−{θ } H E(θ 0 ) − E(θ ) dE
E:RD →R
0
where H is the Heaviside step function, i.e., H E(θ
)−
0
0
E(θ ) = 1 if E(θ ) ≥ E(θ ) and H E(θ ) − E(θ ) = 0
otherwise, and pm,K (E) is the probability of a function E
according to the learned GP mean m and covariance K.
Intuitively, P(θ ) is the expected number of times that θ
happens to be the minimizer of E when E is a function
distributed according to GP density pm,K .
Distribution P from Equation 3 does not have a closedform expression. Therefore, a Monte Carlo (MC) sampling
is employed for estimating P. The process samples vectors
[E(θ 0 )]θ 0 ∈Θ containing values that E could take, according to
the learned Gaussian process, in the discretized space Θ. Then
P(θ ) is estimated by counting the ratio of sampled vectors
of the values of simulation error E where θ happens to make
the lowest error, as indicated in Equation 2 in Algorithm 2.
Finally, the computed distribution P is used to select
the next vector θk+1 to use as a physical model in the
simulator. This process is repeated until the entropy of P
drops below a certain threshold, or until the algorithm runs
out of the allocated time budget.
The entropy of P is given as
∑θ ∈Θ −Pmin (θ ) log Pmin (θ ) . When the entropy of P is close
to zero, the mass of distribution P is concentrated around a
single vector θ , corresponding to the physical model that best
explains the observations. Hence, next θk+1 should be selected
so that the entropy ofP would decrease after adding the data
point θk+1 , E(θk+1 ) } to train the GP and re-estimate P
using the new mean m and covariance K in Equation 3.
Entropy Search methods follow this reasoning and use
MC again to sample, for each potential choice of θk+1 , a
number of values that E(θk+1 ) could take according to the
GP in order to estimate the expected change in the entropy
of P and choose the parameter vector θk+1 that is expected
to decrease the entropy of P the most. The existence of a
secondary nested process of MC sampling makes this method
unpractical for our online optimization. Instead, we present a
simple heuristic for choosing the next θk+1 . In this method,
that we call Greedy Entropy Search, the next θk+1 is chosen
as the point that contributes the most to the entropy of P,
θk+1 = arg max −P(θ ) log P(θ ) .
θ ∈Θ
This selection criterion is greedy because it does not anticipate
how the output of the simulation using θk+1 would affect the
entropy of P. Nevertheless, this criterion selects the point that
is causing the entropy of P to be high. That is a point θk+1
with a good chance P(θk+1 ) of being the realmodel, but
also with a high uncertainty P(θk+1 ) log P(θ1 ) . We found
k+1
out from our first experiments that this heuristic version of
Entropy Search is more practical than the original Entropy
Search method because of the computationally expensive
nested MC sampling loops used in the original method.
The stopping condition of Algorithm 2 depends on the
predicted value of a reference policy π. The reference policy
is one that will be used in the main algorithm (Algorithm 1) as
a starting point in the policy search with the identified model.
That is also the policy executed in the previous round of
the main algorithm. Many policy search algorithms (such as
REPS and TRPO) guarantee that the KL divergence between
consecutive policies π and π 0 is minimal. Therefore, if the
difference |V π (θ ) −V π (θ ∗ )| for two given models θ and θ ∗
0
is smaller than a threshold ε, then the difference |V π (θ ) −
0
V π (θ ∗ )| should also be smaller than a threshold that is a
function of ε and KL(πkπ 0 ). A full proof of this conjecture
is the subject of an upcoming work. In practice, this means
that if θ and θ ∗ are two models with high probabilities, and
|V π (θ ) − V π (θ ∗ )| ≤ ε then there is no point in continuing
the Bayesian optimization to find out which one of the two
models is actually the most accurate because both models will
result in similar policies. The same argument could be used
when there are more than two models with high probabilities.
In some tasks, such as the one in the motivation example
in Figure 1, the policy used for data collection is significantly
different from the policy used for actually performing the
task. The policy used to collect data consists in moving the
object slowly without risking to make it move away from the
reachable workspace of the Motoman. Otherwise, a human
intervention would be needed. The optimal policy, on the
other hand, consists in striking the object with a certain high
velocity. Therefore, the data-collecting policy cannot be used
as a proxy for the optimal policy in Algorithm 2. Instead, we
use the actual optimal policy with respect to the most likely
model, i.e. π = arg maxπi V πi (θ ∗ ). It turns out that finding
the optimal policy for a given model in this specific task can
be performed quickly in simulation by searching in the space
of discretized striking velocities. This is not the case in more
complex systems where searching for an optimal policy is
computationally expensive, which is the reason we use the
previous best policy π as a surrogate for the next best policy
π 0 when checking the stopping condition.
IV. E XPERIMENTAL R ESULTS
The proposed Value-Guided Model Identification (VGMI)
approach is validated both in simulation and on a real robotic
manipulation task, and compared to other RL methods.
A. Experiments on RL Benchmarks in Simulation
Setup: The simulation experiments are done in OpenAI gym
(Figure 3) [43] with the MuJoCo physics simulator [3]. The
space of unknown physical models θ is described below.
Inverted Pendulum: A pendulum is connected to a cart,
which moves linearly. The dimensionality of space Θ is two,
one for the mass of the pendulum and one for the cart.
Swimmer: The swimmer is a 3-link planar robot. Space Θ
has three dimensions, one for the mass of each link.
Hopper: The hopper is a 4-link planar mono-pod robot. Thus,
dimensionality of the parameter space Θ is four.
Walker2D: The walker is a 7-link planar biped robot. Thus,
dimensionality of the parameter space Θ is seven.
For each of the environments, we use the simulator with
default mass as the real system, and increase or decrease the
masses by ten to fifty percent randomly to create inaccurate
simulators to use as prior models. In this section, all the
policies are trained with Trust Region Policy Optimization
(TRPO) [13] implemented in rllab [44]. The policy network
has two hidden layers with 32 neurons each.
Inverted Pendulum
Swimmer
Hopper
Fig. 3.
Walker2D
OpenAI Gym systems used in the experiments
1.6
Entropy Search
Greedy Entropy Search
Trajectory Error (meters)
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0
5
10
15
20
Time (seconds)
Fig. 4. Model identification in Inverted Pendulum environment using two
variants of Entropy Search.
We start by comparing Greedy Entropy Search (GES) with
the original Entropy Search (ES) on the problem identifying
the mass parameters of the Inverted Pendulum system. Rollout
trajectories are collected using optimal policies learned with
the real system. Given inaccurate simulators and the control
sequence from rollouts, we try to identify the mass parameters
which enables the simulator to generate trajectories most
close to the real ones. Figure 4 shows that GES converges
faster than ES. Similar behaviors were observed on the other
systems, but not reported here for space’s sake.
We refer to the main algorithm detailed in Algorithm 1
as TRPO+VGMI in this section. TRPO+VGMI starts with
the inaccurate simulator and VGMI gradually increases
the accuracy of the simulator. We compare TRPO+VGMI
against a) TRPO trained directly with the real system and
b) TRPO trained with inaccurate simulators. Depending on
problem difficulty, we vary the number of iterations for
policy optimization. For TRPO both with the real system and
with the inaccurate simulators, we run Inverted Pendulum
for 100 interactions, Swimmer for 400 iterations, Hopper
for 200 iterations and Walker2D for 1000 iterations. For
TRPO+VGMI, we run VGMI as detailed in Algorithm 2
every 10 iterations, i.e. H = 10 in Algorithm 1. We run
TRPO+VGMI 20 iterations for Inverted Pendulum, 100
iterations for Swimmer, 100 iterations for Hopper and 200
iterations for Walker2D. All the results are the mean and
variance of 20 independent trials, for statistical significance.
Results: We report performance both in terms of the number
of rollouts on the real system and the total training time. The
number of rollouts represents the data efficiency of the policy
search algorithms and corresponds to the actual number of
trajectories in the real system. The total training time is the
total simulation and policy optimization time used for TRPO
to converge. For TRPO+VGMI, it also includes the time spent
on model identification. Figure 5 shows the mean cumulative
reward per rollout (trajectory) on the real systems as functions
of the number of rollouts used for training. For all four tasks,
TRPO+VGMI requires less rollouts. The rollouts are used
by VGMI to identify the optimal mass parameter of the
simulator for policy search, while they are used directly for
policy search by TRPO. The results show that the models
identified by VGMI are accurate enough for TRPO to find a
good policy by using the same amount of data.
Figure 6 shows the cumulative reward per trajectory on the
real system as a function of the total time in seconds. We also
report the performance of TRPO when trained with inaccurate
simulators, which is worse then when it is trained directly
on the real system (the real system here is also a simulator,
but with different physical parameters). This clearly shows
the advantage of model identification from data for policy
search. TRPO+VGMI is slower than TRPO because of all
the extra time spent by TRPO+VGMI on model identification
and policy search in the learned simulator.
In summary, VGMI boosts the data-efficiency of TRPO
by identifying parameters of the objects and using a physics
engine with the identified parameters to search for a policy
before deploying it on the real system. On the other hand,
VGMI adds a computational burden to TRPO.
B. Non-prehensile Manipulation Experiments on a Real Robot
Setup: The task in this experiment is to push the bottle
one meter away from one side of a table to the other, as
shown in Figure 7. The goal is to find an optimal policy
with parameter η representing the pushing velocity of the
robotic hand. The pushing direction is always towards
the target position and the hand pushes the object at its
geometric center. During data collection, no human effort
is needed to reset the scene. The velocity and pushing
direction are controlled such that the object is always in
the workspace of the robotic hand. Specifically, a pushing
velocity limit is set and the pushing direction is always
towards the center of the workspace. The proposed approach
iteratively searches for best pushing velocity by uniformly
sampling 20 different velocities in simulation, and identifies
the object model parameters θ ∗ (the mass and the friction
coefficient) using trajectories from rollouts by running
VGMI as in Algorithm 2. In this experiment, we run
VGMI after each rollout, i.e., H = 1 in Algorithm 1. The
method is compared to two reinforcement learning methods:
PoWER [12] and PILCO [21]. For PoWER, the reward
function is r = e−dist , where dist is the distance between
the object position after pushing and the desired target
Fig. 5. Cumulative reward per trajectory as a function of the number of constant-length trajectories on the real system. Trajectories on a second simulator
with identified models are not counted here, as they do not occur on the real system.
Fig. 6.
Cumulative reward per trajectory as a function of total time in seconds, including search and optimization times.
Fig. 7.
Examples of experiment where the Motoman pushes the object into Baxter’s workspace.
0.6
Figure 8 provides the real robotic experiment with a
Motoman robot. The proposed method achieves both lower
final object location error and fewer number of object drops
comparing to alternatives. The reduction in object drops is
especially important for autonomous robot learning as it
minimizes human effort during learning. The model-free
approach such as PoWER results in higher location error and
more object drops. PILCO performs better than PoWER as it
also learns a dynamical model in addition to the policy, but
the model may not be as accurate as a physics engine with
identified parameters. As only a very simple policy search
method is used for VGMI, the performance is expected to
be better is more advanced policy search methods, such as
combining PoWER with VGMI.
PoWER
PILCO
VGMI
PoWER
PILCO
VGMI
0.5
# of Times Object
Falls Off the Table
Results: Two metrics are used for evaluating the performance:
1) The distance between the final object location after being
pushed and the desired goal location; 2) The number of times
the object falls off the table. A video of these experiments
can be found in the supplementary video or on https:
//goo.gl/dX2wPV.
Location Error (meters)
1.5
position. For PILCO, the state space is the 3D object position.
1
0.5
0.4
0.3
0.2
0.1
0
0
5
10
15
20
25
30
3
6
9
Number of trials
12
15
18
21
24
27
30
Number of trials
Fig. 8. Pushing policy optimization results using a Motoman robot. Our
method VGMI achieves both lower final object location error and fewer
object drops comparing to alternatives. Best viewed in color.
V. C ONCLUSION
This paper presents a practical approach that integrates
a physics engine and Bayesian optimization for model
identification to increase the data efficiency of reinforcement
learning algorithms. The model identification process is taking
place in parallel with the reinforcement learning loop. Instead
of searching for the most accurate model, the objective is
to identify a model that is accurate enough so as to predict
the value function of a policy that is not too different from
the current optimal policy. Therefore, the proposed approach
can be used in combination with any policy search algorithm
that guarantees smooth changes in the learned policy. Both
simulated and real robotic manipulation experiments show that
the proposed technique for model identification can decrease
the number of rollouts needed to learn optimal policy.
Future works include performing an analysis of the properties for the proposed Value-Guided Model Identification
method, such as expressing the conditions under which the
inclusion of the model identification approach reduces the
needs for physical rollouts and the speed-up in convergence
in terms of physical rollouts. It is also interesting to consider
alternative physical tasks, such as locomotion challenges,
which can benefit by the proposed framework.
R EFERENCES
[1] T. Erez, Y. Tassa, and E. Todorov, “Simulation tools for model-based
robotics: Comparison of bullet, havok, mujoco, ODE and physx,” in
IEEE International Conference on Robotics and Automation, ICRA,
2015, pp. 4397–4404.
[2] “Bullet physics engine,” [Online]. Available: www.bulletphysics.org.
[3] “MuJoCo physics engine,” [Online]. Available: www.mujoco.org.
[4] “DART physics egnine,” [Online]. Available: http://dartsim.github.io.
[5] “PhysX physics engine,” [Online]. Available: www.geforce.com/
hardware/technology/physx.
[6] “Havok physics engine,” [Online]. Available: www.havok.com.
[7] R. S. Sutton and A. G. Barto, Introduction to Reinforcement Learning,
1st ed. Cambridge, MA, USA: MIT Press, 1998.
[8] D. P. Bertsekas and J. N. Tsitsiklis, Neuro-Dynamic Programming,
1st ed. Athena Scientific, 1996.
[9] J. Kober, J. A. D. Bagnell, and J. Peters, “Reinforcement learning in
robotics: A survey,” International Journal of Robotics Research, July
2013.
[10] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G.
Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski,
S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran,
D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through
deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529–533,
02 2015. [Online]. Available: http://dx.doi.org/10.1038/nature14236
[11] J. Peters, K. Mülling, and Y. Altün, “Relative entropy policy search,”
in Proceedings of the Twenty-Fourth AAAI Conference on Artificial
Intelligence (AAAI 2010), 2010, pp. 1607–1612.
[12] J. Kober and J. R. Peters, “Policy search for motor primitives in
robotics,” in Advances in neural information processing systems, 2009,
pp. 849–856.
[13] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, “Trust
region policy optimization,” in Proceedings of the 32nd International
Conference on Machine Learning (ICML-15), D. Blei and F. Bach, Eds.
JMLR Workshop and Conference Proceedings, 2015, pp. 1889–1897.
[Online]. Available: http://jmlr.org/proceedings/papers/v37/schulman15.
pdf
[14] R. Calandra, A. Seyfarth, J. Peters, and M. P. Deisenroth, “Bayesian
optimization for learning gaits under uncertainty,” Annals of Mathematics and Artificial Intelligence (AMAI), vol. 76, no. 1, pp. 5–23,
2016.
[15] A. J. Ijspeert, “Central pattern generators for locomotion control in
animals and robots: A review.” Neural Networks, vol. 21, no. 4, pp.
642–653, 2008.
[16] M. Dogar, K. Hsiao, M. Ciocarlie, and S. Srinivasa, “Physics-Based
Grasp Planning Through Clutter,” in Robotics: Science and Systems
VIII, July 2012.
[17] K. M. Lynch and M. T. Mason, “Stable pushing: Mechanics, controllability, and planning,” IJRR, vol. 18, 1996.
[18] T. Merili, M. Veloso, and H. Akin, “Push-manipulation of Complex
Passive Mobile Objects Using Experimentally Acquired Motion
Models,” Autonomous Robots, pp. 1–13, 2014.
[19] J. Scholz, M. Levihn, C. L. Isbell, and D. Wingate, “A Physics-Based
Model Prior for Object-Oriented MDPs,” in Proceedings of the 31st
International Conference on Machine Learning (ICML), 2014.
[20] J. Zhou, R. Paolini, J. A. Bagnell, and M. T. Mason, “A convex
polynomial force-motion model for planar sliding: Identification and
application,” in 2016 IEEE International Conference on Robotics and
Automation, ICRA 2016, Stockholm, Sweden, May 16-21, 2016, 2016,
pp. 372–377.
[21] M. Deisenroth, C. Rasmussen, and D. Fox, “Learning to Control a
Low-Cost Manipulator using Data-Efficient Reinforcement Learning,”
in Robotics: Science and Systems (RSS), 2011.
[22] J. Hamrick, P. W. Battaglia, T. L. Griffiths, and J. B. Tenenbaum,
“Inferring mass in complex scenes by mental simulation,” Cognition,
vol. 157, pp. 61–76, 2016.
[23] M. B. Chang, T. Ullman, A. Torralba, and J. B. Tenenbaum, “A
compositional object-based approach to learning physical dynamics,”
Under review as a conference paper for ICLR, 2017.
[24] P. Battaglia, R. Pascanu, M. Lai, D. J. Rezende, and K. Koray,
“Interaction networks for learning about objects, relations and physics,”
in Advances in Neural Information Processing Systems, 2016.
[25] L. Ljung, Ed., System Identification (2Nd Ed.): Theory for the User.
Upper Saddle River, NJ, USA: Prentice Hall PTR, 1999.
[26] K. Yu, J. J. Leonard, and A. Rodriguez, “Shape and pose recovery
from planar pushing,” in 2015 IEEE/RSJ International Conference
on Intelligent Robots and Systems, IROS 2015, Hamburg, Germany,
September 28 - October 2, 2015, 2015, pp. 1208–1215.
[27] P. Agrawal, A. Nair, P. Abbeel, J. Malik, and S. Levine, “Learning
to poke by poking: Experiential learning of intuitive physics,” NIPS,
2016.
[28] K. Fragkiadaki, P. Agrawal, S. Levine, and J. Malik, “Learning visual
predictive models of physics for playing billiards,” in ICLR, 2016.
[29] T. D. Ullman, A. Stuhlmüller, N. D. Goodman, and J. B. Tenenbaum,
“Learning physics from dynamical scenes,” in Proceedings of the ThirtySixth Annual Conference of the Cognitive Science Society, 2014.
[30] J. Wu, I. Yildirim, J. J. Lim, B. Freeman, and J. Tenenbaum, “Galileo:
Perceiving physical object properties by integrating a physics engine
with deep learning,” in Advances in Neural Information Processing
Systems, 2015, pp. 127–135.
[31] A. Byravan and D. Fox, “Se3-nets: Learning rigid body motion using
deep neural networks,” CoRR, vol. abs/1606.02378, 2016.
[32] C. Finn and S. Levine, “Deep visual foresight for planning robot
motion,” ICRA 2017.
[33] R. Zhang, J. Wu, C. Zhang, W. T. Freeman, and J. B. Tenenbaum,
“A comparative evaluation of approximate probabilistic simulation
and deep neural networks as accounts of human physical scene
understanding,” CoRR, vol. abs/1605.01138, 2016.
[34] W. Li, S. Azimi, A. Leonardis, and M. Fritz, “To fall or not to fall: A
visual approach to physical stability prediction,” vol. arXiv:1604.00066
[cs.CV], 2016.
[35] A. Lerer, S. Gross, and R. Fergus, “Learning physical intuition of
block towers by example,” in Proceedings of the 33nd International
Conference on Machine Learning, ICML 2016, New York City, NY,
USA, June 19-24, 2016, 2016, pp. 430–438.
[36] L. Pinto, D. Gandhi, Y. Han, Y. Park, and A. Gupta, “The curious
robot: Learning visual representations via physical interactions,” CoRR,
vol. abs/1604.01360, 2016.
[37] W. Li, A. Leonardis, and M. Fritz, “Visual stability prediction and its
application to manipulation,” CoRR, vol. abs/1609.04861, 2016.
[38] M. Denil, P. Agrawal, T. D. Kulkarni, T. Erez, P. Battaglia, and
N. de Freitas, “Learning to Perform Physics Experiments via Deep
Reinforcement Learning,” 2016.
[39] W. Yu, C. K. Liu, and G. Turk, “Preparing for the unknown:
Learning a universal policy with online system identification,”
CoRR, vol. abs/1702.02453, 2017. [Online]. Available: http:
//arxiv.org/abs/1702.02453
[40] A. Marco, F. Berkenkamp, P. Hennig, A. P. Schoellig, A. Krause,
S. Schaal, and S. Trimpe, “Virtual vs. real: Trading off simulations
and physical experiments in reinforcement learning with bayesian
optimization,” in 2017 IEEE International Conference on Robotics
and Automation, ICRA 2017, Singapore, Singapore, May 29
- June 3, 2017, 2017, pp. 1557–1563. [Online]. Available:
https://doi.org/10.1109/ICRA.2017.7989186
[41] P. Hennig and C. J. Schuler, “Entropy Search for Information-Efficient
Global Optimization,” Journal of Machine Learning Research, vol. 13,
pp. 1809–1837, 2012.
[42] C. E. Rasmussen and C. K. I. Williams, Gaussian Processes for
Machine Learning. The MIT Press, 2005.
[43] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “Openai gym,” arXiv preprint
arXiv:1606.01540, 2016.
[44] Y. Duan, X. Chen, R. Houthooft, J. Schulman, and P. Abbeel,
“Benchmarking deep reinforcement learning for continuous control,” in
ICML, 2016, pp. 1329–1338.
| 2cs.AI
|
September 12, 2017
©
Multivariate Density Modeling for Retirement Finance
CHRISTOPHER J. ROOK
ABSTRACT
Prior to the financial crisis mortgage securitization models increased in sophistication as did products
built to insure against losses. Layers of complexity formed upon a foundation that could not support
it and as the foundation crumbled the housing market followed. That foundation was the Gaussian
copula which failed to correctly model failure-time correlations of derivative securities in duress. In
retirement, surveys suggest the greatest fear is running out of money and as retirement decumulation
models become increasingly sophisticated, large financial firms and robo-advisors may guarantee
their success. Similar to an investment bank failure the event of retirement ruin is driven by outliers
and correlations in times of stress. It would be desirable to have a foundation able to support the
increased complexity before it forms however the industry currently relies upon similar Gaussian (or
lognormal) dependence structures. We propose a multivariate density model having fixed marginals
that is tractable and fits data which are skewed, heavy-tailed, multimodal, i.e., of arbitrary complexity
allowing for a rich correlation structure. It is also ideal for stress-testing a retirement plan by fitting
historical data seeded with black swan events. A preliminary section reviews all concepts before they
are used and fully documented C/C++ source code is attached making the research self-contained.
Lastly, we take the opportunity to challenge existing retirement finance dogma and also review some
recent criticisms of retirement ruin probabilities and their suggested replacement metrics.
TABLE OF CONTENTS
Introduction ............................................................................................................................................ 1
I. Literature Review ............................................................................................................................. 2
II. Preliminaries.................................................................................................................................... 3
III. Univariate Density Modeling ....................................................................................................... 29
IV. Multivariate Density Modeling w/out Covariances ..................................................................... 37
V. Multivariate Density Modeling w/Covariances ............................................................................ 40
VI. Expense-Adjusted Real Compounding Return on a Diversified Portfolio .................................. 47
VII. Retirement Portfolio Optimization ............................................................................................. 49
VIII. Conclusion ................................................................................................................................ 51
References ............................................................................................................................................ 52
Data Sources/Retirement Surveys........................................................................................................ 55
IX. Appendix with Source Code ........................................................................................................ 56
Keywords: variance components, EM algorithm, ECME algorithm, maximum likelihood, PDF,
CDF, information criteria, finite mixture model, constrained optimization, retirement decumulation,
probability of ruin, static/dynamic glidepaths, financial crisis
Contact: [email protected]
A financial security that is purchased for $pt-1 at time t-1 with all distributions reinvested yields a value at
time t called the adjusted price, say $Pt, for t = 1,2,...,T. The total return at time t is Rt = (Pt - pt-1)/pt-1 and the
total compounding return is 1+Rt = Pt/pt-1 so that $pt-1(1+Rt) = $Pt. If the inflation rate between times t-1 and t
is It then 1+Rt = (1+It)(1+rt), where rt = (1+Rt)/(1+It) – 1 is the real return at time t. The real price at time t is
the value $pt such that (pt-pt-1)/pt-1 = rt, which upon solving yields $pt = $Pt/(1+It). In an efficient market real
prices are governed by a geometric random walk (GRW), that is, ln($pt) = ln($pt-1) + St, where St ~ N(μ,σ2). A
value of μ > 0 represents a drift and is the expected log-scale price increase sufficient to compensate the
investor for risk (σ2) between times t-1 and t. In a random walk, the next value is the current value plus a
random normal step, St, and the best predictor of it is the current value + μ. Exponentiating both sides of the
GRW model yields the alternative form $pt = ($pt-1)e , where e = (1+rt) ~ lognormal(μ,σ2). Under strict
conditions, the normally distributed step, St, can be justified. Decompose the time between t-1 and t into a
series of smaller segments, say, d = 1,2,...,D and let 1+rd be independent and identically distributed (iid)
random variables (RVs) for the compounding real return between times d-1 and d, so that the ln(1+rd) are
also iid RVs. The compounding real return at time t is (1+rt) = ∏
∑
ln 1
r
1
r
e∑
= e where St =
~ N(μ,σ2) when d ≥ 30 by the Central Limit Theorem (CLT). Here, t can represent years and
d days so that the compounding yearly return is the product of compounding daily returns. The lognormal
assumption for (1+rt) breaks down when the (1+rd) are not iid for d=1,2,...,D and there is ample research to
indicate that the correlation between returns increases as the time length decreases. We also find that
compounding returns on liquid securities used in retirement finance are often better fit by the normal
probability density function (PDF) than the lognormal, suggesting that short term real compounding returns
may not be iid (see §II.C). A further complication is that the normal PDF is generally considered tractable,
whereas the lognormal PDF is not. For example, a diversified portfolio of equities and bonds with real
returns et and bt, respectively, has compounding real return α(1+et) + (1-α)(1+bt) where α is the equity ratio.
Unfortunately, no known PDF exists for the sum of correlated lognormal RVs and we are left to approximate
it for a given α, see Rook & Kerman (2015) for an implementation. Despite the benefits of using normal
RVs to model compounding real returns in finance many practitioners and researchers will not, due primarily
to the lack of skewness and heavy tail, but also because the normal PDF can generate negative prices. The
spectacular failure of Gaussian copulas during the financial crisis reinforces the skepticism. Unfortunately,
those who reject the normal PDF do not benefit from discrete-time finance models optimized using it. This
research is motivated by the dilemma, particularly the desire for skewed, heavy-tailed, multimodal, tractable
PDFs to model the compounding real return on a diversified portfolio in discrete-time finance applications.
Of particular interest is the 1895 claim by Karl Pearson that the moments of a lognormal PDF are virtually
indistinguishable from a mixture of normals (McLachlan & Peel, 2000).
1
I. Literature Review
During the housing boom residential mortgages were packaged and sold as securities. The price of a
security is the present value of future cash flows, which here are the mortgage payments. The products were
partitioned into tranches so that as borrowers defaulted low-level holders suffered first, followed by midlevel and then top-level (MacKenzie & Spears, 2014). Cash flows and timings are needed to price a tranche
which is a function of which loans have defaulted by each time point. Default times can be modeled using an
exponential PDF with the probability of default before some time returned by its cumulative distribution
function (CDF). The probability of simultaneous defaults before given times is computed from the copula or
multivariate CDF and depends on the correlation between default times. There is no way to estimate the true
correlation between default times of residential borrowers due to lack of data.
Li (2000) suggested
translating the copula on simultaneous defaults to an equivalent expression using normal RVs.
The
correlation between these RVs is pulled from a measure on the underlying debt instrument for which the
normal assumption is reasonable and sufficient data exists. Using these correlations the Gaussian copula can
return the probability of simultaneous defaults before specific times. Samples on the correlated exponential
failure times can then be simulated from the Gaussian copula and used to value the security.
Loan pools held ≅5000 mortgages with equity tranches acting like a stock and senior tranches like a
safe bond. Low interest rates led to excess liquidity and produced an insatiable appetite from pension and
sovereign wealth funds for AAA-rated senior tranches which yielded more than U.S. Treasurys (Kachani,
2012). A fatal flaw in the system was that economists have assumed for decades that financial data
originates from regimes and correlations change during crises (Hamilton, 1989). Since housing busts follow
housing booms it was unwise, in hindsight, to measure correlation with one value. As witnessed, defaulttime correlations increase in a crisis and senior tranches sold as safe bonds behaved more like a stock which
devastated the insurers, who by 2007 had underwritten $62.2 trillion of credit default contracts (up from $1.6
trillion in 2002). Blame for the crisis has focused on the Gaussian copula (Salmon, 2009), with a takeaway
being that ‘normal returns are not appropriate in finance’ (Nocera, 2009). Researchers and practitioners who
warned against using the normal distribution were vindicated. Paolella (2013) subsequently declared ‘the
race is on’ to find more suitable multivariate PDFs for financial applications and provides an overview of
mixture densities, which are often used to model economic regimes and form the basis for this research.
The PDF we develop is a multivariate normal mixture having fixed normal mixture marginals. It is
tractable when used in discrete time retirement decumulation models, and intuitive to understand. In §II we
detail the needed theory/techniques, and in §III we fit generic univariate normal mixtures to sets of returns.
In §IV we form the multivariate PDF and add correlations in §V. Finally in §VI and §VII we derive the
expense-adjusted real compounding return on a diversified portfolio and use it within optimal decumulation
models. Supporting proofs, derivations, and a full C/C++ implementation are included in the Appendix.
2
II. Preliminaries
Foundational concepts needed for the density model developed in §III thru §V are presented here.
A. Probability Density & Cumulative Distribution Functions
Let X be a continuous RV and f(x) a function such that f(x) ≥ 0 ∀ x ∈
with
= 1. The
function f(x) is a valid PDF for X (Casella & Berger, 1990). The CDF for X is defined as F(x) = P(X x) =
. By the 2nd Fundamental Theorem of Calculus (Anton, 1988), F′(x) = f(x) - f(-∞) = f(x), that is, the
PDF of an RV X is the 1st derivative of its CDF. Note that X may be defined on a subset of , and f(·)
usually depends on a vector of parameters, say θ, which may represent the mean and variance of X. Other
common expressions for the PDF include f(x,θ), f(x;θ), f(x|θ), and it may also be denoted by fX(·) to indicate
the RV governed, written as X ~ fX(·). For a single RV X, f(x) is a univariate PDF, but the above also applies
n
to an n-dimensional vector of RVs, = (X1, ..., Xn)′, defined on
≥0 ∀
∈
n
, and
Ω = { :⋂
= 1. The multivariate CDF of
}.
. Here, f ( ) is a multivariate PDF with f ( )
is defined as F( ) = P(
)=
, where
Similar to the univariate case, differentiating a multivariate CDF yields the
multivariate PDF, that is,
…
F
=
, and the marginal PDF of one RV, say, X1 is obtained by
integrating out all other RVs. That is, f(x1) =
.
∙∙∙
B. Finite Mixture Densities
Let X be a continuous RV and let f 1(x), ..., f g(x) be g functions that satisfy the univariate PDF
conditions in §II.A. Also, let π1, ..., πg be probabilities (0 < πi 1, i=1,...,g) such that π1 + ... + πg = 1. Then f(x) =
π1f 1(x) + ... + πg f g(x) also satisfies the PDF conditions in §II.A and is called a finite mixture density
∑gi 1 πi
(Titterington, et al., 1985). If X ~ f(x) then F(x) = P(X x) =
CDF of X. Let μ
∑
π
i
= π1F1(x) + ... + πgF g(x) is the
be the rth moment for f i(x). Then E(Xr) = μ(r) =
=∑
π μ . Thus, E(X) = ∑
π μ and V(X) = E(X2) – E(X)2 = ∑
When = (X1, ..., Xn)′ is an n-dimensional vector of RVs on
n
,
=
∑
π μ
=
π
- ∑
π μ
.
= π1f 1( ) + ... + πg f g( ) is a multivariate
mixture PDF and satisfies the multivariate PDF conditions set forth in §II.A.
A mixture PDF, f(x), has two distinct interpretations: (1) f(x) is a function that accurately models the
PDF’s shape/form for an RV X, or (2) the RV X originates from component density f i(x) with probability πi,
i=1,2,...,g, and the components have labels. While parameter estimation is unaffected by the interpretation,
the underlying math is. During parameter estimation we adopt the interpretation that simplifies the math.
Each component density f i(x) may depend on a parameter vector θi, and let = (π1, ..., πg)′ be the vector of
component probabilities. When iid observations from f(x) are drawn, say xt, t=1,2,...,T, the objective is to
estimate the parameters, θi, i=1,2,...,g and . Once estimated the PDF is fully specified and can be used.
3
Under interpretation (2) above, components have meaning. Mixture PDFs model serially correlated
data via the components. Each component is considered a state and there are g states at each time t. If we
assume that observation xt+1 depends only on the prior observation xt, and that the long-run probabilities of
transitioning between states are stationary, then state transitions evolve over time as a Markov chain (Hillier
& Lieberman, 2010). Define the gxg matrix{πij} i,j=1,2,...,g as the conditional probability of being in state j
at time t+1 given that we are in state i at time t. Serially correlated data originating from a mixture PDF thus
requires estimation of the transition probabilities {πij} in addition to θi and , where πi is now interpreted as
the unconditional probability of being in state i at time t (used at time t=0) (McLachlan & Peel, 2000). In
finance/economics, the g states of a mixture PDF are called regimes and the process by which dependent
observations transition between states over time is termed regime switching (Hamilton, 1989)1.
As noted, the underlying math differs between interpretations (1) and (2) above. Under interpretation
(2) components have labels, thus observations from a mixture PDF can be viewed as coming in pairs (xt, zt),
where xt is the actual value and zt is the component that generated it, t=1,2,...,T. It is common to replace zt by
a vector
= { }, for t=1,2,...,T & j=1,2,...,g which has a 1 in the component slot and 0’s elsewhere, e.g.,
component=2 at time t=1 can be expressed as z1 = 2 or
= (0,1,0,...,0)′ (see Figure I).
Figure I. Mixture Data Collection when Components have Labels
Note: Each ztj = 0 or 1 and ∑
=1, for t=1,2,...,T.
This representation applies to dependent and independent data. Mixture PDFs for dependent data are
termed hidden Markov models (HMM) (Rabiner & Juang, 1986) because the state vector
generally cannot
be observed (it is hidden). Thus a critical task in HMM model building is determining which state generated
each observation, xt, t=1,2,...,T. Under interpretation (2), the PDF incorporates the RV Zt, observed as
, as
f(xt, ) = f(xt | )·p[Zt= ] or f(xt, zt1, ..., ztg) = f(xt | zt1, ..., ztg)·p[Zt=(zt1, ..., ztg)′], which is given by:
,
∙
∙
,…,
∙
1
,
,
,
⋮
for
for
1
1
for
1
The regimes in Hamilton (1989) were time series models. In this research they are PDFs.
4
,
(2.B.1)
or, more compactly:
,
,…,
∙
∙ …∙
∙
∙
.
(2.B.2)
For example, suppose N returns on a financial security are observed over time and appear symmetric
around some overall mean, do not exhibit serial correlation, but do include black swan events2 at a frequency
greater than their corresponding tail probabilities under either a normal or lognormal PDF (Taleb, 2010).3
An intuitive tractable PDF for such returns is Tukey’s contaminated normal PDF (Huber, 2002), which is a
mixture of two normals with equal means but unequal variances. It can be used to thicken the tail of a
normal PDF. The density with larger variance generates outliers and has a smaller probability.4 We can
proceed intuitively by partitioning the returns into two sets with one holding the n1 non-outliers, and the
other holding the n2 outliers. A normal PDF can be fit to each set using MLEs, for example with the mixture
weights set to πi= ni/N, i=1,2.5
If X is an RV representing these returns, then X ~ f(x) = π1 f 1(x) + π2 f 2(x),
where f i(x) = N(μ,σi), i=1,2 and σ2 > σ1. After replacing all parameters by their estimates, suppose 75% of the
returns originate from a “Common” PDF (non-outliers) which is N(0.08,0.10), and 25% from a “Gray Swan”
PDF (outliers) which is N(0.08,0.30), where μ=0.08 is the overall mean, estimated by
(see Figure II).
Figure II. Example of Tukey’s Contaminated Normal PDF
By labeling components we are using interpretation (2). Note that a mixture of normal PDFs is generally not
normally distributed, and not symmetric (this being an exception). A note of caution is not to mistake the
mixture PDF in Figure II with the RV Z = 0.75*X1 + 0.25*X2, where X1 ~ N(0.08,0.10) and X2 ~ N(0.08,0.30),
as clearly Z is normally distributed. In practice, a g-component normal mixture can model any PDF and is
tractable (McLachlan & Peel, 2000). For example, a mixture of two normals can closely approximate the
2
Observed outliers are called gray swan events. Black swans are extreme events that have not occurred before.
Taleb (2010) discourages use of the normal PDF in finance/economics and refers to the lognormal PDF as a “bad compromise”.
4
Fama (1965) discussed using mixture PDFs to explain non-normality in stock prices, and Taleb (2010) also used mixtures in practice to
add a heavy tail to the normal PDF as an alternative to the lognormal PDF.
5
This ad-hoc procedure is for illustration only. Such methods were common prior to the advent of the EM algorithm (see, Johnson, et al.,
1994). As will be seen, mixture PDFs are almost always calibrated today using either the EM Algorithm or a gradient procedure.
3
5
lognormal PDF (Titterington, et al., 1985). Lastly, since the mixture PDF in Figure II does not model black
swan events it may be deemed unsatisfactory. A solution could be to add a 3rd component labeled “Black
Swan” as, say N(0.08,0.50), with small probability such as π3 = 0.05, then adjust π1 and π2 so that ∑
π = 1.6
C. The Central Limit Theorem (CLT)
Let Xt
~
f(x), t=1,2,...,T, with E(Xt)=μ and V(Xt)=σ . For large T, ∑
X ~ N(Tμ, √Tσ) per the
central limit theorem (Freund, 1992). In words, the sum of iid RVs from any PDF, f(x), is approximately
normal (denoted ~) for large samples and the rule-of-thumb is T ≥ 30. Unfortunately, this rule does not
always apply to mixture PDFs. Consider f(x) = π1 f 1(x) + π2 f 2(x), where π1=(1-10-5) and π2=10-5, and let f 1(x) =
N(0,1) and f 2(x) = N(1010,1). If X ~ f(x), then μ = E(X) = 105 and σ = V X ≅ 3.16E7 (from §II.B). In an iid
sample of size T=30 from f(x) we are unlikely to draw an observation from f 2(x) leaving ∑
in violation of the CLT, which ensures ∑
f(x) and ∑
X ~ N(0, √30)
X ~ N(30·105, √30·3.16E7). Such a sample was generated from
X = 6.77 with all observations originating from f 1(x). This value is within a σ of the CLT PDF
mean thus is a valid value. Repeating the process 9 times should produce an iid sample of size N=10 from
the CLT PDF. It does not as all 10 values are < the CLT PDF mean.7
Caution is therefore advised when invoking the CLT on RVs from mixture PDFs. For example, let rd
~ f(r) be an RV for the daily real return8 on the S&P 500 Index, d=1,2,...,D, where D=# of trading days/year.
The amount $B invested on January 1st, will grow to $E
$B ∙ ∏
The annual real compounding return is A = ∏
and ln(A) = ∑
1
r
1
the CLT when the rd are iid (D ≫ 30). By definition, the RV Y = e
r
on December 31st, in real dollars.
ln 1
r , which is ~ normal per
= A would then be ~ lognormal,
making the historical collection of annual real S&P 500 Index returns a lognormal random sample. This
hypothesis was tested and rejected using the Anderson-Darling test (Rook & Kerman, 2015).
One
explanation is that daily returns are not independent, a claim supported by academic studies of short-term
index returns (Baltussen et al., 2016).9 Another is that daily returns are not identically distributed, or that
daily returns originate from a mixture PDF with D not large enough for the CLT approximation.
D. The Density of a Future Observation
Under certain assumptions the PDF of a future value can be derived before it is observed. Let Xt
~
N(μ,σ) be compounding returns on a financial security at time t=1,2,...,T+1 (μ and σ unknown). Suppose xt
is the observed value of Xt for t=1,2,...,T, with XT+1 reflecting the unobserved next value. Note that X ~
6
Such an approach is considered for stress-testing a retirement plan.
Full sample (N=10): 6.77, 11.71, 3.04, -6.75, 5.65, -0.49, -3.53, 4.58, 7.95, 5.55 (mean=3.45, standard deviation=5.56).
8
The daily (real) compounding return can be approximated by 1+rd = (1 + Rd)/ 1 I , where Rd is the total daily return calculated as
(end value – start value)/start value, and I is the annual inflation rate.
9
While short term (i.e., daily) index returns have historically exhibited positive serial correlation, Baltussen, et al., (2016) suggest the
ubiquity of index products may have eliminated the signal or even turned it negative.
7
6
X is the sample mean and (T-1)S2/σ ~
∑
N(μ,σ/√T), where X
and S2 are independent RVs (Ross, 2009). Since X and X
X)/σ 1
are also independent, X
1/T ~ N(0,1). Thus (2.D.1) has form N(0,1)/
/ T
X
is the
X
X
X
1
T
σ 1
S
σ
~X
S 1
X
X
σ 1
1
T
∗
σ
S
X
X
S 1
1
T
X ~ N(0,σ 1
1/T)
1 which follows a Student’s
:11
t-distribution with T-1 degrees of freedom (Ross, 2009), denoted by Ŧ
Consequently, X
∑
is a chi-square RV with T-1 degrees of freedom.10 It has been established that X
sample variance, and
and X
, where S2 =
~Ŧ
.
(2.D.1)
, or a future observation has the PDF of a scaled Student’s t-
1/T Ŧ
distribution centered at the sample mean. In a simulation study, this would be the preferred PDF for
evaluating a financial plan. Simulating from a Student’s t-distribution with T-1 degrees of freedom is
straight-forward if we can simulate from the normal distribution. First, generate a N(0,1) random value for
random value
the numerator then generate, square, and sum T-1 additional N(0,1) values to construct a
for the denominator.12 Finally, use the ratio from the definition of a Ŧ
RV given above (Law & Kelton,
2000). The PDF of a future value is
which can be derived by differentiating the CDF (see §II.A).
The CDF for a future observation XT+1 is F
= P(XT+1 ≤ x) = P Ŧ
X / S 1
1/T , which is given
by (Freund, 1992):
/
X
FŦ
S 1
1
T
Γ
T
T
2
1 Γ
T
2
1
1
T
1
,
(2.D.2)
where Γ(·) is the gamma function. Finally the PDF for a future observation XT+1 is
X / S 1
1/T
FŦ
which is derived using §II.A and the chain rule as:
Γ
T
The PDF for future values X
T
2
1 Γ
T
2
, …,X
1
1
X
S 1 1/T
T 1
1
S 1
1
T
, for
∞
x
∞.
(2.D.3)
can be derived similarly. The multivariate PDF of N future values
is the product of the univariate PDFs, assuming independence. Note that the technique just described breaks
10
The distributions of X and S2 here are not approximations (i.e., the CLT is not involved), therefore are valid for any sample size T.
The normal numerator and chi-square denominator must be independent RVs.
12
If X1, X2, ..., Xn ~ N(0,1), then ∑ X ~ , (Ross, 2009).
11
7
down when a 2nd asset is added. Let Xt
~
N(μx,σ ) and Yt
N(μy,σ ) be compounding returns on two
~
uncorrelated financial securities at times t=1,2,...,T+1. Interest is in modeling a future unobserved value on a
diversified portfolio using these securities, say, RT+1 = αXT+1 + (1- α)YT+1, where 0 α 1. It follows that
α(XT+1-X) + (1-α)(YT+1-Y) ~ N(0,
1/T α σ
1
α X
X
1
1
T
1
1
α σ
α Y
1
), thus:
α σ
Y
(2.D.4)
~ N 0,1 .
α σ
As with 1 asset, suppose a quantity Q(σx,σy) exists such that a function of it and (2.D.4) has a known PDF that
is not a function of (σx, σy), and it can be solved for the numerator of (2.D.4). This would suggest a solution to
the Behren’s-Fisher Problem which is a famous unsolved problem in statistics (Casella & Berger, 1990).13
E. Maximum Likelihood (ML) Estimation
Let Xt ~ f(x,θ), t=1,2,...,T, be continuous RVs and xt be the observed value of Xt. The likelihood of xt
is f(xt,θ), which is the PDF of the observed value. The vector θ holds unknown parameters such as θ=(μ,σ)′,
where E(Xt)=μ and V(Xt)=σ2. A likelihood value is not a probability and can be >1, however, it is a similar
measure since values with higher likelihoods are more likely to be observed. The likelihood function, (·), is
the PDF written as a function of θ, i.e., (θ|xt)=f(xt,θ). Extending this to the entire sample, the multivariate
PDF of =(X1,..., XT)′ evaluated at x = (x1, ..., xT)′ is f( ,θ) = f(x1,..., xT,θ) which can be written as (θ| ) =
(θ|x1, ..., xT), the likelihood of the entire sample. An appealing estimate for θ is that which maximizes
(θ| ) and its value, denoted , is called the maximum likelihood estimator (MLE). Since the natural log
function is increasing, maximizing (θ| ) and ln( (θ| )) are equivalent problems and it is often easier to deal
with the latter. If the RVs Xt are independent, t=1,2,...,T, then (θ| ) = f(x1, ..., xT,θ) = ∏
ln( (θ| )) = ∑
ln
,
,
, and
. MLEs possess many desirable statistical qualities, such as consistency,
efficiency, asymptotic normality, and invariance (to functional transformations) thus are often considered the
gold standard for parameter estimation. Such qualities, however, depend on certain regularity conditions
being satisfied (Hogg, et al., 2005), see §II.G. Finding parameter estimates in statistics therefore enters the
purview of engineering disciplines that specialize in constrained optimization techniques, see §II.K. The
likelihood function for an iid sample originating from a g-component mixture PDF depends on the
interpretation, see §II.B. Let
(1)
(θ| ) = f( ,θ) and
(2)
(θ| ,
,…,
) = f( ,
,…,
,θ) be the likelihoods of
θ=( ,θ1,θ2, ...,θg) under interpretations (1) and (2), respectively. Then,
13
The Behren’s-Fisher Problem tests for equal means in two normal populations with unequal (and unknown) variances. Let
X1,...,Xn1
~
N(μx,σ ) and Y1,...,Yn2
~
N(μy,σy) be independent samples. Under Ho: μx = μy, X
Q(σ , σ ) to make this Ŧ or any other fully specified PDF is unknown.
8
Y / σ ⁄n
σ ⁄n ~ N(0,1). A
|
|
→ ln
⋯
,
ln
,
⋯
,
,
,
(2.E.1)
ln
,
,
(2.E.2)
and, using (2.B.2),
| ,
,…,
,
→ ln
| ,
∙ …∙
,
z ∙ ln
,…,
,
ln
,
(2.E.3)
.
(2.E.4)
The goal is to collect data and maximize the log-likelihood from either (2.E.2) or (2.E.4) with respect to ,
obtaining the MLE . Unfortunately, the log-likelihood for interpretation (2) in (2.E.4) cannot be maximized
directly because the component indicator RVs Zt = (zt1, zt2, ..., ztg)′ are missing (i.e., not observable). The loglikelihood from interpretation (1) in (2.E.2) can be maximized directly, however this function is unpleasant
for a variety of reasons. For example, it can have multiple local maximums thus finding stationary points
does not guarantee an MLE with its desirable properties. It also is unbounded for normal components thus
given any value , no matter how large, we can always find a setting for
such that ln(
(1)
(θ| )) > , see
Appendix A. When maximizing the mixture log-likelihood for normal components we will therefore restrict
the parameter space for
to a region where ln(
(1)
(θ| )) is finite and search for all local maximums, declaring
as the argmax. To restrict the parameter space for a g-component mixture PDF, let σ be the variance for
component i, i=1,...,g. The variance ratio constraint is max(σ )/min(σ < C, where C is a given constant
(McLachlan & Peel, 2000). A good choice for C will eliminate spurious maximizers which are optimal
values that lack a meaningful interpretation and can occur when one component fits a small # of observations.
F. The Expectation-Maximization (EM) Algorithm
Several researchers had been using a 2-step process to obtain MLEs in studies with missing data.
The process was observed to possess many interesting properties and became formalized with proofs and a
name by Dempster et al. (1977) in what has become one of the most influential statistics papers ever written.
The procedure, termed the Expectation-Maximization (EM) algorithm, generates MLEs as follows: Let
(Xt,Yt) ~ f(xt, yt,θ) and suppose Xt is observed as xt but Yt is missing, for t=1,...,T. When (Xt,Yt) are iid over
time the joint PDF of =(X1, ..., XT)′ and =(Y1, ..., YT)′ is f( ,y,θ)= ∏
,
,
.14 Depending on θ, the
marginal PDF of Xt may be obtained by integrating yt out of f(xt, yt, θ) when Yt is continuous, or summing it
out when Yt is discrete, see §II.A. That is, f(xt,θ) =
14
,
,
or f(xt,θ) = ∑
,
,
for t=1,...,T, then
The EM algorithm does not require iid observations and can also be used to estimate parameters in HMM models (see §II.B).
9
f( ,θ) = ∏
. This results in 2 likelihood functions for θ, one that includes the missing data and one
,
that does not, i.e., (θ| ,y) = f( ,y,θ) and (θ| ) = f( ,θ). The EM-algorithm computes , the MLE of θ as:
Begin:
Initialize
to starting values
E-Step: Compute EY ln
| ,
M-Step: Maximize EY ln
End:
| ,
.
only when taking expectations wrt Y.
. Use
with respect to . Go to E-Step using = in the expectations.
Terminate when ( | ) stops increasing (use %-change below some threshold).
The E-Step replaces the missing yt, t=1,...,T, with constants as a result of taking expectations and the M-Step
therefore only has
unknown. The value of (θ| ) will not decrease while iterating and it will end at the
local maximum nearest to the starting values for , given (θ| ) is bounded in this region. If (θ| ) has
multiple local maximums we use a variety of starting
values and take
as the argmax. This value will
exhibit the desirable qualities noted in §II.E (McLachlan & Peel, 2000). The EM-algorithm can be used to
find MLEs for a wide variety of models. The trick is to reformulate so that some RVs appear missing.
Applications to mixture PDFs is straightforward. Under interpretation (2) the component indicator RVs Zt =
(zt1, ..., ztg)′ are missing (see §II.B, §II.E). The multivariate PDF of Xt and Zt is given in (2.B.2) and the
corresponding marginal of Xt is f(xt,θ)=∑
⋯
,
,
∑
,…,
∙
∙ …∙
∙
, which is the PDF used in interpretation (1), see (2.E.1). (The Zt are discrete
,
RVs.) The likelihood function for an iid sample (x1,..., xT) from a g-component mixture PDF including the
| ,
missing data RVs is
is
|
from (2.E.3), and the corresponding likelihood without missing data
,…,
| ,
from (2.E.1). The EM-algorithm E-Step uses EZ ln
by their expected value. Notice that ln
E ln
| ,
| ,
,…,
is linear in the
E z
,…,
∙ ln
which replaces the
,…,
=(zt1, ..., ztg)′ (see 2.E.4), thus:
ln
,
(2.F.1)
,
where E(zti) = E(zti| , ), i.e., computed using all available non-missing data along with the current settings
for =( , θ1, ..., θg). Since zti is a discrete RV that equals 1 when xt originates from component i, and 0
otherwise, E(zti| , ) = 1·P[zti=1| , ] + 0·P[zti=0| , ] = P[zti=1| , ], which is:
E z | ,
P
1| ,
1,
,
,
, |
1 ∙P
,
,
1
,
,
(2.F.2)
for t=1,...,T and i=1,...,g. When
is given, this value is completely known and replaces E(zti) in (2.F.1).
The resulting function with only
unknown is optimized in the M-Step. Initial values for
randomly using simulation (McLachlan & Peel, 2000) and, since
the strategy is to apply the EM-algorithm to a variety of starting
10
(1)
can be set
(θ| ) may have many local optimums,
values and select
as the argmax.
G. Regularity Conditions
Statistical tests, models and theorems, are built upon sets of assumptions and in statistical inference
these assumptions are called the regularity conditions. Hogg, et al., (2005) Appendix A describe 9 such
conditions and it is usually the case that only a subset need be satisfied for a given result. The 1st regularity
condition applies to PDFs and deals with uniqueness, namely for PDF f(·), if θ
then f(x,θ) f(x, ). This
condition clearly holds for N(μ,σ) PDFs since changing the mean or variance changes the distribution.
However, consider the mixture PDF f(x,θ) = π1 f 1(x) + π2 f 2(x) where f i(x) ~ N(μi,σi), i=1,2. The vector of
unknown parameters is θ=(π1, μ1, σ1, π2, μ2, σ2)′. Define =(π2, μ2, σ2, π1, μ1, σ1)′ and note that θ
but f(x,θ) =
f(x, ) violating regularity condition #1. In general, mixture PDFs do not satisfy all regularity conditions and
caution is advised when using results that requires them.
H. The Likelihood Ratio Test (LRT)
Let (x,θ) be an arbitrary statistical model and (x,θR) be the same model with some parameters
dropped. For example, (x,θ) may be a linear regression model with θ holding the coefficients and (x,θR)
is the corresponding reduced model that excludes some predictor variables. The principle of parsimony
favors statistical models with fewer parameters and the likelihood ratio test (LRT) checks for a significant
difference between the likelihood value of a full model vs. its reduced version. If the likelihoods are not
significantly different the reduced model (x,θR), having fewer parameters, is preferred. The LRT tests for
equivalence of the likelihoods, namely, Ho: ( |x) = ( |x) vs Ha: ( |x) < ( |x), where
and
are the
MLEs for θ and θR, respectively. The test statistic is given by (Hogg, et al., 2005):
λ
Note that ( |x)
|
2ln
(2.H.1)
|
( |x) since adding parameters to a model does not reduce its likelihood and under Ho
the test statistic λ is close to zero since ln(1) = 0. When ( |x) ≫
, λ takes a large positive value and Ho
is therefore rejected when λ > C, for some critical value C. Under Ho and select regularity conditions
(including the 1st, see §II.G), λ ~
, where v is the number of parameters dropped from θ to create θR. A
test of Ho with Type I Error probability α will define C such that P[
C] = α, and a Type I Error means Ho
is rejected when true. We are interested in using the LRT to test for the optimal # of components in a mixture
PDF however, λ is not
under Ho since the applicable regularity conditions are violated. McLachlan
(1987) suggests approximating the null distribution of λ with a bootstrap. The hypothesis test is:
vs.
Ho: Data vector
Ha: Data vector
Under Ho we estimate the PDF of
f(x, ), where
is the MLE of
originates from a g-component mixture PDF, say f(x,θ)
originates from a (g+k)-component mixture PDF, say f(x, )
as f(x, ), where
is the MLE of θ and under Ha we estimate the PDF as
. A value for the test statistic λ from (2.H.1) is computed using the
11
corresponding likelihood functions of these PDFs. If we implicitly assume that f(x, ) generated our sample
vector , a value from the distribution of λ under Ho can be simulated by generating a random sample from
f(x, ) and fitting both a g-component and a (g+k)-component mixture PDF using MLEs. The sample should
be of the same size as our data vector . Repeating this process K times will simulate values λ1, ..., λK which
estimate of the distribution of λ under Ho and the p-value for this test is approximated by (# of λj ≥ λ)/K.
I. Variance-Covariance Matrices
Let = (X1, X2, ..., XN)′ be RVs for the compounding return on N financial assets at a given time point.
If V(Xi)=σ , Cov(Xi,Xj)=σ , and Corr(Xi,Xj)=σ / σ σ =ρ are the variances, covariances, and correlations
for i,j=1,2,...,N then V( )=E[( - E( ))( - E( ))′] is the variance-covariance (VC) matrix of , written as:
V
σ
σ
⋮
σ
σ
σ
⋮
σ
… σ
… σ
⋱
⋮
… σ
σ
ρ σ σ
⋮
ρ σ σ
ρ σ σ
σ
⋮
ρ σ σ
…
…
⋱
…
ρ σ σ
ρ σ σ
⋮
σ
(2.I.1)
The diagonals are the variances σ and the off-diagonals are the covariances σ =ρ σ σ . Note that V( ) is
square and symmetric so that V( )=V( )′, but not every square symmetric matrix is a VC matrix. To qualify,
all variances and correlations must satisfy σ > 0 and -1< ρ < 1. The values must also make statistical sense,
for example the following do not: ρ ≅0.95, ρ ≅0.95, and ρ ≅-0.95. The strong positive correlation of X1
with both X2 and X3 implies that X2 and X3 should also have a strong positive correlation. It turns out that all
conditions for a square symmetric matrix to be a valid VC matrix are met if it is positive-definite (+definite),
that is,
V
> 0 for any constant vector
=(a1,a2,...,aN) ≠ (Wothke, 1993). Since V(
)= V
, the
condition simply means that any non-zero linear combination of the RVs Xi, i=1,2,...,N is an RV with
variance > 0. A matrix is +definite if all eigenvalues are > 0. The eigenvalues of V( ) are the constants λi that
satisfy |V( ) - λi | = 0, i=1,2,...,N (Meyer, 2000). Since V( ) - λi has determinant = 0 it is singular which
implies the equation (V( ) - λi )·ui =
can be solved by a ui
, thus λi and ui come in pairs. The ui with
length = 1 is referred to as the eigenvector for λi. Note that V( )·ui = λiui so that ui′V( )ui = λiui′ui = λi, which
reveals why each λi must be > 0 for V( ) to be +definite (otherwise Z=ui′ has V(Z) < 0). Note that the
determinant of a matrix is the product of its eigenvalues, |V( )| = ∏
λ , and since each λi > 0, |V( )| > 0
which ensures V( )-1 exists (Guttman, 1982). A matrix with determinant 0 thus cannot be +definite. Also,
a correlation of ρ =1 (or ρ =-1) is not allowed in a VC matrix. If two RVs Xi and Xj are perfectly correlated
we can question why both are needed but beyond that V( ) cannot be +definite. Set ai=1, aj=-σi/σj and all
other elements in =(a1,...,aN)′ to 0 and note that V(Z) = V
a = 0 where Z =
, since V(Z) = 12 σ +
(σi/σj)2σ – 2(σi/σj)ρ σ σ = 2σ - 2ρ σ = 0, with ρ =1. (When ρ =-1 use constants ai=1 and aj=σi/σj.)
12
I.1 Repairing a Broken VC Matrix
A VC matrix that is not +definite is said to be broken and can occur for a variety of reasons including
missing data, ad-hoc estimation procedures, and iterative optimization methods. If encountered we can end
the analysis with an error or repair the broken VC matrix and continue. We take the latter approach and
perform a ridge repair (Wothke, 1993). A ridge is added to V( ) by multiplying the diagonal by a constant
K > 1. Start with K=1+ε and increase ε until the modified matrix, say VR( ) with diagonals Kσ , is +definite.
The entire matrix is then divided by K, which revert the diagonals back to σ and forces the covariances to
ρ /K σ σ , which approach 0 (along with the correlations) as K increases (a diagonal matrix with elements
>0 is +definite). The scaled matrix will be +definite since
> 0 when
1/K V
V
> 0, and K > 0.
I.2 Useful Derivatives for VC Matrices
Let
t = (X1t, ..., XNt)′
be compounding returns on N financial assets at times t=1,...,T where
t ~
f(xt,θ)
~ N(μ, V( )), with θ=(μ, V( )), μ=(μ1,...,μN)′ and μi=E(Xit). The terms in V( ) from (2.I.1) are unknown
parameters that can be estimated as MLEs after collecting data (see §II.E). MLEs maximize (θ|x1, ..., xT),
which is the multivariate PDF of the data. The multivariate normal PDF for
,
, ,V
/
2
1
|V
The multivariate PDF for the entire iid RV sample ( 1, ...
,…,
,
,…,
, ,V
e
T)
is:
/
2
is (Guttman, 1982):
’
/
|
t
1
|V
,
|
/
∈
(2.I.2)
’
e
(2.I.3)
The log-likelihood function for the unknown parameters θ=(μ, V( )) is given by:
ln
,V
|
T∙N
ln 2
2
,…,
T
∙ ln |V
2
|
1
2
(2.I.4)
’V
The MLE for θ = (μ, V( )) is found by maximizing (2.I.4) wrt θ. A critical point is where the vector of 1st
derivatives equals
which would reveal the maximum or it could be found iteratively using a gradient
method also requiring said derivatives.15 Let Aij be an NxN matrix having a 1 in the i-j and j-i positions and a
0 in all other positions, and let Vij( ) be the VC matrix from (2.I.1) with a 0 in the i-j and j-i positions. It
follows that V
=σ
and via the product rule V
V
and
V
. By definition, V
V
V
V
= 0 so that
V
, thus
V
V
et al., 1992). Denote σ =σ and the above holds for all elements. The determinant of V
V
V
V
(Searle,
can be expressed
using cofactor expansion with respect to the i-th row, i=1,...,N, as (Meyer, 2000):
15
As noted in §II.I.1, iterative optimization methods are a source of broken VC matrices as we may step into an infeasible region.
13
|V
where Vij
is V
|
|
σ |V
∑
1
σ V
,
with the i-th row and j-th column removed, thus
matrix with aij=aij(t). The determinant is a function of the aij, say, | |
chain rule (N N dimensions),
|V
|
|V
and
|V
|
|V
for i≠j,
|
1
|
2
|V
∑
V
∑
∙
|V
1, … , N,
|
2V
is symmetric, V
, where V
(2.I.5)
|. Let
|V
a ,…,a ,…,a
t (Anton, 1988). If =V
|, with aij=aji=aij(σ )=aji(σ )=σ . Using (2.I.5),
. Since V
V
1
| |
for i
={aij} be a
. Via the
,…,a
and t=σ , i≠j, then
|V
=V
|
1
′=V
V
′=V
, so that V
and
is V
with a 1 in position i-j and a 0 in all
other row i and column j positions (Searle, et al., 1992).16
J. Serial Correlation
Let Xt, t=1,...,T be RVs for the compounding return on a financial asset at time t. If the unconditional
mean and variance are constants let E(Xt)=μ and V(Xt)=σ2 ∀ t, respectively. When σij=Cov(Xi,Xj) ≠ 0 for i≠j,
the returns are serially correlated and cannot be assumed iid. Serially correlated data are modeled using a
time series model such as the autoregressive (AR) or moving average (MA) process. The AR model of order
p is Zt = ϕ1Zt-1+...+ ϕpZt-p + εt, and the MA model of order q is Zt = εt – θ1εt-1 –... – θqεt-q, where Zt = Xt – μ and
εt
~
N 0, σ
∀ t (Box, et al., 1994). The autoregressive moving average (ARMA) model includes AR(p) and
MA(q) terms. The appropriate model for a data set is identified by the signature of the autocorrelation (ACF)
and partial autocorrelation (PACF) functions. The ACF at lag k is ρt,t-k = Corr(Xt,Xt-k) and the PACF at lag k
is the correlation remaining after accounting for all lag < k correlations, namely ϕ in an AR(p) model. The
classic AR(p) signature has the PACF cut off abruptly and the ACF decay after lag p, with the 1st PACF
being > 0. The classic MA(q) signature has the ACF cut off abruptly but the PACF decay after lag q, with
the 1st PACF being < 0. One of these patterns often emerges (perhaps after log or power transforms) and
usually max(p,q) ≤ 3 (Nau, 2014). A series with fixed μ, σ2 and σ is said to be stationary. A time series that
drifts either up or down over time has a trend and is not stationary since E(Xt) is not fixed. A non-stationary
series can often be made stationary by differencing (Box, et al., 1994). Usually 1 or 2 differences will suffice
where Dt = Zt – Zt-1 for t=2,...,T and D2t = Dt – Dt-1 for t=3,...,T are the 1st & 2nd differences. When Dt ~ AR(0), Zt
= Zt-1 + εt is referred to as a random walk (RW). Each observation in a RW is the prior value plus a random
step governed by εt. An alternative to differencing for a trend is to fit a regression with time as a predictor
then model the stationary residuals. Define the backshift operator B as BkZt = Zt-k so that the AR(p) model
can be written as Zt = ϕ1BZt +...+ ϕpBpZt + εt or ϕ(B)Zt = εt, where ϕ(B) = (1-ϕ1B-...- ϕpBp) is the characteristic
16
The notation V
is ours and introduced since dealing only with NxN matrices simplifies the code. A new NxN matrix V
will
nd
with a 1 in position r-s and 0 in all other row r and column s positions.
be introduced (§V.A.2) when taking 2 derivatives which is V
14
polynomial. An AR(p) process is stationary if all roots of ϕ(B) are >1 in magnitude (Box, et al., 1994). The
MA(q) model is stationary by design since E(Zt) = 0 → E(Xt)=μ, σ2=V(Zt) = σ (1 + θ12 +... + θq2), and σt,t-k =
Cov(Zt,Zt-k) = σ
θ
∑
are the fixed μ, σ2 and σt,t-k for k ≤ q. The covariance was derived using
θθ
the model’s definition as σt,t-k = Cov(-θkεt-k, εt-k) + Cov(θk+1εt-(k+1), θ1εt-(k+1)) + ... + Cov(θqεt-q,θq-kεt-q) = -θkV(εt-k) +
θk+1θ1V(εt-(k+1)) + ... + θqθq-kV(εt-q) for k ≤ q and 0 for k > q. The correlation is ρt,t-k = Corr(Zt,Zt-k) = σt,t-k/σ2 =
θ
∑
θθ
/ 1
∑
θ .
Consider the AR(1) process, Zt = ϕ1Zt-1+ εt. Since Zt-1 = ϕ1Zt-2+ εt-1 and Zt-2 = ϕ1Zt-3+ εt-2, each centered
observation can be rewritten as Zt = ϕ1(ϕ1(ϕ1Zt-3+ εt-2)+ εt-1) + εt = ϕ13Zt-3 + ϕ12εt-2 + ϕ1εt-1 + εt, and repeating without
end yields Zt = ϕ1∞Zt-∞ + ∑
ϕ ε
. The AR(1) model thus only works if |ϕ1| ≤ 1, otherwise ϕ1∞Zt-∞ is infinite
and E(Zt), V(Zt) are not defined. Further, if |ϕ1| < 1, ϕ1∞Zt-∞ → 0 and an alternate form for the AR(1) model is
Zt = ∑
ϕ ε
, with E(Zt)=0, E(Xt)=μ. The AR(1) model has characteristic polynomial ϕ(B) = (1-ϕ1B) with
root B = 1/ϕ1 and is hence stationary when |1/ϕ1|>1 or |ϕ1| <1, justifying the condition noted for a fixed E(Zt)
and V(Zt). The conditional variance of a new observation, Xt, given all prior observations is V(Xt|Xt-1, Xt-2, ...)
= V(Zt|Zt-1, Zt-2, ...) = V(εt) = σ , by definition of the AR(1) model. Using the alternative AR(1) model form, the
unconditional variance of each Xt, given |ϕ1| <1, is σ2=V(Zt)=∑
ϕ V ε
=σ ∑
ϕ =σ / 1
ϕ .17 The
AR(1) model does not imply that each new value Xt depends only on the prior value Xt-1. In fact, it assumes
a non-zero correlation between Xt and all prior observations, Xt-1, Xt-2, ..., which is a reason why p>3 is rarely
needed in practice. To see this use the AR(1) alternative form and note that σt,t-k = Cov(Zt,Zt-k) = Cov(ϕ1Kεt-k,
εt-k) + Cov(ϕ1K+1εt-(k+1), ϕ11εt-(k+1)) +... = ϕ1KV(εt-k) + ϕ1K+2V(εt-(k+1)) +... = σ (ϕ1K + ϕ1K+2 +...) = σ ϕ1K/(1 - ϕ12), since
the εt are iid. The correlation between values k time points apart is Corr(Xt,Xt-k) = Cov(Xt,Xt-k)/ V X V X
= ϕ1K (exponential decline in k). The RW is an AR(1) model with unit root, i.e., ϕ1=1, having alternative
form Zt = Zt-k + ∑
ε
. The unconditional variance is V(Zt) = V(Zt-k) + kσ and since V(Zt) > V(Zt-k) the RW is
not stationary (V(Zt) → ∞ as t → ∞). Parameters in an AR(1) model can be estimated by least squares (with
adjustments to account for serial correlation), maximum likelihood, or the Yule-Walker equations (method of
moments) (Box, et al., 1994). The likelihood function of an observed sample is the multivariate PDF of the
data, written as a function of the parameters ϕ1 and σ . Using Zt, this is the likelihood from (2.I.4) with μ = 0,
and V( )=σ / 1
ϕ
is the VC matrix with
having diagonals =1 and i-j off-diagonals ϕ1|j-i| for i ≠ j.
A time series mean reverts if |E(Zt+k|Zt, Zt-1, ...)| → 0 and V(Zt+k) < ∞ as k→∞. In words, future centered
values, Zt+k, come from a PDF with fixed σ2 and μ → 0 as k increases. Stationary time series thus mean revert
due to their E(Zt)=0 and fixed V(Zt), but RWs do not. Mean reversion strength is measured by speed or halflife. This is the k required for E(Zt+k|Zt, Zt-1, ...) = zt/2, i.e., the time before the process’s conditional mean
17
Let S=ϕ10+ϕ12+ϕ14+..., so that ϕ12S=ϕ12+ϕ14+ϕ16..., and S ϕ12S=ϕ10, resulting in S=ϕ10/(1-ϕ12)=1/(1-ϕ12), iff |ϕ1| < 1, otherwise S=∞.
15
equals ½ of the last observed value Zt=zt. In an AR(1) process, Zt+k = ϕ1kZt + ∑
so that E(Zt+k|Zt=zt)
ϕ ε
=ϕ1kzt, and the half-life is k such that ϕ1kzt = zt/2, or k = ln(0.5) / ln(|ϕ1|) (Tsay, 2011). Mean reversion speed
thus strengthens as ϕ1→0 and weakens as |ϕ1|→1, which is intuitive since ϕ1 is the damping factor applied to
prior values. The fastest mean reverting AR(1) process has ϕ1= 0, which is a random sample without serial
correlation. The half-life of k=0 implies the process generates Zt=zt then instantly reverts to a mean 0, finite
σ2 PDF. As ϕ1→1 the AR(1) process approaches a RW with half-life k=∞, and does not mean revert. Thus
in an AR(1) process as serial correlation strengthens mean reversion weakens, and vice versa. This is
intuitive since as mean reversion strengthens (i.e., ϕ1→0) new values increasingly depend on the mean, but as
serial correlation strengthens (i.e., |ϕ1|→1) new values increasingly depend on the prior value.
Data from 18 time series have been retrieved and tested for serial correlation (see Table I, Appendix
B). Conclusions are draft and subject to relevant diagnostics (Box, et al., 1994). The correct significance
level, α = P(Type I Error), for testing multiple simultaneous hypotheses to control the family-wise error rate,
α* = P(≥ 1 Type I Error), is discussed below. Market efficiency is the default and each test is formed as Ho:
No Serial Correlation vs. Ha: Serial Correlation. Evidence is needed to reject market efficiency for liquid
assets as it may suggest a profitable arbitrage trade for retirement advisors who can predict the path of future
prices. A Type 1 Error occurs if we reject Ho when it is true, i.e., we falsely reject efficiency for a security.
Table I. Tests for Serial Correlation in 18 Finance/Economics Time Series
Data Returns (Annual)
CPI-U (Inflation Rate)
Real S&P 500
Total S&P 500
Real Small Cap Equity
Total Small Cap Equity
Real U.S. 10-Year T-Bond
Total U.S. 10-Year T-Bond
Real U.S. 3-Month T-Bill
Real Gold Return
P-Value1
<0.00001
0.90219
0.96352
0.77708
0.58475
0.58194
0.47060
<0.00001
0.00430
Process
AR(3)
RS
RS
RS
RS
RS
RS2
AR(3)
RS3
Data Returns (Annual)
Cash (No Interest)
Real S&P 500-to-Bond RP
Real S&P 500-to-Small Cap RP
Real Small Cap-to-Bond RP
10-Year Avg. Real S&P 500
10-Year Avg. Total S&P 500
1st Diff. Shiller CAPE Ratio
1st Diff. Log Shiller CAPE Ratio
S&P 10-Year Avg. Real Earnings
P-Value1
<0.00001
0.71082
0.00359
0.80516
<0.00001
<0.00001
0.39993
0.96966
<0.00001
Process
AR(3)
RS
RS3
RS
AR(1)4
AR(1)4
RW5,6
GRW5,7
ARMA(2,1)4,8
Abbreviations: RS = random sample, (G)RW = (geometric) random walk w/Drift, RP = risk premium.
1
P-Value is for test Ho: No Serial Correlation vs. Ha: Serial Correlation, not for test Ho: AR(p) vs Ha: AR(p-1).
2
3
Non-normal. Non-normal, possible serial correlation.
4
Near unit roots by design. Let Xt ~ f(x), t=1,...,T, with E(Xt)=μ, V(Xt)=σ . Avg. 10-year returns are Yt =(Xt+...+Xt+9)/10, for t=1,...,T-9.
2
If f(x) ~ N(μ,σ2), Yt+1=Yt – Xt + Xt+10 = Yt + (Xt+10 - Xt) = Yt +εt , εt ~N(0,√ )≠RW as Cov(Yt,εt)=
σ . If f(x) ~ Ln(μ,σ ), ln(Xt) is similar.
5
6,7
Possible non-normal “steps”.
Dickey-Fuller unit root test statistics -2.067, -2.187, respectively, with critical value -2.900 cannot reject
hypothesis of unit root. Note: Since CAPE Ratio ≥ 0, GRW may be more plausible. Non-zero drifts are 0.0885 (RW) & 0.0040 (GRW).
8
Preliminary model would be ARMA(2,1) on detrended data: Yt – [μ + βt] = ARMA(2,1) → Yt = μ + βt + ϕ1Yt-1 + ϕ2Yt-2 - θ1εt-1 + εt.
The goal is to test these hypotheses such that the family of conclusions is replicable with (1-α*) %
confidence, where α* = P(≥ 1 Type I Error). In general, with N independent tests and α = P(Type I Error) for
each test, the probability of making k (≤ N) Type I Errors is P(B = k) =
N
α 1
α
with B ~ Binomial(N, α).
Thus the probability of making ≥ 1 Type I Errors in N independent tests is P(B ≥ 1) = ∑
16
N
α 1
α
= 1–
P(B = 0) = 1 – (1–α)N. For the N=18 tests conducted above with α=0.05 the probability of making ≥ 1 Type I
Errors is ∑
18
0.05 0.95
= 1-P(B = 0) = 1-0.9518 ≅ 0.60. Being 95% confident on each of 18 independent
tests translates into a 40% chance of replicating the family of conclusions with new data. To replicate the
family with 95% confidence, i.e., α* = P(≥ 1 Type I Error) = 0.05, we adjust the α used for each test. The
Bonferonni adjustment uses α = α*/N and does not require independent tests but does ensure P(≥ 1 Type I
Error) ≤ α* (Westfall, et al., 1999). With α* = 0.05, the adjusted p-value for each test in Table I is α =
0.002778. The conclusions in Table I lead to the Table II concerns about current retirement finance dogma.
Table II. Questionable Claims in the Retirement Finance Literature
Claim
Shiller’s CAPE Ratio (annual) can be used to
time markets. Investors/retirees should sell
when the CAPE ratio is high and buy when it
is low relative to its historical average.
Concern1
The hypothesis that the annual CAPE ratio behaves as a (G)RW
cannot be rejected. Random walks have unit roots and do not
mean revert. The best predictor of any future value in a
random walk is the current value + drift (use logs for GRW).
Annual S&P 500 Returns (real or total) mean
revert therefore are serially correlated and
should be fit using an autoregressive model.
Serial correlation and mean reversion are opposites (i.e., as one
strengthens the other weakens). Annual S&P 500 returns exhibit
no serial correlation, they are random samples and mean revert.
Shiller’s CAPE Ratio (annual) can be used to
predict average future long-term returns.3 A
(log-transformed) linear regression predicting
10-year average future S&P 500 returns using
CAPE values has a highly significant R2.
The 10-year average S&P 500 return is strongly serial
correlated (see Table I and Appendix B).2 Fitting a regression
line through these points is inappropriate and will result in
underestimated variances and inflated Type I Error rates which
commonly lead to false claims that predictors are significant.
1
Findings are preliminary. The fitted (linear) models are subject to appropriate diagnostics (see, Box, et al., 1994).
This is by design. Let Xt ~ f(x), t=1,...,T, with E(Xt)=μ, V(Xt)=σ . Average 10-year returns are constructed as Yt = (Xt+...+Xt+9)/10, for
t=1,...,T-9, so that Cov(Yt,Yt+k) = (1/10)2[V(Xt+k) +... + V(Xt+9)] = [(9-k+1)/100]·σ , for 0 ≤ k ≤ 9 and Cov(Yt,Yt+k) = 0, for k ≥ 10.
3
A similar claim is made about safe withdrawal rates (SWR) in retirement via regression with CAPE, and the exact same concern arises.
2
K. Constrained Optimization
K.1 Linear Programming
A linear program (LP) is an optimization problem where the objective and all constraints are linear
functions of the decision variables. LPs are solved using the simplex algorithm and the standard form for an
LP with n decision variables and k linearly independent constraints is (Jensen & Bard, 2003):
Maximize:
Subject to:
Z = c1x1 + c2x2 + ... + cnxn
a11x1 + a12x2 + ... + a1nxn ≤ b1
a21x1 + a22x2 + ... + a2nxn ≤ b2
′
(2.K.1)
(feasible region)
⋮
Where:
(objective function)
ak1x1 + ak2x2 + ... + aknxn ≤ bk
xi ≥ 0, i=1,2,...,n
(non-negativity constraints)
All LPs come in pairs with the dual being an equivalent minimization problem. When the primary LP is
solved the dual is solved and vice-versa. Whereas the primary LP has n decision variables and k constraints,
17
the dual has k decision variables and n constraints and since min(Z) ≡ max(-Z), any LP can be solved assuming
k ≤ n. The simplex algorithm recognizes that when Z is linear, global solutions must occur at corner points of
the feasible region.18 A constraint binds when ‘≤’ becomes ‘=’ for given values of the decision variables.
When m constraints bind, m decision variables are fixed by the constraints and there are
k
m
ways to select the
these constraints, m=0, ..., k. The remaining n-m decision variables must equal 0 at a corner point, and there
are
n
n
m
ways this can occur. The total # of corner points is thus ∑
solution. The simplex algorithm partitions A as
k
m
n
n
m
=
n
k
k
, with each a potential
such that A11 is mxm and of full rank and let c′ =
[c1|c2], x′ = [x1|x2], b′ = [b1|b2] be the corresponding vector partitions. It follows that A11·x1 +A12·x2 = b1 and x1
= A11-1·b1 - A11-1·A12·x2 reflects the m decision variables fixed by the binding constraints. The objective Z
becomes c1·x1 + c2·x2 = c1·A11-1·b1– (c1·A11-1·A12 – c2)·x2 which is a constant less a linear combination of the
decision variables in x2. If any coefficients in this linear combination are negative the problem is unbounded
and has no solution. Simply increase the corresponding decision variable in x2 to increase the objective Z.
To have a solution all coefficients for x2 in Z must be ≤ 0 and when this occurs, all decision variables in x2
must equal 0 to maximize Z. A constant less a quantity ≥ 0 is maximized when that quantity equals 0.
Finally, when x2 = 0, x1 = A11-1·b1 and if all constraints are satisfied then x′ = [x1|x2] is a basic feasible solution
(BFS) to the LP (Jensen & Bard, 2003). The simplex algorithm starts with a corner point of the feasible
region and cycles through adjacent corner points (i.e., bases defined by
) such that Z does not decrease.
Thus the LP can be solved with a small number of evaluations and the algorithm ends at a global maximizer.
Note that when k ≤ n both A21 and A22 vanish so that each basis is formed by setting n-m decision variables
equal to 0 and solving for the remaining variables.
If any quantity is random then (2.K.1) is a stochastic linear program (SLP). While solving SLPs
using theory is involved (Kall & Mayer, 2010), simulation can be a practical alternative. For example
suppose b = (b1,..., bk)′ is a set of RVs with b ~ fB(b), E(b)=μB.19 Since any solution is a function of b it must
also be an RV, say x ~ fX(x), E(x)=μX. When fB(b) is known/approximated a heuristic SLP solution is obtained
by generating random values for b, say bi = (b1i, ..., bki), where sample i yields an LP BFS, say xi = (x1i,..., xni)′,
i=1,...,N. The solution is then taken as
and as N→∞,
or E
E
=
∑
. Since each xi satisfies
it follows that
̅
. Simulated solutions therefore asymptotically satisfy the constraint
set, in expectation. An exceedingly large number of real-life problems can be formulated and solved as LPs
and some non-linear programs can also be closely approximated by an LP.
18
The solution to an LP may or may not be unique, but it is global. The feasible region of an LP is referred to as a polyhedron which sits
in the 1st quadrant since all decision variables are ≥ 0. A technically incorrect but useful visualization tool when n=2 is that four people of
different heights are holding a flat board (objective function is a plane) over a stop sign laying flat on the ground (feasible region in 1st
quadrant). The highest point on the board inside the sign will be directly above a corner of the sign, it cannot be above an interior point.
19
Here, b may reflect any randomly occurring quantity such as supply, demand, temperature, sales revenue, profit, etc...
18
K.2 Quadratic Programming
When the objective function in (2.K.1) is of the form Z =∑
ax
∑
∑
∑
b xx
c x , the
surface being maximized is quadratic (not linear) and the resulting optimization is referred to as a quadratic
program (QP). If bij = 0 ∀ i < j=1,...,n, Z is said to be separable as the objective separates into a sum of 1decision variable functions, i.e., Z = ∑
x
(Hillier & Lieberman, 2010).
A separable QP can be
approximated by an LP. To begin, convert the QP to a minimization problem noting that max(Z) ≡ min(–Z),
then write each function in –Z as
x
x
posing no issue since minimizing -Z and -Z + ∑
c x
a ⁄2c
which adds the constant ∑
a ⁄4c to -Z
a ⁄4c are equivalent problems. Proceed by partitioning
each xi-axis into equidistant constants [xi1,..., xiS+1] which define S line segments that trace out
x . These
S+1 values are chosen and allow xi to be replaced by a weight vector
= (αi1, ..., αiS+1)′, where ∑
Any xi1 ≤ xi ≤ xiS+1 is reachable by a weighted sum of the constants using
, namely, xi = ∑
function
x ≅∑
α
α = 1.
, and each
where the approximation sharpens as S increases when adjacent weights are
α
used (Jensen & Bard, 2003), see Figure III. Since the objective is min(–Z), adjacent weights must be used in
an optimal solution, compare the blue dot objectives for the dashed and red lines in Figure III.
Figure III. Separable QP Approximation by an LP (S = 4) (Minimization Objective)
The QP max(Z), Z = ∑
Minimize:
-Z ≅ ∑
Subject to:
am1∑
∑
Where:
∑
ax
∑
α
c x , is then approximated by the following LP (linear in ):20
′
α
+ ... + amn∑
α ≤ 1, -∑
α
≤ bm,
α ≤ -1
m=1,2,...,k
i=1,2,...,n
αij ≥ 0, i=1,2,...,n & j=1,2,...,S+1
(objective function)
(feasible region)
(2.K.2)
(non-neg. constraints)
20
The standard form LP in (2.K.1) requires all constraints to have the form fm(x) ≤ bm, m=1,2,...,k where fm(x) is linear in x. Express an
equality constraint fm(x) = bm using only “≤” as (fm(x) ≤ bm ∩ fm(x) ≥ bm) ≡(fm(x) ≤ bm ∩ -fm(x) ≤ -bm).
19
K.3 Classical Convex Programming
A classical convex program (CCP) is a non-linear optimization having objective of either minimizing
a convex, or maximizing a concave function over a set of linear equality constraints (Jensen & Bard, 2003).
A function f( ) is convex iff f(α 1 + (1-α) 2) ≤ αf( 1) + (1-α)f( 2) ∀
1,
2
∈
n
, and if f( ) is convex then -f( ) is
concave. Further, max[-f( )] ≡ max[e-f(x)] thus maximizing a concave function and a log-concave function
are equivalent problems as e-f(x) is log-concave since ln(e-f(x)) = -f( ) is concave (Lovasz & Vempala, 2006).
A CCP has the desirable property that local optimums are global optimums, thus the problem reduces to
finding any local optimum. The following formulation is of interest:
f(x1, x2, ..., xn)
f(x1, x2, ..., xn)
Minimize:
(or) Maximize:
Subject to:
(convex objective function)
(concave objective function)
a11x1 + a12x2 + ... + a1nxn = b1
a21x1 + a22x2 + ... + a2nxn = b2
(2.K.3)
(feasible region)
⋮
ak1x1 + ak2x2 + ... + aknxn = bk
xi ≥ 0, i=1,2,...,n
Where:
(non-negativity constraints)
As with LPs, redundant constraints are removed. If k > n the feasible region is empty and there is no solution.
If k = n the feasible region consists of the point =
∗
such that
f
∗
= 0.
If 0 < k < n,
∗
f
, which is also the solution. If k = 0 the solution is
∗
= 0 and
=
then the unconstrained solution solves the
constrained problem. Most likely none of the above will hold, and we are left to optimize
with 0 < k < n,
= , and rank( ) = k. To solve this problem in 1-step we locate the critical point of the Lagrangian, L(·),
-∑
defined as, L(x1,..., xn, λ1,...,λk)=L( , )=
incorporates all constraints into
λ a x
a x
⋯
a x
b .21 The Lagrangian, L( , ),
and k new decision variables are introduced, λi, i=1,2,...,k, called the
Lagrange multipliers. The solution occurs at L ( , ) = 0, namely:
∑
λa
∑
λa
⋯
a x
b
⋯
a x
b
⋮
L( , ) =
a x
a x
a x
a x
⋮
0
⋮
0
.
0
⋮
0
(2.K.4)
This non-linear system can be solved using Newton’s method which approximates L ( , ) linearly in the
neighborhood of
solving yields
21
as L ( , ) ≅ L ( ,
=
-
L( ,
)+
)]-1 L ( ,
L( ,
)
. Setting the approximation = 0 and
). Repeating the process generates the iterative solution
A multi-step solution would write
= as
= with
k k and rank(
in the objective,
, making it an unconstrained function of only. After solving
20
) = k then solve =
( and replace
∗
= 0 we use the above to determine ∗ .
from
derivatives
and convergence occurs when | L ( ∗ ,
∗
)| < ε. The (n+k) (n+k) symmetric matrix of 2nd
L ( , ) is called the bordered Hessian which is given by:
⋯
⋮
⋱
⋯
L( , )
a
⋮
a
where
⋮
⋯
⋱
⋯
a
⋮
a
= implies
implies = , thus
removed,
= and
= implies
a
⋮
⋱
⋮
a
⋯
a
0
⋮
0
⋯
⋱
⋯
0
⋮
0
= . Since f( ) is convex,
= . (Same for -f( ) concave,
is n k with rank( )=k and
L( , )
= thus
⋯
(2.K.5)
,
L ( , )]-1 thus (2.K.5) must be invertible. Note that
is the Hessian of f( ). Newton’s method uses [
L( , )
a
is +definite and
′
=
–definite.) Since redundant constraints have been
= implies =
.
Therefore when f( ) is convex [-f( ) is concave],
= and the bordered Hessian is full-rank thus invertible (Border, 2013).
K.4 General Non-Linear Programming
A general non-linear program (NLP) seeks to minimize or maximize a smooth function
subject to g
≤ bi, i=1,2,...,k, where
is not necessarily convex or concave and g
,
∈
n
,
is generic. Such
problems can have several local optimums and the goal is to find the best among these. Little can be said
about NLP problems in general and the optimization strategy depends on the nature of the problem. In some
cases the Lagrangian can be used to find local optimums. In others a metaheuristic, such as tabu search,
simulated annealing, or a genetic algorithm can be used (Hillier & Lieberman, 2010). If all else fails we can
generate random values and evaluate
∙ when the constraint set is satisfied, keeping a record of the optimal
value. A better approach would generate random values that satisfy all constraints. Alternatively, we can
take a random setting that is infeasible and project it to a point inside the feasible region then evaluate
22
that point.
∙ at
Random starts can be effective when many local optimums exist and strategies for generating
values have been developed for specific problems such as mixture likelihoods (see McLachlan & Peel, 2000).
L. Copula Modeling
Let X be any continuous RV with PDF f(x) and CDF F(x) = P(X ≤ x) =
. Copula modeling is
based on a fact that initially surprises, but is intuitive upon reflection. Namely that F(X) ~ uniform(0,1). The
proof is straightforward, let U = F(X) then FU(u) = P(U≤ u) = P(F(X)≤ u) = P(X ≤ F-1(u)) = F(F-1(u)) = u, for 0 ≤ u ≤ 1.
Random variables having the same CDF are identically distributed and the CDF for U is from a uniform(0,1)
22
Consider maximizing a generic likelihood function that includes the VC matrix V
from (2.I.1). A constraint on the variances and
covariances is that V
must be +definite. An alternative to discarding any point with a broken V
matrix is to repair it.
21
distribution. Let = (X1, ..., XN)′ be RVs for the compounding return on N financial securities at a given time
point. The marginal PDF and CDF of Xi are fi(xi) and Fi(xi) respectively, and the multivariate PDF and CDF
are f( ) and F( ) = P( ≤ ) = F(x1, ..., xN) = P(X1 ≤ x1 ∩ ... ∩ XN≤ xN) respectively, see §II.A. As above, let Ui
of
= Fi(Xi) where Ui ~ uniform(0,1) and Xi = Fi-1(Ui). The CDF of = (U1, ..., UN)′ is G( ) =P( ≤ ) = G(u1, ..., uN) =
P(U1 ≤ u1 ∩ ... ∩ UN ≤ uN). Since G( ) is a valid CDF, its derivative is the multivariate PDF of , namely
…
G
= g( ) - g( ) = g( ), see §II.A. Note the relationship between F( ) and G( ) (Nelson, 2006):
F
F(x1, ..., xN)
The multivariate PDF of
= P(X1 ≤ x1 ∩ ... ∩ XN ≤ xN)
(2.L.1)
= P(F1-1(U1) ≤ x1 ∩ ... ∩ FN-1(UN) ≤ xN)
(2.L.2)
= P(U1 ≤ F1(x1) ∩ ... ∩ UN ≤ FN(xN))
(2.L.3)
= G(F1(x1), ..., FN(xN)).
(2.L.4)
can then be derived by differentiating F(x1, ..., xN), using the chain rule on Fi(xi) as:
f( ) = f (x1, ..., xN)
=
=
…
…
F
(2.L.5)
[G(F1(x1), ..., FN(xN))]
…
g F
Multivariate PDF
g
,…,F
Multivariate Copula PDF
(2.L.6)
[from (2.L.4)]
,…,
(2.L.7)
…
…
(2.L.8)
Product of Marginal PDFs
The literature refers to G(·) as the copula and g(·) as the copula density. The term indicates a coupling
between the multivariate and marginal PDFs for a set of RVs (Nelson, 2006). When X1,...,XN are independent
U1, ..., UN are also independent and g( ) = 1 ... 1 = 1 so that f (x1, ..., xN) =
...
N
N
. The copula term
therefore models the dependence between a set of RVs, with the breakthrough being that marginal PDFs can
be modeled separately. This quality is appealing, particularly for the field of retirement finance. When
modeling the multivariate PDF for a set of real compounding returns, the marginal return on a given security
should not depend on which other securities are involved. For this reason copula modeling is a standard for
multivariate PDF modeling in finance and many forms have been proposed. For example if g(·) is Gaussianinduced then the copula will model dependence after mapping to normal RVs.
Since the unknown
parameters in a copula exist in the likelihood they can be estimated as MLEs. When building a multivariate
PDF we can choose between candidates copulas by taking the one with smallest root-mean-square error
using the empirical copula, which is constructed via the empirical CDF. In this research we propose a
generic fixed-marginal tractable alternative to copula modeling that is well suited to retirement finance.
22
Assuming marginal PDFs f1(x1), ..., fn(xn) have been arbitrarily fit to the data set, a distribution with
corresponding CDF H(·) is selected to model the dependence structure of RVs X1, ..., Xn. For example,
during the housing boom securitized mortgage products modeled default times of named residential
borrowers as exponential RVs, see §I. Here, fi(xi) would be an exponential PDF for i=1, ..., n. Further, the
dependence structure was chosen as Gaussian so that H(·) becomes Φ(·). The multivariate PDF in (2.L.8)
requires g(·) for full specification, and we say that it is induced by the choice of H(·) using transformations
U1 = H1(Y1), ..., Un = Hn(Yn), where Hj(·) is the corresponding marginal CDF for element j. Following the
standard procedure for transforming RVs (Freund, 1992), the multivariate PDF g(·) has form:
…
g
H
,…,
,…,H
∙
⋮
⋱
,
⋮
∈ 0,1
(2.L.9)
…
where h(·) = H′(·) is the corresponding PDF for our chosen dependence structure. All off-diagonals in the
Jacobian term are zero since each transformation involves only one RV. The diagonal terms are derived for
i=1, ..., n as follows:
(2.L.10)
Using rules from calculus, we can treat
as a ratio when Hi(·) is invertible, thus:
1
1
(2.L.11)
H
The copula density g(·) thus takes form of (2.L.12) when induced by the dependence structure of CDF H(·):
g
,…,
H
,…,H
∙
1
H
,
∈ 0,1
(2.L.12)
When H(·) is Gaussian, h(·) = ϕ(·) and hi(·) = Hi′(·) are univariate standard normal PDFs. The copula density
g(·) from (2.L.12) is used in (2.L.8) to complete the multivariate PDF of our sample data, f (x1, ..., xN). Note
that Sklar’s Theorem guarantees the PDF in (2.L.8) has marginals fi(xi) when using g(·) (Nelson, 2006).
Copula parameters can then be estimated as MLEs with only covariance terms unknown, since it is assumed
that the univariate marginals are fully specified.
Tran, et al., (2013) caution that a straight forward
optimization of (2.L.8) may work for basic copula forms, but can fail when using commercial software and a
complicated dependence structure. Giordoni, et al., (2009) solve a problem similar to the one we address,
but in different ways. Namely, using a normal mixture copula along with normal mixture marginals, and
also with a non-copula adaptive estimator that attempts to smooth out the differences between normal
mixture marginals and the implied marginals of a multivariate normal mixture fit directly to the data.
23
M. Information Criteria
In addition to the LRT from §II.H, fitted models in statistics can be compared using a metric called
the information criteria (IC). Such values quantify the information lost by a model vs. the true state of
nature. Smaller IC values are preferred as they indicate less information loss. Models with too many
parameters are said to lack parsimony. When choosing amongst candidate models, IC metrics attempt to
strike a balance between the likelihood value and # of parameters. Among the most widely used IC metrics
is Akaike’s information criteria (AIC), calculated as (Akaike, 1974):
(2.M.1)
AIC = 2 (# Parameters) - 2 ln( ( | ))
The parameters counted must be free. A model with π1 and π2 constrained linearly by π1 + π2 = 1 only counts
as 1 parameter since estimating π1 estimates π2. Thus, (# parameters) = (total # parameters) – (# independent
constraints). The term ln( ( | )) is the log-likelihood using the MLE, see §II.E. The AIC works well when
interest is in controlling Type I Errors in subsequent hypothesis tests (Tao, et al., 2002), but it tends to over
fit in small samples leading to models that lack parsimony. Thus the corrected AIC (AICC) has been
proposed and includes a penalty that increases with the # of parameters (Hurvich & Tsai, 1989):
2
# Parameters
Sample Size
1
# Parameters
# Parameters
2
2
(2.M.2)
As the sample size increases the penalty decreases, thus AICC ≈ AIC. Whereas α = P(Type I Error) = P(Reject
Ho | Ho is True), β = P(Type II Error) = P(Accept Ho | Ho is False). The power of a hypothesis test is 1 – β =
P(Reject Ho | Ho is False). If interest is in controlling the power of subsequent hypothesis tests then Bayesian
information criteria (BIC) is recommended (Tao, et al., 2002), and calculated as (Schwarz, 1978):
BIC = -2 ln( ( | )) + (# Parameters) ln(Sample Size)
(2.M.3)
While determining the optimal # of components in a finite mixture PDF (see §II.B) is an unsolved problem
in statistics, IC are often used as a heuristic to compare mixture PDFs of various sizes. Titterington, et al.,
(1985) caution that, in theory, their validity often relies on the unmet regularity conditions of §II.G.
N. The Probability of Ruin in Retirement
Let t=1,2,...,T be the time points of a retirement horizon where the 1st withdrawal is made at time t=1
and the last withdrawal at time t=T, which can be fixed (TF) or random (TR). The PMF for TR is defined as
P(TR = t), for t=0,1,...,T, and can be derived using lifetables published at SSA.gov for an individual or a group
(Rook, 2014). The safe withdrawal rate (SWR) is a heuristic that suggests retirees withdraw (100%)WR in
real terms from their savings at each time point (Bengen, 1994). A retirement plan often couples the
withdrawal rate (WR) with an asset allocation. If N securities are involved let Rti, rti, Ei, αti, and It be the total
and real returns, expense ratio, proportion allocated to security i, and inflation rate, respectively, all at time t.
The total compounding return for security i at time t is:
24
(2.N.1)
(1+ Rti) = (1+ rti)(1+ It)
The expense-adjusted total compounding return for security i at time t is:
(2.N.2)
(1- Ei)(1+ Rti) = (1- Ei)(1+ rti)(1+ It)
The expense-adjusted real compounding return for security i at time t, denoted ȓti, is:
ȓti = (1- Ei)(1+ rti) = (1- Ei)(1+ Rti)/(1+ It)
(2.N.3)
Since ȓti is a continuous RV it is governed by a univariate PDF, say fti(ȓti). When Cov(rki, rsj) = 0 for times
k≠s=1,...,T and securities i,j=1,...,N, then ȓti and fti(ȓti) are independent of time. If we drop the time index they
become ȓi and fi(ȓi), respectively. The marginal PDF for security i, fi(ȓi), is modeled using historical data.23 A
retirement plan succeeds or fails based on the return of a diversified portfolio, not that of a single security.
The expense-adjusted real compounding return for the portfolio at time t is:
(2.N.4)
ȓ(t) = (αt-1,1)ȓ1 + ... + (αt-1,N)ȓN,
where αt-1,1 + ... + αt-1,N = 1. Consequently, ȓ(t) is a function of time via the portfolio weights (i.e., the asset
allocation) set at time t-1 and is derived as a linear transform (
N
→
1
) of ȓ = (ȓ1, ..., ȓN)′. The univariate PDF
for ȓ(t), say ht(ȓ), is used to evaluate/optimize a retirement plan. In some cases this PDF is easily derived
(i.e., normal), and in others there is no solution (i.e., lognormal). Various methods exist to derive the PDF of
a transformed RV and one uses the multivariate PDF of the random vector ȓ = (ȓ1,...,ȓN)′, f(ȓ) (Freund, 1992).
Our goal is to model ht(ȓ) using f(ȓ), while maintaining the individual security marginals, that is, subject to
fi(ȓi) =
…
ȓ ,...,ȓ
ȓ
…
ȓ
. Here, ht(ȓ) may be skewed, heavy-tailed and multimodal (i.e., generally non-
normal) allowing higher PDF moments to aid in determining a plan’s success or failure.
N.1 Retirement Surveys and Alternative Metrics
Retiree surveys reveal that the #1 concern is running out of money.24 A retiree who runs out of
money experiences financial ruin. The probability of this event occurring can be computed and shared with
the retiree for a given decumulation strategy. Probabilities are bounded by [0, 1] and multiplying by 100%
yields a percentage which is bounded by [0, 100]. Percentages are a ubiquitous metric and universally
understood. For example, a probability of 0.5 translates to 50% which can be described in words as a “coin
flip”. Retirees make withdrawals at time t and they experience the event of ruin at time t, denoted Ruin(t), iff
the time t-1 withdrawal is successful but the account does not support, or is completely emptied by the time t
withdrawal. The compliment of this event is avoiding ruin at time t, denoted RuinC(t), which occurs iff the
withdrawal at time t is successful, leaving a > $0 balance. Define Ruin(≤ t) as the event of ruin occurring on
or before time t and let RuinC(≤ t) be its compliment.
23
24
The tools described in §II.J can be used to detect serial correlation within and between securities over time.
A number are referenced at the conclusion of this research.
25
Despite being extensively researched, the probability of ruin as a metric is not universally accepted
with criticisms leveled from all directions. The most common being that the retirement ruin event could
occur as Ruin(1) or Ruin(30), and there is a substantial real-life difference for the retiree between these
events. This criticism argues that ruin as a binary outcome is too simplistic and a more nuanced approach
would consider varying degrees of failure. A separate criticism is that the ruin metric is overly complicated
and better left to actuaries at insurance companies. Under this argument, the metric is misunderstood and
being abused by financial planners who lack the ability to properly calibrate the computation and/or fail to
understand its inherent flaws such as the impact of covariances and higher order PDF moments.
While a retirement strategy can have varying degrees of failure, we are primarily interested in the
compliment of the ruin event, which is success. Unlike retirement ruin, the event of retirement success does
not have varying time point-degrees attached to it. Further, a decumulation model that maximizes the
probability of success will also minimize the probability of ruin as these are equivalent optimization
problems. None-the-less, a model that maximizes the probability of success may in fact fail, and in this case
it is reasonable to try and limit the damage. Harlow and Brown (2016) introduce two downside risk metrics
which do precisely this. Their approach uses fully stochastic discounting to compute a retirement present
value (RPV) for withdrawals (cash flows) from a decumulation plan. The RPV is an RV and its PDF can be
estimated via simulation. Values of RPV below zero indicate the account did not support all withdrawals
and retirement ruin has occurred. A strategy that minimizes downside risk recognizes that the ruin section of
the RPV’s PDF can be markedly different for retirement plans having similar failure probabilities. The goal
is to make this section of the PDF as palatable as possible if ruin occurs. Both the mean and standard
deviation of negative RPV values are used as minimization metrics in the optimization and corresponding
asset allocations are found. Harlow and Brown (2016) report that far lower equity ratios are optimal in the
context of minimizing downside risk. This finding has the benefit of being intuitive as we can generally
think of a retiree’s bequest distribution as having a spread (variance) that increases with the equity ratio and a
negative bequest (i.e., RPV < 0) indicates that the retiree has exhausted their savings while still alive.
Milevsky (2016) takes the opposing view that ruin probabilities are being routinely abused and/or
misunderstood by retirement planners, and advocates for replacing it altogether by a different metric, namely
the portfolio longevity (PL). Since investments are volatile PL is an RV that measures the length of time a
retirement portfolio lasts. It takes values l = 0, 1, 2, ..., ∞ in discrete time and has PMF defined by P(PL = l).
Note that (l = 0) ↔ (0 successful withdrawals are made) ↔ Ruin(1), (l =1) ↔ (only 1 successful withdrawal is
made) ↔ Ruin(2), ..., (l = T-1) ↔ (exactly T-1 successful withdrawals are made) ↔ Ruin(T). Finally, (l ≥ T) ↔
(All T withdrawals are made successfully) ↔ RuinC(≤T) ≡ Retirement Success (given horizon length T).
Therefore, P(PL = l) = P[Ruin(l+1)] for l = 0, 1, 2, ...,T-1 and P(l ≥ T) = P[RuinC(≤T)], see Table III. The mean,
26
median and mode of PL will thus be functions of ruin probabilities, and any flaws inherent in their
construction will propagate through to these statistics.
Table III. The Portfolio Longevity (PL ) PMF and Corresponding Statistics
Portfolio Longevity (PL= l)
0
1
2
⋮
i
⋮
j
⋮
k
⋮
1
2
P(PL = l)
P[Ruin(1)]
P[Ruin(2)]
P[Ruin(3)]
Mean (μ=E[PL])
0
1*P[Ruin(2)]
2*P[Ruin(3)]
⋮
P[Ruin(i+1)]
⋮
P[Ruin(j+1)]
⋮
P[Ruin(k+1)]
⋮
∑ 1
⋮
i*P[Ruin(i+1)]
⋮
j*P[Ruin(j+1)]
⋮
k*P[Ruin(k+1)]
⋮
∑
μ EP
Median1
Mode2
X
X
X
∑
0.5
argmax[P(PL = l)]
Sum probabilities in either direction and stop when ≥ 0.5 The corresponding l-value is the median, shown as j above. If the sum
is exactly 0.5 at l = j then median(PL) = (2j+1)/2.
Locate the maximum probability(s) P(PL= l). All corresponding l-values are the mode(s), shown above as mode(PL) = {i,k}.
Table III applies to any non-negative withdrawal rate (WR), and ∑
P Ruin t =1 implies that no
investment account lasts in perpetuity, including when WR = 0. It is suggested that financial advisors examine
the event PL < TR as it implies the retiree outlives their savings. The probability of this event, P(PL < TR), is the
probability of ruin (with T = TR). We can compute P(PL < TR) using conditional probabilities as follows:
PL < TR
≡ P
T
∩ S
≡ P
T
∩ T
0 ∪ T
∩ T
0 ∪ ⋯∪ P
≡ P
→ P(PL < TR)
T
[where S is any generic sample space, i.e., P(S)=1]
1 ∪ ⋯∪ T
=∑
P P
T ∩ T
=∑
P P
T |T
=∑
P P
t|T
=∑
1
P P
=∑
1
P Ruin
=∑
1
P Retirement Success | T
[replace S by the sample space for TR]
T
∩ T
T ] [distributive property for sets]
[probabilities for mutually exclusive events are summed]
t
t P T
t P T
t|T
T
(2.N.5)
t
[conditional probability uses fixed horizon length, TF]
t
t P T
T |T
[since P(A|B) = P(A ∩ B)/P(B)]
[drop t=0 from prior step since PL ≥ 0]
t
t P T
t
t P T
[compute using success probabilities]
t
(2.N.6)
As noted, P(PL < TR) is, by definition, the probability of ruin using a random time horizon. With respect to the
portfolio longevity (PL), the probability said to be of most interest is derived entirely using ruin probabilities,
as shown in (2.N.6) where probabilities for the RV PL do not appear.
27
N.2 Computing Ruin Probabilities
Assume the retiree has made successful withdrawals at times t=1,...,t-1. The event Ruin(t) occurs
when ȓ(t) ≤ RF(t-1), where RF(t) = RF(t-1)/[ȓ(t) – RF(t-1)] for t=1,...,T and RF(0) = WR (Rook, 2014). Thus,
RuinC(t) occurs when ȓ(t) > RF(t-1). RF(t) is called the ruin factor and reflects the retiree’s funded status at
time t with 1/RF(t) equal to the # of real withdrawals remaining (Rook, 2014). The event of achieving
retirement success using an SWR, is thus defined for fixed and random horizons, respectively, as:
RuinC(≤ TF) ≡ (Retirement Success | T=TF) ≡ ⋂
ȓ t
RF t
1 , ȓ(t) ~ ht(ȓ)
(2.N.7)
RuinC(≤ TR) ≡ (Retirement Success | T=TR) ≡ ⋂
ȓ t
RF t
1 , ȓ(t) ~ ht(ȓ) and TR ~ P(TR = t)
(2.N.8)
Recall that ȓ(t) is the expense-adjusted real compounding return on a diversified portfolio of N securities and
a function of the asset allocation set at time t-1. It follows that the corresponding success probability for
fixed T=TF in (2.N.7), assuming independence of ȓ(t) across time, is (
T
P Ruin
…
…
replaces ȓ(t) as vbl of integration):
…
(2.N.9)
The success probability for random T=TR in (2.N.8) is 1 – P[Ruin(≤ TR)], where P[Ruin(≤ TR)] was derived in
(2.N.6) above and also uses (2.N.9). For any events A and B, if A ⊂ B then P(B ∩ AC) = P(B) - P(A). Since
RuinC(≤ t) ⊂ RuinC(≤ t-1) and Ruin(t)≡[RuinC(≤t-1)∩Ruin(≤t)], it follows that P[Ruin(t)] = P[RuinC(≤ t-1) ∩
Ruin(≤t)] = P[RuinC(≤ t-1)] - P[RuinC(≤t)].25 Consequently, for ȓ(t) ~ ht(ȓ) and t ≤ TF:
P Ruin t
P ⋂
…
ȓ i
RF i
1
P ⋂
…
ȓ i
RF i
…
1
(2.N.10)
…
…
…
Given values for the security weights (αt-1,1, ..., αt-1,N) at time t (i.e., the asset allocation), the terms in (2.N.10)
are estimated using simulation or approximated recursively with a dynamic program (DP), see Rook (2014)
and Rook (2015). Subsequently, Table III can be populated with these probabilities. Usually we are not
given the weights but are tasked with deriving them according to some optimality criteria. Rook (2014)
derives the weights that minimize the probability of ruin for a stock and bond portfolio using a dynamic
glidepath (for both TF and TR), and Rook (2015) derives the corresponding weights to minimize the
probability of ruin using a static glidepath. Both solutions assume normally distributed compounding returns
and, as noted, many financial practitioners/researchers reject this assumption. The primary purpose of this
research is to extend these and other models to non-normal compounding returns. It was noted above that
investment accounts do not last forever regardless of the withdrawal rate (WR). When WR = 0, RF(t) = 0 ∀ t =
25
See Rook (2014) for corresponding Venn diagrams.
28
0, ..., T. As T → ∞, Ruin(t) occurs when ȓ(t) ≤ 0 for any t. Consequently, if P[ȓ(t) ≤ 0] > 0, the event of ruin will
eventually occur under an infinite time horizon. Unfortunately, the lognormal PDF is defined for values ≥ 0
with f(0) = 0 and it does not allow compounding returns of zero. The PDF we develop for ȓ(t) assigns a > 0
probability to the event ȓ(t) ≤ 0 as compounding returns and prices of securities can and do take values of zero.
III. Univariate Density Modeling
The expense-adjusted real compounding return on a diversified portfolio determines the success or
failure of a retirement strategy (§II.N). We assume independence across time and use securities from Table I
that are random samples.26 The multivariate PDF for the expense-adjusted real compounding return on S&P
500 (L), Small Cap Equities (S), and U.S. 10-Year T-Bonds (B) will be developed in this research. The RVs
representing these returns are (L, S, B)′ and the multivariate PDF is f(l,s,b). A diversified portfolio using
these securities generates time t expense-adjusted real compounding return ȓ(t) =αt-1L(1-EL)L + αt-1S(1-ES)S+
αt-1B(1-EB)B, where αt-1,L, αt-1,S, αt-1,B are the portfolio weights (≥0) set at time t-1 with αt-1,L + αt-1,S + αt-1,B = 1,
and EL, ES, EB are the expenses. Similar to copula modeling (§II.L), we first build univariate PDFs fL(l),
fS(s) and fB(b) for L, S and B, respectively. The multivariate PDF f(l,s,b) (built in §IV and §V) will preserve
the marginals, i.e., fL(l) =
, ,
, fS(s) =
, ,
and fB(b) =
, ,
.
Univariate PDFs for (L, S, B)′ are fit to finite normal mixtures using the EM algorithm with random starts and
a variance ratio constraint to eliminate spurious maximizers (§II.E). A novel forward-backward procedure
is introduced to find the optimal # of univariate components, generally considered an unsolved problem in
statistics (§II.M). The forward portion tests 1 vs. 2 components using a bootstrapped LRT (§II.H), then (1
or 2) vs. 3 components, up to the maximum # of univariate components allowed. If the forward procedure
ends with the last significant test being g components, the backward portion tests g vs. g-1 components, then
g vs. g-2, etc..., until a significant difference is found ending the procedure. For example, if backward test g
vs. g-k yields a significant difference then the optimal # of components is g-k+1. Note that AndersonDarling normality tests for (L, S, B)′ yield p-values 0.7206, 0.0984 and 0.2607, respectively, indicating that
the normality assumption is not rejected at α=0.05 for these securities. The fact that all 3 can be assumed to
originate from univariate normal distributions should not be lost in the forthcoming analysis.
A. Univariate PDFs
All univariate tests of g vs. g+k components will use 1,000 LRT bootstrap samples. Each sample fits
data to both g and g+k component mixtures using 6,000*(g-1) random starts for each execution of the EM
algorithm when g>1. Random starts use values generated from the nearest fitted mixture for the same data
but with fewer components. Each LRT sample value thus requires 6,000*(2g+k-2) EM executions.
26
Retirement research that uses serially correlated assets should account for the dependence in the multivariate PDF, otherwise valid
doubts may be raised. This would include strategies that use certain short-term cash equivalents such as U.S. T-Bills (see Table I).
29
A.1 Univariate PDF for S&P 500 (L): fL(l)
Figure IV. Annual Real S&P 500 Compounding Returns (L) Histogram
Table IV. Univariate Mixture PDFs for Annual Real S&P 500 Compounding Returns (L)
Number of Components
2
5
1
3
4
LL =
VR =
AIC =
AICC =
BIC =
17.9019
1.0000
-31.8038
-31.5181
-26.8491
LL =
VR =
AIC =
AICC =
BIC =
18.6383
46.7606
-27.2766
-26.2396
-14.8899
LL = 20.1144
VR = 68.4102
AIC = -24.2289
AICC = -21.9211
BIC = -4.4102
LL =
VR =
AIC =
AICC =
BIC =
25.1860
84.6646
-28.3721
-24.2120
-1.1214
LL = 26.7351
VR = 42.6613
AIC = -25.4702
AICC = -18.8035
BIC = 9.2125
μ(L) =
σ(L) =
γ(L) =
κ(L) =
1.082
0.1974
0.00000
3.00000
μ(L) =
σ(L) =
γ(L) =
κ(L) =
1.082
0.1974
-0.10701
2.87231
μ(L) =
1.082
σ(L) = 0.1974
γ(L) = -0.26986
κ(L) = 3.02576
μ(L) =
σ(L) =
γ(L) =
κ(L) =
1.082
0.1974
-0.20207
2.76519
μ(L) =
1.082
σ(L) = 0.1974
γ(L) = -0.14463
κ(L) = 2.98090
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
1.000
1.082
0.197
0.056
1.287
0.029
0.833
1.055
0.204
0.092
0.998
0.016
0.114
0.998
0.017
0.944
1.070
0.197
0.078
1.286
0.029
0.708
1.170
0.149
0.045
0.630
0.022
0.089
1.157
0.025
0.155
0.860
0.046
0.191
0.864
0.049
0.045
0.629
0.022
0.037
1.505
0.064
0.613
1.174
0.110
Abbreviations: LL = log-likelihood, VR = Variance Ratio, γ = skewness, κ = kurtosis.
27
27
See §II.M for AIC, AICC, BIC.
Skewness(X) = γ(X) = E[((X-μ)/σ)3] = [E(X3) - 3μE(X2) + 2μ3]/σ3 = 0 for the normal PDF, thus E(X3) = 3μE(X2) - 2μ3. Kurtosis(X) =
κ(X) = E[((X-μ)/σ)4] = [E(X4) - 4μE(X3) + 6μ2E(X2) - 3μ4]/σ4 = 3 for the normal PDF, thus E(X4) = 3σ4 + 4μE(X3) -6μ2E(X2) + 3μ4. (See,
Hogg, et al., 2005 for definitions and Johnson, et al., 1994 for moment details.) Note that higher moments such as these can be difficult to
interpret in multimodal distributions.
30
Figure V. LRT Sampling Distribution for Testing the Optimal # of Univariate Mixture Components
Annual Real S&P 500 Compounding Returns (L)
Forward ↓
Backward ↑
Figure IV plots annual real compounding returns for the S&P 500 Index from 1928-2015 (n=88) in a
histogram and Table IV fits these returns to univariate normal mixtures (≤ 5 components). Tests for the
optimal # of components are done in Figure V using the forward-backward procedure described above. All
displayed values are rounded throughout and unrounded values are used in calculations. The VR constraint
is 162 and a significance level of α=0.2500 is used for each test. Some evidence of non-normality will lead to
rejection of Ho. As shown in Table IV, there is insufficient evidence to reject normality for annual real
compounding S&P 500 Index returns. The forward-backward bootstrapped LRT procedure is in agreement
with all 3 information criteria values (AIC, AICC, BIC) that a univariate normal PDF is appropriate for these
returns, which also agrees with the AD test for normality (p=0.7206).
31
A.2 Univariate PDF for Small Cap (S): fS(s)
Figure VI. Annual Real Small Cap Compounding Returns (S) Histogram
Table V. Univariate Mixture PDFs for Annual Real Small Cap Compounding Returns (S)
Number of Components
2
5
1
3
4
LL =
VR =
AIC =
AICC =
BIC =
-24.7605
1.0000
53.5211
53.8067
58.4758
LL =
VR =
AIC =
AICC =
BIC =
-23.0510
148.4732
56.1021
57.1390
68.4888
μ(S) =
σ(S) =
γ(S) =
κ(S) =
1.132
0.3206
0.00000
3.00000
μ(S) =
σ(S) =
γ(S) =
κ(S) =
1.132
0.3206
0.13061
3.02488
LL = -18.3915
VR = 241.2802
AIC = 52.7829
AICC = 55.0907
BIC = 72.6016
μ(S) =
σ(S) =
γ(S) =
κ(S) =
LL =
VR =
AIC =
AICC =
BIC =
-15.9514
241.5788
53.9028
58.0628
81.1535
μ(S) =
σ(S) =
γ(S) =
κ(S) =
1.132
0.3206
0.15342
4.60462
1.132
0.3206
0.23954
3.71740
LL = -14.4531
VR = 198.4935
AIC = 56.9062
AICC = 63.5729
BIC = 91.5889
μ(S) =
σ(S) =
γ(S) =
κ(S) =
1.132
0.3206
0.61679
4.52729
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
1.000
1.132
0.321
0.928
1.150
0.326
0.164
0.945
0.065
0.483
1.155
0.426
0.194
1.185
0.029
0.072
0.908
0.027
0.707
1.165
0.367
0.114
1.338
0.051
0.149
1.338
0.054
0.129
1.192
0.024
0.167
1.187
0.027
0.105
0.662
0.139
0.237
0.948
0.074
0.258
1.370
0.415
0.295
0.954
0.076
Abbreviations: LL = log-likelihood, VR = Variance Ratio, γ = skewness, κ = kurtosis. See §II.M for AIC, AICC, BIC.
32
Figure VII. LRT Sampling Distribution for Testing the Optimal # of Univariate Mixture Components
Annual Real Small Cap Compounding Returns (S)
Forward ↓
Backward ↑
Annual real small cap compounding returns from 1928-2015 (n=88) are plotted in Figure VI and
Table V fits these returns to univariate normal mixtures (≤ 5 components). Tests for the optimal # of
components are shown in Figure VII using the forward-backward procedure described above with VR
constraint of 162 and significance level α = 0.2500. The test of 1 vs. 3 components yields a significant pvalue and Ho is rejected.
Backward processing begins by testing 2 vs. 3 components which is also
significant, ending the procedure. A 3-component normal mixture PDF is therefore found appropriate for
these returns. The coefficient γ(S) = 0.23954 > 0 indicates the PDF has positive skew and κ(S) = 3.71740 > 3
indicates a positive excess kurtosis which implies a heavier tail than the normal distribution. The fitted
marginal PDF is evidently skewed, heavy-tailed, and multimodal (Table V).
33
A.3 Univariate PDF for U.S. 10-Year T-Bonds (B): fB(b)
Figure VIII. Annual Real U.S. 10-Year T-Bond Total Compounding Returns (B) Histogram
Table VI. Univariate Mixture PDFs for Annual Real U.S. 10-Year T-Bond Total Compounding Returns (B)
Number of Components
2
5
1
3
4
LL = 94.9092
VR = 1.0000
AIC = -185.8184
AICC = -185.5327
BIC = -180.8637
μ(B) =
σ(B) =
γ(B) =
κ(B) =
LL = 99.4731
VR = 16.7844
AIC = -188.9463
AICC = -187.9092
BIC = -176.5596
1.022
0.0823
0.00000
3.00000
μ(B) =
σ(B) =
γ(B) =
κ(B) =
1.022
0.0823
0.47529
3.36595
LL =
VR =
AIC =
AICC =
BIC =
100.2688
20.2289
-184.5377
-182.2299
-164.7190
μ(B) =
σ(B) =
γ(B) =
κ(B) =
1.022
0.0823
0.50035
3.26913
LL = 102.2525
VR =
8.8448
AIC = -182.5049
AICC = -178.3450
BIC = -155.2542
μ(B) =
σ(B) =
γ(B) =
κ(B) =
1.022
0.0823
0.49208
3.14599
LL =
VR =
AIC =
AICC =
BIC =
104.1529
125.3318
-180.3057
-173.6391
-145.6230
μ(B) =
σ(B) =
γ(B) =
κ(B) =
1.022
0.0823
0.49047
3.14076
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
π
μ
σ
1.000
1.022
0.082
0.948
1.011
0.070
0.893
1.004
0.065
0.089
1.118
0.020
0.350
0.991
0.029
0.052
1.220
0.017
0.055
1.220
0.017
0.118
0.903
0.021
0.385
1.059
0.041
0.052
1.120
0.015
0.057
1.220
0.017
0.042
1.128
0.004
0.736
1.014
0.051
0.167
0.908
0.024
0.057
1.220
0.017
Abbreviations: LL = log-likelihood, VR = Variance Ratio, γ = skewness, κ = kurtosis. See §II.M for AIC, AICC, BIC.
34
Figure IX. LRT Sampling Distribution for Testing the Optimal # of Univariate Mixture Components
Annual Real U.S. 10-Year T-Bond Total Compounding Returns (B)
Forward ↓
Backward ↑
Annual real U.S. 10-Year T-Bond total compounding returns from 1928-2015 (n=88) are shown in
Figure VIII and are fit to univariate normal mixtures in Table VI. The optimal # of components is found via
the forward-backward procedure detailed in Figure IX with VR constraint of 162 and significance level of α =
0.2500. The test of 1 vs. 2 components yields a significant p-value and Ho is rejected. Backward processing
begins by testing 1 vs. 2 components which is a repeat test and not performed. A 2-component normal
mixture PDF is thus appropriate for these returns. The coefficient γ(B) = 0.47529 > 0 indicates this PDF has
positive skew and κ(B) =3.36595 > 3 indicates a positive excess kurtosis which implies a heavier tail than the
normal distribution. The fitted marginal PDF is evidently skewed, heavy-tailed, and multimodal (Table VI).
35
B. Univariate PDF Summary
Security
S&P 500 (L)
Small Cap (S)
U.S. 10-Year T-Bond (B)
Table VII. Full Univariate PDF Parameterization
π
Component
μ
σ
1
1.000000000000000
1.082139318181818
0.197430382245555
1
0.163796557010864
0.944667188140307
0.065233151408053
2
0.707369571461765
1.165057494177362
0.366529325768043
3
0.128833871527371
1.191903886074301
0.023596517545339
1
2
0.947744576049301
0.052255423950700
1.011164539967906
1.220436091927283
0.069579917666149
0.016983666409906
Note: Use these estimates along with the historical data to reproduce the log-likelihood values.
Tests for the optimal # of components use α = 0.2500. Large values are common defaults in forwardbackward testing procedures. A variance ratio constraint of VR = 256 was used to eliminate spurious
optimizers while finding MLEs (§II.E) with the EM algorithm (§II.F). Reducing this value or adding new
constraints on either the probabilities or means will alter the PDF shapes, for example increasing or
decreasing the # of modes/peaks (see Figure X). To add a constraint simply discard any random start that
violates it.
Skewed unimodal PDFs such as the lognormal were available but did not optimize the
constrained objectives. Under interpretation (2) from §II.B components are assigned labels. The data
suggest that annual real compounding S&P 500 returns originate from one regime, but that small cap equity
returns originate from 3. Namely, a dominant N(1.165, 0.367) PDF generates about 70% of returns (including
outliers) and a low N(0.945, 0.065) PDF generates >15%, with <15% originating from a high N(1.192, 0.024)
PDF. The high/low regimes add shoulders above and below the mean with the PDF evidently heavier tailed
than a normal. Annual real U.S. 10-Year T-Bond total compounding returns originate from 2 regimes with
the dominant N(1.011, 0.070) PDF generating 95% of returns, and a high-outlier N(1.220, 0.017) PDF regime
generating the other 5%. Note that these returns averaged 1.033 from 2010 - 2015, which is above the
dominant regime μ and overall historical mean of μ(B) = 1.022. Consequently, widespread claims that current
low yields invalidate retirement heuristics such as the “4% rule” should be met with skepticism.
Figure X. Univariate Mixture PDFs with Probability Weighted Component Regimes
36
IV. Multivariate Density Modeling w/out Covariances
The multivariate PDF for (L, S, B)′ is built in two steps. First, dependence is introduced without
correlations and the result is the starting point for a final step of estimating correlations. Under interpretation
(2) in §II.B, mixture observations come from labeled regimes as seen in §III.B for (L,S,B)′. An observation
from the multivariate PDF can be viewed as originating from some combination of these regimes. There are
1, 3, 2 regimes governing L, S, B, respectively, thus at most 1·3·2 = 6 regimes will govern the multivariate PDF.
A parsimonious multivariate PDF may call for eliminating combinations that have not produced data. The
goal is to perform regime selection in an optimal manner, accounting for the sample size and total # of
multivariate PDF parameters. We must also preserve the marginal PDFs that were derived in §III.A.
A. Multivariate Regimes
Under mixture PDF interpretation (2), regimes produce observations. After estimating the PDF
parameters in a mixture (2.B.2), we can estimate the probability that an observation is from a given regime.
Let Xt and Zt = (zt1, ..., ztg)′ be the time t observation and component indicator RVs for a g-component
univariate mixture PDF, for t=1,...,T. Assume all parameters have been estimated as in §III.A and that xt is
the observed value at time t. The probability that xt is produced by component i, i=1,...,g is:
P
Since
1|
has been estimated and
P
P
1,
|
P
=
|
with
1 ∗P
P
∗
1
∑
(4.A.1)
also estimated, this quantity can be computed for
each observation xt, t=1,...,T. Bayes decision rule assigns each observation to the component with largest
probability and is considered an optimal allocation scheme (McLachlan & Peel, 2000). Let nijk be the # of
observations in regimes i, j, k of RVs L, S, B, respectively, using the assignment rule above. The probability
that an observation on (L, S, B)′ originates from a given multivariate component can be estimated as
̂ =nijk/T, where
will be the true (unknown) probability, see Figure XI.
Figure XI. Multivariate Regime Combinations for (L, S, B) with Estimated Probabilities
Note: Observations on the trivariate (L, S, B) exist in each cell and correlations may change across regimes. For example
the correlation between S and B may be strongly positive in one cell and strongly negative in another (see §V).
37
If L, S and B are independent, the multivariate PDF is the product of the marginals, f(l,s,b) = fL(l)·fS(s)·fB(b).
The marginals were fit to mixtures in §III.A and their product is a 6-component multivariate normal mixture
that yields the fitted marginals. Under independence the probability that a multivariate observation (l, s, b) is
from a given regime is the product of the marginal probabilities. There is no basis to assume independence
and Figure XI shows the multivariate component probabilities estimated from the data, namely ̂
.
Dependence for multivariate mixture PDFs takes 2 forms: between component and within component.
Between component dependence is modeled via the probabilities and within component dependence is
modeled via the covariances. Both must preserve the marginals from §III.A, however the ̂
estimates in
Figure XI do not, thus they are infeasible. For example, using fS(s) the probability that S is from the
“dominant” regime equals 0.707 (Table V) whereas using the data it is 60/88 = 0.682 (Figure XI). Conway
(1979) gives conditions that guarantee both univariate and bivariate PMFs are enforced for probabilities in a
3-way contingency table. If the bivariate PDF is of interest we suggest deriving it from the trivariate PDF.
B. Multivariate Regime Selection via Linear Programs (LPs)
Our goal is to parsimoniously model between component dependence while preserving the marginal
PDFs from §III.A. Two approaches are presented and become initial solutions for a final step of estimating
covariances and updating component probabilities. We limit the discussion to the problem at hand, however,
the methods are completely generic and easily extendable to arbitrarily dimensional problems.
B.1 Minimize Maximum Distance (Minimax) LP
Define the maximum distance between true (
S, B from Figure XI as Z = max{|
- ̂
|, |
- ̂
|, |
) and estimated ( ̂ ) component probabilities for L,
- ̂
|, |
- ̂
|, |
|, |
- ̂
- ̂
|}. Values for
that minimize Z and preserve the marginal PDFs from §III.A would be of interest. Namely, solving:
Minimize:
Subject to:
Z = max{|
- ̂
|, |
- ̂
|, |
- ̂
|, |
1.000
0.164
0.707
0.129
|, |
|, |
- ̂
(L-marginals)
(S-marginals)
- ̂
|}
(4.B.1)
(B-marginals)
0.948
0.052
1.000
Where:
- ̂
≥ 0, i=1, j=1,2,3, k=1,2
(Probabilities sum to 1)
(non-negativity constraints)
Minimizing the maximum of a set is a minimax objective. While Z is not linear, the constrained optimization
problem in (4.B.1) can be formulated and solved as an LP. Note that of the 7 constraints, the 4th and 6th are
redundant since all probabilities sum to 1 as is the 7th since L has 1 component. They are thus dropped, and
since the maximum of a set must be greater-than or equal-to all set elements, (4.B.1) becomes:
38
Minimize:
Subject to:
Z
Z≥|
- ̂
|
i=1, j=1,2,3, k=1,2
(L-marginals)
(S-marginals)
1.000
0.164
0.707
(B-marginals)
0.948
Where:
≥ 0, i=1, j=1,2,3, k=1,2
Z,
(4.B.2)
(non-negativity constraints)
Constraints in (4.B.2) include absolute values and are non-linear, however note that X ≥ |Y| iff X ≥ Y and X ≥
-Y. To promote parsimony we penalize the objective when a cell from Figure XI containing 0 observations
yields a > 0 probability. This will ensure it only occurs when needed for feasibility. The constraints are
nijk +Xijk, and the penalty is M· ∑
∑
∑
, where M is arbitrarily large. When nijk = 0 and
X
≤
> 0, Xijk
must be > 0 to satisfy the constraint and Z suffers a penalty. The final LP formulation is:
Minimize:
Z+M∙ ∑
∑
Subject to:
Z≥
and Z ≥ ̂
- ̂
∑
X
for i=1, j=1,2,3, k=1,2
for i=1, j=1,2,3, k=1,2
(L-marginals)
(4.B.3)
(S-marginals)
-
≤ nijk + Xijk
1.000
0.164
0.707
(B-marginals)
0.948
Where:
, Xijk ≥ 0, for i=1, j=1,2,3, k=1,2
M and n for i=1, j=1,2,3, k=1,2
(non-negativity constraints)
(known constants)
Z,
This LP was solved using the techniques from §II.K.1. The solution is
= 0.1212,
= 0.0000,
= 0.1638,
= 0.6627,
= 0.0076, yielding the minimization objective Z = 0.0151.
= 0.0447,
B.2 Minimum Sum of Squared Distances LP
Define the sum of squared distances between true (
for L, S, B from Figure XI as Z = (
Values of
- ̂
2
- ̂
+
2
+(
) and estimated ( ̂ ) component probabilities
- ̂
2
+(
- ̂
2
+(
- ̂
2
+(
- ̂
2
.
that minimize Z and preserve the marginal PDFs in §III.A are of interest. Namely, solving:
Minimize:
Z=∑
∑
∑
̂
Subject to:
1.000
0.164
0.707
0.948
Where:
≥ 0, i=1, j=1,2,3, k=1,2
(L-marginals)
(S-marginals)
(4.B.4)
(B-marginals)
(non-negativity constraints)
The objective Z is quadratic and separable therefore (4.B.4) can be approximated by an LP as shown in
§II.K.2. Each decision variable
appears in one term of Z and has a convex shape similar to Figure III.
Each horizontal axis range is 0 ≤
≤ 1 and (4.B.4) is converted to an LP by replacing the quadratic terms
39
with connected line segments. The horizontal axis is partitioned into S contiguous sections and any value
is reachable as
linearly by ∑
α p
=∑
α p
p/S
p/S with ∑
α p
=1. Each term
, and the same penalty from (4.B.3) is applied thus (4.B.4) becomes:
̂
∑
∑
Subject to:
∑
α p
∑
∑
∑
∑
α p
p/S
1.000
∑
∑
∑
∑
α p
p/S
0.164
∑
∑
∑
∑
α p
p/S
0.707
∑
∑
∑
∑
α p
p/S
0.948
The decision variables
∑
∑
+M∙ ∑
Minimize:
Where:
is then approximated
̂
α p
p/S
̂
∑
p/S ≤ nijk + Xijk
, Xijk ≥ 0, for i=1, j=1,2,3, k=1,2, p=0,...,S
M, S, ̂ , n for i=1, j=1,2,3, k=1,2
α p
in (4.B.4) are replaced by α p
∑
X
for i=1, j=1,2,3, k=1,2
(L-marginals)
(S-marginals)
(4.B.5)
(B-marginals)
(non-negativity constraints)
(known constants)
in (4.B.5) and note that a minimization objective
ensures only adjacent α’s are > 0 for each i,j,k. We use S=500 and solve the LP in (4.B.5) to obtain
0.1638,
= 0.6614,
= 0.1226,
= 0.0000,
= 0.0460,
=
= 0.0063, yielding the objective Z = 0.0008.
The LPs solved in §IV.B.1 and §IV.B.2 were customized for the current exercise of modeling the
multivariate PDF of (L, S, B)′ and the concepts are easily extendable to any collection of securities. The code
supplied in Appendix C models an arbitrary number of securities. Our purpose is twofold: (1) find a feasible
solution to initialize the final step, and (2) eliminate as many unnecessary multivariate components as
possible. In larger problems this step can eliminate > 80% of multivariate cells. Note that dependence for (L,
S, B)′ has been introduced without having estimated any covariances as the multivariate PDFs using these LP
solutions are not the product of the marginals from §III.A. Specifically, between component dependence has
been introduced using the data to guide which components occur together and at what frequency. Note that
randomly reordering the returns on each of the 3 assets separately would not change the results of §III.A,
however would change the probability estimates in §IV.B.1 and §IV.B.2. The unknown
from Figure XI
now have 2 sets of initial values that are feasible and preserve the marginals.
V. Multivariate Density Modeling w/Covariances
Let X1, ..., XN be RVs for the real compounding return on N financial securities that are not serially
correlated. Historical data on each Xj will be a random sample over time, say, xtj, t=1,...,T and j=1,...,N, where
Xtj = (1 + rtj) with rtj being the real return from §II.N. The marginal density for each Xj can be modeled as a
normal mixture, fj(xj), having gj components with E(Xj|zji=1)=μji, V(Xj|zji=1)=σji2 and P(zji=1)=πji for j=1,...,N
and i=1,...,gj. The multivariate PDF for (X1,...,XN)′ will be modeled as a G-component normal mixture with
Cov(Xj,Xk|zc =1) = σcjk for j≠k=1,...,N and P[zc=1] = c, c=1,...,G. As seen in Figure XI, each zc defines a
combination of univariate components. The multivariate PDF for (X1,...,XN)′ is:
40
,…,
√2π
| |
e
,
∈
(5.1)
which, to maintain the fitted marginals, must satisfy:
…
,…,
…
…
,
j
1, … , N
(5.2)
The log-likelihood of the unknown parameters given the historical sample and known parameters is:
ln
,…,
,
,…,
,…,
,
,…,
,
ln
,…,
√2π
| |
(5.3)
e
MLEs for the unknown parameters are found by solving the following general NLP (see, §II.K.4):
Maximize:
Z=∑
(5.4)
ln
∞
Subject to:
j
∞
…
∞
1
∞
…
j 1
j 1
…
N
,
j
Eigenj( ) > 0, for j=1,...,N and c=1,...,G
∑
Where:
1, σ
0
|
√
|
e
,
∈
= Diag( ), c=1,...,G
,
(uphold marginals)
(VC matrices are +definite)
, for j≠k=1,...,N and c=1,...,G
∈
1, … , N
(multivariate PDF)
(decision variables)
(known constants)
The log-likelihood for a G-component multivariate mixture is maximized in (5.4) with respect to the
probabilities and off-diagonal VC elements (covariances). Linear probability constraints maintain marginals
as in §IV.B and covariance constraints will ensure +definite VC matrices (see §II.I). Researchers found that
mixture PDF parameters can be estimated conditionally by first deriving the probabilities holding other
parameters constant, then estimating the remaining parameters holding probabilities constant (repeating until
convergence). The ECME algorithm is one such approach that makes use of the actual or incomplete loglikelihood in (2.E.2) (see, Liu & Rubin, 1994; McLachlan & Krishnan, 2008) and is the technique we will
employ in optimizing (5.4). We will define the indicator function I(c,j,i) as 1 when univariate component i of
security j exists in multivariate component c, for j=1,...,N, i=1,...,gj, c=1,...,G and 0 otherwise.
A. Multivariate PDF Optimization
A.1 Step 1: Optimize wrt Probabilities ( ) Holding Covariances Constant
Maximize:
Z=∑
Subject to:
∑
Where:
∑
ln
I c, j, i ∙
ln ∑
∙
= πji, for j=1,...,N and i=1,...,gj
(5.A.1)
(uphold marginals)
≥ 0, for c=1,...,G
(non-negativity constraints)
e
, for t=1,...,T & c=1,...,G (known constants)
√
|
|
Since, (1) linear functions are concave, (2) log of a concave function is concave, (3) sum of concave functions
is concave, Z is concave in
(Boyd & Vandenberghe, 2009). Marginals are enforced with independent
linear constraints, thus (5.A.1) is a CCP from §II.K.3 and local optimums are global optimums. A critical
41
point of the Lagrangian for (5.A.1) yields the maximum where L( , λ) = Z - ∑
(Jensen & Bard, 2003). The 1st derivatives are ∂L( ,λ)/∂ = ∑
I k, j, i
=π -∑
∂2L( ,λ)/
-∑
, the 2nd derivatives are ∂2L( ,λ)/∂ =∑
, ∂2L( ,λ)/
∑
∑
λ ∑
πji ,
I k, j, i
λ I c, j, i , and ∂L( ,λ)/∂λ
, and
=∑
λ =-I c, j, i for c≠d=1,...,G, j=1,...,N, i=1,...,gj. All 2nd derivatives wrt the Lagrange multipliers
are zero and non-negativity constraints on
are enforced by dropping any components having
< 0.
A.2 Step 2: Optimize wrt Covariances (σ ) Holding Probabilities Constant
Maximize:
Z=∑
Subject to:
Eigenj( ) > 0, for j=1,...,N and c=1,...,G
Where:
σ
ln
∑
ln ∑
√
, for j≠k=1,...,N and c=1,...,G
≥ 0, for c=1,...,G
∈
|
|
(5.A.2)
e
(VC matrices are +definite)
(decision variables)
(known constants)
Maximizing a multivariate log-likelihood function with respect to variance components is a difficult general
NLP (§II.K.4) as there may be multiple local optimums or saddle points with zero gradient as well as
boundary optimums with non-zero gradient. Searle, et al., (1992) recommend a hill-climbing procedure
based at a good starting point. The gradient (g) helps inform on direction and the Hessian (H) on step size.
Levenberg (1944) suggests a modification to Newton’s method that iterates as θi+1 = θi – [H(θi) + si·I]-1·g(θi)
where si adjusts the step size and climbing angle. Marquardt (1963) derived a similar modification iterating
as θi+1 = θi – [H(θi) + si·Diag(H(θi))]-1·g(θi), which is considered an optimal compromise between Newton’s
method (which often diverges) and gradient ascent/descent (which converges too slowly). A class of
Levenberg-Marquardt techniques based on these approaches has since been published, see Gavin (2017) for
an overview. While designed to find least-squares estimates in non-linear models (a constrained minimization
problem), Searle, et al. (1992) note that they are also useful in finding MLEs for variance components.
Both gradient ascent and Newton’s method failed to optimize (5.A.2) for the reasons stated and a
Levenberg-Marquardt approach was taken, namely iterations defined by θi+1 = θi –λi [H(θi) + si·I]-1·g(θi) with |si|
≤ Max(|Diag[H(θi)]|) and line-search parameter λi ∈ (0,1). A large # of random si and λi are generated at each
iteration and we select randomly among the top performers. Varying si and λi prevents divergence with large
si mimicking gradient ascent and small si Newton’s method. This ensures the nearest maximum is found,
relative to an informed start, while scanning nearby regions for better values. Iterating into an infeasible
region is addressed by performing a ridge repair on the offending VC matrix (§II.I.1). Exact 1st and 2nd
derivatives of the log-likelihood in (5.A.2) are derived below wrt covariance terms only (§II.I.2).
A.2.1 First Derivatives for (5.A.2): (Gradient Terms)
There are G·N·(N-1)/2 covariance terms in a G-component multivariate mixture PDF for N securities.
Using the chain rule along with the results derived in §II.I.2 and Z from (5.A.2), the 1st derivatives are:
42
Z
σ
| |
√2π
e
σ
∑
∙Q
∙
e
|
|
√2π
(5.A.3)
e
1
2
, where Q
| |
σ
| |
which is a scalar for c
(5.A.4)
,
1, … , G and j
k
1, … , N.
A.2.2 Second Derivatives for (5.A.2): (Hessian Terms)
c
p
where j < k and j ≤ r are as follows:
r
s
The 2nd derivatives wrt terms σ , σ
A.2.2.1 Case 1: wrt σ , σ
where c ≠ p (covariances are from different multivariate components)
A.2.2.2 Case 2: wrt σ , σ
σ
∙
∙
∙Q
(5.A.5)
where c = p (covariances are from the same multivariate component)
∙
Z
σ
∙Q
∙
Z
σ
σ
∙
∙
Q
σ
∙
∙
σ
Q
∙
∙
σ
∙
Q
∙Q
∙
∙Q
∙
σ
∙Q
(5.A.6)
,
where,
1
2
Q
σ
σ
The last term in (5.A.7) depends on the location of σ
∙| |
2
∙
| |
∙| |
σ
| |
2
∙
| |
∙| |
2
in
r and/or k
s
, for k
r
above diagonal in
∙
. . (5.A.7)
and is derived piecewise as:
, for j
i. e., σ
, otherwise
| |
| |
i. e., σ
i. e., σ
below diagonal in
(5.A.8)
above/below diagonal in
Whereas quasi-Newton methods approximate the Hessian, this Levinson-Marquardt approach will
use exact 2nd derivatives, supplied in (5.A.[5–8]). The multivariate PDF log-likelihood is maximized by
iterating over (5.A.1) and (5.A.2).28 The objective is non-decreasing across steps and iterations. We begin
with the informed starts from §IV.B.1 and §IV.B.2 then iterate until the log-likelihood stops increasing.
Non-border solutions are maximums if the region around (5.A.2) is concave, i.e., the Hessian is negative
28
The 2-step algorithm from (5.A.1) and (5.A.2) is an ECME approach, e.g. see Liu & Sun (1997) who apply Newton’s method to the EM
algorithm M-step for faster convergence. See also Liu & Rubin (1994) and McLachlan & Krishnan (2008) for ECME algorithm details.
43
semi-definite (all eigenvalues ≤ 0).
Border solutions may exist in non-concave regions.
Note that
eigenvalues computed from an ill-conditioned matrix are unstable and should not be used. Sparse matrices
containing extremely large and small (diagonal) entries are often ill-conditioned. The Hessian derived above
in §V.A.2.2 may appear problematic via inspection, however, it is real and symmetric thus its eigenvectors
are orthogonal and the condition # is 1 (in theory). Bauer-Fike give bounds for the accuracy of eigenvalues
as κ(U)·‖ ‖ , where the Hessian H is subject to error E and U is the matrix of column eigenvectors for H
with ‖∙‖ denoting the Euclidean norm (2-norm). The condition # is κ(U) = ‖ ‖ and when κ(U) ≫ 1 the
matrix is ill-conditioned and calculated eigenvalues are suspect (Meyer, 2000). We will select the PDF
having min(AIC) across the informed starts from §IV.B and conduct an in-depth analysis of the surface’s
properties at the maximizer, considering also whether or not it is spurious and/or ill-conditioned.
B. Multivariate PDF for (L, S, B)′
The approach from §V.A was used to estimate the probabilities and covariances for f(l,s,b), the
multivariate PDF of (L, S, B)′. The result is a 5-component multivariate mixture PDF (see Table VIII) which
supplements the univariate PDF estimates in Table VII. The univariate regime labels propagate through to
the multivariate PDF. Without a doubt many disciplines would discard this solution as a spurious maximizer
since components 3 & 5 model only a few observations and much of the log-likelihood improvement over a
multivariate normal is via these components.29 Finance, however, differs from other industries in that a
primary focus is studying risk. Extreme events drive risk and instead of discarding outliers, finance assigns
them labels such as ‘gray’ or ‘black’ swans (Taleb, 2010). Conventional wisdom suggests, for example, that
such outliers can cause a bank to fail or a retiree to experience financial ruin thus they must be accounted for.
It is also a reason why the normal distribution may be rejected in financial research. Our model accounts for
risk by either explicitly modeling low outliers which adds density to the tail or by modeling high outliers
which shifts the dominant regimes left (along with their tails). In this application the latter occurs. There is a
tradeoff with either approach as the within regime variance shrinks when observations are separated. The
kurtosis indicates whether or not the mixture PDF is heavier tailed than a normal PDF as described in §III.
As an aside, the predictive modeler may accuse us of memorizing the training data and suggest that this is
always possible with a model of sufficient complexity but that such models are poor at prediction. A best
practice when predicting is to partition the data into 3 sets then train models on the 1st set, evaluate/select the
best using the 2nd, and report results after applying the chosen model to the 3rd. Unfortunately, there is
insufficient data to use this practice on annual historical returns in finance. Lastly, using information criteria
such as AIC from §II.M, the 5-component multivariate mixture from Table VIII (LL=159.09 & 28 free
parameters) is superior to a multivariate normal (LL=134.43 & 9 free parameters), since -262.18 < -250.86.
29
The solution would be to tighten the marginal constraints (lower the variance ratio or add constraints on the means and/or probabilities).
44
Table VIII. Full Multivariate PDF Parameterization
Component
(L, S, B)
π
ρLS
ρLB
ρSB
Det(Σ)
1
(1, 1, 1)
0.163796557010864
0.718841320548123
0.162032898398328
0.156156396700733
0.000000376625410
2
(1, 2, 1)
0.683298902828332
0.846153215066181
-0.095463348011069
-0.082959075055513
0.000007134870354
3
(1, 2, 2)
0.024070668633433
0.047531637731748
-0.703796286452937
-0.743051448755313
0.000000000000100
4
(1, 3, 1)
0.100649116210105
0.577235331054455
0.555733717419093
0.698818203652309
0.000000033408954
5
(1, 3, 2)
0.028184755317266
0.962441058378833
0.487911866790132
0.673278286734948
0.000000000091838
Note: Use these estimates with those in Table VII (and the historical data) to reproduce the multivariate log-likelihood values.
Table VIII estimates were generated using the 2-step procedure from §V.A, which converged in 2
iterations. Step 1 from (5.A.1) required 5 and 4 sub-iterations while Step 2 from (5.A.2) required 54 then 1,
respectively. The optimization in (5.A.1) is convex and will converge to a global maximum, while that in
(5.A.2) is not. Starting at the LP solution from §IV.B.2, (5.A.2) instantly requires 5 ridge repairs and lands
in a non-concave region. The final VC matrix repair occurs at sub-iteration 6/54 and between sub-iterations
8/54 and 47/54 the procedure finds a concave region and methodically climbs it. The Hessian matrix
condition # begins at 1 and slowly increases, ending at 1.001278, perhaps revealing some numerical
instability. Between sub-iterations 47/54 and 54/54 a Hessian eigenvalue turns positive and Step 1 ends at a
saddle point. Step 2 ends nearby and is a boundary solution since Σ3 is borderline +definite, which is a
constraint in (5.A.2). To enforce it we require that all VC matrix eigenvalues are >0 with determinants
>0.113. Table VIII reveals that det(Σ3) is at the threshold and note that the log-likelihood value can be driven
higher at this solution by lowering the threshold, however the condition # becomes large and the result so
unstable that minor rounding of Σ3 produces a non +definite matrix. Such is the nature of a border solution.
The covariance between RVs X and Y is σXY = E[(X-E(X))(Y-E(Y))] =E(XY) - μ μ . Using the law
of total expectation, E(X) = EZ[E(X|Z)] for any RV Z, thus for a G-component multivariate mixture PDF:
σXY = EZ[E(XY | Z)] - μ μ
= ∑
= ∑
E XY|Z
σ
1 ∙P Z
= 0.014, and
–μ μ
–μ μ
(5.B.2)
(5.B.3)
= σXY/(σXσY). Using the multivariate PDF,
=
= -0.030. Similar values are common in retirement research, likely derived using
the unbiased sample estimator, σ =∑
correlations are
1
∙
μ μ
The correlation between RVs X and Y is then derived as
0.693,
(5.B.1)
= 0.801,
x
= 0.046, and
X y
Y ⁄ ∑
x
X
∑
y
Y
. The corresponding sample
= -0.056. The MLEs produced by iterating over (5.A.1) and
(5.A.2) are subject to constraints that maintain the marginals from §III.30 Further, the LPs solved in §IV set
the initial PDF structure and thereby introduce between-regime dependence. For this reason, ρ and ρ are
30
In general MLEs for variance components are biased because the degrees of freedom are not discounted for the estimated means. A
procedure such as restricted maximum likelihood (REML) corrects this by estimating the variances/covariances after removing the means.
45
unlikely to perfectly align. Since variances and covariances are defined as expectations (averages) they can
(and
be skewed by extreme values. The small positive estimate for
) masks the fact that in over 70%
of years the correlation between L and B is negative (regimes 2 & 3), and in another 12% it is strongly
positive (regimes 4 & 5). Mixture modeling thus uncovers insights previously not known as the within
component means/variances are used to derive these correlations. As witnessed during the financial crisis,
extreme value correlations can invalidate models as simple RV dependence structures do not hold during
times of stress. In hindsight, many blame the 2008 crisis on the Gaussian copula and its failure to accurately
model failure time correlations of derivative securities under duress. We should not expect other simple
structures, such as a single-parameter copula family, to perform any better. Retirement research increasingly
advocates for use of complex instruments to improve outcomes. Coupled with the near universal use of a
Gaussian (or lognormal) copula to model dependence reveals a situation that sounds all too familiar.31
C. Simulating from a Multivariate Mixture PDF
Let = (X1,..., XN)′ be RVs for the compounding return on N financial securities at a given time point
where ~ f( ) = ·f 1( ) + ... +
·f G( ) is a G-component multivariate normal mixture PDF. Simulating values
from f( ) is a 2-step process. First generate a uniform(0,1) random value, say u, to determine the component.
If u ≤ , the observation is from regime 1, else if ∑
<u≤∑
, the observation is from regime k,
k=2,...,G. Next, generate a value from the selected regime, say f k( ). If Σk has non-zero covariances we
apply a decorrelating transformation. Recall that the N eigenvalues (λi) and eigenvectors (ui) of Σk satisfy
Σk·ui = λiui so that ui′Σkui = λiui′ui = λi (see §II.I). Also from §V.A.2, the eigenvectors of a real/symmetric
matrix are orthogonal, thus ui′uj=0 for i≠j=1,...,N. Let U=[u1|...|uN] and make the linear transform Z = UX,
where E(Z) = U·E(X) and V(Z) = U′·Σk·U = is a diagonal matrix with variances λi. If the independent and
normally distributed Z = (Z1,...,ZN)′ are simulated individually as
then = U′ is a sample on X ~ f( ).
D. Stress-Testing a Retirement Plan
Let = (X1,..., XN)′ ~ f( ) be RVs for the compounding return on N financial securities at a given time
point. If f( ) has been developed using the historical sample then it accounts for “gray” swans (observed
outliers). A stress-test will be defined as evaluating a retirement strategy using a multivariate PDF, g( ), that
can produce “black” swans (unobserved outliers), a subjective determination. This is accomplished by
seeding the historical data
with extreme events. Note that the model proposed here can be fit to such data
whereas the normal (or lognormal) PDF cannot. The proposed retirement strategy is then subjected to g( ).
31
The problem is not avoided by backtesting a given strategy as the success/failure at any retirement start year is a Bernoulli RV which is
highly correlated with the success/failure in nearby years. This correlation is rarely, if ever, accounted for in retirement research.
46
VI. Expense-Adjusted Real Compounding Return on a Diversified Portfolio
Assume a retiree holds N securities and let pt-1,i and Pt,i be the price (value) of their security i holdings
at times t-1 and t, respectively, for i=1,...,N. The time t total return on security i is Rt,i = [Pt,i - pt-1,i]/pt-1,i and the
compounding return is 1+Rt,i = Pt,i/pt-1,i, so that (1+Rt,i)·pt-1,i = Pt,i. The real compounding return for security i is
1+rt,i, where 1+Rt,i = (1+rt,i)(1+It) and It is the inflation rate between times t-1 and t. Solving yields 1+rt,i =
pt,i/pt-1,i, where pt,i = Pt,i/(1+It) is the real price (value) of security i holdings at time t and (1+ rt,i)·pt-1,i = pt,i. If
security i includes an expense ratio, Ei, then Ei·Pi,t is paid as a cost at time t and the expense-adjusted price
(value) is Pt,i - Ei·Pt,i = (1-Ei)·Pi,t, the expense-adjusted compounding return is (1-Ei)(1+Rt,i ) = (1-Ei)·Pt,i/pt-1,i, and
the expense-adjusted real compounding return is (1-Ei)(1+rt,i) = (1-Ei)·pt,i/pt-1,i. Adding up all holdings, denote
the expense-adjusted total account values at times t-1 and t by vt-1 = ∑
p
,
1
and Vt = ∑
Ei ∙ P , ,
respectively. The expense-adjusted total return on the account between times t-1 and t is Rt = (Vt - vt-1)/vt-1 and
the expense-adjusted compounding return is 1 + Rt = Vt/vt-1. The time t expense-adjusted real compounding
return on the account must satisfy 1+ Rt = (1+ rt)(1+It) so that ȓ(t)=1+ rt = (1+Rt)/(1+It) = vt/vt-1, where vt = Vt/(1+It)
is the expense-adjusted real account value at time t. Combining it all,
r t
1
r
∑
V
1
I
∙v
E ∙ P,
I ∙∑ p
1
p
∑
,
p
∑
1
1
E
∑
,
1
,
1
r,
E ∙p,
p ,
α
,
∙ 1
∑
1
E
∑
1
E
1
p
r, ∙p
,
(6.1)
,
(6.2)
r, ,
where αt-1,i > 0 is the proportion invested in security i at time t-1 and {αt-1,i, ..., αt-1,N} is the asset allocation with
∑
α
,
1, which proves (2.N.4). When modeling returns as RVs that are not serially correlated the time
index on (1+rt,i) can be dropped, however ȓ(t) remains a function of time through the asset allocation.
A. Expense-Adjusted Real Compounding Return for Portfolio using (L, S, B)
Let EL, ES, and EB be annual expenses for the S&P 500 (L), Small Cap Equities (S), and U.S. 10Year T-Bonds (B). Define the asset allocation, set at time t-1, as {αt-1L, αt-1S, αt-1B} where αt-1L+ αt-1S + αt-1B = 1.
Applying (6.2), the expense-adjusted real compounding return, ȓ(t), is a linear transform of (L, S, B)′, namely:
ȓ(t) = αt-1L(1-EL)L + αt-1S(1-ES)S+ αt-1B(1-EB)B
(6.A.1)
The expenses and asset allocations are constants or decision variables. Using means (μcl, μcs, μcb) and variances
(σ , σ , σ ) from Table VII and probabilities ( ) and covariances (σ , σ , σ
, ,
∙
, ,
√2π
| |
e
,
) from Table VIII, f(l, s, b) is:
σ
σ
σ
As in §II.N, let ht(ȓ) be the univariate PDF of ȓ(t) at time t, then ht(ȓ) satisfies
where Ht(ȓ) is the CDF of ȓ(t) and
ȓ
t
σ
σ
σ
σ
σ
σ
(6.A.2)
,
, ,
ȓ = P[ȓ(t) ≤ ȓ] =
∈
ȓ
ȓ = ht(ȓ). Define the sample space S = {S1 ∪ S2 ∪ S3 ∪ S4 ∪ S5}, where
47
,
Sc is the event that Z
1 and c=1,...,5. These are mutually exclusive with P(S) = 1 and,
ȓ
Pȓ t
ȓ
P ȓ t
ȓ ∩S
P ȓ t
ȓ ∩ S ∪S ∪S ∪S ∪S
(6.A.3)
ȓ
P
ȓ t
ȓ ∩S
where ϕ , μ , σ
P ȓ t
ȓ ∩S
P ȓ t
ȓ |Z
1 ∙P Z
1
ϕ ,μ ,σ
,
is a normal PDF having the following mean and variance for t=1,...,T, c=1,...,5:
μtc = αt-1L(1-EL)μcl + αt-1S(1-ES)μcs + αt-1B(1-EB)μcb
(6.A.4)
σ = αt-1L2(1-EL)2σ + αt-1S2(1-ES)2σ + αt-1B2(1-EB)2σ
(6.A.5)
+ 2αt-1L(1-EL)αt-1S(1-ES)σ
+ 2αt-1L(1-EL)αt-1B(1-EB)σ + 2αt-1S(1-ES)αt-1B(1-EB)σ
The PDF for ȓ(t) is a 5-component univariate Gaussian mixture PDF, derived as:
ȓ
ȓ
ȓ
∙ ϕ ȓ, μ , σ
, ȓ∈
(6.A.6)
Various PDFs using (6.A.6) are shown in Figure XII and feature asset allocations that are dominant in one of
the securities (L, S, B). They are non-normal, skewed (γ > 0) and evidently (mostly) heavier-tailed (κ > 3) than
a normal PDF. Each approaches its univariate shape as the proportion for that security increases (see §VI.B).
Figure XII. PDF for ȓ(t) (Expense-Adjusted Real Compounding Return on a Diversified Portfolio)
Notes:
1.) Expenses for all asset allocations are EL = 0.0015 (0.15%), ES = 0.0025 (0.25%), and EB = 0.0020 (0.20%).
2.) Minimum variance portfolio (σ =0.076) occurs at asset allocation αL= 0.141 , αS= 0.003, αB= 0.856.
B. Marginal Distributions
B.1: S&P 500 (L): fL(l)
Set EL=0, αtL=1, αtS=0, αtB=0 in (6.A.1), (6.A.4) and (6.A.5), then ȓ(t) = L ~ fL(l) = ∑
where μ = 1.082, σ = 0.1972, c=1,...,5, and ∑
∙ ϕ ȓ, μ , σ
,
= 1. Thus, matching §III.A.1:
L ~ fL(l) = ϕ(1.082, 0.1972)
48
(6.A.7)
B.2: Small Cap (S): fS(s)
Set ES=0, αtL=0, αtS=1, αtB=0 in (6.A.1), (6.A.4) and (6.A.5), then ȓ(t) = S ~ fS(s) = ∑
where μ = 0.945, σ = 0.0652,
= 0.164 for c=1, μ = 1.165, σ = 0.3672 for c=2,3 with
=0.707, and μ = 1.192, σ = 0.0242 for c=4,5 with
∙ ϕ ȓ, μ , σ
,
+ = 0.683+0.024
+ = 0.101+0.028 = 0.129. Thus, matching §III.A.2:
S ~ fS(s) =(0.164)· ϕ(0.945, 0.0652) + (0.707)· ϕ(1.165, 0.3672) + (0.129)· ϕ(1.192, 0.0242)
(6.A.8)
B.3: U.S. 10-Year T-Bonds (B): fB(b)
Set EB=0, αtL=0, αtS=0, αtB=1 in (6.A.1), (6.A.4) and (6.A.5), then ȓ(t) = B ~ fB(b) = ∑
2
where μ = 1.011, σ = 0.070 for c=1, 2, 4 with
2
0.017 for c=3,5 with
∙ ϕ ȓ, μ , σ
,
+ + = 0.164+0.683+0.101=0.948, and μ = 1.220, σ =
+ = 0.024+0.028 = 0.052. Thus, matching §III.A.3:
B ~ fB(b) =(0.948)· ϕ(1.011, 0.0702) + (0.052)· ϕ(1.220, 0.0172)
(6.A.9)
VII. Retirement Portfolio Optimization
Dynamic retirement glidepaths are asset allocations that adapt over time to changes in either a retiree’s
funded status or market conditions, whereas static glidepaths are fixed allocations the retiree can ‘set and
forget’. Glidepaths are often considered in the context of safe withdrawal rates (SWR), see §II.N. The
optimal glidepath outperforms all others with respect to some measure. Rook (2014) and (2015) derive the
dynamic and static glidepaths that minimize the probability of ruin using an SWR, respectively. Both
models use normally distributed returns which is an assumption many practitioners and researchers reject due
to the lack of skewness and heavy tail. The purpose of this research is to extend those models to returns that
are non-normal, skewed, heavy-tailed, multi-model, i.e., of generic complexity. The multivariate PDF we
develop also allows for stress-testing a retirement plan using data seeded with extreme events (see §V.D).
A. Optimal Dynamic Retirement Glidepaths
The dynamic glidepath that minimizes the probability of ruin using an SWR is derived in Rook (2014)
via a dynamic program (DP) for fixed and random time horizons. It applies to an individual or a group. The
DP value function V(·) is defined on 3 dimensions: t, RF(t) and ȓ(t). Both t and RF(t) are discretized (see
§II.N) to construct the corresponding grid that manages V(·) and it can be solved when the CDF for ȓ(t) is
tractable. Under lognormal returns the CDF for ȓ(t) is intractable although methods do exist to approximate
it, see Rook & Kerman (2015) for one implementation. We have derived ȓ(t), the expense-adjusted annual
real compounding return for a diversified portfolio using S&P 500 (L), Small Cap Equity (S), and U.S. 10Year T-Bonds (B) returns. The CDF for ȓ(t) in (6.A.3) is a function of the normal CDF, Φ(·), which is
considered tractable as near exact approximation routines are readily available. C/C++ source code to solve
the 2-asset portfolio problem is supplied in Rook (2014). For example, an SWR plan using small cap stocks
49
(S) and bonds (B) can be optimized using ȓ(t) from (6.A.3) with αtL= 0, αtS ≥ 0, αtB≥ 0 and αtS + αtB=1, so that
αtB =1- αtS. Incorporating additional securities (such as L) into the optimal dynamic glidepath problem is
straight-forward.
B. Optimal Static Retirement Glidepaths
Expressions for the probability of ruin were supplied in §II.N for fixed (2.N.7) and random (2.N.8)
time horizons. Minimizing the probability of ruin and maximizing the probability of success are equivalent
optimization problems. Using ȓ(t) ~
T
P Ruin
ȓ from (6.A.6) and (2.N.9), under a fixed time horizon TF:
…
∙ϕ
,μ ,σ
(7.B.1)
…
The optimal static glidepath is found by maximizing (7.B.1) with respect to the asset allocation, and is solved
for a 2-asset portfolio in Rook (2015) using both gradient ascent and Newton’s method for fixed and random
time horizons. As in §VII.A, assume S and B are used so that αtL= 0, αtS ≥ 0, αtB≥ 0 and αtS+ αtB=1. Since αtB=
1- αtS, the probability of success is a function of {α0S, ..., αTF-1S}. The 1st derivative wrt each αiS, i=0,...,TF-1 is:
∂
P Ruin
∂α
T
…
∙
∙ϕ
…
∙ϕ
,μ ,σ
,μ ,σ
∙
∙
(7.B.2)
∂
∂α
∂
ϕ
∂α
∙ϕ
,μ ,σ
,μ ,σ
TF
…
TF
1
…
,
(7.B.3)
The 2nd derivatives wrt the same αiS, i=0,...,TF-1 are:
∂
∂α
P Ruin
T
∙
1
(7.B.4)
…
∙ϕ
,μ ,σ
∙
∂
∂α
ϕ
,μ ,σ
TF
…
1
The 2nd derivatives wrt αjS, j≠i=0,...,TF-1 are:
∂
∂α ∂α
P Ruin
T
(7.B.5)
…
∙ϕ
,μ ,σ
∙
∂
ϕ
∂α
,μ ,σ
∙
∂
ϕ
∂α
,μ ,σ
TF …
1
,
Each term in the sum of (7.B.3), (7.B.4), and (7.B.5) can be computed to an arbitrary level of precision, see
Rook (2015) which includes the relevant C/C++ source code. As in §VII.A, the corresponding DPs would
use the CDFs of univariate normal mixture PDFs for t other than i,j.
also be generated using simulation.
50
Estimates for these expressions can
VIII. Conclusion
As retirement decumulation models increase in sophistication financial firms may guarantee their
success. The retiree could pay for this as a percentage of funds remaining at death. Decumulation models are
statistical and based on assumptions which if incorrect can render the model unsound. Quantitative mortgage
products sold during the housing boom were priced using cash-flows generated by simulating default times
with a Gaussian copula. In hindsight, the normal assumption was incorrect because default-time correlations
change in a crisis. Since housing booms are followed by housing busts, model assumptions should have
incorporated economic regimes. As our economy transitions from pensions to defined contribution plans,
quantitative retirement products are proliferating. At present, the industry is built on a Gaussian (or lognormal)
foundation which also fails to incorporate regimes or crises when modeling returns and their correlation.
The purpose of this research is to develop a multivariate PDF for asset returns that is suitable for
quantitative retirement plans. The model fits any set of returns, however the curse of dimensionality will
limit the number of securities. We propose a multivariate mixture having fixed mixture marginals using
normal components. The model is motivated by the claim that a lognormal PDF is virtually indistinguishable
from a mixture of normals. Whereas the lognormal PDF is intractable with regard to weighted sums, the
normal mixture is not. The lognormal PDF is only justifiable when short-term returns are iid and the PDF is
CLT-compatible for the given sample size. A typical retiree could endure several market crashes and we
should not expect the historical sample to represent all possible extremes. We can stress test a retirement
plan by subjecting it to a return PDF that has been fit on the historical sample seeded with black swan events.
The normal or lognormal PDF are unhelpful in this regard as neither can accommodate such outliers.
The univariate and multivariate PDFs we have developed fit the historical returns closely and a valid
criticism is that models which memorize the training data project poorly into the future. Adjusting the
variance ratio constraint and/or p-value when bootstrapping the marginal LRTs can loosen the fit. We have
used relatively high values for both. Larger variance ratio constraints and LRT p-values lead to more
marginal peaks and more components. Since the multivariate PDF maintains the marginals, over fitting the
marginals propagates through to the multivariate PDF. The user sets these values as desired. We fit the
multivariate PDF in 3 steps. First generic mixture marginals are derived using the EM algorithm. Second,
the multivariate PDF structure is set using LPs where the number of multivariate regimes is pruned by
penalizing the objective when it includes components with no data. Lastly, covariances are added and
probabilities updated using an ECME likelihood-based approach with the M-step split into convex and
general NLP optimizations. For the NLP we use a Levenberg-Marquardt approach that simulates the step
size and line search parameters while iterating. Lastly, a linear transform on the multivariate PDF forms the
expense-adjusted real compounding return on a diversified portfolio and it is incorporated it into optimal
discrete time retirement decumulation models using both static and dynamic asset allocation glidepaths.
51
References
Akaike, Hirotugu, 1974, A New Look at the Statistical Model Identification, IEEE Transactions on Automatic
Control, Vol. AC-19, No. 6., pp. 716-723.
Anton, Howard, 1988, Calculus with Analytic Geometry 3rd Edition (John Wiley & Sons, Inc., New York,
NY).
Baltussen, Guido, Sjoerd van Bekkum, and Zhi Da, 2016, Indexing and Stock Market Serial Dependence
Around the World, Working Paper Series, <https://ssrn.com/abstract=2786516>.
Bengen, William P., 1994, Determining withdrawal rates using historical data, Journal of Financial
Planning, Vol. 7, No. 4, pp. 171-180.
Border, K. C., 2013, More Than You Wanted to Know About Quadratic Forms (California Institute of
Technology) <http://people.hss.caltech.edu/~kcb/Notes/QuadraticForms.pdf>.
Box, George E. P., Gwilym M. Jenkins, and Gregory C. Reinsel, 1994, Time Series Analysis: Forecasting
and Control 3rd Edition (Prentice-Hall, Inc., Englewood Cliffs, NJ).
Boyd, Stephen and Lieven Vandenberghe, 2004, Convex Optimization (Cambridge University Press, New
York, NY).
Casella, George, and Roger L. Berger, 1990, Statistical Inference (Wadsworth & Brooks/Cole
Statistics/Probability Series, Pacific Grove, CA).
Conway, Deloras A., 1979, Multivariate Distributions with Specified Marginals, Stanford University
Technical Report No. 145.
Dempster, A. P., N. M. Laird, and D. B. Rubin, 1977, Maximum Likelihood from Incomplete Data via the
EM Algorithm, Journal of the Royal Statistical Society, Vol. 39, No. 1, pp. 1-38.
Fama, Eugene F., 1965, The Behavior of Stock-Market Prices, Journal of Business, Vol. 38, Iss. 1, pp. 34105.
Freund, John E., 1992, Mathematical Statistics 5th Edition (Prentice-Hall, Inc., Englewood Cliffs, New
Jersey).
Gavin, Henry P., 2017, The Levenberg-Marquardt Method for Non-Linear Least Squares Curve-Fitting
Problems (Duke University) <http://people.duke.edu/~hpgavin/ce281/lm.pdf>.
Guttman, Irwin, 1982, Linear Models: An Introduction (Wiley Series in Probability and Mathematical
Statistics, New York, NY).
Hamilton, James D., 1989, A New Approach to the Economic Analysis of Nonstationary Time Series and the
Business Cycle, Econometrica, Vol. 57, No. 2: pp. 357-384.
Harlow, W. V., and Keith C. Brown, 2016, Market Risk, Mortality Risk, and Sustainable Retirement Asset
Allocation: A Downside Risk Perspective, Journal of Investment Management, Vol. 14, No. 2, pp. 5-32.
Hillier, Frederick S., and Gerald J. Lieberman, 2010, Introduction to Operations Research 9th Edition
(McGraw-Hill, New York, NY).
Hogg, Robert V., Joseph W. McKean, and Allen T. Craig, 2005, Introduction to Mathematical Statistics
(Pearson Prentice Hall, Upper Saddle River, NJ).
52
Huber, Peter J., 2002, John W. Tukey's Contributions to Robust Statistics, The Annals of Statistics, Vol. 30,
No. 6, pp. 1640-1648.
Hurvich, Clifford M. and Tsai, Chih-Ling, 1989, Regression and Time Series Model Selection in Small
Samples, Biometrika, Vol. 76, No. 2, pp. 297-307.
Jensen, Paul A., and Jonathan F. Bard, 2003, Operations Research – Models and Methods (John Wiley &
Sons, Inc., Hoboken, NJ).
Johnson, Norman L., Samuel Kotz, and N. Balakrishnan, 1994, Continuous Univariate Distributions Volume
1 2nd Edition, (John Wiley & Sons Inc., New York, NY).
Kachani, Soulaymane, 2012, The Housing Bubble and the Financial Crisis in the United States: Causes,
Effects, and Lessons Learned, Industrial Economics (IEOR E4003) Lecture Notes, Columbia University.
Kall, Peter, and Janos Mayer, 2010, Stochastic Linear Programming – Models, Theory, and Computation
2nd Edition (Springer Series in Operations Research and Management Science, New York, NY).
Law, Averill M. and W. David Kelton, 2000, Simulation Modeling and Analysis (McGraw-Hill Series in
Industrial Engineering and Management Science, New York, NY).
Levenberg, Kenneth, 1944, A Method for the Solution of Certain Non-Linear Problems in Least Squares,
Quarterly of Applied Mathematics, Vol. 2, No. 2, pp. 164-168.
Li, David, 2000, On Default Correlation: A Copula Function Approach, The Journal of Fixed Income, Vol.
9, No. 4, pp. 43-54.
Liu, Chuanhai and Donald B. Rubin, 1994, The ECME Algorithm: A Simple Extension of EM and ECM with
Faster Monotone Convergence, Biometrika, Vol. 81, No. 4, pp. 633-648.
Lovász, László and Santosh Vempala, 2006, Fast Algorithms for Logconcave Functions: Sampling,
Rounding, Integration and Optimization, Proceedings of the 47th Annual IEEE Symposium on Foundations
of Computer Science, pp. 57-68.
MacKenzie, D. and Taylor Spears, 2014, ‘The formula that killed Wall Street’: The Gaussian copula and
modeling practices in investment banking, Social Studies of Science, Vol 44, Issue 3, pp. 393-417.
Marquardt, Donald W., 1963, An Algorithm for Least-Squares Estimation of Non-Linear Parameters, Journal
of the Society for Industrial and Applied Mathematics, Vol. 11, No. 2, pp. 431-441.
McLachlan, Geoffrey J., 1987, On Bootstrapping the Likelihood Ratio Test Statistic for the Number of
Components in a Normal Mixture, Appl. Statist., Vol. 36, No. 3, pp. 318-324.
McLachlan, Geoffrey J. and Thriyambakam Krishnan, 2008, The EM Algorithm and Extensions (Wiley
Series in Probability and Statistics, New York, NY).
McLachlan, Geoffrey J. and David Peel, 2000, Finite Mixture Models (Wiley Series in Probability and
Statistics, New York, NY).
Meyer, Carl D., 2000, Matrix Analysis and Applied Linear Algebra (Society for Industrial and Applied
Mathematics (SIAM), Philadelphia, PA).
Milevsky, Moshe A., 2016, It’s Time to Retire Ruin (Probabilities), Financial Analysts Journal, Vol. 72, No.
2, pp. 8-13.
Nau, Robert, 2014, Notes on Non-Seasonal ARIMA Models (Duke University)
<https://people.duke.edu/~rnau/Notes_on_nonseasonal_ARIMA_models--Robert_Nau.pdf>.
53
Nelson, Roger B., 2006, An Introduction to Copulas 2nd Edition (Springer Series in Statistics, New York,
NY).
Nocera, Joe, 2009, Risk Management, The New York Times Magazine,
<http://www.nytimes.com/2009/01/04/magazine/04risk-t.html>.
Paolella, Marc S., 2013, Multivariate Asset Return Prediction with Mixture Models, The European Journal of
Finance, Vol. 21, Iss. 13-14, pp. 1214-1252.
Giordoni, Paolo, Xiuyan Mun, and Robert Kohn, 2009, Flexible Multivariate Density Estimation with
Marginal Adaptation, Working Paper Series, <https://arxiv.org/abs/0901.0225> [stat.ME].
Rabiner, Lawrence R. and Biing-Hwang Juang, 1986, An Introduction to Hidden Markov Models, IEEE
ASSP Magazine, pp. 4-16.
Rook, Christopher J., 2014, Minimizing the Probability of Ruin in Retirement, Working Paper Series,
<http://arxiv.org/abs/1501.00419> [q-fin.GN].
Rook, Christopher J., 2015, Optimal Equity Glidepaths in Retirement, Working Paper Series,
<https://arxiv.org/abs/1506.08400> [q-fin.GN].
Rook, Christopher J. and Mitchell C. Kerman, 2015, Approximating the Sum of Correlated Lognormals: An
Implementation, Working Paper Series, <https://arxiv.org/abs/1508.07582> [q-fin.GN].
Ross, Sheldon M., 2009, Introduction to Probability and Statistics for Engineers and Scientists 4th Edition
(Elsevier, New York, NY).
Salmon, Felix, 2009, Recipe for Disaster: The Formula that Killed Wall Street, Wired.com,
<https://www.wired.com/2009/02/wp-quant/>.
Schwarz, Gideon, 1978, Estimating the Dimension of a Model, The Annals of Statistics, Vol. 6, No. 2, pp.
461-464.
Searle, Shayle R., George Casella, and Charles E. McCulloch, 1992, Variance Components (Wiley Series in
Probability and Mathematical Statistics, New York, NY).
Taleb, Nassim N., 2010, The Black Swan (Random House, New York, NY).
Tao, Jill, 2002, Mixed Models Analysis Using the SAS System (SAS Institute, Cary NC).
Titterington, D. M., A. F. M. Smith, and U. E. Makov, 1985, Statistical Analysis of Finite Mixture
Distributions (Wiley Series in Probability and Mathematical Statistics, New York, NY).
Tran, Minh-Ngoc, Paolo Giordani, Xiuyan Mun, Robert Kohn, and Mike Pitt, 2014, Copula-type Estimators
for Flexible Multivariate Density Modeling using Mixtures, Journal of Computational and Graphical
Statistics, Vol. 23, No. 4, pp. 1163-1178.
Tsay, R. S., 2010, Analysis of Financial Time Series 3rd Edition (John Wiley & Sons, Inc., Hoboken, NJ).
Westfall, Peter H., Randall D. Tobias, Dror Rom, Russell D. Wolfinger, and Yosef Hochberg, 1999, Multiple
Comparisons and Multiple Tests Using SAS (SAS Institute, Cary NC).
Wothke, Werner, 1993, Testing Structural Equation Models (Chapter 11): Nonpositive Definite Matrices in
Structural Modeling (Sage Publications, Inc., Newbury Park, CA).
54
Data Sources
CPI-U (Inflation Rate), Cash (No Interest)
Federal Reserve Bank of Minneapolis, “Consumer Price Index (CPI-U)”, Link:
https://www.minneapolisfed.org/community/teaching-aids/cpi-calculator-information/consumer-price-indexand-inflation-rates-1913, [Accessed: December 19, 2016]
Real S&P 500, Total S&P 500, Real U.S. 10-Year T-Bond, Total U.S. 10-Year T-Bond, Real U.S. 3Month T-Bill
Aswath Damodaran [Updated: January 5, 2016], "Historical Returns on Stocks, Bonds and Bills - United
States", Link: http://pages.stern.nyu.edu/~adamodar/, [Accessed: December 19, 2016], Download:
http://www.stern.nyu.edu/~adamodar/pc/datasets/histretSP.xls
Real Small Cap Equity, Total Small Cap Equity
Roger Ibbotson, Roger J. Grabowski, James P. Harrington, Carla Nunes [September 1, 2016], “2016 Stocks,
Bonds, Bills, and Inflation (SBBI) Yearbook”, John Wiley and Sons, Link:
http://www.wiley.com/WileyCDA/WileyTitle/productCd-1119316405.html, [Accessed: December 19, 2016]
Yearly Shiller CAPE Ratio (January), S&P 500 Earnings (January)
Robert Shiller “Online Data Robert Shiller: U.S. Stock Markets 1871-Present and CAPE Ratio”, Link:
http://www.econ.yale.edu/~shiller/data.htm, [Accessed: December 19, 2016], Download:
http://www.econ.yale.edu/~shiller/data/ie_data.xls
Gold Returns
Kitco Metals [2016], “Historical Gold Prices Gold 1833-1999 London PM Fix US Dollars”, Link:
http://corp.kitco.com/en/index.html, [Accessed: December 19, 2016], Download:
http://www.kitco.com/scripts/hist_charts/yearly_graphs.plx
Retiree Surveys
Steve Vernon, December 26, 2016, “The Top Retirement Fears, and How to Tackle Them”, CBS News
MoneyWatch, Link: http://www.cbsnews.com/news/the-top-retirement-fears-and-how-to-tackle-them/,
[Accessed: January 15, 2017]
Lea Hart, October 6, 2016, “American’s biggest retirement fear: Running out of Money”, Journal of
Accountancy, Link: http://www.journalofaccountancy.com/news/2016/oct/americans-fear-running-out-ofretirement-money-201615242.html, [Accessed: January 15, 2017]
Robert Brooks, July 18, 2016, “A Quarter of Americans Worry About Running Out of Money in Retirement”,
The Washington Post, Link: https://www.washingtonpost.com/news/get-there/wp/2016/07/18/a-quarter-ofamericans-worry-about-running-out-of-money-in-retirement/?utm_term=.7baa6479c08b, [Accessed: January
15, 2017]
Prudential Investments, 2016, “Perspectives on Retirement: 2016 Retirement Preparedness Survey
Findings”, Prudential Financial Inc., Link: https://investment.prudential.com/util/common/get?file=
18CAC0B95B49D959852580790076AF83, [Accessed: January 15, 2017]
Emily Brandon, March 25, 2011, “Baby Boomers Reveal Biggest Retirement Fears”, US News & World
Report, Link: http://money.usnews.com/money/blogs/planning-to-retire/2011/03/25/baby-boomers-revealbiggest-retirement-fears, [Accessed: January 15, 2017]
55
IX. Appendix with Source Code
Appendix A. Proof of Unbounded Likelihood for Normal Mixture PDF
Let Xt
~
f(x) = π1 f 1(x) + π2 f 2(x) + ... + πg f g(x), where f(x) is a g-component mixture PDF with f i(x)
~ N(μi,σi), i=1,2,...,g, t=1,2,...,T, and suppose that an iid sample of size T has been observed as (x1, x2, ..., xT).
The normal component PDFs are given by:
1
,
√2 σ
∞
∞, i
(9.A.1)
1,2, … , g
The vector of unknown parameters for f(x) = f(x,θ) is defined as θ=(π,μ,σ)′ where π=(π1,...,πg), μ=(μ1,...,μg),
(1)
and =(σ1,...,σg). To obtain an arbitrarily large value for
(θ| ) from (2.E.1), take a single component and
dedicate it to one observation. For example, consider component k, f k(x), and observation j, xj. Let ε be an
arbitrarily small number and set πk = 1/T, μk = xj - ε, and σk = ε. The value of f k(xj) is:
∝
1
ε
∝
Recall that we initialize θ to begin the EM-algorithm and
(1)
1
ε
(9.A.2)
(θ| ) is guaranteed to increase at each iteration.
From (9.A.2) above, as ε → 0, f k(xt) → ∞. Therefore by choosing k, j, and ε we can initialize f k(xt) to an
arbitrarily large number making
(1)
(θ| ) unbounded. Such a solution for θ, however, is not meaningful as it
dedicates a near-degenerate component to a single observation. Similarly, a component that is trapped into
fitting a small number of closely clustered observations also leads to a high likelihood value (due to the small
variance) and is referred to as a spurious maximizer of
(1)
(θ| ) when the solution is not meaningful. These
should be identified and removed if they are not the MLE, . To avoid manually evaluating each
(1)
(θ| ) for
spuriousness, we impose a variance ratio constraint as noted in (§II.E). This prevents any single variance
from becoming too small and eliminates both problems noted above (McLachlan & Peel, 2000).
Appendix B. Diagnostic Plots for 18 Finance/Economics Time Series
Diagnostic plots for the 18 finance/economics time series analyzed in §II.J Table I are presented
here. Each time series includes a plot of the uncentered raw observations, Xt, t=1,2,...,T as well as the ACF
and PACF up to N/4 lags, where N = total # data points for the series (Box, et al., 1994). Annual values
from 1928 – 2015 are used therefore N = 88 for all series that are not differenced or averaged. The sources
for all data can be found in the Data Sources section located after References in the main paper. Each test is
formed as Ho: No Serial Correlation vs Ha: Serial Correlation, with the p-value provided on the PACF plot.
Preliminary conclusions about the behavior of each process are supplied in Table I followed by a discussion
of the appropriate family-wise α* to account for multiplicity. All security returns are annual compounding.
56
1.) CPI-U (Inflation Rate)
2.) Real Compounding S&P 500
3.) Total Compounding S&P 500
57
4.) Real Compounding Small Cap Equity
5.) Total Compounding Small Cap Equity
6.) Real Compounding U.S. 10-Year T-Bond
58
7.) Total Compounding U.S. 10-Year T-Bond
8.) Real Compounding U.S. 3-Month T-Bill
9.) Real Compounding Gold Return
59
10.) Cash Compounding (No Interest)
11.) Real S&P 500-to-Bond Risk Premium
12.) Real S&P 500-to-Small Cap Risk Premium
60
13.) Real Small Cap-to-Bond Risk Premium
14.) 10-Year Avg. Real Compounding S&P 500
15.) 10-Year Avg. Total Compounding S&P 500
61
16.) 1st Diff. Shiller CAPE Ratio
17.) 1st Diff. Log Shiller CAPE Ratio
18.) Detrended S&P 10-Year Avg. Real Earnings
62
Appendix C. C/C++ Source Code
The application accepts 2 input files: 1.) a control file with settings, and 2.) a text file of returns.
Samples of each are shown below. The control file sets 7 parameters with the 1st and 2nd being the # of assets
and # of time points, respectively. The 3rd parameter sets the # of random starts per CPU core for each
execution of the EM algorithm when fitting univariate PDFs. The value is multiplied by 1 less than the # of
components in the mixture being fit. For example, if fitting a 4-component univariate mixture the setting
below results in 5·3 random starts/core. Since our machine contains 40 cores this mixture PDF is fit using 600
EM algorithm random starts. The 4th parameter is the maximum # of components for all univariate mixtures
which we set to 5 in this implementation. The 5th parameter is the # of samples to use for each bootstrapped
LRT within the proposed forward-backward framework that determines the optimal # of univariate mixture
components for each asset. The final 2 values in “control.txt” are the forward-backward LRT p-values,
respectively. The returns file contains a column of (compounding) returns for each asset. On our computer the
control file shown can reproduce the multivariate PDF for (L, S, B) from (6.A.2) in a few minutes.
control.txt
returns.txt
3 88 5 5 100 0.25 0.25
1.45557
0.91700
0.76958
0.61647
1.41387
0.48640
0.63566
0.55159
1.02065
1.04200
1.07441
1.06959
Both univariate and multivariate parameter estimation results are written to the screen and to the file
named “output.txt” in the folder supplied by the user upon application launch. The 2 files above must also exist
within this directory. Lastly, a folder to contain error files for issues encountered during the optimization must
exist and be specified by the user using the global constant “errFolder” which is also defined in the header file.
To save time, processing for both the univariate and multivariate PDF is multithreaded.
Our
implementation of the EM algorithm launches each random start in a separate thread (random starts end at
local optimums). The ECME algorithm takes a small or large step in the general direction of steepest ascent
while maximizing the multivariate log-likelihood function with respect to the covariance parameters, see
§V.A. The line search and stepping parameters are simulated and values that yield the largest increase are
used, which is also multithreaded. Note that our ECME implementation uses the actual or incomplete loglikelihood function from (2.E.2), conditioning first on the probabilities holding the covariances constant, and
then on the covariances holding the probabilities constant. These are Steps 1 and 2 in (5.A.1) and (5.A.2).
This application uses the Boost (http://www.boost.org/), LPSolve (http://lpsolve.sourceforge.net/5.5/),
and Eigen (http://eigen.tuxfamily.org/) external C/C++ libraries, which are freely available to download under
terms and conditions described at the given sites. Our code consists of a header file and 20 functions and is
being provided under the GNU General Public License, see http://www.gnu.org/licenses/ for full details.
63
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: stdafx.h
/
/ Summary:
/
/
This is the header file which is included before each function call to define needed C/C++ libraries, global constants, inline functions, and function prototypes. For function prototypes, see
/
headers attached to the code definition for a description of the purpose and parameters.
/
/ Inline Functions:
/
/
getNDens:
Compute/return the supplied normal density function evaluated at a given point, without the constant 1/sqrt(2*pi).
/
getMDens:
Compute/return the supplied normal mixture density function evaluated at a given point, without the constant 1/sqrt(2*pi).
/
getPost:
Compute/return the probability that a given value belongs to a given component of the supplied normal mixture density.
/
getUniMean:
Compute/return the mean of a given set of observations.
/
getUniStd:
Compute/return the standard deviation (MLE) of a given set of observations.
/
getLLVal:
Compute/return the log‐likelihood value for a given observation set from the supplied univariate mixture density.
/
getVRatio:
Compute/return the ratio of the largest‐to‐smallest variance for a supplied set of standard deviations.
/
getEMMean:
Compute/return mean update for the EM algorithm along with a sum‐of‐squares quantity needed for the EM algorithm variance update.
/
getMVNDens:
Compute/return the supplied multivariate normal density function evaluated at a given vector of values for each random variable.
/
getMVNMDens:
Compute/return the supplied multivariate normal mixture density function evaluated at a given vector of values for each random variable.
/
getCovs:
Accept a multivariate normal mixture density and return the vector of covariances starting with VC element (1,2) of component #1, then element (1,3), etc...
/
setCovs:
Accept a vector of all covariances and insert the values into a single component's VC matrix as specified by the user.
/
getCofM:
Accept a matrix and element/position and convert it to one who's determinant is equal to the cofactor of the matrix with respect to that position/position.
/
getIDM:
Accept an empty square matrix and populate it with an identity matrix of the same size.
/
chkSum:
Accept a multivariate normal mixture density and check that no component has a likelihood of zero at all time points.
/
showVals:
Display the parameters (means, standard deviations, component probabilities) of a supplied multivariate mixture density to standard output.
/*********************************************************************************************************************************************************************************************************/
#pragma once
// Include libraries.
//=====================
#include "targetver.h"
#include <stdio.h>
#include <iostream>
#include <iomanip>
#include <string>
#include <fstream>
#include <random>
#include <boost/algorithm/string.hpp>
#include <boost/thread/thread.hpp>
#include <Eigen/Core>
#include <Eigen/Dense>
#include <Eigen/LU>
#include <Eigen/Eigenvalues>
#include "lp_lib.h"
using namespace std;
// Global constants.
//====================
const string cfile="control.txt";
const string rfile="returns.txt";
const string ofile="output.txt";
const string errFolder="c:\\mDensity\\Issues\\";
const long double pi=3.141592653589793;
const long double logrecipsqrt2pi=‐0.918938533204673;
const long double stdRatio=16.0;
const long double epsilon=pow(0.1,15);
const long double LNegVal=‐pow(9.0,10);
const long double LPosVal=pow(9.0,10);
//
//
//
//
//
//
//
//
//
//
Name of parameter control input file.
Name of input data file.
Name of output file.
Folder where issues/errors will be written to files (mandatory).
Constant representation of pi.
ln(1.00/square root of 2*pi).
Square root of the variance ratio constraint threshold.
Convergence criteria.
Large negative value for use as an invalid value indicator.
Arbitrarily large positive value.
64
const
const
const
const
const
const
const
const
const
const
const
const
const
const
long double rrMultMin=2.0;
long double rrMultMax=10.0;
long double pdmineval=pow(0.1,13);
long double detminval=pow(0.1,13);
int mIters=1000000000;
int K=1000000;
int dLvl=500;
long double minHessAdd=0.00;
int mItersH=1000;
int nCorMult=2;
int nBeats=80;
int dbug=0;
int cDown=2000;
int nECMEs=1;
//
//
//
//
//
//
//
//
//
//
//
//
//
//
Minimum multiple when performing a ridge repair on a broken variance‐covariance matrix.
Maximum multiple when performing a ridge repair on a broken variance‐covariance matrix.
Threshold for minimum eigenvalue to ensure a positive definite variance‐covariance matrix.
Threshold for determinant to ensure a positive definite variance‐covariance matrix.
Maximum # of EM iterations allowed per single optimization.
Large constant for use in LP objective function to enforce feasibility constraint.
Discretization level for separable quadratic objective function in solveLP().
Minimum additive factor to use for Hessian stepping.
Maximum # Hessian steps per thread. (Set to 1 for non‐random ECME stepping.)
Multiply by # cores to determine the total # threads used for Hessian stepping. (Set to 1 for non‐random ECME stepping.)
# of log‐likelihood beats to randomly select from during the ECME algorithm 2nd Step. (Set to 1 for non‐random ECME stepping.)
Debug level (0,1,2) for output window details (higher # = more details).
Processor cool down time in between EM iterations in msec (value of 100000 = 100 seconds).
Number of outer loops for ECME processing. (# of times to repeat the ECME procedure.)
// Inline functions.
//====================
inline long double getNDens(const long double val, const long double mn, const long double std)
{return (1.00/std)*exp(‐0.5*pow((val‐mn)/std,2));}
inline long double getMDens(const long double val, const int g, const long double **inMDist)
{long double dval=0.00; for (int c=0; c<g; ++c) dval=dval + inMDist[0][c]*getNDens(val,inMDist[1][c],inMDist[2][c]); return dval;}
inline long double getPost(const long double val, const int g, const long double **inMDist, const int cID)
{return inMDist[0][cID]*getNDens(val,inMDist[1][cID],inMDist[2][cID])/getMDens(val,g,inMDist);}
inline long double getUniMean(const int T, const long double *r)
{long double mn=0.00; for (int t=0; t<T; ++t) mn=mn + r[t]; return mn/T;}
inline long double getUniStd(const int T, const long double *r, const long double mn)
{long double var=0.00; for (int t=0; t<T; ++t) var=var + pow(r[t]‐mn,2); return sqrt(var/T);}
inline long double getLLVal(const int T, const long double *r, const int g, const long double **inMDist)
{long double llval=((long double) T)*logrecipsqrt2pi; for (int t=0; t<T; ++t) llval=llval + log(getMDens(r[t],g,inMDist)); return llval;}
inline long double getVRatio(const int g, const long double *stds)
{long double minstd=LPosVal, maxstd=LNegVal; for (int c=0; c<g; ++c){if (stds[c] < minstd) minstd=stds[c]; if (stds[c] > maxstd) maxstd = stds[c];}
return pow(maxstd/minstd,2);}
inline long double getEMMean(const int T, const long double *r, const long double *pprbs, const long double cprbs, long double *ssq)
{long double mn=0.00, atrm; ssq[0]=0.00; for (int t=0; t<T; ++t) {atrm=pprbs[t]*r[t]; mn=mn+atrm; ssq[0]=ssq[0]+atrm*r[t];}; return mn/(cprbs);}
inline long double getMVNDens(const Eigen::VectorXd vals, const Eigen::VectorXd mn, const Eigen::MatrixXd vcmi, const long double sqrdet, const long double picnst)
{return (sqrdet/picnst)*exp(‐0.5*((vals‐mn).transpose())*vcmi*(vals‐mn));}
inline long double getMVNMDens(const int uCells, const Eigen::VectorXd vals, const Eigen::VectorXd *mns, const Eigen::MatrixXd *vcmis, const Eigen::VectorXd prbs, const long double *sqrdets,
const long double picnst){long double dval=0.00; for (int u=0; u<uCells; ++u) {dval=dval+prbs[u]*getMVNDens(vals,mns[u],vcmis[u],sqrdets[u],picnst);} return dval;}
inline void getCovs(const int uCells, const Eigen::MatrixXd *inVCs, Eigen::VectorXd *inDVarsM){int itr=0; for (int v=0; v<uCells; ++v) for (int r=0; r<(int) inVCs[v].rows(); ++r)
for (int c=r+1; c<(int) inVCs[v].cols(); ++c)
inDVarsM[0](itr++)=inVCs[v](r,c);}
inline void setCovs(const int uCell, Eigen::MatrixXd *inVCs, const Eigen::VectorXd *inDVarsM){int itr=uCell*((int) (((int) inVCs[uCell].rows())*((int) inVCs[uCell].rows()‐1))/2);
for (int r=0; r<(int) inVCs[uCell].rows(); ++r)
for (int c=r+1; c<(int) inVCs[uCell].cols(); ++c)
{inVCs[uCell](r,c)=inDVarsM[0](itr++);
inVCs[uCell](c,r)=inVCs[uCell](r,c);}}
inline void getCofM(const int numA, const int inr, const int inc, const Eigen::MatrixXd inE, Eigen::MatrixXd *inEjk){inEjk[0]=inE; for (int r=0; r<numA; ++r)
for (int c=0; c<numA; ++c)
{if (r==inr && c==inc) inEjk[0](r,c)=1.00;
else if (r==inr || c==inc) inEjk[0](r,c)=0.00;}}
inline void getIDM(Eigen::MatrixXd *ID){for (int r=0; r<(int) ID[0].rows(); ++r) for (int c=r; c<ID[0].cols(); ++c)
{if (r==c) ID[0](r,c)=1.00; else ID[0](c,r)=ID[0](r,c)=0.00;}}
65
inline void chkSum(const int T, const int nUCmps, const long double **fvals){for (int v=0; v<nUCmps; ++v){long double tmpsum=0.00; for (int t=0; t<T; ++t) tmpsum=tmpsum + fvals[t][v];
if (tmpsum <= 0.00){cout << endl << "ERROR: Unique component # " << v << " has a likelihood that is zero for each time point." << endl
<< "
This will eliminate the corresponding component probability from the stage 2 objective function." << endl
<< "
The component probability should be treated as a constant and moved to the RHS constraint vector" << endl
<< "
and eliminated from the objective function. The code for this has not yet been implemented." << endl
<< "EXITING...ECMEAlg()..." << endl; cin.get(); exit (EXIT_FAILURE);}}}
inline void showVals(const int T, const long double *r, const int g, const long double **inMDist, ostream& oVar=cout)
{oVar.setf(ios_base::fixed, ios_base::floatfield); oVar.precision(16); for (int x=0; x<T; ++x) oVar << r[x] << endl; for (int c=0; c<g; ++c){oVar << string(37,' ')
<< "Prob[" << c << "]=" << inMDist[0][c] << " Mean[" << c << "]=" << inMDist[1][c] << " Std.Dev.[" << c << "]=" << inMDist[2][c] << endl;}};
// Function prototypes.
//=======================
int fitMixDist(const int a, const int T, const long double *r, const int maxcmps, const int nsmpls, const int nstrts, const long double sl[2], long double **fnlMDst, string rdir);
void getRVals(const int N, const int g, const long double **inMDist, long double *rvls);
void getRPrbsStds(const int T, const long double * r, const int g, long double *prbs, const long double *mns, long double *stds);
void EMAlg(const int T, const long double *r, const int g, long double **prbs, long double **mns, long double **stds, long double *llVal, const long double **inMDist, int *rprms);
int ECMEAlg(const int T, const long double **r, const int numA, const int nUCmps, const Eigen::MatrixXd cMtrx, const Eigen::VectorXd cVctr, long double *muPrbs, Eigen::VectorXd *muMns, Eigen::MatrixXd
*muVCs, int *uCellIDs, const string rdir);
long double ThrdEMAlg(const int T, const long double *r, const int rs, const int ing, const long double **inMDist, const int outg, long double **outMDist);
void mapCells(const int totCells, const int numA, int **inCellAry, const int curAst, const int *inCmps, int *cID, int *tmpAry);
int getCell(const int **inCellAry, const int totCells, const int *cmpLvls, const int numA);
void asgnObs(int *inAry, const int T, const long double *r, const int g, const long double **inMDst);
void getCor(const int T, const long double **r, const int **asgn, const int **inCellAry, const int cellID, Eigen::VectorXd Mn, Eigen::MatrixXd *VC, const int vCell);
int solveLP(const int totCells, const int numA, const int **inCellAry, const int *cmps, const long double **prbs, const int *nCellObs, const long double *cellProb, double *outPrbs, const int type);
void getCMtrx(const int totrows, const int nUCmps, const int sol, const int numA, const int *nCmps, const int *vCIDs, const int **inCellAry, const long double **prbs, Eigen::MatrixXd *fLHS,
Eigen::VectorXd *fRHS);
long double getHessE(const int T, const int inUCmps, const long double **infVals, const long double *inDNoms, const Eigen::MatrixXd inCMtrx, const Eigen::VectorXd inCVctr, Eigen::MatrixXd *inHess,
Eigen::MatrixXd *inLHS, Eigen::VectorXd *inRHS);
long double getHessM(const int T, Eigen::VectorXd *rs, const int inUCmps, const int numA, const long double **infVals, const long double *inDNoms, const long double *inPrbs, const Eigen::VectorXd *muMns,
const Eigen::MatrixXd *E, const Eigen::MatrixXd *Einv, const Eigen::MatrixXd *inA, Eigen::MatrixXd *inHess);
void getGradE(const int T, const int inUCmps, const long double **infVals, const long double *inDNoms, const Eigen::MatrixXd inLHS, const Eigen::VectorXd inRHS, const Eigen::VectorXd inDVars,
Eigen::VectorXd *inGrad);
void getGradM(const int T, Eigen::VectorXd *rs, const int inUCmps, const int numA, const long double **infVals, const long double *inDNoms, const long double *inPrbs, Eigen::VectorXd *muMns, const
Eigen::MatrixXd *E, const Eigen::MatrixXd *Einv, const Eigen::MatrixXd *inA, Eigen::VectorXd *inGrad, int chk=0);
long double getLFVals(const int T, const int numA, const int nUCmps, const Eigen::VectorXd *rs, const Eigen::VectorXd uPrbs, const Eigen::VectorXd *inMns, const Eigen::MatrixXd *inVCIs, const long double
*insqs, const long double inpicst, long double *denoms, long double **lfVals);
void wrtDens(const string typ, const int nUCmps, const int *uCells, const long double *muPrbs, const Eigen::VectorXd *muMns, const Eigen::MatrixXd *muVCs, ostream& oVar);
void stepHessM(int *inHess1, long double *inHess2, const Eigen::VectorXd *rs, Eigen::VectorXd *inDvars, const Eigen::VectorXd inGrad, const Eigen::MatrixXd inHess, const Eigen::VectorXd uPrbs, const
Eigen::VectorXd *inMns, const Eigen::MatrixXd *inVCs);
int ridgeRpr(const int uCell, Eigen::MatrixXd *E, long double mult);
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: mDensity.cpp
/
/ Function: main()
/
/ Summary:
/
/
This function defines the entry point for the console application and drives the analysis using 4 major sections. Section #1.) The contents of the control file (control.txt: see global constant
/
cfile in the header program), and the data file (returns.txt: see global constant rfile in the header program) are read in and stored as variables. Section #2.) Build the univariate mixture PDFs
/
for each asset using the EM algorithm with random starts and bootstrapped likelihood ratio test for determining the optimal # of components for each asset. Section #3.) Combine the univariate
/
mixtures into a multivariate mixture PDF without disturbing the marginals and without correlations. At this point, a multivariate mixture PDF is estimated with dependence but without regime
/
correlations. (There are 2 multivariate mixture densities at this stage, one for each type of LP solved. The 2 LP objectives for building the multivariate mixture PDF are minimax and minimum
/
squared distance.) Section #4.) Use an ECME type algorithm to estimate the correlations and refine the component probabilities with maximum likelihood as the objective. This 4th step is a 2‐step
/
iterative procedure similar to the EM algorithm and random starts are used in the 2nd step where the covariances are estimated. The 4th step is repeated nECMEs times and the optimal multivariate
/
mixture PDF for each is written to the output file. Once all 4 steps have been completed there are 2*nECMEs multivariate mixture densities with fixed marginals and one is chosen based on some
/
criteria such as higher likelihood or higher information criteria. This final step is left to the user (we use AIC) and the decision would account for the likelihood value as well as the total # of
/
parameters. All PDFs and their log‐likelihood values are written to the output file (output.txt: see global constant ofile in the header program), along with the univariate mixture PDF details.
66
/
/ Inputs:
/
/
No input arguments are processed by this function. Critical inputs are supplied via the control.txt file (see global constant cfile in the header program), and data is supplied via returns.txt (see
/
global constant rfile in the header program). In addition to control.txt, other inputs are set to global constants in the header file.
/
/ Outputs:
/
/
This function writes details of fitting univariate & multivariate mixture PDFs for the supplied assets to the screen and to the file output.txt (see ofile in the header program).
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int main(int argc, char *argv[])
{
// Declare/initialize local variables.
//======================================
string rootdir;
int nBoots, nTPoints, nAssets, nRStarts, mComps, *nComps=nullptr, **asgnmnt=nullptr, uCell;
long double **rtrn=nullptr, **prob=nullptr, **mean=nullptr, **stdev=nullptr, alpha[2], **optMDst=new long double *[3];
ofstream fout;
// Ensure that an error/issues folder has been provided. (Set in the header file.)
//===================================================================================
if ("" == errFolder)
{
cout << "ERROR: An error/issues folder has not been provided. Problems found during optimization will be written to files in this folder." << endl
<< "
Use global variable errFolder in the header file to set this destination." << endl << "EXITING...main()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Retrieve directory location of setup files.
//==============================================
cout << "Enter the directory where the setup files reside (eg, c:\\mDensity\\): " << endl; cin >> rootdir; cin.get(); boost::algorithm::trim(rootdir); cout << endl;
// Read in control file which contains:
//
1.) # of asset classes.
//
2.) # of timepoints with return data.
//
3.) # of random starts for maximizing the likelihood of a g‐component univariate mixture, as a multiple of # independent processing units and components (i.e., value of 2 with 20 independent
//
processing units will use 2x20x1=40 random starts for a 2‐component mixture, 2x20x2=80 random starts for a 3‐component mixture, and 2x20x(g‐1) random starts for a g‐component mixture.)
//
4.) Maximum # of components appropriate for this data set when fitting univariate mixture densities (i.e., the marginals).
//
5.) Sample size to use when bootstrapping the LRT test statistic.
//
6.) Alpha‐level for LRTs for the # of components using a forward‐backward selection algorithm, where the 1st alpha is used for forward selection and the 2nd alpha is used for backward selection.
//=====================================================================================================================================================================================================
ifstream getparams(rootdir+cfile);
if (getparams.is_open())
{
getparams >> nAssets >> nTPoints >> nRStarts >> mComps >> nBoots >> alpha[0] >> alpha[1]; getparams.close();
}
else
{
cout << "ERROR: Could not open file: " << rootdir + cfile << endl << "EXITING...main()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Instantiate 2‐dimensional arrays to hold means, standard deviations, and proportion weights for mixtures.
//============================================================================================================
nComps = new int [nAssets];
// Array to hold # components for each asset.
rtrn = new long double *[nAssets];
// One return per asset and time point up to time nTPoints.
prob = new long double *[nAssets];
// One prob per asset and component for component sizes up to mComp.
mean = new long double *[nAssets];
// One mean per asset and component for component sizes up to mComp.
stdev = new long double *[nAssets];
// One stdev per asset and component for component sizes up to mComp.
asgnmnt = new int *[nAssets];
// One assignment per asset and time point up to time nTPoints.
for (int a=0; a<nAssets; ++a)
{
nComps[a] = 1;
// Start with 1 component for each asset.
rtrn[a] = new long double [nTPoints];
// Array of returns for each asset. Returns are then rtrns[a][0], rtrns[a][1], etc...
asgnmnt[a] = new int [nTPoints];
// Array of component assignments for each asset after density has been determined.
}
67
// Read in returns file which has a column of returns for each asset, and store in a 2‐dimensional array.
//=========================================================================================================
ifstream getrtrns(rootdir+rfile);
if (getrtrns.is_open())
{
int r=0;
while (!getrtrns.eof() && r < nTPoints)
{
for (int a=0; a<nAssets; ++a)
getrtrns >> rtrn[a][r];
r=r+1;
}
getrtrns.close();
if (r < nTPoints)
{
cout << "ERROR: File " << rootdir + rfile << " should have " << nTPoints << " rows of returns, for " << nAssets << " assets, but it has fewer." << endl
<< "EXITING...main()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
}
else
{
cout << "ERROR: Could not open file: " << rootdir + rfile << endl << "EXITING...main()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Build a large array to temporarily hold the optimal distribution for each asset.
//===================================================================================
for (int m=0; m<3; ++m)
optMDst[m]=new long double [mComps];
// Calculate the component probability, mean and standard deviation (Normal MLE version). (Note: MLE standard deviation divides by n, not n‐1 and is a biased estimator.)
//===========================================================================================================================================================================
for (int a=0; a<nAssets; ++a)
{
// Initialize all probabilities in the large array to zero.
//===========================================================
for (int c=0; c<mComps; ++c)
optMDst[0][c]=0.00;
// Find the best fitting mixture distribution for this asset.
//=============================================================
nComps[a]=fitMixDist(a,nTPoints,rtrn[a],mComps,nBoots,nRStarts,alpha,optMDst,rootdir);
// Assign each observation to a component. The most likely one using Bayes rule.
//=================================================================================
asgnObs(asgnmnt[a],nTPoints,rtrn[a],nComps[a],(const long double **) optMDst);
// Build arrays to hold the optimal solution and transfer it to these arrays.
//=============================================================================
prob[a]=new long double [nComps[a]];
mean[a]=new long double [nComps[a]];
stdev[a]=new long double [nComps[a]];
for (int c=0; c<nComps[a]; ++c)
{
prob[a][c]=optMDst[0][c];
mean[a][c]=optMDst[1][c];
stdev[a][c]=optMDst[2][c];
}
}
// Write out the assignment of each observation time point to the corresponding components during debug mode.
//=============================================================================================================
if (dbug >= 2)
{
cout << endl << string(97,'=') << endl << "Assignment of observations (at each time point) to a set of components using Bayes Decision Rule:" << endl << string(97,'=');
for (int t=0; t<nTPoints; ++t)
68
{
cout << endl << "Time t = " << setfill('0') << setw(to_string((long long) nTPoints).size()) << t << ":";
for (int a=0; a<nAssets; ++a)
cout << " Asset #" << a+1 << "/Component=" << asgnmnt[a][t];
} cout << endl << endl;
}
// Mixture distribution has been fit for all assets. Assemble the multivariate density. Start by computing the total # of cells
// that need to be mapped (dealing with a k‐dimensional cube). Also compute the total # of components, summing across all assets.
//==================================================================================================================================
int nCells=1, totCmps=0;
for (int a=0; a<nAssets; ++a)
{
nCells = nCells*nComps[a];
totCmps = totCmps + nComps[a];
}
// Initialize variables.
//========================
int **allCells=new int *[nCells], cellID=0, *tmpVals=new int [nAssets];
for (int i=0; i<nCells; ++i)
allCells[i] = new int [nAssets];
// Call function to map each cell of the k‐dimensional cube to a single list value.
//===================================================================================
mapCells(nCells, nAssets, allCells, 0, nComps, &cellID, tmpVals);
// Derive the unique cell ID for each time point and count the # of obs per unique cell.
//========================================================================================
int *tmpCombo=new int [nAssets], *cellAsgn=new int [nTPoints], *cellCnt=new int [nCells];
long double *cellPrb=new long double [nCells];
for (int c=0; c<nCells; ++c)
cellPrb[c]=(long double) (cellCnt[c]=0);
if (dbug>=2)
cout << endl << endl << string(40,'=') << endl << "Assignment of each time point to a cell:" << endl << string(40,'=') << endl;
for (int t=0; t<nTPoints; ++t)
{
if (dbug >= 2)
cout << "Time t = " << setfill('0') << setw(to_string((long long) nTPoints).size()) << t << ": ";
for (int a=0; a<nAssets; a++)
tmpCombo[a]=asgnmnt[a][t];
cellAsgn[t]=getCell((const int **) allCells, nCells, tmpCombo, nAssets);
cellCnt[cellAsgn[t]]=cellCnt[cellAsgn[t]] + 1;
cellPrb[cellAsgn[t]]=cellPrb[cellAsgn[t]] + 1.00/((long double) nTPoints);
}
// Build and solve the corresponding LP that determines the structure of the multivariate density using both a minimax and minimum squared distance objective. Then build an array of cell IDs that
// have non‐zero probabilities attached using each method. These are the cells we must derive a correlation for, ensuring that the resulting VC matrix is positive‐definite. In solveLP(), type is
// defined as: (These become the index values for all arrays that hold results from both methods to compare and select the better performer.)
//
0.) Minimax objective
//
1.) Minimum sum of squared distances objective
//=====================================================================================================================================================================================================
string type[2]; type[0]="Minimax"; type[1]="Minimum Sum of Squared Distances (SSD)";
int numCmps [2], **valcIDs=new int *[2], ctr;
double **estPrbs=new double *[2];
Eigen::MatrixXd *mLHS=new Eigen::MatrixXd [2];
Eigen::VectorXd *mRHS=new Eigen::VectorXd [2];
cout << endl << string(120,'=') << endl << "Running " << type[0] << " and " << type[1] << " LPs to set the initial multivariate density structure." << endl << string(120,'=') << endl << endl;
for (int i=0; i<2; ++i)
{
// Use an LP to approximate the structure of the multivariate density.
//======================================================================
estPrbs[i]=new double [nCells];
69
numCmps[i]=solveLP(nCells,nAssets,(const int **) allCells,nComps,(const long double **) prob,cellCnt,cellPrb,estPrbs[i],i);
valcIDs[i]=new int [numCmps[i]];
// Identify and store the unique cells with non‐zero probabilities.
//===================================================================
ctr=0;
for (int c=0; c<nCells; ++c)
if (estPrbs[i][c] > 0.00)
valcIDs[i][ctr++]=c;
// Retrieve the effective constraint matrix that applies to the LP solution and is full row rank. Ensure it does not have more rows than non‐zero components.
//==============================================================================================================================================================
getCMtrx(totCmps‐nAssets+1,numCmps[i],i,nAssets,nComps,valcIDs[i],(const int **)allCells,(const long double **)prob,mLHS,mRHS);
cout << "‐‐‐‐‐> " << type[i] << " Done. There are " << numCmps[i] << " unique cells with non‐zero probabilities using " << type[i] << "." << endl;
}
// Assemble multivariate densities using the minimax and minimum squared distance objectives as starting points and select the one with a better fit after
// iterating with the ECME algorithm until convergence. (Where the 1st Step is replaced by a convex programming problem.) Better fit means higher likelihood.
//=============================================================================================================================================================
int *numCmpsF=new int [2*nECMEs], **vCIDs=new int *[2*nECMEs];
long double **mProbs=new long double *[2*nECMEs];
Eigen::VectorXd **mMeans=new Eigen::VectorXd *[2*nECMEs];
Eigen::MatrixXd **mVCs=new Eigen::MatrixXd *[2*nECMEs];
for (int i=0; i<2*nECMEs; ++i)
{
// Initialize the arrays to their proper size based on the # of components resulting from the above LPs.
//========================================================================================================
mProbs[i]=new long double [numCmps[(i%2)]];
mMeans[i]=new Eigen::VectorXd [numCmps[(i%2)]];
mVCs[i]=new Eigen::MatrixXd [numCmps[(i%2)]];
vCIDs[i]=new int [numCmps[(i%2)]];
// Initialize each type and cell specific mean vector and VC matrix to the appropriate size, which is the # of assets.
//======================================================================================================================
for (int v=0; v<numCmps[(i%2)]; ++v)
{
mMeans[i][v]=Eigen::VectorXd(nAssets);
mVCs[i][v]=Eigen::MatrixXd(nAssets,nAssets);
vCIDs[i][v]=valcIDs[(i%2)][v];
}
// Populate probabilities, means, and the initial variance‐covariance matrix for each valid cell. Off‐diagonal elements (i.e., covariances) are set to zero at this stage.
//===========================================================================================================================================================================
for (int v=0; v<numCmps[(i%2)]; ++v)
{
uCell=vCIDs[i][v];
mProbs[i][v]=estPrbs[(i%2)][uCell];
for (int a1=0; a1<nAssets; ++a1)
{
mMeans[i][v](a1)=mean[a1][allCells[uCell][a1]];
for (int a2=0; a2<nAssets; ++a2)
{
if (a1 == a2)
mVCs[i][v](a1,a2)=pow(stdev[a1][allCells[uCell][a1]],2);
else
mVCs[i][v](a1,a2)=0.00;
}
}
}
// If debugging mode is on, write out the corresponding parameters of the distribution. These are the initial LP values before improving with covariances.
//===========================================================================================================================================================
if (dbug >= 2)
wrtDens(type[(i%2)],numCmps[(i%2)],vCIDs[i],mProbs[i],mMeans[i],mVCs[i],cout);
70
// Improve the initial estimate using the ECME algorithm. Stop when the multivariate likelihood is maximized.
//==============================================================================================================
cout << endl << string(85,'=') << endl << "ECME Algorithm #" << i+1 << " for " << type[(i%2)] << " initial solution." << endl << string(85,'=') << endl << endl;
fout.open(rootdir+ofile, ios_base::out | ios_base::app);
fout << endl << string(85,'=') << endl << "ECME Algorithm #" << i+1 << " for " << type[(i%2)] << " initial solution." << endl << string(85,'=') << endl << endl;
fout.close();
numCmpsF[i]=ECMEAlg(nTPoints,(const long double **) rtrn,nAssets,numCmps[(i%2)],mLHS[(i%2)],mRHS[(i%2)],mProbs[i],mMeans[i],mVCs[i],vCIDs[i],rootdir);
// Write final density to the output file and display to user (which contains maximum likelihood estimates).
//============================================================================================================
fout.open(rootdir+ofile, ios_base::out | ios_base::app);
wrtDens(type[i],numCmpsF[i],vCIDs[i],mProbs[i],mMeans[i],mVCs[i],fout); fout.close();
wrtDens(type[i],numCmpsF[i],vCIDs[i],mProbs[i],mMeans[i],mVCs[i],cout);
cout << endl << "Done processing " << type[i] << " initial LP solution. Final density shown above and written to file." << endl;
}
// Free temporary memory allocations.
//=====================================
for (int i=0; i<2*nECMEs; ++i)
{
delete [] mProbs[i]; mProbs[i]=nullptr;
delete [] mMeans[i]; mMeans[i]=nullptr;
delete [] mVCs[i]; mVCs[i]=nullptr;
delete [] vCIDs[i]; vCIDs[i]=nullptr;
}
for (int i=0; i<2; ++i)
{
delete [] valcIDs[i]; valcIDs[i]=nullptr;
delete [] estPrbs[i]; estPrbs[i]=nullptr;
}
delete [] mProbs; mProbs=nullptr;
delete [] mMeans; mMeans=nullptr;
delete [] mVCs; mVCs=nullptr;
delete [] valcIDs; valcIDs=nullptr;
delete [] mLHS; mLHS=nullptr;
delete [] mRHS; mRHS=nullptr;
delete [] estPrbs; estPrbs=nullptr;
for (int c=0; c<nCells; ++c)
{
delete [] allCells[c]; allCells[c]=nullptr;
}
delete [] allCells; allCells=nullptr;
delete [] tmpVals; tmpVals=nullptr;
delete [] tmpCombo; tmpCombo=nullptr;
delete [] cellAsgn; cellAsgn=nullptr;
delete [] cellCnt; cellCnt=nullptr;
delete [] cellPrb; cellPrb=nullptr;
for (int m=0; m<3; ++m)
{
delete [] optMDst[m]; optMDst[m]=nullptr;
}
for (int a=0; a<nAssets; ++a)
{
delete [] rtrn[a]; rtrn[a]=nullptr;
delete [] asgnmnt[a]; asgnmnt[a]=nullptr;
delete [] prob[a]; prob[a]=nullptr;
delete [] mean[a]; mean[a]=nullptr;
delete [] stdev[a]; stdev[a]=nullptr;
}
delete [] optMDst; optMDst=nullptr;
delete [] rtrn; rtrn=nullptr;
delete [] asgnmnt; asgnmnt=nullptr;
delete [] prob; prob=nullptr;
delete [] mean; mean=nullptr;
71
delete [] stdev; stdev=nullptr;
delete [] nComps; nComps=nullptr;
delete [] numCmpsF; numCmpsF=nullptr;
// Exit.
//========
cout << endl << "Done.
return 0;
(Hit return to exit.)" << endl; cin.get()
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: fitMixDist.cpp
/
/ Function: fitMixDist()
/
/ Summary:
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
This function fits a univariate normal mixture density to a set of observations using the EM algorithm within a forward‐backward iterative procedure that tests for the optimal # of components.
First, a 1‐component density is fit using standard maximum likelihood (ML) estimates and tested against a 2‐component mixture density fit using the EM algorithm with a (user specified) large # of
random starts. (Note: The # of random starts is specified by the 3rd parameter in the input control file as a multiple of the # of cores/independent processing units on the computer running the
application. In addition this value is multiplied by # components ‐ 1, which uses an increasing # of random starts for densities with more components.) When fitting a particular density, the one
with largest likelihood value from all random starts and obeying the variance ratio constraint (set in the header file using global constant: stdRatio) is selected as the ML estimator. A hypothesis
test of Ho: 1‐component vs Ha: 2‐components is then conducted using the forward alpha (significance level) set by the user in the input control file (parameter #6). The likelihood ratio test
statistic (LRT) is used and derived as ‐2Log(Lo/La), where Lo and La are the maximized likelihood values under Ho and Ha, respectively. The observed LRT statistic value is then compared to the
critical value of the LRT's distribution under Ho that yields area to the right equal to the forward alpha. The distribution of the LRT statistic under Ho is approximated via bootstrapping (see,
McLachlan (1987)). Assuming the MLE's fit under Ho reflect the true distribution of the data (under Ho) random samples from the distribution under Ho are generated. A 1‐component and 2‐component
univariate mixture density is fit to each sample which together generate a single null value from the true null distribution of the LRT statistic (again, assuming the true null distribution is that
fit by the MLEs). Using a large # of samples yields a set of values used to approximate the null distribution of the LRT statistic, with corresponding critical value for rejecting Ho based on the
forward alpha. If Ho is rejected then a 2‐component univariate mixture density becomes the new testing basis and the forward procedure continues with a test of 2‐components vs 3 components.
Otherwise, if Ho is not rejected then the forward procedure testing basis remains a 1‐component density, which is tested against 3‐components in the exact same manner. The key to bootstrapping the
LRT statistic's null distribution is to assume that the density with MLE parameters (under Ho) is the true null density and random samples are generated from it. The forward procedure continues in
this manner testing Ho: go components vs Ha: ga components up to Ha = max # components specified in the control file (parameter #4). If the maximum # of components is set to N, then the forward
procedure will conduct N‐1 hypotheses tests sequentially, where the alternatives are Ha=2 components, 3 components, 4 components, ..., N components, respectively. The forward procedure will end
with a certain # of mixture components being suggested as optimal. It is followed by a backward procedure that tests a univariate mixture density with 1 less component, and repeats the test until
Ho is rejected, ending the forward‐backward procedure. If the forward procedure ends by suggesting that F components are needed to properly fit the data, then the backward procedure begins by
testing Ho: F‐1 components vs Ha: F components using the backward alpha specified by the user in the control file (parameter #7). Note that the forward procedure conducts a series of tests with Ha
having 2, 3, 4, ..., maxcmps (maxcmps is the maximum # of components to consider in the application, which is set by the 4th parameter in the control file) components, respectively. Therefore, once
the forward procedure ends we have optimal univariate mixtures fit with ML estimates for all components from 1 to maxcmps. During backward processing these are used to generate the random samples
needed under Ho to approximate the null distribution of the LRT. If this hypothesis is rejected then the forward‐backward procedure ends and F components are considered optimal for the given
observation set. Else, if Ha is not rejected then the backward procedure basis becomes F‐1 components (since it is not significantly different from F components, and applying the principle of
parsimony). In this case, the backward procedure continues by testing Ho: F‐2 components vs Ha: F‐1 components. If this test is not rejected, a mixture density with F‐2 components becomes the new
basis and the procedure continues by testing Ho: F‐3 components vs Ha: F‐2 components. If the test is rejected then F‐1 components are considered optimal and the forward‐backward procedure ends.
The backward portion continues in this manner until a test is rejected. Consider an example with a maximum of N=10 components specified by the user in the control file (4th parameter) and a forward
alpha = 0.15 (6th parameter) and backward alpha = 0.20 (7th parameter). The following details a possible sequence of testing events and illustrates how the procedure iterates:
Test #
======
1
2
3
4
5
6
7
8
9
10
11
12
======
Ho:
============
1‐Component
1‐Component
1‐Component
4‐Components
4‐Components
4‐Components
7‐Components
7‐Components
7‐Components
6‐Components
5‐Components
4‐Components
============
Ha:
=============
2‐Components
3‐Components
4‐Components
5‐Components
6‐Components
7‐Components
8‐Components
9‐Components
10‐Components
7‐Components
6‐Components
5‐Components
=============
Alpha
=======
0.15
0.15
0.15
0.15
0.15
0.15
0.15
0.15
0.15
0.20
0.20
0.20
=======
Type
========
Forward
Forward
Forward
Forward
Forward
Forward
Forward
Forward
Forward
Backward
Backward
Backward
========
Result (Ho)
===========
Accept
Accept
Reject
Accept
Accept
Reject
Accept
Accept
Accept
Accept
Accept
Reject
===========
72
/
/
Final Result: Here, a 5‐component mixture is considered to provide best fit based on the user‐specified forward and backward alpha values (significance levels). Forward processing ended with a 7‐
/
component mixture producing the best fit and backward processing found no significant difference between 6 & 7 components (a test that was not performed during forward processing), and no difference
/
between 5 & 6 components (a test that was also not performed during forwared processing). The backward test of 4 vs 5 components was performed during forward processing but with a smaller alpha and
/
was not rejected, but is rejected using the larger alpha for backward processing. By tuning the forward/backward alphas the user customizes the procedure to favor a larger/smaller # of components.
/
/ Inputs:
/
/
1.) The asset # being processed (ranges from 0 thru (total # assets‐1)). The asset # is used for debugging and displaying results to the output window and output file. (a)
/
/
2.) The total # of observations (i.e., time points) in the current application. Specified by the user via the control file (2nd parameter). (T)
/
/
3.) An array of size T holding the observations/returns at each time point for the asset currently being processed. (r)
/
/
4.) The maximum # of components to allow in the univariate mixture distribution fit by this function. Specified by the user via the control file (4th parameter). (maxcmps)
/
/
5.) The # of bootstrap samples to use for each LRT when approximating the statistic's null distribution. Specified by the user via the control file (5th parameter). (nsmpls)
/
/
6.) The # of random starts to use when finding the ML estimates for a specific mixture density. Specified by the user via the control file (3rd parameter). Note that this value is taken as a
/
multiple of the # of cores and is also increased by the multiple (# components ‐ 1). (nstrts)
/
/
7.) An array of size 2 to hold the forward and backward alphas, respectively for the LRTs. (sl)
/
/
8.) An empty double array to hold the fitted univariate mixture distribution that results from applying the forward‐backward procedure detailed here. The array is indexed as [s][c], where s = 0, 1,
/
2, and c = component #. Here, s=0 refers to the component's probability, s=1 refers to the component's mean, and s=2 refers to the component's standard deviation. (fnlMDst)
/
/
9.) A string for the output directory. The final univariate mixture density for each # of components considered is written to the output file specified by the global constant ofile. (rdir)
/
/ Outputs:
/
/
This function populates the empty 2‐dimensional array that is supplied via parameter #8 with the optimal univariate mixture density fit by the forward‐backward procedure detailed above. The number
/
of components in this density is returned at the function call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int fitMixDist(const int a, const int T, const long double *r, const int maxcmps, const int nsmpls, const int nstrts, const long double sl[2], long double **fnlMDst, string rdir)
{
// Declare/initialize local variables.
//======================================
int hnum=0, sas, strt, end, incr, rCmpsH0=1, cbfsol=1, cbbsol, curopt, tstid, nBootsAdj, pValCntr, fprms, cmpord;
string yn, f_b, tmptxt;
long double LRT[2], *StatLRT=new long double [nsmpls+1], **orProb=new long double *[maxcmps], **orMean=new long double *[maxcmps], **orStd=new long double *[maxcmps],
*orLLVal=new long double [maxcmps], **orH0MDst=new long double *[3], *tmpRtrn=new long double[T], **outMDstH0=new long double *[3], **outMDstH1=new long double *[3],
**rMDstH0=new long double *[3], **oriMDst=new long double *[3], **oroMDst=new long double *[3];
vector<long double> pValue, alpha;
vector<int> h0, h1;
// Derive the 1‐component solution for the incoming sample (i.e., using the MLE estimates).
//===========================================================================================
for (int m=0; m<3; ++m)
{
oriMDst[m]=new long double [1];
rMDstH0[m]=new long double [1];
// 1‐component mixture used for sample random starts.
}
oriMDst[0][0]=1.00;
oriMDst[1][0]=getUniMean(T,r);
oriMDst[2][0]=getUniStd(T,r,oriMDst[1][0]);
// For each asset 2‐dimensional arrays holds the optimal g‐component solutions for the original sample. This will avoid having to rebuild optimal solutions during forward‐backward procedure.
//===============================================================================================================================================================================================
orProb[0]=new long double [1]; orProb[0][0]=oriMDst[0][0];
orMean[0]=new long double [1]; orMean[0][0]=oriMDst[1][0];
orStd[0]=new long double [1]; orStd[0][0]=oriMDst[2][0];
orLLVal[0]=getLLVal(T,r,1,(const long double **) oriMDst);
cout << string(20,'=') << endl << string(20,'=') << endl << "PROCESSING ASSET #" << a+1 << endl << string(20,'=') << endl << string(20,'=') << endl << endl;
strt=2; end=maxcmps+1; incr=1; f_b="Forward";
73
for (int fb=0; fb<2; ++fb)
{
// Iteration limits depend on the forward‐backward procedure.
//=============================================================
if (1 == fb)
{
strt=cbfsol‐1; end=0; incr=‐1; f_b="Backward"; cbbsol=cbfsol;
}
for (int g=strt; g != end; g=g+incr)
{
// Set null and alternative hypotheses.
//=======================================
hnum=hnum+1;
if (0 == fb)
{
h0.push_back(cbfsol); h1.push_back(g); sas=‐1;
alpha.push_back(sl[0]);
}
else if (1 == fb)
{
h0.push_back(g); h1.push_back(cbfsol);
alpha.push_back(sl[1]);
}
// Write details of this iteration.
//===================================
cout << string(62,'‐') << endl << string(62,'‐') << endl << "Asset #" << a+1 << " ‐ Hypothesis Test #" << hnum << " (" << h0[hnum‐1] << " vs " << h1[hnum‐1]
<< ") " << " [Direction: " << f_b << "]" << endl << string(62,'‐') << endl << string(62,'‐') << endl << endl << " " << string(62,'‐') << endl << " " << string(62,'‐')
<< endl << " H0: Asset #" << a+1 << " is best fit by a(n) " << h0[hnum‐1] << " component Normal mixture." << endl << "
vs" << endl << " H1: Asset #"
<< a+1 << " is best fit by a(n) " << h1[hnum‐1] << " component Normal mixture." << endl << " " << string(62,'‐') << endl << " " << string(62,'‐') << endl;
// If backward processing, check whether or not this hypothesis has already been tested. If so, issue a warning then retrieve and use the existing solution.
//=============================================================================================================================================================
for (int h=0; h<hnum‐1; ++h)
{
if (h0[h]==h0[hnum‐1] && h1[h]==h1[hnum‐1])
{
if (sl[0]==sl[1])
tmptxt=" the same ";
else
tmptxt=" a different ";
cout << endl << " WARNING: This hypothesis has already been tested using" << tmptxt << "alpha. (See test #" << h+1 << ".)"; sas=h+1;
}
}
// Rebuild input array under h0. (We need the null dist. to generate bootstrap samples.)
// Note that h0[hnum‐1] = # components under h0 for test # hnum and h0[hnum‐1]‐1 is the index of the optimal solution under h0 # of components.
//====================================================================================================================================================
for (int m=0; m<3; ++m)
orH0MDst[m]=new long double [h0[hnum‐1]];
for (int c=0; c<h0[hnum‐1]; ++c)
{
// In general, the g‐component solution is stored in orXYZ[g‐1][c]. For example:
orH0MDst[0][c]=orProb[h0[hnum‐1]‐1][c];
// ‐‐> orProb[0] is a 1‐dimensional array with single element orProb[0][0].
orH0MDst[1][c]=orMean[h0[hnum‐1]‐1][c];
// ‐‐> orMean[1] is a 2‐dimensional array with elements orMean[1][0], orMean[1][1].
orH0MDst[2][c]=orStd[h0[hnum‐1]‐1][c];
// ‐‐> orStd[2] is a 3‐dimensional array with elements orStd[2][0], orStd[2][1], orStd[2][2].
}
// Report the current optimal solution for the actual sample under H0 when forward processing.
//==============================================================================================
if (0 == fb && dbug >= 1)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20);
cout << endl << " ‐‐> " << h0[hnum‐1] << "‐Component Optimal Solution: Log‐Likelihood: " << orLLVal[h0[hnum‐1]‐1] << endl << "
<< getVRatio(h0[hnum‐1], orH0MDst[2]) << endl;
showVals(0,r,h0[hnum‐1],(const long double **) orH0MDst);
74
Variance Ratio: "
} cout << endl;
// Simulation is used to approximate the null distribution of the LRT statistic. If a sample is generated but no local optimum is found then reduce the simulation
// count by 1. Also, reduce the simulation count by 1 when the LRT statistic is < 0. This can happen if a local optimum is found under the full model that has a
// smaller likelihood than the optimum value under the reduced model (which may itself be a local optimum). The adjusted count is held in nBootsAdj.
//===================================================================================================================================================================
nBootsAdj=nsmpls;
// LRT value for current solution. Numerator uses the parameters already estimated. We need to fit a new solution using h1 components on the incoming data for the
// denominator of the LRT. This is done here. Once fit, the estimates are transferred to the double arrays that hold optimal solutions for all component sizes.
// (Note: When forward processing, hnum == h1[hnum‐1]‐1 == g‐1 so that hnum+1 == h1[hnum‐1] == g.)
//====================================================================================================================================================================
if (0 == fb)
{
for (int m=0; m<3; ++m)
oroMDst[m]=new long double [h1[hnum‐1]]; /* (2) Note that oriMDst has exactly hnum components during forward processing which is 1 less than the # of comps in H1. */
orLLVal[hnum]=ThrdEMAlg(T,r,3*nstrts,hnum,(const long double **) oriMDst,h1[hnum‐1],oroMDst);
// Instantiate the arrays of size h1 and transfer the optimal solution for storage across component sizes.
//==========================================================================================================
orProb[hnum] = new long double [h1[hnum‐1]]; orMean[hnum] = new long double [h1[hnum‐1]]; orStd[hnum] = new long double [h1[hnum‐1]];
for (int c=0; c<h1[hnum‐1]; ++c)
{
orProb[hnum][c]=oroMDst[0][c];
orMean[hnum][c]=oroMDst[1][c];
orStd[hnum][c]=oroMDst[2][c];
}
// If the algorithm fails to converge on the original data we must exit the program.
//====================================================================================
if (orLLVal[hnum] <= LNegVal)
{
cout << "ERROR: Asset #" << a+1 << " did not converge to a local optimum when attempting to fit " << h1[hnum‐1] << " components." << endl
<< "Try increasing the # of random starts, decreasing the # of components, or increasing the variance ratio constraint." << endl
<< "EXITING...fitMixDist()..." << endl; cin.get();
exit (EXIT_FAILURE);
}
// Also exit if the algorithm finds an inferior optimum when compared to those with fewer #'s of components.
//============================================================================================================
if (orLLVal[hnum‐1] > orLLVal[hnum])
{
cout << "ERROR: Likelihood for Asset #" << a+1 << " when fitting " << h1[hnum‐1] << " components is less than the likelihood when " << endl
<< "fitting " << hnum << " components. An inferior local optimum has been found. (Increase the # of random starts to prevent this.)" << endl
<< "EXITING...fitMixDist()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
}
StatLRT[nsmpls] = ‐2.00*(orLLVal[h0[hnum‐1]‐1] ‐ orLLVal[h1[hnum‐1]‐1]);
// Report the LRT statistic value for the actual sample.
//========================================================
if (dbug >= 1)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20);
cout << " ‐‐> LRT Details for Original Sample: H0 Log‐L=" << orLLVal[h0[hnum‐1]‐1] << ", H1 Log‐L=" << orLLVal[h1[hnum‐1]‐1] << ", LRTS for Asset #" << a+1 << "="
<< StatLRT[nsmpls] << endl << endl;
}
// Bootstrap the LRT statistic to determine its sampling dist. The reduced model is the one with MLE's just derived for the given # of components. This test is for # of components + 1.
//==========================================================================================================================================================================================
if (‐1==sas)
{
if (0 == dbug)
cout << " Processing bootstrap samples ";
75
// Instantiate the arrays to hold the optimal solutions under H0 and H1. These will be reused for each sample.
//===============================================================================================================
for (int m=0; m<3; ++m)
{
outMDstH0[m]=new long double [h0[hnum‐1]];
outMDstH1[m]=new long double [h1[hnum‐1]];
}
// Test the hypothesis.
//=======================
for (int b=0; b<nsmpls; ++b)
{
if (0 == dbug)
{
cout << ".";
if (b+1 != nsmpls && 0 == (b+1) % 100)
cout << endl << string(31,' ');
}
else if (dbug >= 1)
cout << string(162,'‐') << endl << string(162,'‐') << endl << string(40,' ') << "Asset #" << a+1 << ": Hypothesis Test #" << hnum << " (" << h0[hnum‐1]
<< " vs " << h1[hnum‐1] << "): Start Processing Bootstrap Sample #" << b+1 << " of " << nsmpls << "." << endl << string(162,'‐') << endl << string(162,'‐') << endl << endl;
// Generate sample of size nTPoints from the reduced model under the null hypothesis that the reduced model (using MLEs) is correct for this set of returns.
// When bootstrapping the LRT statistic, use 1‐component values (probs, means, stds) to generate the random starts under h0 (because we do not have access to
// the solution using H0 # components less 1 on each sample) and use the model fitted under h0 to generate random starts when fitting the model specified by h1.
//================================================================================================================================================================
getRVals(T,h0[hnum‐1],(const long double **) orH0MDst,tmpRtrn);
// Fit both H0 (# components) and H1 (# components) models and form the LRT statistic. [Random start for H0 (always a 1‐component mixture)].
//=============================================================================================================================================
rMDstH0[0][0]=1.00;
rMDstH0[1][0]=getUniMean(T,tmpRtrn);
rMDstH0[2][0]=getUniStd(T,tmpRtrn,rMDstH0[1][0]);
// H0.
//=====
LRT[0]=ThrdEMAlg(T,tmpRtrn,nstrts,rCmpsH0,(const long double **) rMDstH0,h0[hnum‐1],outMDstH0);
// H1
//=====
if (LRT[0] > LNegVal)
LRT[1]=ThrdEMAlg(T,tmpRtrn,nstrts,h0[hnum‐1],(const long double **) outMDstH0,h1[hnum‐1],outMDstH1);
// Random observation from the null distribution of the following LRT: (Reduce the simulation size count by 1 if the LRT statistic is negative.)
// H0: Data originate from mixture with h0[hnum‐1] components vs. H1: Data originate from mixture with h1[hnum‐1] components
//=================================================================================================================================================
if (LRT[0] > LNegVal && LRT[1] > LNegVal)
StatLRT[b] = ‐2.00*(LRT[0] ‐ LRT[1]);
else
StatLRT[b] = ‐1.00;
// The denominator gets decremented in 2 ways:
//
1.) The LRT statistic is negative which means that an inferior local optimum was found.
//
2.) No local optimum was found when fitting either the null component or the alternative component distributions, all were spurious (i.e., leading to unboundedness).
//==========================================================================================================================================================================
if (StatLRT[b] < 0.00)
nBootsAdj=nBootsAdj‐1;
// Report the LRT statistic value for the bootstrap sample.
//===========================================================
if (dbug >= 1)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20);
if (LRT[0] > LNegVal && LRT[1] > LNegVal && StatLRT[b] >= 0.00)
cout << " ‐‐> LRT Details for Sample #" << b+1 << ": LRT[0]=" << LRT[0] << ", LRT[1]=" << LRT[1] << ", LRTS for bootstrap sample "
76
<< b+1 << " of " << nsmpls << "=" << StatLRT[b] << endl << endl;
else
{
if (LRT[0] <= LNegVal || LRT[1] <= LNegVal)
{
tstid=1;
if (LRT[0] <= LNegVal)
tstid=0;
cout << " ‐‐> The EM algorithm did not find a local optimum under H" << tstid << ". (Sample " << b+1 << " will be discarded.)" << endl << endl;
}
else
cout << " ‐‐> Negative LRT test statistic: " << StatLRT[b] <<". (Sample " << b+1 << " will be discarded.)" << endl
<< "
[An inferior (not the largest) local optimum was found under H1.]" << endl << endl;
}
}
}
// Clear out sample solution arrays.
//====================================
for (int m=0; m<3; ++m)
{
delete [] outMDstH0[m]; outMDstH0[m]=nullptr;
delete [] outMDstH1[m]; outMDstH1[m]=nullptr;
}
}
// Clear the array that holds the optimal solution for the original sample under H0.
//====================================================================================
for (int m=0; m<3; ++m)
{
delete [] orH0MDst[m]; orH0MDst[m]=nullptr;
}
if (‐1==sas && 0 == dbug)
cout << " (" << nsmpls << ")" << endl;
// Determine p‐value & test result. If H0 is rejected then move the temporary values into their permanant placeholders. If null hypothesis is not rejected then set variable stop to 1.
//=========================================================================================================================================================================================
if (‐1 == sas)
{
pValCntr=1;
for (int n=0; n<nsmpls; ++n)
if (StatLRT[n] > 0.00 && StatLRT[n] >= StatLRT[nsmpls])
pValCntr=pValCntr+1;
pValue.push_back((long double) pValCntr/(nBootsAdj+1));
// Uncomment to write the LRT[] array to a file in the error folder. These values form the null distribution of the test statistic.
//====================================================================================================================================
// ofstream fout;
// fout.open(errFolder + "Asset_" + to_string((long long) a+1) + "_LRT_" + to_string((long long) h0[hnum‐1]) + "_vs_" + to_string((long long) h1[hnum‐1]) + ".txt");
// fout.setf(ios_base::fixed, ios_base::floatfield); fout.precision(16);
// fout << "‐‐> LRT statistic for original data: " << StatLRT[nsmpls] << endl << endl;
// fout << "‐‐> Null distribution LRT statistic values for " << nsmpls << " bootstrap samples (set values of ‐1.00 to missing):" << endl << endl;
// for (int b=0; b<nsmpls; ++b)
//
fout << "StatLRT[" << b << "] = " << StatLRT[b] << endl;
// fout.close();
}
else
pValue.push_back(pValue[sas‐1]);
// Write out the hypothesis test result. If this is a retest, report the prior result. (Note that it is a retest when sas >= 0, and sas holds the prior test number.)
//=======================================================================================================================================================================
if (‐1 == sas)
{
77
cout << endl << " " << string(37,'=') << endl << " Hypothesis test #" << hnum << " (" <<
<< endl << " ‐‐> The hypothesis test uses " << nBootsAdj << " valid LRTS values from
<< nBootsAdj+1 << ")" << endl << " ‐‐> Of these, there are " << pValCntr << " values
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20);
cout << " ‐‐> The resulting p‐value for testing H0 vs H1 is: " << pValue[hnum‐1] << " vs.
h0[hnum‐1] << " vs " << h1[hnum‐1] << ") result:" << endl << " " << string(37,'=')
resampling along with the 1 LRTS value from the original sample. (Total="
>= the sample LRT statistic." << endl;
" << f_b << " alpha: " << sl[fb] << "." << endl;
}
else
{
cout << endl << " " << string(67,'=') << endl << " Hypothesis test #" << hnum << " (" << h0[hnum‐1] << " vs " << h1[hnum‐1] << ") is a retest of Hypothesis test # " << sas
<< "." << endl << " " << string(67,'=') << endl;
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20); cout << " ‐‐> The resulting p‐value for testing H0 vs H1 is: " << pValue[hnum‐1] << " vs. " << f_b
<< " alpha: " << sl[fb] << "." << endl;
}
// Accept or reject the hypothesis. If Ho is rejected then Ha is considered a better fit. When forward testing, Ha has additional components (i.e., the integer value of h1).
// When backward testing Ho is considered an acceptable fit when not rejected (i.e., integer value of h0). Here we keep track of the best solutions for both forward and backward
// testing. The index of the optimal solution when accessing the orXYZ[][] double arrays is (cbbsol‐1) once finished processing. That is, orXYZ[cbbsol‐1][0] ‐ orXYZ[cbbsol‐1][cbbsol‐1].
//=========================================================================================================================================================================================
yn=" NOT ";
if (pValue[hnum‐1] <= sl[fb])
{
yn=" ";
if (0 == fb)
cbfsol=h1[hnum‐1];
// H0 was rejected, set CBFSOL to H1 # components.
else
// Stop backward processing after first H0 is rejected.
end=g+incr;
}
else if (1 == fb)
cbbsol=h0[hnum‐1];
curopt = (int) (0==fb)*cbfsol + (int) (1==fb)*cbbsol;
// For displaying current solution to user.
// If forward processing and not at the max # components we should reset oriMDst[][] and clear out oroMDst[][].
//===============================================================================================================
if (0 == fb && g+incr != end)
{
// Clear/update the input and output arrays. The oriMDst[][] densities will generate the random starts when fitting the original
// sample to one additional component. (Note that the original double mixture arrays are only needed during forward processing.)
//=================================================================================================================================
for (int m=0; m<3; ++m)
{
delete [] oriMDst[m];
oriMDst[m]=new long double [h1[hnum‐1]];
for (int c=0; c<h1[hnum‐1]; ++c)
oriMDst[m][c]=oroMDst[m][c];
delete [] oroMDst[m]; oroMDst[m]=nullptr;
}
}
// Report the current best solution after testing this hypothesis.
//==================================================================
cout << " ‐‐> The null hypothesis is" << yn << "rejected in favor of the alternative hypothesis." << endl << " ‐‐> At this stage a(n) " << curopt
<< "‐component Normal mixture provides the best fit for Asset #" << a+1 << "." << endl << endl << string(162,'‐') << endl << string(162,'‐') << endl << endl;
// Processor cool down.
//=======================
if (dbug >= 2)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(4);
cout << endl << "***** Processor cool Down: " << ((double) cDown/1000.00)/60.00 << " minutes. *****" << endl;
} Sleep(cDown);
}
}
// Display p‐values then clear out the containers.
//==================================================
78
cout << " All p‐values for Asset #" << a+1 << " by test number:" << endl << " " << string(41,'=') << endl;
for (int m=0; m<pValue.size(); ++m)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(10);
cout << " P‐value for hypothesis test #" << setfill('0') << setw(2) << m+1 << " (" << setw(2) << h0[m] << " vs " << setw(2) << h1[m] << "): " << pValue[m]; cout.precision(5);
cout << " (alpha = " << alpha[m] << ")"; cout.precision(10); cout << " (Log‐Likelihood Values: LL[H0]=" << orLLVal[h0[m]‐1] << ", LL[H1]=" << orLLVal[h1[m]‐1] << ")" << endl;
}
// Insert the optimal solution into the double array passed to this function. The # components is returned by the function. 1st order the mixture returned by the respective means.
//====================================================================================================================================================================================
for (int c1=0; c1<cbbsol; ++c1)
{
cmpord=0;
for (int c2=0; (c2<cbbsol); ++c2)
if (orMean[cbbsol‐1][c2] < orMean[cbbsol‐1][c1])
cmpord++;
// Account for ties.
//====================
for (int c3=cmpord; (c3<cbbsol && fnlMDst[0][cmpord] > 0.00); ++c3)
cmpord++;
// Populate the outgoing array.
//===============================
fnlMDst[0][cmpord]=orProb[cbbsol‐1][c1];
fnlMDst[1][cmpord]=orMean[cbbsol‐1][c1];
fnlMDst[2][cmpord]=orStd[cbbsol‐1][c1];
}
// Display the final result for each asset.
//===========================================
fprms = (3*(cbbsol)‐1); // Free parameters.
cout << endl << " ‐‐> Distribution for Asset #" << a+1 << " is a(n) " << cbbsol << "‐Component Normal Mixture (full details below)." << endl
<< endl << "
Log‐Likelihood: " << orLLVal[cbbsol‐1]
<< endl << "
Variance Ratio: " << getVRatio(cbbsol, orStd[cbbsol‐1])
<< endl << "
AIC: " << (long double) 2.00*fprms ‐ 2.00*orLLVal[cbbsol‐1]
<< endl << "
BIC: " << ‐2.00*orLLVal[cbbsol‐1] + (long double) fprms*log((long double) T)
<< endl << "
AICC: " << (long double) 2.00*fprms ‐ 2.00*orLLVal[cbbsol‐1] + ((long double) 2.00*(fprms + 1)*(fprms + 2))/((long double) (T ‐ fprms ‐ 2))
<< endl << "
Density Parameters: " << endl;
showVals(0,r,cbbsol,(const long double **) fnlMDst);
cout << endl << string(162,'=') << endl << string(162,'=') << endl;
// Write all densities to output file for this asset. Along with all p‐values.
//==============================================================================
ofstream fout;
if (0 == a)
fout.open(rdir+ofile, ios_base::out);
else
fout.open(rdir+ofile, ios_base::out | ios_base::app);
fout << string(28,'=') << endl << "Asset #" << a+1 << " Optimal Densities:" << endl << string(28,'=') << endl;
for (int i=0; i<maxcmps; ++i)
{
fout << endl << " " << i+1 << "‐Component Optimal Solution: " << endl << " " << string(43,'=');
fout.setf(ios_base::fixed, ios_base::floatfield); fout.precision(16);
fprms = (3*(i+1)‐1); // Free parameters.
fout << endl << " ‐‐> Log‐Likelihood: " << orLLVal[i]
<< endl << "
Variance Ratio: " << getVRatio(i+1, orStd[i])
<< endl << "
AIC: " << (long double) 2.00*fprms ‐ 2.00*orLLVal[i]
<< endl << "
BIC: " << ‐2.00*orLLVal[i] + (long double) fprms*log((long double) T)
<< endl << "
AICC: " << (long double) 2.00*fprms ‐ 2.00*orLLVal[i] + ((long double) 2.00*(fprms + 1)*(fprms + 2))/((long double) (T ‐ fprms ‐ 2))
<< endl << "
Density Parameters: " << endl;
for (int c=0; c<i+1; ++c)
fout << string(37,' ') << "Prob[" << c << "]=" << orProb[i][c] << " Mean[" << c << "]=" << orMean[i][c] << " Std.Dev.[" << c << "]=" << orStd[i][c] << endl;
}
fout << endl << endl << " All p‐values for Asset #" << a+1 << " by test number:" << endl << " " << string(41,'=') << endl;
79
for (int m=0; m<pValue.size(); ++m)
{
fout.setf(ios_base::fixed, ios_base::floatfield); fout.precision(10);
fout << " P‐value for hypothesis test #" << setfill('0') << setw(2) << m+1 << " (" << setw(2) << h0[m] << " vs " << setw(2) << h1[m] << "): " << pValue[m]; fout.precision(5);
fout << " (alpha = " << alpha[m] << ")"; fout.precision(10); fout << " (Log‐Likelihood Values: LL[H0]=" << orLLVal[h0[m]‐1] << ", LL[H1]=" << orLLVal[h1[m]‐1] << ")" << endl;
}
fout << endl << endl << " Optimal Density for Asset #" << a+1 << endl << " " << string(31,'=') << endl;
showVals(0,r,cbbsol,(const long double **) fnlMDst, fout); fout << endl << endl;
fout.close();
// Clear vectors and free temporary memory allocations.
//=======================================================
pValue.clear(); pValue.shrink_to_fit(); alpha.clear(); alpha.shrink_to_fit();
h0.clear(); h0.shrink_to_fit(); h1.clear(); h1.shrink_to_fit();
for (int m=0; m<3; ++m)
{
delete [] oriMDst[m]; oriMDst[m]=nullptr;
delete [] oroMDst[m]; oroMDst[m]=nullptr;
delete [] rMDstH0[m]; rMDstH0[m]=nullptr;
}
delete [] oriMDst; oriMDst=nullptr;
delete [] oroMDst; oroMDst=nullptr;
delete [] rMDstH0; rMDstH0=nullptr;
for (int c=0; c<maxcmps; ++c)
{
delete [] orProb[c]; orProb[c]=nullptr;
delete [] orMean[c]; orMean[c]=nullptr;
delete [] orStd[c]; orStd[c]=nullptr;
}
delete [] orProb; orProb=nullptr;
delete [] orMean; orMean=nullptr;
delete [] orStd; orStd=nullptr;
delete [] tmpRtrn; tmpRtrn=nullptr;
delete [] StatLRT; StatLRT=nullptr;
delete [] orLLVal; orLLVal=nullptr;
delete [] orH0MDst; orH0MDst=nullptr;
delete [] outMDstH0; outMDstH0=nullptr;
delete [] outMDstH1; outMDstH1=nullptr;
// Return the optimal # of components.
//======================================
return cbbsol;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: ThrdEMAlg.cpp
/
/ Function: ThrdEMAlg()
/
/ Summary:
/
/
This function threads the EM algorithm. An individual thread is used to maximize the likelihood function for 1 random start. Likelihood functions for mixture PDFs have multiple local optimums, and
/
they are unbounded. The MLEs for the parameters are those that yield the largest of the local maximums, after removing spurious optimizers. Spurious optimizers are those where a single component
/
is used to fit one or a small number of closely clustered observations. In such cases the variance of the corresponding component becomes very small or approaches zero, which drives the likelihood
/
value to infinity. Spurious optimizers can be eliminated by imposing a variance ratio constraint which does not allow the ratio of the largest to smallest variance (across components) to exceed
/
some given constant. This will prevent a variance from approaching zero or a small number. The user must set the variance ratio constraint to a value that is appropriate for the data in their
/
application. If the intent is to construct a density function that memorizes the training data then this value can be set very high. If the intent is to build a density that extends well to out‐
/
of‐sample data then this value can be set very low. The global constant stdRatio is set in the header file as the square root of the desired variance ratio constraint. The # of threads is set by
/
the # of independent processing units on the computer running the application as well as the # of components in the mixture being fit and the user‐specified # of random starts parameter setting in
80
/
the control file (3rd parameter). The # of random starts is equal to: rs*(# cores)*(g‐1), where rs = # of random starts specified by the 3rd parameter in control.txt, # cores = # of indpendent
/
processing units on the computer running the application, and g = # of components in the univariate mixture density being fit. The # of random starts therefore increases with the size of the
/
density being fit. Each random start is assigned its own thread and launches the EMAlg() function to find the likelihood maximizer based on that random start (this is a unique value and reflects
/
the local maximum nearest to the parameter settings in the random start). Once all threads finish, the parameter settings from the random start which yields the largest likelihood value (and
/
obeying the variance ratio constraint) are taken as the MLEs. An empty double array is populated with these values and the maximum log‐likelihood function value is returned at the call.
/
/ Inputs:
/
/
1.) The total # of observations (i.e., time points) in the current application. Specified by the user via the control file (2nd parameter). (T)
/
/
2.) An array of size T holding the observations/returns at each time point for the asset currently being processed. (r)
/
/
3.) The # of random starts to use when finding the ML estimates for a specific mixture density. Specified by the user via the control file (3rd parameter). Note that this value is taken as a
/
multiple of the # of cores and is also increased by the multiple (# components ‐ 1), so that an increasing # of random starts is used as the # of components increases. (nstrts)
/
/
4.) The # of components in the univariate mixture distribution used to generate random starts (i.e., means) for the EM algorithm. For example if fitting a 5‐component univariate mixture
/
distribution to an asset and an optimal 4‐component univariate mixture distribution is available for that asset, then a 4‐component mixture distribution is used to generate outg (argument #6 to
/
this function) means as the starting point for constructing the random starts. If no such 4‐component mixture distribution is available (for example during bootstrapping of the LRT with Ho
/
having 4 components), then a 1‐component mixture will be used to generate the means for an outg‐component random start. (Note: Recall that a g‐component random start is built by generating g‐
/
random values which serve as the component means. Each observation is then attached to the nearest mean (component). The probability for that component is the # attached divided by the total #
/
of observations, and the standard deviation for that component is the sample standard deviation for all observations assigned to it.) (ing)
/
/
5.) A 2‐dimensional array to hold the univariate mixture distribution used to generate means for a random start. This density function has ing components as mentioned above and the double array is
/
indexed as [s][c], where s = 0, 1, 2, and c = component #. Here, s=0 refers to the component's probability, s=1 refers to the component's mean, and s=2 refers to the component's standard
/
deviation. As noted above, when generating random starts for the EM algorithm a density function that is most similar to the # of components being fit is desirable. Therefore, if fitting a 5‐
/
component mixture density and an optimal 4‐component mixture density is available then it will be used to generate the random starts. If an optimal 4‐component mixture density is not available
/
then a univariate normal distribution (i.e., 1‐component mixture) can be used to generate the means for a random start. (inMDst)
/
/
6.) The # of components in the outgoing univariate mixture density that is being fit using the EM algorithm. (outg)
/
/
7.) An empty double array to hold the fitted univariate mixture distribution fit using the EM algorithm. The double array is indexed as [s][c], where s = 0, 1, 2, and c = component #. Here, s=0
/
refers to the component's probability, s=1 refers to the component's mean, and s=2 refers to the component's standard deviation. (outMDist)
/
/ Outputs:
/
/
This function returns the log‐likelihood for the optimal univariate mixture fit using the EM algorithm at the function call. It also populates the supplied empty double array (outMDist) with the
/
corresponding optimal univariate mixture distribution.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
long double ThrdEMAlg(const int T, const long double *r, const int rs, const int ing, const long double **inMDist, const int outg, long double **outMDist)
{
// Local variables.
//===================
long double llval;
// Handle trivial case first (i.e., 1‐component), then non‐trivial case (i.e., > 1‐component).
//==============================================================================================
if (outg == 1)
{
for (int m=0; m<3; ++m)
outMDist[m][0]=inMDist[m][0];
// Transfer probabilities, means, and standard deviations.
llval = getLLVal(T,r,outg,(const long double **) outMDist);
// Compute likelihood.
}
else
{
// Declare/initialize local variables. Find # of independent processing units and create array of thread objects using the random start multiplication factor specified in the control file.
// The # of random starts specified in the control file is a multiple of the # of independent processing units.
//=============================================================================================================================================================================================
int p = (int) (outg‐1)*rs*boost::thread::hardware_concurrency(), fstgrp=1, **runprms=new int *[p];
long double **rPrbs=new long double *[p], **rMns=new long double *[p], **rStds=new long double *[p], *llVals=new long double [p];
boost::thread *t=new boost::thread[p];
string cmt, spc;
// Display message with details of the optimization process.
//============================================================
81
if (dbug >= 1)
{
cout << " Threading EM Algorithm will use " << p << " random starts for the current optimization." << endl;
if (dbug >=2)
{
cout << " [Note: When the variance ratio constraint is violated, or the maximum # of iterations has been reached, or an invalid probability detected," << endl
<< "
the log‐likelihood is set to the arbitrarily large negative #: "; cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(7); cout << LNegVal << "]" << endl;
}
cout << endl;
}
// Create arrays to hold the random starts and return values from the EM algorithm. A random start must specify all unknown parameters for
// a ng‐component mixture. (i.e., component probabilities, means, standard deviations) These p arrays will be reused within each run group.
//============================================================================================================================================
for (int j=0; j<p; ++j)
{
rPrbs[j] = new long double [outg];
rMns[j] = new long double [outg];
rStds[j] = new long double [outg];
runprms[j] = new int [4]; runprms[j][0]=ing; runprms[j][1]=j;
}
// Iterate over # of random starts and determine the start, then launch optimization calls.
//===========================================================================================
for (int j=0; j<p; ++j)
{
// Launch call to EMAlg() for each thread. Once all threads finish, we scan the likelihood values across the p solutions (inner loop) and select the largest local optimum as the optimum.
// The outer loop then repeats this process rs*(outg‐1) number of times.
//============================================================================================================================================================================================
t[j] = boost::thread(EMAlg,T,boost::cref(r),outg,boost::ref(rPrbs),boost::ref(rMns),boost::ref(rStds),boost::ref(llVals), boost::cref(inMDist),boost::ref(runprms[j]));
}
// Pause until all sub‐threads finish, and save the optimal solution from this group of runs.
//=============================================================================================
for (int j=0; j<p; ++j)
{
t[j].join();
// Report results when requested.
//=================================
if (dbug >= 2)
{
spc=""; cmt="";
if (0 == runprms[j][2])
cmt="[Variance ratio constraint violated, solution not used.]";
else if (1 == runprms[j][2])
cmt="[Maximum # of iterations reached, solution not used.]";
else if (2 == runprms[j][2])
{
if (llVals[j] > 0.00 && llVals[j] < 10.00) spc=" ";
else if (llVals[j] >= 10.00 && llVals[j] < 100.00) spc=" ";
}
else if (3 == runprms[j][2])
cmt="[Invalid probability encountered, solution not used.]";
cout << " ‐‐> " << outg << "‐Component Log‐Likelihood for random start # " << setfill('0') << setw(to_string((long long) p).size()) << j+1 << " = ";
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(8*(1+(2 == runprms[j][2]))); cout << spc << llVals[j]; cout.precision(0); cout << " (Iterations = " << setfill('0')
<< setw(to_string((long long) mIters).size()) << runprms[j][3] << ") " << cmt << endl;
}
// Retrieve optimal solution: Scan random starts and locate the one associated with the highest likelihood value. The optimal probabilities, means, standard deviations
// are transferred to the placeholders passed to this function. If no local optimum has been found then set the return code to 0, otherwise use a return code of 1.
//=========================================================================================================================================================================
if (0 == j && 1 == fstgrp)
{
fstgrp=0;
llval=llVals[0];
82
for (int c=0; c<outg; ++c)
{
outMDist[0][c]=rPrbs[0][c];
outMDist[1][c]=rMns[0][c];
outMDist[2][c]=rStds[0][c];
}
}
else if (llVals[j] > llval)
{
llval=llVals[j];
for (int c=0; c<outg; ++c)
{
outMDist[0][c]=rPrbs[j][c];
outMDist[1][c]=rMns[j][c];
outMDist[2][c]=rStds[j][c];
}
}
}
if (dbug >= 1)
cout << endl;
// Free temporary memory allocations.
//=====================================
for (int j=0; j<p; ++j)
{
delete [] rPrbs[j]; rPrbs[j]=nullptr;
delete [] rMns[j]; rMns[j]=nullptr;
delete [] rStds[j]; rStds[j]=nullptr;
delete [] runprms[j]; runprms[j]=nullptr;
}
delete [] rPrbs; rPrbs=nullptr;
delete [] rMns; rMns=nullptr;
delete [] rStds; rStds=nullptr;
delete [] runprms; runprms=nullptr;
delete [] llVals; llVals=nullptr;
delete [] t; t=nullptr;
}
// Report the optimal solution.
//===============================
if (dbug >= 1)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(16);
cout << " ‐‐> " << outg << "‐Component Optimal Solution: Log‐Likelihood: " << llval << endl << "
if (llval > LNegVal)
showVals(0,r,outg,(const long double **) outMDist); cout << endl;
}
Variance Ratio: " << getVRatio(outg, outMDist[2]) << endl;
// Return the log‐likelihood value corresponding to the optimal solution.
//=========================================================================
return llval;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: EMAlg.cpp
/
/ Function: EMAlg()
/
/ Summary:
83
/
/
This function implements the EM algorithm for estimating parameters of a univariate mixture PDF. The observations from a univariate mixture PDF can be viewed as an incomplete data problem, where
/
the component # for each observation/time point is missing or unobserved. That is, at each time point 2 random variables produce a value: 1.) the component #, and 2.) the value from it. Note that
/
both sets of RVs will have parameters to be estimated. If viewed this way, the likelihood function can be expressed using both the missing and non‐missing random variables and their corresponding
/
parameters. For example, the parameters for the unobserved component random variable that each observation originates from are the component probabilities, and the parameters for the observed
/
density values are means and standard deviations from each component distribution. The EM algorithm then estimates parameters for both random variables iteratively as follows:
/
/
Step 1: Select starting values for all parameters (for both the missing/unobserved and non‐missing/observed random variables).
/
Step 2: Compute the expected values for all missing/unobserved random variables using the most recent parameter estimates.
/
Step 3: Replace all instances of the missing random variables with their expected values in the likelihood or log‐likelihood function.
/
Step 4: Optimize the resulting likelihood or log‐likelihood function with respect to the parameters for the non‐missing random variables.
/
Step 5: Check the %‐change in likelihood or log‐likelihood value maximized in step 4 for convergence (i.e., small %‐change) or error/exit condition.
/
Step 6: If no convergence or error/exit condition is met in Step 5 then return to Step 2.
/
/
Starting values for Step 1 are computed as random starts for all parameter values. This is acheived by first generating g values from the closest distribution to the one being fit (i.e., if fitting
/
a 5‐component univariate mixture and an optimal 4‐component univariate mixture is available, then use the 4‐component mixture to generate random starts, and if no other density is available then use
/
a 1‐component mixture). These g values are taken as the means, and each observation is attached to the closest mean. The standard deviation of the set of obserations attached to each mean is
/
computed as the component standard deviation, and the proportion of observations attached to each mean is the component probability for the random start. Convergence in Step 5 is checked using the
/
global constant epsilon set in the header file. In Step 5 we also check for error/exit conditions and exit the optimization if any of the following error conditions are met. These are:
/
/
Error/Exit 1.) Variance ratio constraint is violated (control with global constant stdRatio set in the header file)
/
Error/Exit 2.) Maximum # of iterations reached (control with global constant mIters set in the header file)
/
Error/Exit 3.) Any component probability becomes zero or negative while iterating
/
/ Inputs:
/
/
1.) The total # of observations (i.e., time points) in the current application. Specified by the user via the control file (2nd parameter). (T)
/
/
2.) An array of size T holding the observations/returns at each time point for the asset currently being processed. (r)
/
/
3.) The # of components in the univariate mixture density being fit. (g)
/
/
4.) An empty array of size g to hold the probability for each component in the univariate mixture. This array is updated during each iteration of the EM algorithm and therefore holds the optimal
/
probabilities upon convergence, which are returned in this array to the calling function. (Note that this parameter is a double array indexed as [t][g], where t is the thread # assigned by
/
ThrdEMAlg(), which invokes this function. The EM algorithm is implemented within threaded calls where a random start is assigned to its own thread.) (prbs)
/
/
5.) An empty array of size g to hold the mean for each component in the univariate mixture. This array is updated during each iteration of the EM algorithm and therefore holds the optimal means
/
upon convergence, which are returned in this array to the calling function. (Note that this parameter is a double array indexed as [t][g], where t is the thread # assigned by ThrdEMAlg(), which
/
invokes this function. The EM algorithm is implemented within threaded calls where a random start is assigned to its own thread.) (mns)
/
/
6.) An empty array of size g to hold the standard deviation for each component in the univariate mixture. This array is updated during each iteration of the EM algorithm and therefore holds the
/
optimal standard deviations upon convergence, which are returned in this array to the calling function. (Note that this parameter is a double array indexed as [t][g], where t is the thread #
/
assigned by ThrdEMAlg(), which invokes this function. The EM algorithm is implemented within threaded calls where a random start is assigned to its own thread.) (stds)
/
/
7.) An array to hold the log‐likelihood value that the EM algorithm converges to. It is the optimal value for the given random start. The array is indexed by thread #, therefore a single call to
/
this function generates 1 optimal log‐likelihood value which is returned in the element # for the current thread. (llVal)
/
/
8.) A double array to hold the univariate mixture density that is used to generate the random starts. This function begins by generating a random start, then optimizes the log‐likelihood function
/
based on that start (i.e., finds the nearest local maximum). This double array is indexed as [s][c], where s = 0, 1, 2, and c = component #. Here, s=0 refers to the component's probability,
/
s=1 refers to the component's mean, and s=2 refers to the component's standard deviation. (inMDist)
/
/
9.) An array of integer parameters holding values passed to and returned by this function. Element [0] holds the # of components in the mixture density that is used to generate the random starts,
/
which is parameterized in inMDist (parameter #8 to this function); element [1] is the thread # of the current call; element [2] is a return code which takes the following values: 0=variance
/
ratio constraint is violated, 1=maximum # of iterations reached, 2=convergence based on change to log‐likelihood value, 3=A component probability is not between 0.0 and 1.0; element [3] is the
/
iteration # of convergence or algorithm termination for one of the reasons decoded in element [2]. (rprms)
/
/ Outputs:
/
/
This function does not return a value at the call but updates several empty arrays that are supplied by the user. The prbs array is updated with the final estimated component probabilities, the mns
/
array is updated with the final estimated component means, the stds array is updated with the final estimated standard deviations, the llval array is updated with the final log‐likelihood value, and
/
the rprms array is updated at element [2] with the functions return code, and element [3] with the # iterations before convergence or termination.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void EMAlg(const int T, const long double *r, const int g, long double **prbs, long double **mns, long double **stds, long double *llVal, const long double **inMDist, int *rprms)
{
84
// Declare/initialize local variables.
//======================================
const int ing=rprms[0], thrd=rprms[1];
int stop;
long double oldllval, newllval, minstd, maxstd, cprbs, ssqrs[1], var, **pdens=nullptr, *postprbs=new long double [T], *mdens=nullptr, psum, **orMDst=new long double *[3],
**uMDst=new long double *[3];
// Generate g random obs from solution with rprms[0] components which are specified in inMDist[][]. Use these means to generate probabilities and standard deviations for each random start.
//=============================================================================================================================================================================================
do
{
stop=1; minstd=LPosVal; maxstd=LNegVal;
getRVals(g,ing,(const long double **) inMDist,mns[thrd]);
getRPrbsStds(T,r,g,prbs[thrd],mns[thrd],stds[thrd]);
// Check that none of the probabilities are zero in the random sample and that the variance ratio constraint is not violated in the random sample. Note that any
// component having standard deviation = 0 will disqualify that sample. Generate a new sample if a probability is zero or the variance ratio constraint is violated.
//=====================================================================================================================================================================
for (int c=0; c<g; ++c)
{
if (prbs[thrd][c] <= 0.00)
stop=0;
if (stds[thrd][c] < minstd)
minstd=stds[thrd][c];
if (stds[thrd][c] > maxstd)
maxstd=stds[thrd][c];
}
if (maxstd > stdRatio*minstd)
stop=0;
} while (stop == 0);
// Store random start and updated mixture distribution as 2‐dimensional array for debugging and compliance with other functions.
//================================================================================================================================
for (int m=0; m<3; ++m)
{
orMDst[m]=new long double [g];
uMDst[m]=new long double [g];
}
for (int c=0; c<g; ++c)
{
orMDst[0][c]=prbs[thrd][c];
orMDst[1][c]=mns[thrd][c];
orMDst[2][c]=stds[thrd][c];
}
// Store (component likelihood)x(component probability) (in pdens[][]) for each observation/component using the current solution. Store the mixture likelihood value (in mdens[]) for each observation
// using the current solution. These are needed to implement the updating equations. The log‐likelihood for the initial random start parameters is also computed here and used below.
//=====================================================================================================================================================================================================
pdens = new long double *[T];
mdens = new long double [T];
oldllval = ((long double) T)*logrecipsqrt2pi;
for (int t=0; t<T; ++t)
{
pdens[t] = new long double [g];
mdens[t] = 0.00;
for (int c=0; c<g; ++c)
{
pdens[t][c]=prbs[thrd][c]*getNDens(r[t],mns[thrd][c],stds[thrd][c]);
mdens[t]=mdens[t] + pdens[t][c];
}
oldllval = oldllval+log(mdens[t]);
}
// Iterate using the EM algorithm. Component probabilities are updated first and independently of the means/variances.
//=======================================================================================================================
85
stop=0;
int itcntr=1;
do
{
// Get updated component probabilities, means, and standard deviations. Store in temporary placeholders. The mean and standard deviation update formulas will not
// work (as written) when any component probability is zero because there will be a division by zero. If this happens end the optimization with an error.
//===================================================================================================================================================================
psum=0.00;
for (int c=0; (c<g && 0==stop); ++c)
{
// Derive the posterior probabilities for this component along with the component probabilities.
//================================================================================================
uMDst[0][c]=0.00;
if (c < g‐1)
// Derive component manually if not at last component.
{
for (int t=0; t<T; ++t)
{
postprbs[t]=(pdens[t][c])/mdens[t];
uMDst[0][c] = uMDst[0][c] + postprbs[t];
}
uMDst[0][c] = uMDst[0][c]/T;
psum = psum + uMDst[0][c];
}
else
// Final component probability is: 1.00 ‐ (sum of all others).
{
for (int t=0; t<T; ++t)
postprbs[t]=(pdens[t][c])/mdens[t];
uMDst[0][c] = 1.00 ‐ psum;
}
// Exit the optimization with appropriate code if any single probability is not between 0.0 and 1.0.
//====================================================================================================
if (uMDst[0][c] <= 0.00 || uMDst[0][c] >= 1.00)
{
stop=1; llVal[thrd]=LNegVal;
rprms[2]=3; rprms[3]=itcntr;
}
else
{
// New updating equations (faster processing).
//==============================================
cprbs=uMDst[0][c]*T;
uMDst[1][c]=getEMMean(T,r,postprbs,cprbs,ssqrs);
var=(ssqrs[0] ‐ uMDst[1][c]*uMDst[1][c]*cprbs)/cprbs;
// Updating equations can result in zero (but stored value is negative) variance. If it happens the variance ratio constraint is automatically violated.
//=========================================================================================================================================================
if (var > 0.00)
uMDst[2][c]=sqrt(var);
else
{
stop=1; llVal[thrd]=LNegVal;
rprms[2]=0; rprms[3]=itcntr;
}
}
}
// Find maximum and minimum StDev values. The EM algorithm will stop when the ratio of largest to smallest variance exceeds a constant (variable is stdRatio set in the header file). This
// prevents an unbounded likelihood value and once the constraint is violated we conclude that the solution is spurious. (The likelihood will be set to a large negative # to ensure it is never
// the maximum across random starts.) We also stop when the maximum # of iterations exceeds some value (mIters set in the header file) or when any component probability is not between 0 and 1.
//=================================================================================================================================================================================================
if (stop == 0)
{
minstd=LPosVal; maxstd=LNegVal;
for (int c=0; c<g; ++c)
86
{
if (uMDst[2][c] < minstd)
minstd = uMDst[2][c];
if (uMDst[2][c] > maxstd)
maxstd = uMDst[2][c];
}
if (maxstd > stdRatio*minstd)
{
stop=1; llVal[thrd]=LNegVal;
rprms[2]=0; rprms[3]=itcntr;
}
else if (itcntr > mIters)
{
stop=1; llVal[thrd]=LNegVal;
rprms[2]=1; rprms[3]=itcntr;
}
}
// If the variance ratio constraint is not violated then proceed as usual.
//==========================================================================
if (stop == 0)
{
// Transfer values to permanant placeholders (variance ratio constraint has not been violated).
//===============================================================================================
for (int c=0; c<g; ++c)
{
prbs[thrd][c]=uMDst[0][c];
mns[thrd][c]=uMDst[1][c];
stds[thrd][c]=uMDst[2][c];
}
// Store the (component likelihood)x(component probability) (in pdens[][]) for each component/observation using the updated solution. Store the mixture likelihood value (in mdens[]) for each
// observation using the updated solution. These are to implement the updating equations. The log‐likelihood for the initial random start parameters is also computed here and used below.
//=============================================================================================================================================================================================
newllval = ((long double) T)*logrecipsqrt2pi;
for (int t=0; t<T; ++t)
{
mdens[t] = 0;
for (int c=0; c<g; ++c)
{
pdens[t][c]=prbs[thrd][c]*getNDens(r[t],mns[thrd][c],stds[thrd][c]);
mdens[t]=mdens[t] + pdens[t][c];
}
newllval = newllval+log(mdens[t]);
}
// Terminate the algorithm when %‐change criteria is met.
//=========================================================
if ((newllval‐oldllval) < epsilon*abs(oldllval))
{
stop=1; llVal[thrd]=newllval;
rprms[2]=2; rprms[3]=itcntr;
}
else
{
oldllval=newllval;
++itcntr;
}
}
} while (stop == 0);
// Write error/issue files for debugging.
//=========================================
if (1 == rprms[2])
{
87
ofstream fout;
fout.open(errFolder + "Max_iterations_reached_" + to_string((long long) g) + "_" + to_string((long long) thrd) + ".txt");
fout << "‐‐> Maximum # of iterations (" << mIters << ") has been reached for this optimization. (Check the solution below.)" << endl << endl;
fout.setf(ios_base::fixed, ios_base::floatfield); fout.precision(25);
fout << "‐‐> New log‐likelihood value: " << newllval << endl;
fout << "‐‐> Epsilon = " << epsilon << endl << endl;
fout << endl << "‐‐> Original observation vector and parameter starting values for this EMAlg() call: " << endl << endl;
for (int x=0; x<T; ++x) fout << r[x] << endl; fout << endl;
for (int cc=0; cc<g; ++cc)
{
fout << "Prob[" << cc << "]=" << orMDst[0][cc] << " Mean[" << cc << "]=" << orMDst[1][cc] << " Std.Dev.[" << cc << "]=" << orMDst[2][cc] << endl;
}
fout.close();
}
else if (3 == rprms[2])
{
ofstream fout;
fout.open(errFolder + "Prob_of_0_or_1_in_" + to_string((long long) g) + "_" + to_string((long long) thrd) + ".txt");
fout << "‐‐> Probability of 0.00 or 1.00 has been encountered and updating equations will not work. (Check the solution below.)" << endl << endl;
fout.setf(ios_base::fixed, ios_base::floatfield); fout.precision(25);
fout << "‐‐> Original observation vector and parameter starting values for this EMAlg() call: " << endl << endl;
for (int x=0; x<T; ++x) fout << r[x] << endl; fout << endl;
for (int cc=0; cc<g; ++cc)
{
fout << "Prob[" << cc << "]=" << orMDst[0][cc] << " Mean[" << cc << "]=" << orMDst[1][cc] << " Std.Dev.[" << cc << "]=" << orMDst[2][cc] << endl;
}
fout.close();
}
// Delete temporary memory allocations.
//=======================================
for (int t=0; t<T; ++t)
{
delete [] pdens[t]; pdens[t]=nullptr;
}
delete [] pdens; pdens=nullptr;
delete [] mdens; mdens=nullptr;
for (int m=0; m<3; ++m)
{
delete [] orMDst[m]; orMDst[m]=nullptr;
delete [] uMDst[m]; uMDst[m]=nullptr;
}
delete [] orMDst; orMDst=nullptr;
delete [] uMDst; uMDst=nullptr;
delete [] postprbs; postprbs=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getRVals.cpp
/
/ Function: getRVals()
/
/ Summary:
/
/
This function accepts a mixture distribution as input and generates a random sample of observations from that distribution. The # of observations to generate is specified by the user in the call
/
and the sample is placed into an empty array supplied to this function by the user. Generating a random sample from a mixture distribution is a 2‐step process. (1) Generate a uniform random value
/
and compare it to the component probabilities to determine the component, and once the component is selected (2) generate an observation from the corresponding component density. Let p(i) be the
/
probability for component i and let there be a total of C components, then i=0,1,2,...,C‐1. If the uniform random value U <= p(0) then component 0 is selected, else if U <= p(0)+p(1) then component
/
1 is selected, else if U <= p(0)+p(1)+p(2) then component 2 is selected, etc... Once a component is selected an observation is generated from the corresponding density function. The result is an
88
/
observation generated from the supplied mixture distribution. Here, there are 2 reasons for generating random observations from a univariate mixture distribution as described below:
/
/
1.) In this application, univariate mixture distributions are fit using the EM algorithm with random starts. A random start must specify a value for all parameters in the given univariate mixture
/
density which is of known size, but with unknown parameters. The EM algorithm continues to increase the likelihood function until a local optimum is found based on the given parameter settings
/
from the random start. To optimize a g‐component mixture density when a g‐1 component mixture density is available from the same data set, we generate a single random start by producing g
/
random observations from the g‐1 component density. These g values are taken as means for the g‐component mixture. The observations are then assigned to the component with nearest mean by
/
using a standard distance computation. Once all observations are assigned to the closest mean, the standard deviation for each component (i.e., mean) is computed by taking the standard
/
deviation of the corresponding set of observations assigned to that mean. The probability assigned to each component is the # of observations assigned to the component divided by the total # of
/
observations in the data set. At this point, values for all parameters (i.e., g means, g standard deviations, g component probabilities) have been derived and the EM algorithm can be applied
/
starting at the given point in the parameter space. The EM algorithm then climbs to the top of the nearest hill and declares a local optimum. The maximum likelihood estimators would be the set
/
of parameters that yields the largest value for the likelihood function amongst the set of local optimums found via a large # of random starts.
/
/
2.) The likelihood ratio test (LRT) is used to select the optimal # of components for a univariate mixture density fit to a given observation set. The null hypothesis is that g0 components are
/
optimal vs the alternative of ga components being optimal (here, ga > g0). By fitting univariate mixtures of both sizes (g0 and ga) to the data we can generate a single value for the LRT
/
statistic. The value is compared to the null distribution of the LRT statistic, which is the distribution of the LRT under the assumption that a g0 component mixture is the appropriate size.
/
The distribution of the LRT under H0 is not known and is not asymptotically chi‐square due to the relevant regularity conditions not being satisfied. We can estimate the null distribution of
/
the LRT by bootstrapping (see, McLachlan (1987)). To bootstrap the LRT distribution we generate a random sample from the distribution specified by H0 of the same size as our data set. (Note
/
that H0 does not specify a particular univariate mixture, but rather just a # of components, namely g0. We take the ML Estimates of our data under g0 components as the distribution governed by
/
the null hypothesis, and this distribution is used to generate the sample of observations.) This sample is then fit to a univariate mixture under both g0 and ga components by applying the EM
/
algorithm using random starts, and a single value of the LRT statistic is produced. By repeating the process a large # of times we can approximate the null distribution of the LRT statistic.
/
We then compare the value derived from our data set and reject the null hypothesis when the LRT is large, where the critical point is determined by the user's choice of alpha = type 1 error for
/
the test. (Type 1 error probability = probability that the null hypothesis is rejected when it is true.) This function is used to generate the random samples used for the LRT just described.
/
/ Inputs:
/
/
1.) The # of random observations to generate from the supplied univariate mixture density. The observations are inserted into an empty array of the same size that is supplied by the user as the
/
last parameter to this function. (N)
/
/
2.) The # of components in the univariate mixture density that a sample will be generated from. (g)
/
/
3.) A double array holding the univariate mixture distribution definition from which a sample of size N will be generated. The array is indexed as [s][c], where s = 0, 1, 2, and c = component #.
/
Here, s=0 refers to the component's probability, s=1 refers to the component's mean, and s=2 refers to the component's standard deviation. (inMDst)
/
/
4.) An empty array of size N to be populated by this function as a random sample from the supplied univariate mixture distribution. (rvls)
/
/ Outputs:
/
/
This function populates an empty array with a random sample of user‐specified size from the supplied univariate mixture distribution. No value is returned at the function call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void getRVals(const int N, const int g, const long double **inMDist, long double *rvls)
{
// Generate N random observations from the current optimal solution that uses g components. These can be the means to use as a random start for fitting a specific model or observations used to
// approximate the null distribution of the LRT statistic.
//===============================================================================================================================================================================================
random_device rd;
default_random_engine gen(rd());
normal_distribution<long double> *ndist = new normal_distribution<long double> [g];
uniform_real_distribution<long double> udist(0.0,1.0);
// Define the array of normal distribution objects, one for each component.
//===========================================================================
for (int c=0; c<g; ++c)
ndist[c] = normal_distribution<long double> (inMDist[1][c],inMDist[2][c]);
// Generate N observations from the existing g‐component mixture distribution.
//==============================================================================
int cid;
long double uval, psum;
for (int i=0; i<N; ++i)
{
// Initialize variables.
//========================
cid=‐1;
psum=inMDist[0][0];
89
// Generate uniform random value.
//=================================
uval = udist(gen);
// Find the corresponding component.
//====================================
for (int c=0; (c<g && cid<0); ++c)
{
if (uval <= psum || c == g‐1)
cid=c;
else
psum = psum + inMDist[0][c+1];
}
// Generate a single obs from that component and store in array provided.
//=========================================================================
rvls[i] = ndist[cid](gen);
}
// Free up temporary memory allocations.
//========================================
delete [] ndist; ndist=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getRPrbsStds.cpp
/
/ Function: getRPrbsStds()
/
/ Summary:
/
/
This function accepts a set of T observations along with a set of g means for a g‐component mixture distribution and it derives corresponding values for the component probabilities and standard
/
deviations. Each observation is assigned to a single component (i.e., mean) and the # of observations assigned to a component divided by the total # of observations is the corresponding estimate
/
for that component's probability. Observations are assigned to components using a simple distance function, and specifically each observation is assigned to the component with nearest mean. The
/
standard deviation for each component is then estimated as the sample standard deviation of the set assigned to that component assuming the mean is known. Empty arrays of size g are supplied to
/
this function to hold the set of component probabilities and set of standard deviations, respectively. Combined with the existing set of g means (derived as a random sample via the function
/
getRVals()) these 3 arrays (i.e., mean, standard deviation, and component probability) completely define a g‐component mixture distribution. If the means were generated as a random sample, then it
/
defines a single random start for fitting a g‐component mixture distribution using the EM algorithm.
/
/ Inputs:
/
/
1.) The total # of time points with data collected. (T)
/
/
2.) An array holding the set of observations (i.e., returns) for the asset being processed. (r)
/
/
3.) The # of means (i.e., components) in the array provided (by parameter #5) used to generate a random start for a mixture density. (g)
/
/
4.) An empty array of size g to be populated by this function as component probabilities for the univariate mixture density being constructed. (prbs)
/
/
5.) An array of means (i.e., components) to use as the basis for generating a g‐component mixture distribution. (mns)
/
/
6.) An empty array of size g to be populated by this function as standard deviations for the univariate mixture density being constructed. (stds)
/
/ Outputs:
/
/
This function populates two empty arrays of size g supplied (prbs & stds) with component probability estimates and standard deviation estimates, for each of the g components defined by the means
/
supplied in mns. No value is returned at the function call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
90
void getRPrbsStds(const int T, const long double *r, const int g, long double *prbs, const long double *mns, long double *stds)
{
// Local variables.
//===================
long double *cssqrs=new long double [g], mindist;
int cid, *ccntr=new int [g];
// Initialize component counter and sum‐of‐squares arrays to all zeros.
//=======================================================================
for (int c=0; c<g; ++c)
{
ccntr[c]=0;
cssqrs[c]=0.0;
}
// Iterate over all observations and classify each into the component whose mean it is closest to.
//==================================================================================================
for (int n=0; n<T; ++n)
{
mindist=abs(r[n]‐mns[0]);
cid=0;
for (int c=1; c<g; ++c)
if (abs(r[n]‐mns[c]) < mindist)
{
cid=c;
mindist=abs(r[n]‐mns[c]);
}
++ccntr[cid];
cssqrs[cid]=cssqrs[cid]+pow(mindist,2);
}
// Compute the estimated probabilities and standard deviations (MLEs).
//======================================================================
for (int c=0; c<g; ++c)
{
prbs[c] = ((long double) ccntr[c]/(long double) T);
stds[c] = sqrt((long double) (cssqrs[c]/ccntr[c]));
}
// Delete temporary memory allocations.
//=======================================
delete [] ccntr; ccntr=nullptr;
delete [] cssqrs; cssqrs=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: asgnObs.cpp
/
/ Function: asgnObs()
/
/ Summary:
/
/
This function assigns each observation (which is a time point in this application) to one of the univariate mixture components for a given asset. Each asset has been fit to a univariate mixture
/
distribution containing a certain # of components. Each component can be viewed as a generator of observations for that asset, with the corresponding component probability that the observation
/
originates from any single component's density. Bayes' decision rule is used to assign each observation to a corresponding component by computing the posterior probability that each observation
/
originates from each component. The observation is then assigned to the component with the highest posterior probability. This function performs that task. The user supplies an empty array for
/
the given asset of size equal to the # of time points. Inserted into the array at each position is the component that the observation most likely originates from, and therefore is assigned to.
/
/ Inputs:
91
/
/
1.) An empty array of size T that will hold the component that the observation is assigned to. (Determined by this function.) (inAry)
/
/
2.) The total number of time points with data collected. (T)
/
/
3.) An array of size T holding the returns for the asset being processed. (r)
/
/
4.) The number of univariate components for the asset being processed. (g)
/
/
5.) The (optimal) univariate mixture distribution fit for the given asset as a 2‐dimensional array indexed as [c][s], where c = component #, and s = 0, 1, 2. Here, s=0 refers to the component's
/
probability, s=1 refers to the component's mean, and s=2 refers to the component's standard deviation. (inMDst)
/
/ Outputs:
/
/
This function populates an empty array of size T that is supplied with the component that the given observation most likely originates from. Bayes decision rule is used and the observation is
/
assigned to the component with highest posterior probability. This function returns no value at the call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void asgnObs(int *inAry, const int T, const long double *r, const int g, const long double **inMDst)
{
// Declare local variables.
//===========================
long double maxProb, tmpProb;
// Assign each observation per time point to the component with highest posterior probability (Bayes' rule).
//============================================================================================================
for (int t=0; t<T; ++t)
{
maxProb=‐1.00;
for (int c=0; c<g; ++c)
{
tmpProb=getPost(r[t], g, (const long double **) inMDst, c);
if (tmpProb > maxProb)
{
inAry[t]=c;
maxProb=tmpProb;
}
}
}
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: mapCells.cpp
/
/ Function: mapCells()
/
/ Summary:
/
/
A multidimensional grid is formed using the components from the univariate marginal mixture densities for all assets. If there are "a" assets with asset i having c[i] univariate mixture components,
/
for i=0,1,2,...,a‐1, then the multidimensional grid used as the basis for the multivariate density will have a total of c[0]xc[1]xc[2]x...xc[a‐1] cells. This function converts the multidimensional
/
grid to a single array/list holding all cells. Each value in this array/list represents one cell in the grid and therefore contains a set of components, one per asset. For example, the first
/
element of this list contains all assets at their first component, the 2nd element of this list contains all assets at their first component, but the final asset at it's 2nd component. The 3rd
/
element of this list contains all assets at their first component but the last asset at its 3rd component, etc...
The list is ordered as a design matrix for a full factorial experiment, from left
/
to right, with values to the right repeatedly cycling through all of their components for each set of values to the left. The function populates an empty list that is supplied. To keep track of
/
which component levels are used in any given cell, a 2nd term is added to the list making it a 2‐dimensional array indexed as [c][a], where c = unique cell # and a = asset #. That is, element
/
[0][0] would indicate the component of the first asset within the first cell of the multidimensional grid, and element [0][1] would indicate the component level for the 2nd asset within the first
/
cell of the multidimensional grid. If there are a total of 4 assets, then elements at positions [5][0], [5][1], [5][2], [5][3] would indicate the set of components for each of the 4 assets within
/
the 6th cell of the multidimensional grid. Essentially, this function converts the multidimensional grid to a list which is easier to manage and each cell in the list contains a 2nd term to
/
identify the contents of the list item (which is a single cell). Example: Suppose there are 3 assets with c[0]=2, c[1]=3, c[2]=2 components in the respective univariate mixture densities. The
92
/
multidimensional grid is formed by crossing all univariate components across assets and will contain 2x3x2=12 total cells. This grid forms the basis for building the multivariate mixture density.
/
A list of length 12 will be used to represent each cell and the components that are contained within each cell as follows:
/
/
=======
========
========
========
/
Cell ID
Asset #0
Asset #1
Asset #2
/
=======
========
========
========
/
0
0
0
0
/
1
0
0
1
/
2
0
1
0
/
3
0
1
1
/
4
0
2
0
/
5
0
2
1
/
6
1
0
0
/
7
1
0
1
/
8
1
1
0
/
9
1
1
1
/
10
1
2
0
/
11
1
2
1
/
=======
================================
/
Array Index:
[ ]
[ ]
/
/
For example, last row is defined as: inCellAry[11][0]=1, inCellAry[11][1]=2, inCellAry[11][2]=1.
/
/
Note that the list can be derived by starting with the last asset and repeatedly cycling through all component levels, then proceeding to the 2nd last asset and cycling through all component levels
/
for each set just defined, then proceeding to the 3rd last asset and cycling through all component levels for the sets defined to the right. This is the strategy used to convert the
/
multidimensional grid to a single list. This function is recursive with a single call for each asset. That call cycles through all levels of the given asset, invoking a call for the asset to the
/
right at each level. When at the final asset, no additional recursive calls are made and all levels of that asset are posted. In this manner, the function begins at asset #0, burrows inward to
/
asset #(a‐1), then expands outward back to the 1st asset. When the debug level is set to a value >= 2 all details of the mapping are printed for review, similar to the table shown above.
/
/ Inputs:
/
/
1.) The total # of cells in the multidimensional grid, which is: c[0]xc[1]xc[2]x...xc[a‐1], where there are "a" assets with asset i having c[i] components in the corresponding univariate mixture
/
density. (totCells)
/
/
2.) The number of assets in the current application. (numA)
/
/
3.) A 2‐dimensional array that is populated by this function indexed as [c][a], where c = unique cell #, and a = asset #. Here, both “c” and “a” begin at 0. The value contained is the component
/
level of asset a within unique cell c. (inCellAry)
/
/
4.) The current asset that is being processed. This function processes each asset separately and iterates over all of its component levels. At each component level the function recursively invokes
/
itself for the next asset, which similarly processes each component in order. (curAst)
/
/
5.) An array to hold the # of components for each asset in their respective univariate mixture density. This array will hold the values for c[i] as described above. (inCmps)
/
/
6.) An integer value to hold the current unique cell #. Value begins at 0 and ends at totCells‐1. After each cell is defined and written to the inCellAry[][] array this value is incremented. (cID)
/
/
7.) An array to hold the component level being processed for each asset. This function starts at the first asset (i.e., a=0) and iterates over all c[0] levels of the corresponding univariate
/
mixture. At each level, the function is invoked recursively to process the next asset (i.e., a=1). The function then iterates over all c[1] levels of the 2nd asset and recursively invokes
/
itself to process the next asset (i.e., a=2). When at the final asset a cell is completely defined and the result is appended to the list. This array is persistant and of size numA to hold the
/
current component level being processed for each asset. When at the final asset (i.e., final recursive call) the numA components held in this array defines the given unique cell. (tmpAry)
/
/ Outputs:
/
/
This function converts the multidimensional grid formed by crossing all assets and their univariate component levels to a single list where each element in the list defines one cell of the
/
multidimensional grid. It is more straightforward to navigate the grid in this manner. The list has 2 elements, first is the unique cell # and 2nd is the array of component levels that define the
/
given cell. This function is recursive (invokes itself) and returns no value at the call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void mapCells(const int totCells, const int numA, int **inCellAry, const int curAst, const int *inCmps, int *cID, int *tmpAry)
{
// Output cell mappings when debugging mode is on.
//==================================================
if (0 == curAst && dbug >= 2)
cout << string(14,'=') << endl << "Cell Mappings: " << endl << string(14,'=') << endl;
93
// Iterate over all components of the current asset. At each component recursively invoke this function to process the next asset.
//===================================================================================================================================
for (int i=0; i<inCmps[curAst]; ++i)
{
// Store the component level for the current asset.
//===================================================
tmpAry[curAst]=i;
// Recursive call, if not processing the final asset. If processing final asset, populate the array and increment the cell counter.
//====================================================================================================================================
if (curAst < numA‐1)
mapCells(totCells, numA, inCellAry, curAst+1, inCmps, cID, tmpAry);
else
{
if (dbug >= 2)
cout << "Cell #" << setfill('0') << setw(to_string((long long) totCells).size()) << *cID;
for (int j=0; j<numA; ++j)
{
if (dbug >= 2)
cout << ", Asset #" << j+1 << " = " << tmpAry[j];
inCellAry[*cID][j] = tmpAry[j];
}
if (dbug >= 2)
cout << endl;
*cID = *cID + 1;
}
}
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getCell.cpp
/
/ Function: getCell()
/
/ Summary:
/
/
This function accepts a set of component levels (one per asset) and returns the unique cell ID # from the multidimensional grid that contains this set. The multi‐dimensional grid contains a cell
/
for each combination of univariate mixture components across all assets. If no match is found a value of ‐1 is returned. This function iterates over the 2‐dimensional array that is supplied until
/
a match is found, then returns the cell ID # and exits. The 2‐dimensional array supplied is indexed as [c][a], where c = unique cell #, and a = asset #.
/
/ Inputs:
/
/
1.) 2‐dimensional array as [c][a], where c=unique cell #, a=asset #. Here, both "c" & "a" begin at 0. The value contained is the component level of asset "a" within unique cell "c". (inCellAry)
/
/
2.) The total # cells in the multidimensional grid: c[0]xc[1]xc[2]x...xc[a‐1], there are "a" assets with asset i having c[i] components in the corresponding univariate mixture PDF. (totCells)
/
/
3.) An array of size numA containing the set of asset component levels we are attempting to match. The unique cell # for the match is returned at the call. (cmpLvls)
/
/
4.) The number of assets in the current application. (numA)
/
/ Outputs:
/
/
This function searches for a match on a set of numA asset component levels and returns the unique cell ID #. The cell ID # ranges from 0 ‐ (totCells‐1). Return ‐1 if no match is found.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int getCell(const int **inCellAry, const int totCells, const int *cmpLvls, const int numA)
{
// Iterate over all mapped values and find the match.
//=====================================================
94
int cntr;
for (int i=0; i<totCells; ++i)
{
cntr=0;
for (int j=0; j<numA; ++j)
if (inCellAry[i][j] == cmpLvls[j])
++cntr;
// Check for match, then return the cell position of the match.
//===============================================================
if (cntr == numA)
{
if (dbug >= 2)
cout << "Cell = " << i << endl;
return i; // Match found.
}
}
return ‐1;
// No match found.
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: solveLP.cpp
/
/ Function: solveLP()
/
/ Summary:
/
/
This function solves 2 linear programs (LPs) which become (feasible) initial solutions for maximizing the multivariate mixture distribution likelihood. The univariate mixture densities for each
/
asset have already been fit using the EM algorithm and then combined into a multi‐dimensional grid where the grid levels in each dimension are the corresponding asset/mixture component
/
combinations. This multidimensional grid forms the basis for the multivariate density function, which is also a mixture PDF. Each cell in the multivariate grid defines a unique combination of
/
assets and their components. Using Bayes' decision rule we can assign the observation for a given asset at each time point to a corresponding univariate component based on the component with
/
highest probability of membership. Using these individual component memberships we combine them and assign each multivariate observation to a single cell in the multidimensional grid. The
/
probability of an observation originating from that grid cell is then the # of observations in the given cell divided by the total # of observations (or time points). We now have an estimated
/
probability that a new observations originates from each multi‐dimensional grid cell. (Refer to this estimate as ek, applicable to cell k.) Note that each grid cell defines a multivariate density
/
function using the corresponding means and variances for the assets/components that define the cell, but at this point all covariances are undefined (i.e., zero). An important aspect of this
/
research is that we must maintain the univariate marginals that have already been fit. To accomplish this, the sum of probabilities for each cell containing a given asset/component must equal the
/
corresponding probability for that component in the univariate density. This implies that we can use linear constraints on the grid cell probabilities to maintain the marginals. Since the sum of
/
all grid cell probabilites must equal 1.00, there will be (# univariate components) ‐ 1 constraints needed to maintain the univariate marginal for each asset. (Once these constraints are enforced
/
the final (sum of probabilities) constraint for each component is automatically enforced by the fact that the sum of all probabilities equals 1.00.) This means that there will be a total of (Total
/
# of components across all assets ‐ Total # of assets + 1) constraints required to maintain the marginals. A set of linear equality constraints can be formulated in matrix notation as Ax = b. Here,
/
A is a matrix with one row holding the coefficients (0 or 1) for a single constraint and the matrix A has a column for each decision variable (i.e., single grid cell probability which is a
/
probability for the multivariate mixture density). We will refer to A as the LHS constraint matrix and b as the RHS constraint vector. The vector b will hold the component probabilities from the
/
univariate marginals, with only the first (# components) ‐ 1 needed since the probability for the final component within each asset is automatically enforced by the last row of A (i.e., final
/
constraint) that the sum of all probabilities equals 1.00. Note that the LHS constraint matrix should be of full row rank, meaning that we include the minimum # of constraints needed to enforce the
/
marginals. The rows of A must be linearly independent. (This requirement is needed for a future optimization that uses this matrix.) The cell probabilities that have been estimated using the data
/
(i.e., the ek via Bayes' decision rule, see above) will generally not satisfy the marginal constraints in Ax = b, therefore these estimates will not (in general) maintain the marginals. The purpose
/
of this function is to find cell probabilities that do maintain the marginals, and that are in some way as close as possible to the estimated probabilities. We offer 2 methods here. First, to
/
formulate the problem we assign an unknown decision variable to each unique cell in the multidimensional grid and that represents the true probability of membership in that cell. These decision
/
variables are then estimated using the following 2 LP objectives: 1.) Minimize the maximum distance between all estimated unique cell probabilties and the decision variable that represents each
/
cell, and 2.) Minimize the sum of squared distances between the decision variables and the estimated unique cell probabilities. LP #1 is a classic minimax objective of the form:
/
Min{Max[|p1‐e1|,|p2‐e2|,|p3‐e3|, ..., |pu‐eu|]}, where there are u unique cells in the multidimensional grid and pk is the decision variable (true cell probability) for cell k, and ek is the
/
corresponding estimated cell probability. The distance between the two is: |pk‐ek|. In LP #1 our objective is to select the pk, k=1,2,...,u that minimizes the maximum of these distance values such
/
that the constraints Ax = b are satisfied (i.e., the marginal densities are maintained). As written, the objective in LP #1 contains absolute values and therefore is not linear, however, it can be
/
rewritten as an equivalent linear program. For example, note that the objective can be rewritten as: Min{Max[p1‐e1,e1‐p1, p2‐e2,e2‐p2, p3‐e3,e3‐p3, ..., pu‐eu,eu‐pu]} since |pk‐ek| = pk‐ek or
/
ek‐pk, for k=1,2,...,u. The absolute values have now been removed. Next, let Z = Max[p1‐e1,e1‐p1, p2‐e2,e2‐p2, p3‐e3,e3‐p3, ..., pu‐eu,eu‐pu], then the objective becomes Min{Z}, and note that the
/
following inequalities must hold: Z >= p1‐e1, Z >= e1‐p1, Z >= p2‐e2, Z >= e2‐p2, ... Z >= pu‐eu, Z >= eu‐pu. This is because Z is the maximum of a set of values therefore it must be >= all
/
members of the set. Further, since the objective is to minimize Z, it must take one of the values that bounds it below at an optimal solution. Using the new objective and added constraints, the
/
problem is now an equivalent linear program. Lastly, it will be important to keep the multivariate density as parsimonious as possible, meaning carrying fewer unique cells into the multivariate
/
mixture density is desirable. Fewer components translates to fewer parameters and also we will run into problems when optimizing the multivariate mixture likelihood when a component is included
95
/
that generates a zero likelihood value for the set of observations at all time points. A future gradient/Hessian computation will see the component probabilities prefixed by the corresponding
/
likelihood value in the objective. If the likelihood of this multivariate component density is zero for all time points then the decision variable is effectively removed from the problem and the
/
corresponding Hessian will not be of full rank. This will be a problem when applying Newton's method, for example. To prevent components with zero likelihood from being carried merely to satisfy
/
the constraints, we will add a 2nd decision variable to each unique cell in the multidimensional grid that penalizes the objective function when a cell with zero observations is included in the LP
/
solution. This will help to guarantee that we only keep cells that contain actual data points and it prevents likelihoods from being zero at all time points. LP #2 is similar to LP #1 but the
/
objective is to minimize the sum of squared distances between the actual and estimated probabilities, subject to the marginal constraints. That is, in LP #2 the objective takes the form:
/
Min {(p1‐e1)^2 + (p2‐e2)^2 + (p3‐e3)^2 + ... + (pu‐eu)^2}. Note that each term in this sum (pk‐ek)^2, k=1,2,...,u is quadratic and concave (U‐shaped) in pk and centered at ek. Furthermore, no
/
decision variable exists in more than 1 term of the sum. This objective is known as a quadratic program that is separable, which by definition means it is not linear. It can, however, be
/
approximated arbitrarily close by a linear program. Since 0 <= pk <= 1, for k=1,2,...,u, we will define a set of decision variables to approximate each concave quadratic term in the sum over this
/
pk range. The approximation will consist of line segments that trace out the U‐shaped term. For the range 0 <= pk <= 1, we first set the number of line segments to use when tracing out the curve,
/
and this is done via the global constant dLvl defined in the header file (current value=500). Next we compute the pk values on the horizontal axis (i.e., the probability axis) that are equidistant
/
and cover the region between 0 and 1. Note that there will be dLvl+1 such points. At the current setting of 500 line segments, these points will be: 0, 1/500, 2/500, 3/500, ..., 1. They are
/
fixed once dLvl is known and they also do not change per term in the sum. That is, these points on the probability axis are used to trace out each concave term in the objective. Once these points
/
are known we can compute the function evaluated at each point. For example, considering the first term in the sum, the function evaluated at each point is: (0‐e1)^2, (1/500‐e1)^2, (2/500‐e1)^2,
/
..., (1‐e1)^2. These values are fixed and constant once dLvl is determined and the estimates ek are known, but they do change per term in the sum. If there are a total of U components then there
/
will be U*(dLvl+1) such quantities defined. (Connecting the dots of these function values will trace out the curve.) In LP #1 the decision variables were the true probabilities, for each unique
/
cell but in LP #2 the decision variables are dLvl+1 alpha values for each term in the sum that are >= 0 and sum to 1. Each such alpha variable is attached to a segment boundary on the horizontal
/
axis (i.e., 0, 1/500, 2/500, ..., 1). A probability value pk is then defined by using the 2 alpha variables that bound the pk value. For example, we can create the value (1.3)/500 by using
/
(alpha1)*(1/500) + (alpha2)*(2/500), where alpha1=0.7 and alpha2=0.3, and all other alpha values = 0.0.
That is, in general, we define pk=(alpha_k_0)*(0) + (alpha_k_1)*(1/500) +
/
(alpha_k_2)*(2/500) + (alpha_k_3)*(3/500) + ... + (alpha_k_500)*(1). Note that the alpha variables are specific to a component of the sum. Lastly, the quadratic objective component (pk‐ek)^2 is
/
estimated arbitrarily close by (alpha_k_0)*(0‐ek)^2 + (alpha_k_1)*(1/500‐ek)^2 + (alpha_k_2)*(2/500‐ek)^2 + (alpha_k_3)*(3/500‐ek)^2 + ... + (alpha_k_500)*(1‐ek)^2, where ek is the known cell
/
estimate. These constant multipliers were defined above and are stored in variables. Now, both the objective function and constraints are linear in the U*(dLvl+1) alpha decision variables. Within
/
a component, all alpha variables must sum to 1.00, that is: alpha_k_0 + alpha_k_1 + alpha_k_2 + ... + alpha_k_500 = 1.00. Once the pk are defined, we use these variables to build the constraints
/
that maintain the marginals, that is Ax = b. We now have a linear objective, with linear constraints that approximates the quadratic separable program. Linear programs can be solved fast for a
/
global solution using the simplex algorithm. Here, we use the lp_solve free library of functions to solve both LPs.
/
/ Variable Summary:
/
/
Minimax Objective:
/
==================
/
/
Total # of decision variables defined in the array vbl[]: 2*totCells + 1
/
/
vbl[0] ‐ vbl[totCells‐1] will hold the (totCells ‐ 1 ‐ 0 + 1) = totCells "probability" decision variables (pk, for k=1,2,...,u) as detailed above. These are referenced in the LP as column #'s: 1,
/
2, 3, ..., totCells.
/
/
vbl[totCells] ‐ vbl[2*totCells‐1] will hold the (2*totCells ‐ 1 ‐ totCells + 1) = totCells "feasibility factor" decision variables (fk, for k=1,2,...,u). These are referenced in the LP as column
/
#'s: totCells+1, totCells+2, totCells+3, ..., 2*totCells. One constraint is that the true cell probability (i.e., value of the decision variable) must be <= the total # of observations that fall
/
in a cell plus the cell's corresponding feasiblity factor. Thus, when a cell with zero actual observations assigned to it needs a non‐zero probability for an optimal solution this feasibility
/
factor must be set to a positive value to satisfy the constraint. In the objective we then add each of these feasibility factors multiplied by the large constant K (set in the header file). This
/
applies a penalty to the objective when a cell with no observations is included in the optimal solution, and it makes the occurence rare. As noted above we want to avoid including grid cells that
/
have a zero likelihood across all time points since it will lead to the Hessian of a future optimization being non‐singular.
/
/
vbl[2*totCells] holds the objective function decision variable. This is referenced in the LP as column #: 2*totCells+1.
/
/
Minimum Sum of Squared Distances Objective:
/
===========================================
/
/
Total # of decision variables defined in the array vbl[]: totCells*(dLvl+2)
/
/
vbl[0] ‐ vbl[dLvl] will hold the alpha values (which sum to 1.00) within the 1st unique cell (index=0) of the multidimensional grid, say, alpha_0_0, alpha_0_1, alpha_0_2, ..., alpha_0_500. The
/
corresponding optimal probability value p0 is then derived using the 2 non‐zero values combined with the corresponding 2 segment endpoints.
/
/
vbl[dlvl+1] ‐ vbl[2dLvl+1] will hold the alpha values (which sum to 1.00) within the 2nd unique cell (index=1) of the multidimensional grid, say, alpha_1_0, alpha_1_1, alpha_1_2, ...,
/
alpha_1_500. The corresponding optimal probability value p1 is then derived using the 2 non‐zero values combined with the corresponding 2 segment endpoints.
... Etc ... (cont. below)
/
/
vbl[(totCells‐1)*dlvl+(totCells‐1)] ‐ vbl[(totCells)*dLvl+(totCells‐1)] will hold the alpha values (which sum to 1.00) within the last unique cell (index=totCells‐1) of the multidimensional grid,
/
say, alpha_(totCells‐1)_0, alpha_(totCells‐1)_1, alpha_(totCells‐1)_2, ..., alpha_(totCells‐1)_500. The corresponding optimal probability value p2 is then derived using the 2 non‐zero values
/
combined with the corresponding 2 segment endpoints.
/
/
vbl[(totCells)*dLvl+totCells] ‐ vbl[(totCells)*(dLvl+2)‐1] will hold the single feasiblity factor assigned per cell. The corresponding cell probability must be <= the # observations which are
/
assigned to that cell + this variable. When there are 0 observations in a cell and it is required to be non‐zero in an optimal solution then the feasibility factor must be forced to a non‐zero
/
(positive) value. The objective function contains a term for each feasibility factor multiplied by a large constant (see, K=1000000 in the header file). This serves as a penalty on the objective
/
when a grid cell with no observations from the actual data is included in the optimal solution. We want to avoid such solutions whenever possible since it can lead to the Hessian matrix of an
/
upcoming optimization being non full rank.
96
/
/ Inputs:
/
/
1.) The total # of unique cells in the multidimensional grid formed by combining all components across the estimated univariate mixture density functions for all assets. For example, if there are a
/
total of 4 assets being considered and the univariate mixture densities have 2, 3, 4, 2 levels, respectively, the complete multidimensional grid will have 2x3x4x2 = 48 unique cells. This
/
function provides 2 methods for determining which cells are important and needed in the multivariate mixture density. (totCells)
/
/
2.) The total number of assets being considered in the problem. (numA)
/
/
3.) A double array indexed as [c][a], with "c" being a unique cell ID (values are from 0 ‐ totCells‐1), and "a" being an asset ID (values are from 0 ‐ numA‐1). The value held at this position is
/
the univariate mixture component level for asset "a" within unique cell "c". Recall that the multidimensional grid is formed by crossing all assets/components with all other assets/components.
/
(Here component refers back the the univariate mixture density for the asset.) (inCellAry)
/
/
4.) An array of size numA holding the # of univariate mixture components for each asset. Index "a" of this array will return the number of components that were needed to fit the univariate mixture
/
for component "a". (Range is a=0 to numA‐1.) (cmps)
/
/
5.) A double array indexed as [a][g], with "a" being an asset ID (from 0 ‐ numA‐1), and "g" being a component ID for asset "a" ("g" ranges from 0 ‐ cmps[a]‐1). The corresponding univariate mixture
/
component probability is stored at the indexed position. (prbs)
/
/
6.) An array of size totCells that holds the number of observations (i.e., time points) that are assigned to the given unique cell of the multidimensional grid. Observations are assigned to
/
specific components of an asset using Bayes' decision rule, that is they are assigned to the component with highest probability of membership. Once a time point has been processed, it is
/
assigned to a component for each asset and this defines the cell of the multidimensional grid that it is assigned to. For example, nCellObs[4] = 11 implies that there are 11 observations which
/
fall into unique cell ID #4. (nCellObs)
/
/
7.) An array of estimated cell probabilities, ek, k=1,2,...,totCells. These are derived as nCellObs[c]/T and drive both LP optimizations. (cellProb)
/
/
8.) An empty array of true probabilities estimated by the 2 LPs and populated by this function. This array will be of size totCells. (outPrbs)
/
/
9.) The type of LP to use for a given function call, where 0 = Minimax objective, 1 = Minimum Sum of Squared Distances (SSD) objective. (type)
/
/ Outputs:
/
/
This function counts the number of non‐zero probabilities in the array outPrbs and returns this value at the call. This function also populates the empty array outPrbs with the estimated true
/
probability for each unique cell which is "close" to the estimated ek values (using Bayes' decision rule), but that satisfies the marginal constraints.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int solveLP(const int totCells, const int numA, const int **inCellAry, const int *cmps, const long double **prbs, const int *nCellObs, const long double *cellProb, double *outPrbs, const int type)
{
// Initialize local variables. (Variable dVars=total # of decision variables for the given LP and it depends on type.)
//=======================================================================================================================
lprec *lp=NULL;
int dVars=(int) (2*totCells+1)*(0==type) + (totCells*(dLvl+2))*(1==type), *vnum=new int [dVars], j, non0cmps=0;
double *vbl=new double [dVars], objval, *epnt=new double [dLvl+1], *fpnt=new double [dVars‐totCells];
string strlabel;
char **vlabels=new char *[dVars];
// Build LP model to derive the joint density with zero covariances.
//====================================================================
lp = make_lp(0,dVars);
if (lp == NULL)
{
cout << "ERROR: LP Model did not build. Something is wrong with the lp_solve setup (type=" << type << ")." << endl << "EXITING...solveLP()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Add labels to the decision variables. Include the cell index and asset/component values for minimax objective. Include the cell index and alpha index for minimum squared distance objective.
//==================================================================================================================================================================================================
if (0 == type)
// Minimax objective.
{
for (int c=0; c<totCells; ++c)
{
strlabel="Cell_" + to_string((long long) c) + "_";
for (int a=0; a<numA; ++a)
{
strlabel=strlabel + "A" + to_string((long long) a+1) + "C" + to_string((long long) inCellAry[c][a]+1);
if (a < (numA‐1))
97
strlabel=strlabel + "_";
}
vlabels[c]=new char [strlabel.size()+1];
strcpy_s(vlabels[c], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, c+1, vlabels[c]);
// Now label the corresponding feasibility factor. The decision variables come in (probability, feasibility factor) pairs.
//===========================================================================================================================
strlabel="F_" + strlabel;
vlabels[totCells+c]=new char [strlabel.size()+1];
strcpy_s(vlabels[totCells+c], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, totCells+c+1, vlabels[totCells+c]);
}
strlabel="Z";
vlabels[dVars‐1]=new char [strlabel.size()+1];
strcpy_s(vlabels[dVars‐1], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, (int) dVars, vlabels[dVars‐1]);
}
else if (1 == type)
// Sum of squared distances objective.
for (int c=0; c<totCells; ++c)
{
for (int s=0; s<=dLvl; ++s)
{
// Build array of constants for the piecewise linear function endpoints. There are dLvl segments, therefore dLvl+1 endpoints and these
// do not change by cell. Build an array of corresponding function endpoints, which do change by cell since the est. prob changes.
//======================================================================================================================================
if (0 == c)
epnt[s]=(double) (s*(1.00/dLvl));
fpnt[c*(dLvl+1)+s]=pow((epnt[s] ‐ (double) cellProb[c]),2);
// Labels for alpha decision‐variables using minimum squared distance objective.
//================================================================================
strlabel="Cell_" + to_string((long long) c) + "_Alpha_" + to_string((long long) s);
vlabels[c*(dLvl+1)+s]=new char [strlabel.size()+1];
strcpy_s(vlabels[c*(dLvl+1)+s], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, c*(dLvl+1)+s+1, vlabels[c*(dLvl+1)+s]);
}
// Label feasibility factor decision‐variables using minimum squared distance objective.
//========================================================================================
strlabel="F_Cell_" + to_string((long long) c) + "_";
for (int a=0; a<numA; ++a)
{
strlabel=strlabel + "A" + to_string((long long) a+1) + "C" + to_string((long long) inCellAry[c][a]+1);
if (a < (numA‐1))
strlabel=strlabel + "_";
}
vlabels[totCells*(dLvl+1) + c]=new char [strlabel.size()+1];
strcpy_s(vlabels[totCells*(dLvl+1) + c], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, totCells*(dLvl+1) + (c+1), vlabels[totCells*(dLvl+1) + c]);
}
// Add marginal constraints on the cell probabilities. Sum of all probabilities attached to an Asset/Component must equal that Asset/Component probability. For each asset with G components, once the
// first G‐1 constraints have been satisfied the G‐th constraint is set since probabilities sum to 1.00. We have not added the sum to 1.00 constraint thus will keep all G for each component.
//=====================================================================================================================================================================================================
set_add_rowmode(lp, TRUE);
for (int a=0; a<numA; ++a)
for (int g=0; g<cmps[a]; ++g)
{
j=0;
for (int c=0; c<totCells; ++c)
if (inCellAry[c][a] == g)
{
if (0 == type)
// Minimax objective.
{
98
vnum[j]=c+1;
vbl[j++]=1.00;
}
else if (1 == type)
// Sum of squared distances objective.
for (int s=0; s<=dLvl; ++s)
{
vnum[j]=(c*(dLvl+1)+s)+1;
vbl[j++]=epnt[s];
}
}
if (!add_constraintex(lp, j, vbl, vnum, EQ, (double) prbs[a][g]))
{
cout << "ERROR: [LP Issue] Marginal probability constraint for Asset=" << a << "/Component=" << g << " failed to load (type=" << type << ")." << endl
<< "EXITING...solveLP()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
}
// Add feasibility constraints on the cell probabilities when using a minimax or minimum squared distance objective. When an individual cell has zero observations assigned to it, force the cell
// probability to zero (relax if there is no feasible solution).
//=================================================================================================================================================================================================
for (int c=0; c<(int) totCells; ++c)
{
j=0;
if (0 == type)
{
vnum[j]=c+1; vbl[j++]=1.00;
vnum[j]=totCells+c+1; vbl[j++]=‐1.00;
}
else if (1 == type)
{
for (int s=0; s<=dLvl; ++s)
{
vnum[j]=c*(dLvl+1)+s+1;
vbl[j++]=epnt[s];
}
vnum[j]=totCells*(dLvl+1) + (c+1);
vbl[j++]=‐1.00;
}
if (!add_constraintex(lp, j, vbl, vnum, LE, (double) nCellObs[c]))
{
cout << "ERROR: [LP Issue] Feasibility factor constraint for Cell=" << c << " failed to load (type=" << type << ")." << endl << "EXITING...solveLP()..." << endl; cin.get();
exit (EXIT_FAILURE);
}
}
// Add inequality constraints on the minimax objective function. Objective is transformed into an LP using appropriate inequality constraints.
//===============================================================================================================================================
if (0 == type)
for (int c=0; c<(int) totCells; ++c)
{
// First absolute value constraint for the objective function pertaining to this cell probability.
//==================================================================================================
j=0;
vnum[j]=(int) dVars; vbl[j++]=1.00;
vnum[j]=c+1; vbl[j++]=‐1.00;
if (!add_constraintex(lp, j, vbl, vnum, GE, ‐1.00*cellProb[c]))
{
cout << "ERROR: [LP Issue] Minimax objective function absolute value constraint (#1) for Cell=" << c << " failed to load." << endl << "EXITING...solveLP()..." << endl; cin.get();
exit (EXIT_FAILURE);
}
// Second absolute value constraint for the objective function pertaining to this cell probability.
//===================================================================================================
j=0;
99
vnum[j]=(int) dVars; vbl[j++]=1.00;
vnum[j]=c+1; vbl[j++]=1.00;
if (!add_constraintex(lp, j, vbl, vnum, GE, cellProb[c]))
{
cout << "ERROR: [LP Issue] Minimax objective function absolute value constraint (#2) for Cell=" << c << " failed to load." << endl << "EXITING...solveLP()..." << endl; cin.get();
exit (EXIT_FAILURE);
}
}
// The sum of squared distances objective requires a constraint that the decision variables sum to 1 within each cell.
//======================================================================================================================
if (1 == type)
for (int c=0; c<(int) totCells; ++c)
{
j=0;
for (int s=0; s<=dLvl; ++s)
{
vnum[j]=(c*(dLvl+1)+s)+1;
vbl[j++]=1.00;
}
if (!add_constraintex(lp, j, vbl, vnum, EQ, 1.00))
{
cout << "ERROR: [LP Issue] Sum of squared distances constraint that sum of " << dLvl+1 << " decision variables = 1.00 within Cell=" << c << "failed to load."
<< endl << "EXITING...solveLP()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
}
// Add minimization objective. Output the entire LP formulation when requested.
//================================================================================
set_add_rowmode(lp, FALSE);
j=0;
if (0 == type)
// Minimax objective.
{
vnum[j]=(int) dVars; vbl[j++]=1.00;
for (int c=0; c<totCells; ++c)
{
vnum[j]=totCells+c+1;
vbl[j++]=K;
}
}
else if (1 == type)
// Sum of squared distances objective.
for (int c=0; c<totCells; ++c)
{
for (int s=0; s<=dLvl; ++s)
{
vnum[j]=c*(dLvl+1)+s+1;
vbl[j++]=fpnt[c*(dLvl+1)+s];
}
vnum[j]=totCells*(dLvl+1) + (c+1);
vbl[j++]=K;
}
if (!set_obj_fnex(lp, j, vbl, vnum))
{
cout << "ERROR: [LP Issue] Objective failed to load. (Type=" << type << ")." << endl << "EXITING...solveLP()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
set_minim(lp);
if (dbug >= 2)
{
// Note: Uncomment to write out full LP details when debugging. (Note: Output can be large.)
//==============================================================================================
/* cout << string(11,'=') << endl << "LP Details:" << endl << string(11,'=') << endl;
write_LP(lp, stdout); */
}
100
// Solve the LP and retrieve the results.
//=========================================
set_verbose(lp, IMPORTANT);
if (solve(lp) != OPTIMAL)
{
cout << "ERROR: [LP Issue] No solution found, something has gone wrong.
}
(Type=" << type << ")." << endl << "EXITING...solveLP()..." << endl; cin.get(); exit (EXIT_FAILURE);
// Get the objective function value as well as the value of the unknown probabilities that define the multivariate density. (Output the values.)
//=================================================================================================================================================
objval=get_objective(lp);
get_variables(lp, vbl);
// Compute the probabilities. For Minimax these are the values from the first (Total # Cells) decision variables. For Min SSD objective, these are the weighted sum of the dLvl+1 decision variables.
//=====================================================================================================================================================================================================
for (int c=0; c<totCells; ++c)
{
if (0 == type)
outPrbs[c] = vbl[c];
else if (1 == type)
{
outPrbs[c] = 0.00;
for (int s=0; s<=dLvl; ++s)
outPrbs[c] = outPrbs[c] + vbl[c*(dLvl+1)+s]*epnt[s];
}
}
// Write out the LP estimated probabilities along with the empirical values and corresponding distance between the estimated probabilities LP solution probabilities when debugging is requested.
//=================================================================================================================================================================================================
if (dbug >= 2)
{
// Add correct labels if using sum of squared distances objective.
//==================================================================
if (1 == type )
for (int c=0; c<totCells; ++c)
{
strlabel="Cell_" + to_string((long long) c) + "_";
for (int a=0; a<numA; ++a)
{
strlabel=strlabel + "A" + to_string((long long) a+1) + "C" + to_string((long long) inCellAry[c][a]+1);
if (a < (numA‐1))
strlabel=strlabel + "_";
}
vlabels[c]=new char [strlabel.size()+1];
strcpy_s(vlabels[c], strlabel.size()+1, strlabel.c_str());
set_col_name(lp, c+1, vlabels[c]);
}
// Output the estimated probabilities along with the actual and absolute distance between the values.
//=====================================================================================================
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(20);
cout << endl << "Objective function (Type=" << type << ") value = " << objval << endl;
cout << "LP Solution:" << endl << endl;
for (int c=0; c<totCells; ++c)
cout << get_col_name(lp, c+1) << " = " << outPrbs[c] << " vs. (actual) " << cellProb[c] << " and abs(diff) = " << abs(outPrbs[c]‐cellProb[c]) << endl;
if (0 == type)
for (int c=totCells; c<(int) 2*totCells+1; ++c)
cout << get_col_name(lp, c+1) << " = " << vbl[c] << endl;
else if (1 == type)
for (int c=totCells*(dLvl+1); c<totCells*(dLvl+2); ++c)
cout << get_col_name(lp, c+1) << " = " << vbl[c] << endl;
}
// Issue a warning if a cell with zero observations is assigned a non‐zero probability.
//=======================================================================================
101
for (int c=0; c<totCells; ++c)
if (cellProb[c] <= 0.00 && outPrbs[c] > 0.00)
cout << endl << "WARNING: [LP Issue] Unique Cell # " << c << " has no observations but is assigned a non‐zero probability. (Type=" << type << ")." << endl
<< "
The danger is that the likelihood function using this cell density could be zero at all time points. If it occurs then the stage 2" << endl
<< "
optimization will eliminate this decision variable (i.e., the unique cell probability) in the objective function (of the 1st Step)" << endl
<< "
causing the corresponding Hessian to be singular because the upper left block is singular (it is a border matrix). A solution is" << endl
<< "
to change the forward‐backward alphas so that a simpler solution is used as there may be too many unique cells. Also, the variance" << endl
<< "
ratio constraint may be too large resulting in spurious solutions being combined across assets resulting in cells with no obs." << endl << endl;
// Free the memory allocated for the LP and the labels array.
//=============================================================
delete_lp(lp);
delete [] vnum; vnum=nullptr;
delete [] vbl; vbl=nullptr;
delete [] epnt; epnt=nullptr;
delete [] fpnt; fpnt=nullptr;
for (int c=0; c<dVars; ++c)
{
delete [] vlabels[c]; vlabels[c]=nullptr;
}
delete [] vlabels; vlabels=nullptr;
// Count the # of non‐zero cell probability decision variables. This is the # of components in the multivariate density and is returned by this function.
//==========================================================================================================================================================
for (int c=0; c<totCells; ++c)
if (outPrbs[c] > 0.00)
non0cmps++;
return non0cmps;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getCMtrx.cpp
/
/ Function: getCMtrx()
/
/ Summary:
/
/
This function ensures that the constraint matrix used to maintain the marginal mixture densities is of full row rank using only the non‐zero component probabilities from the given LP solution. A
/
full row rank component matrix is required for a future optimization. The marginal probability densities are maintained via a set of linear constraints on the multidimensional grid cell
/
probabilities. Since the sum of all probabilities must equal 1.00, each asset will require: (# univariate mixture components for that asset) ‐ 1 constraints to maintain the marginals. Note that
/
the constraint on the last univariate component for each asset will be automatically satisfied by the final constraint that the sum of all cell probabilities equals 1.00. If there are "a" assets,
/
and C[i] components for asset i in the univariate mixture density, i=1,2,...,a, then there will be: (C[0]‐1) + (C[1]‐1) + (C[2]‐1) + ... + (C[a‐1]‐1) + 1 = C[0] + C[1] + C[2] + ... + C[a‐1] ‐ a + 1
/
total rows in the original constraint matrix used to solve the LPs. The constraints can be written in matrix form as Ax = b, where the vector b contains the corresponding probabilities for each row
/
of A, and x is a vector that holds the unique cell probabilities (i.e., decision variables) for the multidimensional grid. Before the LP has been solved rows(A) <= elements(x), since elements(x) =
/
C[0]xC[1]xC[2]x...xC[a‐1]. Therefore, the matrix A will be of full row rank. After each LP has been fit many elements of x will be zero. The effective constraint matrix that applies to the LP
/
solution will be the constraint matrix A with all columns that correspond to zeros of the vector x removed. This effective constraint matrix is not necessarily of full row rank after the LP has
/
been solved. Consider for example the case of 2 assets with 3 components in their corresponding univariate mixture densities. In this scenario, the multidimensional grid is 3x3. Suppose the
/
optimal LP solution contains non‐zero probabilities on the diagonal of this grid and zeros in all off‐diagonal positions. The matrix A will have (3‐1)+(3‐1)+1 = 5 rows initially, however, after the
/
LP has been fit only 3 decision variables (i.e., columns of A) remain. Therefore, the effective constraint matrix A will be of dimension 5x3, and not of full row rank. That is, not all 5 rows of A
/
are needed to enforce the marginal constraints given the current LP solution and 2 rows may be dropped. This function determines the rows that can be dropped and removes them from A producing a
/
full row rank effective constraint matrix, which is needed for an upcoming optimization. To promote parsimony in the fitted multivariate mixture density, components with zero probabilities are
/
always permanantly eliminated at any point in any optimization. Once dropped, a component is not permitted to return to the multivariate density.
/
/ Inputs:
/
/
1.) The number of rows in the original constraint matrix A before solving the LP (either Minimax or Minimum Sum of Squared Distances (SSD)). (totrows)
/
/
2.) The number of non‐zero probabilities (i.e., decision variables) after solving the LP (either Minimax or Minimum Sum of Squared Distances (SSD)). (nUCmps)
/
102
/
3.) Type of LP objective, 0=Minimax or 1=Minimum Sum of Squared Distances (SSD). (type)
/
/
4.) The total number of assets under consideration. (numA)
/
/
5.) Array to hold the optimal/chosen # of univariate mixture components for each asset. (nCmps)
/
/
6.) Array of unique cell IDs for the multidimensional grid w/non‐0 probs in the LP solution, i.e., cells that are used to structure the initial multivariate mixture PDF (w/out covariances). (vCIDs)
/
/
7.) A double array indexed as [c][a], with "c" being a unique cell ID (values are from 0 ‐ totCells‐1), and "a" being an asset ID (values are from 0 ‐ numA‐1). The value held at this position is
/
the univariate mixture component level for asset "a" within unique cell "c". Recall that the multidimensional grid is formed by crossing all univariate assets/components with all other
/
univariate assets/components. (Here component refers back the univariate mixture density for the asset.) (inCellAry)
/
/
8.) A double array indexed as [a][r], with "a" being the asset indicator (values 0 thru numA‐1), and "r" being the component of the optimal univariate mixture distribution for that asset (values 0
/
thru nCmps[a]‐1). The corresponding univariate mixture component probability is held by the array. Note that these probabilities are used to construct all but the last element of the vector b,
/
where the marginals are maintained via: Ax = b. (prbs)
/
/
9.) An empty matrix to hold the full row rank version of A for the given LP solution. This function derives and returns the corresponding matrix. (inLHS)
/
/
10.) An empty vector to hold the corresponding probabilities for the new full row rank version of A. This is the b vector and it is derived here as the original b vector without the corresponding
/
rows that were dropped from A to make it full row rank. (inRHS)
/
/ Outputs:
/
/
This function returns no value at the call, but it derives and populates the empty LHS matrix and RHS vector for the full row rank version of the constraint set that maintains the marginals. The
/
constraints are linear of the form Ax = b.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void getCMtrx(const int totrows, const int nUCmps, const int type, const int numA, const int *nCmps, const int *vCIDs, const int **inCellAry, const long double **prbs, Eigen::MatrixXd *inLHS,
Eigen::VectorXd *inRHS)
{
// Declare/initialize local variables.
//======================================
int rnk, trnk, *remrows=nullptr, rr, kr, rw;
Eigen::MatrixXd oLHS(totrows,nUCmps);
Eigen::VectorXd oRHS(totrows), tVec(nUCmps);
// Build modified constraint matrix that applies to the LP solution.
//====================================================================
rw=0;
for (int a=0; a<numA; ++a)
for (int r=0; r<(nCmps[a]‐1); ++r)
{
for (int c=0; c<nUCmps; ++c)
oLHS(rw,c) = (int) (inCellAry[vCIDs[c]][a] == r);
oRHS(rw++) = prbs[a][r];
}
for (int c=0; c<nUCmps; ++c)
oLHS(rw,c) = 1;
oRHS(rw) = 1.00;
// Rank of constraint matrix with non‐basic and zero‐basic columns removed.
//===========================================================================
rnk = (int) Eigen::FullPivLU<Eigen::MatrixXd>(oLHS).rank();
// Full row rank modified constraint matrix must have # rows equal to its rank.
//===============================================================================
inLHS[type] = Eigen::MatrixXd(rnk,nUCmps);
inRHS[type] = Eigen::VectorXd(rnk);
// If the constraint matrix is not of full row rank, identify (totrows ‐ rnk) rows that can be removed from the constraint matrix to make it full row rank.
//===========================================================================================================================================================
if (rnk < totrows)
{
rr=0;
remrows = new int [totrows‐rnk];
103
for (int r=0; r<totrows; ++r)
{
// Store the values at this row and then set the row to all zeros.
//==================================================================
for (int c=0; c<nUCmps; ++c)
{
tVec[c]=oLHS(r,c);
oLHS(r,c)=0.00;
}
// Recheck the rank. If it changes replace the row with its original values. Otherwise, leave it as all zeros and store the row # that can be dropped.
//========================================================================================================================================================
trnk = (int) Eigen::FullPivLU<Eigen::MatrixXd>(oLHS).rank();
if (trnk < rnk)
for (int c=0; c<nUCmps; ++c)
oLHS(r,c)=tVec(c);
else
remrows[rr++]=r;
}
}
// Populate the full row rank modified constraint matrix.
//=========================================================
rw=0;
for (int r=0; r<totrows; ++r)
{
kr=1;
for (int k=0; k<(totrows‐rnk); ++k)
if (r == remrows[k])
kr=0;
if (1 == kr)
{
for (int c=0; c<nUCmps; ++c)
{
inLHS[type](rw,c)=oLHS(r,c);
inRHS[type](rw)=oRHS(r);
}
rw++;
}
}
// Print out the modified constraint matrix when debugging is on.
//=================================================================
if (dbug >= 2)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(0);
cout << endl << "Modified Full Row‐Rank LHS Constraint Matrix: " << endl << inLHS[type] << endl;
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(10);
cout << endl << "Modified RHS Constraint Vector: " << endl << inRHS[type] << endl;
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(0);
int chkrnk; chkrnk = (int) Eigen::FullPivLU<Eigen::MatrixXd>(inLHS[type]).rank();
cout << endl << "The rank of this modified constraint matrix is: " << chkrnk << endl;
}
// Free up temporary memory allocations.
//========================================
delete [] remrows; remrows=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
104
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
/
PARTICULAR PURPOSE.
See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
Filename:
ECMEAlg.cpp
Function:
ECMEAlg()
Summary:
An extension of the ECME algorithm (Liu & Rubin, 1994) is implemented by this function. The multivariate mixture likelihood is optimized with respect to the mixing proportions and the covariances.
The means and variances are held fixed to maintain the mixture marginals. The 1st Step optimization is convex in the mixing proportions and constrained to maintain the mixture marginals. The
corresponding Lagrangian is formed (1st Step optimization) and its zeros are determined iteratively using Newton's method. The resulting mixture proportions are unique global optimizers of the 1st
Step likelihood (all other parameters are fixed at this stage). Once 1st Step convergence is achieved, processing is passed to the 2nd Step where the likelihoood is maximized with respect to only the
covariances. This optimization is constrained by the positive‐definiteness of the corresponding estimated VC matrices (there is 1 VC matrix per multivariate density component). It is also non‐
convex as there may be multiple local optimums of the likelihood function with only the covariances unknown, and the positive‐definite constraint is not convex. The goal of the 2nd Step is to find
the largest local optimum given that the means, variances, and mixing proportions are fixed and only the covariances are unknown. We will attempt to climb to the top of the current hill (local
optimum) using both the gradient and Hessian while also searching for larger hills in the general direction of steepest ascent (see Marquardt, 1963). This is considered a compromise between
strictly applying Newton's Method and Gradient Ascent and is useful when a single‐step Newton's Method overshoots or a single step lands in an infeasible region. Once no further progress is made
during the 2nd Step, we return to the 1st Step with the newly estimated covariances and repeat the optimization over the mixing proportions. Convergence is achieved when the 2nd Step fails to improve
the likelihood function returned from the 1st Step. Both the 1st and 2nd Steps are iterative with the corresponding gradients and Hessians updated repeatedly during a single corresponding iteration
of the extended ECME algorithm. Solutions found here will not be considered spurious, which differs from the search for a solution when dealing with univariate mixtures. This is due to the fact
that the variances have already been fixed and do not change. The corresponding variance ratio constraint specified by the user remains in force. We can justify this approach by noting that
commercial software packages such as SAS (R) use a line‐search based quasi‐Newton algorithm to find the MLE for a univariate mixture density (instead of the EM algorithm). We have not proven that
the extended ECME method used here will guarantee convergence to the largest local optimum, only that we have located the nearest local optimum in the vicinity of the informed start.
Important Note: The 1st Step imposes equality constraints that maintain the mixture marginals. Greater‐than‐zero constraints on the mixing proportions (i.e., probabilities) are not explicitly
imposed, therefore negative probabilities may maximize the 1st Step likelihood function. The 1st Step likelihood function treats the mixing probabilities as unknowns and all other parameters (means,
variances, covariances) as known constants. Any components that require a negative probability to maximize the likelihood function are dropped during the 1st Step and the entire problem is resized
accordingly (fewer components). This may result in a likelihood that decreases, however the overall objective is to balance parsimony with maximizing the likelihood function.
Inputs:
1.) The total number of time points with data collected. (T)
2.) The double array of returns for each asset and at each time point, indexed as r[a][t].
3.) The number of assets with returns collected. (numA)
4.) The number of unique components in the multivariate mixture that results from either the Minimax or Minimum SSD LP optimizations. Each multivariate mixture has fixed means and variances and
this function will optimize the mixing probabilities and covariance terms. All covariance terms begin the optimization at zero. (nUCmps)
5.) The LHS matrix required to enforce the constraint that the marginal density for each asset equals its fixed univariate mixture. This matrix is built during the LP optimization and resized there
accordingly to ensure it is of full row rank. The LHS matrix has a column for each component in the multivariate density. (cMtrx)
6.) The RHS vector required to enforce the constraint that the marginal density for each asset equals its fixed univariate mixture. This vector is built during the LP optimization and contains the
marginal mixture component probabilities for each asset (less the last probability for each asset, which is fixed once all others have been fixed for that asset). (cVctr)
7.) The array of multivariate mixture probabilities returned from the corresponding LP optimization (either Minimax of Minimum SSD).
converted to a vector within this program as other functions require the values to be stored in a vector. (muPrbs)
Note that these probabilities are passed as an array but
8.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector
of means for the first multivariate component, etc... All means supplied to this function are fixed and do not change which is required to maintain the marginals, with the exception that
components may be dropped. When a component is dropped the corresponding mean vector for that component is dropped. (muMns)
9.) The array of VC matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension (numA)x(numA).
The diagonals of each VC matrix are the corresponding variances for that asset within that component. All variances supplied to this function and are fixed and do not change which is required
to maintain the marginals, with the exception that components may be dropped. When a component is dropped the corresponding VC matrix for that component is dropped. (muVCs)
10.) The array of unique cell IDs that link each component of the multivariate density back to the full factorial of components. The full factorial of components represents each cell in the
multidimensional grid formed by considering all combinations of assets and their levels. Note that the full factorial would be required to build a multivariate mixture density with given
marginals under the assumption that the assets were all mutually independent random variables. (uCellIDs)
11.) A string to hold the directory where the output file resides.
(rdir)
Outputs:
105
/
This function updates the arrays of multivariate mixture probabilities (muPrbs), mean vectors (muMns), and VC matrices (muVCs). Note that muMns is updated only when components are dropped, and
/
muVCs is updated when components are dropped and when covariances are estimated. This function returns the total number of unique multivariate mixture components in the final density.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int ECMEAlg(const int T, const long double **r, const int numA, const int nUCmps, const Eigen::MatrixXd cMtrx, const Eigen::VectorXd cVctr, long double *muPrbs, Eigen::VectorXd *muMns, Eigen::MatrixXd
*muVCs, int *uCellIDs, const string rdir)
{
// Local variables.
//===================
long long mHessMag;
int uCmps=nUCmps, ecnvg, mcnvg, cnvg, n0vals, nLMs=(int) cVctr.size(), ld=uCmps+nLMs, itr, nCovs, nCores=(int) boost::thread::hardware_concurrency(), nUpdts, nThrds=(int) nCorMult*nCores,
**sHess1=new int *[nThrds], ecmeItr=1, eItrs, mItrs, sumval, n_beats=min(nBeats,nThrds), *beat=new int [n_beats], mtch, npos;
long double **fVals=new long double *[T], *dnom=new long double [T], picst=exp(((double) (numA/2.00))*log(2.00*pi)), *sqdets=new long double [uCmps], LL, curmlt, oldLL, **sHess2=new long double
*[nThrds], *maxHLL=new long double [n_beats], eLL, mLL, lBound, sumLL, uval, max1, max2, cnum;
Eigen::VectorXd *gradE=new Eigen::VectorXd [1], *dvarsE=new Eigen::VectorXd [2], *rts=new Eigen::VectorXd [T], *RHS=new Eigen::VectorXd [1], *tmpCVctr=new Eigen::VectorXd [1], *tmpDv=new
Eigen::VectorXd [1], *gradM=new Eigen::VectorXd [1], *dvarsM=new Eigen::VectorXd [4], **tmpDvarsM=new Eigen::VectorXd *[nThrds];
Eigen::MatrixXd *hessE=new Eigen::MatrixXd [1], *VCMInv=new Eigen::MatrixXd [nUCmps], *LHS=new Eigen::MatrixXd [1], *tmpCMtrx=new Eigen::MatrixXd [1], *hessM=new Eigen::MatrixXd [1], *A=new
Eigen::MatrixXd [(int) numA*(numA‐1)/2], *p1Opt=new Eigen::MatrixXd [1], *P=new Eigen::MatrixXd [1], *Pinv=new Eigen::MatrixXd [1];
string lblvar, ndef;
boost::thread *t1=new boost::thread[nThrds];
random_device rd; default_random_engine gen(rd());
uniform_real_distribution<long double> udist(0.0,1.0);
Eigen::EigenSolver<Eigen::MatrixXd> egnslvr, normslvr, normslvrt;
ofstream fout;
// Populate an array of vectors with the returns at each time point indexed as [t][a] to ease computations.
//===========================================================================================================
for (int t=0; t<T; ++t)
{
fVals[t] = new long double [nUCmps];
rts[t]=Eigen::VectorXd(numA);
for (int a=0; a<numA; ++a)
rts[t](a)=r[a][t];
}
// Derive VC inverses and corresponding determinants.
//=====================================================
for (int v=0; v<uCmps; ++v)
{
VCMInv[v]=Eigen::MatrixXd(numA,numA);
VCMInv[v]=muVCs[v].inverse();
sqdets[v]=sqrt(VCMInv[v].determinant());
}
// Populate the covariance identifier matrices for use in the 2nd Step.
//======================================================================
itr=0;
for (int a1=0; a1<numA; ++a1)
for (int a2=a1+1; a2<numA; ++a2)
{
A[itr]=Eigen::MatrixXd(numA,numA);
for (int r=0; r<numA; ++r)
for (int c=0; c<numA; ++c)
{
if (r==a1 && c==a2 || r==a2 && c==a1)
A[itr](r,c)=1.0;
else
A[itr](r,c)=0.0;
}
itr++;
}
// Initialize decision variables for 1st step. Probabilities will be set to their values as determined by solving the corresponding LP and the Lagrange multipliers will be initialized to zeros.
//================================================================================================================================================================================================
for (int i=0; i<2; ++i)
106
dvarsE[i]=Eigen::VectorXd(ld);
for (int d=0; d<uCmps; ++d)
dvarsE[0](d)=muPrbs[d];
for (int d=uCmps; d<ld; ++d)
dvarsE[0](d)=0.00;
// Populate a double array with the likelihood values for each timepoint and component. The multivariate likelihood of each observation is also stored in an array (dnom[]).
// Any component with zero likelihood for all time points is a variable that does not exist in the objective function. It should be treated as a constant and moved to the right
// hand side of each constraint and the problem needs to be resized accordingly. This check is made via the function call chkSum() below.
//=================================================================================================================================================================================
LL=getLFVals(T,numA,uCmps,rts,dvarsE[0],muMns,VCMInv,sqdets,picst,dnom,fVals);
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15); cout << "Initial Log‐Likelihood value is: " << LL << endl << endl;
chkSum(T,uCmps,(const long double **) fVals);
// Build the corresponding Hessian (of the Lagrangian). The matrix is stored in hessE[0](.,.). When building the Hessian, iterate until it is invertible by multiplying the constraint LHS and RHS by
// a constant value (multiple of 10) until full rank. (Note: A matrix with large and small eigenvalues may be ill‐conditioned and there may be computational issues when attempting to invert it.)
//=====================================================================================================================================================================================================
hessE[0]=Eigen::MatrixXd(ld,ld);
curmlt=getHessE(T,uCmps,(const long double **) fVals,dnom,cMtrx,cVctr,hessE,LHS,RHS);
// Build the gradient (of the Lagrangian) (using the modifed constraint matrix/vector as required above).
//=========================================================================================================
gradE[0]=Eigen::VectorXd(ld);
getGradE(T,uCmps,(const long double **) fVals,dnom,LHS[0],RHS[0],dvarsE[0],gradE);
// Initialize arrays used and reused during the 2nd Step.
//========================================================
for (int h=0; h<nThrds; ++h)
{
sHess1[h]=new int [8];
sHess2[h]=new long double [7];
tmpDvarsM[h]=new Eigen::VectorXd [4];
}
// Iterate using the ECME algorithm until convergence.
// Iterate and update the probabilities (the 1st Step is a maximization problem with concave objective and convex constraints). Stationary points for the Lagrangian will therefore be taken as global
// optimizers and these are determined using Newton's method. (The Hessian here is a bordered matrix which is invertible under certain met conditions on the non‐zero sections.) The 2nd Step is a
// constrained maximization problem having multiple local optimums. We attempt to find the largest local optimum using an iterative technique that steps in the general direction of steepest ascent.
//=====================================================================================================================================================================================================
cnvg=0;
do
{
//==================
// (ECME) 1st STEP:
//==================
oldLL=LL; ecnvg=0; n0vals=0; eItrs=0;
cout << endl << "1st Step Start: ECME Algorithm (Iteration=" << ecmeItr << ") (Beginning LL = ";
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15); cout << oldLL << ")." << endl << endl << string(15,' ') << "‐‐‐‐‐> Iterating: ";
do
{
// 1st Step iteration counter.
//=============================
eItrs++;
cout << ".";
// Solve for new component probabilities, which are the decision variables in the 1st Step.
//==========================================================================================
dvarsE[1]=hessE[0].colPivHouseholderQr().solve(‐1.00*gradE[0]);
dvarsE[0]=(dvarsE[1]+dvarsE[0]);
// Check for zero or negative probabilities and prepare for next iteration.
//===========================================================================
n0vals=0;
for (int v=0; v<uCmps; ++v)
if (dvarsE[0](v) <= 0.00)
107
n0vals++;
// Resize the problem (if needed) and perform another iteration.
//================================================================
uCmps=uCmps‐n0vals;
nLMs=(int) RHS[0].size();
ld=uCmps+nLMs;
// Update constraint matrix/vector. Undo the multiplier and adjust size if multivariate component probabilities have been set to zero.
//=======================================================================================================================================
tmpCMtrx[0]=Eigen::MatrixXd(nLMs,uCmps);
for (int r=0; r<(int) LHS[0].rows(); ++r)
{
itr=0;
for (int c=0; c<(int) LHS[0].cols(); ++c)
if (dvarsE[0](c) > 0.00)
tmpCMtrx[0](r,itr++)=(1.00/curmlt)*LHS[0](r,c);
}
tmpCVctr[0]=Eigen::VectorXd(nLMs);
tmpCVctr[0]=(1.00/curmlt)*RHS[0];
// If a component is dropped then update mean vectors, VC matrices, and unique cell IDs.
//========================================================================================
if (n0vals > 0)
{
itr=0;
for (int v=0; v<(int) LHS[0].cols(); ++v)
if (dvarsE[0](v) > 0.00)
{
muMns[itr]=muMns[v];
muVCs[itr]=muVCs[v];
VCMInv[itr]=muVCs[itr].inverse();
sqdets[itr]=sqrt(VCMInv[itr].determinant());
uCellIDs[itr++]=uCellIDs[v];
}
// Resize the internal array that holds the likelihood values for each timepoint and component when the # of components changes.
//================================================================================================================================
for (int t=0; t<T; ++t)
{
delete [] fVals[t]; fVals[t]=new long double [uCmps];
}
// Update decision variable vectors, dropping relevant rows.
//============================================================
tmpDv[0]=Eigen::VectorXd((int) dvarsE[0].size());
tmpDv[0]=dvarsE[0];
delete [] dvarsE; dvarsE=new Eigen::VectorXd [2];
for (int i=0; i<2; ++i)
dvarsE[i]=Eigen::VectorXd(ld);
itr=0;
for (int v=0; v<(int) LHS[0].cols(); ++v)
if (tmpDv[0](v) > 0.00)
dvarsE[0](itr++)=tmpDv[0](v);
for (int v=(int) LHS[0].cols(); v<(int) tmpDv[0].size(); ++v)
dvarsE[0](itr++)=tmpDv[0](v);
delete [] tmpDv; tmpDv=new Eigen::VectorXd [1];
}
// Update density function values grid timepoint‐specific likelihood function values.
//=====================================================================================
LL=getLFVals(T,numA,uCmps,rts,dvarsE[0],muMns,VCMInv,sqdets,picst,dnom,fVals);
// Check for 1st Step convergence. Need no negative/zero component probabilities and unchanged LL. Otherwise, rebuild Hessian/gradient and iterate again.
//==========================================================================================================================================================
if (0 == n0vals && (LL ‐ oldLL) < (epsilon)*oldLL)
108
{
ecnvg=1;
itr=0;
oldLL=LL;
for (int v=0; v<uCmps; ++v)
muPrbs[itr++]=dvarsE[0](v);
}
// No 1st Step convergence, iterate again.
//=========================================
if (0 == ecnvg)
{
// Reset Log‐Likelihood, which is only valid when there are no negative probabilities.
//======================================================================================
if (0 == n0vals)
oldLL=LL;
// Rebuild Hessian.
//===================
delete [] LHS; LHS=new Eigen::MatrixXd [1];
delete [] RHS; RHS=new Eigen::VectorXd [1];
delete [] hessE; hessE=new Eigen::MatrixXd [1]; hessE[0]=Eigen::MatrixXd(ld,ld);
curmlt=getHessE(T,uCmps,(const long double **) fVals,dnom,tmpCMtrx[0],tmpCVctr[0],hessE,LHS,RHS);
// Rebuild Gradient then delete temporary memory allocations.
//=============================================================
delete [] gradE; gradE=new Eigen::VectorXd [1]; gradE[0]=Eigen::VectorXd(ld);
getGradE(T,uCmps,(const long double **) fVals,dnom,LHS[0],RHS[0],dvarsE[0],gradE);
delete [] tmpCMtrx; tmpCMtrx=new Eigen::MatrixXd [1];
delete [] tmpCVctr; tmpCVctr=new Eigen::VectorXd [1];
}
} while (0 == ecnvg);
eLL=oldLL;
cout << " (Done.)" << endl << endl << "1st Step Done:
ECME Algorithm (Iteration=" << ecmeItr << ") Converged in " << eItrs << " iterations.
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15); cout << eLL << ")." << endl << endl;
//==================
// (ECME) 2nd STEP:
//==================
nCovs=(int) uCmps*numA*(numA‐1)/2;
p1Opt[0]=Eigen::VectorXd(nCovs);
for (int i=0; i<4; ++i)
dvarsM[i]=Eigen::VectorXd(nCovs);
getCovs(uCmps,muVCs,dvarsM);
dvarsM[2]=dvarsM[0];
gradM[0]=Eigen::VectorXd(nCovs);
hessM[0]=Eigen::MatrixXd(nCovs,nCovs);
P[0]=Eigen::MatrixXd(nCovs,nCovs);
Pinv[0]=Eigen::MatrixXd(nCovs,nCovs);
for (int h=0; h<nThrds; ++h)
for (int i=0; i<4; ++i)
tmpDvarsM[h][i]=Eigen::VectorXd(nCovs);
mcnvg=0; mItrs=0;
cout << endl << "2nd Step Start: ECME Algorithm (Iteration=" << ecmeItr << ") (Beginning LL = ";
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15); cout << oldLL << ")." << endl << endl;
do
{
// 2nd Step iteration counter.
//=============================
mItrs++;
// Build gradient for 2nd step.
//==============================
getGradM(T,rts,uCmps,numA,(const long double **) fVals,dnom,muPrbs,muMns,muVCs,VCMInv,A,gradM);
109
(New LL = ";
// Build Hessian for 2nd step. Find the length (# digits) of the element with largest magnitude.
//================================================================================================
mHessMag=(long long) getHessM(T,rts,uCmps,numA,(const long double **) fVals,dnom,muPrbs,muMns,muVCs,VCMInv,A,hessM);
// Write out the max eigenvalue, condition # and # of negative/positive eigenvalues of the Hessian just derived.
//================================================================================================================
egnslvr.compute(hessM[0], true);
P[0]=egnslvr.eigenvectors().real();
Pinv[0]=P[0].inverse();
normslvr.compute(P[0].transpose()*P[0],false);
normslvrt.compute(Pinv[0].transpose()*Pinv[0],false);
max1=LNegVal; max2=LNegVal; npos=0;
for (int a=0; a<(int) hessM[0].rows(); ++a)
{
if (egnslvr.eigenvalues()[a].real() > 0)
npos++;
if (abs(normslvr.eigenvalues()[a].real()) > max1)
max1=abs(normslvr.eigenvalues()[a].real());
if (abs(normslvrt.eigenvalues()[a].real()) > max2)
max2=abs(normslvrt.eigenvalues()[a].real());
}
cnum = sqrt(max1)*sqrt(max2);
cout << endl << string(15,' ') << "Total # of (‐,+) eigenvalues: ("<< nCovs‐npos << "," << npos << ") (Hessian condition # = " << cnum << ")" << endl;
// The function stepHessM() uses the Hessian to step in the direction of the gradient. To step we add a (random) constant to the diagonal
// with larger (random) constants translating to smaller steps, and smaller (random) constants translating to larger steps.
//==========================================================================================================================================
Eigen::initParallel();
lblvar=string(15,' ') + "‐‐‐‐‐> Run #" + to_string((long long) mItrs) + " ‐ Threads Launched: ";
cout << lblvar;
for (int h=0; h<nThrds; ++h)
{
sHess1[h][0]=T; sHess1[h][1]=uCmps; sHess1[h][2]=h; sHess1[h][3]=0; sHess1[h][4]=(int) to_string(mHessMag).size() + 2; sHess1[h][6]=0; sHess1[h][7]=0;
sHess2[h][0]=oldLL; sHess2[h][1]=0.00; sHess2[h][2]=(long double) (minHessAdd + h*mItersH*sHess2[h][1]); sHess2[h][5]=0.00;
tmpDvarsM[h][0]=dvarsM[0];
tmpDvarsM[h][2]=dvarsM[2];
t1[h]=boost::thread(stepHessM,boost::ref(sHess1[h]),boost::ref(sHess2[h]),boost::cref(rts),boost::ref(tmpDvarsM[h]),gradM[0],hessM[0],dvarsE[0],boost::cref(muMns),boost::cref(muVCs));
}
// Conditionally output a line feed and alignment spaces once all threads have successfully launched.
//=====================================================================================================
if (dbug <= 1)
{
do
{
Sleep(1000); sumval=0;
for (int h=0; h<nThrds; ++h)
sumval=sumval+sHess1[h][3];
} while (sumval < nThrds); cout << string(10,' ') << "(Pr LL = " << oldLL << ")" << endl << string(lblvar.size()‐18,' ') << "Threads Finished: ";
sumval=0;
do
{
Sleep(100);
for (int h=0; h<nThrds; ++h)
if (1 == sHess1[h][7])
{
cout << ".";
sHess1[h][7]=0;
sumval++;
}
} while (sumval < nThrds);
}
// Pause until all threads finish.
//==================================
110
for (int j=0; j<nThrds; ++j)
t1[j].join();
// Randomly select one of the top "nBeats" performers to begin the next iteration. (Weight values by their LL to favor higher values.)
// The value of "nBeats" is set in the header file and only LL values that exceed the current LL are considered beats.
//=======================================================================================================================================
nUpdts=0; sumLL=0.00;
for (int b=0; b<n_beats; ++b)
maxHLL[b]=oldLL;
for (int h1=0; h1<nThrds; ++h1)
{
// No LL values returned should be less than the existing maximum.
//==================================================================
if (sHess2[h1][3] < oldLL)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15);
cout << "ERROR: ECME Algorithm 2nd Step stepping function has returned a LL value inferior to the current maximum, which should not happen." << endl
<< "
Must inspect and fix. Thread # = " << h1 << endl
<< "
The current maximum LL = " << oldLL << endl
<< "
Maximum LL value returned from stepping function = " << sHess2[h1][3] << endl
<< "EXITING...ECMEAlg()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Find and process the improvements.
//=====================================
itr=0;
for (int h2=0; h2<nThrds; ++h2)
if (sHess2[h1][3] < sHess2[h2][3])
itr++;
for (int b=0; b<n_beats; ++b)
// Deal with LL ties.
if (sHess2[h1][3] > oldLL && sHess2[h1][3] == maxHLL[b])
itr++;
if (sHess2[h1][3] > oldLL && itr < n_beats)
{
maxHLL[itr]=sHess2[h1][3];
// Store the log‐likelihood for the given beat.
beat[itr] = h1;
// Store the thread index for the given beat.
sumLL = sumLL + (maxHLL[itr]‐oldLL);
// Sum the magnitude of log‐likelihood improvements.
nUpdts++;
// Record the # of improvements.
}
}
// Randomly select one of the LL beats to begin the next 2nd Step iteration.
//===========================================================================
if (nUpdts > 0)
{
uval = udist(gen)*sumLL; mtch=0;
do
{
lBound = sumLL ‐ (maxHLL[nUpdts‐1]‐oldLL);
if (uval >= lBound)
mtch = 1;
else
sumLL = sumLL ‐ (maxHLL[(nUpdts‐‐)‐1]‐oldLL);
} while (0 == mtch);
dvarsM[0]=tmpDvarsM[beat[nUpdts‐1]][3];
}
else
nUpdts=1;
if (0 == dbug)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15);
cout << " (Done.) (RC LL = " << maxHLL[nUpdts‐1] << ", #(RRs)=" << sHess1[beat[nUpdts‐1]][6] << ")" << endl << endl; cout.precision(15); /* RC = Randomly Chosen */
}
111
// Update the VC and inverse VC matrices along with the vector of corresponding determinants.
//=============================================================================================
for (int v=0; v<uCmps; ++v)
{
setCovs(v,muVCs,dvarsM);
VCMInv[v]=muVCs[v].inverse();
sqdets[v]=sqrt(VCMInv[v].determinant());
}
LL=getLFVals(T,numA,uCmps,rts,dvarsE[0],muMns,VCMInv,sqdets,picst,dnom,fVals);
// QC check that Max LL equals the beat value chosen above.
//===========================================================
if (abs(LL ‐ maxHLL[nUpdts‐1]) > epsilon*maxHLL[nUpdts‐1])
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15);
cout << "ERROR: ECME Algorithm has derived LL not equal to the beat LL chosen randomly, which should not happen. Must inspect and fix." << endl
<< "
Value of maxHLL[nUpdts‐1] = " << maxHLL[nUpdts‐1] << endl
<< "
Value of LL = " << LL << endl << "EXITING...ECMEAlg()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
// Check for convergence of the 2nd step.
//========================================
if ((LL ‐ oldLL) < pow(10.0,6)*(epsilon)*oldLL)
mcnvg=1;
oldLL=LL; dvarsM[2]=dvarsM[0];
} while (0 == mcnvg);
mLL=oldLL;
cout << endl << "2nd Step Done:
ECME Algorithm (Iteration=" << ecmeItr << ") Converged in " << mItrs << " iterations. (New LL=";
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(15); cout << mLL << ")." << endl << endl;
// Check ECME convergence. (2nd Step did not improve 1st Step likelihood.) If non‐convergence, prepare another 1st Step iteration. If convergence, find/display eigenvalues to check concavity.
//================================================================================================================================================================================================
if ((mLL ‐ eLL) < pow(10.0,6)*(epsilon)*eLL)
cnvg=1;
// Free temporary memory allocations.
//=====================================
delete [] gradM; gradM=new Eigen::VectorXd [1];
delete [] hessM; hessM=new Eigen::MatrixXd [1];
delete [] P; P=new Eigen::MatrixXd [1];
delete [] Pinv; Pinv=new Eigen::MatrixXd [1];
delete [] dvarsM; dvarsM=new Eigen::VectorXd [4];
delete [] p1Opt; p1Opt=new Eigen::MatrixXd [1];
for (int h=0; h<nThrds; ++h)
{
delete [] tmpDvarsM[h]; tmpDvarsM[h]=new Eigen::VectorXd [4];
}
if (0 == cnvg)
{
// Populate the dnom and fVals arrays.
//======================================
LL=getLFVals(T,numA,uCmps,rts,dvarsE[0],muMns,VCMInv,sqdets,picst,dnom,fVals);
chkSum(T,uCmps,(const long double **) fVals);
// Build Hessian and Gradient.
//==============================
delete [] LHS; LHS=new Eigen::MatrixXd [1];
delete [] RHS; RHS=new Eigen::VectorXd [1];
delete [] hessE; hessE=new Eigen::MatrixXd [1]; hessE[0]=Eigen::MatrixXd(ld,ld);
curmlt=getHessE(T,uCmps,(const long double **) fVals,dnom,tmpCMtrx[0],tmpCVctr[0],hessE,LHS,RHS);
delete [] gradE; gradE=new Eigen::VectorXd [1]; gradE[0]=Eigen::VectorXd(ld);
getGradE(T,uCmps,(const long double **) fVals,dnom,LHS[0],RHS[0],dvarsE[0],gradE);
112
// Delete temporary memory allocations.
//=======================================
delete [] tmpCMtrx; tmpCMtrx=new Eigen::MatrixXd [1];
delete [] tmpCVctr; tmpCVctr=new Eigen::VectorXd [1];
ecmeItr++;
}
// Processor cool down.
//=======================
if (dbug >= 2)
{
cout.setf(ios_base::fixed, ios_base::floatfield); cout.precision(4);
cout << endl << "***** Processor cool Down: " << ((double) cDown/1000.00)/60.00 << " minutes. *****" << endl;
} Sleep(cDown);
} while (0 == cnvg);
cout << "ECME Algorithm Converged in " << ecmeItr << " Iterations (Maximum LL=" << LL << ")." << endl;
fout.open(rdir+ofile, ios_base::out | ios_base::app);
fout << "ECME Algorithm Converged in " << ecmeItr << " Iterations (Maximum LL=" << LL << ")." << endl;
fout.close();
// Delete temporary memory allocations.
//=======================================
delete [] dnom; dnom=nullptr;
delete [] sqdets; sqdets=nullptr;
delete [] gradE; gradE=nullptr;
delete [] gradM; gradM=nullptr;
delete [] rts; rts=nullptr;
delete [] VCMInv; VCMInv=nullptr;
delete [] hessE; hessE=nullptr;
delete [] hessM; hessM=nullptr;
delete [] P; P=nullptr;
delete [] Pinv; Pinv=nullptr;
delete [] dvarsE; dvarsE=nullptr;
delete [] dvarsM; dvarsM=nullptr;
delete [] LHS; LHS=nullptr;
delete [] RHS; RHS=nullptr;
delete [] tmpCMtrx; tmpCMtrx=nullptr;
delete [] tmpCVctr; tmpCVctr=nullptr;
delete [] tmpDv; tmpDv=nullptr;
delete [] A; A=nullptr;
delete [] beat; beat=nullptr;
delete [] maxHLL; maxHLL=nullptr;
for (int t=0; t<T; ++t)
{
delete [] fVals[t]; fVals[t]=nullptr;
}
delete [] fVals; fVals=nullptr;
for (int h=0; h<nThrds; ++h)
{
delete [] sHess1[h]; sHess1[h]=nullptr;
delete [] sHess2[h]; sHess2[h]=nullptr;
delete [] tmpDvarsM[h]; tmpDvarsM[h]=nullptr;
}
delete [] sHess1; sHess1=nullptr;
delete [] sHess2; sHess2=nullptr;
delete [] tmpDvarsM; tmpDvarsM=nullptr;
delete [] t1; t1=nullptr;
// Count and return the final number of non‐zero unique cell probabilities for this solution.
//=============================================================================================
return uCmps;
}
113
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getLFVals.cpp
/
/ Function: getLFVals()
/
/ Summary:
/
/
This function decomposes the multivariate mixture likelihood as a 2‐dimensional grid of values with time (T) on the vertical axis and component (U) on the horizontal axis. Each cell in the 2‐
/
dimensional grid is a likelihood value for the data at that timepoint using the corresponding multivariate density function for that component. This function computes each value in the grid and
/
stores the value in the 2‐dimensional array supplied by parameter #11. In addition, if the values in each row are summed using the component probabilities as weights, the value is the multivariate
/
mixture likelihood at the given time point. These values are derived and stored in the single array supplied by parameter #10. Lastly, if the log of the values computed for parameter #10 are taken
/
and summed across all time points then this value is the log‐likelihood for the data using the supplied multivariate mixture density. This log‐likelihood value is computed/returned at the call.
/
/ Inputs:
/
/
1.) The total number of time points with data collected. (T)
/
/
2.) The number of assets with returns. (numA)
/
/
3.) The number of unique components in the multivariate mixture. (Initial value is from either the Minimax or Minimum SSD LP optimizations.) (inUCmps)
/
/
4.) The array of vector returns at each time point indexed as rs[t](a). There are T vectors of returns and each is of size numA, where numA=# assets. (rs)
/
/
5.) The current vector of multivariate mixture probabilities. (Initial values are from either the Minimax or Minimum SSD LP optimizations.) (uPrbs)
/
/
6.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector
/
of means for the first multivariate component, etc... (inMns)
/
/
7.) The array of VC inverse matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is invertible and of
/
dimension (numA)x(numA). The diagonals of each VC matrix are the corresponding variances for that asset within that component. (inVCIs)
/
/
8.) The array of square roots of the determinants of the VC inverse matrices from parmeter #7. This term is required to construct the multivariate normal density. (insqs)
/
/
9.) This parameter equals (2*pi)^(numA/2), where numA = total # of assets in the application. (inpicst)
/
/
10.) A single array of T values that sum the double array in parameter #11 across the components at each time point, weighting each component by its corresponding estimated probability. Each value
/
in this array is the multivariate mixture likelihood value for the data at each individual time point. This parameter is supplied empty and populated/returned by this function. (denoms)
/
/
11.) A double array of likelihood values indexed by time and component. At each time point the likelihood for each component is computed and stored in this double array for reuse. This forms a 2‐
/
dimensional grid of values of size TxU, where T=# of time points, and U=# components. This parameter is supplied empty and populated/returned by this function. (lfVals)
/
/ Outputs:
/
/
This function returns the log‐likelihood value at the call and also populates the 2 incoming arrays denoms (see parameter #10) and lfVals (parameter #11).
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
long double getLFVals(const int T, const int numA, const int inUCmps, const Eigen::VectorXd *rs, const Eigen::VectorXd uPrbs, const Eigen::VectorXd *inMns, const Eigen::MatrixXd *inVCIs,
const long double *insqs, const long double inpicst, long double *denoms, long double **lfVals)
{
// Local variables.
//===================
long double LL=0.00;
// Populate 2 containers: 1.) 2‐dimensional grid of all component likelihoods evaluated at each time point. 2.) 1‐dimensional array of all full likelihood values evaluated at each time point.
//=================================================================================================================================================================================================
for (int t=0; t<T; ++t)
{
denoms[t]=0.00;
for (int v=0; v<inUCmps; ++v)
114
{
lfVals[t][v]=getMVNDens(rs[t],inMns[v],inVCIs[v],insqs[v],inpicst);
denoms[t]=denoms[t] + uPrbs(v)*lfVals[t][v];
}
if (denoms[t] > 0.00)
LL=LL+log(denoms[t]);
else
LL=LL+LNegVal;
}
// Return the log‐likelihood value.
//===================================
return LL;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getGradE.cpp
/
/ Function: getGradE()
/
/ Summary:
/
/
This function derives the Gradient for the ECME algorithm 1st Step optimization, which is convex in the decision variables (multivariate mixture component probabilities). All means, variances and
/
covariances are treated as constants. The objective is to maximize the corresponding log‐likelihood function subject to constraints that the marginal densities are fixed and known univariate
/
mixtures. The marginal constraints can be enforced via linear functions on the decision variables (component probabilities). By incorporating the constraints into the objective we form the
/
Lagrangian. Stationary points of the Lagrangian will be unique global optimizers of the constrainted convex optimization problem. These points are found by applying Newton's method to the
Lagrangian. Newton's method requires that the gradient and Hessian of the Lagrangian be constructed during each iteration. The optimization problem converges when the log‐likelihood fails to
/
/
improve. In this function we compute the Gradient of the Lagrangian for the 1st Step. The Gradient is the vector of 1st partial derivatives of the Lagrangian. The number of elements is the sum of
/
the # of unique components (multivariate mixture probabilities) and the # of constraints (Lagrange multipliers).
/
/ Inputs:
/
/
1.) The total number of time points with data collected. (T)
/
/
2.) The number of unique components in the multivariate mixture. (Initial value is from either the Minimax or Minimum SSD LP optimizations.) (inUCmps)
/
/
3.) A double array of likelihood values indexed by time and component. At each time point the likelihood for each component is computed and stored for reuse. This forms a 2‐dimensional grid of
/
values of size TxU, where T=# of time points, and U=# components. (infVals)
/
/
4.) A single array of T values that sum the double array in parameter #3 across the components at each time point, weighting each likelihood value by its corresponding estimated component
/
probability. Therefore, each value in this array is the multivariate mixture likelihood value for the data at each individual time point. (inDNoms)
/
/
5.) The LHS matrix needed to enforce the marginal mixture density constraints. These constraints are linear in the component probabilities therefore can be represented using a LHS matrix and RHS
/
vector. This matrix has 1's & 0's. It may be necessary to multiply both sides of each constraint by a constant factor to make the corresponding Hessian full rank (computationally). The # of
/
columns is equal to the # of components and the # of rows is equal to the # of constraints (i.e., Lagrange multipliers) needed to maintain the marginal univariate mixtures (inLHS).
/
/
6.) The RHS constraint vector required to enforce the fixed marginal density constraints using actual univariate marginal mixture probabilities. The constraints needed to ensure that given fixed
/
mixture marginals add multivariate component probabilities up to equal the given marginal mixture probabilities for each component of each asset. This vector is scaled when the corresponding
/
LHS matrix above is scaled to ensure the Hessian of the Lagrangian is full rank (computationally). (inRHS)
/
/
7.) Current values of the decision variables. Values for the probabilities are not needed given that we have the double array infVals[t][u] above. However we do need the current values of the
/
Lagrange multipliers as these change during each iteration. Therefore we will pull these from the vector of all decision variables passed via this parameter. (inDVars)
/
/
8.) The empty Gradient vector to be filled by this function. The vector is of dimension equal to the sum of the # of unique components and the # of Lagrange multipliers. The # of Lagrange
/
multipliers equals the # of rows in the LHS constraint matrix, which equals the # of constraints. (inGrad)
/
/ Outputs:
/
/
This function populates the empty gradient vector supplied to it but does not return any other output at the function call.
/*********************************************************************************************************************************************************************************************************/
115
#include "stdafx.h"
void getGradE(const int T, const int inUCmps, const long double **infVals, const long double *inDNoms, const Eigen::MatrixXd inLHS, const Eigen::VectorXd inRHS, const Eigen::VectorXd inDVars,
Eigen::VectorXd *inGrad)
{
// Local variables.
//===================
int ld=inUCmps + (int) inRHS.size();
// Gradient: Partials wrt Probabilities.
//=========================================
for (int v=0; v<inUCmps; ++v)
{
inGrad[0](v)=0.00;
for (int t=0; t<T; ++t)
inGrad[0](v)=inGrad[0](v) + infVals[t][v]/inDNoms[t];
for (int r=0; r<(int) inRHS.size(); ++r)
inGrad[0](v)=inGrad[0](v) ‐ inDVars(inUCmps+r)*inLHS(r,v);
}
// Gradient: Partials wrt Multipliers.
//=======================================
for (int m=inUCmps; m<ld; ++m)
{
inGrad[0](m)=0.00;
for (int v=0; v<inUCmps; ++v)
inGrad[0](m)=inGrad[0](m) + ((double) inLHS(m‐inUCmps,v))*inDVars(v);
inGrad[0](m)=‐1.00*(inGrad[0](m) ‐ inRHS(m‐inUCmps));
}
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getHessE.cpp
/
/ Function: getHessE()
/
/ Summary:
/
/
This function derives the Hessian for the ECME algorithm 1st Step optimization, which is convex in the decision variables (multivariate mixture component probabilities). All means, variances and
/
covariances are treated as constants during this optimization (i.e., ECME 1st Step). The objective is to maximize the corresponding log‐likelihood function subject to constraints that the marginal
/
densities are fixed and known univariate mixtures. The marginal constraints can be enforced via linear functions on the decision variables (component probabilities). By incorporating the
/
constraints into the objective we form the Lagrangian. Stationary points of the Lagrangian will be unique global optimizers of the constrained convex optimization problem. These points are found
/
by applying Newton's method to the Lagrangian. Newton's method requires that the gradient and Hessian of the Lagrangian be constructed during each iteration. The optimization problem converges
/
when the log‐likelihood fails to improve (i.e., at a zero of the Lagrangian). In this function we compute the Hessian of the Lagrangian for the 1st Step. The Hessian is a border matrix since the 2nd
/
derivative WRT the Lagrange multipliers is always zero, therefore there will be a block matrix of zeros in the lower right corner. A border matrix is invertible under certain conditions on the 3
/
block matrices that border the zero block. These conditions will be met for this optimization, however it may be necessary to inflate the constraint matrix by using a constant larger than 1. This
/
would be needed when the Hessian is ill‐conditioned using 0/1 indicator variables to enforce the constraints.
/
/ Inputs:
/
/
1.) The total number of time points with data collected. (T)
/
/
2.) The number of unique components in the multivariate mixture. (Initial values are from either the Minimax or Minimum SSD LP optimizations.) (nUCmps)
/
/
3.) A double array of likelihood values indexed by time and component. At each time point the likelihood for each component is computed and stored for reuse. This forms a 2‐dimensional grid of
/
values of size TxU, where T=# of time points, and U=# components. (infVals)
/
/
4.) A single array of T values that sum the double array in parameter #3 across the components at each time point, weighting each component by its corresponding estimated probability. Therefore,
/
each value in this array is the multivariate mixture likelihood value for the data at each individual time point. (inDNoms)
/
116
/
5.) The LHS matrix (of 1's and 0's) needed to enforce the marginal mixture density constraints. These constraints are linear in the decision variables (component probabilities) therefore can be
/
represented using a LHS matrix and RHS vector. This matrix has 1's and 0's but these may be multiplied by a constant to ensure the Hessian returned is full rank (computationally). The # of
/
columns is equal to the # of components and the # of rows is equal to the # of constraints (Lagrange multipliers) that are required to ensure that the marginals match their fixed mixtures as
/
found earlier. The matrix must be of full rank, therefore if multivariate components are set to zero we should check that it remains full rank, if not force it to be full rank by removing rows
/
one at a time until it is. (This code is yet to be implemented. Problem has not been encountered. The function getCMtrx() can be used to perform the task.) (inCMtrx)
/
/
6.) The RHS constraint vector required to enforce the fixed marginal density constraints using actual marginal probabilities. The constraints needed to ensure given fixed mixture marginals add
/
multivariate component probabilities up to equal the given marginal mixture probabilities for each component of each asset. This vector is scaled when the corresponding LHS matrix above is
/
scaled to ensure the Hessian of the Lagrangian is invertible. (inCVctr)
/
/
7.) The empty Hessian matrix to be filled by this function. The matrix is square with dimension equal to the # of unique components plus the # of Lagrange multipliers. The # of Lagrange
/
multipliers equals the # of rows in the LHS constraint matrix, which equals the # of constraints. (inHess)
/
/
8.) An empty matrix to be filled with the updated LHS constraint matrix once scaled to ensure the resulting Hessian is invertible. As noted, the Hessian is for the Lagrangian which is a border
/
matrix with a block of zeros in the lower right corner. The upper right corner is the constraint matrix of 1's and 0's, while the upper left corner is the Hessian of the original objective
/
function (i.e., without the Lagrange multipliers). In rare cases, large values in the upper left matrix coupled with 1's and 0's in the upper right matrix can cause the matrix to be ill‐
/
conditioned, therefore not invertible. We have found that a solution is to scale the constraint matrix up by a large constant. That is, we multiple the LHS and RHS of each constraint by a
/
given large constant. This fixes the singularity of the Hessian. Note that the gradient also uses the constraints therefore when the constraint matrix is scaled we must use the same scaled
/
version when constructing the gradient. This parameter returns the scaled LHS constraint matrix. Note that the scale factor is returned by the function. (inLHS)
/
/
9.) An empty vector to be filled with the updated RHS constraint values when the constraint matrix is scaled to be invertible. (inRHS)
/
/ Outputs:
/
/
This function returns the the multiplier used to scale the constraint matrix and vector to ensure that the resulting Hessian is computationally invertible. It also populates the empty Hessian
/
matrix supplied to it along with the (scaled) LHS matrix and RHS vector.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
long double getHessE(const int T, const int inUCmps, const long double **infVals, const long double *inDNoms, const Eigen::MatrixXd inCMtrx, const Eigen::VectorXd inCVctr, Eigen::MatrixXd *inHess,
Eigen::MatrixXd *inLHS, Eigen::VectorXd *inRHS)
{
// Local variables.
//===================
int rnk, ld=inUCmps + (int) inCVctr.size();
long double mult=1.00;
Eigen::MatrixXd *ulHess=new Eigen::MatrixXd [1];
ulHess[0]=Eigen::MatrixXd(inUCmps,inUCmps);
// Hessian: Upper left.
//========================
for (int r=0; r<inUCmps; ++r)
for (int c=r; c<inUCmps; ++c)
{
inHess[0](r,c)=0.00;
for (int t=0; t<T; ++t)
inHess[0](r,c) = inHess[0](r,c) + infVals[t][r]*infVals[t][c]/pow(inDNoms[t],2);
inHess[0](r,c)=‐1.00*inHess[0](r,c);
ulHess[0](r,c)=inHess[0](r,c);
if (c > r)
{
inHess[0](c,r)=inHess[0](r,c);
ulHess[0](c,r)=ulHess[0](r,c);
}
}
// Before proceeding with the UR, LL, and LR sections, check that the UL Hessian is full rank and put out a warning if it is not.
// (This may or may not prevent the optimization from working. Often it does not prevent the optimization from working.)
//=================================================================================================================================
rnk=(int) Eigen::FullPivLU<Eigen::MatrixXd>(ulHess[0]).rank();
if (rnk < inUCmps && dbug >=2)
cout << endl << "WARNING: The UL Hessian matrix is singular which may prevent the component probabilities from being optimized." << endl
<< "
This can happen for various reasons, two of which are:" << endl
<< "
1.) The likelihood of a single component is zero at all time points, which eliminates the decision variable." << endl
<< "
2.) The Upper Left matrix is ill‐conditioned having large and small elements at different diagonal positions." << endl
<< "Message from ... getHessE() ..." << endl;
117
// Hessian: Build upper right, lower left, and lower right sections of the Hessian. Iterate until the entire Hessian is full rank so that Newton's method may be applied. This may require
//
multiplying all constraints by a constant (both LHS and RHS). First, make sure that the constraint matrix is of full rank since components may be dropped during the 1st Step.
//=============================================================================================================================================================================================
rnk=(int) Eigen::FullPivLU<Eigen::MatrixXd>(inCMtrx).rank();
if (rnk < (int) inCVctr.size())
{
cout << "ERROR: Detection of a non‐full rank constraint matrix rank during the ECME 1st Step, which is likely due to" << endl
<< "
components with < 0 probabilities being dropped. The function getCMtrx() can be used to fix this" << endl
<< "
by sequentially removing linearly dependent rows until the constraint matrix becomes full rank." << endl
<< "EXITING...getHessE()..." << endl; cin.get(); exit (EXIT_FAILURE);
}
inLHS[0]=Eigen::MatrixXd((int) inCVctr.size(),inUCmps); inLHS[0]=mult*inCMtrx;
inRHS[0]=Eigen::VectorXd((int) inCVctr.size());
inRHS[0]=mult*inCVctr;
do
{
// Hessian: Upper right & Lower left.
//======================================
for (int r=0; r<inUCmps; ++r)
for (int c=inUCmps; c<ld; ++c)
{
inHess[0](r,c)=‐1.00*inLHS[0](c‐inUCmps,r);
inHess[0](c,r)=inHess[0](r,c);
}
// Hessian: Lower right.
//=========================
for (int r=inUCmps; r<ld; ++r)
for (int c=r; c<ld; ++c)
{
inHess[0](r,c)=0.00;
if (c > r)
inHess[0](c,r)=inHess[0](r,c);
}
// Before proceeding, ensure that the entire Hessian is full rank.
//==================================================================
rnk=(int) Eigen::FullPivLU<Eigen::MatrixXd>(inHess[0]).rank();
// If another iteration is needed, update the multiplier and the LHS/RHS constraints.
//=====================================================================================
if (rnk < ld)
{
mult=mult*10.00;
inRHS[0]=mult*inRHS[0];
inLHS[0]=mult*inLHS[0];
}
} while (rnk < ld && mult < LPosVal);
// If the entire Hessian is not full rank then put out a warning but do not exit, the optimization may still succeed.
//=====================================================================================================================
if (rnk < ld && dbug >=2)
{
cout << endl << "WARNING: The 1st Step Hessian matrix is singular which may prevent the multivariate component probabilities from being optimized." << endl
<< "
This can happen for various reasons, including:" << endl
<< "
1.) There are very large and very small eigenvalues, resulting in an ill‐conditioned matrix." << endl
<< "
2.) Components have been dropped and the constraint matrix is no longer of full row rank. (Note: Already checked above.)" << endl
<< "Message from ... getHessE() ..." << endl;
}
// Free temporary memory allocations.
//=====================================
delete [] ulHess; ulHess=nullptr;
// Return the multiplier used to correct an ill‐conditioned Hessian.
//====================================================================
118
return mult;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getGradM.cpp
/
/ Function: getGradM()
/
/ Summary:
/
/
This function derives the Gradient for the ECME algorithm 2nd Step optimization, which is (evidently) NOT convex in the decision variables (covariances). In general mixture density likelihoods are
/
not concave functions and have many local optimums. We are dealing with a multivariate mixture density here. All means, variances and component probabilities are treated as constants during this
/
ECME 2nd Step optimization. The objective is to maximize the corresponding log‐likelihood function (only covariances are unknown) subject to constraints that all variance‐covariance matrices are
/
positive definite. That is, we seek the covariances that maximize the multivariate mixture log‐likelihood function with the variance‐covariance matrix at each component being positive definite
/
(and all means, variances, component probabilities being fixed and known). These points are found by applying a modified Newton's method to the log‐likelihood (in which only the covariances are
/
unknown). Newton's method requires that the gradient and Hessian of the log‐likelihood be constructed during each iteration. The gradient is the vector of first order partial derivatives WRT each
/
covariance term, and is derived in this function. The Hessian is the matrix of second order partial derivatives WRT all covariance terms. If the problem has "A" total assets with returns measured,
/
and "U" components in the multivariate mixture density then there will be a total of U*A*(A‐1)/2 unique covariance terms that require estimation. Clearly this problem suffers from the curse of
/
dimensionality and will work best with a limited number of assets relative to the number of observations (i.e., time points). The optimization problem converges when the log‐likelihood fails to
/
improve. The method used to find the largest local optimum is due to Marquardt (1963), which uses the Hessian to step in the general direction of the gradient by adding a constant to the Hessian
/
diagonals prior to solving the updating equation. In practice we will iterate over a large number of random step sizes searching for the local/global log‐likelihood function maximizer. Once the
/
maximizer is found we recompute the gradient and Hessian and iterate again. Note that large additive quantities (i.e., added to the Hessian diagonal) translate into small steps and small quantities
/
translate into large steps. An additive factor of zero translates into using Newton's method without modification, which is assumed here to overshoot the local optimizer. This method is approprate
/
when strictly applying Newton's method overshoots. Here, the goal is to find the largest local optimum in the vicinity of the carefully constructed starting point (i.e., LP solution), but also to
/
search outside the current region/hill in an attempt to find a better solution. The constraints on the resulting variance‐covariance matrices are enforced implictly. At each step, the resulting
/
matrix is decomposed and the eigenvalues are inspected. If none are <= 0 the resulting variance‐covariance matrix is positive definite. Otherwise, it is not and a ridge repair is immediately
/
performed. Stepping continues using the repaired variance‐covariance matrix. In general, we find that there are large regions of the covariance set where stepping proceeds without the need for
/
repairs, and other large regions of the covariance set where repairs are needed after each step. The variance‐covariance matrix of each multivariate component is examined and repaired (if
/
necessary) by the function ridgeRpr(). The feasible region is any covariance set that results in all component variance‐covariance matrices being valid (i.e., positive definite).
/
/ Inputs:
/
/
1.) The total number of time points with data collected. (T)
/
/
2.) The array of vector returns at each time point indexed as rs[t](a). There are T vectors of returns and each is of size numA, where numA=# assets. (rs)
/
/
3.) The number of unique components in the multivariate mixture. (Initial value is from either the Minimax or Minimum SSD LP optimizations.) (inUCmps)
/
/
4.) The number of assets with returns collected. (numA)
/
/
5.) A double array of likelihood values indexed by time and component. At each time point the likelihood for each component is computed and stored for reuse. This forms a 2‐dimensional grid of
/
values of size TxU, where T=# of time points, and U=# components. (infVals)
/
/
6.) A single array of T values that sum the double array in parameter #5 across the components at each time point, weighting each likelihood value by its corresponding estimated component
/
probability. Therefore, each value in this array is the multivariate mixture likelihood value for the data at each individual time point. (inDNoms)
/
/
7.) The current array of multivariate mixture probabilities. (Initial values from either the Minimax or Minimum SSD LP optimizations.) (inPrbs)
/
/
8.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector
/
of means for the first multivariate component, etc... (muMns)
/
/
9.) The array of VC matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension (numA)x(numA).
/
The diagonals of each VC matrix are the corresponding variances for that asset within that component. (E)
/
/
10.) The array of VC matrix inverses. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension
/
(numA)x(numA). The diagonals of each VC matrix are the corresponding variances for that asset within that component. This parameter holds the corresponding array of inverses of the VC
/
matrices. (Einv)
/
/
11.) The array of identifier matrices for each covariance term. A VC matrix can be decomposed into the sum of a matrix of diagonal elements and a term for each unique covariance. The constant
119
/
matrix multiplied by the covariance term has a 1/0 in each element where 1 is in the location of the corresponding covariance term. The constant matrices are contained in this array. The
/
constant matrices are identical across components and can be reused. The # of these matrices is the number of unique covariance terms for a single multivariate mixture: numA*(numA‐1)/2. (inA)
/
/
12.) The empty Gradient vector to be filled by this function. The vector is of dimension equal to the total # of covariances in the problem. If the problem has "A" total assets with returns
/
measured, and "U" components in the multivariate mixture density then there are a total of U*A*(A‐1)/2 covariance terms. (inGrad)
/
/ Outputs:
/
/
This function populates the empty gradient vector supplied to it but does not return any other output at the function call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void getGradM(const int T, Eigen::VectorXd *rs, const int inUCmps, const int numA, const long double **infVals, const long double *inDNoms, const long double *inPrbs, Eigen::VectorXd *muMns, const
Eigen::MatrixXd *E, const Eigen::MatrixXd *Einv, const Eigen::MatrixXd *inA, Eigen::VectorXd *inGrad, int chk)
{
// Local variables.
//===================
int itr=0, itrA;
long double Qtijk;
Eigen::MatrixXd *Ejk=new Eigen::MatrixXd [1];
Ejk[0]=Eigen::MatrixXd(numA,numA);
// Populate gradient vector.
//============================
for (int i=0; i<inUCmps; ++i)
{
itrA=0;
for (int j=0; j<numA; ++j)
for (int k=j+1; k<numA; ++k)
{
getCofM(numA,j,k,E[i],Ejk);
inGrad[0](itr)=0.00;
for (int t=0; t<T; ++t)
{
Qtijk = (0.50)*((rs[t]‐muMns[i]).transpose())*Einv[i]*inA[itrA]*Einv[i]*(rs[t]‐muMns[i]) ‐ Ejk[0].determinant()/E[i].determinant();
inGrad[0](itr)=inGrad[0](itr) + (inPrbs[i]*infVals[t][i]*Qtijk)/inDNoms[t];
}
itr++;
itrA++;
}
}
// Delete temporary memory allocations.
//=======================================
delete [] Ejk; Ejk=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: getHessM.cpp
/
/ Function: getHessM()
/
/ Summary:
/
/
This function derives the Hessian matrix for the ECME algorithm 2nd Step optimization, which is (evidently) NOT convex in the decision variables (covariances). In general a univariate/multivariate
/
mixture density likelihood function is not strictly concave and has many local optimums. We are dealing with a multivariate mixture density here where all means, variances and component
/
probabilities fixed during the ECME 2nd Step optimization. The objective is to maximize the corresponding log‐likelihood function (only covariances are unknown) subject to constraints that all
/
resulting variance‐covariance matrices are positive definite. That is, we seek the covariances that maximize the multivariate mixture log‐likelihood function with the variance‐covariance matrix at
each component being positive definite (and all means, variances, component probabilities are fixed). These points are found by applying a modified Newton's method to the log‐likelihood (in which
/
/
only the covariances are unknown). Newton's method requires that the gradient and Hessian of the log‐likelihood be constructed during each iteration. The gradient is the vector of first order
120
/
partial derivatives WRT each covariance term, and is derived in the function getGradM(). The Hessian is the matrix of second order partial derivatives WRT all covariance terms and is derived in
/
this function. If the problem has "A" total assets with returns measured, and "U" components in the multivariate mixture density then there will be a total of U*A*(A‐1)/2 unique covariance terms
/
that require estimation. This problem suffers from the curse of dimensionality and will work best with a limited number of assets relative to the # of observations (i.e., time points). The
/
optimization problem converges when the log‐likelihood fails to improve. The method used to find the largest local optimum is due to Marquardt (1963), which uses the Hessian to step in the general
/
direction of the gradient by adding a constant to the Hessian diagonals prior to solving the updating equation. In practice we will iterate over a large number of random step sizes searching for
/
the best local and global log‐likelihood function maximizer. Once the maximizer is found we recompute the gradient and Hessian and iterate again. Note that large additive quantities translate into
/
small steps and small quantities translate into large steps. An additive factor of zero translates into using Newton's method without modification. This method is appropriate when strictly applying
/
Newton's method overshoots. Here, the goal is to find the largest local optimum in the vicinity of the carefully constructed starting point (LP solution) using small step sizes, but also to search
/
outside the current region/hill in an attempt to find a better solution (using large step sizes). The constraints on the resulting variance‐covariance matrices are enforced implictly. At each
/
step, the resulting matrix is decomposed and the eigenvalues are inspected. If none are <= 0 the resulting variance‐covariance matrix is positive definite. Otherwise, it is not and a ridge repair
/
is immediately performed and stepping continues using the repaired variance‐covariance matrix. In general, we find that there are large regions of the covariance set where stepping proceeds without
/
the need for repairs, and other large regions of the covariance set where repairs are needed after each step. The variance‐covariance matrix of each multivariate component is examined and repaired
/
(if necessary) by the function ridgeRpr(). The feasible region is any covariance set that results in all component variance‐covariance matrices being valid (i.e., positive definite).
/
/ Inputs:
/
/
1.) The total number of time points with data collected. (T)
/
/
2.) The array of vector returns at each time point indexed as rs[t](a). There are T vectors of returns and each is of size numA, where numA=# assets. (rs)
/
/
3.) The number of unique components in the multivariate mixture. (Initial value is from either the Minimax or Minimum SSD LP optimizations.) (inUCmps)
/
/
4.) The number of assets with returns collected. (numA)
/
/
5.) A double array of likelihood values indexed by time and component. At each time point the likelihood for each component is computed and stored for reuse. This forms a 2‐dimensional grid of
/
values of size TxU, where T=# of time points, and U=# components. (infVals)
/
/
6.) A single array of T values that sum the double array in parameter #5 across the components at each time point, weighting each likelihood value by its corresponding estimated component
/
probability. Therefore, each value in this array is the multivariate mixture likelihood value for the data at each individual time point. (inDNoms)
/
/
7.) The current array of multivariate mixture probabilities. (Initial values from either the Minimax or Minimum SSD LP optimizations.) (inPrbs)
/
/
8.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector
/
of means for the first multivariate component, etc... (muMns)
/
/
9.) The array of VC matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension (numA)x(numA).
/
The diagonals of each VC matrix are the corresponding variances for that asset within that component. (E)
/
/
10.) The array of VC matrix inverses. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension
/
(numA)x(numA). The diagonals of each VC matrix are the corresponding variances for that asset within that component. This parameter holds the inverses of the VC matrices. (Einv)
/
/
11.) The array of identifier matrices for each covariance term. A VC matrix can be decomposed into the sum of a matrix of diagonal elements and a term for each unique covariance. The constant
/
matrix multiplied by the covariance term has a 1/0 in each element where 1 is in the location of the corresponding covariance term. The constant matrices are contained in this array. The
/
constant matrices are identical across components and can be reused. The # of these matrices is the number of unique covariance terms for a single multivariate mixture: numA*(numA‐1)/2. (inA)
/
/
12.) The empty Hessian matrix to be filled by this function. The matrix is square and of dimension equal to the total # of covariances in the problem. If the problem has "A" total assets with
/
returns measured, and "U" components in the multivariate mixture density then there are a total of U*A*(A‐1)/2 covariance terms. (inHess)
/
/ Internal Variables:
/
/
Note: Each Hessian element is a partial WRT Sigma(i,j,k), then another partial WRT Sigma(p,r,s). Where i is the component index for the 1st partial derivative, and p is the component index for the
/
2nd partial derivative. The paired index (j,k) identifies the covariance term from component i, whereas the paired index (r,s) identifies the covariance term from component p. Note that the
/
covariance term at (j,k) is equivalent to the covariance term at (k,j) therefore we will also assume that j < k, and r < s.
/
/
itrAjk:
Index of the indicator matrix array A[] for the 1st covariance term in the partial.
/
itrArs:
Index of the indicator matrix array A[] for the 2nd covariance term in the partial.
/
fti:
Product of the likelihood value for the observations at time t using density for component i (of C) and the corresponding component probability.
/
fti_p:
Partial derivative of fti (defined above) WRT a covariance term from component p (of C) where (p == i). [Note that fti is just a constant when (p != i) since it does not contain
/
the 2nd covariance term in its function.]
/
ftp:
Product of the likelihood value for the observations at time t using density for component p (of C) and the corresponding component probability.
/
Qtijk:
This is the extra term that arises in the numerator when differentiating the density of component i WRT the covariance term (j,k).
/
Qtprs:
This is the extra term that arises in the numerator when differentiating the density of component p WRT the covariance term (r,s).
/
Qtijk_prs: Partial derivative of Qtijk WRT Sigma(p,r,s).
/
gt:
Overall likelihood of all data points using the full multivariate mixture density.
/
gt_prs:
Partial derivative of gt WRT Sigma(p,r,s).
121
/
‐‐‐‐> Note:
If there are n assets then there are n(n‐1)/2 distinct covariance terms within each component and C*n*(n‐1)/2 total distinct covariance terms.
/
/ Outputs:
/
/
This function populates the empty Hessian matrix supplied. The magnitude of the largest element is returned at the call and used to help determine the best step size.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
long double getHessM(const int T, Eigen::VectorXd *rs, const int inUCmps, const int numA, const long double **infVals, const long double *inDNoms, const long double *inPrbs, const Eigen::VectorXd *muMns,
const Eigen::MatrixXd *E, const Eigen::MatrixXd *Einv, const Eigen::MatrixXd *inA, Eigen::MatrixXd *inHess)
{
// Local variables.
//===================
int hr=0, hc, itrAjk, itrArs;
long double fti, fti_p, ftp, Qtijk, Qtprs, Qtijk_prs, gt, gt_prs, q1, q2, q3, maxmag=0.00;
Eigen::MatrixXd *Eijk=new Eigen::MatrixXd [1], *Eprs=new Eigen::MatrixXd [1], *Ejkrs=new Eigen::MatrixXd [1], *Ejksr=new Eigen::MatrixXd [1];
Eijk[0]=Eigen::MatrixXd(numA,numA); Eprs[0]=Eigen::MatrixXd(numA,numA); Ejkrs[0]=Eigen::MatrixXd(numA,numA); Ejksr[0]=Eigen::MatrixXd(numA,numA);
Eigen::MatrixXd q4=Eigen::MatrixXd(numA,numA);
// Populate Hessian matrix.
//===========================
for (int i=0; i<inUCmps; ++i)
{
itrAjk=0;
for (int j=0; j<numA; ++j)
for (int k=j+1; k<numA; ++k)
// 1st covariance term fixed: Cov(i,j,k)
{
hc=0;
// Hessian column entry indicator (0 thru nCovs‐1).
getCofM(numA,j,k,E[i],Eijk);
// Build Eijk.
for (int p=0; p<inUCmps; ++p)
{
itrArs=0;
for (int r=0; r<numA; ++r)
for (int s=r+1; s<numA; ++s)
// 2nd covariance term fixed: Cov(p,r,s)
{
if (hr <= hc)
{
// Unconditional quantities that are not functions of time.
//===========================================================
getCofM(numA,r,s,E[p],Eprs);
// Build Ers.
// Conditional quantities that are not functions of time.
//=========================================================
if (i == p)
/* 2nd covariance term is from the same component. */
{
getCofM(numA,r,s,Eijk[0],Ejkrs);
// Build Ejkrs, conditionally.
getCofM(numA,s,r,Eijk[0],Ejksr);
// Build Ejksr, conditionally.
q1=2.00*(Eijk[0].determinant())*(Eprs[0].determinant())/pow(E[i].determinant(),2);
q2=(Ejksr[0].determinant())/(E[i].determinant());
q3=(Ejkrs[0].determinant())/(E[i].determinant());
q4=Einv[i]*inA[itrAjk]*Einv[i]*inA[itrArs]*Einv[i] + Einv[i]*inA[itrArs]*Einv[i]*inA[itrAjk]*Einv[i];
}
// Initialize the element.
//==========================
inHess[0](hr,hc)=0.00;
// Iterate over the time dimension.
//===================================
for (int t=0; t<T; ++t)
{
// Derive unconditional quantities that are functions of time.
//==============================================================
fti=inPrbs[i]*infVals[t][i];
Qtijk=(0.50)*((rs[t]‐muMns[i]).transpose())*Einv[i]*inA[itrAjk]*Einv[i]*(rs[t]‐muMns[i]) ‐ Eijk[0].determinant()/E[i].determinant();
ftp=inPrbs[p]*infVals[t][p];
122
Qtprs=(0.50)*((rs[t]‐muMns[p]).transpose())*Einv[p]*inA[itrArs]*Einv[p]*(rs[t]‐muMns[p]) ‐ Eprs[0].determinant()/E[p].determinant();
gt=inDNoms[t];
// Derive conditional quantities that are functions of time, including the Hessian value itself.
//================================================================================================
if (i == p)
// 2nd correlation term is from same component.
{
fti_p=fti*Qtprs;
gt_prs=ftp*Qtprs;
if (j == r || k == s)
// 2nd partial WRT term below diagonal only.
Qtijk_prs=q1 ‐ q2 ‐ 0.50*((rs[t]‐muMns[i]).transpose())*(q4)*(rs[t]‐muMns[i]);
else if (k == r)
// 2nd partial WRT term above diagonal only.
Qtijk_prs=q1 ‐ q3 ‐ 0.50*((rs[t]‐muMns[i]).transpose())*(q4)*(rs[t]‐muMns[i]);
else
// 2nd partial WRT term above and below diagonal.
Qtijk_prs=q1 ‐ (q2+q3) ‐ 0.50*((rs[t]‐muMns[i]).transpose())*(q4)*(rs[t]‐muMns[i]);
inHess[0](hr,hc)=inHess[0](hr,hc) + (gt*(fti*Qtijk_prs + fti_p*Qtijk) ‐ (fti*Qtijk)*gt_prs)/pow(gt,2);
}
else
inHess[0](hr,hc)=inHess[0](hr,hc) ‐ (fti*Qtijk*ftp*Qtprs)/pow(gt,2);
}
// Populate corresponding element below the diagonal.
//=====================================================
if (hr < hc)
inHess[0](hc,hr)=inHess[0](hr,hc);
}
hc++;
itrArs++;
// Hessian column entry indicator (0 thru nCovs‐1).
}
}
hr++;
itrAjk++;
}
}
// Delete temporary memory allocations.
//=======================================
delete [] Eijk; Eijk=nullptr;
delete [] Eprs; Eprs=nullptr;
delete [] Ejkrs; Ejkrs=nullptr;
delete [] Ejksr; Ejksr=nullptr;
// Return the magnitude of the largest element.
//===============================================
for (int i=0; i<(int) inHess[0].rows(); ++i)
for (int j=0; j<(int) inHess[0].cols(); ++j)
if (abs(inHess[0](i,j)) > maxmag)
maxmag=abs(inHess[0](i,j));
return maxmag;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: stepHessM.cpp
/
/ Function: stepHessM()
/
/ Summary:
/
/
This function steps in the general direction of steepest ascent for the multivariate mixture likelihood that maintains the marginal mixture densities, using Marquardt (1963). The multivariate
/
mixture likelihood has fixed means, variances, and component probabilities but unknown covariances. The multivariate mixture here is therefore a function of only the unknown covariances, and we
123
/
seek to maximize it. A problem is that the function can have multiple local optimums. Using Marquardt (1963) a constant term of random size is added/subtracted to the diagonals of the
/
correpsonding Hessian matrix prior to solving the updating equations when implementing Newton's method. Adding a large constant results in taking a small step, and adding a small constant results
/
in taking a large step. This method will allow us to take small steps towards the local optimum of the current hill and simultaneously search for larger hills in the general direction of steepest
/
ascent. The minimum step size is set (to zero) in the header file (see minHessAdd) and the maximum step size is set to 10^X where X = # digits in the maximum Hessian element + 2. Within the
/
min/max range a step size is randomly generated for each iteration. The stepping is multi‐threaded (see the wrapper that invokes this function) to save time and the total # of threads used is equal
/
to the # of independent processing units on the PC running the application multiplied by the global constant nCorMult (also set in the header file). Each thread will take a total of mItersH random
/
sized steps, which is also a global constant set in the header file. To cover the step size range from min‐to‐max using a total of nCorMult*(# of independent processing units) threads we first
/
generate a random value of X, between 1 and the maximum # of digits in the largest Hessian element + 2, then generate a step size randomly between 0 and 10^X. At each iteration, a step is taken by
/
adding/subtracting the random step size to the Hessian diagonals and solving the Newton's method updating equations. The decision variables are the U*A*(A‐1)/2 unique covariances, where U = total #
/
multivariate mixture components and A = total # assets. After obtaining the new solution, all corresponding variance‐covariance matrices are confirmed to be positive definite, and if not a ridge
/
repair is immediately performed using a random multiplier between the 2 values (rrMultMin and rrMultMax) specified as global constants in the header file. The decision variables that maximize the
/
likelihood function are returned, as is the maximum log‐likelihood value along with the random step size that generates the maximum. The best solution across all threaded calls is then used for the
/
current iteration of the ECME 2nd Step. After stepping finishes, processing returns to the top of the 2nd Step and the 2nd Step gradient and Hessian are rebuilt for another iteration (in ECMEAlg()).
/
/ Inputs:
/
/
1.) Input supplied to this function as a 4‐element integer array. The element at position [0] is the # of time points with data. The element at position [1] is the # of unique componenents in the
/
current multivariate mixture solution. The element at position [2] is the current thread # determined by the function that generates threaded calls. The thread # is only used within this
/
function for reporting results to the output window, for example during debugging. The element at position [3] is an indicator that the current thread has launched. (inHess1)
/
/
2.) Input/output both supplied to this function and generated by this function as a 5‐element long double array. Elements at positions [0], [1], [2] are input parameters and elements at positions
/
[3] and [4] are output generated by this function and returned to the calling function. The input at position [0] is the current optimal log‐likelihood value we are attempting to improve upon
/
during this ECME 2nd Step iteration. Input at position [1] is the step size, and the input at position [2] is the starting value for stepping in this threaded call. Elements at positions [3]
/
and [4] are placeholders for return values. The maximum log‐likelihood value found during this stepping iteration is returned in element [3], and the step size multiplier that generates this
/
maximum is returned in element [4]. (inHess2)
/
/
3.) The array of vector returns at each time point indexed as rs[t](a). There are T vectors of returns and each is of size numA, where numA=# assets. (rs)
/
/
4.) The current vector of 2nd Step decision variables (i.e., all covariances), as an array of 4 elements. The vector at indices [0] and [2] hold the current covariance estimates. The vector at
/
index [3] holds the returned estimates that maximize the log‐likelihood for this function call. The vector at position [1] holds the updated covariance estimates derived at each step. (inDvars)
/
/
5.) The current gradient vector evaluated at the current values of the covariance estimates (i.e., decision variables). (inGrad)
/
/
6.) The current Hessian matrix evaluated at the current values of the covariance estimates (i.e., decision variables). (inHess)
/
/
7.) The current vector of multivariate mixture probabilities. (Initial values are from either the Minimax or Minimum SSD LP optimizations.) (uPrbs)
/
/
8.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector
/
of means for the first multivariate component, etc... (muMns)
/
/
9.) The array of VC matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension (numA)x(numA).
/
The diagonals of each VC matrix are the corresponding variances for that asset within that component. (inVCs)
/
/ Outputs:
/
/
This function updates element [3] of incoming parameter inDvars[] array with the covariance estimates that maximize the log‐likelihood during this stepping iteration. In addition, elements at
/
positions [3] and [4] of incoming array inHess2 are updated with the maximum value of the log‐likelihood, and the step size multiplier that maximizes the log‐likelihood, respectively.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void stepHessM(int *inHess1, long double *inHess2, const Eigen::VectorXd *rs, Eigen::VectorXd *inDvars, const Eigen::VectorXd inGrad, const Eigen::MatrixXd inHess, const Eigen::VectorXd uPrbs, const
Eigen::VectorXd *inMns, const Eigen::MatrixXd *inVCs)
{
// Declare/initialize local variables.
//======================================
Eigen::initParallel();
int rpr, nA=(int) inMns[0].size(), inT=inHess1[0], inUCmps=inHess1[1], hMaxLen=inHess1[4], expon, curExp;
long double hIdx=1.00, hmult, curMaxLL=inHess2[0], LL, picst=exp(((double) (nA/2.00))*log(2.00*pi)), *sq_dets=new long double [inUCmps], *d_nom=new long double [inT], **f_vals=new long double
*[inT], strt=inHess2[2], jmp, curMaxMult=strt, mult;
random_device rd; default_random_engine gen(rd());
uniform_real_distribution<long double> udist(0.0,1.0);
Eigen::MatrixXd *Idm=new Eigen::MatrixXd [1], tHessM(inHess.rows(),inHess.cols()), *VC_Inv=new Eigen::MatrixXd [inUCmps], *V_C=new Eigen::MatrixXd [inUCmps];
Idm[0]=Eigen::MatrixXd(inHess.rows(),inHess.cols()); getIDM(Idm);
// Size the array holding all component likelihood values at all time points.
//=============================================================================
124
for (int t=0; t<inT; ++t)
f_vals[t]=new long double [inUCmps];
// Size the local VC and inverse matrices.
//==========================================
for (int v=0; v<inUCmps; ++v)
{
V_C[v]=Eigen::MatrixXd(nA,nA);
V_C[v]=inVCs[v];
VC_Inv[v]=Eigen::MatrixXd(nA,nA);
}
// Initialize the covariance maximizers to the starting values. (Write details when debugging.)
//================================================================================================
inDvars[3]=inDvars[0];
if (dbug <= 1)
{
cout << ".";
inHess1[3]=1;
}
else if (dbug >=2)
{
if (0 == inHess1[2])
cout << endl;
cout << "Launching thread # " << inHess1[2] << endl;
}
// Set the multiple for variance‐covariance matrix repairs. This applies to all steps.
//=======================================================================================
mult = rrMultMin + udist(gen)*(rrMultMax ‐ rrMultMin);
// Iterate using the Hessian and solve for new covariances.
//===========================================================
for (int i=0; i<mItersH; ++i)
{
// Hessian has element with maximum length equal to hMaxLen digits. Randomly select value between 1 and this # to use as the max for stepping.
//===============================================================================================================================================
expon = 1 + (int) (udist(gen)*((double) hMaxLen));
jmp = pow(10.0,expon);
hmult = minHessAdd + udist(gen)*jmp;
if (udist(gen) <= 0.50)
{
hmult=‐1.00*hmult;
expon=0‐expon;
// Indicate the sign/direction. (Value of ‐10 means exponent of 10 used for backward stepping.)
}
tHessM=inHess + hmult*Idm[0];
inDvars[1]=tHessM.colPivHouseholderQr().solve(‐1.00*udist(gen)*inGrad);
inDvars[0]=(inDvars[1]+inDvars[2]);
// Check that the new covariance estimates yield PD VC matrices. (Repair if broken, and update the corresponding vector of current decision variables.)
//========================================================================================================================================================
rpr=0;
for (int v=0; v<inUCmps; ++v)
{
setCovs(v,V_C,inDvars);
rpr=rpr + ridgeRpr(v,V_C,mult);
VC_Inv[v]=V_C[v].inverse();
sq_dets[v]=sqrt(VC_Inv[v].determinant());
}
LL=getLFVals(inT,nA,inUCmps,rs,uPrbs,inMns,VC_Inv,sq_dets,picst,d_nom,f_vals);
if (rpr > 0)
getCovs(inUCmps,V_C,inDvars);
125
// Check vs existing maximum LL. If larger, then update the current maximum and covariance array.
//==================================================================================================
if (LL > curMaxLL)
{
curMaxLL=LL;
curMaxMult=hmult;
inDvars[3]=inDvars[0];
curExp=expon;
inHess1[6]=rpr;
}
}
// Return the maximum log‐likelihood value along with the mutliplier that generated it. (Write details when debugging.)
//========================================================================================================================
inHess2[3]=curMaxLL;
inHess2[4]=curMaxMult;
inHess2[6]=mult;
inHess1[5]=curExp;
if (dbug <= 1)
inHess1[7]=1;
else if (dbug >= 2)
cout << "Done with thread # " << inHess1[2] << " which started at: " << strt << " maximum LL is: " << curMaxLL << endl;
// Delete temporary memory allocations.
//=======================================
delete [] Idm; Idm=nullptr;
delete [] d_nom; d_nom=nullptr;
delete [] sq_dets; sq_dets=nullptr;
delete [] V_C; V_C=nullptr;
delete [] VC_Inv; VC_Inv=nullptr;
for (int t=0; t<inT; ++t)
{
delete [] f_vals[t]; f_vals[t]=nullptr;
}
delete [] f_vals; f_vals=nullptr;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: ridgeRpr.cpp
/
/ Function: ridgeRpr()
/
/ Summary:
/
/
This function accepts a variance‐covariance matrix as input and determines whether or not it is positive‐definite. If not, it performs a ridge repair to make the matrix positive definite. To
/
determine whether or not the matrix is positive definite, an eigenvalue decomposition is performed. If all eigenvalues are > 0 then the matrix is positive definite. A necessary and sufficient
/
condition for a variance‐covariance matrix to be valid is that it is positive definite. To perform a ridge repair, the diagonal elements are all multiplied by the same constant value (> 1). Since
/
the diagonals are the variances, this implies that we increase the variances. Doing so will automatically reduce the size of the covariances relative to the variances. It will also reduce the
/
magnitude of the correlations. The covariances are the correlations multiplied by the standard deviations. That is, Cov(X,Y) = rho*std(X)*std(Y). When std(X) and std(Y) increase and Cov(X,Y)
/
remains constant the correlations decrease in magnitude. If the constant multiplier is large enough we will drive the correlations to near zero, and at this point the covariances will be extremely
/
small relative to the variances. The resulting matrix approaches a diagonal matrix which is positive definite. The point is to use a small multiplier and increase it iteratively until the repaired
/
matrix becomes positive definite and then scale it back so that the variances are undisturbed, but the off‐diagonal elements are smaller relative to the diagonal elements. If a matrix is positive
/
definite, then that matrix multiplied by a constant is also positive definite (easily proven with the definition of positive definite). The initial multiplier is randomly generated in the calling
/
function and bounded by the global constants rrMultMin and rrMultMax, which are set in the header file. This function iterates, multiplying the diagonals by (1.00 + i*rrmult) and checking for
/
positive definiteness after each iteration ("i" is the iteration index). A variance‐covariance matrix that is badly broken (for example with a correlation term > 100, when all such quantities
/
should be between ‐1.00 and +1.00) may require a large number of iterations to repair. Therefore, to speed up processing we increase the multiplier by a factor of 10 after each 100 iterations.
/
That is, after 100 iterations rrmult is multiplied by 10, and again after 200 iterations, etc... Note that The variance‐covariance matrix supplied to this function is a modifiable value and is
/
updated in place with care taken to ensure that the diagonals are not disturbed when it is returned.
/
126
/ Inputs:
/
/
A single variance‐covariance matrix is supplied to this function using 2 inputs: (1) An array of variance‐covariance matrices and (2) an index to identify the one we are checking for positive‐
/
definiteness, and repairing if necessary.
/
/
1.) The index that identifies the variance‐covariance matrix to be checked/repaired. (uCell)
/
/
2.) The array of variance‐covariance matrices for the current multivariate mixture solution. (E)
/
/
3.) The multiple used to add a ridge. (mult)
/
/ Outputs:
/
/
This function returns an integer value of 1 or 0. A 1 is returned if the variance‐covariance matrix is repaired, and a 0 is returned if it is not in need of repair.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
int ridgeRpr(const int uCell, Eigen::MatrixXd *E, long double mult)
{
// Local variables.
//===================
Eigen::initParallel();
int pd, retvar=0;
long double sclfctr, rr_mult=mult, det;
Eigen::EigenSolver<Eigen::MatrixXd> egnslvr;
Eigen::MatrixXd newVCm((int) E[uCell].rows(),(int) E[uCell].cols());
// Is the VC matrix Positive‐definite as required? If not, repair it.
//======================================================================
egnslvr.compute(E[uCell], false);
pd=1; det=1.0;
for (int a=0; a<(int) E[uCell].rows(); ++a)
{
if (egnslvr.eigenvalues()[a].real() <= pdmineval)
pd=0;
det=det*egnslvr.eigenvalues()[a].real();
}
if (det <= detminval)
pd=0;
// Repair is necessary.
//=======================
if (0 == pd)
{
// VC matrix is not positive definite and a ridge repair will be performed.
//===========================================================================
newVCm=E[uCell];
// Repair the matrix.
//=====================
for (int i=1; 0==pd; ++i)
{
// Increase each variance by a factor > 1 and rescale.
//======================================================
sclfctr=(1 + ((double) i)*rr_mult);
for (int r=0; r<(int) E[uCell].rows(); ++r)
for (int c=0; c<(int) E[uCell].cols(); ++c)
if (r != c)
newVCm(r,c)=E[uCell](r,c)/sclfctr;
// Check if the updated VC matrix is positive‐definite.
//=======================================================
egnslvr.compute(newVCm, false);
pd=1; det=1.0;
for (int a=0; a<(int) E[uCell].rows(); ++a)
{
127
if (egnslvr.eigenvalues()[a].real() <= pdmineval)
pd=0;
det=det*egnslvr.eigenvalues()[a].real();
}
if (det <= detminval)
pd=0;
// Increase the scale factor by a multiple of 10 each 100 iterations.
//=====================================================================
if (0 == (int) (i % 100))
rr_mult=10*rr_mult;
}
// Replace the original VC matrix with the repaired version.
//============================================================
E[uCell]=newVCm;
retvar=1;
}
// If the VC matrix has been repaired return a 1, otherwise return a 0.
//=======================================================================
return retvar;
}
/*
/ Copyright (C) 2016 Chris Rook
/
/ This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License,
/ or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
/ PARTICULAR PURPOSE. See the GNU General Public License for more details: <http://www.gnu.org/licenses/>
/
/ Filename: wrtDens.cpp
/
/ Function: wrtDens()
/
/ Summary:
/
/
This function writes out the structure of the final multivariate density along with the actual values contained in that structure. The output has 2 parts: 1.) the structure, which is a sum of
/
multivariate normal densities weighted by the corresponding component probability using parameters, and 2.) details of the parameter values. For each component, the definitions include the actual
/
component probability, the mean vector, the variance‐covariance matrix, and the unique cell that this component originates from with respect to the full factorial of combinations defined after
/
fitting the univariate mixture densities. Finally, the inverse of the variance‐covariance matrix is printed as is the rank and the determinant. This function can be used to write the supplied
/
density to either standard output (using cout in the last parameter), or a file (using an ofstream fout definition in the last parameter).
/
/ Inputs:
/
/
1.) Specify the type of starting point when transitioning between univariate marginal densities and the multivariate mixture PDF. In this application we offer 2 transition methods: Minimax or
/
Minimum Sum of Squared Distances (SSD). Each are solved as constrained linear programs (LPs). The constraints maintain the univariate marginal mixture densities. This string variable takes
/
one of two values: "Minimax" or "Minimum Sum of Squared Distances (SSD)". This is for display purposes only and informs the user which transition method was used for the given multivariate
/
density function being written. (typ)
/
/
2.) The number of unique components in the multivariate mixture that results from either the Minimax or Minimum SSD LP optimization starting points. (nUCmps)
/
/
3.) The array of unique cell IDs that link each component of the multivariate density back to the full factorial of components. The full factorial of components represents each cell in the
/
multidimensional grid formed by considering all combinations of assets and their levels. (Note that the full factorial would be required to build a multivariate mixture density with given
/
marginals under the assumption that the assets were all mutuallly independent random variables (RVs).) (uCellIDs)
/
/
4.) The array of final multivariate mixture component probabilities. (muPrbs)
/
/
5.) The array of mean vectors. Each component of the multivariate mixture density is a multivariate density function which has its own set of means. The first element of this array is the vector of
/
means for the first multivariate component, etc... (muMns)
/
/
6.) The array of VC matrices. Each component in the multivariate mixture density is a multivariate density function with a corresponding VC matrix. Each VC matrix is of dimension (numA)x(numA),
/
where numA = total # of assets. The diagonals of each VC matrix are the corresponding variances for that asset within that component. (muVCs)
/
/
7.) The output destination as a variable reference. Use either "cout" for display to the screen or a valid ofstream output object. (oVar)
128
/
/ Outputs:
/
/
The density supplied to this function is written to the output desination supplied in 2 parts. First, the structure of the density is written using parameters followed by a detailed definition of
/
those parameters. No value is returned at the call.
/*********************************************************************************************************************************************************************************************************/
#include "stdafx.h"
void wrtDens(const string typ, const int nUCmps, const int *uCells, const long double *muPrbs, const Eigen::VectorXd *muMns, const Eigen::MatrixXd *muVCs, ostream& oVar)
{
// Local variables.
//===================
long long w=to_string((long long) nUCmps).size();
string ps="";
int ln=135;
if ("Minimum Sum of Squared Distances (SSD)" == typ)
ln=ln+20;
// Start by writing the structure of the multivariate density, without actual details of the values for the means, variances, covariances, and component probabilities.
//=======================================================================================================================================================================
oVar << endl << string(ln,'=') << endl << "The structure of the multivariate density function for the supplied assets using an initial " << typ << " LP objective is given by: "
<< endl << string(ln,'=') << endl << endl;
for (int v=0; v<nUCmps; ++v)
{
oVar << ps << "p" << setfill('0') << setw(w) << v << "*" << "f(m" << setfill('0') << setw(w) << v << ",V" << setfill('0') << setw(w) << v << ")";
ps=" + ";
if (0 == ((1 + v) % 5))
oVar << endl;
}
// Write out the details for each component. That is, the means, variances, covariances, and component probabilities.
//=======================================================================================================================
oVar << endl << endl << string(112,'=') << endl << "Where values for the multivariate normal pdfs f(m,V) with mean vector m and " << "variance‐covariance matrix V, are:"
<< endl << string(112,'=') << endl;
oVar.setf(ios_base::fixed, ios_base::floatfield); oVar.precision(30);
for (int v=0; v<nUCmps; ++v)
oVar << endl << string(35,'=') << endl << "Component #" << v << " is unique cell #" << uCells[v] << endl << string(35,'=') << endl << "p" << setfill('0') << setw(w)
<< v << " = " << muPrbs[v] << endl << "m" << setfill('0') << setw(w) << v << " = " << endl << muMns[v] << endl << "V" << v << " = " << endl << muVCs[v] << endl
<< "V" << v << setfill('0') << setw(w) << "^(‐1) = " << endl << muVCs[v].inverse() << endl << "‐‐> The rank of V" << v << " is: "
<< (int) Eigen::FullPivLU<Eigen::MatrixXd>(muVCs[v]).rank() << " and the determinant is: " << muVCs[v].determinant() << endl;
}
129
| 5cs.CE
|
Almost Hyperbolic Groups with Almost Finitely
Presented Subgroups
arXiv:1802.01658v1 [math.GR] 5 Feb 2018
Robert Kropholler
February 7, 2018
Abstract
We construct new examples of CAT(0) groups containing non finitely
presented subgroups that are of type F P2 , these CAT(0) groups do not
contain copies of Z3 . We also give a construction of groups which are of
type Fn but not Fn`1 with no free abelian subgroups of rank greater than
r n3 s.
1
Introduction
Subgroups of CAT(0) groups can be shown to exhibit interesting finiteness properties. In [2] there are examples of groups satisfying Fn but not Fn`1 for all n as
well as the first examples of groups of type F P2 that are not finitely presented
answering a question of Brown.
Definition 1.1. A group G satisfies property Fn , if there is a classifying space
KpG, 1q with finite n skeleton.
Definition 1.2. A group G is of type F Pn if there is a partial resolution of the
trivial ZG module,
Pn
...
P1
P0
Z
0
where Pi is a finitely generated projective ZG module.
It can easily be seen that F1 is equivalent to finite generation and F2 is
equivalent to finite presentability. In [2] the groups are kernels of maps φ : G Ñ
Z where G is a right angled Artin group, the finiteness properties of the kernel
depend solely on the defining flag complex for the right angled Artin group.
The construction of groups of type F P2 that are not finitely presented are
contained in right angled Artin groups having free abelian subgroups of rank 3.
Since [2] there have been other constructions of groups with subgroups of type
F P2 not F2 see [?, 10]. In many cases these groups are subgroups of groups G
of type F , in these cases the maximal rank of a free abelian subgroup of G is at
least 3. We construct the first examples of groups of type F containing no free
abelian subgroups of rank 3 containing a subgroup of type F P2 that is not F2 .
Theorem A. There exists a non positively curved space X such that π1 pXq
contains no subgroups isomorphic to Z3 , which contains a subgroup of type F P2
not F2 .
1
The groups constructed in [2] which are of type Fn´1 not Fn all contain a free
abelian subgroup of rank n ´ 1. Brady asked in [?, Question 8.5] whether there
exist groups of type Fn´1 but not Fn which do not contain Z2 ; he notes that
the known examples all contain Zn´2 . While we are not able to find examples
without Z2 , we reduce the bound to a fraction of n.
Theorem B. For every positive integer n, there exists a group of type Fn´1
but not Fn that contains no abelian subgroups of rank greater than r n3 s.
In the future, we would like to extend these theorems, giving examples of
hyperbolic groups with such subgroups.
The author would like to thank Federico Vigolo for kindly helping draw
several of the figures contained within. The author would also like to thank
Martin Bridson and Gareth Wilkes for reading earlier drafts of this paper and
providing helpful and constructive comments.
2
Preliminaries
2.1
Cube Complexes
A cube complex can be constructed by taking a collection of disjoint cubes and
gluing them together by isometries of their faces.
There is a standard way in which cube complexes can be endowed with
metrics. There is a well known criterion of [7] that characterises those cube
complexes that are locally CAT(0).
The precise definition of a cube complex is given in [5, Def. 7.32] as follows.
Definition
2.1. A cube complex X is a quotient of a disjoint union of cubes
Ů
K “ C r0, 1snc by an equivalence relation „. The restrictions χc : r0, 1snc Ñ X
of the natural projection χ : K Ñ X “ K{ „ are required to satisfy:
• for every c P C the map χc is injective;
• if χc pr0, 1snc q X χc1 pr0, 1snc1 q ‰ H, then there is an isometry hc,c1 from a
1
face Tc Ă r0, 1snc onto a face Tc1 Ă r0, 1snc such that χc pxq “ χc1 px1 q if
1
and only if x “ hc,c1 pxq.
Definition 2.2. A metric space is non-positively curved if its metric is locally
CAT(0)
Gromov’s insight allows us to easily check whether a P E complex is nonpositively curved.
Lemma 2.3 (Gromov, [7], p. 120). A P E complex is non-positively curved if
and only if the link of each vertex is a CAT(1) space.
In a cube complex the link of each vertex is a spherical complex built from
all-right speherical simplices (see [5, §5.18]). Gromov realised that an all-right
spherical complex is CAT(1) if and only if it is flag.
Definition 2.4. A complex L is a flag complex if it is simplicial and every set
tv1 , . . . , vn u of pairwise adjacent vertices spans a simplex.
2
Flag complexes are completely determined by their 1-skeleta.
Thus we arrive at the following combinatorial condition for cube complexes.
Lemma 2.5 (Gromov, [7]). A cube complex is non-positively curved if the link
of every vertex is a flag complex.
We wish to limit the rank of free abelian subgroups, in fact we will limit
the largest dimension of an isometrically embedded flat. The following theorem
shows that in the CAT(0) world the maximal dimension of a flat is the same as
the maximal rank of a free abelian subgroup.
Theorem 2.6 (Bridson-Haefliger, [5], II.7). Let X be a compact non positively
curved cube complex, X̃ its universal cover and G “ π1 pXq. If H “ Zn is a
subgroup of G, then there is an isometrically embedded copy of i : Rn ãÑ X̃.
Moreover, the quotient ipRn q{H is an n-torus
Finally we would like to know when a cube complex is hyperbolic. Bridson
shows that the only obstruction is containing an isometrically embedded flat.
Theorem 2.7 (Bridson, [4], Theorem A). Let X be a compact non-positively
curved cube complex and X̃ be its universal cover. X̃ is not hyperbolic if and
only if there exists an isometric embedding i : E2 ãÑ X̃.
2.2
Right angled Artin groups
Right angled Artin groups have been at the centre of a lot of recent study,
particularly because of their interesting subgroup structure [1, 2, 8, 12]. In particular, their subgroups have interesting connections with finiteness properties
of groups, as shown in [2]. For completeness, we recall the basic theory of right
angled Artin groups.
Definition 2.8. Given a flag complex Γ, we define the associated right angled
Artin group (RAAG) AΓ , as the group,
ˇ
@
D
AΓ “ Γp0q ˇ rv, ws “ 1 if rv, ws P Γp1q .
These groups have non-positively curved cube complexes as classifying spaces,
which are unions of tori:
Definition 2.9. Given a flag complex Γ the Salvetti complex SΓ is defined as
follows.
For each vertex vi in Γ, let Sv1i “ S 1 be a copy of the circle cubulated with
1 vertex. For each simplex σ “ rv0 , . . . , vn s of Γ there is an associated torus
Tσ “ Sv10 ˆ ¨ ¨ ¨ ˆ Sv1n . If τ ă σ, then there is a natural inclusion Tτ ãÑ Tσ . Now
define
O
ž
SΓ “
Tσ „
σăΓ
where the equivalence relation „ is generated by the inclusions Tτ ãÑ Tσ .
p0q
There is a map SΓ Ñ pS 1 qΓ sending each circle in SΓ to the respective
p0q
circle in pS 1 qΓ and extending linearly over cubes; this map is an inclusion of
cubical complexes.
It is a standard fact that SΓ has fundamental group AΓ and is a nonpositively curved cube complex. Proofs of the following can be found in [2].
3
Lemma 2.10. π1 pSΓ q “ AΓ .
Lemma 2.11. SΓ is non-positively curved.
2.2.1
A new classifying space
We will construct a new classifying space, which will also be a non-positively
curved cube complex and will be more amenable to taking branched covers. We
will build such a classifying space in the case that the flag complex satisfies the
following condition.
Definition 2.12. A simplicial complex Γ has an n-partite structure if Γ is
contained in the join V1 ˚ ¨ ¨ ¨ ˚ Vn of several discrete sets Vi . In the case n “ 2
we will say bipartite and in the case n “ 3, tripartite.
Any finite simplicial complex can be given an n-partite structure as a subcomplex of a simplex.
We define a cube complex KΓ for an n-partite complex Γ as follows.
i
Let Vi “ tv1i , . . . , vm
u and I “ t1, . . . , nu. For each vertex vki , let Sv1i be a
i
k
copy of S 1 cubulated with two vertices labelled 0 and 1, and two edges evki and
ev0i .
Definition 2.13. Given J P PpIq, J “ ti1 , . . . , il u, we say that σ is a J-simplex
if it is of the form rvki11 , . . . , vkill s.
Remark 1. Every simplex is a J simplex for some unique (possibly empty) J.
For each J-simplex σ “ rvki11 , . . . , vkill s in Γ we associate the following space.
˜
¸ ˜
¸
ź
ź
Sv1 ˆ
ev0i .
Tσ “
vPσ
iRJ
This is the product of a torus and a cube.
š
Let TJ “ tTσ : σ is a J-simplexu and T “ JPPpIq TJ . Given simplices
τ ă σ there is a natural inclusion Tτ ãÑ Tσ .
O
ž
KΓ “
T „.
T PT
Where once again the equivalence relation is generated via the inclusions
Tτ Ñ Tσ . We need to prove the two key lemmata: the fundamental group of
KΓ is AΓ , and KΓ is non-positively curved.
Lemma 2.14. π1 pKΓ q “ AΓ .
Proof. Apply the Seifert-van Kampen Theorem repeatedly.
Lemma 2.15. KΓ is non-positively curved.
Proof. There are 2n vertices in KΓ . Given two vertices v, w there is a cellular
isomorphism KΓ Ñ KΓ which sends v to w. Thus we only need to check the
link of one vertex; we will check the link of 0 “ p0, . . . , 0q. Let L “ Lkp0, KΓ q.
Let Vi “ Vi Y tv0i u.
š
There is a vertex in L for each edge at 0, so Lp0q “
Vi . Two distinct
vertices vki11 and vki22 are connected if one of the following conditions holds:
4
1. k1 “ 0, k2 “ 0,
2. k1 “ 0, i1 ‰ i2 ,
3. i1 ‰ i2 , k2 “ 0,
4. rvki11 , vki22 s is an edge of Γ.
These edges come from the following subcomplexes.
1. TH ,
2. Tvi1 ,
k1
3. Tvi2 ,
k2
4. Trvi1 ,vi2 s .
k1
k2
This tells us that L Ă V1 ˚ ¨ ¨ ¨ ˚ Vn . We now want to prove that L is in
fact a flag complex. Given a set W “ tvki11 , . . . vkill u of pairwise adjacent vertices
in L we want to show that rvki11 , . . . vkill s is also in L. We split W into two sets
(
i
W0 “ vkjj : kj “ 0 and W1 “ W r W0 “ tw1 , . . . , wa u. Since the vertices in
W1 are pairwise adjacent there is a simplex σ Ă Γ spanning them. We see that
the subcomplex Tσ contains an pl ` 1q-cell which fills the required simplex in
L.
It should be noted the natural projections are injective on each closed cube
and not just on the interior.
Remark 2. The complex KΓ requires a choice of n-partite structure. Given
two n-partite structures the complexes are homotopy equivalent but are not
isomorphic as cube complexes. This is shown in Figure 1, together with some
examples of the construction.
Remark 3. Let Θi be a “cage graph” with two vertices that has an edge
śn for each
element of Vi , then the complex KΓ constructed above embeds in i“1 Θi .
This fact will come in useful later, as will the fact that in the case n “ 3
there is an embedded copy of Θ1 > Θ2 > Θ3 in KΓ .
2.2.2
Finiteness properties of subgroups of RAAGs
We will require one theorem on the finiteness properties of subgroups of RAAGs
from [?]. A homomorphism f : AΓ Ñ Z can be defined by putting an integer
label on each vertex and sending the corresponding generator of AΓ to its label.
Definition 2.16. Let f : AΓ Ñ Z be a homomorphism. We denote Γ: ă Γ
the full subcomplex spanned by those vertices with label 0. Let L˚ be the full
subcomplex spanned by vertices not in Γ: .
Theorem 2.17 (Bux-Gonzalez [?], Theorem A). Let f : AΓ Ñ Z be a homomorphism. Then the following are equivalent:
• The kernel of f is of type F Pn , respectively F2 .
• For every, possibly empty, dead simplex σ ă Γ: the complex Γ˚ X Lkpσ, Γq
is homologically pn ´ dimpσq ´ 2q-connected, respectively L˚ is simply connected.
5
KΓ
Γ
1
1
1
2
1
2
1
2
Y
Figure 1: Some examples of KΓ . The labels on vertices exhibit the n-partite
structure. In the last example the shaded regions are identified.
6
2.3
Branched Covers of Cube Complexes
We will take branched covers of cube complexes to get rid of high dimensional
flats. The techniques we will use were developed in [3]. The idea is to take an
appropriate subset which intersects the high dimensional flats.
Definition 2.18. Let X be a non-positively curved cube complex. We say that
Y Ă X is a branching locus if it satisfies the following conditions:
1. Y is a locally convex cubical subcomplex,
2. Lkpc, Xq r Y is connected and non-empty for all cubes c in Y .
The first condition is required to prove that non-positive curvature is preserved when taking branched covers. The second is a reformulation of the the
classical requirement that the branching locus has codimension 2 in the theory
of branched covers of manifolds, ensuring that the trivial branched covering of
X is X.
Definition 2.19. A branched cover X̂ of X over the branching locus Y is the
result of the following process.
1. Take a finite covering X r Y of X r Y .
2. Lift the piecewise Euclidean metric locally and consider the induced path
metric.
3. Take the metric completion X̂ of X r Y .
We require some key results from [3] which allow us to conclude that this
process is natural and that the resulting complex is still a non-positively curved
cube complex.
Lemma 2.20 (Brady, [3], Lemma 5.3). There is a natural surjection b : X̂ Ñ X
and X̂ is a piecewise Euclidean cube complex.
Lemma 2.21 (Brady, [3], Lemma 5.5). If Y is a finite graph, then X̂ is nonpositively curved.
3
Hyperbolisation in dimension 2
This section will be a warm up to our key theorems which are all related to
dimension 3. The lower dimensional case carries a lot of the ideas that will be
used later.
Throughout this section Γ will be a bipartite graph; say Γ Ă V1 ˚ V2 . Let
i
Vi “ tv1i , . . . , vm
u. Let Θi be the graph with two vertices, labelled 0 and 1,
i
i
, all directed from 0 to 1. We noted
and |Vi | ` 1 edges, labelled v0i , v1i , . . . , vm
i
in Remark 3 that the complex KΓ constructed in Section 2.2.1 is a subcomplex
of Θ1 ˆ Θ2 .
Theorem 3.1. Let Γ be a bipartite graph, let AΓ be the associated RAAG and
let KΓ be the classifying space constructed in Section 2.2.1. Then there is a
branched cover RΓ of KΓ which has hyperbolic fundamental group.
7
vk1
vl1
vk1
p0, 0q
ù
vl1
2
vm
vn2
2
vm
vn2
Figure 2: Depiction of the deformation retraction T 2 r tp0, 0qu Ñ S 1 _ S 1 .
There are in fact many branched covers delivering this result. During the
course of the proof we will pick a prime p and as this varies different hyperbolic
branched covers are obtained.
Proof. The branching locus will be one of the 4 vertices of KΓ . Given two
vertices v and w there is a homeomorphism of KΓ which sends v to w. Therefore,
it does not matter which vertex we pick to be our branching locus. We will
choose the vertex 0 “ p0, 0q.
KΓ r t0u deformation retracts onto the graph Θ1 _ Θ2 . This can be seen
as follows. We start with a torus cubulated as in Figure 2. From each torus
we remove the center vertex. The complement deformation retracts onto the
graph depicted in Figure 2. We now identify the edges via their labels, which
will result in Θ1 _ Θ2 .
This argument shows that π1 pKΓ r t0uq is a free group on |V1 | ` |V2 | gen1
erators. We will denote these generators ai “ vi´1
vi1 for i P t1, . . . , |V1 |u and
2
bj “ vj´1
vj2 for j P t1, . . . , |V2 |u.
From the deformation described we get a map Lkp0, KΓ q Ñ Θ1 _ Θ2 . Each
loop of length 4 in Lkp0, KΓ q corresponds to a torus as in Figure 2. Under the
deformation retraction this gets
sent to aśloop in Θ1 _ Θ2 corresponding to the
śk`l
m`n
2 2
vn s “ r i“k`1 ai , j“m`1 bj s.
commutator rvk1 vl1 , vm
Let p ą 1 ` maxp|V1 |, |V2 |q be a prime and Sp be the symmetric group on
p letters. Let λ be a p-cycle in Sp and µ an element which conjugates λ to λr ,
where r is a generator of Zˆ
p . We define our cover using the map,
ρ : π1 pΘ1 _ Θ2 q Ñ Sp ,
ai ÞÑ λ,
bj ÞÑ µ,
taking the cover corresponding to the stabiliser of 1 in Sp .
n
Note that the commutator rλm , µn s “ λpr ´1qm is a p-cycle. This means
that the loops of length 4 in the link have connected preimage in the cover.
We take the completion RΓ of the resulting complex; there is a natural map
b : RΓ Ñ KΓ . The link of the vertex which maps to 0 will contain no cycles of
length ă 5.
8
We now prove that the resulting complex has hyperbolic universal cover.
rΓ is a non-positively curved cube complex. So to prove that
We know that R
it is hyperbolic we just have to show that there are no isometric embeddings
E2 Ñ R˜Γ by Theorem 2.7. If there is such an embedding, then it will contain
at least one square 2-cell. However, each square contains one vertex which is a
lift of 0 and in the link of this vertex there are no loops of length 2π. However,
were the flat plane to be isometrically embedded there would be a loop of length
2π in the link of every vertex on the plane. This contradiction completes the
proof.
We use this theorem along with Morse theoretic ideas from Section 5 to find
more examples of hyperbolic groups with finitely generated subgroups that are
not finitely presentable.
Proposition 3.2. Let Γ be a complete bipartite graph on sets A “ ta1 , a2 , a3 u
and B “ tb1 , b2 , b3 u, fix p satisfying the above hypotheses and let b : RΓ Ñ KΓ
be the branched cover constructed above. Then π1 pRΓ q has a finitely generated
subgroup which is not finitely presentable.
Proof. In this case KΓ “ Θ1 ˆ Θ2 where Θi is as above and has 4 edges. Put an
orientation on each edge of Θi such that there are 2 edges oriented towards each
vertex and 2 edges oriented away from each vertex. Cubulating S 1 with one
vertex and one oriented edge. Define maps hi : Θi Ñ S 1 on each edge by their
orientation. Define h : KΓ Ñ S 1 by hpx1 , x2 q “ h1 px1 q ` h2 px2 q. Precomposing
with the branched covering map b and lifting to universal covers, we obtain a
Morse function f : R̃Γ Ñ R, which is phbq˚ -equivariant.
The ascending and descending links of this Morse functions are the preimages, under b, of the ascending and descending links of h. For the Morse function
h the ascending and descending links are joins of the ascending and descending
links for hi . These will be copies of S 0 , so the ascending and descending links
will be copies of S 1 .
There are now two possibilities, if we look at a vertex in RΓ which does not
map to 0 the ascending and descending links will remain unchanged and will
still be copies of S 1 .
If the vertex in question maps to 0 we will study the ascending link the other
case being identical. The ascending link of 0 is a loop of length 4. Taking a
branched cover cause this loop of length 4 to lengthen but in the preimage it
will still be a copy of S 1 . It follows that the kernel of phbq˚ is finitely generated
but not finitely presentable by 5.9.
Sizeable graphs are used in [9] to give examples of hyperbolic groups with
subgroups which are type F2 not F3 . Here we outline a procedure for producing
examples of sizeable graphs.
Proposition 3.3. Let A and B be sets with partitions A “ A` \ A´ and
B “ B ` \ B ´ where A´ , B ´ are non-empty and |A` |, |B ` | ą 1. Let Γ be
the complete bipartite graph on A and B. Let AΓ be the associated right angled
Artin group, KΓ the classifying space from 2.13 and RΓ the branched covering
constructed in Theorem 3.1. Let v P RΓ be a vertex mapping to p0, 0q. Then
Lkpv, RΓ q is sizeable.
9
Proof. The link of a vertex is a cover Λ of the graph Λ. Λ is the complete
bipartite graph on sets A1 “ A Y ta0 u and B 1 “ B Y tb0 u. Define A1` “ A` and
A1´ “ A1 r A1` , defining B 1` and B 1´ similarly.
The graph Λ has the following properties. It is bipartite as it is the cover of
a bipartite graph and it has no cycles of length 4 since the branching process
was designed to remove these. To check the last property we let A`
be the set of
Λ
´
1`
vertices mapping to A Ă Λ and AΛ the complement of these in the bipartite
structure. We define BΛ` and BΛ´ similarly.
We must prove that ΛpAsΛ Y BΛt q is connected. ΛpA1s , B 1t q can be covered
by finitely many loops c1 , c2 , . . . of length 4 such that cm`1 X Ym
i“1 cm ‰ H for
all m. When taking the branched cover each ci has connected preimage and the
intersection will still be non empty so the resulting union will be connected.
4
Almost hyperbolisation in dimension 3
Notation 4.1. Given a tripartite complex L Ă V1 ˚ V2 ˚ V3 , let Lij be the full
subcomplex spanned by the vertices of Vi Y Vj .
The main theorem of this section is the following.
Theorem 4.2. Let Γ be a tripartite flag complex, AΓ the associated RAAG
and KΓ the classifying space constructed in Section 2.2.1. Then there exists a
branched cover X of KΓ , such that π1 pXq contains no subgroups isomorphic to
Z3 .
Proof. Our branching locus will be
B “ pΘ1 ˆ t0u ˆ t1uq Y pt1u ˆ Θ2 ˆ t0uq Y pt0u ˆ t1u ˆ Θ3 q,
where Θi is the graph with 2 vertices, 0 and 1, and |Vi | ` 1 edges, each directed
from 0 to 1. We have three maps
p1 : KΓ r B Ñ KΓ12 r tp0, 1qu,
p2 : KΓ r B Ñ KΓ23 r tp0, 1qu,
p3 : KΓ r B Ñ KΓ31 r tp0, 1qu,
which are the restrictions of the projections Θ1 ˆ Θ2 ˆ Θ3 Ñ Θi ˆ Θj . The
maps from Section 3 give us three maps
π1 pKΓ12 r tp0, 1quq Ñ Sq12
π1 pKΓ23 r tp0, 1quq Ñ Sq23
π1 pKΓ31 r tp0, 1quq Ñ Sq31 .
Here, qij are the primes picked in the process of taking a branched cover in
Theorem 3.1. Let q “ q12 q23 q31 . We can combine these permutation representations with the projection maps above to get a map π1 pKΓ r Bq Ñ Sq , which
defines a q-fold cover of KΓ r B by taking the subgroup corresponding to the
stabiliser of 1 in Sq . We complete this cover to get our branched cover X.
Let ti, j, ku “ t1, 2, 3u. The maps p1 , p2 , p3 are retractions to see this consider
the natural map
" *
1
ãÑ KΓij ˆ ev0k ãÑ KΓ .
KΓij ãÑ KΓij ˆ
2
10
v02
v01
vi1
vj2
Figure 3: A loop corresponding to the commutator rvi1 , vj2 s.
It follows that π1 pKΓ r Bq Ñ π1 pKΓij r tp0, 1quq is surjective. We will now
consider what the link of a vertex in the branched cover is. We will restrict our
attention to a vertex mapping to p0, 1, 0q all the other cases are similar.
We will consider the image of the link of p0, 1, 0q under the three maps
p1 , p2 , p3 . In the image of the map p1 , this link is sent surjectively onto the
link of p0, 1q in KΓ12 . By Section 3 we know that KΓ12 r tp0, 1qu deformation
retracts onto the graph Θ1 _ Θ2 . Loops of length 4 are sent to commutators of
the form rvi1 v 1j , vk2 v 2l s. In the map π1 pKΓ12 r tp0, 1quq Ñ Sq this commutator is
sent to a q12 -cycle.
We must now consider the image under the maps p2 , p3 . These maps send the
link to a disjoint union of contractible subsets, so the maps π1 pKΓ r Bq Ñ Sq23
and π1 pKΓ r Bq Ñ Sq31 send the image of the fundamental group of the link to
the identity.
From this we can see that, in the cover of KΓ r B corresponding to the
stabiliser of 1 in Sq , the preimage of one of the loops of length 4 depicted in
Figure 3 will have q23 q31 components, and each component is a loop of length
4q12 . We will now prove that there are no isometrically embedded planes of
dimension ą 2. This, combined with Theorem 2.6, will complete the proof of
the theorem.
Since the resulting cube complex has cubes of dimension at most 3, we can
see that the dimension of an isometrically embedded flat plane is at most 3.
If such a copy E of E3 were isometrically embedded in X̃, it would contain at
least one cube and would in fact be a cubical embedding of the flat. For each
vertex contained in the flat, the link would contain a subcomplex isomorphic to
an octahedron. Let B̃ be the lift of the branching locus B. We can see how this
intersects each cube in X̃ by Figure 4. As such any 3-flat will intersect a vertex
on B̃.
Let x be a vertex on B̃ X E. Then Lkpx, X̃q has a tripartite structure.
If there is an octahedron in this complex it has a tripartite structure of the
form S 0 ˚ S 0 ˚ S 0 . One of the copies of S 0 will be contained in the vertices
corresponding to B̃. The other 4 vertices form a loop of length 4 in the bipartite
graph defined by the edges not in B̃. However, we constructed the branched
11
Figure 4: Intersection pattern of L̃ on a cube in X̃.
cover so that this graph has no cycles of length ă 6.
5
Bestvina-Brady Morse theory
While Bestvina-Brady Morse theory is defined in the more general setting of
affine cell complexes, in this instance we shall only need it for non-positively
curved cube complexes.
For the remainder of this section, let X be a CAT(0) cube complex and let
G be a group which acts freely, cellularly, properly and cocompactly on X. Let
φ : G Ñ Z be a homomorphism and let Z act on R by translations.
Recall that χc is the characteristic map of the cube c.
Definition 5.1. We say that a function f : X Ñ R is a φ-equivariant Morse
function if it satisfies the following 3 conditions.
• For every cube c Ă X of dimension n, the map f χc : r0, 1sn Ñ R extends
to an affine map Rn Ñ R and f χc : r0, 1sn Ñ R is constant if and only if
n “ 0.
• The image of the 0-skeleton of X is discrete in R.
• f is φ-equivariant, i.e. f pg ¨ xq “ φpgq ¨ f pxq.
We will consider the level sets of our function, which we will denote as
follows.
Definition 5.2. For a non-empty closed subset I Ă R we denote by XI the
preimage of I. We also use Xt to denote the preimage of t P R.
The kernel H of φ acts on the cube complex X in a manner preserving each
level set XI . Moreover, it acts properly and cocompactly on the level sets. We
will use the topological properties of the level sets to gain information about
the finiteness properties of the group. We will need to examine how they vary
as we pass to larger level sets.
12
Theorem 5.3 (Bestvina-Brady, [2], Lemma 2.3). If I Ă I 1 Ă R are closed
intervals and XI 1 r XI contains no vertices of X, then the inclusion XI ãÑ XI 1
is a homotopy equivalence.
If XI 1 r XI contains vertices of X, then the topological properties of XI 1 can
be very different from those of XI . This difference is encoded in the ascending
and descending links.
Definition 5.4.ŤThe ascending link of a vertex is
LkÒ pv, Xq “ tLkpw, cq | χc pwq “ v and w is a minimum of f χc u Ă Lkpv, Xq
The descending
Ť link of a vertex is
LkÓ pv, Xq “ tLkpw, cq | χc pwq “ v and w is a maximum of f χc u Ă Lkpv, Xq
Theorem 5.5 (Bestvina-Brady, [2], Lemma 2.5). Let f be a Morse function.
Suppose that I Ă I 1 Ă R are connected, closed and min I “ `min I˘1 (resp.
max I “ max I 1 q, and that I 1 r I contains only one point r of f X p0q . Then
XI 1 is homotopy equivalent to the space obtained from XI by coning off the
descending (resp. ascending) links of v for each v P f ´1 prq.
We can now deduce a lot about the topology of the level and sub-level sets.
We know how they change as we pass to larger intervals and so we have the
following.
Corollary 5.6 (Bestvina-Brady, [2], Corollary 2.6). Let I, I 1 be as above.
1. If each ascending and descending link is homologically pn ´ 1q-connected,
then the inclusion XI ãÑ XI 1 induces an isomorphism on Hi for i ď n ´ 1
and is surjective for i “ n.
2. If the ascending and descending links are connected, then the inclusion
XI ãÑ XI 1 induces a surjection on π1 .
3. If the ascending and descending links are simply connected, then the inclusion XI ãÑ XI 1 induces an isomorphism on π1 .
Knowing that the direct limit of this system is a contractible space allows
us to compute the finiteness properties of the kernel of φ.
Theorem 5.7 (Bestvina-Brady, [2], Theorem 4.1). Let f : X Ñ R be a φequivariant Morse function and let H “ kerpφq. If all ascending and descending
links are simply connected, then H is finitely presented (i.e. is of type F2 ).
We would also like to have conditions which will allow us to deduce that H
does not satisfy certain other finiteness properties. A well known result in this
direction is:
Proposition 5.8 (Brown, [6], p. 193). Let H be a group acting freely, properly,
r i pX, Zq “
cellularly and cocompactly on a cell complex X. Assume further that H
r
0 for 0 ď i ď n ´ 1 and that Hn pX, Zq is not finitely generated as a ZH-module.
Then H is of type F Pn but not F Pn`1 .
In [3], the above result was used to prove that a certain group is not of type
F P3 . However in our theorems not all the links will satisfy the assumptions
of [3, Theorem 4.7] and we require the following.
13
Theorem 5.9 (Kropholler, [9], Theorem 2.23). Let f : X Ñ R be a φ-equivariant
Morse function and let H “ kerpφq. Suppose that for all vertices v the reduced
homology of LkÒ pvq and LkÓ pvq is 0 in dimensions 0, . . . , n´1 and n`1. Further
`
˘
`
˘
r n LkÒ pv 1 q ‰ 0 or H
r n LkÓ pv 1 q ‰ 0
assume that there is a vertex v 1 such that H
(possibly both). Then H is of type F Pn but not of type F Pn`1 .
Finally, we prove a key theorem regarding ascending and descending links.
Theorem 5.10. For i “ 1, 2 let Xi be a CAT(0) cube complex, Gi a group
acting freely, properly and cocompactly on Xi , φi : Gi Ñ Z a surjective homomorphism and fi : Xi Ñ R a φi -equivariant Morse function. Then, there is a
Morse function f : X1 ˆ X2 Ñ R such that
`
˘
LkÒ pv1 , v2 q, X1 ˆ X2 “ LkÒ pv1 , X1 q ˚ LkÒ pv2 , X2 q
and
`
˘
LkÓ pv1 , v2 q, X1 ˆ X2 “ LkÓ pv1 , X1 q ˚ LkÓ pv2 , X2 q
.
We will prove the staement for ascending links, the proof for descending
links is the same. The proof really relies on the following key lemma.
Lemma 5.11. Let X be a CAT(0) cube complex with Morse function f . Then
LkÒ pv, Xq is the full subcomplex of Lkpv, Xq spanned by the vertices in LkÒ pv, Xq.
At first sight this does not appear to be an improvement, however, it allows
for simple calculation of the ascending and descending links once the link of a
vertex is known.
Proof. Let v1 , . . . , vn be n pairwise adjacent vertices in LkÒ pv, Xq, proving that
the simplex rv1 , . . . , vn s is in LkÒ pv, Xq will prove the claim.
Let ei be the edge in X corresponding to vi . We must prove that v is a
minimum for f restricted to c “ e1 ˆ ¨ ¨ ¨ ˆ en . We note that since f extends
to an affine map, c is foliated by level sets f ´1 ptq for t P R. Each of these
level sets corresponds to a linear subspace of dimension n ´ 1 not containing
any subcubes of dimension ą 0. When intersected with the cube c there are
exactly two subspaces where the intersection is a single vertex. One of these
corresponds to the minimum of f |c and the other to the maximum of f |c . One
of these must be at the vertex mapping to v and this is the minimum.
Proof of Theorem 5.10. We define a Morse function f : X1 ˆ X2 Ñ R by
f px1 , x2 q “ f1 px1 q ` f2 px2 q. This satisfies all the conditions of being a Morse
function and is φ-equivariant with respect to the map φpg1 , g2 q “ φ1 pg1 q`φ2 pg2 q.
The link of a vertex pv1 , v2 q in X1 ˆ X2 is Lkpv1 , X1 q ˚ Lkpv2 , X2 q. It is easy
to see that the ascending link is the full subcomplex spanned by LkÒ pv1 , X1 q \
LkÒ pv2 , X2 q Ă Lkpv1 , X1 q˚Lkpv2 , X2 q, which is equal to LkÒ pv1 , X1 q˚LkÒ pv2 , X2 q.
14
5.1
Groups of type Fn´1 but not Fn
In [?], Question 8.5 Brady asks whether there exist groups of type Fn´1 but
not Fn which do not contain Z2 ; he notes that the known examples all contain
Zn´2 . While we are not able to find examples without Z2 , we can drastically
reduce the rank of a free abelian subgroup as shown by the following.
Theorem B. For every positive integer n, there exists a group of type Fn´1
but not Fn that contains no abelian subgroups of rank greater than r n3 s.
Proof. For our general construction, we require 4-tuples pGi , Xi , φi , fi q for i “
1, 2, 3, where:
• Gi is a hyperbolic group,
• Xi is a CAT(0) cube complex with a free cocompact Gi action,
• φi : Gi Ñ Z is a surjective homomorphism,
• fi is a φi -equivariant Morse function.
We also require that all ascending and descending links of fi are pi ´ 2qconnected but at least one is not pi ´ 1q-connected.
The existence of such a Morse function would show by Theorem 5.9 that Gi
has a subgroup of type Fi´1 not Fi , namely kerpφi q.
Let G1 be a free group of rank 2, X1 a 4-valent tree (the Cayley graph of Gi
with respect to generators a and b), φ1 the exponent sum homomorphism with
respect to a and b, and f1 the map that is linear on edges and whose restriction
to the vertices of X1 is φ1 .
Let pG2 , X2 , φ2 , f2 q be the group, classifying space, homomorphism and
Morse function from Proposition 3.2.
Let pG3 , X3 , φ3 , f3 q be the group, classifying space, homomorphism and
Morse function from [3]. We could also use the examples of groups from [9, 11].
An important point to note is that for i “ 1, 2, 3, the ascending and descending links for pGi , Xi , φi , fi q are isomorphic to S i´1 .
Now let n be an integer, let l “ t n3 u, and let pGl , Xl , φl , fl q be the 4-tuple
with
l
l
ź
ź
Gl “
G3 , Xl “
X3 ,
i“1
φl “
l
ÿ
i“1
φ3 , fl “
i“1
l
ÿ
f3 .
i“1
By Theorem 5.10 the ascending and descending links of fl are ˚li“1 S 2 “ S 3l´1
which is p3l ´ 2q-connected but not p3l ´ 1q-connected. If n ” 0 mod 3, then
3l “ n, and we define G “ Gl . If not, let m be the residue of n mod 3 and
consider the 4-tuple pG “ Gl ˆ Gm , X “ Xl ˆ Xm , φ “ φl ` φm , f “ fl ` fm q.
The φ-equivariant Morse function f on the cube complex X has ascending and
descending links that are all copies of S n´1 . In all cases, Kerpφq is of type Fn´1
not Fn by Theorem 5.9.
The group G contains no free abelian subgroups of rank greater than r n3 s
since Gi as a hyperbolic group contains no copy of Z2 .
15
6
Subgroups of type F P2 that cannot be finitely
presented
We will now apply our branched covering technique with carefully chosen flag
complexes Γ to prove the following.
Theorem A. There exists a non positively curved space X such that π1 pXq
contains no subgroups isomorphic to Z3 , which contains a subgroup of type F P2
not F2 .
We will do this using the following steps.
1. Start with a connected 2-dimensional tripartite flag complex L such that
π1 pLq is a perfect group with the the link of every vertex connected and not
a point. Take an auxiliary complex Γ “ RpLq. Then build the complex
KΓ in Section 2.2.1.
2. Define a function f : KΓ Ñ S 1 which lifts to a Morse function on universal covers. Examine the ascending and descending links of this Morse
function.
3. Take a branched covering of KΓ as in Section 4 to get a complex X with an
associated Morse function. Examine the ascending and descending links
of this Morse function. This will show that the kernel of f˚ is of type F P2 .
4. Prove that the kernel of f˚ is not finitely presented.
6.1
The complex RpLq
The construction of this complex is required, the key point of this complex is that
it satisfies Proposition 6.4. This is required to make sure that the fundamental
group of the links is not changed in the branching process.
Firstly, we prove that none of the assumptions from the first step are restrictive. We can realise any finitely presented group as the fundamental group of
a finite connected 2-dimensional simplicial complex. It is also well known that
the barycentric subdivision of a 2-dimensional simplicial complex is flag and we
can put an obvious tripartite structure, on the barycentric subdivision, labelling
vertices by the dimension of the corresponding cell. So we can realise any group
as the fundamental group of a connected 2-dimensional tripartite flag complex.
Given a connected 2-dimensional tripartite flag complex there is a homotopy equivalent complex of the same form such that the link of every vertex is
connected and not a point.
To see this, first note that if there is a vertex where the link is just a single
point, then we can contract this edge without changing the homotopy type of
the complex. Next, note that if there is a vertex x with disconnected link,
then we perform the following procedure: pick two vertices v, w in two different
components of the link, add an extra vertex y and connect it to v, w and x
while also adding two triangles rv, y, xs and ru, y, xs. The result is that we have
reduced the number of components of the link at x, without adding any extra
components to the link of v or w, and the link of y is connected. We have also
not changed the homotopy type, since we have added a contractible space glued
16
along a contractible subspace. Repeating this procedure, we can make sure that
the link of every vertex is connected.
Definition 6.1. For a simplicial complex L the octahedralisation SpLq is defined
as follows. For each vertex v of L, let Sv0 “ tv ` , v ´ u be a copy of S 0 . For every
simplex σ of L take Sσ “ ˚vPσ Sv0 . If τ ă σ, then there is a natural map
Sτ Ñ Sσ . SpLq “ YσăL Sσ { „, where the equivalence relations „ is generated
by the inclusions Sτ Ñ Sσ .
The complex SpLq can also be seen as the link of the vertex in the Salvetti
complex for the RAAG defined by L. If L has an n-partite structure then there
is a natural n-partite structure on SpLq. Let L be contained in V1 ˚ ¨ ¨ ¨ ˚ Vn .
Then SpLq is contained in SpV1 q ˚ ¨ ¨ ¨ ˚ SpVn q.
Remark 4. The map defined by Sv0 Ñ tvu extends to a retraction of SpLq to L
(not a deformation retraction).
In particular, if π1 pLq ‰ 0, then π1 pSpLqq ‰ 0.
It is proved in [2] that if L is a flag complex, then SpLq is a flag complex.
Lemma 6.2. Assume L is a connected simplicial complex and Lkpv, Lq is connected and not equal to a point for all vertices v P L. If H1 pLq “ 0, then
H1 pSpLqq “ 0.
Proof. Let L` be the full subcomplex of SpLq spanned by the set tv ` : v P Lu,
and let L´ be defined similarly. Let N ` be the interior of the star of L` and
N ´ the interior of the star of L´ . It is clear that SpLq is contained in N ´ YN ` .
We can also see that N ` X N ´ is the union of the open simplices which are
contained in neither L` nor L´ . Here we are using the fact that N ` is homotopy
equivalent to L and similarly N ´ is homotopy equivalent to L.
By considering the Mayer-Vietoris sequence for N ´ and N ` we see that
H1 pSpLqq is isomorphic to H̃0 pN ` X N ´ q, so we must prove that N ` X N ´ is
connected.
Let x, y be points of N ` XN ´ . We can always connect x and y to open edges
contained in N ` X N ´ ; label these edges ex and ey . Let vx be the end point of
ex in L` and wx the end point in L´ ; define vertices vy and wy similarly. Let
v “ pvx “ v0 , v1 , . . . , vn “ vy q be a sequence of vertices in L` corresponding to
a geodesic in Lp1q from vx to vy .
The vertices wx and wy have corresponding vertices in L` , these are adjacent
to vx and vy respectively. We split into 4 cases by whether these vertices are on
the geodesic v. We will define a path P in each of these four cases.
1. If wx “ v1 and wy “ vn´1 , then let P “ v0 , v1 , . . . , vn .
2. If wx “ v1 and wy ‰ vn´1 , then let P “ v0 , v1 , . . . , vn , wy .
3. If wx ‰ v1 and wy “ vn´1 , then let P “ wx , v0 , v1 , . . . , vn .
4. If wx ‰ v1 and wy ‰ vn´1 , then let P “ wx , v0 , v1 , . . . , vn , wx .
We will now describe how to get a sequence of edges ex “ e0 , . . . , em “ ey
such that ei , ei`1 are on a 2-simplex in L. The corresponding sequence of edges
in N ` X N ´ will give a path from ex to ey . Thus completing the proof of the
lemma.
17
b1
b3
b2
a1
a3
a2
Figure 5: The key idea from Lemma 6.2.
A path in the link of a vertex v can be viewed as a sequence of edges in L
such that adjacent edges are on a 2-simplex.
We will relabel the path P to a1 , b1 , . . . . Let Ai be a path in the link of bi from
ai to ai`1 . Let Bi be a path in the link of ai from bi to bi`1 . This can be done
since the link of every vertex is connected. The sequence of edges A1 , B2 , A2 , . . .
defines the sequence of edges we require. The key idea is encapsulated in Figure
5, where the curved arcs correspond to the paths Ai , Bi .
We can now define the complex RpLq.
1. Let L be a 2-dimensional tripartite flag complex with π1 pLq perfect, such
that the link of every vertex in L is connected and not a point. Label the
sets of vertices from the tripartite structure V0 , V1 , V2 .
2. Construct SpLq as above. This is a 2-dimensional tripartite flag complex
with π1 pLq perfect.
3. Add 3 extra vertices v 0 , v 1 , v 2 which are of type 0, 1, 2 respectively. Connect v i to all the vertices not of type i. Define RpLq to be the flag completion of the resulting complex.
Take a simplicial complex L as above and let Γ “ RpLq. We can now
construct the cubical complex KΓ from Section 2.2.1.
6.2
The Morse function f
As noted in Remark 3, we can view KΓ as a subcomplex of Θ1 ˆ Θ2 ˆ Θ3 , where
Θi is a graph that has two vertices 0, 1 and has edges labelled by the vertices of
Γ as well as one extra edge labelled v0i , such that each edge runs from 0 to 1.
We define a Morse function on the product by putting an orientation on each
edge of Θi as follows: if it is an edge corresponding to a vertex of SpLq Ă Γ we
orient it towards 1, while for the vertices v i and v0i we orient towards 0. Now
put an orientation on S 1 and map each graph by its orientation; we then extend
linearly across cubes. Restricting this map to KΓ we get a map g, and by lifting
to universal covers we get a Morse function f : K̃Γ Ñ R which is g˚ -equivariant.
Let the vertices of type i in SpLq Ă Γ be the set Vi` and let Vi´ “ tv i , v0i u “:
0
Si .
18
a4
Vertex
p0, 0, 0q
p0, 0, 1q
p0, 1, 0q
p0, 1, 1q
p1, 0, 0q
p1, 0, 1q
p1, 1, 0q
p1, 1, 1q
LkÒ
ΓpV1` Y V2` Y V3` q “ SpLq
ΓpV1` Y V2` q ˚ S30
ΓpV1` Y V3` q ˚ S20
V1` ˚ S20 ˚ S30
ΓpV2` Y V3` q ˚ S10
V2` ˚ S10 ˚ S30
V3` ˚ S10 ˚ S20
2
S “ S10 ˚ S20 ˚ S30
LkÓ
S 2 “ S10 ˚ S20 ˚ S30
V3` ˚ S10 ˚ S20
V2` ˚ S10 ˚ S30
ΓpV2` Y V3` q ˚ S10
V1` ˚ S20 ˚ S30
ΓpV1` Y V3` q ˚ S20
ΓpV1` Y V2` q ˚ S30
`
ΓpV1 Y V2` Y V3` q “ SpLq
Table 1: The ascending and descending links of the Morse function f : KΓ Ñ S 1 .
The ascending and descending links of this Morse function are given in Table
1.
Notation 6.3. Given a simplicial complex Λ and a subset S of the vertices of
Λ, ΛpSq denotes the full subcomplex of Λ spanned by the vertices in S.
Proposition 6.4. Given a complex of the form Vi` ˚Vj´ ˚Sk0 or ΓpVi` YVj` q˚Sk0
there is an ordering v1 , v2 , . . . , vq on the set Vjs such that Stpvm qXpYlăm Stpvl qq
is connected and π1 pStpvm q X pYlăm Stpvl qqq ‰ 0 for 1 ă m ď q.
Proof. In the case Vjs “ Vj´ “ Sj0 we have 2 vertices v1 and v2 and Stpv1 q X
Stpv2 q “ Sk0 ˚ Vi` , which is connected but not simply connected. So either
ordering will do.
In the case Vjs “ Vj` , let Vj “ tw1 , . . . , wn u and Vj` “ tw1` , w1´ , w2` , . . . , wn´ u.
The subgraph Z “ LpVi Y Vj q is connected since L is connected and the
link of every vertex is connected and not a point. We can thus assume that
Stpwl , Zq X pYmăl Stpwm qq ‰ H.
Noting that Stpv s , SpLqq “ CpSpLkpv, Lqqq. We can see that Stpv s , SpLqq X
Stpwt , SpLqq “ SpStpv, LqXStpw, Lqq. Thus ordering Vj` “ tw1` , w1´ , w2` , . . . , wn´ u
we can see that Vi` X Stpvl , SpLqq X pYmăl Stpvm , SpLqqq contains at least two
points.
Noting that
¯¯
´
´
¯ ´ ď
ď
Stpvl , SpLqqXp
Stpvm , SpLqqq “ Vi` X Stpvl , SpLqq X
Stpvm , SpLqq ˚Sk0
măl
măl
we can see that this is connected but not simply connected.
Remark 5. In the above proof we are gluing contractible complexes along connected complexes. This shows that all the complexes in the statement of Proposition 6.4 are simply connected.
Remark 6. At each stage in the above proof Stpvl , SpLqq X pYmăl Stpvm , SpLqqq
could be covered by cycles of length 4, since it is the join of a discrete set and
a copy of S 0 .
19
6.3
Almost hyperbolisation and the Morse function h :
X Ñ S1
We use the almost hyperbolisation technique from Section 4 to get a branched
cover X of KΓ , recall that there is a natural length preserving map b : X Ñ KΓ .
We define a function h “ g ˝ b : X Ñ S 1 , lifting to universal covers we get a
Morse function. In what follows G is the fundamental group of X which does
not contain any copies of Z3 . In what follows, H :“ kerph˚ : G Ñ Zq.
It is worth noting that in the almost hyperbolisation procedure we ensured
that loops of length 4 in the link of a vertex have connected preimage.
6.3.1
Ascending and descending links of h
Recall that KΓ Ă Θ1 ˆ Θ2 Θ3 . We distinguish between 4 types of vertices in X
and label them as follows:
Vertices of type A are those which map to p0, 0, 0q or p1, 1, 1q.
Vertices of type B are those which map to p0, 0, 1q or p1, 0, 1q.
Vertices of type C are those which map to p1, 0, 0q or p1, 1, 0q.
Vertices of type D are those which map to p0, 1, 0q or p0, 1, 1q.
For a vertex in X the ascending (descending) link is the preimage of the
ascending (descending) link of the corresponding vertex in KΓ .
Type A vertices are disjoint from the branching locus, so a small neighbourhood of each lifts to X and the ascending and descending links are isomorphic
to those of the corresponding vertex of KΓ . We claim that for vertices not of
type A the ascending and descending links are simply connected. We will prove
this in the case of a vertex of type B and the ascending link; the other cases are
similar.
A vertex x of type B is on a lift of Θ1 . We may assume that x maps to
p0, 0, 1q. Now, LkÒ p0, 0, 1q “ ΓpV1` Y V2` q ˚ S30 . Let us consider the preimage of
LkÒ p0, 0, 1q in Lkpx, Xq. To envisage this, we start by removing the vertices in
V1 and taking the covering of the remaining space coming from the derivative
map b˚ of b, then add back the vertices of V1 . By Remark 5 we can cover
ΓpV1` Y V2` q ˚ S30 by Stpvq as v runs over vertices in V1` . We noted in Remark 6
that we can construct this cover in such a way that each Stpvm q X pYlăm Stpvl qq
is connected and covered by loops of length 4 with non-empty intersection.
Specifically, if S41 and S411 are 2 loops of length 4 in Stpvm q X pYlăm Stpvl qq, then
S41 X S411 Ą S30 .
In the procedure of passing to the branched covering b : X Ñ KΓ . Associated
to each vertex of X there is a derivative map bLkpvq : Lkpv, Xq Ñ Lkpbpvq, KΓ q.
Each of the 4 cycles above has connected preimage under the map bLkpvq , therefore the preimage of Stpvm q X pYlăm Stpvl qq is connected. Upon taking the
completion, we replace the vertices in V1` ; this corresponds to coning off the
lifts of their links. Thus we see LkÒ pxq is made from a sequence of contractible
spaces glued along connected subspaces, and so is simply connected.
6.4
Proof that the kernel is not finitely presented
To prove that H is not finitely presented we need the following lemma.
Lemma 6.5. π1 pXI q ‰ 0 for all compact intervals I Ă R
20
ĂΓ . In our case Γ: “ S 2 and
Proof. For the purposes of this proof let Y “ K
Γ˚ “ RpLq. By Theorem 2.17 we know that the kernel of g˚ is not finitely
presented. Since the kernel of g˚ acts cocompactly on the level set YI we can
see that YI is not simply connected.
Let γ : S 1 Ñ YI be a non-trivial loop. Then there is a larger interval J 1 such
that γ is trivial in YJ 1 . Let J 1 “ ra, bs. We can assume that γ is non-trivial
in J “ ra ` , bs or J “ ra, b ´ s for all ą 0; assume the latter. There is
a sequence of vertices v1 , v2 , . . . such that YJ 1 “ YJ Y CpLkÒ pvi qq. There is
an integer m such that γ is trivial in YJ Yiďm CpLkÒ pvi qq but not trivial in
xJ .
YJ Yiăm CpLkÒ pvi qq “ Y
Since adding CpLkÒ pxm qq changes the fundamental group (γ becomes trivxJ which is also contained in LkÒ pxm q.
ial), we can find a non-trivial loop γ̂ in Y
We also know that it is contained in YJ since LkÒ pvi q X LkÒ pvj q “ H whenever
i ‰ j, as the restriction of an affine map to a cube can only have one maximum
and one minimum or it is constant on a subcube.
Since adding CpLkÒ pxm qq changes π1 , we see that π1 pLkÒ pxm qq ‰ 0, so xm is
a vertex mapping to p0, 0, 0q and γ̂ bounds a disc in Y which does not intersect
L̃. It follows that this disc lifts to X̃ under the branched covering b̃ coming from
the following commutative diagram:
X̃
b̃
Y
X
b
KΓ
The boundary of this lifted disc is in XJ ; call this loop µ. If µ bounded
a disc in XJ , then we could map this to YJ via b̃, but this would imply that
b̃pµq “ γ̂ bounds a disc in YJ , which it does not. Thus µ is non-trivial in
π1 pXJ q ‰ 1. Since π1 pXI q Ñ π1 pXJ q is surjective we deduce by Theorem 5.7
that π1 pXI q ‰ 1.
Theorem 6.6. H “ kerpg˚ q is not finitely presentable.
Proof. Assume that H is finitely presented.
H acts cocompactly on XI so we can add finitely many 2-cells to the quotient
to gain fundamental group H. Taking a universal cover of the space obtained in
this way we arrive at XI with finitely many H-orbits of 2-cells attached, which
is simply connected. In other words, there are finitely many H-orbits of loops
which generate π1 pXI q.
The direct limit of all the XI is the space X which is CAT(0) and in particular
contractible. We can pass to a larger interval such that the H-orbits of loops
which generate π1 pXI q are trivial. In other words, the map π1 pXI q Ñ π1 pXJ q
is trivial. But it is also surjective by Theorem 5.7 and because we have assumed
that all the ascending and descending links are connected. Thus π1 pXJ q is
trivial, however we know this not to be the case by Lemma 6.5.
References
[1] I. Agol. The virtual Haken conjecture. Doc. Math., 18:1045–1087, 2013.
With an appendix by Agol, Daniel Groves, and Jason Manning.
21
[2] M. Bestvina and N. Brady. Morse theory and finiteness properties of groups.
Invent. Math., 129(3):445–470, 1997.
[3] N. Brady. Branched Coverings of Cubical Complexes and Subgroups of
Hyperbolic Groups. J. Lond. Math. Soc., 60(2):461–480, 1999.
[4] M. R. Bridson. On the existence of flat planes in spaces of nonpositive
curvature. Proceedings of the American Mathematical Society, 123(1):223–
223, January 1995.
[5] M. R. Bridson and A. Haefliger. Metric spaces of non-positive curvature,
volume 319 of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1999.
[6] K. S. Brown. Cohomology of Groups, volume 87 of Graduate Texts in
Mathematics. Springer New York, New York, NY, 1982.
[7] M. Gromov. Hyperbolic Groups. In S. S. Chern, I. Kaplansky, C. C. Moore,
I. M. Singer, and S. M. Gersten, editors, Essays in Group Theory, volume 8,
pages 75–263. Springer New York, 1987.
[8] F. Haglund and D. T. Wise. Special cube complexes. Geom. Funct. Anal.,
17(5):1551–1620, 2008.
[9] R. Kropholler. Hyperbolic groups with almost finitely presented subgroups.
In Preparation, 2015.
[10] Ian J. Leary. Uncountably many groups of type $FP$. arXiv:1512.06609
[math], December 2015. arXiv: 1512.06609.
[11] Yash Lodha. A hyperbolic group with a finitely presented subgroup that is
not of type FP3, page 67?81. London Mathematical Society Lecture Note
Series. Cambridge University Press, 2017.
[12] D. T. Wise. The Structure of Groups with a Quasiconvex Hierarchy. 2011.
22
| 4math.GR
|
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
Automated Classification of L/R Hand Movement
EEG Signals using Advanced Feature Extraction and
Machine Learning
Mohammad H. Alomari, Aya Samaha, and Khaled AlKamha
Applied Science University
Amman, Jordan
Abstract— In this paper, we propose an automated computer
platform for the purpose of classifying Electroencephalography
(EEG) signals associated with left and right hand movements
using a hybrid system that uses advanced feature extraction
techniques and machine learning algorithms. It is known that
EEG represents the brain activity by the electrical voltage
fluctuations along the scalp, and Brain-Computer Interface (BCI)
is a device that enables the use of the brain’s neural activity to
communicate with others or to control machines, artificial limbs,
or robots without direct physical movements. In our research
work, we aspired to find the best feature extraction method that
enables the differentiation between left and right executed fist
movements through various classification algorithms. The EEG
dataset used in this research was created and contributed to
PhysioNet by the developers of the BCI2000 instrumentation
system. Data was preprocessed using the EEGLAB MATLAB
toolbox and artifacts removal was done using AAR. Data was
epoched on the basis of Event-Related (De) Synchronization
(ERD/ERS) and movement-related cortical potentials (MRCP)
features. Mu/beta rhythms were isolated for the ERD/ERS
analysis and delta rhythms were isolated for the MRCP analysis.
The Independent Component Analysis (ICA) spatial filter was
applied on related channels for noise reduction and isolation of
both artifactually and neutrally generated EEG sources. The
final feature vector included the ERD, ERS, and MRCP features
in addition to the mean, power and energy of the activations of
the resulting Independent Components (ICs) of the epoched
feature datasets. The datasets were inputted into two machinelearning algorithms: Neural Networks (NNs) and Support Vector
Machines (SVMs). Intensive experiments were carried out and
optimum classification performances of 89.8 and 97.1 were
obtained using NN and SVM, respectively. This research shows
that this method of feature extraction holds some promise for the
classification of various pairs of motor movements, which can be
used in a BCI context to mentally control a computer or machine.
Keywords—EEG; BCI; ICA; MRCP; ERD/ERS; machine
learning; NN; SVM
I.
INTRODUCTION
The importance of understanding brain waves is increasing
with the ongoing growth in the Brain-Computer Interface
(BCI) field, and as computerized systems are becoming one of
the main tools for making people’s lives easier, BCI or BrainMachine Interface (BMI) has become an attractive field of
research and applications, BCI is a device that enables the use
of the brain’s neural activity to communicate with others or to
control machines, artificial limbs, or robots without direct
physical movements [1-4].
The term “Electroencephalography” (EEG) is the process
of measuring the brain’s neural activity as electrical voltage
fluctuations along the scalp that results from the current flows
in brain’s neurons [5]. In a typical EEG test, electrodes are
fixed on the scalp to monitor and record the brain’s electrical
activity [6]. BCI measures EEG signals associated with the
user’s activity then applies different signal processing
algorithms for the purpose of translating the recorded signals
into control commands for different applications [7].
The most important application for BCI is helping disabled
individuals by offering a new way of communication with the
external environment [8]. Many BCI applications were
described in [9] including controlling devices like video games
and personal computers using thoughts translation. BCI is a
highly interdisciplinary research topic that combines medicine,
neurology, psychology, rehabilitation engineering, HumanComputer Interaction (HCI), signal processing and machine
learning [10].
The strength of BCI applications lies in the way we
translate the neural patterns extracted from EEG into machine
commands. The improvement of the interpretation of these
EEG signals has become the goal of many researchers; hence,
our research work explores the possibility of multi-trial EEG
classification between left and right hand movements in an
offline manner, which will enormously smooth the path leading
to online classification and reading of executed movements,
leading us to what we can technically call “Reading Minds”.
In this work, we introduce an automated computer system
that uses advanced feature extraction techniques to identify
some of the brain activity patterns, especially for the left and
right hand movements. The system then uses machine learning
algorithms to extract the knowledge embedded in the recorded
patterns and provides the required decision rules for translating
thoughts into commands (as seen in Fig. 1).
This article is organized as follows: a brief review of related
research work is provided in Section II. In Section III, the
dataset used in this study is described. The automated feature
extraction process is described in Section IV. The generation of
our training/testing datasets and the practical implementation
and system evaluation are discussed in Section V. Conclusions
and suggested future work are provided in Section VI.
207 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
II. LITERATURE REVIEW
The idea of BCI was originally proposed by Jaques Vidal in
[11] where he proved that signals recorded from brain activity
could be used to effectively represent a user’s intent. In [12],
the authors recorded EEG signals for three subjects while
imagining either right or left hand movement based on a visual
cue stimulus. They were able to classify EEG signals into right
and left hand movements using a neural network classifier with
an accuracy of 80% and concluded that this accuracy did not
improve with increasing number of sessions.
vector consisting of the ERD/ERS patterns of the mu and beta
rhythms and the coefficients of the autoregressive model.
Artificial Neural Networks (ANNs) is applied to two kinds of
testing datasets and an average recognition rate of 93% is
achieved.
The strength of BCI applications depends lies in the way
we translate the neural patterns extracted from EEG into
machine commands. The improvement of the interpretation of
these EEG signals has become the goal of many researchers;
hence, our research work explores the possibility of multi-trial
EEG classification between left and right hand movements in
an offline manner, which will enormously smooth the path
leading to online classification and reading of any executed
movements, leading us to what we can technically call
“Reading Minds”.
III.
THE PHYSIONET EEG DATA
A. Description of the Dataset
The EEG dataset used in this research was created and
contributed to PhysioNet [19] by the developers of the
BCI2000 [20] instrumentation system. The dataset is publically
available at http://www.physionet.org/pn4/eegmmidb/.
Fig. 1. Feature extraction and translation into machine commands
The author of [13] used features produced by Motor
Imagery (MI) to control a robot arm. Features such as the band
power in specific frequency bands (alpha: 8-12Hz and beta: 1330Hz) were mapped into right and left limb movements. In
addition, they used similar features with MI, which are the
Event Related Desynchronization and Synchronization
(ERD/ERS) comparing the signal’s energy in specific
frequency bands with respect to the mentally relaxed state. It
was shown in [14] that the combination of ERD/ERS and
Movement-Related Cortical Potentials (MRCP) improves EEG
classification as this offers an independent and complimentary
information.
In [15], a hybrid BCI control strategy is presented. The
authors expanded the control functions of a P300 potential
based BCI for virtual devices and MI related sensorimotor
rhythms to navigate in a virtual environment. Imagined
left/right hand movements were translated into movement
commands in a virtual apartment and an extremely high testing
accuracy results were reached.
A three-class BCI system was presented in [16] for the
translation of imagined left/right hands and foot movements
into commands that operates a wheelchair. This work uses
many spatial patterns of ERD on mu rhythms along the
sensory-motor cortex and the resulting classification accuracy
for online and offline tests was 79.48% and 85.00%,
respectively. The authors of [17] proposed an EEG-based BCI
system that controls hand prosthesis of paralyzed people by
movement thoughts of left and right hands. They reported an
accuracy of about 90%.
A single trial right/left hand movement classification is
reported in [18]. The authors analyzed both executed and
imagined hand movement EEG signals and created a feature
The dataset consists of more than 1500 EEG records, with
different durations (one or two minutes per record), obtained
from 109 healthy subjects. Subjects were asked to perform
different motor/imagery tasks while EEG signals were recorded
from 64 electrodes along the surface of the scalp. Each subject
performed 14 experimental runs:
A one-minute baseline runs (with eyes open)
A one-minute baseline runs (with eyes closed)
Three two-minute runs of each of the four following
tasks:
o
The left or right side of the screen shows a target.
The subject keeps opening and closing the
corresponding fist until the target disappears. Then
he relaxes.
o
The left or right side of the screen shows a target.
The subject imagines opening and closing the
corresponding fist until the target disappears. Then
he relaxes.
o
The top or bottom of the screen. A target appears
on either. The subject keeps opening and closing
either both fists (in case of a top-target) or both feet
(in case of a bottom-target) until the target
disappears. Then he relaxes.
o
The top or bottom of the screen A target appears on
either. The subject imagines opening and closing
either both fists (in case of a top-target) or both feet
(in case of a bottom-target) until the target
disappears. Then he relaxes.
The 64-channels EEG signals were recorded according to
the international 10-20 system (excluding some electrodes) as
seen in Fig. 2.
208 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
B. The Subset used in the Current Work
From this dataset, we selected the three (two-minute) runs
of the first task described above (opening and closing the
left/right fist based on a target that appears on left or right side
of the screen). These runs include EEG data for executed hand
movements.
We created an EEG data subset corresponding to the first
six subjects (S001, S002, S003, S004, S005, and S006)
including three runs of executed movement specifically per
subject for a total of 18 two-minute records.
IV.
AUTOMATED ANALYSIS OF EEG SIGNALS FOR FEATURE
EXTRACTION
A. Channel Selection
According to [6], many of the EEG channels appeared to
represent redundant information. It is shown in [21, 22] that the
neural activity that is correlated to the executed left and right
hand movements is almost exclusively contained within the
channels C3, C4, and CZ of the EEG channels of Fig. 2. This
means that there is no need to analyze all 64 channels of data.
On the other hand, only eight electrode locations are
commonly used for MRCP analysis covering the regions
between frontal and central sites (FC3, FCZ, FC4, C3, C1, CZ,
C2, and C4) [14]. These channels were used for the
Independent Component Analysis (ICA) discussed later in the
current section (Fig. 3).
Fig. 3. Schematic diagram for the proposed system.
B. Filtering
Because EEG signals are known to be noisy and nonstationary, filtering the data is an important step to get rid of
unnecessary information from the raw signals. EEGLAB [23],
which is an interactive MATLAB toolbox, was used to filter
EEG signals.
A band pass filter from 0.5 Hz to 90 Hz was applied to
remove the DC (direct current) shifts and to minimize the
presence of filtering artifacts at epoch boundaries. A Notch
filter was also applied to remove the 50 Hz line noise.
C. Automatic Artifact Removal (AAR)
The EEG data of significance is usually mixed with huge
amounts of useless data produced by physiological artifacts that
masks the EEG signals [24]. These artifacts include eye and
muscle movements and they constitute a challenge in the field
of BCI research. AAR automatically removes artifacts from
EEG data based on blind source separation and other various
algorithms.
The AAR toolbox [25] was implemented as an EEGLAB
plug-in in MATLAB and was used to process our EEG data
subset on two stages: Electrooculography (EOG) removal using
the Blind Source Separation (BSS) algorithm then
Electromyography (EMG) Removal using the same algorithm
[26].
Fig. 2. Electrodes of the International 10-20 system for EEG
D. Epoch Extraction (Splitting)
After the AAR process, the continuous EEG data were
epoched by extracting data epochs that are time locked to
specific event types.
209 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
When no sensory inputs or motor outputs are being
processed, the mu (8–12 Hz) and beta (13–30 Hz) rhythms are
said to be synchronized [4, 27]. These rhythms are
electrophysiological features that are associated with the
brain’s normal motor output channels [4, 27]. While preparing
for a movement or executing a movement, a desynchronization
of the mu and beta rhythms occurs which is referred to as ERD
and it can be extracted 1-2 seconds before onset of movement
(as depicted in Fig. 4). Later, these rhythms synchronize again
within 1-2 seconds after movement, and this is referred to as
ERS.
On the other hand, delta rhythms can be extracted from the
motor cortex, within the pre-movement stage, and this is
referred to MRCP. The slow (less than 3 Hz) MRCP is
associated with an event-related negativity that occurs 1-2
seconds before the onset of movement [28, 29].
In our experiments, we extracted time-locking events with
type = 3 (left hand) or type = 4 (right hand) with different
epoch limits and types of analysis:
ERD analysis: epoch limits from -2 to 0 seconds.
ERS analysis: epoch limits from 4.1 to 5.1 seconds.
MRCP analysis: epoch limits from -2 to 0 seconds.
each run: ERD/ERS and MRCP for both left and right hand
movements for each subject.
V.
PRACTICAL IMPLEMENTATION AND RESULTS
A. Feature Vectors Construction and Numerical
Representation
After the EEG datasets were analyzed as described in the
previous section, the activation vectors were calculated for
each of the resulted epochs’ datasets as the multiplication of the
ICA weights and ICA sphere for each dataset subtracting the
mean of the raw data from the multiplication results.
Then, the mean, power, and energy of the activations were
calculated to construct the feature vectors. For each subject’s
single run, 6 feature vectors were extracted as <Power (8
features), Mean (8 features), Energy (8 features), Type (1
feature: ERS/ERD/MRCP), Side (1 target: Left/Right)>
resulting in a 10826 feature matrix.
The constructed features were represented in a numerical
format that is suitable for use with machine learning algorithms
[33, 34]. Every column in the features matrices was normalized
between 0.1 and 0.9 such that the datasets could be inputted to
the learning algorithms described in the next subsection.
B. Machine Learning Algorithms
In this work, Neural Networks (NNs) and Support Vector
Machines (SVMs) algorithms were optimized for the purpose
of classifying EEG signals into right and left hand movements.
A detailed description of these learning algorithms can be
found in [35] and [36].
The MATLAB neural networks toolbox was used for all
NN experiments. The number of input features (25 features)
determined the number of input nodes for NN and the number
of different target functions (1 output: left or right) determined
the number of output nodes. Training was handled with the aid
of the back-propagation learning algorithm [37].
Fig. 4. Epoch Extraction (ERS/ERD and MRCP)
E. Independent Component Analysis (ICA)
After the AAR process, ICA was used to parse the
underlying electrocortical sources from EEG signals that are
affected by artifacts [30, 31]. Data decomposition using ICA
changes the basis linearly from data that are collected at single
scalp channels to a spatially transformed virtual channel basis.
Each row of the EEG data in the original scalp channel data
represents the time course of accumulated differences between
source projections to a single data channel and one or more
reference channels [32].
EEGLAB was used to run ICA on the described epoched
datasets (left and right ERD, ERS, and MRCP) for the channels
FC3, FCZ, FC4, C3, C1, CZ, C2, and C4.
F. Rhythm Isolation
A short IIR band pass filter from 8 to 30 Hz was applied on
the ERD/ERS epoched datasets of the experiment for the
purpose of isolating mu/beta rhythms. Another short IIR
lowpass filter of 3 Hz was applied on MRCP epoched datasets
for isolating delta rhythms. The result of this was 6 files for
All SVM experiments were carried out using the
“MySVM” software [38]. SVM can be performed with
different kernels and most of them were reported to provide
similar results for similar applications [6]. So, the AnovaKernel SVM was used in this work.
C. Optimisation and Results
In all experiments, 80% samples were randomly selected
and used for training and the remaining 20% for testing. This
was repeated 10 times, and in each time the datasets were
randomly mixed.
For each experiment, the number of hidden nodes for NN
varied from 1 to 20. In SVM, each of the degree and gamma
parameters varied from 1 to 10. The mean of the accuracy was
calculated for each ten training-testing pairs.
The features that were used as inputs to NN and SVM are
symbolized as follows:
P: the power.
M: the mean.
E: the energy.
210 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
X: the sample type (ERS/ERD/MRCP).
[3]
The results of the experiment are summarized in the Table
[4]
I.
TABLE I.
NN
Features
All
P, X
M, X
E, X
P, M, X
M, E, X
P, E, X
Accuracy
%
88.9
80.4
68.5
82.1
79.8
82.7
89.8
[5]
RESULTS FOR EXPERIMENT B.
SVM
Hidden
Layers
3
15
11
11
3
9
4
Accuracy
%
85.3
88.2
91.2
94.1
80.6
82.4
97.1
Degree
Gamma
1
3
3
4
8
5
4
5
4
10
5
3
4
4
[6]
[7]
[8]
[9]
It is clear from the testing results that SVM outperforms
NN in most experiments. An SVM topology of degree = 4 and
gamma = 4 provides an accuracy of 97.1% if tested with the
power, energy and type inputs of the experiment. A NN of 10
hidden layers can provide an accuracy of 86.5% if all features
are used. These results clearly show that the use of advanced
feature extraction techniques provides good and clear
properties that can be translated using machine learning into
machine commands.
[10]
The next best SVM performance (94.1%) is achieved using
the energy and type features. In general, there has been an
increase in the classification performance with the use of more
discriminative features, such as the total energy, compared to
the power and mean inputs.
[13]
VI. CONCLUSIONS AND FUTURE RESEARCH
This paper focuses on the classification of EEG signals for
right and left fist movements based on a specific set of features.
Very good results were obtained using NNs and SVMs
showing that offline discrimination between right and left
movement, for executed hand movements, is comparable to
leading BCI research. Our methodology is not the best, but is
somewhat a simplified efficient one that satisfies the needs for
researchers in field of neuroscience.
[15]
In the near future, we aim to develop and implement our
system in online applications, such as health systems and
computer games. In addition, more datasets has to be analyzed
for a better knowledgeable extraction and more accurate
decision rules.
[11]
[12]
[14]
[16]
[17]
[18]
[19]
ACKNOWLEDGMENT
The authors would like to acknowledge the financial
support received from Applied Science University that helped
in accomplishing the work of this article.
[1]
[2]
References
J. P. Donoghue, "Connecting cortex to machines: recent advances in
brain interfaces," Nature Neuroscience Supplement, vol. 5, pp. 1085–
1088, 2002.
S. Levine, J. Huggins, S. BeMent, R. Kushwaha, L. Schuh, E. Passaro,
M. Rohde, and D. Ross, "Identification of electrocorticogram patterns as
the basis for a direct brain interface," Journal of Clinical
Neurophysiology, vol. 16, pp. 439-447, 1999.
[20]
[21]
[22]
A. Vallabhaneni, T. Wang, and B. He, "Brain—Computer Interface," in
Neural Engineering, B. He, Ed.: Springer US, 2005, pp. 85-121.
J. Wolpaw, N. Birbaumer, D. McFarland, G. Pfurtscheller, and T.
Vaughan, "Brain-computer interfaces for communication and control,"
Clinical Neurophysiology, vol. 113, pp. 767-791, 2002.
E. Niedermeyer and F. H. L. da Silva, Electroencephalography: Basic
Principles, Clinical Applications, and Related Fields: Lippincott
Williams & Wilkins, 2005.
J. Sleight, P. Pillai, and S. Mohan, "Classification of Executed and
Imagined Motor Movement EEG Signals," Ann Arbor: University of
Michigan, 2009, pp. 1-10.
B. Graimann, G. Pfurtscheller, and B. Allison, "Brain-Computer
Interfaces: A Gentle Introduction," in Brain-Computer Interfaces:
Springer Berlin Heidelberg, 2010, pp. 1-27.
A. E. Selim, M. A. Wahed, and Y. M. Kadah, "Machine Learning
Methodologies in Brain-Computer Interface Systems," in Biomedical
Engineering Conference, 2008, CIBEC 2008. Cairo, 2008, pp. 1-5.
E. Grabianowski, "How Brain-computer Interfaces Work,"
http://computer.howstuffworks.com/brain-computer-interface.htm, 2007.
M. Smith, G. Salvendy, K. R. Müller, M. Krauledat, G. Dornhege, G.
Curio, and B. Blankertz, "Machine Learning and Applications for BrainComputer Interfacing," in Human Interface and the Management of
Information. Methods, Techniques and Tools in Information Design. vol.
4557: Springer Berlin Heidelberg, 2007, pp. 705-714.
J. J. Vidal, "Toward Direct Brain-Computer Communication," Annual
Review of Biophysics and Bioengineering, vol. 2, pp. 157-180, 1973.
G. Pfurtscheller, C. Neuper, D. Flotzinger, and M. Pregenzer, "EEGbased discrimination between imagination of right and left hand
movement," Electroencephalography and Clinical Neurophysiology, vol.
103, pp. 642-651, 1997.
F. Sepulveda, "Brain-actuated Control of Robot Navigation," in
Advances in Robot Navigation, A. Barrera, Ed.: InTech, 2011.
A.-K. Mohamed, "Towards improved EEG interpretation in a
sensorimotor BCI for the control of a prosthetic or orthotic hand," in
Faculty of Engineering. Master of Science in Engineering,
Johannesburg: Universityof Witwatersrand, 2011, p. 144.
Y. Su, Y. Qi, J.-x. Luo, B. Wu, F. Yang, Y. Li, Y.-t. Zhuang, X.-x.
Zheng, and W.-d. Chen, "A hybrid brain-computer interface control
strategy in a virtual environment," Journal of Zhejiang University
SCIENCE C, vol. 12, pp. 351-361, 2011.
Y. Wang, B. Hong, X. Gao, and S. Gao, "Implementation of a BrainComputer Interface Based on Three States of Motor Imagery," in 29th
Annual International Conference of the IEEE Engineering in Medicine
and Biology Society, EMBS2007, 2007, pp. 5059-5062.
C. Guger, W. Harkam, C. Hertnaes, and G. Pfurtscheller, "Prosthetic
Control by an EEG-based Brain- Computer Interface (BCI)," in AAATE
5th European Conference for the Advancement of Assistive Technology,
Düsseldorf, Germany, 1999.
J. A. Kim, D. U. Hwang, S. Y. Cho, and S. K. Han, "Single trial
discrimination between right and left hand movement with EEG signal,"
in Proceedings of the 25th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, 2003., Cancun, Mexico,
2003, pp. 3321-3324 Vol.4.
A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C.
Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E.
Stanley, "PhysioBank, PhysioToolkit, and PhysioNet: Components of a
New Research Resource for Complex Physiologic Signals," Circulation,
vol. 101, pp. e215-e220, 2000.
G. Schalk, D. J. McFarland, T. Hinterberger, N. Birbaumer, and J. R.
Wolpaw, "BCI2000: a general-purpose brain-computer interface (BCI)
system," IEEE Transactions on Biomedical Engineering, vol. 51, pp.
1034-1043, 2004.
L. Deecke, H. Weinberg, and P. Brickett, "Magnetic fields of the human
brain accompanying voluntary movements: Bereitschaftsmagnetfeld,"
Experimental Brain Research, vol. 48, pp. 144-148, 1982.
C. Neuper and G. Pfurtscheller, "Evidence for distinct beta resonance
frequencies in human EEG related to specific sensorimotor cortical
areas," Clinical Neurophysiology, vol. 112, pp. 2084-2097, 2001.
211 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 4, No. 6, 2013
[23] A. Delorme and S. Makeig, "EEGLAB: an open source toolbox for
analysis of single-trial EEG dynamics," Journal of Neuroscience
Methods, vol. 134, pp. 9-21, 2004.
[24] G. Bartels, S. Li-Chen, and L. Bao-Liang, "Automatic artifact removal
from EEG - a mixed approach based on double blind source separation
and support vector machine," in 2010 Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC),
2010, pp. 5383-5386.
[25] G. Gómez-Herrero, "Automatic Artifact Removal (AAR) toolbox for
MATLAB," in Transform methods for Electroencephalography (EEG):
http://kasku.org/projects/eeg/aar.htm, 2008.
[26] C. Joyce, I. Gorodnitsky, and M. Kutas, "Automatic removal of eye
movement and blink artifacts from EEG data using blind component
separation," Psychophysiology, vol. 41, pp. 313-325, 2004.
[27] A. Bashashati, M. Fatourechi, R. Ward, and G. Birch, "A survey of
signal processing algorithms in brain-computer interfaces based on
electrical brain signals," Journal of Neural Engineering, vol. 4, pp. R3257, 2007.
[28] A. Vuckovic and F. Sepulveda, "Delta band contribution in cue based
single trial classification of real and imaginary wrist movement,"
Medical and Biological Engineering and Computing, vol. 46, pp. 529 –
539, 2008.
[29] Y. Gu, K. Dremstrup, and D. Farina, "Single-trial discrimination of type
and speed of wrist movements from EEG recordings," Clinical
Neurophysiology, vol. 20, pp. 1596–1600, 2009.
[30] J. T. Gwin and D. Ferris, "High-density EEG and independent
component analysis mixture models distinguish knee contractions from
ankle contractions," in 2011 Annual International Conference of the
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
IEEE Engineering in Medicine and Biology Society, EMBC, boston,
USA, 2011, pp. 4195-4198.
S. Makeig, A. J. Bell, T. P. Jung, and T. J. Sejnowski, "Independent
component analysis of electroencephalographic data," Advances in
Neural Information Processing Systems, vol. 8, pp. 145-151, 1996.
A. Delorme and S. Makeig, "Single subject data processing tutorial:
Decomposing Data Using ICA," in The EEGLAB Tutorial:
http://sccn.ucsd.edu/wiki/EEGLAB, 2013.
M. Al-Omari, R. Qahwaji, T. Colak, and S. Ipson, "Machine learningbased investigation of the associations between cmes and filaments,"
Solar Physics, vol. 262, pp. 511-539, 2010.
R. Qahwaji, T. Colak, M. Al-Omari, and S. Ipson, "Automated machine
learning based prediction of CMEs based on flare associations," Sol.
Phys, vol. 248, 2007.
R. Qahwaji and T. Colak, "Automatic Short-Term Solar Flare Prediction
Using Machine Learning and Sunspot Associations," Solar Phys., vol.
241, pp. 195-211, 2007.
R. Qahwaji, M. Al-Omari, T. Colak, and S. Ipson, "Using the Real,
Gentle and Modest AdaBoost Learning Algorithms to Investigate the
Computerised Associations between Coronal Mass Ejections and
Filaments," in Mosharaka International Conference on Communications,
Computers and Applications (MIC-CCA 2008), Mosharaka for
Researches and Studies, Amman, Jordan, 2008, pp. 37-42.
S. E. Fahlmann and C. Lebiere, "The cascade-correlation learning
architecture," in Advances in Neural Information Processing Systems 2
(NIPS-2) Denver, Colorado, 1989.
S. Rüping, "mySVM-Manual ": University of Dortmund, Lehrstuhl
Informatik 8, 2000.
212 | P a g e
www.ijacsa.thesai.org
| 9cs.NE
|
arXiv:1702.05063v2 [math.ST] 9 Mar 2018
A concentration inequality for the excess risk in least-squares
regression with random design and heteroscedastic noise
Adrien Saumard
Crest-Ensai, Université Bretagne Loire
March 12, 2018
Abstract
We prove a new and general concentration inequality for the excess risk in least-squares regression
with random design and heteroscedastic noise. No specific structure is required on the model, except the
existence of a suitable function that controls the local suprema of the empirical process. So far, only the
case of linear contrast estimation was tackled in the literature with this level of generality on the model.
We solve here the case of a quadratic contrast, by separating the behavior of a linearized empirical process
and the empirical process driven by the squares of functions of models.
keywords: regression, least-squares, excess risk, empirical process, concentration inequality, margin
relation.
AMS2000 : 62G08, 62J02, 60E15.
1
Introduction
The excess risk of a M-estimator is a fundamental quantity of the theory of statistical learning. Consequently,
a general theory of rates of convergence as been developed in the nineties and early 2000 ([Mas07, Kol11]).
However, it has been recently identified that some theoretical descriptions of learning procedures need finer
controls than those brought by the classical upper bounds of the excess risk. In this case, the derivation of
concentration inequalities for the excess risk is a new and exiting axis of research, of particular importance for
obtaining satisfying oracle inequalities in various contexts, especially linked to high dimension.
In the field of model selection, it has been indeed remarked that such concentration inequalities allow to
discuss the non-asymptotic optimality of model selection procedures ([BM07, AM09, Ler11, Sau12]). More
precisely, concentration inequalities for the excess risk and for the excess empirical risk ([BM11]) are central
tools to access the optimal constants in the oracle inequalities describing the model selection accuracy. Such
results have put to evidence the optimality of the so-called slope heuristics ([BM07, BMM12]) and more
generally of selection procedures based on the estimation of the minimal penalty ([AB09]), in statistical
frameworks linked to regularized quadratic estimation. Under similar assumptions, it is also possible to
discuss optimality of resampling and cross-validation type procedures ([Arl08, AL15, Cel14, NS17]).
In high dimension, convex methods allow to design and compute efficient estimators. This is the reason why
Chatterjee [Cha14] has recently focused on the estimation of the mean of a high dimensional Gaussian vector
under convex constraints. By getting a concentration inequality for the excess risk of a projected least-squares
estimator, Chatterjee [Cha14] has proved the universal admissibility of this estimator. The concentration
inequality has then been sharpened and extended to the excess risk of estimators minimizing penalized convex
criteria ([MvdG15, vdGW17]).
It is also well known (see for instance [BLT17]) that a weakness of the theory of regularized estimators
in a sparsity context is that classical oracle inequalities, such as in [BRT09], describe the performance of
estimators with an amount of regularization that actually depends on the confidence level considered in the
oracle inequality. This does not correspond to any practice with this kind of estimators, whose regularization
parameter is usually fixed using a cross-validation procedure. Recently, Bellec and Tsybakov [BT17], building
on [BLT18], have established more satisfying oracle inequalities, describing the performance of regularized
estimators such as LASSO, group LASSO and SLOPE with a confidence level independent of the regularization
1
parameter. In particular, the oracle inequalities can be integrated. Again, the central tool to obtain such bound
is a concentration inequality for the excess risk estimator at hand.
In this paper, we extend the technology developed in [Cha14, MvdG15, vdGW17] in order to establish a new
concentration inequality for the excess risk in least-squares regression with random design and heteroscedastic
noise. This is appealing since [Cha14] and [MvdG15] cover only least-squares regression with fixed design and
homoscedastic Gaussian noise, while [vdGW17] have to assume that the law of the design is known in order
to perform a ”linearized least-squares regression” (see Section 6.2 of [vdGW17]).
Our strategy is as follows. We first remark that the empirical process of interest splits into two parts: a
linear process and a quadratic one. Then we prove that the linear process achieves a second order margin
condition as defined in [vdGW17] and put meaningful conditions on the quadratic process in order to handle it.
Techniques from empirical process theory such as Talagrand’s type concentration inequalities and contraction
arguments are at the core of our approach.
The paper is organized as follows. The regression framework as well as some properties linked to margin
relations are described in Section 2. Then we state our main result in Section 4. The proofs are deferred to
Section 4.
2
Least-squares heteroscedastic regression with random design
2.1
Setting
n
Let {(Xi , Yi )}i=1 be an i.i.d. sample taking values in X × R, where X is a measurable space - typically a
subset of Rp . We assume that the following relation holds
Yi = g∗ (Xi ) + σ (Xi ) εi
for i = 1, ..., n ,
where g∗ is the regression function, σ is the heteroscedastic
noise level and E [εi |Xi ] = 0, E ε2i |Xi = 1.
We take a closed convex model G ⊂ L2 P X , where P X is the common distribution of the Xi′ s, and set
g 0 = arg min {P (γ (g))} ,
g∈G
where P is the common distribution of the pairs (Xi , Yi ) and γ is the least-squares contrast, defined by
γ (g) (x, y) = (y − g (x))2 .
We will also denote fg := γ (g) , f∗ = γ (g∗ ), f 0 = γ g 0 , fˆ = γ (ĝ) and0 F := γ (G). The function g 0 will be
called the projection of the regression function onto the model G. Indeed, if we denote k·k the quadratic norm
in L2 P X , it holds
g 0 − g∗ = min kg − g∗ k .
g∈G
We consider the least-squares estimator ĝ over G, defined to be
ĝ ∈ arg min {Pn (γ (g))} ,
g∈G
where Pn := 1/n
Pn
i=1 δ (Xi ,Yi )
is the empirical measure associated to the sample.
We want to assess the concentration of the quantity P fˆ − f 0 = P γ (ĝ) − γ g 0 , called the excess risk
of the least-squares estimator on G, around a single deterministic point. To this end, it is easy to see ([Sau12],
Remark 1, [NS17], Section 5.2 or van de Geer and Wainwright [vdGW17]) that the following representation
formula holds for the excess risk on G in terms of empirical process,
r
n
o
(1)
ŝ := P fˆ − f 0 = arg min s2 − Ên (s) ,
s≥0
where
Ên (s) = max (Pn − P ) f 0 − f
f ∈Fs
2
,
with Fs = f ∈ F ; P f − f 0 ≤ s2 . It is shown in [vdGW17] for various settings that include linearized
regression that the quantity ŝ actually concentrates around the following point,
i
h
(2)
s0 := arg min s2 − E (s) , E (s) := E Ên (s) .
s≥0
2.2
On a margin-like relation pointed on the projection of the regression function
In order to prove concentration inequalities for the excess risk on G, we will need to check the following
margin-like relation, also called ”quadratic curvature condition” in [vdGW17]: there exists a constant C > 0
such that
σ2 f − f 0
0
for all f ∈ F ,
(3)
P f −f ≥
C2
where
2
σ 2 (f ) := Ef 2 (X1 ) − (Ef (X1 ))
is the variance of f with respect to P X .
A very classical relation in statistical learning, called margin relation, consists in assuming that (3) holds
with f 0 is replaced by the image of the target f∗ = γ (g∗ ). Such relation is satisfied in least-squares regression
whenever the response variable Y is uniformly bounded. Here we do not assume that f∗ belongs to F thus f 0
may be different from f∗ .
Condition 1 There exists A1 > 0 such that |Y | ≤ A1 a.s.
From Condition 1 we deduce that kg∗ k∞ ≤ A1 and kσk∞ ≤ 2A1 .
Condition 2 There exists A2 > 0 such that
sup kgk∞ ≤ A2 < ∞ .
g∈G
From Conditions 1 and 2, we deduce that the image model F is also uniformly bounded: there exists K > 0
such that
sup f − f 0 ∞ ≤ K < ∞ .
f ∈F
More precisely, K = 2 (A1 + A2 ) is convenient.
The following proposition shows that relation (3) is satisfied in our regression setting whenever the response
variable is bounded and the model G is convex and uniformly bounded. It can also be found in [Lec11],
Proposition 1.3.3.
Proposition 3 If the model G is convex and Conditions 1 and 2 hold, then there exists a constant C > 0 such
that
σ2 f − f 0
P f − f0 ≥
for all f ∈ F ,
(4)
C2
Furthermore, C = 2 (A1 + A2 ) is convenient.
The major gain brought by Proposition 3 over the classical margin relation is that the bias of the model,
that is the quantity P f 0 − f∗ that is implicitly contained in the excess risk appearing in the classical margin
relation, is pushed away from Inequality (4). Proposition 3 is thus a refinement over the classical notion of
margin relation. It is stated for the least-squares contrast but it is easy to see that it can be extended to more
general situations, where the contrast γ is convex and regular in some sense (see Proposition 2.1, Section 2.2.3
in [Sau10]). For completeness, the proof of Proposition can be found in Section 4.1.
3
2.3
Second order quadratic margin condition
First notice that the arguments of the empirical process of interest can be decomposed into a linear and a
quadratic part. It holds, for any f = fg ∈ F and any (x, y) ∈ X ×R,
fg (x, y) − f 0 (x, y) = γ (g) (x, y) − γ g 0 (x, y)
2
(5)
= ψ (x, y) · g − g 0 (x) + g − g 0 (x) ,
where ψ (x, y) = −2 y − g 0 (x) .
To this contrast expansion around the projection g 0 of the regression function onto G, we can associate
two empirical processes, that we will call respectively the linear and
the quadratic empirical process, and we
will be more precisely interested by their local maxima on Gs := g ∈ G ; g − g 0 ≤ s , s ≥ 0,
n
2 o
.
and Ên,q (s) = max (Pn − P ) g − g 0
Ên,ℓ (s) = max (Pn − P ) ψ · g − g 0
g∈Gs
g∈Gs
In what follows, we will not directly show that the excess risk concentrates around s0 defined in (2), but rather
around a point s̃0 , defined to be,
h
i
s̃0 = arg min s2 − Eℓ (s) , Eℓ (s) = E Ên,ℓ (s) .
s≥0
It holds around s̃0 a relation of the type of a second order margin relation, introduced in [vdGW17], as proved
in the following Lemma, which proof is available in Section 4.1.
Lemma 4 For any s ≥ 0, it holds
and also
s2 − Eℓ (s) − s̃20 − Eℓ (s̃0 ) ≥ (s − s̃0 )2 ,
(6)
h
i
2
s2 − Ên,ℓ (s) − s̃20 − Ên,ℓ (s̃0 ) ≥ (s − s̃0 ) .
(7)
The fundamental difference with the second order margin relation stated in [vdGW17] is that we require
in Lemma 4 conditions on the linear part of empirical process and not on the empirical process of origin, that
takes in arguments contrasted functions. Indeed, it seems that for the latter empirical process, the second order
margin relation does not hold or is hard to check in least-squares regression in general. This difficulty indeed
forced van de geer and Wainwright [vdGW17], Section 6.2, to work in a ”linearized least-squares regression”
context, under the quite severe restriction that the distribution of design is known from the statistitian. On
contrary, our main result stated in Section 3 below is stated for a general regression situation, where the
distribution of the design is unknown and the noise level is heteroscedastic.
3
Main result
Before stating our new concentration inequality, we describe the required assumptions. In order to state the
next condition, let us denote
h
i
, E1 (s) = E Ên,1 (s) .
Ên,1 (s) = max (Pn − P ) g − g 0
g∈Gs
Condition 5 There is a sequence mn and a strictly increasing function J1 such that the function
is strictly convex and such that
2
ΦJ1 (u) := J1−1 (u) , u > 0,
E1 (s) ≤
J1 (s)
, s ≥ 0.
mn
4
Condition 6 Take K > 0. For any s > 0, there exists a constant D (s) ∈ (0, K] such that
sup g − g 0
∞
g∈Gs
≤ D (s) .
Notice that if Conditions 1, 2 and 5 hold, then by the use of classical symmetrization and contraction
arguments, for all s ≥ 0,
J1 (s)
J1 (s)
Eℓ (s) ≤ CA1 ,A2
and Eq (s) ≤ 4D (s)
,
mn
mn
where CA1 ,A2 is a positive constant that only depends on A1 and A2 . From now on, we set
J := max {CA1 ,A2 ; 4 (K ∨ 1)} J1 ,
so that max {Eℓ (s) ; Eq (s)} ≤ J (s) /mn for any s > 0 and also Eq (s) ≤ D (s) J (s) /mn .
We are now able to state our main result.
Theorem 7 If
J (s) = AJ s ; D (s) = A∞ s ; mn =
then it holds, for any t ≥ 0,
P
ĝ − g
Hence, if moreover
0
− s̃0
√
AJ A∞ s̃0
∨ c0
≥
n1/4
r
√
A0
n ; s̃0 ≍ √
n
√ !!
√
t + ln (1 + K n) t + ln (1 + K n)
≤ e−t .
+
n
n
(8)
√
√
√
ln n ≪ A0 ≪ n and AJ A∞ ≪ n then,
!
√
√
AJ A∞
ln n
= oP (1) .
∨
A0
n1/4
ĝ − g 0 − s̃0
= OP
s̃0
Inequality (8) of Theorem 7 is a new concentration inequality related least-squares regression with random
design and heteroscedastic noise on a convex, uniformly bounded model. In particular, it extends results of
[vdGW17] related to linearized regression, which is a simplified framework for regression, to the classical and
general least-squares framework described in Section 2.1 above. The proof of Theorem 7 is detailed in Section
4.2.
The following corollary provides an generic example entering into the assumptions of Theorem 7 and that
is related to linear aggregation via empirical risk minimization.
D
Corollary 8 Define M = Span (ϕ1 , ...,
ϕD ) the linear span generated by (ϕk )k=1 , an orthonormal dictionary
in L2 P X . Take G = B(M,L∞ ) g 0 , 1 ⊂ M the unit ball in sup-norm of M centered on g 0 , the projection of
g∗ onto M. Assume that
√
sup
kgk∞ ≤ cM D
(9)
g∈M,kgk2 =1
and
2
(ln n) ≤ D ≤
Then, if s̃0 ≍
p
D/n,
ĝ − g 0 − s̃0
= OP
s̃0
√
D
n1/4
√
n
.
ln n
∨
r
ln n
D
!
= oP (1) .
Note that Inequality (9) relating the sup-norm to the quadratic norm of the functions in the linear span
of the dictionary, is classical in non-parametric estimation and is satisfied for the usual functional bases such
as the Fourier basis, wavelets or piecewise polynomials over a regular partition (including histograms), see for
instance [BBM99]. In particular, Corollary 8 extends a concentration inequality recently obtained in [Sau18],
for the excess risk of the ERM in the linear aggregation problem when the dictionary at hand is the Fourier
dictionary.
5
4
4.1
Proofs
Proofs related to Section 2
Proof of proposition 3. Take f = fg ∈ F . Then, on the one hand,
2
≤ P f − f0
σ2 f − f 0
h
i
2
= E γ (g) − γ g 0
(X, Y )
= E (g − g 0 )2 (X) (2Y − g (X) − g0 (X))2
≤ 4 (A1 + A2 )
2
g − g0
γ (g) − γ g 0
(X, Y )
2
.
(10)
On the other hand,
P f − f0
=
E
=
E (g 0 − g) (X) (2Y − g (X) − g0 (X))
E (g 0 − g) (X) (2 Y − g 0 (X) + g0 (X) − g (X))
2
g − g 0 − 2E Y − g 0 (X) g − g 0 (X)
2
g − g 0 − 2E g∗ − g 0 (X) g − g 0 (X)
=
=
=
≥
g − g0
2
.
(11)
The latter inequality, which corresponds to E g∗ − g (X) g − g (X) ≤ 0 comes from the fact that G is
convex and so, g 0 being the projection of g∗ onto G, the scalar product in L2 P X between the functions
g∗ − g 0 and g − g 0 is nonpositive. Combining (10) and (11) now gives the result.
0
0
Proof of Lemma 4. Inequality (4) derives from (7) by taking expectation on both sides. Concerning the
proof of (4), it is easily seen that the function s 7→ Ên,ℓ (s) is concave. Indeed, take for si ≥ 0, i ∈ {1, 2},
gn,si = arg max (Pn − P ) ψ 1 · g − g 0
.
g∈Gsi
For any α, β ∈ (0, 1) , α + β = 1, if gb = αgn,s1 + βgn,s2 , then by the triangular inequality, gb ∈ Gαs1 +βs2 ,
which gives,
Ên,ℓ (αs1 + βs2 ) ≥ (Pn − P ) ψ 1 · gb − g 0 = αÊn,ℓ (s1 ) + β Ên,ℓ (s2 ) .
Now, from the concavity of s 7→ Ên,ℓ (s), we deduce that the function s 7→ s2 − Ên,ℓ (s) is 1-strongly convex,
which implies (6).
4.2
Proofs related to Section 3
Proof of Theorem 7. We prove the concentration of ĝ − g 0 at the right of s̃0 and arguments will be of
the same type for the deviations at the left. Take t, ε > 0, J := ⌈K/ε⌉ and set, for any j ∈ {1, ..., J} , the
intervals
Ij := ((j − 1) ε + δ + s̃0 , s̃0 + δ + jε] .
We also set
z (t) = 2C s̃0
It holds, for any δ > 0 (to be chosen later),
≤
≤
r
t
+ r0
n
r
Kt
t
+
.
n
n
P (s̃0 + δ < ŝ ≤ K)
P ∃s ∈ (s̃0 + δ, K) , s2 − Ên (s) ≤ s̃20 − Ên (s̃0 )
P ∃s ∈ (s̃0 + δ, K) , s2 − Ên (s) ≤ s̃20 − E (s̃0 ) + z (t) + e−t ,
6
where in the last inequality we used Lemma 10. Furthermore, by setting for all j ∈ {1, ..., J},
Pj := P ∃s ∈ Ij , s2 − Ên (s) ≤ s̃20 − E (s̃0 ) + z (t) ,
a union bound gives,
J
X
Pj .
P ∃s ∈ (s̃0 + δ, K) , s2 − Ên (s) ≤ s̃20 − E (s̃0 ) + z (t) ≤
j=1
Now, for each index j and for all s ∈ Ij , we have
2
s2 − Ên (s) ≥ ((j − 1) ε + δ + s̃0 ) − Ên (jε + δ + s̃0 ) .
Furthermore, it holds for all u > 0, with probability 1 − e−u ,
r
u
+ z (u)
n
Eℓ (jε + δ + s̃0 ) + Eq (jε + δ + s̃0 )
r
u
+2C (δ + jε)
+ z (u) ,
n
Ên (jε + δ + s̃0 ) ≤
E (jε + δ + s̃0 ) + 2C (δ + jε)
≤
where the first inequality comes from Lemma 10. By Lemma 4, we then have
2
((j − 1) ε + δ + s̃0 ) − Eℓ (jε + δ + s̃0 ) − Eq (jε + δ + s̃0 )
2
≥
s̃20 − Eℓ (s̃0 ) + (δ + jε) − D (jε + δ + s̃0 ) J (jε + δ + s̃0 ) /mn
≥
s̃20 − E (s̃0 ) + (δ + jε)2 − D (jε + δ + s̃0 ) J (jε + δ + s̃0 ) /mn
2
2
+ ((j − 1) ε + δ + s̃0 ) − (jε + δ + s̃0 )
−2ε (s̃0 + δ + jε) + ε2 .
Putting the previous estimates in (12), we get, for all s ∈ Ij ,
s2 − Ên (s) ≥
≥
2
s̃20 − E (s̃0 ) + (δ + jε) − D (jε + δ + s̃0 ) J (jε + δ + s̃0 ) /mn
r
u
−2 C
+ ε (δ + jε) − 2εs̃0 + ε2 − z (u)
n
1
2
s̃20 − E (s̃0 ) + (δ + jε) − D (jε + δ + s̃0 ) J (jε + δ + s̃0 ) /mn
2
2
r
u
+ ε − 2εs̃0 + ε2 − z (u) ,
−2 C
n
with probability 1 − e−u . We require that
2
(δ + jε) ≥ 4D (jε + δ + s̃0 ) J (jε + δ + s̃0 ) /mn .
Using the assumptions, it is equivalent to require
√
4 AJ A∞ (jε + δ + s̃0 )2
√
.
(δ + jε) ≥
n
2
Whenever AJ A∞ ≤
√
n, the last display is true if δ ∼ AJ A∞ s̃0 /n1/4 . We also require that
!
r
2
u
2
2
(δ + jε) ≥ 4 2 C
+ ε + 2εs̃0 − ε + z (u) + z (t) .
n
7
(12)
√
√
√
To finish the proof, we fix ε = 1/ n and u = t + ln (1 + K n). In particular, J ≤ 1 + K n and
J
X
j=1
Pj ≤ e−t .
Our conditions on δ = δ (t) become, for a constant c0 only depending on C and K,
r
√
√ !
√
AJ A∞ s̃0
t + ln (1 + K n) t + ln (1 + K n)
.
+
δ (t) ≥
∨ c0
n
n
n1/4
since s̃0 ∨ r0 ≪ 1.
Proof of Corollary 8. Under Assumption (9), we have
"
r
#
2
D
0
1/2
≤s
max (Pn − P ) g − g
E1 (s) ≤ E
,
g∈Gs
n
where we used two times Cauchy-Schwarz inequality. Hence, mn =
Furthermore, by Assumption (9), we have
√
sup g − g 0 ∞ ≤ cM s D .
√
√
n and J (s) = s D are convenient.
g∈Gs
√
√
√
We can thus apply Theorem 7 with D (s) = cM s D. Consequently, condition ln n ≪ A0 ≪ n turns into
ln n ≪ D ≪ n
√
√
and condition AJ A∞ ≪ n is satisfied whenever D ≪ n.
In the following theorem (see Theorem 8.1 in [vdGW17]), the right-tail inequalities are direct applications
of Bousquet [Bou02] and the left-tail inequalities are deduced from Klein and Rio [KR05].
Theorem 9 If Conditions 1 and 2 are satisfied and we set
K := sup f − f 0
f ∈Fs
∞
, σ s := sup σ f − f 0
f ∈Fs
then it holds
!
t
2Kt
P Ên (s) ≥ E (s) + (8KE (s) + 2σ 2s ) +
n
3n
!
r
Kt
t
2
P Ên (s) ≤ E (s) − (8KE (s) + 2σ s ) −
n
n
r
,
≤ e−t ,
≤ e−t ,
Using Proposition 3 and Conditions 1, 2, 5, we can simplify the bounds given in Theorem 9 as follows.
Lemma 10 If Conditions 1, 2 and 5 are satisfied, then with the same notations as in Theorem 9 and also
8K
2
2 ∗
r0 := 2C Φ
mn C 2
we have
!
r
2Kt
t
t
≤
P Ên (s) ≥ E (s) + 2Cs
+ r0
+
n
n
3n
!
r
r
t
t
Kt
P Ên,ℓ (s) ≤ E (s) − 2Cs
≤
− r0
−
n
n
n
r
8
e−t ,
e−t ,
Proof. We have σ 2s ≤ C 2 s2 , where the constant C is defined in proposition 3. Furthermore, using the fact
that uv ≤ ΦJ (u) + Φ∗J (v) for any u, v > 0, we get
8K
16K
2 2
2 ∗
= 2C 2 s2 + r02 .
J (s) ≤ 2C s + 2C ΦJ
8KE (s) ≤
mn
mn C 2
√
√
√
The conclusion is then easy to obtain by using a + b ≤ a + b.
References
[AB09]
S. Arlot and F. Bach. Data-driven calibration of linear estimators with minimal penalties. In
Y. Bengio, D. Schuurmans, J. Lafferty, C. K. I. Williams, and A. Culotta, editors, Advances in
Neural Information Processing Systems 22, pages 46–54, 2009.
[AL15]
S. Arlot and M. Lerasle. Choice of V for V -fold cross-validation in least-squares density estimation.
J. Mach. Learn. Res., 2015. To appear. arXiv:1210.5830.
[AM09]
S. Arlot and P. Massart. Data-driven calibration of penalties for least-squares regression. J. Mach.
Learn. Res., 10:245–279 (electronic), 2009.
[Arl08]
S. Arlot. V -fold cross-validation improved: V -fold penalization, February 2008. arXiv:0802.0566v2.
[BBM99]
A. Barron, L. Birgé, and P. Massart. Risk bounds for model selection via penalization. Probab.
Theory Related Fields, 113(3):301–413, 1999.
[BLT17]
P. C. Bellec, G. Lecué, and A. B. Tsybakov. Towards the study of least squares estimators with
convex penalty. arXiv preprint arXiv:1701.09120, 2017.
[BLT18]
P. C. Bellec, G. Lecué, and A. B. Tsybakov. Slope meets lasso: improved oracle bounds and
optimality. Ann. Stat., 2018. to appear, arXiv:1605.08651.
[BM07]
L. Birgé and P. Massart. Minimal penalties for Gaussian model selection. Probab. Theory Related
Fields, 138(1-2):33–73, 2007.
[BM11]
S. Boucheron and P. Massart. A high-dimensional Wilks phenomenon. Probab. Theory Related
Fields, 150(3-4):405–433, 2011.
[BMM12] J.-P. Baudry, C. Maugis, and B. Michel. Slope heuristics: overview and implementation. Stat.
Comput., 22(2):455–470, 2012.
[Bou02]
O. Bousquet. A Bennett concentration inequality and its application to suprema of empirical
processes. C. R. Math. Acad. Sci. Paris, 334(6):495–500, 2002.
[BRT09]
P. J. Bickel, Y. Ritov, and A. B. Tsybakov. Simultaneous analysis of lasso and Dantzig selector.
Ann. Statist., 37(4):1705–1732, 2009.
[BT17]
P. C. Bellec and A. B. Tsybakov. Bounds on the prediction error of penalized least squares
estimators with convex penalty. In Vladimir Panov, editor, Modern Problems of Stochastic Analysis
and Statistics, Selected Contributions In Honor of Valentin Konakov. Springer, 2017. To appear.
[Cel14]
A. Celisse. Optimal cross-validation in density estimation with the l2 -loss. Ann. Statist., 42(5):1879–
1910, 10 2014.
[Cha14]
S. Chatterjee. A new perspective on least squares under convex constraint.
42(6):2340–2381, 12 2014.
[Kol11]
V. Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems,
volume 2033 of Lecture Notes in Mathematics. Springer, Heidelberg, 2011. Lectures from the 38th
Probability Summer School held in Saint-Flour, 2008, École d’Été de Probabilités de Saint-Flour.
[Saint-Flour Probability Summer School].
9
Ann. Statist.,
[KR05]
T. Klein and E. Rio. Concentration around the mean for maxima of empirical processes. Ann.
Probab., 33(3):1060–1077, 2005.
[Lec11]
G. Lecué. Interplay between concentration, complexity and geometry in learning theory with applications to high dimensional data analysis. Habilitation à diriger des recherches, Université Paris-Est,
December 2011.
[Ler11]
M. Lerasle. Optimal model selection for density estimation of stationary data under various mixing
conditions. Ann. Statist., 39(4):1852–1877, 2011.
[Mas07]
P. Massart. Concentration inequalities and model selection, volume 1896 of Lecture Notes in Mathematics. Springer, Berlin, 2007. Lectures from the 33rd Summer School on Probability Theory
held in Saint-Flour, July 6–23, 2003, With a foreword by Jean Picard.
[MvdG15] A. Muro and S. van de Geer. Concentration behavior of the penalized least squares estimator.
arXiv preprint arXiv:1511.08698, 2015. To appear in Statistica Neerlandica.
[NS17]
F. Navarro and A. Saumard. Slope heuristics and v-fold model selection in heteroscedastic regression
using strongly localized bases. ESAIM Probab. Stat., 21:412–451, 2017.
[Sau10]
A. Saumard. Regular Contrast Estimation and the Slope Heuristics. PhD thesis, Université Rennes
1, October 2010. https://tel.archives-ouvertes.fr/tel-00569372.
[Sau12]
A. Saumard. Optimal upper and lower bounds for the true and empirical excess risks in heteroscedastic least-squares regression. Electron. J. Statist., 6(1-2):579–655, 2012.
[Sau18]
Adrien Saumard. On optimality of empirical risk minimization in linear aggregation. Bernoulli,
24(3):2176–2203, 2018.
[vdGW17] S. van de Geer and M. J. Wainwright. On concentration for (regularized) empirical risk minimization. Sankhya A, 79(2):159–200, 2017.
10
| 10math.ST
|
ON GEODESIC RAY BUNDLES IN BUILDINGS
arXiv:1708.08431v3 [math.GR] 25 Mar 2018
TIMOTHÉE MARQUIS∗
Abstract. Let X be a building, identified with its Davis realisation. In this
paper, we provide for each x ∈ X and each η in the visual boundary ∂X of
X a description of the geodesic ray bundle Geo(x, η), namely, of the union of
all combinatorial geodesic rays (corresponding to infinite minimal galleries in
the chamber graph of X) starting from x and pointing towards η. When X is
locally finite and hyperbolic, we show that the symmetric difference between
Geo(x, η) and Geo(y, η) is always finite, for x, y ∈ X and η ∈ ∂X. This gives
a positive answer to a question of Huang, Sabok and Shinko in the setting of
buildings. Combining their results with a construction of Bourdon, we obtain
examples of hyperbolic groups G with Kazhdan’s property (T) such that the
G-action on its Gromov boundary is hyperfinite.
1. Introduction
This paper is motivated by a question of Huang, Sabok and Shinko ([HSS17,
Question 1.5]), asking whether in a proper and cocompact hyperbolic space X,
the symmetric difference between two geodesic ray bundles pointing in the same
direction is always finite (see below for precise definitions).
This is motivated by the study of Borel equivalence relations for the action of
a hyperbolic group G on its Gromov boundary: the authors of [HSS17] give a
positive answer to the above question when X is a CAT(0) cube complex, and
deduce that if G is a hyperbolic cubulated group (namely, if G acts properly
and cocompactly on a CAT(0) cube complex), then the G-action on its Gromov
boundary ∂G is hyperfinite, that is, it induces a hyperfinite equivalence relation
([HSS17, Corollary 1.2]).
As it turns out, the answer to [HSS17, Question 1.5] is “no” in full generality:
counter-examples are constructed in [Tou17]. The purpose of this paper is to give a
positive answer to this question when X is a hyperbolic locally finite building. We
underline that the class of groups acting properly and cocompactly on hyperbolic
locally finite buildings includes groups with Kazhdan’s property (T), and is thus
significantly different from the class of cubulated hyperbolic groups considered in
[HSS17] (see the fixed point theorem [NR97]). We now give a precise statement
of our main result.
By a classical result of M. Davis ([Dav98]), any building ∆ can be realised as a
complete CAT(0) metric space (X, d), which can be viewed as a subcomplex of the
barycentric subdivision of the standard geometric realisation of ∆. Let X (0) ⊆ X
denote the set of barycenters of chambers of X, that is, X (0) is the 0-skeleton of
the chamber graph of ∆. The boundary ∂X of X is the set of equivalence classes of
asymptotic geodesic rays in X (see Section 2 for precise definitions). We denote,
for each x ∈ X (0) and η ∈ ∂X, by Geo(x, η) ⊆ X (0) the union of all combinatorial
geodesic rays Γ = (xn )n∈N ⊆ X (0) (i.e. if xn is the barycenter of the chamber Cn ,
∗
F.R.S.-FNRS Postdoctoral Researcher.
1
2
T. MARQUIS
then (Cn )n∈N in an infinite minimal gallery in ∆) starting at x0 = x and pointing
towards η, in the sense that Γ is contained in a tubular neighbourhood of some
geodesic ray towards η. The sets Geo(x, η) are called geodesic ray bundles.
In this paper, we give a description of geodesic ray bundles in arbitrary buildings
(see Section 3 and Proposition 4.1). When the building X is (Gromov) hyperbolic
and locally finite, we deduce from this description the following theorem.
Theorem A. Let X be a locally finite hyperbolic building. Let x, y ∈ X (0) and let
η ∈ ∂X. Then the symmetric difference of Geo(x, η) and Geo(y, η) is finite.
As an immediate consequence of Theorem A and of [HSS17, Theorem 1.4], we
deduce the following corollary.
Corollary B. Let G be a group acting cocompactly on a locally finite hyperbolic
building X, and assume that G acts freely on the chambers of X. Then the natural
action of G on its Gromov boundary is hyperfinite.
In [Bou00, §1.5.3] (see also [Świ01]), M. Bourdon constructs a family of groups
G with property (T) acting cocompactly on some hyperbolic building X. These
groups G are defined as fundamental groups of some complexes of groups (a standard reference to this topic is [BH99]), and it follows straightaway from the form
of the complexes of groups involved that G acts freely on the set of chambers of
X and that X is locally finite. Another example of such a group with an explicit
short presentation also recently appeared in [Cap17].
In particular, Corollary B yields examples of hyperbolic groups with property
(T) whose boundary action is hyperfinite.
Corollary C. There exist (infinite) hyperbolic groups G with property (T) such
that the G-action on its Gromov boundary is hyperfinite.
Note that any group with property (T) that acts on a CAT(0) cube complex
has a global fixed point ([NR97]). In particular, Theorem A covers situations that
are not covered by [HSS17] (see also the last paragraph in the introduction of
[HSS17]).
Acknowledgement. I would like to thank Pierre-Emmanuel Caprace for bringing [HSS17, Question 1.5] to my attention and for suggesting to explore it in the
context of buildings. I would also like to thank the anonymous referee for his/her
precious comments.
2. Preliminaries
2.1. CAT(0)-spaces and Gromov hyperbolic spaces. The standard reference for this paragraph is [BH99].
Let (X, d) be a complete CAT(0)-space, namely, a complete geodesic metric
space in which every triangle ∆ is at least as thin as the corresponding triangle
∆0 in Euclidean space E with same side lengths, in the sense that any two points
x, y of ∆ are at distance at most dE (x0 , y 0 ) from one another, where x0 , y 0 are the
points on ∆0 corresponding to x, y respectively, and dE is the Euclidean distance
on E.
Given two points x, y ∈ X, there is a unique geodesic segment from x to y, which
we denote [x, y]. A geodesic ray based at x ∈ X is an isometry r : R≥0 → X with
r(0) = x. Two geodesic rays r, r0 are called asymptotic if supt∈R≥0 d(r(t), r0 (t)) <
ON GEODESIC RAY BUNDLES IN BUILDINGS
3
∞. Equivalently, identifying r, r0 with their image in X, they are asymptotic if
they are at bounded Hausdorff distance from one another, that is, if r (resp.
r0 ) is contained in a tubular neighbourhood of r0 (resp. r). We recall that a
tubular neighbourhood of a subset S of X is just an -neighbourhood of S for
some > 0. The boundary of X, denoted ∂X, is the set of equivalence classes
[r] of geodesic rays r ⊆ X, where two geodesic rays are equivalent if they are
asymptotic. We then say that the geodesic ray r points towards η := [r] ∈ ∂X.
For each x ∈ X and η ∈ ∂X, there is a unique geodesic ray starting at x and
pointing towards η, which we denote [x, η).
The space X is called (Gromov) hyperbolic if there is some δ > 0 such that
each triangle ∆ in X is δ-slim, in the sense that each side of ∆ is contained in a
δ-neighbourhood of the other two (X is then also called δ-hyperbolic). Hyperbolic spaces can be thought of as fattened versions of trees, and their behavior is
somehow opposite to that of a Euclidean space: if the CAT(0) space X is proper
(every closed ball of X is compact) and cocompact (there is some compact subset C ⊆ X such that Isom(X).C = X), then X is hyperbolic if and only if it does
not contain a subspace isometric to the Euclidean plane.
There is a notion of Gromov boundary of a hyperbolic space; in the context
of CAT(0) spaces, it coincides with the boundary defined above, endowed with
the cone topology (see [BH99, II.8]).
2.2. Buildings. The standard reference for this paragraph is [AB08].
Let ∆ be a building, viewed as a simplicial complex (see [AB08, Chapter 4]).
Let Ch(∆) denote the set of chambers (i.e. maximal simplices) of ∆. A panel
is a codimension 1 simplex of ∆. Two chambers are adjacent if they share a
common panel. A gallery between two chambers C, D ∈ Ch(∆) is a sequence
Γ = (C0 = C, C1 , . . . , Ck = D) of chambers such that Ci−1 and Ci are distinct
and adjacent for each i = 1, . . . , k. The integer k is called the length of Γ. If Γ
is a gallery of minimal length between C and D, it is called a minimal gallery
and its length is denoted dCh (C, D). The map dCh : Ch(∆) × Ch(∆) → N is then
a metric, called the chamber distance on ∆. An infinite sequence Γ = (Ci )i∈N
of chambers is called a minimal gallery if (C0 , . . . , Cn ) is a minimal gallery for
each n ∈ N. Any such Γ is contained in an apartment A of ∆.
Let A be an apartment of ∆ and let C, D ∈ Ch(A) be distinct adjacent chambers
in A. Then no chamber of A is at equal (chamber) distance from C and D;
˙
this yields a partition Ch(A) = Φ(C, D)∪Φ(D,
C), where Φ(C, D) is the set of
chambers that are closer to C than to D. The subcomplexes of A with underlying
chamber sets Φ(C, D) and Φ(D, C) are called half-spaces or roots, and their
intersection is called the wall separating C from D. If m is a wall delimiting the
half-spaces Φ+ , Φ− of A, we say that two subsets S+ ⊆ Φ+ and S− ⊆ Φ− are
separated by m.
A gallery Γ = (C0 , . . . , Ck ) (resp. Γ = (Ci )i∈N ) contained in an apartment A
is said to cross a wall m of A if m is the wall separating Ci−1 from Ci for some
i ∈ {1, . . . , k} (resp. i ∈ N>0 ). The gallery Γ ⊆ A is then minimal if and only if
it crosses each wall of A at most once. Moreover, if C, D ∈ Ch(A), then the set
of walls crossed by a minimal gallery Γ from C to D depends only on C, D, i.e. it
is independent of the choice of Γ.
If A is an apartment of ∆ and C ∈ Ch(A), there is a simplicial map ρA,C : ∆ →
A, called the retraction onto A centered at C, with the following properties:
4
T. MARQUIS
ρA,C is the identity on A and its restriction to any apartment A0 containing C
is an isomorphism, with inverse ρA0 ,C |A : A → A0 (in particular, ρA,C preserves
the minimal galleries from C). Moreover, ρA,C does not increase the distance:
dCh (ρA,C (D), ρA,C (E)) ≤ dCh (D, E) for all D, E ∈ Ch(∆).
The set of all panels of ∆ is denoted Res1 (∆). The star of a panel σ ∈ Res1 (∆),
denoted St(σ), is the set of chambers containing σ. For any panel σ ∈ Res1 (∆)
and any chamber C ∈ Ch(∆), there is a unique chamber C 0 in St(σ) minimising
the gallery distance from C to St(σ); it is called the projection of C on σ and
is denoted projσ (C) := C 0 . It has the following gate property: dCh (C, D) =
dCh (C, C 0 ) + dCh (C 0 , D) for all D ∈ St(σ).
The building ∆ is called locally finite if St(σ) is a finite set of chambers for
each σ ∈ Res1 (∆).
2.3. Davis realisation of a building. The standard reference for this paragraph
is [AB08, Chapter 12] (see also [Dav98]).
Let ∆ be a building. Then ∆ admits a CAT(0)-realisation (X, d), called the
Davis realisation of ∆, which is a complete CAT(0) space. It can be viewed as
a subcomplex of the barycentric subdivision of the standard geometric realisation
of ∆, and contains the barycenter of each chamber and panel of ∆. In the sequel,
we will often identify ∆ with its Davis realisation X, and all related notions
(apartment, chamber, panel, gallery, wall,. . . ) with their realisation in X (viewed
as closed subspaces of X).
We set
X (0) := {xC | C ∈ Ch(∆) = Ch(X)} ⊆ X,
where xC ∈ X denotes the barycenter of the chamber C. If A is an apartment of
X, we also set A(0) := A ∩ X (0) .
A combinatorial path between xC , xD ∈ X (0) is a piecewise geodesic path
Γ ⊆ X which is the union of the geodesic segments [xCi−1 , xCi ] (i = 1, . . . , k) for
some gallery (C = C0 , C1 , . . . , Ck = D); we then write Γ = (xCi )0≤i≤k . Thus
combinatorial paths in X are in bijection with galleries in ∆. A combinatorial
geodesic is a combinatorial path corresponding to a minimal gallery in ∆. One
defines similarly infinite combinatorial paths and combinatorial geodesic rays
(abbreviated CGR) by replacing galleries in ∆ with infinite galleries. If η ∈ ∂X,
then a combinatorial geodesic ray from x ∈ X (0) to η is a combinatorial
geodesic ray starting at x and at bounded Hausdorff distance from some (any)
geodesic ray pointing towards η. We denote by CGR(x, η) the set of CGR from
x ∈ X (0) to η ∈ ∂X. If Γxy is a combinatorial geodesic from some x ∈ X (0) to
some y ∈ X (0) , and if Γyz is a combinatorial geodesic (resp. ray) from y to some
z ∈ X (0) (resp. z ∈ ∂X), we denote by Γxy · Γyz the combinatorial path obtained
as the concatenation of Γxy and Γyz .
Each geodesic segment (resp. geodesic ray) of X is contained in some minimal
gallery, and hence also in some apartment A of X; in particular, ∂X is covered
by the boundaries ∂A of all apartments A of X. Conversely, the uniqueness of
geodesic rays implies that if x ∈ A and η ∈ ∂A for some apartment A of X,
then [x, η) ⊆ A. Of course, any combinatorial geodesic ray is also contained in
some apartment of X. For every apartment A and x = xC ∈ A(0) , the retraction
ρA,C : ∆ → A induces a retraction ρA,x : X → A with the same properties as the
ON GEODESIC RAY BUNDLES IN BUILDINGS
5
ones described in §2.2. Moreover, d(ρA,x (y), ρA,x (z)) ≤ d(y, z) for all y, z ∈ X,
with equality if y belongs to the closed chamber C ⊆ X.
Let A be an apartment of X. Here are a few important properties of walls in
A, which can be found in [Nos11] (see also [NV02]). A wall m of A that intersects
a geodesic (resp. geodesic ray) in more than one point entirely contains that
geodesic (resp. geodesic ray); in particular, m is convex. The subset A\m of A has
two connected components (the open half-spaces corresponding to m), and those
components are convex. As we saw in §2.2, a combinatorial path Γ = (xCi )0≤i≤k
(resp. Γ = (xCi )i∈N ) contained in A is a combinatorial geodesic (resp. a CGR) if
and only if it crosses each wall of A at most once.
Note that X is a proper CAT(0) space if and only if it is locally finite. The
building X is called hyperbolic if it is hyperbolic in the sense of §2.1 when
equipped with the CAT(0) metric d. Equivalently, X is hyperbolic if and only if
(A, d) is hyperbolic for some (resp. for each) apartment A of X, as readily follows
from the properties of retractions onto apartments. Note that Moussong gave a
characterisation of the hyperbolicity of X in terms of the type (W, S) of ∆ (see
[Mou88, Theorem 17.1]): X is hyperbolic if and only if (WJ := hJi, J) is not an
affine Coxeter system whenever |J| ≥ 3, and there is no pair of disjoint subsets
I, J ⊆ S such that WI and WJ are infinite and commute. The only fact that we
will need about hyperbolic buildings, however, is the following.
Lemma 2.1. Assume that the building X is hyperbolic. Then there is a constant
K > 0 such that for any x ∈ X (0) and η ∈ ∂X, any Γ ∈ CGR(x, η) is contained
in a K-neighbourhood of [x, η).
Proof. By [BH99, Proposition I.7.31], combinatorial geodesics are quasi-geodesics
(for the CAT(0) metric d), so that the lemma follows from [BH99, Theorem III.1.7].
Here is also a basic useful fact about combinatorial geodesic rays.
Lemma 2.2. Let x ∈ X (0) , and let Γ = (xn )n∈N be a CGR. Then there exists
some k ∈ N such that Γxxk · (xn )n≥k is a CGR for any combinatorial geodesic Γxxk
from x to xk .
Proof. Reasoning inductively, we may assume that x and x0 are adjacent. If
(x, x0 ) · Γ is a CGR, the claim is clear with k = 0. Otherwise, there is some m ≥ 1
such that the combinatorial path (x, x0 , . . . , xm ) is not a combinatorial geodesic,
so that dCh (x, xm ) ≤ m. Let Γxxm be any combinatorial geodesic from x to xm .
If Γxxm · (xn )n≥m is a CGR, we are done with k = m. Otherwise, there is some
k > m such that the combinatorial path Γxxm · (xm , . . . , xk ) is not a combinatorial
geodesic, so that dCh (x, xk ) ≤ k − 1. Let Γxxk be any combinatorial geodesic from
x to xk . We claim that Γxxk · (xn )n≥k is a CGR, yielding the lemma. Indeed,
otherwise there is some ` > k such that the combinatorial path Γxxk · (xn )k≤n≤` is
not a combinatorial geodesic, and hence
`−1 = dCh (x0 , x` )−1 ≤ dCh (x, x` ) < dCh (x, xk )+dCh (xk , x` ) ≤ k−1+`−k = `−1,
a contradiction.
3. Combinatorial bordification of a building
In this section, we recall the notion of combinatorial bordification of a building
introduced in [CL11], and relate it to the notions introduced in Section 2.
6
T. MARQUIS
Let ∆ be a building, as in §2.2. Recall that for each panel σ ∈ Res1 (∆), we have
a projection map projσ : Ch(∆) → St(σ) ⊆ Ch(∆) associating to each chamber
C the unique chamber of St(σ) closest to C. This defines an injective map
πCh : Ch(∆) →
Y
St(σ) : C 7→ σ 7→ projσ (C) .
σ∈Res1 (∆)
Q
We endow σ∈Res1 (∆) St(σ) with the product topology, where each star St(σ) is a
discrete set of chambers. The (minimal) combinatorial bordification of ∆ is
then defined as the closure
Y
C1 (∆) := πCh (Ch(∆)) ⊆
St(σ).
σ∈Res1 (∆)
Since πCh is injective, we may identify Ch(∆) with a subset of C1 (∆), and it
thus makes sense to say that a sequence of chambers (Cn )n∈N converges to some
ξ : Res1 (∆) → Ch(∆) in C1 (∆). If ∆ is reduced to a single apartment A, this
notion of convergence is transparent: (Cn )n∈N converges in C1 (A) if and only if
for every wall m of A, the sequence (Cn )n∈N eventually remains on the same side
of m. On the other hand, back to a general ∆, one can identify C1 (A) (A an
apartment) with the subset of C1 (∆) consisting of limits of sequences of chambers
in A, and in fact (see [CL11, Proposition 2.4])
C1 (∆) =
[
C1 (A).
A apartment
Let C ∈ Ch(∆), and let (Cn )n∈N be a sequence of chambers converging to some
ξ ∈ C1 (∆). We define the combinatorial sector based at C and pointing towards
ξ as
[ \
Q(C, ξ) :=
Conv(C, Cn ),
k≥0 n≥k
where Conv(C, Cn ) denotes the union of all minimal galleries from C to Cn . Then
Q(C, ξ) indeed only depends on C and ξ (and not on the choice of sequence
(Cn )n∈N converging to ξ), and is contained in some apartment. Note also that if
C 0 ∈ Ch(∆) is contained in Q(C, ξ), then
Q(C 0 , ξ) ⊆ Q(C, ξ).
Example 3.1. Let X be a building of type Ae2 . The apartments of X are then
Euclidean planes tesselated by congruent equilateral triangles. If A is an apartment of X, its bordification C1 (A) consists of 6 “lines of points” and 6 “isolated
points” (see [CL11, Example 2.6]), which can be seen as follows. Let x ∈ A(0) and
η ∈ ∂A.
If the direction η is non-singular, in the sense that [x, η) is not contained in a
tubular neighbourhood of any wall of A, then for any Γ = (xn )n∈N ∈ CGR(x, η)
contained in A, the sequence of (barycenters of) chambers (xn )n∈N converges in
C1 (A), to a unique ξ ∈ C1 (A). The sector Q(x, ξ) in A is shown on Figure 1.
If η is singular, that is, if [x, η) is contained in a tubular neighbourhood of
some wall of A, then the set of ξ ∈ C1 (A) obtained as above as the limit of
some Γ ∈ CGR(x, η) are the vertices of some simplicial line “at infinity” (see the
dashed line on Figure 2), and the combinatorial sectors Q(x, ξ) for ξ on this line
are represented on Figure 2.
ON GEODESIC RAY BUNDLES IN BUILDINGS
7
Figure 1. Non-singular direction
To see what combinatorial sectors look like, we relate them to the notions
introduced in §2.3. As in that paragraph, we identify ∆ with its Davis realisation
X. To avoid cumbersome notations, we also identify the chambers of ∆ with their
barycenters in X (i.e. Ch(∆) with X (0) ): this thus also identifies the notions of
minimal (resp. infinite) gallery and of combinatorial geodesic (resp. ray). For
each η ∈ ∂X and each apartment A of X, we let WηA denote the set of walls m of
A containing η in their boundary (i.e. m contains a geodesic ray towards η). We
also let Aη be the set of apartments of X with η ∈ ∂A.
For η ∈ ∂X, we next define an equivalence relation ∼η on X (0) as follows. For
x, y ∈ X (0) distinct adjacent chambers, we write x ≈η y if, for any apartment A
containing x and y, the wall of A separating x from y does not belong to WηA .
We also write x ≈η x for x ∈ X (0) , so that ≈η becomes a symmetric and reflexive
relation on X (0) . We then let ∼η be the transitive closure of ≈η . For any x ∈ X (0) ,
we now let
Φη (x) ⊆ X
be the subcomplex of X obtained as the union of all chambers y ∈ X (0) with
y ∼η x. Note that
x ∼η y ⇐⇒ y ∈ Φη (x) ⇐⇒ Φη (y) = Φη (x) for any y ∈ X (0) .
We start by making some useful observations about the relation ∼η .
Lemma 3.2. Let x, y ∈ X (0) and η ∈ ∂X, and assume that there exists an
apartment of Aη containing x, y. Then x ≈η y if and only if there exists an
apartment A ∈ Aη containing x, y such that the wall of A separating x from y
does not belong to WηA .
Proof. The implication ⇒ is clear. Conversely, let A ∈ Aη be an apartment
containing x, y and such that the wall m of A separating x from y does not
belong to WηA , and assume for a contradiction that there is an apartment A0 ∈ Aη
0
containing x, y such that the wall m0 of A0 separating x from y belongs to WηA . Let
z ∈ A∩A0 be the barycenter of the common panel of x and y. Then [z, η) ⊆ A∩A0 .
By definition of buildings, there is a simplicial isomorphism φ : A0 → A fixing
8
T. MARQUIS
A ∩ A0 pointwise. Since [z, η) ⊆ m0 by assumption and φ(m0 ) = m, we deduce
that [z, η) = φ([z, η)) ⊆ m. Hence m ∈ WηA , a contradiction.
The next lemma introduces some important terminology and notations. For
x ∈ X (0) and η ∈ ∂X, we call a CGR Γ ∈ CGR(x, η) straight if the infinite
gallery corresponding to Γ contains the geodesic ray [x, η).
Lemma 3.3. Let x ∈ X (0) and η ∈ ∂X. Then the following assertions hold:
(1) Let Γ = (xn )n∈N ∈ CGR(x, η). Then the sequence of chambers (xn )n∈N
converges in C1 (X). We denote its limit by ξΓ ∈ C1 (X) \ X (0) and we say
that Γ converges to ξΓ .
(2) Let y ∈ X (0) , and let Γx ∈ CGR(x, η) and Γy ∈ CGR(y, η) be contained in
some apartment A. Then ξΓx = ξΓy if and only if Γx and Γy eventually lie
on the same side of any given wall m ∈ WηA .
(3) If Γ ∈ CGR(x, η) is straight, it is contained in Φη (x) and in every apartment A ∈ Aη with x ∈ A(0) . Moreover, ξx,η := ξΓ ∈ C1 (X) \ X (0) is
independent of the choice of a straight Γ ∈ CGR(x, η).
Proof. (1) Since the CGR Γ is contained in some apartment A, this readily follows
from the above description of convergence in C1 (A).
(2) We have to show that if Γx and Γy eventually lie on different sides of a wall
m of A, then m ∈ WηA . But if Γ0x ⊆ Γx and Γ0y ⊆ Γy are CGRs separated by m,
then [x, η) (which is contained in a tubular neighbourhood of Γ0x and Γ0y ) must be
contained in a tubular neighbourhood of m, as claimed.
(3) Let Γ ∈ CGR(x, η) be straight, and let A ∈ Aη with x ∈ A(0) . Thus
A also contains [x, η). Since [x, η) is not contained in any wall of A, we deduce
that A must contain infinitely many chambers of Γ, and hence also Γ by convexity.
Moreover, since [x, η) does not intersect any wall in WηA , the CGR Γ does not cross
any wall in WηA , and hence Γ ⊆ Φη (x) by Lemma 3.2. Finally, if Γ0 ∈ CGR(x, η)
is straight, then both Γ and Γ0 are contained in a common apartment A ∈ Aη by
the above discussion, and hence ξΓ = ξΓ0 by (2).
We next give an alternative description of the sets Φη (x).
Proposition 3.4. Let x ∈ X (0) and η ∈ ∂X. Then
Φη (x) ∩ X (0) = {y ∈ X (0) | ξx,η = ξy,η }.
Proof. Let y ∈ X (0) . We have to show that x ∼η y if and only if ξx,η = ξy,η .
Assume first that x ∼η y. Reasoning inductively on the length of a gallery (x =
x0 , x1 , . . . , xn = y) from x to y such that xi−1 ≈η xi for all i ∈ {1, . . . , n}, there is
no loss of generality in assuming that x ≈η y. Let Γx = (xn )n∈N ∈ CGR(x, η) be
straight. By Lemma 2.2, there is some k ∈ N and some combinatorial geodesic
Γyxk from y to xk such that Γyxk η := Γyxk · (xn )n≥k ∈ CGR(y, η). Let A ∈ Aη be
an apartment containing Γyxk η . Let Γy ∈ CGR(y, η) be straight, so that Γy ⊆ A
by Lemma 3.3(3). We claim that Γyxk does not cross any wall in WηA : this will
imply that Γyxk η and Γy do not cross any wall in WηA , and hence that
ξy,η = ξΓy = ξΓyxk η = ξΓx = ξx,η
by Lemma 3.3(2), as desired. Indeed, if m ∈ WηA separates y from xk and if
A0 ∈ Aη is an apartment containing Γx , then the wall m0 := ρA0 ,xk (m) of A0
0
belongs to WηA (because ρA0 ,xk does not increase the distance, and hence [xk , η) =
ON GEODESIC RAY BUNDLES IN BUILDINGS
9
ρA0 ,xk ([xk , η)) ⊆ A ∩ A0 is contained in a tubular neighbourhood of both m and
m0 ) and separates xk from y 0 := ρA0 ,xk (y). But since m0 cannot be crossed by
(xn )0≤n≤k , it must separate the adjacent chambers x = ρA0 ,xk (x) and y 0 . Now, if α
is the half-apartment of A0 containing x and delimited by m0 , we find by [AB08,
00
Exercise 5.83(a)] an apartment A00 containing α and y. Then m0 ∈ WηA separates
x from y in A00 , contradicting our hypothesis x ≈η y.
Conversely, assume that ξx,η = ξy,η , and let us show that x ∼η y. Let Γx =
(xn )n∈N ∈ CGR(x, η) and Γy ∈ CGR(y, η) be straight. By Lemma 2.2, there
is some k ∈ N and some combinatorial geodesic Γyxk from y to xk such that
Γyxk η := Γyxk · (xn )n≥k ∈ CGR(y, η). Let A ∈ Aη be an apartment containing
Γyxk η . Then Γy ⊆ A by Lemma 3.3(3). Note that the walls of A separating xk
from y do not belong to WηA , for otherwise the CGRs (xn )n≥k and Γy (which do
not cross any wall in WηA ) would be separated by some wall of WηA , contradicting
our assumption that ξΓx = ξx,η = ξy,η = ξΓy . Hence x ∼η xk ∼η y by Lemma 3.2,
yielding the claim.
We conclude our round of observations about the relation ∼η with the following
consequences of Proposition 3.4.
Lemma 3.5. Let x ∈ X (0) and η ∈ ∂X. Then the following assertions hold:
(1) Let A ∈ Aη be an apartment containing x and let y ∈ A(0) . Then x ∼η y
if and only if the walls of A separating x from y do not belong to WηA .
(2) Let Γ ∈ CGR(x, η) and let A be an apartment containing Γ. Then Γ
converges to ξx,η ⇐⇒ the walls of A crossed by Γ do not belong to WηA
⇐⇒ Γ ⊆ Φη (x).
Proof. (1) The implication ⇐ follows from Lemma 3.2. Conversely, assume that
x ∼η y. Then ξx,η = ξy,η by Proposition 3.4, that is, ξΓx = ξΓy for some straight
Γx ∈ CGR(x, η) and Γy ∈ CGR(y, η). By Lemma 3.3(3), Γx and Γy are contained
in A. If now m ∈ WηA separates x from y then it also separates Γx from Γy and
hence ξΓx 6= ξΓy , a contradiction.
(2) This readilfy follows from (1).
To get a better understanding of combinatorial sectors, we first show that, given
an element ξ ∈ C1 (X) \ X (0) , one can choose a sequence of chambers (xn )n∈N of
X converging to ξ in a nice and controlled way. Here, by “nice” we mean that
Γ := (xn )n∈N can be chosen to be a CGR, and by “controlled” we mean that we
may impose further restrictions on Γ.
Lemma 3.6. Let ξ ∈ C1 (X) \ X (0) . Then there is some η ∈ ∂X and some straight
Γ = (yn )n∈N ∈ CGR(y0 , η) such that ξΓ = ξ.
Proof. Let A be an apartment of X with ξ ∈ C1 (A). Let (xn )n∈N be a sequence
of chambers of A converging to ξ. Since the space A is proper, the sequence of
geodesic segments ([x0 , xn ])n∈N subconverges to some geodesic ray r0 := [x0 , η) for
some η ∈ ∂A. In other words, up to extracting a subsequence, we may assume
that (xn )n∈N is contained in an -neighbourhood N (r0 ) ⊆ A of r0 for some > 0.
We claim that there exists a finite subset S ⊆ WηA such that for each m ∈
WηA \ S, the neighbourhood N (r0 ) is entirely contained in one of the half-spaces
delimited by m. Indeed, any wall in WηA intersecting N (r0 ), say in y, contains
the geodesic ray [y, η) ⊆ N (r0 ) (see e.g. [BH99, Theorem II.2.13]), and hence a
10
T. MARQUIS
Figure 2. Singular direction
subray of r0 in an -neighbourhood. On the other hand, since A is locally finite,
there is some N ∈ N such that any ball of radius in A intersects at most N walls.
In particular, there are at most N walls intersecting N (r0 ), whence the claim.
Recall that for any wall m of A, the sequence (xn )n∈N eventually remains on
the same side of m; in particular, there is some k ∈ N such that (xn )n≥k is entirely
contained in some half-space associated to m for each m ∈ S. Hence for any n ≥ k,
the walls separating xk from xn do not lie in WηA . Let Γ = (yn )n∈N ∈ CGR(xk , η)
be straight (thus y0 = xk ). Then by Lemma 3.3(3), we know that Γ ⊆ Φη (xk ) ∩ A.
Since for any n ≥ k and any wall m ∈ WηA , the chambers xn , xk , yn all lie on
the same side of m, we conclude as in the proof of Lemma 3.3(2) that ξΓ = ξ, as
desired.
Proposition 3.7. Let x ∈ X (0) and ξ ∈ C1 (X) \ X (0) . Then
Q(x, ξ) = {y ∈ X (0) | y is on a CGR starting from x and converging to ξ}.
Proof. Let Γ = (xn )n∈N be a CGR with x0 = x and converging to ξ. To prove
the inclusion ⊇, we have to show that for any k ∈ N, the chamber xk belongs to
Q(x, ξ). But as xk ∈ Conv(x, xn ) for every n ≥ k, this is clear.
Conversely, let y ∈ Q(x, ξ). By Lemmas 2.2 and 3.6, there exists a CGR
Γ = (xn )n∈N starting from x and converging to ξ. Let n ∈ N be such that
y ∈ Conv(x, xn ). Since replacing the portion of Γ between x and xn by some
combinatorial geodesic from x to xn passing through y still yields a CGR, the
lemma follows.
We next wish to prove a refinement of Proposition 3.7 by relating the combinatorial bordification to the visual boundary of X.
ON GEODESIC RAY BUNDLES IN BUILDINGS
11
For η ∈ ∂X, we define the transversal graph to X in the direction η as the
graph X η with vertex set
Cη := {ξx,η | x ∈ X (0) } ⊆ C1 (X) \ X (0)
and such that ξx,η , ξy,η ∈ Cη with ξx,η 6= ξy,η are adjacent (i.e. connected by an
edge) if and only if there exist adjacent chambers x0 , y 0 ∈ X (0) such that ξx0 ,η = ξx,η
and ξy0 ,η = ξy,η . The elements of Cη will also be called chambers, and we define
the notions of galleries and chamber distance in X η as in §2.2. Note that by
Lemma 3.6, any ξ ∈ C1 (X) \ X (0) is of the form ξ = ξx,η for some x ∈ X (0) and
η ∈ ∂X, that is,
[
C1 (X) = X (0) ∪
Cη .
η∈∂X
Example 3.8. In the context of Example 3.1, assume that X consists of a single
apartment A and that η ∈ ∂A is a singular direction. Then X η is a simplicial line
(the dashed line on Figure 2). The Φη (y) for y ∈ A(0) are stripes obtained as the
convex hull of two “adjacent” walls of WηA , namely, of walls of A in the direction
of η.
Here is another description of Cη , in terms of CGR’s.
Proposition 3.9. Let x ∈ X (0) and η ∈ ∂X. Then
Cη = {ξ ∈ C1 (X) \ X (0) | ξ = ξΓ for some Γ ∈ CGR(x, η)}.
Proof. The inclusion ⊆ is clear. Conversely, if ξ = ξΓ for some Γ = (xn )n∈N ∈
CGR(x, η), then by Lemma 3.10 below, there is some k ∈ N such that Γ0 :=
(xn )n≥k ⊆ Φη (xk ). Hence ξ = ξΓ0 = ξxk ,η ∈ Cη by Lemma 3.5(2).
Lemma 3.10. Let x ∈ X (0) and η ∈ ∂X. For any Γ = (xn )n∈N ∈ CGR(x, η),
the sequence (ξxn ,η )n∈N ⊆ Cη is eventually constant. In other words, there is some
k ∈ N such that (xn )n≥k ⊆ Φη (xk ).
Proof. Let A be an apartment containing Γ (thus A ∈ Aη ), and assume for a
contradiction that (ξxn ,η )n∈N ⊆ Cη is not eventually constant. Thus there are
infinitely many walls {mi | i ∈ N} ⊆ WηA crossed by Γ (see Lemma 3.5).
Since [x, η) ⊆ A is not contained in any wall, it does not intersect any wall mi ,
i ∈ N. On the other hand, Γ is contained in an -neighbourhood of [x, η) for some
> 0. For each y ∈ Γ, let y 0 ∈ [x, η) with d(y, y 0 ) ≤ . Then there is some n ∈ N
such that [y, y 0 ] intersects at most n walls of A for any y ∈ Γ. Now, let y ∈ Γ∩A(0)
be such that the walls m1 , . . . , mn+1 intersect the combinatorial geodesic from x
to y contained in Γ. Then the walls m1 , . . . , mn+1 must all intersect the geodesic
segment [y, y 0 ], yielding the desired contradiction.
We can now formulate the announced refinement of Proposition 3.7.
Theorem 3.11. Let x ∈ X (0) and ξ ∈ Cη for some η ∈ ∂X. Then
Q(x, ξ) = {y ∈ X (0) | y is on a CGR from x to η and converging to ξ}.
Proof. The inclusion ⊇ follows from Proposition 3.7. The converse inclusion is
proved exactly as in the proof of Proposition 3.7, the existence of a CGR from x
to η and converging to ξ following from Proposition 3.9.
12
T. MARQUIS
Given x ∈ X (0) and η ∈ ∂X, we next wish to show that the combinatorial sector
Q(x, ξx,η ) is “minimal” in the direction η, in the sense that it is contained in every
other combinatorial sector Q(x, ξ) with ξ ∈ Cη (see Proposition 3.13 below). To
this end, we first need a more precise version of Lemma 3.6, by further improving
our control of CGR’s converging to a given ξ ∈ C1 (X) \ X (0) .
Lemma 3.12. Let x ∈ X (0) and η ∈ ∂X. If y ∈ X (0) is on some CGR from x to
η, then y is on some Γ ∈ CGR(x, η) converging to ξy,η .
Proof. Let Γxyη be a CGR from x to η passing through y, and let Γxy (resp. Γyη )
be the combinatorial geodesic from x to y (resp. CGR from y to η) contained in
Γxyη , so that Γxyη = Γxy · Γyη . Let A be an apartment containing Γxyη , and let
Γyξ ⊆ A be a straight CGR from y to η and converging to ξ := ξy,η . In particular,
Γyξ does not cross any wall in WηA . We claim that Γ := Γxy ·Γyξ is a CGR, yielding
the lemma.
Otherwise, there is a wall m of A that is crossed by both Γxy and Γyξ . Since
m cannot be crossed by Γyη , it separates Γyη from Γy0 ξ for some CGR Γy0 ξ to
η contained in Γyξ . Since Γy0 ξ and Γyη are at bounded Hausdorff distance from
one another, this implies that m ∈ WηA , and hence that Γyξ cannot cross m, a
contradiction.
Proposition 3.13. Let x ∈ X (0) , η ∈ ∂X and ξ ∈ Cη . Then
Q(x, ξx,η ) = Q(x, ξ) ∩ Φη (x) = {y ∈ Q(x, ξ) | ξy,η = ξx,η }.
Proof. Note that the second equality holds by Proposition 3.4. Let y ∈ Q(x, ξ) ∩
Φη (x). Then y lies on some CGR from x to η by Theorem 3.11, and hence also
on some Γ ∈ CGR(x, η) converging to ξy,η by Lemma 3.12. But since ξy,η = ξx,η
by Proposition 3.4, we then have y ∈ Q(x, ξx,η ) by Theorem 3.11.
Conversely, if y ∈ Q(x, ξx,η ) then certainly y ∈ Φη (x) by Theorem 3.11 and
Lemma 3.5(2), and it remains to show that y ∈ Q(x, ξ). Let A be an apartment
containing Q(x, ξ) (thus A ∈ Aη ).
Note first that A contains Q(x, ξx,η ). Indeed, let Γ = (xn )n∈N ∈ CGR(x, η)
be straight, Sso that
Γ converges to ξx,η and Γ ⊆ A by Lemma 3.3(3). Then
T
Q(x, ξx,η ) = k∈N n≥k Conv(x, xn ) ⊆ A, as claimed.
Let now Γxyη be a CGR from x to η passing through y and converging to ξx,η
(see Theorem 3.11), and let us show that y is also on a CGR from x to η converging
to ξ (and hence that y ∈ Q(x, ξ) by Theorem 3.11, as desired).
Write Γxyη = Γxy · Γyη , where Γxy is a combinatorial geodesic from x to y and
Γyη a CGR from y to η. By Lemma 2.2, there is a CGR Γyξ ⊆ A from y to η
converging to ξ. We claim that Γxy ·Γyξ is still a CGR, as desired. Otherwise, there
is a wall m of A that is crossed by both Γxy and Γyξ . Since m cannot be crossed
by Γyη , it separates Γyη from Γy0 ξ for some CGR Γy0 ξ contained in Γyξ . Since Γy0 ξ
and Γyη are at bounded Hausdorff distance from one another, this implies that
m ∈ WηA , so that Γxyη cannot cross m, a contradiction.
To conclude this section, we give a consequence of hyperbolicity for the building
X in terms of the sets Cη .
Lemma 3.14. Assume that the building X is hyperbolic. Then Cη is a bounded
set of chambers for each η ∈ ∂X. In particular, if X is, moreover, locally finite,
then Cη is finite for all η ∈ ∂X.
ON GEODESIC RAY BUNDLES IN BUILDINGS
13
Proof. Let K be as in Lemma 2.1. Note that there exist constants δ1 ≥ 1 and
δ2 ≥ 0 such that
1
dCh (y, z) − δ2 ≤ d(y, z) for all y, z ∈ X (0)
δ1
(see [BH99, Proposition I.7.31]). We also let δ3 > 0 be such that for any x ∈ X (0) ,
the closed chamber C of X of which x is the barycenter is contained in a δ3 neighbourhood of x (see §2.3). We fix some N ∈ N such that N ≥ δ1 (2K +δ2 +δ3 ).
Let x ∈ X (0) and η ∈ ∂X. We claim that any Γ = (xn )n∈N ∈ CGR(x, η) crosses
at most N walls in WηA , where A is any apartment containing Γ (hence η ∈ ∂A).
This will imply that the chambers {ξxn ,η | n ∈ N} of X η are at gallery distance at
most N from ξx,η , and hence the lemma will follow from Proposition 3.9.
Let thus Γ = (xn )n∈N ∈ CGR(x, η) and let A be an apartment containing Γ.
Let Γ0 = (yn )n∈N ⊆ A be a straight CGR from x to η. By Lemma 2.1, we know
that Γ and Γ0 are contained in a 2K-neighbourhood of one another. Assume for a
contradiction that the combinatorial geodesic (xn )0≤n≤k crosses N +1 walls in WηA
for some k ∈ N. Let ` ∈ N be such that d(xk , y` ) ≤ 2K +δ3 . Thus dCh (xk , y` ) > N .
But then
N
1
2K + δ3 ≤
− δ2 < dCh (xk , y` ) − δ2 ≤ d(xk , y` ) ≤ 2K + δ3 ,
δ1
δ1
a contradiction.
Remark 3.15. Note that, although Lemma 3.14 will be sufficient for our purpose,
it is not hard to see (using Moussong’s characterisation of hyperbolicity for Coxeter
groups, see [Mou88, Theorem 17.1]) that its converse also holds: the building X
is hyperbolic if and only if each transversal graph X η (η ∈ ∂X) is bounded.
4. Geodesic ray bundles in buildings
Throughout this section, we let X be a building, identified with its Davis realisation as in Section 3, and we keep all notations introduced in Sections 2 and 3.
We also fix some η ∈ ∂X.
We denote for each x ∈ X (0) by Geo(x, η) the ray bundle from x to η, that is,
Geo(x, η) := {y ∈ X (0) | y lies on some Γ ∈ CGR(x, η)}.
The description of combinatorial sectors provided in Section 3 then yields the
following description of ray bundles.
Proposition 4.1. Geo(x, η) =
S
ξ∈Cη
Q(x, ξ).
Proof. This inclusion ⊇ is clear by Theorem 3.11. Conversely, if Γ ∈ Geo(x, η),
then Γ converges to some ξ ∈ Cη by Proposition 3.9, and Γ ⊆ Q(x, ξ) by Theorem 3.11, yielding the converse inclusion.
We first establish Theorem A inside an apartment. If x ∈ A(0) for some apartment A ∈ Aη , we set
GeoA (x, η) := {y ∈ A(0) | y lies on some Γ ∈ CGR(x, η) with Γ ⊆ A}.
Lemma 4.2. Let A ∈ Aη . Then for any x ∈ A(0) ,
GeoA (x, η) = Geo(x, η) ∩ A.
14
T. MARQUIS
Proof. The inclusion ⊆ is clear. For the converse inclusion, we have to show that
if y ∈ A(0) lies on some Γ ∈ CGR(x, η), then it also lies on some Γ0 ∈ CGR(x, η)
with Γ0 ⊆ A. But we may take Γ0 = ρA,x (Γ), because ρA,x preserves combinatorial
geodesic rays from x and does not increase the distance.
Lemma 4.3. Let A ∈ Aη and let x ∈ A(0) . If y ∈ GeoA (x, η), then GeoA (y, η) ⊆
GeoA (x, η).
Proof. Reasoning inductively on dCh (x, y), we may assume that x and y are
adjacent. Let m be the wall of A separating x from y. Let z ∈ GeoA (y, η) and let
us show that z ∈ GeoA (x, η).
By Lemma 3.12, we find some CGR Γyzξ ⊆ A from y to η going through z and
converging to ξ := ξz,η . Write Γzξ for the CGR from z to η contained in Γyzξ . Let
also Γxyη ⊆ A be a CGR from x to η going through y, and let Γyη be the CGR
from y to η contained in Γxyη . Finally, let Γxz be a combinatorial geodesic from x
to z. We claim that Γxz · Γzξ is a CGR, yielding the lemma.
Otherwise, there is some wall m0 of A crossed by Γxz and Γzξ . Then m0 cannot
separate z from y because Γyzξ is a CGR, and hence m0 = m. But as Γxyη is a
CGR, the CGR Γyη cannot cross m, so that m separates Γyη from some CGR Γz0 ξ
to η contained in Γzξ . Since Γyη and Γz0 ξ are at bounded Hausdorff distance, we
deduce that m ∈ WηA , and hence that Γzξ cannot cross m, a contradiction.
Lemma 4.4. Assume that X is hyperbolic. Let A ∈ Aη and let x, y ∈ A(0) . Then
the symmetric difference of GeoA (x, η) and GeoA (y, η) is finite.
Proof. Reasoning inductively on dCh (x, y), there is no loss of generality in assuming that x and y are adjacent. Assume for a contradiction that there is an
infinite sequence (yn )n∈N ⊆ GeoA (y, η) \ GeoA (x, η). Choose for each n ∈ N some
Γn ∈ CGR(y, η) passing through yn . Note that if Γn,≤yn denotes the combinatorial
geodesic from y to yn contained in Γn , then Γn,≤yn is disjoint from GeoA (x, η), for
if x0 ∈ GeoA (x, η) ∩ Γn,≤yn , then yn ∈ GeoA (x0 , η) ⊆ GeoA (x, η) by Lemma 4.3, a
contradiction.
Since A is locally finite, the sequence (Γn )n∈N then subconverges to a CGR
Γ = (xn )n∈N ⊆ A that is disjoint from GeoA (x, η). On the other hand, since A is
hyperbolic, Lemma 2.1 yields that Γ ∈ CGR(y, η). But then Lemma 2.2 implies
that xn ∈ GeoA (x, η) for all large enough n, a contradiction.
We now turn to the proof of Theorem A in the building X. For the rest of this
section, we assume that X is hyperbolic and locally finite, so that Cη is finite by
Lemma 3.14.
Lemma 4.5. Let x ∈ X (0) and ξ ∈ Cη , and let A be an apartment containing
Q(x, ξ). Let S be an infinite subset of Q(x, ξ). Then there is some z ∈ Q(x, ξ)
such that Q(z, ξz,η ) ∩ S is infinite.
Proof. Since Cη is finite, there is an infinite subset S1 of S and some ξ1 ∈ Cη
such that ξy,η = ξ1 for all y ∈ S1 . Let z ∈ S1 . By Lemma 4.4, we know that
Q(x, ξ) \ GeoA (z, η) is finite, and hence by Proposition 4.1, there is some infinite
subset S2 of S1 contained in Q(z, ξ2 ) for some ξ2 ∈ Cη . But then Proposition 3.13
implies that S2 ⊆ Q(z, ξz,η ) since ξz,η = ξ1 , as desired.
Lemma 4.6. Let x, y ∈ X (0) . Then Q(y, ξy,η ) \ Geo(x, η) is finite.
ON GEODESIC RAY BUNDLES IN BUILDINGS
15
Proof. Assume for a contradiction that there exists an infinite sequence (yn )n∈N ⊆
Q(y, ξy,η )\Geo(x, η). By [CL11, Proposition 2.30], there is some z ∈ X (0) such that
Q(z, ξy,η ) ⊆ Q(x, ξy,η ) ∩ Q(y, ξy,η ) (note that this amounts to say that Q(x, ξy,η ) ∩
Q(y, ξy,η ) is nonempty, which readily follows from Proposition 3.7 together with
Lemma 2.2). Let A be an apartment containing Q(y, ξy,η ).
By Lemma 4.4, we know that Q(y, ξy,η ) \ GeoA (z, η) is finite, and we may thus
assume, up to taking a subsequence, that (yn )n∈N ⊆ GeoA (z, η). Hence by Proposition 4.1, we may further assume, again up to extracting a subsequence, that
(yn )n∈N ⊆ Q(z, ξ) for some ξ ∈ Cη . Since ξz,η = ξy,η and (yn )n∈N ⊆ Φη (y)
by Proposition 3.13, this proposition implies that (yn )n∈N ⊆ Q(z, ξz,η ). But
Q(z, ξz,η ) ⊆ Q(x, ξz,η ) because z ∈ Q(x, ξy,η ) = Q(x, ξz,η ), yielding the desired
contradiction as Q(x, ξz,η ) ⊆ Geo(x, η) by Proposition 4.1.
Theorem 4.7. Let x, y ∈ X (0) . Assume that X is hyperbolic and locally finite.
Then Geo(y, η) \ Geo(x, η) is finite.
Proof. By Lemma 3.14, we know that Cη is finite. By Proposition 4.1, we have
to prove that if ξ ∈ Cη , then Q(y, ξ) \ Geo(x, η) is finite. Assume for a contradiction that there is an infinite sequence (yn )n∈N ⊆ Q(y, ξ) \ Geo(x, η). Then
by Lemma 4.5, up to extracting a subsequence, we may assume that (yn )n∈N ⊆
Q(z, ξz,η ) for some z ∈ Q(y, ξ). This contradicts Lemma 4.6.
Appendix A. Transversal buildings
Let X be a building and let η ∈ ∂X. In [CL11, Section 5], a construction of
“transversal building” X η to X in the direction η is given. However, as pointed
out to us by the referee, the premises of that construction are incorrect; the correct
construction of X η is the one given in §3 (where we called X η the “transversal
graph” to X in the direction η). Although we did not need this fact in our proof
of Theorem A, one can show that this transversal graph X η is indeed the chamber
graph of a building, and therefore deserves the name of transversal building to
X in the direction η. Since this fact is used in other papers, we devote the present
appendix to its proof. Here, we will follow the “W -metric approach” to buildings
(as opposed to the “simplicial approach” from §2.2); a standard reference for this
topic is [AB08, §5.1].
Let (W, S) be the type of X. Let A ∈ Aη , and view W as a reflection group
acting on A. Let Wη be the reflection subgroup of W generated by the reflections
across the walls in WηA . By a classical result of Deodhar ([Deo89]), Wη is then
itself a Coxeter group. Moreover, the polyhedral structure on A induced by the
walls in WηA can be identified with the Coxeter complex of Wη . More precisely,
if x ∈ A(0) is the fundamental chamber of A (i.e. the chamber whose walls are
associated to the reflections in S), then xη := Φη (x) ∩ A (which coincides with the
intersection of all half-spaces containing x and whose wall belong to WηA ) is the
fundamental chamber of the Coxeter complex associated to the Coxeter system
(Wη , Sη ), where Sη is the set of reflections across the walls in WηA that delimit xη .
Lemma A.1. The group Wη depends only on η, and not on the choice of apartment A.
Proof. This is [CL11, Lemma 5.2] (and the proof of this lemma in [CL11, §5.1]
remains valid in our context).
16
T. MARQUIS
Lemma A.2. Let ξ, ξ 0 ∈ Cη , and let x ∈ X (0) with ξ = ξx,η . Then there exists
some y ∈ X (0) with ξ 0 = ξy,η such that x and y are contained in some apartment
A ∈ Aη .
Proof. Let y ∈ X (0) with ξ 0 = ξy,η , and let Γy = (yn )n∈N ∈ CGR(y, η) be straight.
By Lemma 2.2, there is some k ∈ N such that Γxyk η := Γxyk · (yn )n≥k ∈ CGR(x, η)
for some combinatorial geodesic Γxyk from x to yk . Let A ∈ Aη be an apartment
containing Γxyk η . Then ξy,η = ξyk ,η , so that the claim follows by replacing y with
yk .
Theorem A.3. The transversal graph X η to X in the direction η is the graph of
chambers of a building of type (Wη , Sη ).
Proof. We define a “Weyl distance” function δη : Cη × Cη → Wη as follows. Let
ξ, ξ 0 ∈ Cη . By Lemma A.2, we can write ξ = ξx,η and ξ 0 = ξy,η for some x, y ∈ A(0)
where A ∈ Aη . We then set
δη (ξ, ξ 0 ) := δηA (xη , yη ),
where xη := Φη (x) ∩ A and yη := Φη (x) ∩ A are chambers in the Coxeter complex
of (Wη , Sη ) and δηA is the Weyl distance function on that complex (see [AB08,
§5.1]). Note that this definition is independent of the choice of apartment A (see
Lemma A.1 and its proof) and of chambers x, y ∈ A(0) such that ξ = ξx,η and ξ 0 =
ξy,η . To simplify the notations, we will also simply write δηA (x, y) := δηA (xη , yη ).
To check that X η is a building of type (Wη , Sη ), it then remains to check the
axioms (WD1), (WD2) and (WD3) of [AB08, Definition 5.1]. The axioms (WD1)
and (WD3) are clearly satisfied, because they are satisfied in the building (A, δηA )
for any apartment A ∈ Aη . We now check (WD2).
Let w ∈ Wη , s ∈ Sη and ξ1 , ξ2 , ξ10 ∈ Cη with δη (ξ1 , ξ2 ) = w and δη (ξ10 , ξ1 ) = s.
We have to show that δη (ξ10 , ξ2 ) ∈ {sw, w} and that δη (ξ10 , ξ2 ) = sw if `η (sw) =
`η (w) + 1, where `η : Wη → N is the length function on Wη with respect to the
generating set Sη . Choose some adjacent chambers x, x0 ∈ X (0) such that ξ1 = ξx,η
and ξ10 = ξx0 ,η . Let also y ∈ X (0) be such that ξ2 = ξy,η and such that there is
an apartment A ∈ Aη with x, y ∈ A(0) (see Lemma A.2). Let m be the wall of A
0
containing the s-panel of x. Since x and x0 are separated by a wall m0 ∈ WηA in
some apartment A0 ∈ Aη containing x, x0 , the wall m belongs to WηA (as it is the
image of m0 by the retraction ρA,x |A0 , which fixes the geodesic ray [x, η) ⊆ A ∩ A0
pointwise). On the other hand, by [AB08, Exercise 5.83(a)], there is an apartment
A00 ∈ Aη containing x0 and the half-space α of A delimited by m and that contains
y.
If `η (sw) = `η (w) + 1, then x ∈ α, and hence x, x0 , y ∈ A00 , so that
00
00
00
δη (ξ10 , ξ2 ) = δηA (x0 , y) = δηA (x0 , x)δηA (x, y) = δη (ξ10 , ξ1 )δη (ξ1 , ξ2 ) = sw.
On the other hand, if `η (sw) = `η (w) − 1, so that x ∈
/ α, then letting x00 ∈ A00
00
denote the unique chamber of A that is s-adjacent to x and contained in α, we
00
have δηA (x00 , y) = sw. Hence, in that case,
00
00
00
00
δη (ξ10 , ξ2 ) = δηA (x0 , y) = δηA (x0 , x00 )δηA (x00 , y) = δηA (x0 , x00 )sw ∈ {w, sw},
as desired.
ON GEODESIC RAY BUNDLES IN BUILDINGS
17
References
[AB08]
[BH99]
[Bou00]
[Cap17]
[CL11]
[Dav98]
[Deo89]
[HSS17]
[Mou88]
[Nos11]
[NR97]
[NV02]
[Świ01]
[Tou17]
Peter Abramenko and Kenneth S. Brown, Buildings, Graduate Texts in Mathematics,
vol. 248, Springer, New York, 2008, Theory and applications.
Martin R. Bridson and André Haefliger, Metric spaces of non-positive curvature,
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 319, Springer-Verlag, Berlin, 1999.
Marc Bourdon, Sur les immeubles fuchsiens et leur type de quasi-isométrie, Ergodic
Theory Dynam. Systems 20 (2000), no. 2, 343–364.
Pierre-Emmanuel Caprace, A sixteen-relator presentation of an infinite hyperbolic
Kazhdan group, preprint (2017), arxiv.org/abs/1708.09772.
Pierre-Emmanuel Caprace and Jean Lécureux, Combinatorial and group-theoretic compactifications of buildings, Ann. Inst. Fourier, Grenoble 61 (2011), no. 2, 619–672.
Michael W. Davis, Buildings are CAT(0), Geometry and cohomology in group theory
(Durham, 1994), London Math. Soc. Lecture Note Ser., vol. 252, Cambridge Univ.
Press, Cambridge, 1998, pp. 108–123.
Vinay V. Deodhar, A note on subgroups generated by reflections in Coxeter groups,
Arch. Math. (Basel) 53 (1989), no. 6, 543–546.
Jingyin Huang, Marcin Sabok, and Forte Shinko, Hyperfiniteness of boundary actions
of cubulated hyperbolic groups, preprint (2017), arxiv.org/abs/1701.03969v2.
Gabor Moussong, Hyperbolic Coxeter groups, Thesis of the Ohio State University
(1988).
G. A. Noskov, Asymptotic behavior of word metrics on Coxeter groups, Doc. Math. 16
(2011), 373–398.
Graham Niblo and Lawrence Reeves, Groups acting on CAT(0) cube complexes, Geom.
Topol. 1 (1997), 1–7.
Guennadi A. Noskov and Èrnest B. Vinberg, Strong Tits alternative for subgroups of
Coxeter groups, J. Lie Theory 12 (2002), no. 1, 259–264.
Jacek Światkowski, Some infinite groups generated by involutions have Kazhdan’s property (T), Forum Math. 13 (2001), no. 6, 741–755.
Nicholas Touikan, On geodesic ray bundles in hyperbolic groups, preprint (2017),
arxiv.org/abs/1706.01979.
UCL, 1348 Louvain-la-Neuve, Belgium
E-mail address: [email protected]
| 4math.GR
|
arXiv:1001.3604v1 [cs.SE] 20 Jan 2010
Type-Safe Feature-Oriented
Product Lines
Sven Apel† , Christian Kästner‡ , Armin Größlinger† , and
Christian Lengauer†
†
Department of Informatics and Mathematics, University of Passau
{apel,groesslinger,lengauer}@uni-passau.de
‡
School of Computer Science, University of Magdeburg
[email protected]
Technical Report, Number MIP-0909
Department of Informatics and Mathematics
University of Passau, Germany
June 2009
Type-Safe Feature-Oriented Product Lines
Sven Apel† , Christian Kästner‡ , Armin Größlinger† , and Christian Lengauer†
†
Department of Informatics and Mathematics, University of Passau
{apel,groesslinger,lengauer}@uni-passau.de
‡
School of Computer Science, University of Magdeburg
[email protected]
Abstract. A feature-oriented product line is a family of programs that share a
common set of features. A feature implements a stakeholder’s requirement, represents a design decision and configuration option and, when added to a program,
involves the introduction of new structures, such as classes and methods, and the
refinement of existing ones, such as extending methods. With feature-oriented
decomposition, programs can be generated, solely on the basis of a user’s selection of features, by the composition of the corresponding feature code. A key
challenge of feature-oriented product line engineering is how to guarantee the
correctness of an entire feature-oriented product line, i.e., of all of the member
programs generated from different combinations of features. As the number of
valid feature combinations grows progressively with the number of features, it
is not feasible to check all individual programs. The only feasible approach is
to have a type system check the entire code base of the feature-oriented product
line. We have developed such a type system on the basis of a formal model of a
feature-oriented Java-like language. We demonstrate that the type system ensures
that every valid program of a feature-oriented product line is well-typed and that
the type system is complete.
1
Introduction
Feature-oriented programming (FOP) aims at the modularization of programs in terms
of features [55,13]. A feature implements a stakeholder’s requirement and is typically
an increment in program functionality [55,13]. Contemporary feature-oriented programming languages and tools such as AHEAD [13], Xak [2], CaesarJ [48], Classbox/J [14], FeatureHouse [9], and FeatureC++ [10] provide a variety of mechanisms
that support the specification, modularization, and composition of features. A key idea
is that a feature is implemented by a distinct code unit, called a feature module. When
added to a base program, it introduces new structures, such as classes and methods, and
refines existing ones, such as extending methods [43,11]. A program that is decomposed
into features is called henceforth a feature-oriented program.1
1
Typically, feature-oriented decomposition is orthogonal to class-based or functional
decomposition [63,49,61]. A multitude of modularization and composition mechanisms [17,26,24,45,46,56,65] have been developed in order to allow programmers to decompose a program along multiple dimensions [63]. Feature-oriented languages and tools provide
a significant subset of these mechanisms [11].
Beside the decomposition of programs into features, the concept of a feature is
useful for distinguishing different, related programs thus forming a software product
line [35,21]. Typically, programs of a common domain share a set of features but also
differ in other features. For example, suppose an email client for mobile devices that
supports the protocols IMAP and POP3 and another client that supports POP3, MIME,
and SSL encryption. With a decomposition of the two programs into the features IMAP,
POP3, MIME, and SSL, both programs can share the code of the feature POP3. Since
mobile devices have only limited resources, unnecessary features should be removed.
With feature-oriented decomposition, programs can be generated solely on the basis of a user’s selection of features by the composition of the corresponding feature
modules. Of course, not all combinations of features are legal and result in correct
programs [12]. A feature model describes which features can be composed in which
combinations, i.e., which programs are valid [35,21]. It consists of an (ordered) set
of features and a set of constraints on feature combinations [21,12]. For example, our
email client may have different rendering engines for HTML text, e.g., the Mozilla engine or the Safari engine, but only one at a time. A set of feature modules along wit a
feature model is called a feature-oriented product line [12].
An important question is how the correctness of feature-oriented programs, in particular, and product lines, in general, can be guaranteed. A first problem is that contemporary feature-oriented languages and tools usually involve a code generation step
during composition in which the code is transformed into a lower-level representation.
In previous work, we have addressed this problem by modeling feature-oriented mechanisms directly in the formal syntax and semantics of a core language, called Feature
Featherweight Java (FFJ). The type system of FFJ ensures that the composition of
feature modules is type-safe [8].
In this paper, we address a second problem: How can the correctness of an entire
feature-oriented product line be guaranteed? A naive approach would be to type-check
all valid programs of a product line using a type checker like the one of FFJ [8]. However, this approach does not scale; already for 34 implemented optional features, a variant can be generated for every person on the planet. Noticing this problem, Czarnecki
and Pietroszek [22] and Thaker et al. [64] suggested the development of a type system that checks the entire code base of the feature-oriented product line, instead of all
individual feature-oriented programs. In this scenario, a type checker must analyze all
feature modules of a product line on the basis of the feature model. We will show that,
with this information, the type checker can ensure that every valid program variant that
can be generated is type-safe. Specifically, we make the following contributions:
– We provide a condensed version of FFJ, which is in many respects more elegant
and concise than its predecessor [8].
– We develop a formal type system that uses information about features and constraints on feature combinations in order to type-check a product line without generating every program.
– We prove correctness by proving that every program generated from a well-formed
product line is well-formed, as long as the feature selection satisfies the constraints
of the product line. Furthermore, we prove completeness by proving that the well2
typedness of all programs of a product line guarantees that the product line is welltyped as a whole.
– We offer an implementation of FFJ, including the proposed type system, which can
be downloaded for evaluation and for experiments with further feature-oriented
language and typing mechanisms.
Or work differs in many respects from previous and related work (see Section 5
for a comprehensive discussion). Most notably, Thaker et al. have implemented a type
system for feature-oriented product lines and conducted several case studies [64]. We
take their work further with a formalization and a correctness and completeness proof.
Furthermore, our work differs in many respects from previous work on modeling
and type-checking feature-oriented and related programming mechanisms. Most notably, we model the feature-related mechanisms directly in FFJ’s syntax and semantics,
without any transformation to a lower-level representation, and we stay very close to
the syntax of contemporary feature-oriented languages and tools (see Section 5). We
begin with a brief introduction to FFJ.
2
Feature-Oriented Programs in FFJ
In this section, we introduce the language FFJ. Originally, FFJ was designed for featureoriented programs [8,7]. We extend FFJ in Section 3 to support feature-oriented product
lines, i.e., to support the representation of multiple alternative program variants at a
time.
2.1
An Overview of FFJ
FFJ is a lightweight feature-oriented language that has been inspired by Featherweight
Java (FJ) [32]. As with FJ, we have aimed at minimality in the design of FFJ. FFJ
provides basic constructs like classes, fields, methods, and inheritance and only a few
new constructs capturing the core mechanisms of feature-oriented programming. But,
so far, FFJ’s type system has not supported the development of feature-oriented product
lines. That is, the feature modules written in FFJ are interpreted as a single program.
We will change this in Section 3.
An FFJ program consists of a set of classes and refinements. A refinement extends a
class that has been introduced previously. Each class and refinement is associated with
a feature. We say that a feature introduces a class or applies a refinement to a class.
Technically, the mapping between classes/refinements and the features they belong to
can be established in different ways, e.g., by extending the language with modules representing features [48,14,23] or by grouping classes and refinements that belong to a
feature in packages or directories [13,10].
Like in FJ, each class declares a superclass, which may be the class Object. Refinements are defined using the keyword refines. The semantics of a refinement applied to
a class is that the refinement’s members are added to and merged with the member of
the refined class. This way, a refinement can add new fields and methods to the class
and override existing methods (declared by overrides).
3
On the left side in Figure 1, we show an excerpt of the code of a basic email client,
called E MAIL C LIENT, (top) and a feature, called SSL, (bottom) in FFJ. The feature
SSL adds the class SSL (Lines 7–10) to the email client’s code base and refines the
class Trans in order to encrypt outgoing messages (Lines 11–15). To this effect, the
refinement of Trans adds a new field key (Line 12) and overrides the method send of
class Trans (Lines 13-15).
Feature E MAIL C LIENT
1
2
3
4
5
6
class Msg extends Object {
String serialize() { ... }
}
class Trans extends Object {
Bool send(Msg m) { ... }
}
EmailClient
Object
Feature SSL
7
8
9
10
11
12
13
14
15
16
class SSL extends Object {
Trans trans;
Bool send(Msg m) { ... }
}
refines class Trans {
Key key;
overrides Bool send(Msg m) {
return new SSL(this).send(m);
}
}
SSL
Trans
refines
Trans
inherits
class
refinement
Msg
refinement chain
SSL
feature
Fig. 1. A feature-oriented email client supporting SSL encryption.
Typically, a programmer applies multiple refinements to a class by composing a
sequence of features. This is called a refinement chain. A refinement that is applied
immediately before another refinement in the chain is called its predecessor. The order
of the refinements in a refinement chain is determined by their composition order. On
the right side in Figure 1, we depict the refinement and inheritance relationships of our
email example.
Fields are unique within the scope of a class and its inheritance hierarchy and refinement chain. That is, a refinement or subclass is not allowed to add a field that has
already been defined in a predecessor in the refinement chain or in a superclass. For
example, a further refinement of Trans would not be allowed to add a field key, since
key has been introduced by a refinement of feature SSL already. With methods, this is
different. A refinement or subclass may add new methods (overloading is prohibited)
and override existing methods. In order to distinguish the two cases, FFJ expects the
programmer to declare whether a method overrides an existing method (using the modifier overrides). For example, the refinement of Trans in feature SSL overrides the
method send introduced by feature M AIL; for subclasses, this is similar.
The distinction between method introduction and overriding allows the type system
to check (1) whether an introduced method inadvertently replaces or occludes an existing method with the same name and (2) whether, for every overriding method, there
is a proper method to be overridden. Apart from the modifier overrides, a method in
FFJ is similar to a method in FJ. That is, a method body is an expression (prefixed with
4
return) and not a sequence of statements. This is due to the functional nature of FFJ
and FJ. Furthermore, overloading of methods (introducing methods with equal names
and different argument types) is not allowed in FFJ (and FJ).
As shown in Figure 1, refinement chains grow from left to right and inheritance
hierarchies from top to bottom. When looking up a method body, FFJ traverses the
combined inheritance and refinement hierarchy of an object and selects the right-most
and bottom-most body of a method declaration or method refinement that is compatible. This kind of lookup is necessary since we model features directly in FFJ, instead
of generating and evaluating FJ code [40]. First, the FFJ calculus looks for a method
declaration in the refinement chain of the object’s class, starting with the last refinement
back to the class declaration itself. The first body of a matching method declaration is
returned. If the method is not found in the class’ refinement chain or in its own declaration, the methods in the superclass (and then the superclass’ superclass, etc.) are
searched, each again from the most specific refinement of the class declaration itself.
The field lookup works similarly, except that the entire inheritance and refinement hierarchy is searched and the fields are accumulated in a list. In Figure 2, we illustrate the
processes of method body and field lookup schematically.
Object
Class1
Ref(n,1)
Ref(n,p−1)
Ref(n,p)
Classn−1
Ref(n−1,1)
Ref(n−1,k−1)
Ref(n−1,k)
Classn
Ref(n,1)
Ref(n,m−1)
Ref(n,m)
Fig. 2. Order of method body and field lookup in FFJ.
2.2
Syntax of FFJ
Before we go into detail, let us explain some notational conventions. We abbreviate lists
in the obvious ways:
–
–
–
–
–
–
C is shorthand for C1 , . . . , Cn
C f is shorthand for C1 f1 , . . . , Cn fn
C f; is shorthand for C1 f1 ; . . . ; Cn fn ;
t : C is shorthand for t1 : C1 , . . . , tn : Cn
C <: D is shorthand for C1 <: D1 . . . Cn <: Dn
...
Note that, depending on the context, blanks, commas, or semicolons separate the elements of a list. The context will make clear which separator is meant. The symbol •
5
denotes the empty list and lists of field declarations, method declarations, and parameter
names must not contain duplicates. We use the metavariables A–E for class names, f–h
for field names, and m for method names. Feature names are denoted by Greek letters.
In Figure 3, we depict the syntax of FFJ in extended Backus-Naur-Form. An FFJ
program consists of a set of class and refinement declarations. A class declaration L
declares a class with the name C that inherits from a superclass D and consists of a list
C f; of fields and a list M of method declarations.2 A refinement declaration R consists
of a list C f; of fields and a list M of method declarations.
t ::=
terms:
x
variable
t.f
field access
t.m(t) method invocation
new C(t) object creation
(C) t
cast
L ::=
class declarations:
class C extends D { C f; M }
R ::=
refinement declarations:
refines class C { C f; M }
M ::=
method declarations:
[overrides] C m(C x) { return t; }
v ::=
values:
new C(v) object creation
Fig. 3. Syntax of FFJ in extended BNF.
A method m expects a list C x of arguments and declares a body that returns only
a single expression t of type C. Using the modifier overrides, a method declares that it
intends to override another method with the same name and signature. Where we want
to distinguish methods that override others and methods that do not override others, we
call the former method introductions and the latter method refinements
Finally, there are five forms of terms: the variable, field access, method invocation,
object creation, and type cast, which are taken from FJ without change. The only values
are object creations whose arguments are values as well.
2.3
FFJ’s Class Table
Declarations of classes and refinements can be looked up via a class table CT. The compiler fills the class table during the parser pass. In contrast to FJ, class and refinement
declarations are identified not only by their names but, additionally, by the names of
the enclosing features. For example, in order to retrieve the declaration of class Trans,
introduced by feature M AIL, in our example of Figure 1, we write CT(M AIL.Trans);
in order to retrieve the refinement of class Trans applied by feature SSL, we write
CT(SSL.Trans). We call Φ.C the qualified type of class C in feature Φ. In FFJ, class
and refinement declarations are unique with respect to their qualified types. This property is ensured because of the following sanity conditions: a feature is not allowed
2
The concept of a class constructor is unnecessary in FFJ and FJ [54]. Its omittance simplifies
the syntax, semantics, and type rules significantly without loss of generality.
6
– to introduce a class or refinement twice inside a single feature module and
– to refine a class that the feature has just introduced.
These are common sanity conditions in feature-oriented languages and tools [13,10,9].
As for FJ, we impose further sanity conditions on the class table and the inheritance
relation:
– CT (Φ.C) = class C. . . or refines class C. . . for every qualified type Φ.C ∈
dom(CT ); Feature Base plays the same role for features as Object plays for
classes; it is a symbol denoting the empty feature at which lookups terminate.
– Base.Object ∈
/ dom(CT );
– for every class name C appearing anywhere in CT , we have Φ.C ∈ dom(CT ) for
at least one feature Φ; and
– the inheritance relation contains no cycles (incl. self-cycles).
2.4
Refinement in FFJ
Information about the refinement chain of a class can be retrieved using the refinement
table RT . The compiler fills the refinement table during the parser pass. RT (C) yields
a list of all features that either introduce or refine class C. The leftmost element of
the result list is the feature that introduces the class C and, then, from left to right, the
features are listed that refine class C in the order of their composition. In our example of
Figure 1, RT (Trans) yields the list E MAIL C LIENT, SSL. There is only a single sanity
condition for the refinement table:
– RT (C) = Φ for every type C ∈ dom(CT ), with Φ being the features that introduce and refine class C.
In Figure 4, we show two functions for the navigation of the refinement chain that
rely on RT . Function last returns, for a class name C, a qualified type Ψn .C, in which
Ψn refers to the feature that applies the final refinement to class C; if a class is not
refined at all, Ψn refers to the feature that introduces class C. Function pred returns,
for a qualified type Φ.C, another qualified type Ψn .C, in which Ψn refers to the feature
that introduces or refines class C and that is the immediate predecessor of Φ in the
refinement chain; if there is no predecessor, Base.Object is returned.
Navigating along the refinement chain
RT (C) = Ψ
last(C) = Ψn .C
RT (C) = Ψ, Φ, Ω
Ψ 6= •
pred(Φ.C) = Ψn .C
Fig. 4. Refinement in FFJ.
7
RT (C) = Φ, Ω
pred(Φ.C) = Base.Object
2.5
Subtyping in FFJ
In Figure 5, we show the subtype relation of FFJ. The subtype relation <: is defined
by one rule each for reflexivity and transitivity and one rule for relating the type of a
class to the type of its immediate superclass. It is not necessary to define subtyping over
qualified types because only classes (not refinements) declare superclasses and there is
only a single declaration per class.
C <: D
Subtyping
C <: C
CT(Φ.C) = class C extends D { . . . }
C <: D
C <: D
D <: E
C <: E
Fig. 5. Subtyping in FFJ.
2.6
Auxiliary Definitions of FFJ
In Figure 6, we show the auxiliary definitions of FFJ. Function fields searches the refinement chain from right to left and accumulates the fields into a list (using the comma
as concatenation operator). If there is no further predecessor in the refinement chain,
i.e., we have reached a class declaration, then the refinement chain of the superclass is
searched (see Figure 2). If Base.Object is reached, the empty list is returned (denoted
by •).
Function mbody looks up the most specific and most refined body of a method m. A
body consists of the formal parameters x of a method and the actual term t representing
the content. The search is like in fields. First, the refinement chain is searched from
right to left and, then, the superclasses’ refinement chains are searched, as illustrated in
Figure 2. Note that [overrides] means that a given method declaration may (or may not)
have the modifier. This way, we are able to define uniform rules for method introduction
and method refinement. Function mtype yields the signature B → B0 of a declaration
of method m. The lookup is like in mbody.
Predicate introduce is used to check whether a class has been introduced by multiple features and whether a field or method has been introduced multiple times in a
class. Precisely, it states, in the case of classes, whether C has not been introduced by
any feature other than Φ and whether a method m or a field f has not been introduced
by Φ.C or in any of its predecessors or superclasses. To evaluate it, we check, in the
case of classes, whether CT (Ψ.C) yields a class declaration or not, for any feature Ψ
different from Φ, in the case of methods, whether mtype yields a signature or not and,
in the case of fields, whether f is defined in the list of fields returned by fields.
Predicate refine states whether, for a given refinement, a proper class has been
declared previously in the refinement chain. The predicate override states whether a
method m has been introduced before in some predecessor of Φ.C and whether the
previous declaration of m has the given signature.
8
fields(Φ.C) = C f
Field lookup
fields(Base.Object) = •
CT (Φ.C) = class C extends D { C f; M }
fields(Φ.C) = fields(last(D)), C f
CT (Φ.C) = refines class C { C f; M }
fields(Φ.C) = fields(pred(Φ.C)), C f
mbody(m, Φ.C) = (x, t)
Method body lookup
[overrides] B m(B x) { return t; } ∈ M
CT (Φ.C) = class C extends D { C f; M }
mbody(m, Φ.C) = (x, t)
m is not defined in M
CT (Φ.C) = class C extends D { C f; M }
mbody(m, Φ.C) = mbody(m, last(D))
[overrides] B m(B x) { return t; } ∈ M
CT (Φ.C) = refines class C { C f; M }
mbody(m, Φ.C) = (x, t)
m is not defined in M
CT (Φ.C) = refines class C { C f; M }
mbody(m, Φ.C) = mbody(m, pred(Φ.C))
mtype(m, Φ.C) = C → C
Method type lookup
B0 m(B x) { return t; } ∈ M
CT (Φ.C) = class C extends D { C f; M }
m is not defined in M
CT (Φ.C) = class C extends D { C f; M }
mtype(m, Φ.C) = mtype(m, last(D))
mtype(m, Φ.C) = B → B0
B0 m(B x) { return t; } ∈ M
CT (Φ.C) = refines class C { C f; M }
m is not defined in M
CT (Φ.C) = refines class C { C f; M }
mtype(m, Φ.C) = mtype(m, pred(Φ.C))
mtype(m, Φ.C) = B → B0
introduce(Φ.C)
Valid class introduction
@ Ψ : (CT (Ψ.C) = class C . . . ∧ Φ 6= Ψ)
introduce(Φ.C)
introduce(f, Φ.C)
Valid field introduction
fields(Φ.C) = E h
f∈
/h
introduce(f, Φ.C)
introduce(m, Φ.C)
Valid method introduction
(m, Φ.C) ∈
/ dom(mtype)
introduce(m, Φ.C)
refine(Φ.C)
Valid class refinement
RT (C) = Ψ, Φ, Ω
CT (Ψ1 .C) = class C . . .
refine(Φ.C)
override(m, Φ.C, C → C0 )
Valid method overriding
mtype(m, Φ.C) = B → B0
C=B
override(m, Φ.C, C → C0 )
C0 = B0
9
Fig. 6. Auxiliary definitions of FFJ.
2.7
Evaluation of FFJ Programs
Each FFJ program consists of a class table and a term.3 The term is evaluated using the
evaluation rules shown in Figure 7. The evaluation terminates when a value, i.e., a term
of the form new C(v), is reached. Note that we use a direct semantics of class refinement [40]. That is, the field and method lookup mechanisms incorporate all refinements
when a class is searched for fields and methods. An alternative, which is discussed in
Section 5, would be a flattening semantics, i.e., to merge a class in a preprocessing step
with all of its refinements into a single declaration.
fields(last(C)) = C f
(new C(v)).fi −→ vi
(E-P ROJ N EW)
mbody(m, last(C)) = (x, t0 )
(E-I NVK N EW)
(new C(v)).m(u) −→ [x 7→ u, this 7→ new C(v)] t0
C <: D
(D)(new C(v)) −→ new C(v)
(E-C AST N EW)
t0 −→ t00
t0 .f −→ t00 .f
(E-F IELD)
t0 −→ t00
t0 .m(t) −→ t00 .m(t)
(E-I NVK R ECV)
ti −→ t0i
v0 .m(v, ti , t) −→ v0 .m(v, t0i , t)
(E-I NVK A RG)
ti −→ t0i
new C(v, ti , t) −→ new C(v, t0i , t)
(E-N EWA RG)
t0 −→ t00
(C)t0 .f −→ (C)t00 .f
(E-C AST)
Fig. 7. Evaluation of FFJ programs.
Using the subtype relation <: and the auxiliary functions fields and mbody, the evaluation of FFJ is fairly simple. The first three rules are most interesting (the remaining
rules are just congruence rules). Rule E-P ROJ N EW describes the projection of a field
from an instantiated class. A projected field fi evaluates to a value vi that has been
passed as argument to the instantiation. Function fields is used to look up the fields
of the given class. It receives last(C) as argument since we want to search the entire
refinement chain of class C from right to left (cf. Figure 2).
Rule E-P ROJ I NVK evaluates a method invocation by replacing the invocation with
the method’s body. The formal parameters of the method are substituted in the body for
3
The refinement table is not relevant for evaluation.
10
the arguments of the invocation; the value on which the method is invoked is substituted
for this. The function mbody is called with the last refinement of the class C in order to
search the refinement chain from right to left and return the most specific method body
(cf. Figure 2).
Rule E-C AST N EW evaluates an upcast by simply removing the cast. Of course, the
premise must be that the cast is really an upcast and not a downcast or an incorrect cast.
2.8
Type Checking FFJ Programs
The type relation of FFJ consists of the type rules for terms and the well-formedness
rules for classes, refinements, and methods, shown in Figures 8 and 9.
Γ ` t:C
Term typing
x:C ∈ Γ
Γ ` x:C
Γ ` t0 : C0
Γ ` t0 : C0
Γ ` t:C
Γ ` t:C
(T-VAR)
fields(last(C0 )) = C f
Γ ` t0 .fi : Ci
mtype(m, last(C0 )) = D → C
Γ ` t0 .m(t) : C
fields(last(C)) = D f
Γ ` new C(t) : C
C <: D
Γ ` t0 : D
D <: C
Γ ` (C)t0 : C
Γ ` t0 : D
C <: D
Γ ` (C)t0 : C
Γ ` t0 : D
C 6<: D
D 6<: C
Γ ` (C)t0 : C
(T-F IELD)
C <: D
(T-I NVK)
(T-N EW)
(T-UC AST)
C 6= D
stupid warning
(T-DC AST)
(T-SC AST)
Fig. 8. Term typing in FFJ.
Term Typing Rules. A term typing judgment is a triple consisting of a typing context
Γ, a term t, and a type C (see Figure 8).
Rule T-VAR checks whether a free variable is contained in the typing context. Rule
T-F IELD checks whether a field access t0 .f is well-typed. Specifically, it checks whether
f is declared in the type of t0 and whether the type f equals the type of the entire term.
Rule T-I NVK checks whether a method invocation t0 .m(t) is well-typed. To this end,
it checks whether the arguments t of the invocation are subtypes of the types of the
11
M OK a Φ.C
Method typing
x : B, this : C ` t0 : E0
E0 <: B0
introduce(m, last(D))
CT (Φ.C) = class C extends D { C f; M }
B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E0
CT (Φ.C) = class C extends D { C f; M }
E0 <: B0
override(m, last(D), B → B0 )
overrides B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E0
CT (Φ.C) = refines class C { C f; M }
E0 <: B0
introduce(m, pred(Φ.C))
B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E0
E0 <: B0
CT (Φ.C) = refines class C { C f; M }
override(m, pred(Φ.C), B → B0 )
overrides B0 m(B x) { return t0 ; } OK a Φ.C
L OK a Φ
Class typing
∀ f ∈ f : introduce(f, last(D))
introduce(Φ.C)
M OK a Φ.C
class C extends D { C f; M } OK a Φ
R OK a Φ
Refinement typing
∀ f ∈ f : introduce(f, pred(Φ.C))
refine(Φ.C)
refines class C { C f; M } OK a Φ
Fig. 9. Well-formedness rules of FFJ.
12
M OK a Φ.C
formal parameters of m and whether the return type of m equals the type of the entire
term. Rule T-N EW checks whether an object creation new C(t) is well-typed in that it
checks whether the arguments t of the instantiation of C are subtypes of the types D of
the fields of C and whether C equals the type of the entire term. The rules T-UC AST,
T-DC AST, and T-SC AST check whether casts are well-typed. In each rule, it is checked
whether the type C the term t0 is cast to is a subtype, supertype, or unrelated type of the
type of t0 and whether C equals the type of the entire term.4
Well-Formedness Rules. In Figure 9, we show FFJ’s well-formedness rules of classes,
refinements, and methods.
The typing judgments of classes and refinements are binary relations between a
class or refinement declaration and a feature, written L OK a Φ and R OK a Φ. The
rule of classes checks whether all methods are well-formed in the context of the class’
qualified type. Moreover, it checks whether none of the fields of the class declaration
is introduced multiple times in the combined inheritance and refinement hierarchy and
whether there is no feature other than Φ that introduces a class C (using introduce). The
well-formedness rule of refinements is analogous, except that the rule checks whether
a corresponding class has been introduced before (using refine).
The typing judgment of methods is a binary relation between a method declaration
and the qualified type that declares the method, written M OK a Φ.C. There are four
different rules for methods (from top to bottom in Figure 9)
1.
2.
3.
4.
that do not override another method and that are declared by classes,
that override another method and that are declared by classes,
that do not override another method and that are declared by refinements,
that override another method and that are declared by refinements.
All four rules check whether the type E0 of the method body is a subtype of the declared
return type B0 of the method declaration. For methods that are being introduced, it is
checked whether no method with an identical name has been introduced in a superclass
(Rule 1) or in a predecessor in the refinement chain (Rule 3). For methods that override
other methods, it is checked whether a method with identical name and signature exists
in the superclass (Rule 2) or in a predecessor in the refinement chain (Rule 4).
Well-Typed FFJ Programs. Finally, an FFJ program, consisting of a term, a class
table, and a refinement table, is well-typed if
– the term is well-typed (checked using FFJ’s term typing rules),
– all classes and refinements stored in the class table are well-typed (checked using
FFJ’s well-formedness rules), and
– the class and refinement tables are well-formed (ensured by the corresponding sanity conditions).
4
Rule T-SC AST is needed only for the small step semantics of FFJ (and FJ) in order to be
able to formulate and prove the type preservation property. FFJ (and FJ) programs whose type
derivation contains this rule (i.e., the premise stupid warning appears in the derivation) are
not further considered (cf. [32]).
13
Type Soundness of FFJ. The type system of FFJ is sound. We can prove this using the
standard theorems of preservation and progress [66]:
T HEOREM 2.1 (Preservation) If Γ ` t : C and t −→ t0 , then Γ ` t0 : C0 for some
C0 <: C.
T HEOREM 2.2 (Progress) Suppose t is a well-typed term.
1. If t includes new C0 (t).fi as a subterm, then fields(last(C0 )) = C f for some C and
f.
2. If t includes new C0 (t).m(u) as a subterm, then mbody(m, last(C0 )) = (x, t0 ) and
|x| = |u| for some x and t0 .
We provide the proofs of the two theorems in Appendix A.
3
Feature-Oriented Product Lines in FFJPL
In this section, our goal is to define a type system for feature-oriented product lines –
a type system that checks whether all valid combinations of features yield well-typed
programs. In this scenario, the features in question may be optional or mutually exclusive so that different combinations are possible that form different feature-oriented
programs. Since there may be plenty of valid combinations, type checking all of them
individually is usually not feasible.
In order to provide a type system for feature-oriented product lines, we need information about which combinations of features are valid, i.e., which features are mandatory, optional, or mutually exclusive, and we need to adapt the subtype and type rules
of FFJ to check that there are no combinations/variants that lead to ill-typed terms.
The type system guarantees that every program derived from a well-typed product line
is a well-typed FFJ program. FFJ together with the type system for checking featureoriented product lines is henceforth called FFJPL .
3.1
An Overview of Feature-Oriented Product Lines
A feature-oriented product line is made up of a set of feature modules and a feature model. The feature modules contains the features’ implementation and the feature
model describes how the feature modules can be combined. In contrast to the featureoriented programs of Section 2, typically, some features are optional and some are mutually exclusive (Also other relations such as disjunction, negation, and implication are
possible [12]; they are broken down to mandatory, optional, and mutually exclusive
features, as we will explain.). Generally, in a derivation step, a user selects a valid subset of features from which, subsequently, a feature-oriented program is derived. In our
case, derivation means assembling the corresponding feature modules for a given set of
features. In Figure 10, we illustrate the process of program derivation.
Typically, a wide variety of programs can be derived from a product line [21,19].
The challenge is to define a type system that guarantees, on the basis of the feature
modules and the feature model, that all valid programs are well-typed. Once a program
is derived from such a product line, we can be sure that it is well-typed and we can
evaluate it using the standard evaluation rules of FFJ (see Section 2.7).
14
feature−oriented product line
feature−oriented programs
feature modules
A
B
C
D
E
F
A
B
C
program
program
A
B
E
A
B
D
E
derivation
A
B
D
B
F
E
E F
...
...
...
user’s feature
selection
program
program
A
B
C
E
A
B
C
D
E
...
feature model
Fig. 10. The process of deriving programs from a product line.
3.2
Managing Variability – Feature Models
The aim of developing a product line is to manage the variability of a set of programs
developed for a particular domain and to facilitate the reuse of feature implementations
among the programs of the domain. A feature model captures the variability by (explicitly or implicitly) defining an ordered set of all features of a product line and their legal
feature combinations. A well-defined feature order is essential for field and method
lookup (see Section 3.6).
Different approaches to product line engineering use different representations of
feature models to define legal feature combinations. The simplest approach is to enumerate all legal feature combinations. In practice, commonly different flavors of tree
structures are used, sometimes in combination with additional propositional constraints,
to define legal combinations [21,12], as illustrated in Figure 10.
For our purpose, the actual representation of legal feature combinations is not relevant. In FFJPL , we use the feature model only to check whether feature and/or specific
program elements are present in certain circumstances. A design decision of FFJPL is
to abstract from the concrete representation of the underlying feature model and rather
to provide an interface to the feature model. This has to benefits: (1) we do not need
to struggle with all the details of the formalization of feature models, which is well
understood by researchers [12,22,64,23] and outside the scope of this paper, and (2) we
are able to support different kinds of feature model representations, e.g., a tree structures, grammars, or propositional formulas [12]. The interface to the feature model is
simply a set of functions and predicates that we use to ask questions like “may (or may
not) feature A be present together with feature B” or “is program element m present in
every variant in which also feature A is present”, i.e., “is program element m always
reachable from feature A”.
3.3
Challenges of Type Checking
Let us explain the challenges of type checking by extending our email example, as
shown in Figure 11. Suppose our basic email client is refined to process incoming
text messages (feature T EXT, Lines 1–8). Optionally, it is enabled to process HTML
15
messages, using either Mozilla’s rendering engine (feature M OZILLA, Lines 9–12)
or Safari’s rendering engine (feature S AFARI, Lines 13–16). To this end, the features
M OZILLA and S AFARI override the method render of class Display (Line 11 and 15)
in order to invoke the respective rendering engines (field renderer, Lines 10 and 14)
instead of the text printing function (Line 7).
Feature T EXT
1
2
3
4
5
6
7
8
refines class Trans {
Unit receive(Msg msg) {
return /∗ do something... ∗/ new Display().render(msg);
}
}
class Display {
Unit render(Msg msg) { /∗ display message in text format ∗/ }
}
Feature M OZILLA
9 refines class Display {
10
MozillaRenderer renderer;
11
overrides Unit render(Msg m) { /∗ render HTML message using the Mozilla engine ∗/ }
12 }
Feature S AFARI
13 refines class Display {
14
SafariRenderer renderer;
15
overrides Unit render(Msg m) { /∗ render HTML message using the Safari engine ∗/ }
16 }
Fig. 11. A feature-oriented email client using Mozilla’s and Safari’s rendering engines.
The first thing to observe is that the features M OZILLA and S AFARI rely on class
Display and its method render introduced by feature T EXT. In order to guarantee that
every derived program is well-formed, the type system checks whether Display and
render are always reachable from the features M OZILLA and S AFARI, i.e., whether,
in every program variant that contains M OZILLA and S AFARI, also feature T EXT is
present.
The second thing to observe is that the features M OZILLA and S AFARI both add
a field renderer to Display (Lines 10 and 14), both of which have different types. In
FFJ, a program with both feature modules would not be a well-typed program because
the field renderer is introduced twice. However, Figure 11 is not intended to represent a single feature-oriented program but a feature-oriented product line; the features
M OZILLA and S AFARI are mutually exclusive, as defined in the product line’s feature
model (stated earlier), and the type system has to take this fact into account.
Let us summarize the key challenges of type checking product lines:
– A global class table contains classes and refinements of all features of a product
line, even if some features are optional or mutually exclusive so that they are present
only in some derived programs. That is, a single class can be introduced by multiple
features as long as the features are mutually exclusive. This is also the case for
multiple introductions of methods and fields, which may even have different types.
16
– The presence of types, fields, and methods depends on the presence of the features that introduce them. A reference from the elements of a feature to a type, a
field projection, or a method invocation is valid if the referenced element is always
reachable from the referring feature, i.e., in every variant that contains the referring
feature.
– Like references, an extension of a program element, such as a class or method
refinement, is valid only if the extended program element is always reachable from
the feature that applies the refinement.
– Refinements of classes and methods do not necessarily form linear refinement
chains. There may be alternative refinements of a single class or method that exclude one another, as explained below.
3.4
Collecting Information on Feature Modules
For type checking, the FFJPL compiler collects various information on the feature modules of the product line. Before the actual type checking is performed, the compiler fills
three tables with information: the class table (CT ), the introduction table (IT ), and the
refinement table (RT ).
The class table CT of FFJPL is like the one of FFJ and has to satisfy the same
sanity conditions except that (1) there may be multiple declarations of a class (or field
or method), as long as they are defined in are mutually exclusive features, and (2) there
may be cycles in the inheritance hierarchy, but no cycles for each set of classes which
are reachable from any given feature.
The introduction table IT maps a type to a list Φ of (mutually exclusive) features
that introduce the type. The features returned by IT are listed in the order prescribed
by the feature model. In our example of Figure 11, a call of IT (Display) would return
a list consisting only of the single feature T EXT. Likewise, the introduction table maps
field and method names, in combination with their declaring classes, to features. For
example, a call of IT (Display.renderer) would return the list M OZILLA, S AFARI. The
sanity conditions for the introduction table are straightforward:
– IT (C) = Φ for every type C ∈ dom(CT ), with Φ being the features that introduce
class C.
– IT (C.f) = Φ for every field f contained in some class C ∈ dom(CT ), with Φ
being the features that introduce field f.
– IT (C.m) = Φ for every method m contained in some class C ∈ dom(CT ), with
Φ being the features that introduce method m.
Much like in FFJ, in FFJPL there is a refinement table RT . A call of RT (C) yields
a list of all features that either introduce or refine class C, which is different from the
introduction table that returns only the features that introduce class C. As with IT , the
features returned by RT are listed in the order prescribed by the feature model. The
sanity condition for FFJPL ’s refinement table is identical to the one of FFJ, namely:
– RT (C) = Φ for every type C ∈ dom(CT ), with Φ being the features that introduce and refine class C.
17
3.5
Feature Model Interface
As said before, in FFJPL , we abstract from the concrete representation of the feature
model and define instead an interface consisting of proper functions and predicates.
There are two kinds of questions we want to ask about the feature model, which we
explain next.
First, we would like to know which features are never present together, which features are sometimes present together, and which features are always present together. To
this end, we define two predicates, never and sometimes, and a function always. Predicate never (Ω, Φ) indicates that feature Φ is never reachable in the context Ω, i.e., there
is no valid program variant in which the features Ω and feature Φ are present together.
Predicate sometimes(Ω, Φ) indicates that feature Φ is sometimes present when the
features Ω are present, i.e., there are variants in which the features Ω and feature Φ are
present together and there are variants in which they are not present together. Function
always(Ω, Φ) is used to evaluate whether feature Φ is always present in the context Ω
(either alone or within a group of alternative features). There are three cases: if feature Φ
is always present in the context, always returns the feature again (always(Ω, Φ) = Φ);
if feature Φ is not always present, but would be together with a certain group of mutually exclusive features Ψ (i.e., one of the group is always present), always returns all
features of this group (always(Ω, Φ) = Φ, Ψ). If a feature is not present at all, neither
alone nor together with other mutually exclusive features, always returns the empty
list (always(Ω, Φ) = •). The above predicates and function provide all information we
need to know about the features’ relationships. They are used especially for field and
method lookup.
Second, we would like to know whether a specific program element is always
present when a given set of features is present. This is necessary to ensure that references to program elements are always valid (i.e., not dangling). We need two sources
of information for that. First, we need to know all features that introduce the program
element in question (determined using the introduction table) and, second, we need to
know which combinations of features are legal (determined using the feature model).
For the field renderer of our example, the introduction table would yield the features
M OZILLA and S AFARI and, from the feature model, it follows that M OZILLA and S A FARI are mutually exclusive, i.e., never (M OZILLA , S AFARI ). But it can happen that
none of the two features is present, which can invalidate a reference to the field. The
type system needs to know about this situation.
To this end, we introduce a predicate validref that expresses that a program element
is always reachable from a set of features. For example, validref (Ω, C) holds if type
C is always reachable from the context Ω, validref (Ω, C.f) holds if field f of class C
is always reachable from the context Ω, and validref (Ω, C.m) holds if method m of
class C is always reachable from the context Ω. Applying validref to a list of program
elements means that the conjunction of the predicates for every list element is taken.
Finally, when we write validref (Ω, C) a Ψ, we mean that program element C is always
reachable from a context Ω in a subset Ψ of features of the product line.
In our prototype, we have implemented the above functions and predicates using a
SAT solver that reasons about propositional formulas representing constraints on legal
18
feature combinations (see Section 4), as proposed by Batory [12] and Czarnecki and
Pietroszek [22].
3.6
Refinement in FFJPL
In Figure 12, we show the functions last and pred for the navigation along the refinement chain. The two functions are identical to the ones of FFJ (cf. Figure 4). However,
in FFJPL , there may be alternative declarations of a class and, in the refinement chain,
refinement declarations may even precede class declarations, as long as the declaring
features are mutually exclusive. Let us illustrate refinement in FFJPL by means of the
example shown in Figure 13. Class C is introduced in the features Φ1 and Φ3 . Feature
Φ2 refines class C introduced by feature Φ1 and feature Φ4 refines class C introduced by
feature Φ3 . Feature Φ1 and Φ2 are never present when feature Φ3 or Φ4 are present and
vice versa. A call of RT (C) would return the list Φ1 , . . . , Φ4 , a call of last(C) would
return the qualified type Φ4 .C, and a call of pred (Φ4 .C) would return the qualified type
Φ3 .C and so on.
Navigating along the refinement chain
RT (C) = Ψ
last(C) = Ψn .C
Ψ 6= •
RT (C) = Ψ, Φ, Ω
pred(Φ.C) = Ψn .C
RT (C) = Φ, Ω
pred(Φ.C) = Base.Object
Fig. 12. Refinement in FFJPL .
C
C
C
C
Φ1
Φ2
Φ3
Φ4
mutually exclusive
Fig. 13. Multiple alternative refinements.
3.7
Subtyping in FFJPL
The subtype relation is more complicated in FFJPL than in FFJ. The reason is that
a class may have multiple declarations in different features, each declaring possibly
different superclasses, as illustrated in Figure 14. That is, when checking whether a class
is a subtype of another class, we need to check whether the subtype relation holds in
all alternative inheritance paths that may be reached from a given context. For example,
FooBar is a subtype of BarFoo because BarFoo is a superclass of FooBar in every
19
program variant (since always(Φ1 , Φ2 ) = Φ2 , Φ3 ); but FooBar is not a subtype of
Foo and Bar because, in both cases, a program variant exists in which FooBar is not a
(indirect) subclass of the class in question.
Φ2
Foo
Bar
A a;
A m(D d);
B b;
B m(B b);
BarFoo
BarFoo
D d;
D d;
Φ3
Φ2 and Φ3 are
mutually exclusive and
one of them is always
present together with Φ1
FooBar
E e;
Φ1
Fig. 14. Multiple inheritance chains in the presence of alternative features.
In Figure 15, we show the subtype relation of FFJPL . The subtype relation C <:
E a Ω is read as follows: in the context Ω, type C is a subtype of type E, i.e., type C is
a subtype of type E in every variant in which also the features Ω are present. The first
rule in Figure 15 covers reflexivity and terminates the recursion over the inheritance
hierarchy. The second rule states that class C is a subtype of class E if at least one
declaration of C is always present (tested with validref ) and if every of C’s declarations
that may be present together with Ω (tested with sometimes) declares some type D as
its supertype and D is a subtype of E in the context Ω. That is, E must be a direct or
indirect supertype of D in all variants in which the features Ω are present. Additionally,
supertype D must be always reachable from the context (Ω, Ψ). When traversing the
inheritance hierarchy, in each step, the context is extended by the feature that introduces
the current class in question, e.g., Ω is extended with Ψ.
Interestingly, the second rule subsumes the two FFJ rules for transitivity and direct
superclass declaration because some declarations of C may declare E directly as its
superclass and some declarations may declare another superclass D that is, in turn, a
subtype of E, and the rule must be applicable to both cases simultaneously.
C <: E a Ω
Subtyping
C <: C a Ω
validref (Ω, C)
CT(Ψ.C) = class C extends D { . . . }
∀ Ψ ∈ IT (C) : sometimes(Ω, Ψ) ⇒
validref ((Ω, Ψ), D) D <: E a Ω, Ψ
C <: E a Ω
Fig. 15. Subtyping in FFJPL .
20
Applied to our example of Figure 14, we have FooBar <: FooBar a Φ1 because
of the reflexivity rule. We also have FooBar <: BarFoo a Φ1 because FooBar is
reachable from feature Φ1 and every feature that introduces FooBar, namely Φ1 , contains a corresponding class declaration that declares BarFoo as FooBar’s superclass,
and BarFoo is always reachable from Φ1 . However, we have FooBar 6<: Foo a Φ1
and FooBar 6<: Bar a Φ1 because FooBar’s immediate superclass BarFoo is not
always a subtype of Foo respectively of Bar.
3.8
Auxiliary Definitions of FFJPL
Extending FFJ toward FFJPL makes it necessary to add and modify some auxiliary
functions. The most complex changes concern the field and method lookup mechanisms.
Field Lookup. The auxiliary function fields collects the fields of a class including the
fields of its superclasses and refinements. Since alternative class or refinement declarations may introduce alternative fields (or the same field with identical or alternative
types), fields may return different fields for different feature selections. Since we want
to type-check all valid variants, field returns multiple field lists (i.e., a list of lists) that
cover all possible feature selections. Each inner list contains field declarations collected
in an alternative path of the combined inheritance and refinement hierarchy.
For legibility, we separate the inner lists using the delimiter ‘◦’. For example, looking up the fields of class FooBar in the context of feature Φ1 (Figure 14) yields the
list A a, D d, E e ◦ B b, D d, E e because the features Φ2 and Φ3 are mutually exclusive and one of them is present in each variant in which also Φ1 is present. For
readability, we use the metavariables F and G when referring to inner field lists. We
abbreviate a list of lists of fields F1 ◦ . . . ◦ Fn by F. Analogously, F is shorthand for
F11 ◦ . . . ◦ Fn1 ◦ . . . ◦ F1m ◦ . . . ◦ Fnm .
Function fields receives a qualified type Φ.C and a context of selected features Ω.
If we want all possible field lists, the context is empty. If we want only field lists for
a subset of feature selections, e.g., only the fields that can be referenced from a term
in a specific feature module, we can use the context to specify one or more features of
which we know that they must be selected.
The basic idea of FFJPL ’s field lookup is to traverse the combined inheritance and
refinement hierarchy much like in FFJ. There are four situations that are handled differently:
1. The field lookup returns the empty list when it reaches Base.Object.
2. The field lookup ignores all fields that are introduced by features that are never
present in a given context.
3. The field lookup collects all fields that are introduced by features that are always
present in a given context. References to these fields are always valid.
4. The field lookup collects all fields that are introduced by features that may be
present in a given context but that are not always present. In this case, a special
21
marker @ is added to the fields in question because we cannot guarantee that a reference to this field is safe in the given context.5 It is up to the type system to decide,
based on the marker, whether this situation may provoke an error (e.g., the type system ignores the marker when looking for duplicate fields but reports an error when
type checking object creations).
5. A special situation occurs when the field lookup identifies a group of alternative
features. In such a group each feature is optional and excludes every other feature
of the group and at least one feature of the group is always present in a given
context. Once the field lookup identifies a group of alternative features, we split the
result list, each list containing the fields of a feature of the group and the fields of
the original list.
fields(Ω, Φ.C) = C f
Field lookup
fields(Ω, Φ.Object) = •
(FL-1)
never (Ω, Φ)
fields(Ω, Φ.C) = fields(Ω, pred (Φ.C))
(FL-2)
sometimes(Ω, Φ)
always(Ω, Φ) = Φ
CT (Φ.C) = class C extends D { C f; M }
fields(Ω, Φ.C) = append (fields(Ω, last(D)), C f)
sometimes(Ω, Φ)
always(Ω, Φ) = Φ
CT (Φ.C) = refines class C { C f; M }
fields(Ω, Φ.C) = append (fields(Ω, pred (Φ.C)), C f)
sometimes(Ω, Φ)
always(Ω, Φ) = •
CT (Φ.C) = class C extends D { C f; M }
fields(Ω, Φ.C) = append (fields(Ω, last(D)), C f@)
sometimes(Ω, Φ)
always(Ω, Φ) = •
CT (Φ.C) = refines class C { C f; M }
fields(Ω, Φ.C) = append (fields(Ω, pred (Φ.C)), C f@)
(FL-3.1)
(FL-3.2)
(FL-4.1)
(FL-4.2)
sometimes(Ω, Φ)
always(Ω, Φ) = Ψ
(FL-5)
fields(Ω, Φ.C) = fields((Ω, Ψ1 ), Φ.C) ◦ . . . ◦ fields((Ω, Ψn ), Φ.C)
Fig. 16. Field lookup in FFJPL .
In order to distinguish the different cases, we use the predicates and functions defined in Section 3.5 (especially never , sometimes, and always). The definition of func5
Note that the marker @ is generated during type checking, so we do not include it in the syntax
of FFJ.
22
tion fields, shown in Figure 16, follows the intuition described above: Once Base.Object
is reached, the recursion terminates (FL-1). When a feature is never reachable in the
given context, fields ignores this feature and resumes with the previous one (FL-2).
When a feature is mandatory (i.e., always present in a given context), the fields in question are added to each alternative result list, which were created in Rule FL-5 (FL-3.1
and FL-3.2).6 When a feature is optional, the fields in question, annotated with the
marker @, are added to each alternative result list (FL-4.1 and FL-4.2). When a feature
is part of an alternative group of features, we cannot immediately decide how to proceed. We split the result list in multiple lists (by means of multiple recursive invocations
of fields), in which we add one of the alternative features to each context passed to an
invocation of fields (FL-5).
mtype(Ω, m, Φ.C) = B → B0
Method type lookup
mtype(Ω, m, Base.Object) = •
sometimes(Ω, Φ)
B0 m(B x) { . . . } ∈ M
CT (Φ.C) = class C extends D { C f; M }
mtype(Ω, m, Φ.C) = mtype(Ω, m, pred (Φ.C)), mtype(Ω, m, last(D)), B → B0
B0 m(B x) { . . . } ∈ M
sometimes(Ω, Φ)
CT (Φ.C) = refines class C { C f; M }
mtype(Ω, m, Φ.C) = mtype(Ω, m, pred (Φ.C)), B → B0
(m is not defined in M ∨ never (Ω, Φ))
CT (Φ.C) = class C extends D { C f; M }
mtype(Ω, m, Φ.C) = mtype(Ω, m, pred (Φ.C)), mtype(Ω, m, last(D))
(m is not defined in M ∨ never (Ω, Φ))
CT (Φ.C) = refines class C { C f; M }
mtype(Ω, m, Φ.C) = mtype(Ω, m, pred (Φ.C))
(ML-1)
(ML-2)
(ML-3)
(ML-4)
(ML-5)
Fig. 17. Method Lookup in FFJPL .
Method Type Lookup. Like in field lookup, in method lookup, we have to take alternative definitions of methods into account. But the lookup mechanism is simpler
than in fields because the order of signatures found in the combined inheritance and
refinement hierarchy is irrelevant for type checking. Hence, function mtype yields a
simple list B → B0 of signatures for a given method name m. For example, calling
mtype(Φ1 , m, Φ1 .C) in the context of Figure 14 yields the list D → A, B → B.
6
Function append adds to each inner list of a list of field lists a given field. Its implementation
is straightforward and omitted for brevity.
23
In Figure 17, we show the definition of function mtype. For Base.Object, the empty
list is returned (ML-1). If a class that is sometimes reachable introduces a method in
question (ML-2), its signature is added to the result list and all possible predecessors
in the refinement chain (using pred ) and all possible subclasses are searched (using
last). Likewise, if a refinement that is sometimes reachable introduces a method with
the name searched (ML-3), its signature is added to the result list and all possible predecessors in the refinement chain are searched (using pred ). If a class or refinement does
not declare a corresponding method (ML-4 and ML-5) or the a class is never reachable,
the search proceeds with the possible superclasses or predecessors.
The current definition of function mtype returns possibly many duplicate signatures.
A straightforward optimization would be to remove duplicates before using the result
list, which we omitted for simplicity.
introduce(Ω, Φ.C)
Valid class introduction
CT (Ψ.C) = class C extends D { C f; M }
@Ψ :
Ψ 6= Φ
sometimes(Ω, Ψ)
introduce(Ω, Φ.C)
introduce(Ω, f, Φ.C)
Valid field introduction
∀ E h ∈ fields(Ω, Φ.C) : f ∈
/h
introduce(Ω, f, Φ.C)
introduce(Ω, m, Φ.C)
Valid method introduction
mtype(Ω, m, Φ.C) = •
introduce(Ω, m, Φ.C)
refine(Ω, Φ.C)
Valid class refinement
RT (C) = Ψ, Φ, Π
validref (Ω, C) a Ψ
refine(Ω, Φ.C)
Valid method overriding
override(Ω, m, Φ.C, C → C0 )
RT (C) = Ψ, Φ, Π
validref (Ω, C.m) a Ψ, Φ
∀ B → B0 ∈ mtype(Ω, m, Φ.C) : C = B ∧ C0 = B0
override(Ω, m, Φ.C, C → C0 )
Fig. 18. Valid introduction, refinement, and overriding in FFJPL .
24
Valid Introduction, Refinement, and Overriding. In Figure 18, we show predicates
for checking the validity of introduction, refinement, and overriding in FFJPL . Predicate
introduce indicates whether a class with the qualified type Φ.C has not been introduced
by any other feature Ψ that may be present in the context Ω. Likewise, introduce holds
if a method m or a field f has not been introduced by a qualified type Φ.C (including
possible predecessors and superclasses) that may be present in the given context Ω.
To this end, it checks either whether mtype yields the empty list or whether f is not
contained in every inner list returned by fields.
For a given refinement, predicate refine indicates whether a proper class, which is
always reachable in the given context, has been declared previously in the refinement
chain. We write validref (Ω, C) a Ψ in order to state that a declaration of class C
has been introduced in the set Ψ of features, which is only a subset of the features of
the product line, namely the features that precede the feature that introduces class C.
Predicate override indicates whether a declaration of method m has been introduced
(and is always reachable) in some feature introduced by before the feature that refines
m and whether every possible declaration of m in any predecessor of a Φ.C has the
same signature.
3.9
Type Relation of FFJPL
The type relation of FFJPL consists of type rules for terms and well-formedness rules
for classes, refinements, and methods, shown in Figure 19 and Figure 20.
Term Typing Rules. A term typing judgment in FFJPL is a quadruple, consisting of
a typing context Γ, a term t, a list of types C, and a feature Φ that contains the term
(see Figure 19). A term can have multiple types in a product line because there may
be multiple declarations of classes, fields, and methods. The list C contains all possible
types a term can have.
Rule T-VARPL is standard and does not refer to the feature model. It yields a list
consisting only of the type of the variable in question.
Rule T-F IELDPL checks whether a field access t0 .f is well-typed in every possible variant in which also Φ is present. Based on the possible types E of the term t0
the field f is accessed from, the rule checks whether f is always reachable from Φ
(using validref ). Note that this is a key mechanism of FFJPL ’s type system. It ensures that a field, being accessed, is definitely present in every valid program variant
in which the field access occurs – without generating all these variants. Furthermore,
all possible fields of all possible types E are assembled in a nested list F, C f, G in
which C f denotes a declaration of the field f; the call of fields(Φ, last(E)) is shorthand
for fields(Φ, last(E1 )) . . . fields(Φ, last(En )), in which the individual result lists are
concatenated. Finally, the list of all possible types C11 , . . . , Cn1 , . . . , C1m , . . . , Cnm of
field f becomes the list of types of the overall field access. Note that the result list may
contain duplicates, which could be eliminated for optimization purposes.
Rule T-I NVKPL checks whether a method invocation t0 .m(t) is well-typed in every
possible variant in which also Φ is present. Based on the possible types E of the term
t0 the method m is invoked on, the rule checks whether m is always reachable from Φ
25
Γ ` t:CaΦ
Term typing
x:C∈Γ
Γ ` x:CaΦ
(T-VARPL )
∀ E ∈ E : validref (Φ, E.f)
fields(Φ, last(E)) = F, C f, G
Γ ` t0 : E a Φ
Γ ` t0 .f : C11 , . . . , Cn1 , . . . , C1m , . . . , Cnm a Φ
∀ E ∈ E : validref (Φ, E.m)
Γ ` t0 : E a Φ
∀ C ∈ C, ∀ D ∈ D ∈ D : C <: D a Φ
Γ ` t:CaΦ
mtype(Φ, m, last(E)) = D → B
Γ ` t0 .m(t) : B11 , . . . , Bn1 , . . . , B1m , . . . , Bnm a Φ
validref (Φ, C)
∀ D g ∈ F , ∀ C ∈ C : C <: D a Φ
Γ ` t:CaΦ
fields(Φ, last(C)) = F
@∈
/F
Γ ` new C(t) : C a Φ
Γ ` t0 : E a Φ
(T-F IELDPL )
validref (Φ, C)
∀ E ∈ E : (E <: C a Φ ∨ C <: E a Φ)
Γ ` (C)t0 : C a Φ
validref (Φ, C)
stupid warning
Γ ` t0 : E a Φ
∃ E ∈ E : (C 6<: E a Φ ∧ E 6<: C a Φ)
Γ ` (C)t0 : C a Φ
Fig. 19. Term typing in FFJPL .
26
(T-I NVKPL )
(T-N EWPL )
(T-UDC ASTPL )
(T-SC ASTPL )
M OK a Φ.C
Method typing
x : B, this : C ` t0 : E a Φ
∀ E ∈ E : E <: B0 a Φ
introduce(Φ, m, last(D))
validref (Φ, B)
CT (Φ.C) = class C extends D { C f; M }
B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E a Φ
∀ E ∈ E : E <: B0 a Φ
validref (Φ, B)
override(Φ, m, last(D), B → B0 )
CT (Φ.C) = class C extends D { C f; M }
overrides B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E a Φ
∀ E ∈ E : E <: B0 a Φ
validref (Φ, B)
introduce(Φ, m, pred (Φ.C))
CT (Φ.C) = refines class C { C f; M }
B0 m(B x) { return t0 ; } OK a Φ.C
x : B, this : C ` t0 : E a Φ
∀ E ∈ E : E <: B0 a Φ
override(Φ, m, pred (Φ.C), B → B0 )
validref (Φ, B)
CT (Φ.C) = refines class C { C f; M }
overrides B0 m(B x) { return t0 ; } OK a Φ.C
L OK a Φ
Class typing
validref (Φ, D)
∀ f ∈ f : introduce(Φ, f, last(D))
validref (Φ, C)
introduce(Φ, Φ.C)
M OK a Φ.C
class C extends D { C f; M } OK a Φ
R OK a Φ
Refinement typing
validref (Φ, C)
∀ f ∈ f : introduce(Φ, f, pred (Φ.C))
refine(Φ, Φ.C)
refines class C { C f; M } OK a Φ
Fig. 20. Well-formedness rules of FFJPL .
27
M OK a Φ.C
(using validref ). As with field access, this check is essential. It ensures that in generated
programs only methods are invoked that are also present. Furthermore, all possible
signatures of m of all possible types E are assembled in the nested list D → B and it is
checked that all possible lists C of argument types of the method invocation are subtypes
of all possible lists D of parameter types of the method (this implies that the lengths of
the two lists must be equal). A method invocation has multiple types assembled in a list
that contains all result types of method m determined by mtype. As with field access,
duplicates should be eliminated for optimization purposes.
Rule T-N EWPL checks whether an object creation new C(t) is well-typed in every
possible variant in which also Φ is present. Specifically, it checks whether there is a
declaration of class C always reachable from Φ. Furthermore, all possible field combinations of C are assembled in the nested list F, and it is checked whether all possible
combinations of argument types passed to the object creation are subtypes of the types
of all possible field combinations (this implies that the number of arguments types must
equal the number of field types). The fields of the result list must not be annotated with
the marker @ since optional fields may not be present in every variant and references
may become invalid (see field lookup).7 An object creation has only a single type C.
Rules T-UDC ASTPL and T-SC ASTPL check whether casts are well-typed in every
possible variant in which also Φ is present. This is done by checking whether the type
C the term t0 is cast to is always reachable from Φ and whether this type is a subtype,
supertype, or unrelated type of all possible types E the term t0 can have. We have only
a single rule T-UDC ASTPL for up- and downcasts because the list E of possible types
may contain super- and subtypes of C simultaneously. If there is a type in the list which
leads to a stupid case, we flag a stupid warning. A cast yields a list containing only a
single type C.
Well-Formedness Rules. In Figure 20, we show the well-formedness rules of classes,
refinements, and methods.
Like in FFJ, the typing judgment of classes and refinements is a binary relation between a class or refinement declaration and a feature. The rule of classes checks whether
all methods are well-formed in the context of the class’ qualified type. Moreover, it
checks whether the class declaration is unique in the scope of the enclosing feature Φ,
i.e., whether no other feature, that may be present together with feature Φ, introduces
a class with an identical name (using introduce). Furthermore, it checks whether the
superclass and all field types are always reachable from Φ (using validref ). Finally, it
checks whether none of the fields of the class declaration have been introduced before
(using introduce). The well-formedness rule of refinements is analogous, except that
the rule checks that there is at least one class declaration reachable that is refined and
that has been introduced before the refinement (using refine).
The typing judgment of methods is a binary relation between a method declaration
and the qualified type that declares the method. Like in FFJ, there are four different
rules for methods (from top to bottom in Figure 20)
1. that do not override another method and that are declared by classes,
7
The treatment of @ is semiformal but simplifies the rule.
28
2. that override another method and that are declared by classes,
3. that do not override another method and that are declared by refinements,
4. that override another method and that are declared by refinements.
All four rules check whether all possible types E of the method body are subtypes of
the declared return type B0 of the method and whether the argument types B are always
reachable from the enclosing feature Φ (using validref ).
For methods that are introduced, it is checked, using introduce, whether no method
with identical name has been introduced in any possible superclass (Rule 1) or in any
possible predecessor in the refinement chain (Rule 3). For methods that override other
methods, it is checked, using override, whether a method with identical name and
signature exists in any possible superclass (Rule 2) or in any possible predecessor in the
refinement chain (Rule 4).
Well-Typed FFJPL Product Lines. An FFJPL product line, consisting of a term, a
class table, an introduction table, and a refinement table, is well-typed if
– the term is well-typed (checked using FFJPL ’s term typing rules),
– all classes and refinements stored in the class table are well-formed (checked using
FFJPL ’s well-formedness rules), and
– the class, introduction, and refinement tables are well-formed (ensured by the corresponding sanity conditions).
3.10
Type Safety of FFJPL
Type checking in FFJPL is based on information contained in the class table, introduction table, refinement table, and feature model. The first three are filled by the compiler
that has parsed the code base of the product line. The feature model is supplied directly
by the user (or tool). The compiler determines which class and refinement declarations
belong to which features. The classes and refinements of the class table are checked using their well-formedness rules which, in turn, use the well-formedness rules for methods and the term typing rules for method bodies. Several rules use the introduction and
refinement tables in order to map types, fields, and methods to features and the feature model to navigate along refinement chains and to check the presence of program
elements.
What does type safety mean in the context of a product line? The product line itself
is never evaluated; rather, different programs are derived that are then evaluated. Hence,
the property we are interested in is that all programs that can be derived from a welltyped product line are in turn well-typed. Furthermore, we would like to be sure that
all FFJPL product lines, from which only well-typed FFJ programs can be derived, are
well-typed. We formulate the two properties as the two theorems Correctness of FFJPL
and Completeness of FFJPL .
29
Correctness
T HEOREM 3.1 (Correctness of FFJPL ) Given a well-typed FFJPL product line pl (including with a well-typed term t, well-formed class, introduction, and refinement tables
CT , IT , and RT , and a feature model FM ), every program that can be derived with a
valid feature selection fs is a well-typed FFJ program (cf. Figure 10).
pl = (t, CT , IT , RT , FM )
pl is well-typed
derive(pl , fs) is well-typed
fs is valid in FM
Function derive collects the feature modules from a product line according to a
user’s selection fs, i.e., non-selected feature modules are removed from the derived
program. After this derivation step, the class table contains only classes and refinements
stemming from the selected feature modules. We define a valid feature selection to be
a list of features whose combination does not contradict the constraints implied by the
feature model.
The proof idea is to show that the type derivation tree of an FFJPL product line
is a superimposition of multiple so-called type derivation slices. As usual, the type
derivation proceeds from the root (i.e., an initial type rule that checks the term and all
classes and refinements of the class table) to the leaves (type rules that do not have a
premise) of the type derivation tree. Each time a term has multiple types, e.g., a method
has different alternative return types, which is caused by multiple mutually exclusive
method declarations, the type derivation splits into multiple branches. With branch we
refer only to positions in which the type derivation tree is split into multiple subtrees
in order to type check multiple mutually exclusive term definitions. Each subtree from
the root of the type derivation tree along the branches toward a leaf is a type derivation
slice. Each slice corresponds to the type derivation of a feature-oriented program.
Let us illustrate the concept of a type derivation slice by a simplified example. Suppose the application of an arbitrary type rule to a term t somewhere in the type derivation. Term t has multiple types C due to different alternative definitions of t’s subterms.
For simplicity, we assume here that t has only a single subterm t0 , like in the case of a
field access (t = t0 .f), in which the overall term t has multiple types depending on t0 ’s
and f’s types; the rule can be easily extended to multiple subterms by adding a predicate
per subterm. The type rule ensures the well-typedness of all possible variants of t on
the basis of the variants of t’s subterm t0 . Furthermore, the type rule checks whether a
predicate predicate (e.g., C <: D) holds for each variant of the subterm with its possible
types E, written predicate(t0 : Ei ). The possible types C of the overall term follow in
some way from the possible types E of its subterm. Predicate validref is used to check
whether all referenced elements and types are present in all valid variants, including
different combinations of optional features. For the general case, this can be written as
follows:
predicate(t0 : E1 ) predicate(t0 : E2 ) . . .
t0 : E
always(. . .)
Γ ` t:CaΦ
predicate(t0 : En )
(T-*PL )
The different uses of predicate in the premise of an FFJPL type rule correspond
to the branches in the type derivation that denote alternative definitions of subterms.
30
Hence, the premise of the FFJPL type rule is the conjunction of the different premises
that cover the different alternative definitions of the subterms of a term.
The proof strategy is as follows. Assuming that the FFJPL type system ensures that
each slice is a valid FFJ type derivation (see Lemma B.1 in Appendix B.1) and that
each valid feature selection corresponds to a single slice (since alternative features have
been removed; see Lemma B.2 in Appendix B.1), each feature-oriented program that
corresponds to a valid feature selection is guaranteed to be well-typed. Note that multiple valid feature selections may correspond to the same slice because of the presence
of optional features. It follows that, for every valid feature selection, we derive a wellformed FFJ program – since its type derivation is valid – whose evaluation satisfies the
properties of progress and preservation (see Appendix A). In Appendix B, we describe
the proof of Theorem 3.1 in more detail.
Completeness
T HEOREM 3.2 (Completeness of FFJPL ) Given an FFJPL product line pl (including
a well-typed term t, well-formed class, introduction, and refinement tables CT , IT ,
and RT , and a feature model FM ), and given that all valid feature selections fs yield
well-typed FFJ programs, according to Theorem 3.1, pl is a well-typed product line
according to the rules of FFJPL .
pl = (t, CT , IT , RT , FM ) ∀ fs : (fs is valid in FM ⇒ derive(pl , fs) is well-typed)
pl is well-typed
The proof idea is to examine three basic cases and to generalize subsequently: (1) pl
has only mandatory features; (2) pl has only mandatory features except a single optional
feature; (3) pl has only mandatory features except two mutually exclusive features.
All other cases can be formulated as combinations of these three basic cases. To this
end, we divide the possible relations between features into three disjoint sets: (1) a
feature is reachable from another feature in all variants, (2) a feature is reachable from
another feature in some, but not in all, variants, (3) two features are mutually exclusive.
From these three possible relations, we can prove the three basic cases in isolation and,
subsequently, construct a general case that can be phrased as a combination of the three
basic cases. The description of the general case and the reduction finish the proof of
Theorem 3.2. In Appendix B, we describe the proof of Theorem 3.2 in detail.
4
Implementation & Discussion
We have implemented FFJ and FFJPL in Haskell, including the program evaluation and
type checking of product lines. The FFJPL compiler expects a set of feature modules
and a feature model both of which, together, represent the product line. A feature module is represented by a directory. The files found inside a feature module’s directory are
assigned to / belong to the enclosing feature. The FFJPL compiler stores this information for type checking. Each file may contain multiple classes and class refinements.
In Figure 21, we show a snapshot of our test environment, which is based on Eclipse
31
and a Haskell plugin8 . We use Eclipse to interpret or compile our FFJ and FFJPL type
systems and interpreters. Specifically, the figure shows the directory structure of our
email system. The file EmailClient.features contains the user’s feature selection and
the feature model of the product line.
Fig. 21. Snapshot of the test environment of the Haskell implementation.
The feature model of a product line is represented by a propositional formula, following the approach of Batory [12] and Czarnecki and Pietroszek [22]. Propositional
formulas are an effective way of representing the relationships between features (e.g., of
specifying which feature implies the presence and absence of other features and of machine checking whether a feature selection is valid). For example, we have implemented
predicate sometimes as follows:
sometimes(FM , Ω, Ψ) = satisfiable(FM ∧ Ω1 ∧ . . . ∧ Ωn ∧ Ψ)
The feature model is an propositional formula; feature are variables; and satisfiable is
a satisfiability solver. Likewise, we have implemented predicate always on the basis of
logical reasoning on propositional formulas:
always(FM , Ω, Ψ) = ¬(satisfiable(¬(FM ⇒ ((Ω1 ∧ . . . ∧ Ωn ) ⇒ Ψ))))
For a more detailed explanation of how propositional formulas relate to feature models
and feature selections, we refer the interest to the work of Batory [12].
In Figure 22, we show the textual specification of the feature model of our email
system, which can be passed directly to the FFJPL compiler.
8
http://eclipsefp.sourceforge.net/haskell/
32
1 features:
2
EmailClient IMAP POP3 MIME SSL Text Mozilla Safari
3
4 model:
5
EmailClient implies (IMAP or POP3);
6
IMAP implies EmailClient;
7
POP3 implies EmailClient;
8
MIME implies EmailClient;
9
SSL implies EmailClient;
10
Text implies (IMAP or POP3);
11
Mozilla implies (IMAP or POP3);
12
Safari implies (IMAP or POP3);
13
Mozilla implies (not Safari);
14
Safari implies (not Mozilla);
Fig. 22. Feature model of an email client product line.
The first section (features:) of the file representing the feature model defines an
ordered set of names of the features of the product line and the second section (model:)
defines constraints on the features’ presence in the derived programs. In our example,
each email client supports either the protocols IMAP, POP3, or both. Furthermore, every
feature requires the presence of the base feature E MAIL C LIENT. Feature T EXT requires
either the presence of IMAP or POP3 or both – the same for M OZILLA and S AFARI.
Finally, feature M OZILLA requires the absence of feature S AFARI and vice versa.
On the basis of the feature modules and the feature model, FFJPL ’s type system
checks the entire product line and identifies valid program variants that still contain
type errors. A SAT solver is used to check whether elements are never, sometimes, or
always reachable. If an error is found, the product line is rejected as ill-formed. If not,
a feature-oriented program guaranteed to be well-formed can derived on the basis of a
user’s feature selection. This program can be evaluated using the standard evaluation
rules of FFJ, which we have also implemented in Haskell.
In contrast to previous work on type checking feature-oriented product lines [64,23],
our type system provides detailed error messages. This is possible due to the finegrained checks at the level of individual term typing and well-formedness rules. For
example, if a field access succeeds only in some program variants, this fact can be
reported to the user and the error message can point to the erroneous field access. Previously proposed type systems compose all code of all features of a product line and
extract a single propositional formula, which is checked for satisfiability. If the formula
is not satisfiable (i.e., a type error has occurred), it is not possible to identify the location
that has caused the error (at least not without further information). See Section 5, for a
detailed discussion of related approaches.
We made several tests and experiments with our Haskell implementation. However,
real-world tests were not feasible because of two reasons. First, in previous work it has
been already demonstrated that feature-oriented product lines require proper type systems and that type checking entire real-world product lines is feasible and useful [64].
Second, like FJ, FFJ is a core language into which all Java programs can be compiled
and which, by its relative simplicity, is suited for the formal definition and proof of
language properties – in our case, a type system and its correctness and completeness.
33
But, a core language is never suited for the development of real-world programs. This
is why our examples and test programs are of similar size and complexity as the FJ examples of Pierce [54]. Type checking our test programs required acceptable amounts of
time (in the order of magnitude of milliseconds per product line). We do not claim to be
able to handle full-sized feature-oriented product lines by hand-coding them in FFJPL .
Rather, this would require an expansion of the type system to full Java (including support for features as provided by AHEAD [13] or FeatureHouse [9]) – an enticing goal,
but one for the future (especially, as Java’s informal language specification [28] has 688
pages). Our work lays a foundation for implementing type systems in that it provides
evidence that core feature-oriented mechanisms are type sound and type systems of
feature-oriented product lines can be implemented correctly and completely.
Still, we would like to make some predictions on the scalability of our approach. The
novelty of our type system is that it incorporates alternative features and, consequently,
alternative definitions of classes, fields, and methods. This leads to a type derivation tree
with possibly multiple branches denoting alternative term types. Hence, performing a
type derivation of product line with many alternative features may consume a significant
amount of computation time and memory. It seems that this overhead is the price for
allowing alternative implementation of program parts.
Nevertheless, our approach minimizes the overhead caused by alternative features
compared to the naive approach. In the naive approach, all possible programs are derived and type checked subsequently. In our approach, we type check the entire code
base of the product line and branch the type derivation only at terms that really have
multiple, alternative types, and not at the level of entire program variants, as done in
the naive approach. Our experience with feature-oriented product lines shows that, usually, there are not many alternative features in a product line, but mostly optional features [42,3,64,37,59,11,9,5,6,57,60]. For example, in the Berkeley DB product line (JE
edition; 80 000 lines of code) there are 99 feature modules, but only two pairs of them
alternative [9,37]; in the Graph Product Line there are 26 feature modules, of which
only three pairs are alternative [42,9]. A further observation is that most alternative features that we encountered do not alter types. That is, there are multiple definitions of
fields and methods but with equal types. For example, GPL and Berkeley DB contain
alternative definitions of a few methods but only with identical signatures. Type checking these product lines with our approach, the type derivation would have almost no
branches. In the naive approach, still many program variants exist due to optional features. Hence, our approach is preferable. For example, in a product line with n features
and c ∗ n variants (with c being a constant), in our approach, the type system would
have to check n feature modules (with some few branches in the type derivation and
solving few simple SAT problems; see below) and, in the naive approach, the type system would have to check, at least, 2 ∗ n feature modules but, commonly, 2 ∗ n ∗ m with
m < n. For product lines with a higher degree of variability, e.g., with n2 or even 2n
variants the benefit of our approach becomes even more significant. We believe that this
benefit can make a difference in real world product line engineering.
A further point is that almost all typing and well-formedness rules contain calls to
the built-in SAT solver. This results in possibly many invocations of the SAT solver
at type checking time. Determining the satisfiability of a propositional formula is in
34
general an N P-complete problem. However, it has been shown that the structures of
propositional formulas occurring in software product lines are simple enough to scale
satisfiability solving to thousands of features [47]. Furthermore, in our experiments, we
have observed that many calls to the SAT solver are redundant, which is easy to see
when thinking about type checking feature-oriented product lines where the presence
of single types or members is checked in many type rules. We have implemented a
caching mechanism to decrease the number of calls to the SAT solver to a minimum.
Finally, the implementation in Haskell helped us a lot with the evaluation of the
correctness of our type rules. It can serve other researchers to reproduce and evaluate
our work and to experiment with further (feature-oriented) language mechanisms. The
implementations of FFJ and FFJPL , along with test programs, can be downloaded from
the Web.9
5
Related Work
We divide our discussions of related work into two parts: the implementation, formal
models, and type systems (1) of feature-oriented programs and (2) of feature oriented
product lines.
5.1
Feature-Oriented Programs
FFJ has been inspired by several feature-oriented languages and tools, most notably
AHEAD/Jak [13], FeatureC++ [10], FeatureHouse [9], and Prehofer’s feature-oriented
Java extension [55]. Their key aim is to separate the implementation of software artifacts, e.g., classes and methods, from the definition of features. That is, classes and refinements are not annotated or declared to belong to a feature. There is no statement in
the program text that defines explicitly a connection between code and features. Instead,
the mapping of software artifacts to features is established via so-called containment hierarchies, which are basically directories containing software artifacts. The advantage
of this approach is that a feature’s implementation can include, beside classes in the
form of Java files, also other supporting documents, e.g., documentation in the form of
HTML files, grammar specifications in the form of JavaCC files, or build scripts and
deployment descriptors in the form of XML files [13]. To this end, feature composition
merges not only classes with their refinements but also other artifacts, such as HTML
or XML files, with their respective refinements [2,9].
Another class of programming languages that provide mechanisms for the definition
and extension of classes and class hierarchies includes, e.g., ContextL [29], Scala [52],
and Classbox/J [14]. The difference to feature-oriented languages is that they provide
explicit language constructs for aggregating the classes that belong to a feature, e.g.,
family classes, classboxes, or layers. This implies that non-code software artifacts cannot be included in a feature [11]. However, FFJ still models a subset of these languages,
in particular, class refinement.
Similarly, related work on a formalization of the key concepts underlying featureoriented programming has not disassociated the concept of a feature from the level of
9
http://www.fosd.de/ffj
35
code. Especially, calculi for mixins [26,16,1,34], traits [41], family polymorphism and
virtual classes [33,25,30,18], path-dependent types [52,51], open classes [20], dependent classes [27], and nested inheritance [50] either support only the refinement of single classes or expect the classes that form a semantically coherent unit (i.e., that belong
to a feature) to be located in a physical module that is defined in the host programming
language. For example, a virtual class is by definition an inner class of the enclosing
object, and a classbox is a package that aggregates a set of related classes. Thus, FFJ
differs from previous approaches in that it relies on contextual information that has been
collected by the compiler, e.g., the features’ composition order or the mapping of code
to features.
A different line of research aims at the language-independent reasoning about features [13,44,9,39]. The calculus gDeep is most closely related to FFJ since it provides a
type system for feature-oriented languages that is language-independent [4]. The idea is
that the recursive process of merging software artifacts, when composing hierarchically
structured features, is very similar for different host languages, e.g., for Java, C#, and
XML. The calculus describes formally how feature composition is performed and what
type constraints have to be satisfied. In contrast, FFJ does not aspire to be languageindependent, although the key concepts can certainly be used with different languages.
The advantage of FFJ is that its type system can be used to check whether terms of
the host language (Java or FJ) violate the principles of feature orientation, e.g., whether
methods refer to classes that have been added by other features. Due to its language
independence, gDeep does not have enough information to perform such checks.
5.2
Feature-Oriented Product Lines
Our work on type checking feature-oriented product lines was motivated by the work
of Thaker et al. [64]. They suggested the development of a type system for featureoriented product lines that does not check all individual programs but the individual
feature implementations. They have implemented an (incomplete) type system and, in a
number of case studies on real product lines, they found numerous hidden errors using
their type rules. Nevertheless, the implementation of their type system is ad-hoc in the
sense that it is described only informally, and they do not provide a correctness and
completeness proof. Our type system has been inspired by their work and we were able
to provide a formalization and a proof of type safety.
In a parallel line of work, Delaware et al. have developed a formal model of a
feature-oriented language, called Lightweight Feature Java (LFJ), and a type system
for feature-oriented product lines [23]. Their work was also influenced by the practical
work of Thaker et al. So, it is not surprising that it is closest to ours. However, there are
numerous differences. First, their formal model of a feature-oriented language is based
on Lightweight Java (LJ) [62] and not on Featherweight Java (FJ). While LJ is more
expressive, it is also more complex. We decided for the simpler variant FJ, omitting,
e.g., constructors and mutable state. Second, Delaware et al. do not model featureoriented mechanisms, such as class or method refinements, directly in the semantics
and type rules of the language. Instead, they introduce a transformation step in which
LFJ code is “compiled down” to LJ code, i.e., they flatten refinement chains to single
classes. Proceeding likewise, we would have to generate first an FJ program from an
36
FFJ product line and type check the FJ program (that consists of some or all possible
features of the product line) subsequently. We refrained from such a transformation
step in order to model the semantics of feature-oriented mechanisms directly in terms
of dedicated field and method lookup mechanisms as well as special well-formed rules
for method and class refinements.
Lagorio et al. have shown that a flattening semantics and a direct semantics are
equivalent [40]. An advantage of a “direct” semantics is that it allows a type checking
and error reporting at a finer grain. In LFJ, all feature modules are composed and a
single propositional formula is generated and tested for satisfiability; if the formula
is not satisfiable, it is difficult to identify precisely the point of failure. In FFJPL , the
individual type rules consult the feature model and can point directly to the point of
failure.
A further advantage of our approach is that it leaves open when feature composition is performed. Currently, feature composition is modeled in FFJ/FFJPL as a static
process done before compilation but, with our approach, it becomes possible to model
dynamic feature composition at run time [58,53] by making the class and feature tables
and the feature model dynamic, i.e., allowing them to change during a computation.
With LFJ this is not possible. Hutchins has shown that feature composition can be performed by an interpreter and partial evaluation can be used to pre-evaluate the parts of
a composition that are static [31]. However, Delaware et al. have developed a machinechecked model of their type system formalized with the theorem prover Coq [15]. Our
proof is hand-written, but we have a Haskell implementation of the FFJ and FFJPL
calculi that we have tested thoroughly.
Even previously to the work of Thaker et al., Czarnecki et al. presented an automatic
verification procedure for ensuring that no ill-structured UML model template instances
will be generated from a valid feature selection [22]. That is, they type check product
lines that consist not of Java programs but of UML models. They use OCL (object
constraint language) constraints to express and implement a type system for model
composition. In this sense, their aim is very similar to that of FFJPL , but limited to
model artifacts – although they have proposed to generalize their work to programming
languages.
Kästner et al. have implemented a tool, called CIDE, that allows a developer to
decompose a software system into features via annotations [38]. In contrast to other
feature-oriented languages and tools, the link between code and features is established
via annotations. If a user selects a set of features, all code that is annotated with features
(using background colors) that are not present in the selection is removed. Kästner et
al. have developed a formal calculus and a set of type rules that ensure that only welltyped programs can be generated from a valid feature selection [36]. For example, if
a method declaration is removed, the remaining code must not contain calls to this
method. CIDE’s type rules are related to the type rules of FFJPL but, so far, mutually
exclusive features are not supported in CIDE. In some sense, FFJPL and CIDE represent
two sides of the same coin: the former aims at the composition of feature modules, the
latter at the annotation of feature-related code.
37
6
Conclusion
A feature-oriented product line imposes severe challenges on type checking. The naive
approach of checking all individual programs of a product line is not feasible because
of the combinatorial explosion of program variants. Hence, the only practical option is
to check the entire code base of a product line, including all features, and, based on the
information of which feature combinations are valid, to ensure that it is not possible to
derive a valid program variant that contains type errors.
We have developed such a type system based on a formal model of a featureoriented Java-like language, called Feature Featherweight Java (FFJ). A distinguishing
property of our work is that we have modeled the semantics and type rules for core
feature-oriented mechanisms directly, without compiling feature-oriented code down to
a lower-level representation such as object-oriented Java code. The direct semantics allows us to reason about core feature-oriented mechanisms in terms of themselves and
not of generated lower-level code. A further advantage is the fine-grained error reporting and that the time of feature composition may vary between compile time and run
time.
We have demonstrated and proved that, based on a valid feature selection, our type
system ensures that every program of a feature-oriented product line is well-formed and
that our type system is complete. Our implementation of FFJ, including the type system
for product lines, indicates the feasibility of our approach and can serve as a testbed for
experimenting with further feature-oriented mechanisms.
Acknowledgment
This work is being funded in part by the German Research Foundation (DFG), project
number AP 206/2-1.
References
1. D. Ancona, G. Lagorio, and E. Zucca. Jam—Designing a Java Extension with Mixins. ACM
Transactions on Programming Languages and Systems (TOPLAS), 25(5):641–712, 2003.
2. F. Anfurrutia, O. Díaz, and S. Trujillo. On Refining XML Artifacts. In Proceedings of the
International Conference on Web Engineering (ICWE), volume 4607 of LNCS, pages 473–
478. Springer-Verlag, 2007.
3. S. Apel and K. Böhm. Towards the Development of Ubiquitous Middleware Product Lines.
In Software Engineering and Middleware, volume 3437 of LNCS, pages 137–153. SpringerVerlag, 2004.
4. S. Apel and D. Hutchins. An Overview of the gDeep Calculus. Technical Report MIP-0712,
Department of Informatics and Mathematics, University of Passau, 2007.
5. S. Apel, F. Janda, S. Trujillo, and C. Kästner. Model Superimposition in Software Product
Lines. In Proceedings of the International Conference on Model Transformation (ICMT),
volume 5563 of LNCS, pages 4–19. Springer-Verlag, 2009.
6. S. Apel, C. Kästner, A. Größlinger, and C. Lengauer. Feature (De)composition in Functional
Programming. In Proceedings of the International Conference on Software Composition
(SC), volume 5634 of LNCS, pages 9–26. Springer-Verlag, 2009.
38
7. S. Apel, C. Kästner, and C. Lengauer. An Overview of Feature Featherweight Java. Technical
Report MIP-0802, Department of Informatics and Mathematics, University of Passau, 2008.
8. S. Apel, C. Kästner, and C. Lengauer. Feature Featherweight Java: A Calculus for FeatureOriented Programming and Stepwise Refinement. In Proceedings of the International Conference on Generative Programming and Component Engineering (GPCE), pages 101–112.
ACM Press, 2008.
9. S. Apel, C. Kästner, and C. Lengauer. FeatureHouse: Language-Independent, Automated
Software Composition. In Proceedings of the International Conference on Software Engineering (ICSE), pages 221–231. IEEE CS Press, 2009.
10. S. Apel, T. Leich, M. Rosenmüller, and G. Saake. FeatureC++: On the Symbiosis of FeatureOriented and Aspect-Oriented Programming. In Proceedings of the International Conference
on Generative Programming and Component Engineering (GPCE), volume 3676 of LNCS,
pages 125–140. Springer-Verlag, 2005.
11. S. Apel, T. Leich, and G. Saake. Aspectual Feature Modules. IEEE Transactions on Software
Engineering (TSE), 34(2):162–180, 2008.
12. D. Batory. Feature Models, Grammars, and Propositional Formulas. In Proceedings of the
International Software Product Line Conference (SPLC), volume 3714 of LNCS, pages 7–20.
Springer-Verlag, 2005.
13. D. Batory, J. Sarvela, and A. Rauschmayer. Scaling Step-Wise Refinement. IEEE Transactions on Software Engineering (TSE), 30(6):355–371, 2004.
14. A. Bergel, S. Ducasse, and O. Nierstrasz. Classbox/J: Controlling the Scope of Change in
Java. In Proceedings of the International Conference on Object-Oriented Programming,
Systems, Languages, and Applications (OOPSLA), pages 177–189. ACM Press, 2005.
15. Y. Bertot and P. Casteran. Interactive Theorem Proving and Program Development –
Coq’Art: The Calculus of Inductive Constructions. Texts in Theoretical Computer Science.
An EATCS Series. Springer-Verlag, 2004.
16. V. Bono, A. Patel, and V. Shmatikov. A Core Calculus of Classes and Mixins. In Proceedings
of the European Conference on Object-Oriented Programming (ECOOP), volume 1628 of
LNCS, pages 43–66. Springer-Verlag, 1999.
17. G. Bracha and W. Cook. Mixin-Based Inheritance. In Proceedings of the European Conference on Object-Oriented Programming (ECOOP) and International Conference on ObjectOriented Programming Systems, Languages, and Applications (OOPSLA), pages 303–311.
ACM Press, 1990.
18. D. Clarke, S. Drossopoulou, J. Noble, and T. Wrigstad. Tribe: A Simple Virtual Class Calculus. In Proceedings of the International Conference on Aspect-Oriented Software Development (AOSD), pages 121–134. ACM Press, 2007.
19. P. Clements and L. Northrop. Software Product Lines: Practices and Patterns. AddisonWesley, 2002.
20. C. Clifton, T. Millstein, G. Leavens, and C. Chambers. MultiJava: Design Rationale, Compiler Implementation, and Applications. ACM Transactions on Programming Languages and
Systems (TOPLAS), 28(3):517–575, 2006.
21. K. Czarnecki and U. Eisenecker. Generative Programming: Methods, Tools, and Applications. Addison-Wesley, 2000.
22. K. Czarnecki and K. Pietroszek. Verifying Feature-Based Model Templates Against WellFormedness OCL Constraints. In Proceedings of the International Conference on Generative
Programming and Component Engineering (GPCE), pages 211–220. ACM Press, 2006.
23. B. Delaware, W. Cook, and D. Batory. A Machine-Checked Model of Safe Composition. In
Proceedings of the International Workshop on Foundations of Aspect-Oriented Languages
(FOAL), pages 31–35. ACM Press, 2009.
39
24. S. Ducasse, O. Nierstrasz, N. Schärli, R. Wuyts, and A. Black. Traits: A Mechanism for FineGrained Reuse. ACM Transactions on Programming Languages and Systems (TOPLAS),
28(2):331–388, 2006.
25. E. Ernst, K. Ostermann, and W. Cook. A Virtual Class Calculus. In Proceedings of the
International Symposium on Principles of Programming Languages (POPL), pages 270–
282. ACM Press, 2006.
26. M. Flatt, S. Krishnamurthi, and M. Felleisen. Classes and Mixins. In Proceedings of the
International Symposium on Principles of Programming Languages (POPL), pages 171–
183. ACM Press, 1998.
27. V. Gasiunas, M. Mezini, and K. Ostermann. Dependent Classes. In Proceedings of the
International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), pages 133–152. ACM Press, 2007.
28. J. Gosling, B. Joy, G. Steele, and G. Bracha. The Java Language Specification. The Java
Series. Addison-Wesley, 3rd edition, 2005.
29. R. Hirschfeld, P. Costanza, and O. Nierstrasz. Context-Oriented Programming. Journal of
Object Technology (JOT), 7(3):125–151, 2008.
30. D. Hutchins. Eliminating Distinctions of Class: Using Prototypes to Model Virtual Classes.
In Proceedings of the International Conference on Object-Oriented Programming, Systems,
Languages, and Applications (OOPSLA), pages 1–19. ACM Press, 2006.
31. D. Hutchins. Pure Subtype Systems: A Type Theory For Extensible Software. PhD thesis,
School of Informatics, University of Edinburgh, 2008.
32. A. Igarashi, B. Pierce, and P. Wadler. Featherweight Java: A Minimal Core Calculus for Java
and GJ. ACM Transactions on Programming Languages and Systems (TOPLAS), 23(3):396–
450, 2001.
33. A. Igarashi, C. Saito, and M. Viroli. Lightweight Family Polymorphism. In Proceedings of
the Asian Symposium on Programming Languages and Systems (APLAS), volume 3780 of
LNCS, pages 161–177. Springer-Verlag, 2005.
34. T. Kamina and T. Tamai. McJava – A Design and Implementation of Java with Mixin-Types.
In Proceedings of the Asian Symposium on Programming Languages and Systems (APLAS),
volume 3302 of LNCS, pages 398–414. Springer-Verlag, 2004.
35. K. Kang, S. Cohen, J. Hess, W. Novak, and A. Peterson. Feature-Oriented Domain Analysis
(FODA) Feasibility Study. Technical Report CMU/SEI-90-TR-21, Software Engineering
Institute, Carnegie Mellon University, 1990.
36. C. Kästner and S. Apel. Type-Checking Software Product Lines – A Formal Approach.
In Proceedings of the International Conference on Automated Software Engineering (ASE),
pages 258–267. IEEE CS Press, 2008.
37. C. Kästner, S. Apel, and D. Batory. A Case Study Implementing Features using AspectJ. In
Proceedings of the International Software Product Line Conference (SPLC), pages 222–232.
IEEE CS Press, 2007.
38. C. Kästner, S. Apel, and M. Kuhlemann. Granularity in Software Product Lines. In Proceedings of the International Conference on Software Engineering (ICSE), pages 311–320. ACM
Press, 2008.
39. C. Kästner, S. Apel, S. Trujillo, M. Kuhlemann, and D. Batory. Guaranteeing Syntactic Correctness for all Product Line Variants: A Language-Independent Approach. In Proceedings of
the International Conference on Objects, Models, Components, Patterns (TOOLS EUROPE),
volume 33 of LNBI, pages 174–194. Springer-Verlag, 2009.
40. G. Lagorio, M. Servetto, and E. Zucca. Featherweight Jigsaw – A Minimal Core Calculus
for Modular Composition of Classes. In Proceedings of the European Conference on ObjectOriented Programming (ECOOP), LNCS. Springer-Verlag, 2009.
41. L. Liquori and A. Spiwack. FeatherTrait: A Modest Extension of Featherweight Java. ACM
Transactions on Programming Languages and Systems (TOPLAS), 30(2):1–32, 2008.
40
42. R. Lopez-Herrejon and D. Batory. A Standard Problem for Evaluating Product-Line Methodologies. In Proceedings of the International Conference on Generative and ComponentBased Software Engineering (GCSE), volume 2186 of LNCS, pages 10–24. Springer-Verlag,
2001.
43. R. Lopez-Herrejon, D. Batory, and W. Cook. Evaluating Support for Features in Advanced Modularization Technologies. In Proceedings of the European Conference on ObjectOriented Programming (ECOOP), volume 3586 of LNCS, pages 169–194. Springer-Verlag,
2005.
44. R. Lopez-Herrejon, D. Batory, and C. Lengauer. A Disciplined Approach to Aspect Composition. In Proceedings of the International Symposium Partial Evaluation and SemanticsBased Program Manipulation (PEPM), pages 68–77. ACM Press, 2006.
45. O. Madsen and B. Moller-Pedersen. Virtual Classes: A Powerful Mechanism in ObjectOriented Programming. In Proceedings of the International Conference on Object-Oriented
Programming, Systems, Languages, and Applications (OOPSLA), pages 397–406. ACM
Press, 1989.
46. H. Masuhara and G. Kiczales. Modeling Crosscutting in Aspect-Oriented Mechanisms. In
Proceedings of the European Conference on Object-Oriented Programming (ECOOP), volume 2743 of LNCS, pages 2–28. Springer-Verlag, 2003.
47. M. Mendonca, A. Wasowski, and K. Czarnecki. SAT-based Analysis of Feature Models
is Easy. In Proceedings of the International Software Product Line Conference (SPLC).
Software Engineering Institute, Carnegie Mellon University, 2009.
48. M. Mezini and K. Ostermann. Variability Management with Feature-Oriented Programming
and Aspects. In Proceedings of the International Symposium on Foundations of Software
Engineering (FSE), pages 127–136. ACM Press, 2004.
49. G. Murphy, A. Lai, R. Walker, and M. Robillard. Separating Features in Source Code: An
Exploratory Study. In Proceedings of the International Conference on Software Engineering
(ICSE), pages 275–284. IEEE CS Press, 2001.
50. N. Nystrom, S. Chong, and A. Myers. Scalable Extensibility via Nested Inheritance. In Proceedings of the International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), pages 99–115. ACM Press, 2004.
51. M. Odersky, V. Cremet, C. Röckl, and M. Zenger. A Nominal Theory of Objects with Dependent Types. In Proceedings of the European Conference on Object-Oriented Programming
(ECOOP), volume 2743 of LNCS, pages 201–224. Springer-Verlag, 2003.
52. M. Odersky and M. Zenger. Scalable Component Abstractions. In Proceedings of the International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), pages 41–57. ACM Press, 2005.
53. K. Ostermann. Dynamically Composable Collaborations with Delegation Layers. In Proceedings of the European Conference on Object-Oriented Programming (ECOOP), volume
2374 of LNCS, pages 89–110. Springer-Verlag, 2002.
54. B. Pierce. Types and Programming Languages. MIT Press, 2002.
55. C. Prehofer. Feature-Oriented Programming: A Fresh Look at Objects. In Proceedings of the
European Conference on Object-Oriented Programming (ECOOP), volume 1241 of LNCS,
pages 419–443. Springer-Verlag, 1997.
56. T. Reenskaug, E. Andersen, A. Berre, A. Hurlen, A. Landmark, O. Lehne, E. Nordhagen,
E. Ness-Ulseth, G. Oftedal, A. Skaar, and P. Stenslet. OORASS: Seamless Support for the
Creation and Maintenance of Object-Oriented Systems. Journal of Object-Oriented Programming (JOOP), 5(6):27–41, 1992.
57. M. Rosenmüller, C. Kästner, N. Siegmund, S. Sunkle, S. Apel, T. Leich, and G. Saake. SQL
á la Carte – Toward Tailor-made Data Management. In Datenbanksysteme in Business, Technologie und Web – Fachtagung des GI-Fachbereichs Datenbanken und Informationssysteme,
volume P-144 of GI-Edition – LNI, pages 117–136. Gesellschaft für Informatik, 2009.
41
58. M. Rosenmüller, N. Siegmund, G. Saake, and S. Apel. Code Generation to Support Static
and Dynamic Composition of Software Product Lines. In Proceedings of the International
Conference on Generative Programming and Component Engineering (GPCE), pages 3–12.
ACM Press, 2008.
59. M. Rosenmüller, N. Siegmund, H. Schirmeier, J. Sincero, S. Apel, T. Leich, O. Spinczyk,
and G. Saake. FAME-DBMS: Talor-made Data Management Solutions for Embedded Systems. In Proceedings of the EDBT Workshop on Software Engineering for Tailor-made Data
Management (SETMDM), pages 1–6. ACM Press, 2008.
60. N. Siegmund, C. Kästner, M. Rosenmüller, F. Heidenreich, S. Apel, and G. Saake. Bridging
the Gap between Variability in Client Application and Database Schema. In Datenbanksysteme in Business, Technologie und Web – Fachtagung des GI-Fachbereichs Datenbanken
und Informationssysteme, volume P-144 of GI-Edition – LNI, pages 297–306. Gesellschaft
für Informatik, 2009.
61. Y. Smaragdakis and D. Batory. Mixin Layers: An Object-Oriented Implementation Technique for Refinements and Collaboration-Based Designs. ACM Transactions on Software
Engineering and Methodology (TOSEM), 11(2):215–255, 2002.
62. R. Strniša, P. Sewell, and M. Parkinson. The Java Module System: Core Design and Semantic Definition. In Proceedings of the International Conference on Object-Oriented Programming, Systems, Languages, and Applications (OOPSLA), pages 499–514. ACM Press,
2007.
63. P. Tarr, H. Ossher, W. Harrison, and S. Sutton, Jr. N Degrees of Separation: MultiDimensional Separation of Concerns. In Proceedings of the International Conference on
Software Engineering (ICSE), pages 107–119. IEEE CS Press, 1999.
64. S. Thaker, D. Batory, D. Kitchin, and W. Cook. Safe Composition of Product Lines. In
Proceedings of the International Conference on Generative Programming and Component
Engineering (GPCE), pages 95–104. ACM Press, 2007.
65. M. VanHilst and D. Notkin. Using Role Components in Implement Collaboration-based
Designs. In Proceedings of the International Conference on Object-Oriented Programming,
Systems, Languages, and Applications (OOPSLA), pages 359–369. ACM Press, 1996.
66. A. K. Wright and M. Felleisen. A Syntactic Approach to Type Soundness. Information and
Computation, 115(1):38–94, 1994.
A
Type Soundness Proof of FFJ
Before giving the main proof, we state and proof some required lemmas.
L EMMA A.1 If mtype(m, last(D)) = C → C0 , then mtype(m, last(C)) = C → C0 for
all C <: D.
Proof Straightforward induction on the derivation of C <: D. There are two cases:
First, if method m is not defined in the declaration or in any refinement of class C,
then mtype(m, last(C)) should be the same as mtype(m, last(E)) where CT (Φ.C) =
class C extends E { . . . } for some Φ. This follows from the definition of mtype that
searches E’s refinement chain from right to left if m is not declared in C’s refinement
chain. Second, if m is defined in the declaration or in any refinement of class C, then
mtype(m, last(C)) should also be the same as mtype(m, last(E)) with CT (Φ.C) =
class C extends E { . . . } for some Φ. This case is covered by the well-formedness
rules for methods that use the predicate override to ensure that m is properly overridden, i.e., the signatures of the overridden and the overriding declaration of m are equal,
and that m is not introduced twice, i.e., overloading is not allowed in FFJ.
t
u
42
L EMMA A.2 (Term substitution preserves typing) If Γ, x : B ` t : D and Γ, s : A,
where A <: B, then Γ ` [x 7→ s] t : C for some C <: D.
Proof By induction on the derivation of Γ, x : B ` t : D.
C ASE (T-VAR ) t = x
x:D∈Γ
If x 6∈ x, then the result is trivial since [x 7→ s] x = x.10 On the other hand, if x = xi
and D = Bi , then, since [x 7→ s] x = si , letting C = Ai finishes the case.
C ASE (T-F IELD ) t = t0 .fi
Γ, x : B ` t0 : D0
fields(last(D0 )) = C f
D = Ci
By the induction hypothesis, there is some C0 such that Γ ` [x 7→ s] t0 : C0 and
C0 <: D0 . It is easy to check that fields(last(C0 )) = (fields(last(D0 )), D g) for some
D g. Therefore, by T-F IELD, Γ ` ([x 7→ s] t0 ).fi : Ci . The fact that the refinements of
a class may add new fields does not cause problems. D g contains all fields that C0 ,
including all of its refinements, add to D0 .
C ASE (T-I NVK ) t = t0 .m(t) Γ, x : B ` t0 : D0
Γ, x : B ` t : D D <: E
mtype(m, last(D0 )) = E → D
By the induction hypothesis, there are some C0 and C such that:
Γ ` [x 7→ s] t0 : C0
C0 <: D0
Γ ` [x 7→ s] t : C C <: D.
By Lemma A.1, we have mtype(m, last(C0 )) = E → D. Moreover, C <: E by the
transitivity of <: . Therefore, by T-I NVK, Γ ` [x 7→ s] t0 .m([x 7→ s] t) : D. The key is
that subclasses and refinements may override methods but the well-formedness rules of
methods ensure that the method’s type is not altered, i.e., there is no overloading in FFJ.
C ASE (T-N EW ) t = new D(t) fields(last(D)) = D f
Γ, x : B ` t : C C <: D
By the induction hypothesis, Γ ` [x 7→ s] t : E for some E with E <: C. We have E <: D
by the transitivity of <: . Therefore, by rule T-N EW, Γ ` new D([x 7→ s] t) : D. Although refinements of class D may add new fields, rule T-N EW ensures that the arguments of the object creation match the overall fields of D, including all refinements, in
number and types. That is, the number of arguments (t) equals the number of fields (f)
which function fields returns.
C ASE (T-UC AST ) t = (D)t0
Γ, x : B ` t0 : C
C <: D
By the induction hypothesis, there is some E such that Γ ` [x 7→ s] t0 : E and E <: C.
We have E <: D by the transitivity of <: , which yields Γ ` (D)( [x 7→ s] t0 ) : D by
T-UC AST.
C ASE (T-DC AST ) t = (D)t0
10
Γ, x : B ` t0 : C
D <: C
D 6= C
Note that [x 7→ s] x is an abbreviation for [x1 7→ s1 , . . . , xn 7→ sn ] x. It means that all occurrences of the variables x1 , . . . , xn in the term x are substituted with the corresponsing terms
s1 , . . . , sn .
43
By the induction hypothesis, there is some E such that Γ ` [x 7→ s] t0 : E and E <: C.
If E <: D or D <: E, then Γ ` (D)([x 7→ s] t0 ) : D by T-UC AST or T-DC AST, respectively. If both D 6<: E and E 6<: D, then Γ ` (D)([x 7→ s] t0 ) : D (with a stupid warning)
by T-SC AST.
C ASE (T-SC AST ) t = (D)t0
Γ, x : B ` t0 : C
D 6<: C
C 6<: D
By the induction hypothesis, there is some E such that Γ ` [x 7→ s] t0 : E and E <: C.
This means that E 6<: D because, in FFJ, each class has just one superclass and, if both
E <: C and E <: D, then either C <: D or D <: C, which contradicts the induction hypothesis. So Γ ` (D)( [x 7→ s] t0 ) : D (with a stupid warning), by T-SC AST.
t
u
L EMMA A.3 (Weakening) If Γ ` t : C, then Γ, x : D ` t : C
Proof Straightforward induction. The proof for FFJ is similar to the proof for FJ.
t
u
L EMMA A.4 If mtype(m, last(C0 )) = D → D, and mbody(m, last(C0 )) = (x, t), then
for some D0 and some C <: D we have C0 <: D0 and x : D, this : D0 ` t : C .
Proof By induction on the derivation of mbody(m, last(C0 )). The base case (in which
m is defined in the most specific refinement of C0 ) is easy since m is defined in
CT (last(C0 )) and the well-formedness of the class table implies that we must have
derived x : D, this : C0 ` t : C by the well-formedness rules of methods. The induction
step is also straightforward: if m is not defined in CT (last(C0 )), then mbody searches
the refinement chain from right to left; if m has not been found, the superclass’ refinement chain is searched. There are two subcases: first, m is defined in the declaration or
in any refinement of C0 ; this case is similar to the base case. Second, m is defined in a
superclass D0 of C0 or in one of D0 ’s refinements; in this case, the well-formedness of
the class table implies that we must have derived x : D, this : D0 ` t : C by the wellformedness rules of methods, which finishes the case.
t
u
Note that this lemma holds because method refinements do not change the types of
the arguments and the result of a method, overloading is not allowed, and this points
always to the class that is introduced or refined.
T HEOREM A.1 (Preservation) If Γ ` t : C and t −→ t0 , then Γ ` t0 : C0 for some
C0 <: C.
Proof By induction on a derivation of t −→ t0 , with a case analysis on the final rule.
C ASE (E-P ROJ N EW ) t = new C0 (v).fi
t0 = vi
fields(last(C0 )) = D f
From the shape of t, we see that the final rule in the derivation of Γ ` t : C must be
T-F IELD, with premise Γ ` new C0 (v) : D0 , for some D0 , and that C = Di . Similarly,
the last rule in the derivation of Γ ` new C0 (v) : D0 must be T-N EW, with premises
Γ ` v : C and C <: D, and with D0 = C0 . In particular, Γ ` vi : Ci , which finishes the
case, since Ci <: Di .
C ASE (E-I NVK N EW ) t = (new C0 (v)).m(u)
mbody(m, last(C0 )) = (x, t0 )
44
t0 = [x 7→ u, this 7→ new C0 (v)] t0
The final rules in the derivation of Γ ` t : C must be T-I NVK and T-N EW, with premises
Γ ` new C0 (v) : C0 , Γ ` u : C, C <: D, and mtype(m, last(C0 )) = D → C. By
Lemma A.4, we have x : D, this : D0 ` t : B for some D0 and B, with C0 <: D0 and
B <: C. By Lemma A.3, Γ, x : D, this : D0 ` t0 : B. Then, by Lemma A.2, we have
Γ [x 7→ u, this 7→ new C0 (v)] t0 : E for some E <: B. By the transitivity of <: , we obtain E <: C. Letting C0 = E completes the case.
C ASE (E-C AST N EW ) t = (D)(new C0 (v))
C0 <: D
t0 = new C0 (v)
The proof of Γ ` (D)(new C0 (v)) : C must end with T-UC AST since ending with TSC AST or T-DC AST would contradict the assumption of C0 <: D. The premises of
T-UC AST give us Γ ` new C0 (v) : C0 and D = C, finishing the case.
The cases for the congruence rules are easy. We show just the case E-C AST.
C ASE (E-C AST ) t = (D)t0
t0 = (D)t00
t0 −→ t00
There are three subcases according to the last typing rule used.
S UBCASE (T-UC AST ) Γ ` t0 : C0
C0 <: D
t00
D=C
By the induction hypothesis, Γ `
:
for some C00 <: C0 . By transitivity of <: ,
0
0
C0 <: C. Therefore, by T-UC AST, Γ ` (C)t0 : C (with no additional stupid warning).
S UBCASE (T-DC AST ) Γ ` t0 : C0
t00
C00
D <: C0
D=C
By the induction hypothesis, Γ ` :
for some C00 <: C0 . If C00 <: C or C <: C00 ,
0
then Γ ` (C)t0 : C by T-UC AST or T-DC AST (without any additional stupid warning).
On the other hand, if both C00 6<: C or C 6<: C00 , then Γ ` (C)t00 : C with a stupid warning
by T-SC AST.
S UBCASE (T-SC AST ) Γ ` t0 : C0
t00
C00
D 6<: C0
C00
C0 6<: D
D=C
C00
By the induction hypothesis, Γ ` :
for some
<: C0 . Then, also C00 6<: C and
0
0
C 6<: C0 . Therefore Γ ` (C)t0 : C with a stupid warning. If C00 6<: C, then C 6<: C00
since C 6<: C0 and, therefore, Γ ` (C)t00 : C with stupid war ning. If C00 <: C, then Γ `
(C)t00 : C by T-UC AST (with no additional stupid warning). This subcase is analogous
to the case T-SC AST of the proof of Lemma A.2.
t
u
T HEOREM A.2 (Progress) Suppose t is a well-typed term.
1. If t includes new C0 (t).fi as a subterm, then fields(last(C0 )) = C f for some C and
f.
2. If t includes new C0 (t).m(u) as a subterm, then mbody(m, last(C0 )) = (x, t0 ) and
|x| = |u| for some x and t0 .
Proof If t has new C0 (t).fi as a subterm, then, by well-typedness of the subterm, it is
easy to check that fields(last(C0 )) is well-defined and fi appears in it. The fact that
refinements may add fields (that have not been defined already) does not invalidate this
conclusion. Note that for every field of a class, including its superclasses and all its
refinements, there must be a proper argument. Similarly, if t has new C0 (t).m(u) as a
subterm, then it is also easy to show that mbody(m, last(C0 )) = (x, t0 ) and |x| = |u|
from the fact that mtype(m, last(C0 )) = C → D where |x| = |C|. This conclusion
holds for FFJ since a method refinement must have the same signature than the method
refined and overloading is not allowed.
t
u
45
T HEOREM A.3 (Type soundness of FFJ) If ∅ ` t : C and t −→∗ t0 with t0 a normal
form, then t0 is either a value v with ∅ ` v : D and D <: C, or a term containing
(D)(new C(t)) in which C <: D.
Proof Immediate from Theorem A.1 and A.2. Nothing changes in the proof of Theorem A.3 for FFJ compared to FJ.
t
u
B
Type Soundness Proof of FFJPL
In this section, we provide proof sketches of the theorems Correctness of FFJPL and
Completeness of FFJPL . A further formalization would be desirable, but we have stopped
at this point. As is often the case with formal systems, there is a trade-off between formal precision and legibility. We decided that a semi-formal development of the proof
strategies are the best fit for our purposes.
B.1
Correctness
T HEOREM B.1 (Correctness of FFJPL ) Given a well-typed FFJPL product line pl (including with a well-typed term t, well-formed class, introduction, and refinement tables
CT , IT , and RT , and a feature model FM ), every program that can be derived with a
valid feature selection fs is a well-typed FFJ program (cf. Figure 10).
pl = (t, CT , IT , RT , FM )
pl is well-typed
derive(pl , fs) is well-typed
fs is valid in FM
The proof strategy is as follows: assuming that the FFJPL type system ensures that
each slice is a valid FFJ type derivation (Lemma B.1) and that each valid feature selection corresponds to a single slice (Lemma B.1), it follows that the corresponding
feature-oriented program is well-formed. Before we prove Theorem B.1 we develop
two required lemmas that cover the two assumptions of our proof strategy.
L EMMA B.1 Given a well-formed FFJPL product line, every slice of the product line’s
type derivation corresponds to a (set of) valid type derivation(s) in FFJ.
Proof (Proof sketch) Given a well-formed FFJPL product line, the corresponding type
derivation consists of possibly multiple slices.
The basic case is easy: there is only a simple derivation without branches due to
mutually exclusive features (optional features may be present). In this case, each term
has only a single type, which is the one that would also be determined by FFJ. Furthermore, FFJPL guarantees that referenced types, methods, and fields are present in all
valid variants, using the predicate validref .
Let us illustrate this with the rule T-F IELDPL ; the other rules are analogous:
∀ E ∈ E : validref (Φ, E.f)
Γ ` t0 : E a Φ
fields(Φ, last(E)) = F, C f, G
Γ ` t0 .f : C11 , . . . , Cn1 , . . . , C1m , . . . , Cnm a Φ
46
(T-F IELDPL )
In the basic case there are no branches in the type derivation and thus the term
t0 has only a single type E1 . For the same reason, fields returns only a simple list of
fields that contains the declaration of field f. Finally, T-F IELDPL checks whether the
declaration of f is present in all valid variants (using validref ). Hence, in the basic
case, an FFJPL derivation that ends at the rule T-F IELDPL is equivalent to a set of
corresponding FFJ derivations, which do not contain alternative and optional features
and thus t0 has a single type, fields returns a simple list of fields that contains the
declaration of f, and the declaration of f is present. The reason that an FFJPL derivation
without mutually exclusive features (i.e., a single slice) corresponds to multiple FFJ
derivations is that the FFJPL derivation may contain optional features whose different
combinations correspond to the different FFJ derivations. Using predicate validref , all
type rules of FFJPL ensure that all possible combinations of optional features are welltyped.
In the case that there are multiple slices in the FFJPL derivation, a term t0 may have
multiple types E. The type rules of FFJPL make sure that every possible shape of a given
term is well-typed. Each possible type of the term leads to a branch in the derivation
tree. The premise of T-F IELDPL checks whether all possible shapes of a given term
are well-typed by taking the conjunction of all branches of the derivation. Hence, if
T-F IELDPL is successful, each individual branch holds, i.e., each slice corresponds to a
well-typed FFJ program. Ensuring that, in the presence of optional features, all relevant
subterms are well-typed (i.e., all referenced elements are present in all valid variants), a
well-typed slice covers a set of well-typed FFJ derivations that correspond to different
combinations of optional features, like in the basic case.
For example, in a field projection t0 .f, the subterm t0 has multiple types E. For all
these types, fields yields all possible combinations of fields declared by the variants
of the types. It is checked whether, for each type of the subterm t0 , each combination
of fields contains a proper declaration of field f. The different types of f become the
possible types of the overall field projection term. Like in the basic case, it is checked
whether every possible type of t0 is present in all valid variants (using validref ), so that
each slice corresponds a valid FFJ derivation, i.e., a whole set of derivations covering
different combinations of optional features.
t
u
L EMMA B.2 Given a well-formed FFJPL product line, each valid feature selection corresponds to a single slice in the corresponding type derivation.
Proof (Proof sketch) By definition, a valid feature selection does not contain mutually
exclusive features. Considering only a single valid feature selection, each term has only
a single type. But the type derivation of the overall product line contains branches corresponding to alternative types of the terms. A successive removal of mutually exclusive
features removes these branches until only a single branch remains. Consequently, a
valid feature selection corresponds to a single slice.
t
u
Proof (Proof sketch of Theorem B.1 (Correctness of FFJPL )) The fact that the FFJPL
type system ensures that each slice is a valid FFJ type derivation (Lemma B.1) and that
each valid feature selection corresponds to a single slice (Lemma B.2), implies that each
feature-oriented program that corresponds to a valid feature selection is well-formed.
t
u
47
B.2
Completeness
T HEOREM B.2 (Completeness of FFJPL ) Given an FFJPL product line pl (including
a well-typed term t, well-formed class, introduction, and refinement tables CT , IT ,
and RT , and a feature model FM ), and given that all valid feature selections fs yield
well-typed FFJ programs, according to Theorem B.1, pl is a well-typed product line
according to the rules of FFJPL .
pl = (t, CT , IT , RT , FM ) ∀ fs : (fs is valid in FM ⇒ derive(pl , fs) is well-typed)
pl is well-typed
Proof (Proof sketch of Theorem B.2 (Completeness of FFJPL )) There are three basic
cases: (1) pl has only mandatory features; (2) pl has only mandatory features except a
single optional feature; (3) pl has only mandatory features except two mutually exclusive features. Proving Theorem B.2 for the first basic case is trivial. Since only mandatory features exist, only a single FFJ program can be derived from the product line. If
the FFJ program is well-typed, the product line is well-typed, too, because all elements
are always reachable and each term has only a single type. In fact, the type rules of
FFJPL and FFJ become equivalent in this case.
In the second basic case, two FFJ programs can be derived from the product line,
one including and one excluding the optional feature. The difference between the two
programs is the content of the optional feature. The feature can add new classes, refine
existing classes by new methods and fields, and refine existing methods by overriding.
If the two programs are well-typed, then the overall product line is well-typed as well
since the reachability checks succeed in every type rule of FFJPL . Otherwise, at least
one of the two programs would not be well-typed since, in this case, the reachability
checks are the only difference between FFJPL ’s and FFJ’s type rules (as in the first case,
each term has only a single type since there are no mutually exclusive features). The
fact that the two FJ programs are well-typed implies that all elements are reachable in
the type derivations of two FFJ programs. Thus, the reachability checks of the FFJPL
derivation succeed in every case, i.e., the product line in question is well-typed.
In the third basic case, two FFJ programs can be derived from the product line,
one including the first alternative and the other including the second alternative of the
feature in question. The difference between the two programs is, on the one hand, the
program elements one feature introduces that are not present in the other and, on the
other hand, the alternative definitions of similar elements, like two alternative definitions of a single class. The first kind of difference is already covered by the second
basic case. Alternative definitions of a program element (second kind of difference)
that are well-typed in the context of their enclosing FFJ programs, are well-typed in
FFJPL because they lead to two new branches in the derivation tree which are handled
separately and the conjunction of their premises must hold. Since the corresponding
FFJ type rule for the element succeeds in both FFJ programs, their conjunction in the
FFJPL type rule always holds, i.e., the product line in question is well-typed.
Finally, we it remains to show that all other cases, i.e., all other combinations of
mandatory, optional, and alternative features, can be reduced to combinations of the
three basic cases, which proves Theorem B.2. To this end, we divide the possible relations between features into three disjoint sets: (1) a feature is reachable from another
48
feature in all variants, (2) a feature is reachable from another feature in some, but not
in all, variants, (3) two features are mutually exclusive. From these three possible relations we construct a general case that can be reduced to a combination of the three basic
cases.
Assume a feature Φ that is mandatory with respect to a set of features Π, that is
optional with respect to a set of features Ω, and that is alternative to a set ∆ of features.
We use arrows to illustrate to which of the three basic cases a pairwise relation between
Φ and each element of a list is reduced:
/Ω
ΦA
AA
AA3
1
AA
A
Π
∆
2
Such an arrow diagram can be created for every feature of a product line. The reason is
that the three kinds of relations are orthogonal and there are no further relations relevant
for type checking. Hence, the general case covers all possible relations between features
and combinations of features. The description of the general case and the reduction
finish the proof of Theorem B.2, i.e., FFJPL ’s type system is complete.
t
u
49
| 6cs.PL
|
arXiv:1708.05845v1 [math.AT] 19 Aug 2017
SPANNING SIMPLICIAL COMPLEXES OF UNI-CYCLIC
MULTIGRAPHS
IMRAN AHMED, SHAHID MUHMOOD
Abstract. A multigraph is a nonsimple graph which is permitted to have
multiple edges, that is, edges that have the same end nodes. We introduce
the concept of spanning simplicial complexes ∆s (G) of multigraphs G, which
provides a generalization of spanning simplicial complexes of associated
simple graphs. We give first the characterization of all spanning trees of a
r
uni-cyclic multigraph Un,m
with n edges including r multiple edges within
and outside the cycle of length m. Then, we determine the facet ideal
r
r
IF (∆s (Un,m
)) of spanning simplicial complex ∆s (Un,m
) and its primary
decomposition. The Euler characteristic is a well-known topological and
homotopic invariant to classify surfaces. Finally, we device a formula for
r
Euler characteristic of spanning simplicial complex ∆s (Un,m
).
Key words: multigraph, spanning simplicial complex, Euler characteristic.
2010 Mathematics Subject Classification: Primary 05E25, 55U10, 13P10,
Secondary 06A11, 13H10.
1. Introduction
Let G = G(V, E) be a multigraph on the vertex set V and edge-set E. A
spanning tree of a multigraph G is a subtree of G that contains every vertex
of G. We represent the collection of all edge-sets of the spanning trees of a
multigraph G by s(G). The facets of spanning simplicial complex ∆s (G) is
exactly the edge set s(G) of all possible spanning trees of a multigraph G.
Therefore, the spanning simplicial complex ∆s (G) of a multigraph G is defined
by
∆s (G) = hFk | Fk ∈ s(G)i,
which gives a generalization of the spanning simplicial complex ∆s (G) of an
associated simple graph G. The spanning simplicial complex of a simple connected finite graph was firstly introduced by Anwar, Raza and Kashif in [1].
Many authors discussed algebraic and combinatorial properties of spanning
simplicial complexes of various classes of simple connected finite graphs, see
for instance [1], [5], [6] and [9].
Let ∆ be simplicial complex of dimension d. We denote fi by the number
of i-cells of simplicial complex ∆. Then, the Euler characteristic of ∆ is given
1
2
Ahmed and Muhmood
by
χ(∆) =
d−1
X
(−1)i fi ,
i=0
which is a well-known topological and homotopic invariant to classify surfaces,
see [4] and [7].
r
The uni-cyclic multigraph Un,m
is a connected graph having n edges including r multiple edges within and outside the cycle of length m. Our aim is
to give some algebraic and topological characterizations of spanning simplicial
r
complex ∆s (Un,m
).
In Lemma 3.1, we give characterization of all spanning trees of a uni-cyclic
r
multigraph Un,m
having n edges including r multiple edges and a cycle of
r
length m. In Proposition 3.2, we determine the facet ideal IF (∆s (Un,m
)) of
r
spanning simplicial complex ∆s (Un,m ) and its primary decomposition. In Theorem 3.3, we give a formula for Euler characteristic of spanning simplicial
r
complex ∆s (Un,m
).
2. Basic Setup
A simplicial complex ∆ on [n] = {1, . . . , n} is a collection of subsets of [n]
satisfying the following properties.
(1) {j} ∈ ∆ for all j ∈ [n];
(2) If F ∈ ∆ then every subset of F will belong to ∆ (including empty set).
The elements of ∆ are called faces of ∆ and the dimension of any face F ∈ ∆
is defined as |F |−1 and is written as dim F , where |F | is the number of vertices
of F . The vertices and edges are 0 and 1 dimensional faces of ∆ (respectively),
whereas, dim ∅ = −1. The maximal faces of ∆ under inclusion are said to be
the facets of ∆. The dimension of ∆ is denoted by dim ∆ and is defined by
dim ∆ = max{dim F | F ∈ ∆}.
If {F1 , . . . , Fq } is the set of all the facets of ∆, then ∆ =< F1 , . . . , Fq >.
A simplicial complex ∆ is said to be pure, if all its facets are of the same
dimension.
A subset M of [n] is said to be a vertex cover for ∆ if M has non-empty
intersection with every Fk . M is said to be a minimal vertex cover for ∆ if no
proper subset of M is a vertex cover for ∆.
Definition 2.1. Let G = G(V, E) be a multigraph on the vertex set V and
edge-set E. A spanning tree of a multigraph G is a subtree of G that contains
every vertex of G.
Definition 2.2. Let G = G(V, E) be a multigraph on the vertex set V and edgeset E. Let s(G) be the edge-set of all possible spanning trees of G. We define
a simplicial complex ∆s (G) on E such that the facets of ∆s (G) are exactly the
elements of s(G), we call ∆s (G) as the spanning simplicial complex of G and
given by
∆s (G) = hFk | Fk ∈ s(G)i.
Spanning Simplicial Complexes of Uni-Cyclic Multigraphs
3
r
Definition 2.3. A uni-cyclic multigraph Un,m
is a connected graph having n
edges including r multiple edges within and outside the cycle of length m.
Let ∆ be simplicial complex of dimension d. Then, the chain complex C∗ (∆)
is given by
∂
∂d−1
∂
∂
∂
d
2
1
→0 0.
→ Cd−1 (∆)−
−−
→ ...−
→ C1 (∆)−
→ C0 (∆)−
0 → Cd (∆)−
Each Ci (∆) is a free abelian group of rank fi . The boundary homomorphism
∂d : Cd (∆) → Cd−1 (∆) is defined by
P
.
∂d (σαd ) = di=0 (−1)i σαd |[v0 ,v1 ,...,vi,...,v
ˆ
d]
Of course,
Hi (∆) = Zi (∆)/Bi (∆) = Ker∂i /Im∂i+1 ,
where Zi (∆) = Ker ∂i and Bi (∆) = Im ∂i+1 are the groups of simplicial icycles and simplicial i-boundaries, respectively. Therefore,
rank Hi (∆) = rank Zi (∆) − rank Bi (∆).
One can easily see that rank Bd (∆) = 0 due to Bd (∆) = 0. For each i > 0
there is an exact sequence
∂
i
→ Bi−1 (∆) → 0.
0 → Zi (∆) → Ci (∆)−
Moreover,
fi = rank Ci (∆) = rank Zi (∆) + rank Bi−1 (∆).
Therefore,
characteristic
of ∆ can be expressed as
Pd
Pdthe Euler
i
i
χ(∆) = i=0 (−1) fi = i=0 (−1) (rank Zi (∆) + rank Bi−1 (∆))
P
P
= di=0 (−1)i rank Zi (∆) + di=0 (−1)i rank Bi−1 (∆).
Changing index of summation in the last sum and using the fact that
rank B−1 (∆) = 0 = rank Bd (∆), we get
P
P
χ(∆) = di=0 (−1)i rank Zi (∆) + di=0 (−1)i+1 rank Bi (∆)
P
P
= di=0 (−1)i (rank Zi (∆) − rank Bi (∆)) = di=0 (−1)i rank Hi (∆).
Thus, the Euler characteristic of ∆ is given by
χ(∆) =
d
X
(−1)i βi (∆),
i=0
where βi (∆) = rank Hi (∆) is the i-th Betti number of ∆, see [4] and [7].
r
3. Topological Characterizations of ∆s (Un,m
)
r
Let Un,m
be a uni-cyclic multigraph having n edges including r multiple
edges within and outside the cycle of length m. We fix the labeling of the edge
r
set E of Un,m
as follows:
E = {e11 , . . . , e1t1 , . . . , er′ 1 , . . . , er′ tr′ , e(r′ +1)1 , . . . , em1 , e(m+1)1 , . . . , e(m+1)tm+1 ,
. . . , e(m+r′′ )1 , . . . , e(m+r′′ )tm+r′′ , e1 , . . . , ev }, where ei1 , . . . , eiti are the multiple
edges of i-th edge of cycle with 1 ≤ i ≤ r ′ while e(r′ +1)1 , . . . , em1 are the
single edges of the cycle and ej1 , . . . , ejtj are the multiple edges of the j-th
edge outside the cycle with m + 1 ≤ j ≤ m + r ′′ , moreover, e1 , . . . , ev are single
edges appeared outside the cycle.
4
Ahmed and Muhmood
r
We give first the characterization of s(Un,m
).
r
Lemma 3.1. Let Un,m
be the uni-cyclic multigraph having n edges including
r multiple edges and a cycle of length m with the edge set E, given above. A
r
) if and only if Twiw = {e1i1 , . . . , er′ ir′ ,
subset E(Twiw ) ⊂ E will belong to s(Un,m
e(r′ +1)1 , . . . , em1 , e(m+1)i(m+1) , . . . , e(m+r′′ )i(m+r′′ ) , e1 , . . . , ev }\{ewiw } for some ih ∈
{1, . . . , th } with 1 ≤ h ≤ r ′ , m + 1 ≤ h ≤ m + r ′′ and iw ∈ {1, . . . , tw } with
1 ≤ w ≤ r ′ or iw = 1 with r ′ + 1 ≤ w ≤ m for some w = h and iw = ih
appeared in Twiw .
r
Proof. By cutting down method [3], the spanning trees of Un,m
can be obtained
by removing exactly th − 1 edges from each multiple edge such that 1 ≤ h ≤ r ′ ,
m + 1 ≤ h ≤ m + r ′′ and in addition, an edge from the resulting cycle need
to be removed. Therefore, the spanning trees will be of the form Twiw =
{e1i1 , . . . , er′ ir′ , e(r′ +1)1 , . . . , em1 , e(m+1)i(m+1) , . . . , e(m+r′′ )i(m+r′′ ) , e1 , . . . , ev }\{ewiw }
for some ih ∈ {1, . . . , th } with 1 ≤ h ≤ r ′ , m + 1 ≤ h ≤ m + r ′′ and
iw ∈ {1, . . . , tw } with 1 ≤ w ≤ r ′ or iw = 1 with r ′ + 1 ≤ w ≤ m for
some w = h and iw = ih appeared in Twiw .
In the following result, we give the primary decomposition of facet ideal
r
IF (∆s (Un,m
).
r
Proposition 3.2. Let ∆s (Un,m
) be the spanning simplicial complex of unir
cyclic multigraph Un,m having n edges including r multiple edges within and
outside the cycle of length m. Then,
r
IF (∆s (Un,m
)=
\
\
\
\
\
(xk1 , xl1 )
(xi1 , . . . , xiti , xk1 )
(xa )
1≤a≤v
\
r ′ +1≤k<l≤m
1≤i≤r ′ ; r ′ +1≤k≤m
\
1≤i<b≤r ′
(xi1 , . . . , xiti , xb1 , . . . , xbtb )
\
\
m+1≤j≤m+r ′′
(xj1 , . . . , xjtj ) ,
where ti with 1 ≤ i ≤ r ′ is the number of multiple edges appeared in the i-th
edge of the cycle and tj with m + 1 ≤ j ≤ m + r ′′ is the number of multiple
edges appeared in the jth-edge outside the cycle.
r
Proof. Let IF (∆s (Un,m
) be the facet ideal of the spanning simplicial complex
r
∆s (Un,m ). From ([8], Proposition 1.8), minimal prime ideals of the facet ideal
IF (∆) have one-to-one correspondence with the minimal vertex covers of the
simplicial complex ∆. Therefore, in order to find the primary decomposition
r
)); it is sufficient to find all the minimal vertex
of the facet ideal IF (∆s (Un,m
r
covers of ∆s (Un,m ).
r
As ea , a ∈ [v] is not an edge of the cycle of uni-cyclic multigraph Un,m
r
and does not belong to any multiple edge of Un,m
. Therefore, it is clear by
definition of minimal vertex cover that {ea }, 1 ≤ a ≤ v is a minimal vertex
r
cover of ∆s (Un,m
). Moreover, a spanning tree is obtained by removing exactly
th − 1 edges from each multiple edge with h = 1, . . . , r ′ , m + 1, . . . , m + r ′′ and
Spanning Simplicial Complexes of Uni-Cyclic Multigraphs
5
r
in addition, an edge from the resulting cycle of Un,m
. We illustrate the result
into the following cases.
r
Case 1. If atleast one multiple edge is appeared in the cycle of Un,m
. Then, we
cannot remove one complete multiple edge and one single edge from the cycle
r
of Un,m
to get spanning tree. Therefore, (xi1 , . . . , xiti , xk1 ) with 1 ≤ i ≤ r ′ ,
′
r + 1 ≤ k ≤ m is a minimal vertex cover of the spanning simplicial complex
r
r
∆s (Un,m
) having non-empty intersection with all the spanning trees of Un,m
.
r
Moreover, two single edges cannot be removed from the cycle of Un,m to get
spanning tree. Consequently, (xk1 , xl1 ) for r ′ + 1 ≤ k < l ≤ m is a minimal
r
vertex cover of ∆s (Un,m
) having non-empty intersection with all the spanning
r
trees of Un,m .
r
Case 2. If atleast two multiple edges are appeared in the cycle of Un,m
. Then,
r
two complete multiple edges cannot be removed from the cycle of Un,m
to get
spanning tree. Consequently, (xi1 , . . . , xiti , xb1 , . . . , xbtb ) for 1 ≤ i < b ≤ r ′ is a
r
minimal vertex cover of ∆s (Un,m
) having non-empty intersection with all the
r
spanning trees of Un,m .
r
Case 3. If atleast one multiple edge appeared outside the cycle of Un,m
. Then,
r
one complete multiple edge outside the cycle of Un,m cannot be removed to get
spanning tree. So, (xj1 , . . . , xjtj ) for m + 1 ≤ j ≤ m + r ′′ is a minimal vertex
r
cover of ∆s (Un,m
) having non-empty intersection with all the spanning trees
r
of Un,m . This completes the proof.
r
We give now formula for Euler characteristic of ∆s (Un,m
).
r
Theorem 3.3. Let ∆s (Un,m
) be spanning simplicial complex of uni-cyclic
r
multigraph Un,m having n edges including r multiple edges and a cycle of
r′
m+r
P
P′′
r
length m. Then, dim(∆s (Un,m
)) = n −
ti −
tj + r − 2 and the Euler
i=1
j=m+1
r
characteristic of ∆s (Un,m
) is given by
r′
n−1
Q
P
′ −m
n
r
i
ti n−α+r
χ(∆s (Un,m )) =
(−1) i+1 −
i+1−m
i=1
i=0
β
ij
β
P
P
Q
P
β
tk
−
−
(−1)l−j
j
1
j=2
−
α+β
P
j=2
α+β
with α =
j
r′
P
i=1
m+1≤i1 <...<ij ≤m+r ′′ k=i1
ij ∈{r
/ ′P
+1,...,m}
ij
Q
−
1≤i1 <...<ij ≤m+r ′′ k=i1
m+r
P′′
l=j
α+β
P
tk
1
l=j
β−j
l−j
n−α+r ′ −m−l
i+1−m−l
(−1)l−j α+β−j
l−j
n−l
i+1−l
tj such that r ′ and r ′′ are the number of multiple
ti and β =
j=m+1
edges appeared within and outside the cycle, respectively.
Proof. Let E = {e11 , . . . , e1t1 , . . . , er′ 1 , . . . , er′ tr′ , e(r′ +1)1 , . . . , em1 , e(m+1)1 , . . . ,
e(m+1)tm+1 , . . . , e(m+r′′ )1 , . . . , e(m+r′′ )tm+r′′ , e1 , . . . , ev } be the edge set of uni-cyclic
r
multigraph Un,m
having n edges including r multiple edges and a cycle of length
′
m such that r and r ′′ are the number of multiple edges appeared within and
outside the cycle, respectively.
6
Ahmed and Muhmood
r
) is of the same dimension
One can easily see that each facet Twiw of ∆s (Un,m
′
′′
m+r
r′
m+r
r
P
P
P
P′′
n − ti −
tj + r − 2 = n − α − β + r − 2 with α =
ti and β =
tj ,
i=1
j=m+1
i=1
j=m+1
see Lemma 3.1.
By definition, fi is the number of subsets of E with i + 1 elements not
r′
Q
′ −m
containing cycle and multiple edges. There are
ti n−α+r
number of
i+1−m
i=1
subsets of E containing cycle but not containing any multiple edge within the
cycle. There are
X
β
n − α + r′ − m − l
β
n − α + r′ − m − β
l−β β − β
(−1)
=
i+1−m−l
l−β
β l=β
i+1−m−β
r
subsets of E containing cycle and β multiple edges of Un,m
outside the cycle
but not containing any multiple edge within the cycle. There are
n − α + r′ − m − β
n − α + r ′ − m − (β − 1)
β
−
i+1−m−β
i + 1 − m − (β − 1)
β−1
X
β
n − α + r′ − m − l
β
l−(β−1) β − (β − 1)
(−1)
=
i+1−m−l
l − (β − 1)
β−1
l=β−1
r
subsets of E containing cycle and β −1 multiple edges of Un,m
outside the cycle
but not containing any multiple edge within the cycle. Continuing in similar
manner, the number of subsets of E containing cycle and two edges from a
multiple edge outside the cycle but not containing any multiple edge within
the cycle
is given by
n−α+r′ −m−3
n−α+r′ −m−4
β−2 n−α+r ′ −m−2
− β−2
+ β−2
+···
0
i+1−m−2
1
i+1−m−3
2
i+1−m−4
β
n−α+r′ −m−l
P
β−2 n−α+r ′ −m−β
.
+ (−1)β−2 β−2
= (−1)l−2 β−2
l−2
i+1−m−l
i+1−m−β
l=2
i2
Q
P
tk
β
choices of two edges from a
There are
−
1
2
m+1≤i1 <i2 ≤m+r ′′ k=i1
multiple edge outside the cycle. Therefore, we obtain
β
i2 X
X
Y
n − α + r′ − m − l
β
tk
l−2 β − 2
−
(−1)
i+1−m−l
l−2
2
1
′′ k=i
l=2
m+1≤i1 <i2 ≤m+r
1
the number of subsets of E with i+1 elements containing cycle and all possible
choices of two edges from a multiple edge outside the cycle but not containing
multiple edges within the cycle.
Now, we use inclusion exclusion principal to obtain
(number of subsets of E containing cycle but not containing multiple edges)=
r′
Q
ti [(number of subsets of E with i + 1 elements containing cycle but not
i=1
containing multiple edges within the cycle)−(number of subsets of E with
i + 1 elements containing cycle and β multiple edges outside the cycle but
Spanning Simplicial Complexes of Uni-Cyclic Multigraphs
7
not containing multiple edges within the cycle)−(number of subsets of E with
i + 1 elements containing cycle and β − 1 multiple edges outside the cycle but
not containing multiple edges within the cycle)− · · · −(number of subsets of
E with i + 1 elements containing cycle and two edges from a multiple edge
outsidethe cycle but not containing multiple edges within the cycle)]
r′
β
Q
′ −m
β P
l−β β−β n−α+r ′ −m−l
=
ti n−α+r
−
(−1)
i+1−m
β
l−β
i+1−m−l
i=1
l=β
β
P
n−α+r′ −m−l
(−1)l−(β−1) β−(β−1)
l−(β−1)
i+1−m−l
l=β−1
β
i2
P
P
Q
β
tk
−···− 2 −
(−1)l−2
1
′′
l=2
m+1≤i1 <i2 ≤m+r k=i1
r′
Q
′
−m
=
ti n−α+r
i+1−m
i=1
β
ij
β
Q
P
P
P
tk
β
(−1)l−j
−
−
1
j
−
β
β−1
j=2
m+1≤i1 <...<ij ≤m+r ′′ k=i1
l=j
β−2
l−2
β−j
l−j
n−α+r ′ −m−l
i+1−m−l
n−α+r ′ −m−l
i+1−m−l
.
Therefore, we compute
fi = (number of subsets of E with i+1 elements)−(number of subsets of E with
i + 1 elements containing cycle but not containing multiple edges)−(number
of subsets of E with i+1 elements containing α+β multiple edges)−(number of
subsets of E with i+1 elements containing α+β−1 multiple edges)− · · · −(number
of subsets of E with i + 1 elements containing two edges from a multiple edge
r
of Un,m
)
r′
Q
′ −m
n
= i+1 −
ti n−α+r
i+1−m
β
i=1
ij
β
Q
P
P
P
tk
β
l−j β−j n−α+r ′ −m−l
(−1)
−
−
1
l−j
i+1−m−l
j
j=2
−
α+β
α+β
m+1≤i1 <...<ij ≤m+r ′′ k=i1
α+β
P
(−1)l−(α+β)
l=α+β
α+β
P
(α+β)−(α+β)
l−(α+β)
l=j
n−l
i+1−l
n−l
(−1)l−(α+β−1) (α+β)−(α+β−1)
l−(α+β−1)
i+1−l
l=α+β−1
ij ∈{r
′
/ P
+1,...,m} Q
α+β
i2
P
tk
− · · · − α+β
−
(−1)l−2
2
1
′′
l=2
1≤i1 <i2 ≤m+r k=i1
r′
Q
′
−m
n
= i+1
−
ti n−α+r
i+1−m
β
i=1
ij
β
P
Q
P
P
tk
β
(−1)l−j
−
−
1
j
−
α+β
α+β−1
j=2
−
α+β
P
j=2
α+β
j
m+1≤i1 <...<ij ≤m+r ′′ k=i1
ij ∈{r
/ ′P
+1,...,m}
ij
Q
−
1≤i1 <...<ij ≤m+r ′′ k=i1
l=j
α+β
P
tk
1
l=j
(−1)l−j
α+β−2
l−2
β−j
l−j
n−l
i+1−l
n−α+r ′ −m−l
i+1−m−l
α+β−j
l−j
n−l
i+1−l
.
8
Ahmed and Muhmood
e
13
e
12
e
e
e
21
42
11
e
e
31
41
2
Figure 1. U7,3
Example 3.4. Let E = {e11 , e12 , e13 , e21 , e31 , e41 , e42 } be the edge set of uni2
cyclic multigraph U7,3
having 7 edges including 2 multiple edges and a cycle of
length 3, as shown in Figure 1. By cutting-down method, we obtain
2
s(U7,3
) = {{e21 , e31 , e41 }, {e11 , e31 , e41 }, {e11 , e21 , e41 }, {e12 , e31 , e41 }, {e12 , e21 , e41 },
{e13 , e31 , e41 }, {e13 , e21 , e41 }, {e21 , e31 , e42 }, {e11 , e31 , e42 }, {e11 , e21 , e42 }, {e12 , e31 , e42 },
{e12 , e21 , e42 }, {e13 , e31 , e42 }, {e13 , e21 , e42 }}.
By definition, fi is the number of subsets of E with i + 1 elements not containing cycle and multiple edges. Since, {e11 }, {e12 }, {e13 }, {e21 }, {e31 }, {e41 }
and {e42 } are subsets of E containing one element. It implies that f0 = 7.
There are {e11 , e21 }, {e11 , e31 }, {e11 , e41 }, {e11 , e42 }, {e12 , e21 }, {e12 , e31 }, {e12 , e41 },
{e12 , e42 }, {e13 , e21 }, {e13 , e31 }, {e13 , e41 }, {e13 , e42 }, {e21 , e31 }, {e21 , e41 }, {e21 , e42 },
{e31 , e41 }, {e31 , e42 } subsets of E containing two elements but not containing
cycle and multiple edges. So, f1 = 17. We know that the spanning trees
2
2
of U7,3
are 2-dimensional facets of the spanning simplicial complex ∆s (U7,3
).
Therefore, f2 = 14. Thus,
2
χ(∆s (U7,3
)) = f0 − f1 + f2 = 7 − 17 + 14 = 4.
2
Now, we compute the Euler characteristic of ∆s (U7,3
) by using Theorem 3.3.
′
′′
We observe that, n = 7, r = 1, r = 1, r = 2, m = 3, α = 3, β = 2 and
2
0 ≤ i ≤ d, where d = n − α − β + r − 2 = 2 is the dimension of ∆s (U7,3
).
By substituting
3.3, we get
these values in Theorem
5
4
7
2
0
fi = i+1
− 3 i+1−3
− [ 52 − 31 21 ] 30 i+1−2
− 22 i+1−5
− 31 i+1−3
+
3
2
2 4
3
2
3
− 33 i+1−4
− 53
− 21 i+1−4
+ 22 i+1−5
−
2 i+1−4
0 i+1−3
2
1 3
2
5
5
1
−
.
−
5 i+1−5
4
0 i+1−4
1 i+1−5
2
Alternatively, we compute (f0 , f1 , f2 ) = (7, 17, 14) and χ(∆s (U7,3
)) = 4.
2
2
We compute now the Betti numbers of ∆s (U7,3 ). The facet ideal of ∆s (U7,3
)
is given by
Spanning Simplicial Complexes of Uni-Cyclic Multigraphs
9
2
IF (∆s (U7,3
)) = hxF21,31,41 , xF11,31,41 , xF11,21,41 , xF12,31,41 , xF12,21,41 , xF13,31,41 , xF13,21,41 ,
xF21,31,42 , xF11,31,42 , xF11,21,42 , xF12,31,42 , xF12,21,42 , xF13,31,42 , xF13,21,42 i.
2
We consider the chain complex of ∆s (U7,3
)
0 ∂3 C2 ∂2 C1 ∂1 C0 ∂0 0
→
−
−
→
−
→
−
→
Ker(∂i )
with i = 0, 1, 2.
The homology groups are given by Hi = Im(∂
i+1 )
2
Therefore, the i-th Betti number of ∆s (U7,3 ) is given by
βi = rank(Hi ) = rank(Ker(∂i )) − rank(Im(∂i+1 )) with i = 0, 1, 2.
Now, we compute rank and nullity of the matrix ∂i of order fi × fi+1 with
i = 0, 1, 2.
2
2
The boundary homomorphism ∂2 : C2 (∆s (U7,3
)) → C1 (∆s (U7,3
)) can be expressed as
∂2 (F21,31,41 ) = F31,41 − F21,41 + F21,31 ; ∂2 (F11,31,41 ) = F31,41 − F11,41 + F11,31 ;
∂2 (F11,21,41 ) = F21,41 − F11,41 + F11,21 ; ∂2 (F12,31,41 ) = F31,41 − F12,41 + F12,31 ;
∂2 (F12,21,41 ) = F21,41 − F12,41 + F12,21 ; ∂2 (F13,31,41 ) = F31,41 − F13,41 + F13,31 ;
∂2 (F13,21,41 ) = F21,41 − F13,41 + F13,21 ; ∂2 (F21,31,42 ) = F31,42 − F21,42 + F21,31 ;
∂2 (F11,31,42 ) = F31,42 − F11,42 + F11,31 ; ∂2 (F11,21,42 ) = F21,42 − F11,42 + F11,21 ;
∂2 (F12,31,42 ) = F31,42 − F12,42 + F12,31 ; ∂2 (F12,21,42 ) = F21,42 − F12,42 + F12,21 ;
∂2 (F13,31,42 ) = F31,42 − F13,42 + F13,31 ; ∂2 (F13,21,42 ) = F21,42 − F13,42 + F13,21 .
2
The boundary homomorphism ∂1 : C1 (∆s (U7,3
)) → C0 (∆s (UM27,3 )) can be
written as
∂1 (F11,21 ) = e21 − e11 ; ∂1 (F11,31 ) = e31 − e11 ; ∂1 (F11,41 ) = e41 − e11 ; ∂1 (F11,42 ) =
e42 − e11 ; ∂1 (F12,21 ) = e21 − e12 ; ∂1 (F12,31 ) = e31 − e12 ; ∂1 (F12,41 ) = e41 − e12 ;
∂1 (F12,42 ) = e42 − e12 ; ∂1 (F13,21 ) = e21 − e13 ; ∂1 (F13,31 ) = e31 − e13 ; ∂1 (F13,41 ) =
e41 − e13 ; ∂1 (F13,42 ) = e42 − e13 ; ∂1 (F21,31 ) = e31 − e21 ; ∂1 (F21,41 ) = e41 − e21 ;
∂1 (F21,42 ) = e42 − e21 ; ∂1 (F31,41 ) = e41 − e31 ; ∂1 (F31,42 ) = e42 − e31 .
Then, by using MATLAB, we compute rank of ∂2 = 11; nullity of ∂2 = 3; rank
of ∂1 = 6; nullity of ∂1 = 11.
Therefore, the Betti numbers are given by β0 = rank(Ker(∂0 ))−rank(Im(∂1 )) =
7 − 6 = 1; β1 = rank(Ker(∂1 )) − rank(Im(∂2 )) = 11 − 11 = 0; β2 =
rank(Ker(∂2 )) − rank(Im(∂3 )) = 3 − 0 = 3.
2
Alternatively, the Euler characteristic of ∆s (U7,3
) is given by
2
χ(∆s (U7,3
)) = β0 − β1 + β2 = 1 − 0 + 3 = 4.
References
[1] I. Anwar, Z. Raza and A. Kashif, Spanning Simplicial Complexes of Uni-Cyclic Graphs,
Algebra Colloquium, 22 (2015), no.4, 707-710.
[2] S. Faridi, The Facet Ideal of a Simplicial Complex, Manuscripta, 109 (2002), 159-174.
[3] F. Harary, Graph Theory, Reading, MA: Addison-Wesley, 1994.
[4] A. Hatcher, Algebraic Topology, Cambridge University Press, 2002.
[5] A. Kashif, I. Anwar and Z. Raza, On the Algebraic Study of Spanning Simplicial Complex
of r-cycles Graphs Gn,r , ARS Combinatoria, 115 (2014), 89-99.
[6] Y. Pan, R. Li and G.Zhu, Spanning Simplicial Complexes of n-Cyclic Graphs with a
Common Vertex, International Electronic Journal of Algebra, 17 (2015), 180-187.
10
Ahmed and Muhmood
[7] J. J. Rotman, An Introduction to Algebraic Topology, Springer-Verlag, New York, 1988.
[8] R. H. Villarreal, Monomial Algebras. Dekker, New York, 2001.
[9] G. Zhu, F. Shi and Y.Geng, Spanning Simplicial Complexes of n-Cyclic Graphs with a
Common Edge, International Electronic Journal of Algebra, 15 (2014), 132-144.
COMSATS Institute of Information Technology, Lahore, Pakistan
E-mail address: [email protected]
COMSATS Institute of Information Technology, Lahore, Pakistan
E-mail address: shahid [email protected]
| 0math.AC
|
arXiv:1212.2653v1 [math.AC] 11 Dec 2012
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
ABED ABEDELFATAH
Abstract. It has been conjectured by Eisenbud, Green and Harris that if I is
a homogeneous ideal in k[x1 , . . . , xn ] containing a regular sequence f1 , . . . , fn
of degrees deg(fi ) = ai , where 2 ≤ a1 ≤ ⋯ ≤ an , then there is a homogeneous
an
1
ideal J containing xa
1 , . . . , xn with the same Hilbert function. In this paper we prove the Eisenbud-Green-Harris conjecture when fi splits into linear
factors for all i.
1. Introduction
Let S = k[x1 , . . . , xn ] be a polynomial ring over a field k. The ring S = ⊕d≥0 Sd
is graded by deg(xi ) = 1 for all i. In 1927, F.Macaulay proved that if I = ⊕d≥0 Id is
a graded ideal in S, then there exists a lex ideal L such that L has the same Hilbert
function as I [13]; i.e., every Hilbert function in S is attained by a lex ideal. Let M
be a monomial ideal in S. It is natural to ask if we have the same result in S/M . In
[5], Clements and Lindström proved that every Hilbert function in S/⟨xa1 1 , . . . , xann ⟩
is attained by a lex ideal, where 2 ≤ a1 ≤ ⋅ ⋅ ⋅ ≤ an . In the case a1 = ⋯ = an = 2, the
result was obtained earlier by Katona [11] and Kruskal [?]. Another generalizations
of Macaulay’s theorem can be found in [17], [15] and [1].
Let f1 , . . . , fn be a regular sequence in S such that 2 ≤ a1 = deg(f1 ) ≤ ⋯ ≤ an =
deg(fn ). A well known result says that ⟨f1 , . . . , fn ⟩ has the same Hilbert function
as ⟨xa1 1 , . . . , xann ⟩ (see Exercise 6.2. of [9]). It is natural to ask what happens if
I ⊆ S is a homogeneous ideal containing a regular sequence in fixed degrees. This
question bring us to the Eisenbud-Green-Harris Conjecture, denoted by EGH.
Conjecture 1.1 (EGH [8]).
If I is a homogeneous ideal in S containing a regular sequence f1 , . . . , fn of degrees
deg(fi ) = ai , where 2 ≤ a1 ≤ ⋯ ≤ an , then I has the same Hilbert function as an
ideal containing xa1 1 , . . . , xann .
The original conjecture (see Conjecture 2.3) is equivalent to Conjecture 1.1 in
the case ai = 2 for all i (see Proposition 2.5). The EGH Conjecture is known to be
true in few cases. The conjecture has been proven in the case n = 2 [16]. Caviglia
and Maclagan [3] have proven that the EGH Conjecture is true if aj > ∑j−1
i=1 (ai − 1)
for all j > 1. Richert [16] says that the EGH Conjecture in degree 2 (ai = 2 for
all i) holds for n ≤ 5, but this result was not published. Herzog and Popescu [10]
proved that if k is a field of characteristic zero and I is minimally generated by
generic quadratic forms, then the EGH Conjecture in degree 2 holds. Cooper [6, 7]
has done some work in a geometric direction. She studies the EGH Conjecture for
some cases with n = 3.
Key words and phrases. Hilbert function, EGH conjecture, Regular sequence.
1
2
A.ABEDELFATAH
Let f1 , . . . , fn be a regular sequence in S such that fi splits into linear factors for
all i. For all 1 ≤ i ≤ n, let pi ∈ S1 such that pi ∣fi . Since p1 , . . . , pn must be a k-linear
independent, it follows that the k-algebra map α ∶ S → S defined by α(xi ) = pi for
all 1 ≤ i ≤ n, is a graded isomorphism. So the Hilbert function is preserved under
this map and we may assume that pi = xi for all i.
In Section 2, we give background information to the EGH Conjecture. In section
3, we study the dimension growth of some ideals containing a regular sequence
x1 l1 , . . . , xn ln , where li ∈ S1 for all i. In section 4, we prove the EGH Conjecture
when fi splits into linear factors for all i. This answers a question of Chen, who
asked if the EGH Conjecture holds when fi = xi li , where li ∈ S1 for all 1 ≤ i ≤ n (see
Example 3.8 of [4]).
2. Background
A proper ideal I in S is called graded or homogeneous if it has a system of
homogeneous generators. Let R = S/I, where I is a homogeneous ideal. The
Hilbert function of I is the sequence H(R) = {H(R, t)}t≥0 , where
H(R, t) ∶= dimk Rt = dimk St /It .
For simplicity, sometimes we denote the dimension of a k-vector space V by ∣V ∣
instead of dimk V . For a k-vector space V ⊆ Sd , where d ≥ 0, we denote by S1 V
the k-vector space spanned by {xi v ∶ 1 ≤ i ≤ n ∧ v ∈ V }. Throughout this paper
A = (a1 , . . . , an ) ∈ Zn , where 2 ≤ a1 ≤ ⋯ ≤ an . For a subset A of S, we denote by
Mon(A) the set of all monomials in A and let Au = {j ∶ xj ∣u}, where u ∈ Mon(S).
The support of the polynomial f = ∑u∈Mon(S) au u, where au ∈ k, is the set
supp(f ) = {u ∈ Mon(S) ∶ au ≠ 0}.
A monomial w ∈ S is called square-free if x2i ∤ w, for all 1 ≤ i ≤ n. We define
the lex order on Mon(S) by setting xb = xb11 ⋯xbnn <lex xc11 ⋯xcnn = xc if either
deg(xb ) < deg(xc ) or deg(xb ) = deg(xc ) and bi < ci for the first index i such that
bi ≠ ci . We recall the definitions of lex ideal and lex-plus-powers ideal.
Definition 2.1.
● A graded ideal is called monomial if it has a system of
monomial generators.
● A monomial ideal I ⊆ S is called lex, if whenever I ∋ z <lex w, where w, z
are monomials of the same degree, then w ∈ I.
● A monomial ideal I is A-lex-plus-powers if there exists a lex ideal L such
that I = ⟨x1a1 , . . . , xann ⟩ + L.
Example 2.2. the ideal I = ⟨x21 , x22 , x1 x2 x3 , x33 ⟩ is a (2, 2, 3)-lex-plus-powers ideal
in k[x1 , x2 , x3 ], because I = ⟨x21 , x22 , x33 ⟩ + ⟨x31 , x21 x2 , x21 x3 , x1 x22 , x1 x2 x3 ⟩ and
⟨x31 , x21 x2 , x21 x3 , x1 x22 , x1 x2 x3 ⟩ is a lex ideal in k[x1 , x2 , x3 ].
By Clements-Lindström’s theorem, we obtain that for any graded ideal containing ⟨xa1 1 , . . . , xann ⟩ there is a (a1 , . . . , an )-lex-plus-powers ideal with the same Hilbert
function.
q−1
) + ⋯ + (s11 ) be the unique Macaulay expansion of p with
Let p ≥ 0 and (sqq ) + (sq−1
sq
) + (sq−1
) + ⋯ + (s21 ). In [8], Eisenbud,
respect to q > 0. Set 0(q) = 0 and p(q) = (q+1
q
Green and Harris made the following conjecture.
Conjecture 2.3. If I ⊂ S is a graded ideal such that I2 contains a regular sequence
of maximal length and d > 0, then H(S/I, d + 1) ≤ H(S/I, d)(d) .
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
3
Conjecture 2.3 is true if the ideal contains the squares of the variables. This
follows from the Kruskal-Katona theorem (see [2]). In the following proposition,
we prove the equivalence of Conjecture 2.3 and the EGH Conjecture in degree 2.
First, we need the following definition.
Definition 2.4. Let M be a monomial ideal in S and d ≥ 0. A monomial vector
space Ld in (S/M )d is called lexsegment if it is generated by the t biggest monomials
(with respect to the lex order) in (S/M )d = Sd /Md , for some t ≥ 0.
For example, if L is a lex ideal in S, then Lj is lexsegment for all j ≥ 0. If Ld
is a lexsegment space in (S/M )d , where M is a monomial ideal in S, then S1 Ld is
lexsegment in (S/M )d+1 (see Proposition 2.5 of [15]).
Proposition 2.5. Let f1 , . . . , fn be a regular sequence of degrees 2 in S. The
following are equivalent:
(a) If I is a graded ideal in S containing f1 , . . . , fn , then there is a graded ideal
J in S containing x21 , . . . , x2n such that H(S/I) = H(S/J).
(b) If I is a graded ideal in S containing f1 , . . . , fn , then
H(S/I, d + 1) ≤ H(S/I, d)(d) for all d > 0.
Proof. First, we prove that (a) implies (b). Let I be a graded ideal in S containing
f1 , . . . .fn . By (a), it follows that there is a graded ideal J in S containing x21 , . . . , x2n
such that H(S/I) = H(S/J). By Kruskal-Katona theorem it follows that H(S/I, d+
1) = H(S/J, d + 1) ≤ H(S/J, d)(d) = H(S/I, d)(d) for all d > 0.
Now, we prove that (b) implies (a). Let I be a graded ideal in S containing
f1 , . . . , fn . Set M = ⟨x21 , . . . , x2n ⟩ and P = ⟨f1 , . . . , fn ⟩. For every d ≥ 0, let Ld be the
k-vector space spanned by the first square-free monomials (in lex order) of Sd such
that ∣Ld ⊕ Md ∣ = ∣Id ∣. Let K = ⊕j≥0 Kj = ⊕j≥0 (Lj + Mj ). We need to show that K is
an ideal. Let d > 0. By Proposition 6.4.3 of [9], we obtain that
∣Sd+1 /Md+1 ∣ − ∣S1 Ld /S1 Ld ∩ Md+1 ∣ = (∣Sd /Md ∣ − ∣Ld ∣)(d) .
By the hypothesis of (b), we obtain (∣Sd /Md ∣ − ∣Ld ∣)(d) = ∣Sd /Id ∣(d) ≥ ∣Sd+1 /Id+1 ∣. So
∣Sd+1 /Md+1 ∣ − ∣S1 Ld /S1 Ld ∩ Md+1 ∣ = ∣Sd+1 /Md+1 ∣ − ∣S1 Ld + Md+1 /Md+1 ∣ ≥ ∣Sd+1 /Id+1 ∣.
This implies that ∣S1 Ld +Md+1 ∣ ≤ ∣Ld+1 +Md+1 ∣. Since Ld+1 and S1 Ld are lexsegments
in (S/M )d+1 , it follows that S1 Ld ⊆ Ld+1 + Md+1 . So S1 Kd ⊆ Kd+1 for all d ≥ 0,
which implies that K is a graded ideal in S. Clearly, H(S/K) = H(S/I).
The following lemma helps us to study the EGH Conjecture in each component
of the homogeneous ideal.
Lemma 2.6. Let I be a graded ideal in S containing a regular sequence f1 , . . . , fn
of degrees deg(fi ) = ai . The following are equivalent:
(a) There exists a graded ideal J in S containing xa1 1 , . . . , xann such that H(S/I) =
H(S/J).
(b) For every d ≥ 0, there exists a graded ideal J in S containing xa1 1 , . . . , xann
such that H(S/I, d) = H(S/J, d) and H(S/I, d + 1) ≤ H(S/J, d + 1).
Proof. Clearly, (a) implies (b). We will show that (b) implies (a). For every
d ≥ 0, there exists an ideal Jd in S containing xa1 1 , . . . , xann such that H(S/I, d) =
H(S/Jd, d) and H(S/I, d + 1) ≤ H(S/Jd, d + 1). By Clements-Lindström’s theorem,
we may assume that Jd is a A-lex-plus-powers ideal for all d. Let J = ⊕j≥0 Jj,j , where
4
A.ABEDELFATAH
Jj,j is the j-th component of Jj . Since dim(Jd,d+1 ) ≤ dim(Id+1 ) = dim(Jd+1,d+1 ),
it follows that Jd,d+1 ⊆ Jd+1,d+1 , for all d. So S1 Jd,d ⊆ Jd,d+1 ⊆ Jd+1,d+1 , for all d.
Thus, J is an ideal. Clearly, H(S/I) = H(S/J).
We will use the following lemma on regular sequences (see [14, Chapter 6]).
Lemma 2.7. Let f1 , . . . , fn be a sequence of homogeneous polynomials in S with
deg(fi ) = ai and P = ⟨f1 , . . . , fn ⟩. Then
(a) If f1 , . . . , fn is a regular sequence, then H(S/P ) = H(S/⟨xa1 1 , . . . , xann ⟩).
(b) f1 , . . . , fn is a regular sequence if and only if the following condition holds:
if g1 f1 + ⋯ + gn fn = 0 for some g1 , . . . , gn ∈ S, then g1 , . . . , gn ∈ P .
(c) If f1 , . . . , fn is a regular sequence and σ ∈ Sn is a permutation, then
fσ(1) , . . . , fσ(n) is a regular sequence.
3. The dimension growth of some ideals containing a reducible
regular sequence
Let f1 = x1 l1 , . . . , fn = xn ln be a regular sequence in S, where li ∈ S1 for all i.
Set P = ⟨f1 , . . . , fn ⟩ and M = ⟨x21 , . . . , x2n ⟩. Let Vd be a vector space spanned by Pd
and square-free monomials w1 , . . . , wt in Sd , and Wd be the vector space spanned
by Md and w1 , . . . , wt . In this section, we prove that dim(S1 Vd ) = dim(S1 Wd ). We
also compute dim(S1 Kd ), where Kd is the space generated by Pd and the biggest
(in lex order) square-free monomials v1 , . . . , vt in Sd .
For a matrix A ∈ Mn×n (k), we denote by A[i1 , . . . , ir ] the submatrix of A formed
by rows i1 , . . . , ir and columns i1 , . . . , ir , where 1 ≤ r ≤ n and 1 ≤ i1 < ⋯ < ir ≤ n.
We begin with the following lemma, which characterize the structure of f1 , . . . , fn .
Lemma 3.1. (Example 3.8 of [4])
Let f1 = x1 l1 , . . . , fn = xn ln be a sequence of homogeneous polynomials in S, where
li = ∑nj=1 aij xj with aij ∈ k and A be the n × n matrix (aij ). Then f1 , . . . , fn
is a regular sequence if and only if det A[i1 , . . . , ir ] ≠ 0 for all 1 ≤ r ≤ n and
1 ≤ i1 < ⋯ < ir ≤ n.
Proof. Assume that f1 , . . . , fn is regular. We prove that det A[i1 , . . . , ir ] ≠ 0 for all
1 ≤ r ≤ n and 1 ≤ i1 < ⋯ < ir ≤ n, by induction on n, starting with n = 1. Let
n > 1. Assume that 1 ≤ i1 < ⋯ < ir ≤ n, where 1 ≤ r ≤ n − 1. Let j ∉ {i1 , . . . , ir }.
Note that xj lj is regular modulo an ideal I if and only if both xj and lj are regular
modulo I. By Lemma 2.7, xj , f1 , . . . , fj−1 , fj+1 , . . . , fn is a regular sequence. So
f1 , . . . , fj−1 , fj+1 , . . . , fn is a regular sequence in S/⟨xj ⟩. By the inductive step
we obtain that det A[i1 , . . . , ir ] ≠ 0. It remains to show that det(A) ≠ 0. From
the permutability property of regular sequences of homogeneous polynomials, we
obtain that l1 , . . . , ln is a regular sequence. So l1 , . . . , ln is k-linearly independent.
Assume now det A[i1 , . . . , ir ] ≠ 0 for all 1 ≤ r ≤ n and 1 ≤ i1 < ⋯ < ir ≤ n.
We prove that f1 , . . . , fn is a regular sequence by induction on n, starting with
n = 1. Let n > 1. By the inductive step, the sequence f1 , . . . , fn−1 is regular in
S/⟨xn ⟩. So f1 , . . . , fn−1 , xn is a regulae sequence in S. It remains to show that
f1 , . . . , fn−1 , ln is a regular sequence. Since det(A) ≠ 0, it follows that the k-algebra
map α ∶ S → S defined by α(xi ) = li , for all i, is an isomorphism. By the inductive
step, α−1 (f1 ), . . . , α−1 (fn−1 ), α−1 (ln ) = xn is a regular sequence. So f1 , . . . , fn−1 , ln
is a regular sequence, as desired.
The special structure of the regular sequence in 3.1 implies the following lemma.
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
5
Lemma 3.2. Let f1 = x1 l1 , . . . , fn = xn ln be a regular sequence of homogeneous
polynomials in S, where li = ∑nj=1 aij xj with aij ∈ k, and P = ⟨f1 , . . . , fn ⟩. If g ∉ P
is a homogeneous polynomial in S, then
g ≡ h (mod P )
where deg(h) = deg(g) and h is a k-linear combination of square-free monomials.
Proof. Since g ∉ P , we have deg(g) ≤ n. It is sufficient to prove the lemma when
g ∉ P is a monomial in ⟨x21 , . . . , x2n ⟩ of degree ≤ n. We prove by induction on deg(g).
The lemma is true when deg(g) = 2, since aii ≠ 0 for all i. Let g be a monomial in
⟨x21 , . . . , x2n ⟩ of degree d > 2 and A be the n × n matrix (aij ). By the inductive step,
we may assume that xgi is a square-free monomial for some i. By Lemma 3.1, we
have det A[j ∶ j ∈ Ag ] ≠ 0. So there exist scalars (cj )j∈Ag , such that ∑j∈Ag cj lj ≡ xi
(mod ⟨xj ∶ j ∉ Ag ⟩). It follows that xi = ∑j∈Ag cj lj + ∑j∉Ag cj xj , where cj ∈ k
for all j ∉ Ag . Then g = ∑j∈Ag cj lj xgi + ∑j∉Ag cj xj xgi . Let h = ∑j∉Ag cj xj xgi . Note
that h ≠ 0 is a k-linear combination of square-free monomials of degree d. Since
g
∑j∈Ag cj lj xi ∈ P , we obtain that g ≡ h (mod P ).
By the proof of Lemma 3.2, we obtain the following.
Remark 3.3. Let P be as in Lemma 3.2 and 0 ≤ d ≤ n. If w is a square-free
monomial in Sd and q ∈ S1 , then
qw = q̃w + q̂w
where q̃, q̂ ∈ S1 , q̂w ∈ P and q̃w is a k-linear combination of square-free monomials.
Example 3.4. Assume that S = C[x1 , x2 , x3 ] and
f1 = x21 + x1 x2 + x1 x3 = x1 (x1 + x2 + x3 )
f2 = −x1 x2 + x22 + x2 x3 = x2 (−x1 + x2 + x3 )
f3 = −x1 x3 − x2 x3 + x23 = x3 (−x1 − x2 + x3 ).
1 1 ⎞
⎛ 1
In this case, A = ⎜ −1 1 1 ⎟ is the matrix that defined in Lemma 3.1.
⎝ −1 −1 1 ⎠
Since det A[i1 , . . . , ir ] ≠ 0 for all 1 ≤ r ≤ 3 and 1 ≤ i1 < ⋯ < ir ≤ 3, we have that
f1 , f2 , f3 is a regular sequence in S. Set P = ⟨f1 , f2 , f3 ⟩ and let g = x31 + x21 x2 . Since
x21 ≡ −x1 x2 − x1 x3 (mod P ), we have x31 ≡ −x21 x2 − x21 x3 (mod P ). So g ≡ −x21 x3
(mod P ). Also, we see that x3 f1 − x1 f3 = 2x21 x3 + 2x1 x2 x3 ∈ P . So g ≡ x1 x2 x3
(mod P ) and x1 x2 x3 ∉ ⟨x21 , x22 , x23 ⟩.
Remark 3.5. Lemma 3.2 is not true if f1 , . . . , fn is an arbitrary regular sequence.
For example, consider the sequence
f1 = x21 + x1 x2 + x22 , f2 = x1 x2 in C[x1 , x2 ].
Note that f1 , f2 is a regular sequence ⇔ f1 , x1 and f1 , x2 are regular sequences ⇔
x1 , f1 and x2 , f1 are regular sequences ⇔ x22 and x21 are regular elements in C[x2 ]
and C[x1 ], respectively. So f1 , f2 is a regular sequence. Let g = x22 . It is easy to
show that g ∉ ⟨f1 , f2 ⟩. If g ≡ ax1 x2 (mod ⟨f1 , f2 ⟩), for some a ∈ C, then there exist
c1 , c2 , c3 ∈ C, not all zero, such that c1 f1 + c2 f2 + c3 (g − ax1 x2 ) = 0. But the equation
c1 x21 + (c1 + c2 − ac3 )x1 x2 + (c1 + c3 )x22 = 0,
implies that c1 = c2 = c3 = 0, a contradiction.
6
A.ABEDELFATAH
As a result of Lemma 3.2, we obtain the following.
Lemma 3.6. If P as in Lemma 3.2, then the set of all square-free monomials form
a k-basis of S/P .
Proof. Denote by A the set of all square-free monomials in S. Lemma 3.2 shows
that S/P generated by A. Let w = x1 ⋯xn . Assume that w ∈ P . Since H(S/P ) =
H(S/⟨x21 , . . . , x2n ⟩), it follows that there is a polynomial f ∈ Sn such that f ∉ P . By
Lemma 3.2, f ≡ bx1 ⋯xn (mod P ), where 0 ≠ b ∈ k. Since w ∈ P , it follows that
f ∈ P , a contradiction. So w ∉ P . Suppose that ∑w∈A aw w ∈ P , where aw ∈ k and
aw = 0 for almost all w ∈ A. Assume that aw ≠ 0 for some w. Let v ∈ A be a
monomial with minimal degree such that av ≠ 0. So v ∈ ⟨fi ∶ i ∈ Av ⟩ in the ring
S/⟨xi ∶ i ∉ Av ⟩, a contradiction.
Lemma 3.7. Let P be as in Lemma 3.2. If w is a square-free monomial in Sd ,
where 0 ≤ d ≤ n, then
(a) ∣S1 (w) ∩ Pd+1 ∣ = d.
(b) ∣S1 (w) ∩ (Pd+1 + S1 (w1 , . . . , wt ))∣ = ∣S1 (w) ∩ Pd+1 ∣ + ∣S1 (w) ∩ S1 (w1 , . . . , wt )∣
for every square-free monomials w1 , . . . , wt of degrees d such that wi ≠ w
for all 1 ≤ i ≤ t.
Proof. (a). Let q = ∑ni=1 ci li ∈ S1 , where ci ∈ k for all i, such that qw ∈ Pd+1 .
Assume that cj ≠ 0 for some j ∉ Aw . Since qw ∏j≠k∉Aw xk ∈ P , it follows that
cj lj w ∏j≠k∉Aw xk ∈ P . Thus, cj lj w ∏j≠k∉Aw xk = h1 f1 + ⋅ ⋅ ⋅ + hn fn , where hi ∈ S for
all 1 ≤ i ≤ n. So
h1 f1 + ⋅ ⋅ ⋅ + hj−1 fj−1 + (xj hj − cj w ∏ xk )lj + hj+1 fj+1 + ⋅ ⋅ ⋅ + hn fn = 0,
j≠k∉Aw
which implies that
xj hj − cj w ∏ xk ∈ ⟨f1 , . . . , fj−1 , fj+1 , . . . , fn ⟩.
j≠k∉Aw
So w ∏j≠k∉Aw xk ∈ ⟨f1 , . . . , fj−1 , fj+1 , . . . , fn ⟩ in the ring S/⟨xj ⟩, a contradiction to
Lemma 3.6. It follows that q belong to the k-vector space (li ∶ i ∈ Aw ). On the
other hand, li w ∈ P , for all i ∈ Aw . So ∣S1 (w) ∩ Pd+1 ∣ = dim(li w ∶ i ∈ Aw ) = d.
(b). First, we show that
S1 (w) ∩ (Pd+1 + S1 (w1 , . . . , wt )) = S1 (w) ∩ Pd+1 + S1 (w) ∩ S1 (w1 , . . . , wt ).
Assume that qw ∈ Pd+1 +S1 (w1 , . . . , wt ), where q ∈ S1 . There exist f ∈ S1 (w1 , . . . , wt )
and g ∈ Pd+1 such that qw = g + f . If f ∈ P , then qw ∈ S1 (w) ∩ Pd+1 . So assume
that f ∉ P . By 3.3, we may assume that f is a k-linear combination of square-free
monomials. Also, we obtain that qw = q̃w + q̂w, where q̃, q̂ ∈ S1 , q̂w ∈ P and q̃w is a
k-linear combination of square-free monomials. So q̃w − f ∈ P , which implies that
q̃w = f ∈ S1 (w1 , . . . , wt ). Hence qw ∈ S1 (w) ∩ Pd+1 + S1 (w) ∩ S1 (w1 , . . . , wt ) and we
obtain that the desired equality.
It remains to show that
S1 (w) ∩ S1 (w1 , . . . , wt ) ∩ Pd+1 = (0).
Let qw ∈ S1 (w1 , . . . , wt ) ∩ Pd+1 , where q ∈ S1 . By (a), we have q = ∑j∈Aw cj lj , where
cj ∈ k for all j ∈ Aw . For every 1 ≤ j ≤ t, let ij ∈ Awj ∖Aw and let B = {ij ∶ 1 ≤ j ≤ t}.
By the hypothesis, we obtain that qw = ∑ti=1 qi wi , where qi ∈ S1 for all 1 ≤ i ≤ t. So
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
7
qw = 0 in the ring S/⟨xj ∶ j ∈ B⟩, which implies that ∑j∈Aw cj lj = 0. By 3.1, we
obtain that cj = 0, for all j ∈ Aw . Thus, qw = 0.
Remark 3.8. Part (b) of Lemma 3.7 is not true if we replace w, w1 , . . . , wt by homogeneous polynomials which are a k-linear combination of square-free monomials
in Sd . For example, let S = k[x1 , x2 , x3 , x4 ] and P = ⟨x21 , x22 , x23 , x24 ⟩. Suppose that
h = x1 x2 + x2 x4 + x3 x4 and h1 = x1 x2 + x1 x3 . Computation with Macaulay2 shows
that
∣S1 (h) ∩ (P3 + S1 (h1 ))∣ = 2 and ∣S1 (h) ∩ P3 ∣ = ∣S1 (h) ∩ S1 (h1 ))∣ = 0.
In the case that w is a homogeneous polynomial in part (a) of Lemma 3.7,
the dimension is always bounded by the degree. This is a result of the following
proposition.
Proposition 3.9. Let P be as in Lemma 3.2. If g ∉ P is a homogeneous polynomial
of degree d, then ∣S1 (g) ∩ Pd+1 ∣ ≤ d.
Proof. We prove by induction on n. If n = 1, then g = ax1 or g ∈ k, where a ∈ k.
If g ∈ k, then ∣S1 (g) ∩ P1 ∣ = 0 and if g = ax1 , then ∣S1 (g) ∩ P2 ∣ = 1. Let n > 1. We
prove by induction on d, starting with d = 0. Let d > 0. If d = n, then Pd+1 = Sd+1
and so ∣S1 (g) ∩ Pd+1 ∣ = n. Assume that d < n. By 3.2, there exists a k-linear
combination of square-free monomials h ∈ Sd such that g ≡ h (mod Pd ). Clearly,
S1 (h)∩Pd+1 = S1 (g)∩Pd+1 . Let h = ∑ti=1 ai wi , where 0 ≠ ai ∈ k and wi ∈ Mon(Sd ) for
all i. Let j ∉ Aw1 . If lj h ∈ Pd+1 , then lj w1 ∈ Pd+1 in the ring S/⟨xi ∶ i ∉ Aw ∧ i ≠ j⟩,
a contradiction. So lj h ∉ Pd+1 for all j ∉ Aw1 . In particular, there exists a variable
xi such that xi h ∉ Pd+1 . We have two cases:
Case 1. h ∉ Pd in the ring S/⟨xi ⟩. Let p1 h, . . . , ps h be a basis of S1 (h) ∩ Pd+1
in the ring S/⟨xi ⟩. By the inductive step, we obtain that s ≤ d. If f ∈ S1 (h) ∩ Pd+1 ,
then f ∈ (p1 h, . . . , ps h, xi q), where q ∈ Sd . Since f ∈ S1 (h), it follows that xi q = rh,
where r ∈ S1 . Since xi ∤ h, it follows that xi ∣r. So f ∈ (p1 h, . . . , ps h, xi h). Therefore,
S1 (h) ∩ Pd+1 ⊆ (p1 h, . . . , ps h, xi h). If ∣S1 (h) ∩ Pd+1 ∣ = s + 1, then xi h ∈ Pd+1 , a
contradiction.
Case 2. h ∈ Pd in the ring S/⟨xi ⟩. So h ≡ xi q (mod Pd ), where q ∈ Sd−1 . Since
h is the unique k-linear combination of square-free monomials such that xi q ≡ h
(mod P ), we obtain that h = xi h1 , where h1 ∈ Sd−1 . If f ∈ S1 (h) ∩ Pd+1 , then
f = pxi h1 , for some p ∈ S1 . Clearly, xfi ∈ S1 (h1 ). Since f ∈ P , it follows that
ph1 ∈ Pd in the ring S/⟨li ⟩. So xfi ∈ S1 (h1 ) ∩ Pd in S/⟨li ⟩. If h1 ∈ Pd−1 in S/⟨li ⟩, then
xi h1 = h ∈ P , a contradiction. Let p1 h1 , . . . , ps h1 be a basis of S1 (h1 ) ∩ Pd . By the
inductive step, we obtain that s ≤ d − 1. So xfi ∈ (p1 h1 , . . . , ps h1 , li q), which implies
that f ∈ (p1 h, . . . , ps h, li xi q). Therefore, ∣S1 (h) ∩ Pd+1 ∣ ≤ s + 1 ≤ d.
Now, we prove the main results of this section.
Theorem 3.10. Let P be as in Lemma 3.2 and M = ⟨x21 , . . . , x2n ⟩. Assume that
V = Pd +(w1 , . . . , wt ) and W = Md +(w1 , . . . , wt ), where wi is a square-free monomial
of degree d, for all i. Then
dim S1 W = dim S1 V.
8
A.ABEDELFATAH
Proof. We may assume that d ≥ 2 and prove by induction on t. If t = 1, then
dim S1 W = dim Md+1 + dim S1 (w1 ) − dim S1 (w1 ) ∩ Md+1
= dim Pd+1 + dim S1 (w1 ) − dim S1 (w1 ) ∩ Pd+1 = dim S1 V.
Let t > 1, and set W1 = Md + (w1 , . . . , wt−1 ), V1 = Pd + (w1 , . . . , wt−1 ) and Z =
S1 (wt ) ∩ S1 (w1 , . . . , wt−1 ). By Lemma (3.7) and the inductive step, we have
dim S1 W = dim S1 W1 + dim S1 (wt ) − dim S1 (wt ) ∩ S1 W1
= dim S1 V1 + dim S1 (wt ) − dim S1 (wt ) ∩ S1 W1
= dim S1 V1 + dim S1 (wt ) − dim S1 (wt ) ∩ Md+1 − dim Z
= dim S1 V1 + dim S1 (wt ) − dim S1 (wt ) ∩ Pd+1 − dim Z
= dim S1 V1 + dim S1 (wt ) − dim S1 (wt ) ∩ S1 V1
= dim S1 V.
Proposition 3.11. Let P be as in Lemma 3.2 and V = Pd + (w1 , . . . , wt ) be the
k-vector space spanned by Pd and the t biggest (in lex order) square-free monomials
in Sd . Then
t
d+n
n
dim S1 V = (
)−(
) + ∑(n − m(wi )),
d+1
d+1
i=1
where m(wi ) = max{j ∶ xj ∣wi }, 1 ≤ i ≤ t.
Proof. We claim that
t
t
t
i=1
i=1
i=2
∣S1 V ∣ = ∣Pd+1 ∣ + ∑ ∣S1 (wi )∣ − ∑ ∣S1 (wi ) ∩ Pd+1 ∣ − ∑ ∣S1 (wi ) ∩ S1 (w1 , . . . , wi−1 )∣.
We prove the claim by induction on t. If t = 1, then
∣S1 V ∣ = ∣Pd+1 ∣ + ∣S1 (w1 )∣ − ∣S1 (w1 ) ∩ Pd+1 ∣.
Let t > 1 and V1 = Pd + (w1 , . . . , wt−1 ). By the inductive step we obtain that ∣S1 V ∣
is equal to
t
t−1
t−1
i=1
i=1
i=2
∣Pd+1 ∣+ ∑ ∣S1 (wi )∣− ∑ ∣S1 (wi )∩Pd+1 ∣− ∑ ∣S1 (wi )∩S1 (w1 , . . . , wi−1 )∣−∣S1 (wt )∩S1 V ∣.
By Lemma 3.7, we have ∣S1 (wt )∩S1 V1 ∣ = ∣S1 (wt )∩Pd+1 ∣+∣S1 (wt )∩S1 (w1 , . . . , wt−1 )∣.
We proved the claim.
Let 2 ≤ j ≤ t. If i < m(wj ) such that xi ∤ wj , then xi wj ∈ S1 (w1 , . . . , wj−1 ). So
∣S1 (wi ) ∩ S1 (w1 , . . . , wi−1 )∣ = m(wj ) − d. Therefore
∣S1 V ∣ = ∣Sd+1 ∣ − (
t
n
) + tn − td − ∑(m(wi ) − d)
d+1
i=2
t
d+n
n
=(
)−(
) + tn − td − ∑(m(wi ) − d)
d+1
d+1
i=2
t
d+n
n
=(
)−(
) + tn − td − ∑(m(wi ) − d)
d+1
d+1
i=1
t
d+n
n
=(
)−(
) + ∑(n − m(wi )).
d+1
d+1
i=1
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
9
4. The main result
In this section we prove that the EGH Conjecture is true if fi splits into linear
factors for all i. We begin with the following lemma.
Lemma 4.1. Let P = ⟨f1 , . . . , fn ⟩ be an ideal of S generated by a regular sequence
with deg(fi ) = ai and n ≥ 2. Assume that fn = q1 ⋯qs , where q1 , . . . , qs ∈ S1 . Then
(a) H(S/P + ⟨qm ⟩) = H(S/P + ⟨qk ⟩) for all 1 ≤ m, k ≤ s.
(b) H(S/(P ∶ q1 ⋯qj ) + ⟨qm ⟩) = H(S/(P ∶ q1 ⋯qj ) + ⟨qk ⟩) for all 1 ≤ j ≤ s − 1 and
j < m, k ≤ s.
Proof. First, we will prove (a). Let 1 ≤ m, k ≤ s. Note that P + ⟨qm ⟩/⟨qm ⟩ and
P +⟨qk ⟩/⟨qk ⟩ are ideals in S/⟨pm ⟩ and S/⟨qk ⟩, respectively, generated by f1 , . . . , fn−1 .
Note also that f1 , . . . , fn−1 , qm and f1 , . . . , fn−1 , qk are regular sequences. By part
(c) of Lemma 2.7, we obtain that f1 , . . . , fn−1 is a regular sequence in S/⟨qm ⟩ and
S/⟨qk ⟩. By part (a) of Lemma 2.7, we obtain that H(S/P + ⟨qm ⟩) = H(S/P + ⟨qk ⟩).
Now, we prove (b). Let 1 ≤ j ≤ s − 1 and j < m, k ≤ s. Assume that
h = h1 + h2 ∈ (P ∶ q1 ⋯qj ) + ⟨qm ⟩,
where h1 ∈ (P ∶ q1 ⋯qj ) and h2 ∈ ⟨qm ⟩. Since h1 q1 ⋯qj ∈ P , it follows that h1 q1 ⋯qj =
g1 f1 + ⋅ ⋅ ⋅ + gn fn , where g1 , . . . , gn ∈ S; i.e.,
g1 f1 + ⋅ ⋅ ⋅ + gn−1 fn−1 + q1 ⋯qj (gn qj+1 ⋯qs − h1 ) = 0.
Since f1 , . . . , fn−1 , q1 ⋯qj is a regular sequence, it follows that gn qj+1 ⋯qs − h1 ∈
⟨f1 , . . . , fn−1 ⟩. So h1 ∈ ⟨f1 , . . . , fn−1 ⟩ in the ring S/⟨qm ⟩, which implies that h ∈
⟨f1 , . . . , fn−1 ⟩. Conversely, fi ∈ (P ∶ q1 ⋯qj ) + ⟨qm ⟩/⟨qm ⟩ for all 1 ≤ i ≤ n − 1. So
(P ∶ q1 ⋯qj ) + ⟨qm ⟩/⟨qm ⟩ is an ideal in S/⟨qm ⟩ generated by f1 , . . . , fn−1 . Similarly,
(P ∶ q1 ⋯qj ) + ⟨qk ⟩/⟨qk ⟩ is an ideal in S/⟨qk ⟩ generated by f1 , . . . , fn−1 . By Lemma
2.7, it follows that H(S/(P ∶ q1 ⋯qj ) + ⟨qk ⟩) = H(S/(P ∶ q1 ⋯qj ) + ⟨qm ⟩).
Theorem 4.2. Assume that the EGH Conjecture holds in k[x1 , . . . , xn−1 ], where
n ≥ 2. If I is a graded ideal in S = k[x1 , . . . , xn ] containing a regular sequence
f1 , . . . , fn−1 , fn = q1 ⋯qs of degrees deg(fi ) = ai such that qi ∈ S1 for all 1 ≤ i ≤ s,
then I has the same Hilbert function as a graded ideal in S containing xa1 1 , . . . , xann .
Proof. We check the property (b) of Lemma 2.6. Let d ≥ 0. We need to find a
graded ideal K in S containing xa1 1 , . . . , xann such that H(S/I, d) = H(S/K, d) and
H(S/I, d + 1) ≤ H(S/K, d + 1). Let J to be the ideal generated by f1 , . . . , fn and
Id . By renaming the linear polynomials q1 , . . . , qs , we may assume without loss of
generality that
∣Jd ∩ ⟨q1 ⟩d ∣ ≥ ∣Jd ∩ ⟨qi ⟩d ∣ for all 2 ≤ i ≤ s,
∣(J ∶ q1 )d−1 ∩ ⟨q2 ⟩d−1 ∣ ≥ ∣(J ∶ q1 )d−1 ∩ ⟨qi ⟩d−1 ∣ for all 3 ≤ i ≤ s,
∣(J ∶ q1 q2 )d−2 ∩ ⟨q3 ⟩d−2 ∣ ≥ ∣(J ∶ q1 q2 )d−2 ∩ ⟨qi ⟩d−2 ∣ for all 4 ≤ i ≤ s,
⋮
∣(J ∶ q1 ⋯qs−2 )d−(s−2) ∩ ⟨qs−1 ⟩d−(s−2) ∣ ≥ ∣(J ∶ q1 ⋯qs−2 )d−(s−2) ∩ ⟨qs ⟩d−(s−2) ∣.
By considering the short exact sequences
10
A.ABEDELFATAH
0 → S/(J ∶ q1 ) Ð→ S/J Ð→ S/J + ⟨q1 ⟩ → 0,
g↦gq1
g↦g
0 → S/(J ∶ q1 q2 ) Ð→ S/(J ∶ q1 ) Ð→ S/(J ∶ q1 ) + ⟨q2 ⟩ → 0,
g↦gq2
g↦g
0 → S/(J ∶ q1 q2 q3 ) Ð→ S/(J ∶ q1 q2 ) Ð→ S/(J ∶ q1 q2 ) + ⟨q3 ⟩ → 0,
g↦gq3
g↦g
⋮
0 → S/(J ∶ q1 ⋯qs−1 ) Ð→ S/(J ∶ q1 ⋯qs−2 ) Ð→ S/(J ∶ q1 ⋯qs−2 ) + ⟨qs−1 ⟩ → 0.
g↦g
g↦gqs−1
we see that H(S/J, t) is equal to
s−2
H(S/J + ⟨q1 ⟩, t) + ∑ H(S/(J ∶ q1 ⋯qi ) + ⟨qi+1 ⟩, t − i) + H(S/(J ∶ q1 ⋯qs−1 ), t − (s − 1))
i=1
for all t ≥ 0. Let J0 = J + ⟨q1 ⟩, Js−1 = (J ∶ q1 ⋯qs−1 ), and for 1 ≤ i ≤ s − 2 let
i+1 ⟩
) = H(S/Ji ) for all
Ji = (J ∶ q1 ⋯qi ) + ⟨qi+1 ⟩. Note that qi+1 ∈ Ji and H( JS/⟨q
i /⟨qi+1 ⟩
0 ≤ i ≤ s − 1. Set S = k[x1 , . . . , xn−1 ]. For all 0 ≤ i ≤ s − 1, S/⟨qi+1 ⟩ is isomorphic
an−1
with the
to S, so by the hypothesis there is an ideal in S containing xa1 1 , . . . , xn−1
same Hilbert function as Ji . For all 0 ≤ i ≤ s − 1, let Li be the lex-plus-powers ideal
an−1
such that H(S/Li ) = H(S/Ji ).
in S containing xa1 1 , . . . , xn−1
Claim: Li,j ⊆ Li+1,j for all 0 ≤ i ≤ s − 2 and j ≤ d − i, where Li,j is the j-th
component of the ideal Li .
Proof of the claim: Assume that i = 0. If j < d, then by part (a) of Lemma
4.1 we obtain
∣J0,j ∣ = ∣Jj + ⟨q1 ⟩j ∣ = ∣Pj + ⟨q1 ⟩j ∣ = ∣Pj + ⟨q2 ⟩j ∣ ≤ ∣J1,j ∣.
If j = d, then by our assumption we obtain
∣J0,d ∣ = ∣Jd ∣ + ∣⟨q1 ⟩d ∣ − ∣Jd ∩ ⟨q1 ⟩d ∣
≤ ∣Jd ∣ + ∣⟨q1 ⟩d ∣ − ∣Jd ∩ ⟨q2 ⟩d ∣
= ∣Jd ∣ + ∣⟨q2 ⟩d ∣ − ∣Jd ∩ ⟨q2 ⟩d ∣
= ∣Jd + ⟨q2 ⟩d ∣
≤ ∣J1,d ∣.
This means that H(S/J0 , j) ≥ H(S/J1 , j) for all j ≤ d. So H(S/L0 , j) ≥ H(S/L1 , j)
for all j ≤ d. Since L0 and L1 are lex-plus-powers ideals, it follows that L0,j ⊆ L1,j
for all j ≤ d.
Let 0 < i ≤ s − 2. If j < d − i, then by part (b) of Lemma 4.1 we obtain
∣Ji,j ∣ = ∣(J ∶ q1 ⋯qi )j +⟨qi+1 ⟩j ∣ = ∣(P ∶ q1 ⋯qi )j +⟨qi+1 ⟩j ∣ = ∣(P ∶ q1 ⋯qi )j +⟨qi+2 ⟩j ∣ ≤ ∣Ji+1,j ∣.
If j = d − i, then by our assumption we obtain
∣Ji,d−i ∣ = ∣(J ∶ q1 ⋯qi )d−i ∣ + ∣⟨qi+1 ⟩d−i ∣ − ∣(J ∶ q1 ⋯qi )d−i ∩ ⟨qi+1 ⟩d−i ∣
≤ ∣(J ∶ q1 ⋯qi )d−i ∣ + ∣⟨qi+1 ⟩d−i ∣ − ∣(J ∶ q1 ⋯qi )d−i ∩ ⟨qi+2 ⟩d−i ∣
= ∣(J ∶ q1 ⋯qi )d−i ∣ + ∣⟨qi+2 ⟩d−i ∣ − ∣(J ∶ q1 ⋯qi )d−i ∩ ⟨qi+2 ⟩d−i ∣
= ∣(J ∶ q1 ⋯qi )d−i + ⟨qi+2 ⟩d−i ∣
≤ ∣Ji+1,d−i ∣.
Similarly, we conclude that Li,j ⊆ Li+1,j for all j ≤ d − i, and proving the claim.
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
11
Let Ks = {zxs+j
∶ z ∈ Mon(S) ∧ j ≥ 0} and Ki = {zxin ∶ z ∈ Mon(Li )} for all
n
0 ≤ i ≤ s − 1. Define K to be the ideal generated by ⋃0≤i≤s Ki . Since xsn ∈ Ks and
xai i ∈ K0 for all 1 ≤ i ≤ n − 1, it follows that xa1 1 , . . . , xann ∈ K.
Claim: If w is a monomial in K of degree t, where 0 ≤ t ≤ d + 1, then w ∈ ⋃0≤i≤s Ki .
Proof of the claim: There exists a monomial u in ⋃0≤i≤s Ki such that u∣w; i.e.,
w = vu for some monomial v ∈ S. If u ∈ Ks , then w ∈ Ks . Assume that u = zxin ∈ Ki ,
where z ∈ Li for some 0 ≤ i ≤ s − 1. If xn ∤ v, then w ∈ ⋃0≤i≤s Ki . Assume that
xn ∣v. Let r = max{j ∶ xjn ∣v}. If i + r ≥ s, then w ∈ Ks . So we may assume that
i + r < s. By the previous claim, we obtain that Li,j ⊆ Li+r,j for all j ≤ d − (i + r − 1).
Since deg(z) ≤ d + 1 − (i + r), it follows that z ∈ Li+r . So xvr z ∈ Li+r , and then
n
v
zxr+i
n
xrn
= w ∈ Ki+r . Hence, we proved the claim.
We conclude that the number of monomials in K of degree t, where 0 ≤ t ≤ d + 1,
t−s
is equal to ∑s−1
i=0 ∣Li,t−i ∣ + ∑i=0 ∣S i ∣. Since ∣St ∣ = ∑0≤i≤t ∣S i ∣, it follows that
∣St ∣ − ∣Kt ∣ =
t
∑
i=t−(s−1)
s−1
s−1
s−1
i=0
i=0
i=0
∣S i ∣ − ∑ ∣Li,t−i ∣ = ∑ ∣S t−i ∣ − ∑ ∣Li,t−i ∣.
s−1
So H(S/K, t) = ∑s−1
i=0 H(S/Li , t − i) = ∑i=0 H(S/Ji , t − i) = H(S/J, t). In particular,
H(S/K, d) = H(S/J, d) = H(S/I, d) and
H(S/K, d + 1) = H(S/J, d + 1) ≥ H(S/I, d + 1).
Corollary 4.3. If I is a graded ideal in S containing a regular sequence f1 , . . . , fn
with deg(fi ) = ai such that fi splits into linear factors for all i, then I has the same
Hilbert function as a graded ideal in S containing xa1 1 , . . . , xann .
Since the EGH Conjecture holds when n = 2, we obtain the following.
Corollary 4.4. Let n ≥ 3. If I is a graded ideal in S containing a regular sequence
f1 , . . . , fn with deg(fi ) = ai such that fi splits into linear factors for all 3 ≤ i ≤ n,
then I has the same Hilbert function as a graded ideal in S containing xa1 1 , . . . , xann .
By 4.3, the EGH Conjecture is equivalent to the following conjecture.
Conjecture 4.5. If I is a homogeneous ideal in S containing a regular sequence
f1 , . . . , fn of degrees deg(fi ) = ai , then I has the same Hilbert function as an ideal
containing a regular sequence g1 , . . . , gn of degrees deg(gi ) = ai , where gi splits into
linear factors for all i.
5
Example 4.6. Let S = C[x1 , . . . , x5 ], fi = xi (∑i−1
j=1 −xj )+xi (∑j=i xj ) for all 1 ≤ i ≤ 5
and
1
1
1 1 ⎞
⎛ 1
−1
1
1
1
1 ⎟
⎜
⎜
⎟
1 1 ⎟.
A = ⎜ −1 −1 1
⎜
⎟
⎜ −1 −1 −1 1 1 ⎟
⎝ −1 −1 −1 −1 1 ⎠
Since det A[i1 , . . . , ir ] ≠ 0 for all 1 ≤ r ≤ 5 and 1 ≤ i1 < ⋯ < ir ≤ 5, it follows that
f1 , . . . , f5 is a regular sequence in S. Assume that I = ⟨f1 , . . . , f5 , x1 x2 + x1 x3 , x21 +
x4 x5 ⟩. In this example, we construct an ideal in S with the same Hilbert function
12
A.ABEDELFATAH
as I, using the Hilbert functions of J0 = I + ⟨x5 ⟩ and J1 = (I ∶ x5 ). Computation
with Macaulay2 shows that
HS/I = (1, 5, 8, 3, 0, 0, . . . ), HS/J0 = (1, 4, 4, 1, 0, 0, . . . ) and HS/J1 = (1, 4, 2, 0, 0, . . . )
are the Hilbert sequence of I, J0 and J1 , respectively. Denote by R the polynomial
ring C[x1 , . . . , x4 ]. Let
L0 = ⟨x21 , . . . , x24 , x1 x2 , x1 x3 ⟩ ⊂ R and L1 = ⟨x21 , . . . , x24 , x1 x2 , x1 x3 , x1 x4 , x2 x3 ⟩ ⊂ R.
Note that L0 and L1 are lex-plus-powers ideals in R. We can see that L0,0 = L0,1 =
(0) and
L0,2 = (x21 , x1 x2 , x1 x3 , x22 , x23 , x24 ),
L0,3 = (w ∶ w ∈ Mon(R3 ) and w ≠ x2 x3 x4 ),
L0,j = Rj for all j ≥ 4.
So we have HR/L0 = HS/J0 . Also, we have L1,0 = L1,1 = (0) and
So we have HR/L1
L1,2 = (x21 , x1 x2 , x1 x3 , x1 x4 , x22 , x2 x3 , x23 , x24 ),
L1,j = Rj for all j ≥ 3.
= HS/J1 . Let K to be the ideal in S generated by
Mon(L0 ) ∪ {wx5 ∶ w ∈ Mon(L1 )} ∪ {x25 }.
Then K = ⟨x21 , x22 , x23 , x24 , x25 , x1 x2 , x1 x3 , x1 x4 x5 , x2 x3 x5 ⟩. It is clear that ∣S0 /K0 ∣ = 1
and ∣S1 /K1 ∣ = 5. Since S2 /K2 = (x1 x4 , x1 x5 , x2 x3 , x2 x4 , x2 x5 , x3 x4 , x3 x5 , x4 x5 ),
it follows that ∣S3 /K3 ∣ = 8. Also we have S3 /K3 = (x2 x3 x4 , x2 x4 x5 , x3 x4 x5 ) and
Kj = Sj for all j ≥ 4. Thus
HS/K = (1, 5, 8, 3, 0, 0, . . . ) = HS/I .
6
Example 4.7. Let S = C[x1 , . . . , x6 ], fi = xi (∑i−1
j=1 −xj )+xi (∑j=i xj ) for all 1 ≤ i ≤ 5
f6
2
and f6 = x6 (−x1 − x2 − x3 − x4 − x5 + x6 ). Since f1 , . . . , f5 , x6 is a regular sequence,
it follows that f1 , . . . , f6 is a regular sequence in S. Assume that
I = ⟨f1 , . . . , f6 , x1 x2 + x3 x4 , x1 x6 + x25 , x22 x3 ⟩.
Computation with Macaulay2 shows that
HS/I =(1, 6, 14, 13, 2, 0, . . . ),
HS/I+⟨x6 ⟩ =(1, 5, 8, 2, 0, . . . ),
HS/(I∶x6 )+⟨x6 ⟩ =(1, 5, 6, 0, . . . ),
HS/(I∶x26 ) =(1, 5, 2, 0, . . . ).
Also we have
∣I2 ∩ ⟨x6 ⟩2 ∣ = ∣I2 ∩ ⟨−x1 − x2 − x3 − x4 − x5 + x6 ⟩2 ∣
and
∣(I ∶ x6 )1 ∩ ⟨x6 ⟩1 ∣ = ∣(I ∶ x6 )1 ∩ ⟨−x1 − x2 − x3 − x4 − x5 + x6 ⟩1 ∣.
We construct an ideal in S with the same Hilbert function as I, using the Hilbert
functions of I + ⟨x6 ⟩, (I ∶ x6 ) + ⟨x6 ⟩ and (I ∶ x26 ). Denote by J0 , J1 and J2 the ideals
I + ⟨x6 ⟩, (I ∶ x6 ) + ⟨x6 ⟩ and (I ∶ x26 ), respectively. Let R = C[x1 , . . . , x5 ] and L0 =
⟨x21 , . . . , x25 , x1 x2 , x1 x3 , x1 x4 x5 , x2 x3 x4 , x2 x3 x5 ⟩ ⊂ R. An easy calculation shows that
L0 is a lex-plus-powers ideal in R and HR/L0 = (1, 5, 8, 0, . . . ) = HS/I+⟨x6 ⟩ . Let
L1 = ⟨x21 , . . . , x25 , x1 x2 , x1 x3 , x1 x4 , x1 x5 , x2 x3 x4 , x2 x3 x5 , x2 x4 x5 , x3 x4 x5 ⟩ ⊂ R. We
can see that L1 is a lex-plus-powers ideal and HR/L1 = (1, 5, 6, 0, . . . ) = HS/J1 . Let
ON THE EISENBUD-GREEN-HARRIS CONJECTURE
13
L2 = ⟨x21 , . . . , x25 , x1 x2 , x1 x3 , x1 x4 , x1 x5 , x2 x3 , x2 x4 , x2 x5 , x3 x4 ⟩ ⊂ R. Also we have
that L2 is a lex-plus-powers ideal in R and HR/L2 = (1, 5, 3, 0, . . . ) = HS/J2 . Let K
to be the ideal in S generated by Mon(L0 ) ∪ {wx6 ∶ w ∈ Mon(L1 )} ∪ {wx26 ∶ w ∈
Mon(L2 )} ∪ {x36 }. The ideal K generated by
{x21 , . . . , x25 , x36 , x1 x2 , x1 x3 , x1 x4 x5 , x2 x3 x4 , x2 x3 x5 , x1 x4 x6 }
⋃
{x1 x5 x6 , x2 x4 x5 x6 , x3 x4 x5 x6 , x2 x3 x26 , x2 x4 x26 , x2 x5 x26 , x3 x4 x26 }.
Computation with Macaulay2 shows that HS/K = (1, 6, 14, 13, 2, 0, . . . ) = HS/I .
References
[1] A. Abedelfatah. Macaulay-Lex rings. Journal of Algebra, 374:122-131, 2013.
[2] A. Aramova, J. Herzog, and T. Hibi. Gotzmann Theorems for Exterior Algebras and Combinatorics. Journal of Algebra, 191(1):174-211, 1997.
[3] G. Caviglia and D. Maclagan. Some cases of the Eisenbud-Green-Harris conjecture. Mathematical Research Letters, 15(3):427-433, 2008.
[4] R.I.X. Chen. Some special cases of the Eisenbud-Green-Harris Conjecture.
[5] G.F. Clements and B. Lindström. A generalization of a combinatorial theorem of Macaulay.
Journal of Combinatorial Theory, 7(3):230-238, 1969.
[6] S.M. Cooper. Growth conditions for a family of ideals containing regular sequences. Journal
of Pure and Applied Algebra, 212(1):122-131, 2008.
[7] S.M. Cooper. The Eisenbud-Green-Harris Conjecture for Ideals of Points. 2008.
[8] D. Eisenbud, M. Green, and J. Harris. Higher Castelnuovo Theory. Astérisque, 218:187, 1993.
[9] J. Herzog and T. Hibi. Monomial Ideals, volume 260. Springer Verlag, 2010.
[10] J. Herzog and D. Popescu. Hilbert functions and generic forms. Compositio Mathematica,
113(1):1-22, 1998.
[11] G.O.H. Katona. A theorem of finite sets. Theory of graphs, pages 187-207, 1968.
[12] J.B. Kruskal. The Number of Simplices in a Complex. Mathematical optimization techniques,
page 251, 1963.
[13] F. Macaulay. Some properties of enumeration in the theory of modular systems. Proceedings
of the London Mathematical Society, 2(1):531, 1927.
[14] H. Matsumura. Commutative ring theory, volume 8 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge,, 1989.
[15] J. Mermin and I. Peeva. Lexifying ideals. Mathematical Research Letters, 13(2/3):409, 2006.
[16] B.P. Richert. A study of the lex plus powers conjecture. Journal of Pure and Applied Algebra,
186(2):169-183, 2004.
[17] DA. Shakin. Piecewise lexsegment ideals. Sbornik: Mathematics, 194:1701, 2003.
Department of Mathematics, University of Haifa, Mount Carmel, Haifa 31905, Israel
E-mail address: [email protected]
| 0math.AC
|
arXiv:1607.06334v1 [math.GR] 21 Jul 2016
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
DANIEL J. WOODHOUSE
Abstract. A group is tubular if it acts on a tree with Z2 vertex stabilizers and
Z edge stabilizers. We prove that a tubular group is virtually special if and only
if it acts freely on a locally finite CAT(0) cube complex. Furthermore, we prove
that if a tubular group acts freely on a finite dimensional CAT(0) cube complex,
then it virtually acts freely on a three dimensional CAT(0) cube complex.
1. Introduction
A tubular group G splits as a graph of groups with Z2 vertex groups and Z edge
groups. Equivalently, G is the fundamental group of a graph of spaces, denoted by X,
with each vertex space homeomorphic to a torus and each edge space homeomorphic
to S 1 × [−1, 1]. The graph of spaces X is a tubular space. In this paper all tubular
groups will be finitely generated and therefore have compact tubular spaces.
Tubular groups have been studied from various persectives: Brady and Bridson
provided tubular groups with isoperimetric function nα for all α in a dense subset
of [2, ∞) in [1]. Cashen determined when two tubular groups are quasi-isometric
[5]. Wise determined whether or not a tubular group acts freely on a CAT(0) cube
complex [11], and classified which tubular groups are cocompactly cubulated. The
author determined a criterion for finite dimensional cubulations [13]. Button has
proven that all free-by-cyclic groups that are also tubular groups act freely on finite
dimensional cube complexes [3]. The main theorem of this paper is:
Theorem 5.10. A tubular group G acts freely on a locally finite CAT(0) cube
complex if and only if G is virtually special.
Haglund and Wise introduced special cube complexes in [7]. The main consequence of a group being special is that it embeds in a right angled Artin group (see
[9] or [10] for a full outline of Wise’s program).
1.1. Structure of the Paper. In [11] Wise obtained free actions of tubular groups
on CAT(0) cube complexes by first finding equitable sets that allow the construction
e W)
of immersed walls. Such a set of immersed walls determines a wallspace (X,
e W) which G acts freely on. Wallspaces were
which yields a dual cube complex C(X,
first introduced by Haglund and Paulin [6], and the dual cube complex construction
1
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
2
was first developed by Sageev [8]. In [13] the author defined a criterion, called
dilation, that determines if an immersed wall produces infinite or finite dimensional
e W) is
cubulations. More precisely, if the immersed walls are non-dilated, then C(X,
finite dimensional. We recall the relevant definitions and background in Section 2.
Section 3 establishes a technical result using techniques from [13]. It is shown that
that immersed walls can be replaced with primitive immersed walls without losing
the finite dimensionality or local finiteness of the associated dual cube complex. The
reader is encouraged to either read this section alongside [13], or skip it on a first
reading.
e W) in the finite dimensional case to establish a set of
In Section 4 we analyse C(X,
e W) is virtually special. We decompose C(X,
e W)
conditions that imply that G\C(X,
e and then, under the assumpas a tree of spaces, with the same underlying tree as X
e W) maps G-equivariantly into
tion that the walls are primitive, we show that C(X,
e where R is the standard cubulation of R, and Γ
e is the underlying graph
Rd × Γ,
e A further criterion, the notion of a fortified immersed wall, determines when
of X.
e W) is locally finite. Combining these results allow us to give criterion for
C(X,
e W) to be virtually special.
G\C(X,
In Section 5 we consider a tubular group acting freely on a CAT(0) cube complex
e
Y . We show that we can obtain from such an action immersed walls that preserve
the important properties of Ye . More precisely, we prove the following:
Proposition 5.9. Let G be a tubular group acting freely on a CAT(0) cube complex Ye . Then there is a tubular space X and a finite set of immersed walls in X.
e W) is the associated wallspace, then
Moreover, if (X,
e W).
(1) G acts freely on C(X,
e W) is finite dimensional if Ye is finite dimensional.
(2) C(X,
e W) is finite dimensional and locally finite if Ye is locally finite.
(3) C(X,
This Proposition is sufficient to allow us to prove 5.10.
In Section 6 we further exploit the results obtained in Section 4 to obtain the
following, demonstrating that the cubical dimension of tubular groups with finite
dimensional cubulations are virtually within 1 of their cohomological dimension.
Theorem 6.2. A tubular group acting freely on a finite dimensional CAT(0) cube
complex has a finite index subgroup that acts freely on a 3-dimensional CAT(0) cube
complex.
Acknowledgements: I would like to thank Dani Wise and Mark Hagen.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
3
2. Background Tubular Groups and their Cubulations
Let G be a tubular group with associated tubular space X and underlying graph
Γ. Given an edge e in a graph we will let −e and +e respectively denote the initial
and terminal vertices of e. Let Xv and Xe denote vertex and edge spaces in this
graph of spaces. Let X−e and X+e be the boundary circles of Xe , and denote the
−
+
+
−
+
attaching maps by ϕ−
e : Xe → X−e , and ϕe : Xe → X+e . Note that ϕe and ϕe
e denote the
respectively represent generators of Ge in Ge− and Ge+ . We will let X
eve and X
eee denote vertex and edge spaces in the universal
universal cover of X. Let X
e and let Γ
e denote the Bass-Serre tree. We will assume that each vertex
cover X,
space has the structure of a nonpositively curved geodesic metric space and that
+
attaching maps ϕ−
e and ϕe define locally geodesic curves in X−e and X+e .
2.1. Equitable Sets and Intersection Numbers. Given a pair of closed curves
in a torus α, β : S 1 → T , the intersection points are the elements (p, q) ∈ S 1 × S 1
such that α(p) = β(q). For a pair of homotopy classes [α], [β] of closed curves in a
torus T , their geometric intersection number # [α], [β] is the minimal number of
intersection points realised by a pair of representatives from the respective classes.
This number is realised by any pair of geodesic representatives of the classes. If
B = {[βi ]} is a finite set of homotopy classes of curves in T , then #[α, B] :=
P
#[α, βi ]. Viewing [α] and [β] as elements of π1 T = Z2 , we can compute that
i
# [α], [β] = det [α], [β] . Given an identification of Z2 with π1 T , the elements of
Z2 are identified with homotopy classes of curves in T , so it makes sense to consider
their geometric intersection number. An equitable set for a tubular group G is a
collection of sets {Sv }v∈Γ , where Sv is a finite set of distinct geodesic curves in Xv
disjoint from the attaching maps of adjacent edge spaces, such that Sv generate
+
a finite index subgroup of π1 Xv = Gv , and # ϕ−
e , Se− = # ϕe , Se+ ]. Note that
equitable sets can also be given with Sv a finite subset of Gv that generates a
finite index subgroup of Gv and satisfies the corresponding equality for intersection
numbers. This is how Wise formulates equitable sets, and its equivalence follows
from exchanging elements of Gv = π1 Xv with geodesic closed curves in Xv that
represent the corresponding elements. An equitable set is fortified if for each edge
−
+
+
−
−
e in Γ, there exists α+
e ∈ S+e and αe ∈ S−e such that #[αe , ϕe ] = #[αe , ϕe ] = 0.
An equitable set is primitive if every element α ∈ Sv represents a primitive element
in Gv .
2.2. Immersed Walls From Equitable Sets. Immersed walls are constructed
from circles and arcs. For each α ∈ Sv , let Sα1 be the domain of α. The disjoint union
F 1
+
Sα over all α ∈ Sv and v ∈ V Γ are the circles. Since #[ϕ−
e , Se− ] = #[ϕe , Se+ ],
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
4
there exists a bijection from the intersection points between curves in S−e and ϕ−
e,
+
−
−
1
and the intersection points between curves in S+e and ϕe . Let (p , q ) ∈ Sα− × Xe−
and (p+ , q + ) ∈ Sα1 + be corresponding intersection points between α± ∈ S±e and ϕ±
e.
−
+
∼
Then an arc a = [0, 1] has its endpoints attached to p and p . The endpoints of a
are mapped into X−e ∩Xe and X+e ∩Xe , so the interior of a can be embedded in Xe .
After attaching an arc for each pair of corresponding intersection points, we obtain a
set {Λ1 , . . . , Λn } of connected graphs that map into X, called immersed walls. Each
graph Λi has its own graph of groups structure with infinite cyclic vertex groups
and trivial edge groups. As in [13], all “immersed walls” in this paper are
immersed walls constructed from equitable sets as above.
This means
that we are free use the results obtained in [13].
ei → X
e is a two sided embedding in X,
e separating X
e into two halfspaces.
A lift of Λ
e i to X
e are horizontal walls Wh . The vertical walls
The images of the lifts of Λ
Wv are obtained from the lifts of curves αe : S 1 → Xe given by the inclusion
S 1 × {0} ֒→ S 1 × [−1, 1]. The set W = Wh ⊔ Wv of all horizonal and vertical walls
e W) where the G-action on X
e also gives an action on W. The
gives a wallspace (X,
main theorem of [11] is that a tubular group acts freely on a CAT(0) cube complex
if and only if there exists an equitable set.
A set of immersed walls is fortified if they are obtained from a fortified equitable
set. A set of immersed walls is primitive if they are obtained from a primitive
equitable set.
e and Λ
e ′ be horizontal walls in X.
e An point x ∈ Λ∩
e Λ
e is a regular intersection
Let Λ
eve, and the lines Λ∩
e X
eve and Λ
e′ ∩ X
eve are non-parallel.
point if it lies in a vertex space X
′
e ∩Λ
e is a non-regular intersection point, and either x ∈ X
eve
Otherwise, a point x ∈ Λ
′
e∩X
eve = Λ
e ∩X
eve, or x ∈ X
eee.
where Λ
An infinite cube in a CAT(0) cube complex Ye , is an sequence of c0 , c1 , . . . , cn , . . .
such that cn is an n-cube in Ye , and cn is a face of cn+1 . In [13] a dilation function
is constructed for each immersed wall R : π1 Λ → Q∗ , and an immersed wall is said
to be dilated if R has infinite image. The following is Thm 1.2 from that paper:
e W) the wallspace obtained from a
Theorem 2.1. Let X be tubular space, and (X,
finite set of immersed walls in X. The following are equivalent:
e W) is infinite dimensional.
(1) The dual cube complex C(X,
e W) contains an infinite cube.
(2) The dual cube complex C(X,
(3) One of the immersed walls is dilated.
The following result is also obtained from [13] by combining Thm 1.2, Prop 4.6,
and Prop 3.4. The last part follows from the last paragraph of the proof of Prop
3.4.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
5
e W) the wallspace obtained from
Proposition 2.2. Let X be tubular space, and (X,
e W) is infinite dimensional, then W
a finite set of immersed walls in X. If C(X,
contains an set of pairwise regularly intersecting walls of infinite cardinality, that
e W). Moreover, the infinite
correspond to the hyperplanes in an infinite cube in C(X,
cube contains a canonical 0-cube.
3. Primitive Immersed Walls
The following result uses the techniques in Section 5. of [13] to compute the
dilation function. Let Λ be an immersed wall in X, and let R : π1 (Λ) → Q∗ be
its dilation function. If R has finite image then Λ is non-dilated. Let q : Λ → Ω
be the quotient map obtained by crushing each circle to a vertex. Note that the
arcs in Ω correspond to the arcs in Λ. The dilation function R factors through
q∗ : π1 Λ → π1 Ω, so there exists a function R̂ : π1 Ω → Q∗ such that R = R̂ ◦ q∗ . We
can therefore determine if Λ is dilated by computing the function R̂.
We orient each arc in Λ so that all arcs embedded in the same edge space are
oriented in the same direction. We orient the arcs in Ω accordingly. We define a
weighting ω : E(Ω) → Q∗ . Let Xe be an edge space in X, and let σ be an arc
mapped into Xe connecting the circles C − and C + . Let α− : C − → X−e and
α+ : C + → X+e be the corresponding elements in the equitable set. Then
ω(σ) =
−
#[ϕ−
e ,α ]
.
+
#[ϕ+
e ,α ]
If γ = σ1ǫ1 · · · σnǫn is an edge path in Ω where σi is an oriented arc in Ω, and ǫ ∈ ±1,
then R̂(γ) = ω(σ1 )ǫ1 · · · ω(σn )ǫn .
Lemma 3.1. Let X be a tubular space, and let G = π1 X. Let Λ1 , . . . Λk be a set
of immersed walls in X obtained from an equitable set {Sv }v∈Γ . Then there exists
a set of primitive immersed walls Λ′1 , . . . , Λ′ℓ in X obtained from an equitable set
{Sv′ }v∈Γ . Moreover:
(1) If Λ1 , . . . , Λk are non-dilated, then so are Λ′1 , . . . Λ′ℓ .
(2) If Λ1 , . . . Λk , are fortified, then so are Λ′1 , . . . Λ′ℓ .
Proof. Each Λi decomposes as the union of disjoint circles, which are the domain of
locally geodesic closed paths in the equitable set and arcs. Suppose that αn ∈ Sv ,
where [α] ∈ π1 Xv = Gv is primitive. Let Λi be the immersed wall containing the
circle Sα1 n corresponding to αn .
A new equitable set is obtained by replacing αn in Sv with n locally geodesic
curves {αi : Si1 → Xv }ni=1 with disjoint images in Xv that are isotopic to α in
ev . This remains an equitable set since #[αn , γ] = n#[α, γ] = Pn #[αi , γ] for
X
i=1
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
6
any locally geodesic curve γ in Xv . New immersed walls are obtained from Λi by
replacing Sg1n with S11 , . . . Sn1 and reattaching the arcs that were attached to the
intersection points in Sα1 n to the corresponding intersection points on S11 , . . . Sn1 . Let
Λi1 , . . . , Λiℓ be the new set of immersed walls obtained in this way. Note that each
arc in Λi1 , . . . , Λiℓ , corresponds to a unique arc in Λi .
Assume that Λi is non-dilated. We claim that the new immersed walls Λi1 , . . . , Λiℓ
are also non-dilated. Let qi : Λi → Ωi and qij : Λij → Ωij be the quotient maps
obtained by crushing the circles to vertices. Let u be the vertex in Ωi corresponding
to Sα1 n . Let Ri : π1 Λi → Q∗ and Rij : π1 Λij → Q∗ be the dilation functions. Let
R̂i : π1 Ωi → Q∗ and R̂ij : π1 Ωij → Q∗ be the unique maps such that Ri = R̂i ◦ qi
and Rij = R̂ij ◦ qij . Let ωi and ωij be the respective weightings of the arcs in Ωi
and Ωij . By assumption, Ri and R̂i have finite image. As the arcs in Λij correspond
to arcs in Λi , there is a map ρi : Ωij → Ωi . We show Λij is non-dilated by showing
that R̂ij (γ) = R̂i (ρi ◦ γ).
−1
Let σ be an oriented arc in Ωij . The edge qij
(σ) embeds in an edge space Xe .
ǫ
−1
If the vertices of σ are disjoint from ρ (u), then ωij (σ ǫ ) = ωi (ρij ◦ σ ǫ ). If the
1
1
endpoints of σ ǫ are contained in ρ−1
ij (v), and correspond to the circles Sι and Sτ
then
ωij (σ ±1 ) =
n
#[ϕ∓
n#[ϕ∓
#[ϕ∓
e , αι ]
e , αι ]
e ,α ]
=
=
= ω(ρij ◦ σ ±1 ).
±
±
n]
#[ϕ±
,
α
]
n#[ϕ
,
α
]
#[ϕ
,
α
e
e
e
τ
τ
ǫ
Suppose that exactly one endpoint of σ ǫ is contained in ρ−1
ij (u). If σ terminates
1
a vertex in ρ−1
ij (u) corresponding to Sτ , and the initial vertex corresponds to a circle
that is the domain of a locally geodesic curve β then
ωij (σ ±1 ) =
n#[ϕ∓
n#[ϕ∓
#[ϕ∓
e , β]
e , β]
e , β]
=
=
= nω(ρij ◦ σ ±1 ).
±
±
±
#[ϕe , ατ ]
n#[ϕe , ατ ]
#[ϕe , αn ]
1
If σ ǫ starts at a vertex in ρ−1
ij (u) corresponding to Sι , and the terminal vertex
correspond to a circle that is the domain of a locally geodesic curve β then
ωij (σ ±1 ) =
n
n#[ϕ∓
#[ϕ∓
1
#[ϕ∓
e , αι ]
e ,α ]
e , αι ]
=
=
= ω(ρij ◦ σ ±1 ).
±
±
±
n
#[ϕe , β]
n#[ϕe , β]
n#[ϕe , β]
Therefore, given an edge path γ in Ωij , since the number of edges exiting vertices
in ρ−1
ij (v) is the same as the number of vertices entering, R̂ij (γ) = R̂i (ρi ◦ γ).
This procedure produces immersed walls with one fewer non-primitive element in
the equitable set. Repeating this procedure for each non-primitive element in the
equitable set produces a primitive set of non-dilated immersed walls. It is also clear,
that if Λ1 , . . . , Λk are fortified, then so are the new immersed walls.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
7
4. Finite Dimensional Dual Cube Complexes
e W) be the wallspace obtained
Let X be a tubular space and let G = π1 X. Let (X,
from a set of non-dilated immersed walls Λ1 , . . . Λk constructed from an equitable
set, and a vertical immersed wall in each edge space. We emphasize that in this
section all immersed walls are assumed to be non-dilated, even when it
e = C(X,
e W) and let Z = G\Z.
e By Theorem 2.1,
is not explicitly stated. Let Z
e being finite dimensional.
the immersed walls being non-dilated is equivalent to Z
e let Λ
e ee denote the vertical wall in X
eee.
For each edge ee in Γ
We refer to [12] for full background on the dual cube complex construction. A
e is a choice of halfspace z[Λ]
e of Λ
e for each Λ
e ∈ W such that
0-cube z in Z
e1, Λ
e 2 ∈ W, then z[Λ
e 1 ] ∩ z[Λ
e 2 ] 6= ∅.
(1) If Λ
e ∈ W such that x ∈
e
(2) If x ∈ X, then there are only finitely many Λ
/ z[Λ].
e = z2 [Λ]
e for all but precisely one hyperplane
Two 0-cubes z1 , z2 are adjacent if z1 [Λ]
e 12 ∈ W. The 1-cube joining z1 and z2 is dual to the hyperplane corresponding to
Λ
e 12 . An n-cube is then present wherever the 1-skeleton of an n-cube appears. We
Λ
e Λ
e ′ face each other in z if z[Λ]
e is not contained in z[Λ
e ′]
say that two disjoint walls Λ,
and vice versa.
e Therefore Ze decomProposition 4.1. There is a G-equivariant map f : Ze → Γ.
−1
−1
eee = f (e
poses as a tree of spaces with Zeve = f (e
v ), and Z
e) is the carrier of the
e ee ∈ Wv .
hyperplane corresponding to Λ
e such that f (c) = e.
By f −1 (e) we mean the union of all cubes c in Z
Proof. As there is a vertical wall in each edge space, and since the vertical walls are
e Wv ) with the Bass-Serre tree Γ
e of X.
e We define a
all disjoint we can identify C(X,
e let z be a 0-cube in Z, then define f (z) by letting f (z)[Λ
e e ] = z[Λ
e e ].
map f : Ze → Γ:
e 6= z2 [Λ]
e for precisely one wall Λ
e ∈ W. If Λ
e
If z1 , z2 are adjacent 0-cubes, then z1 [Λ]
is a horizontal wall then f (z1 ) = f (z2 ) and the 1-cube joining them is also mapped
e=Λ
e e ∈ Wv then f (z1 ) and f (z2 ) are adjacent in Γ,
e so the
to the same vertex. If Λ
1-cube joining z1 and z2 maps to the edge joining f (z1 ) and f (z2 ). As f is defined
e Then
on the 1-skeleton, the map extends uniquely to the entire cube complex Z.
eve = f −1 (e
Z
v ) and Zeee = f −1 (e
e) is the carrier of the hyperplane corresponding to
e
Λe .
Proposition 4.1 implies that Z decomposes as a graph of spaces with vertex spaces
e
Zv , edge spaces Ze , and underlying graph Γ = G\Γ.
The following proposition, which collects the principal consequences of finite dimensionality, is Prop 4.12 in [13].
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
8
Proposition 4.2. Let X be tubular space with geodesic attaching maps, and let
e W) be the wallspace obtained from a finite set of immersed walls in X. If the
(X,
e W) is finite dimensional, then the horizontal walls in W
dual cube complex C(X,
can be partitioned into a collection P of subsets such that:
(1) The partition P is preserved by G,
(2) For each A ∈ P, the walls in A are pairwise non-intersecting.
e ∈ A ∈ P be a wall intersecting X
eve. There exists h ∈ Gve stabilizing an
(3) Let Λ
eve perpendicular to Λ
e ∩X
eve such that A = {hr Λ}
e r∈Z .
axis in X
Any partition of the horizontal walls in W satisfying conditions (1)-(3) in Proposition 4.2 will be called a stable partition.
e W) be the wallspace obtained from a
Lemma 4.3. Let X be a tubular space and (X,
finite set of immersed walls in X. Let P be a stable partition of the horizontal walls
e only finitely many A ∈ P contain walls intersecting
in W. Then for each ve ∈ Γ,
eve.
X
e is a wall intersecting X
eve, then, by condition (3) of a stable
Proof. Suppose that Λ
e ∩X
eve such that
partition, there exists some h ∈ Gve that is perpendicular to Λ
e r∈Z ∈ P. By G-invariance we can deduce that each of the Gve-translates of
{hr Λ}
e r∈Z is also in P. There are only finitely many such translates, therefore each
{hr Λ}
eve is contained in finitely many elements of P. The claim then
Gve-orbit of a wall in X
follows from the fact that there are only finitely many Gve-orbits of walls intersecting
eve.
X
The immersed walls Λ1 , . . . , Λk are non-dilated, and therefore Ze is finite dimensional, so by Proposition 4.2 there exists a stable partition P of the horizontal walls
eve. Let Pee be the
in W. Let Pve be the subpartition containing walls intersecting X
eee. By Lemma 4.3, both Pve and Pee are finite subsubpartition of walls intersecting X
partitions. If ee is incident to the vertex e
v , then Pee ⊆ Pve. Let Pve = {A1 , . . . , Adve }.
e i }r∈Z such that hi ∈ Gve stablizes an
By criterion (3) of a stable partition Ai = {hri Λ
eve perpendicular to Λ
ei ∩ X
eve. The action of Gve preserves both the partition
axis in X
Pve and the ordering of the walls in each Ai .
Let R denote the cubulation of R with a vertex for each integer and an edge
joining consecutive integers. Therefore, each 0-cube in Rd is an element of Zd . We
construct a free action of Gve on Rdve . Let g ∈ Gve and let (α1 , . . . , αdve ) be a 0-cube
e i = hβj Λ
ej.
in Rdve . Define the map g · (α1 , . . . , αdve ) = (β1 , . . . , βdve ) such that g · hiαi Λ
j
d
As g permutes the walls in Pve, the map g is a bijection on the 0-cubes in R ve . If
e i = hβj Λ
e j , then necessarily g · hαi +1 Λ
e i = hβj ±1 Λ
e j , so adjacent 0-cubes are
g · hαi i Λ
j
i
j
mapped to adjacent 0-cubes and the map extends to an isomorphism of Rdve . If
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
9
g · (α1 , . . . , αdve ) = (α1 , . . . , αdve ) then g would stabilize all the walls in Pve, which
eve. Since Gve acts freely on Z
eve this would
would imply that it fixed every 0-cube in Z
imply that g = 1G , and hence Gve acts freely on Rdve .
eve then every wall
We also define an embedding φve : Zeve → Rdve . If z is a 0-cube in Z
e that is either vertical or not in contained in the subpartition Pve has X
eve ⊆ z[Λ].
e
Λ
e for Λ
e in Pve. For 1 ≤ i ≤ de
Therefore z is entirely determined by z[Λ]
v the set
r
ei ∩ X
eve}r∈Z is an infinite collection of disjoint parallel lines in X
eve. As all the
{hi Λ
e for each 0-cube z in Zeve there exists a unique αi ∈ Z
walls in Ai are disjoint in X,
αi e
αi +1 e
such that hi Λi and hi
Λi face each other in z. Let φve(z) = (α1 , . . . , αdve ). Note
that the map is injective and sends adjacent 0-cubes to adjacent 0-cubes, so the
map on the 0-cubes extends to an embedding of the entire cube complex.
eve → Rdve is Gve-equivariant.
Lemma 4.4. The embedding φve : Z
Proof. Let g ∈ Gve. If φve(z) = (α1 , . . . , αdve ) and g · φve(z) = (β1 , . . . , βdve ), then
e i ] = z[ghαi Λ
e i ] = z[hβj Λ
e j ], which implies that φve(gz) = (β1 , . . . , βd ) =
(gz)[hαi i Λ
i
j
v
e
gφve(z).
Let ee be an edge adjacent to e
v . Then either +e
e = ve or −e
e = ve. We define a free
d
e
action of Gee on R × [−1, 1]. After reindexing, let Pee = {A1 , . . . , Ade } ⊆ Pve where
dee ≤ dve. Let (α1 , . . . , αde , ±1) be a 0-cube in Rd × [−1, 1] and let g ∈ Gee. Then
e i = hβj Λ
e j . As in the case
g · (α1 , . . . , αde , ±1) = (β1 , . . . , βde , ±1) such that g · hαi i Λ
j
of vertex spaces, this map extends to an isomorphism of Rde × [−1, 1].
As with the vertex spaces, there is a Gee-equivariant embedding φee : Zeee →
d
R e × [−1, 1]. Let z be a 0-cube in Zeee. Then for each 1 ≤ i ≤ dee there exe i faces hαi +1 Λ
e i in z, and X
e±ee ⊆ z[Λ
e ee]. Define
ists a unique αi such that hαi i Λ
i
φee(z) = (α1 , . . . , αde , ±1).
Let ve = ±e
e. The free action of Gve on Rdve restricts to a free action of Gee. We
claim that we can embed Rde × {±1} into Rdve in a Gee-equivariant way. Let Hee ⊆ Ze
e e . As Zeee is the carrier of Hee, we can identify
be the hyperplane corresponding to Λ
eee with Hee × [−1, 1]. Note that Hee × {±1} embeds as a subspace in Zeve, and φee
Z
de
restricts to an embedding φ±
e = ±e
e.
ee : Hee × {±1} → R , where v
±
d
d
e
v
e
We construct an embedding Ψee : R → R . Recall that Pee = {A1 , . . . , Ade } ⊆
e j ∈ Aj , then X
eee ⊆ z[hr Λ
e
Pve = {A1 , . . . , Adve }. For dee < j ≤ dve if hrj Λ
j j ] for all 0-cubes
αe
αe +1
e j faces h j
z in Zeee. Therefore, there is a unique αeje ∈ Z such that hj j Λ
j
e
0-cube z in Zee and dee < j ≤ dve. Thus we define
ee
ee
Ψ±
ee (α1 , . . . , αde ) = (α1 , . . . , αde , αde +1 , . . . , αdve ).
The Gee-equivariance of Ψ±
ee will require a further assumption:
e j for every
Λ
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
10
Lemma 4.5. The following commutative square is Gee-equivariant provided the immersed walls are primitive.
Hee × {±1}
φ±
e
/ Rde
_
φve
Zeve
Ψ±
e
/ Rdve
Moreover, Ψ±
ee is a Gee-equivariant inclusion that is equivalent to extending the Geed
action on R e by a trivial action on Rdve −de .
Proof. Let z be a 0-cube in Hee × {±1}. Then by construction
±
ee
ee
Ψ±
e(z).
ee ◦ φee (z) = (α1 , . . . , αde , αde +1 , . . . , αdve ) = φv
To verify that Ψee is Gee-equivariant, let g ∈ Gee. For 1 ≤ i ≤ dee there exists
e i = hβj Λ
e j . For dee < i ≤ dve the intersection
1 ≤ j ≤ dee and βj be such that g · hαi i Λ
j
ei ∩ X
eve is a geodesic line parallel to X
eee ∩ X
eve. Thus, Gee stabilizes Λ
ei ∩ X
eve. As the
Λ
e
immersed walls are primitive we can deduce that Gee stabilizes Λi . For dee < i ≤ dve
e i = hα Λ
e i for α ∈ Z, and conclude:
we deduce that g · hα Λ
i
i
g · Ψee(α1 , . . . , αee) = g · (α1 , . . . , αde , αedee +1 , . . . , αedeve )
= (β1 , . . . , βde , αedee +1 , . . . , αedeve )
= Ψee(β1 , . . . , βee)
= Ψee(g · (α1 , . . . , αee)).
Observe that Gee acts trivially on the last dve − dee coordinates.
e which is finite, since there are only finitely many
Let d = max{|Pve| | ve ∈ V Γ},
vertex orbits.
Proposition 4.6. If the immersed walls Λ1 , . . . , Λk are primitive, then G acts freely
e such that the action on the Γ
e factor is the action of G on the Bass-Serre
on Rd × Γ
e
tree. Moreover, there is a G-equivariant embedding φ : Ze → Rd × Γ.
Proof. The Gve and Gee-actions on Rdve and Rde can be equivariantly extended to
actions on Rd such that Gve and Gee act trivially on the additional factors. Therefore,
the Gee-commutative square in Lemma 4.5 can be extended:
Hee × {±1}
φ±
e
/ Rde
/ Rd
_
eve
Z
φve
Ψ±
e
/ Rdve
/ Rd
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
11
e and a G-equivariant embedding of the
Therefore, we obtain a G-action on Rd × Γ
e
tree of spaces Ze → Rd × Γ.
Proposition 4.7. Ze is locally finite if and only if Λ1 , . . . , Λk are fortified.
eve and an
Proof. If Λ1 , . . . , Λk is not fortified, then there exists a vertex space X
eee such that every horizontal wall Λ
e in Pve intersects X
eve as a
adjacent edge space X
e∩X
eve that intersects X
eve ∩ X
eee. Therefore, every horizontal wall intersecting
line Λ
eve intersects X
eee, so Pee = Pve. Let ee1 , . . . , eei , . . . be an enumeration of the Gve-orbit
X
e ee intersects all the horizontal walls in Pve.
of ee. Then Peei = Pve and Λ
i
e = z[Λ]
e for Λ
e 6= Λ
e ee ,
Let z be a 0-cube in Zeve. There is a 0-cube zi such that zi [Λ]
i
e ee,
e
e
and zi [Λeei ] 6= z[Λeei ]. To verify zi is a 0-cube note that every wall in Pve intersects Λ
e that is not Λ
e e has X
eee ⊆ zi [Λ].
e Therefore zi [Λ
e ee] ∩ zi [Λ]
e 6= ∅ for
and every other wall Λ
e ∈ W − {Λ
e ee}. For any walls Λ
e 1, Λ
e 2 ∈ W − {Λ
e ee} the intersection zi [Λ
e 1 ] ∩ zi [Λ
e 2] =
all Λ
e 1 ] ∩ z[Λ
e 2 ] 6= ∅. Finally, if x ∈ X,
e then x ∈ zi [Λ]
e for all but finitely many Λ
e ∈ W,
z[Λ
because it is true for z, which differs from zi on precisely one wall. Each zi is
adjacent to z since they differ on precisely one wall, so z1 , . . . , zi , . . . is an infinite
e is not locally finite.
collection of distinct 0-cubes adjacent to z, and Z
To show that converse, we first observe that the embedding φve : Zeve → Rdve
proves that Zeve is always locally finite, irrespective of whether the immersed walls
eve, and let u
e via an edge ee.
are fortified. Let z be a 0-cube in Z
e be adjacent to ve in Γ
eue such that z[Λ]
e = zee[Λ]
e for
Then z can be adjacent to at most one 0-cube zee in Z
e ∈ W except Λ
e e . This zee may not always define a 0-cube however. Let ee be an
all Λ
e r∈Z ∈ Pve
edge adjacent to ve. As the immersed walls are fortified there exists {hr Λ}
e ∩X
eve}r∈Z is an infinite set of lines parallel to X
eee ∩ X
eve. As {hr Λ}
e is a
such that {hr Λ
e and hr+1 Λ
e are facing in z. There
set of disjoint walls, there exists r such that hr Λ
eg ee ⊆ z[hr Λ]
e ∩ z[hr+1 Λ].
e
are only finitely many edges g1 ee, . . . , gm ee ∈ Gveee such that X
i
r
r+1
eg ee is not contained in z[h Λ]
e ∩ z[h Λ],
e then either
If ge
e is an edge such that X
i
r
r+1
e ∩ zgee [Λ
e e ] = ∅ or zgee [h Λ]
e ∩ zgee [Λ
e e ] = ∅ so zgee is not a 0-cube. As there are
zgee [h Λ]
only finitely many Gve-orbits of edges incident to ve we conclude that zeee is a 0-cube
for finitely many edges ee incident to ve.
Proposition 4.8. If Λ1 , . . . , Λk are primitive, fortified, non-dilated immersed walls,
then G is virtually special.
Proof. By Proposition 4.6, There is a free action of G on Rd × Aut([−1, 1]d ), so G is
e ∼
e Therefore, there is a
a subgroup of Isom(Rd × Γ)
= Zd ⋊ Aut([−1, 1])d × Aut(Γ).
d
d
projection ρ : G → Z ⋊ Aut([−1, 1] ). Each vertex group Gve embeds in Zd , and the
mapping is invariant under conjugation. As there are only finitely orbits of vertices in
e there exists a finite index subgroup (DZ)d 6 Zd such that if ee is incident to ve then
Γ,
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
12
Gee∩(DZ)d is generated by a primitive element in Gve ∩(DZ)d . Let G′ = ρ−1 (DZ)d .
e such that
Then G′ 6 G is a finite index subgroup that embeds in (DZ)d × Aut(Γ)
each edge group is generated by an element that is primitive in the adjacent vertex
e
groups. By Proposition 4.6 there is a G-equivariant embedding φ : Ze → Rd × Γ.
As G′ does not permute the factors of Rd we can deduce that the hyperplanes in
e /G do not self-intersect, so neither do the hyperplanes in Z/G
e ′ . Indeed,
Rd × Γ
they are also 2-sided and cannot inter-osculate.
Let G′′ be a finite index subgroup such that the underlying graph Γ′′ has girth
at least 2. Let ee be an edge such that +e
e = ve. As Λ1 , . . . , Λk are fortified, we
+
e
conclude that dve > dee and Zee is a proper subcomplex of Zeve. As G′′ee is primitive, if
e d 6= hr Λ
e
ee in a direction
g ∈ G′′ve − G′′ee then ghrdve Λ
dve dve as g acts by translation on Xv
v
e
ed ∩ X
eve. Thus, Ze± is not stabilizes by g, so we can deduce that
non-parallel to Λ
v
e
ee
StabG′′ (Ze+ ) = G′′ . Therefore, G′′ \Ze+ embeds in G′′ \Zeve.
v
e
ee
ee
ee
ee
v
e
e Let z be a 0-cube in Zv′′ . Let He be the vertical hyperplane
Let Z ′′ = G′′ \Z.
contained in Ze′′ , and dual to an edge incident to v. As the attaching maps of Ze′′ are
embeddings, and Γ′′ has girth at least s, we deduce that z can only be incident to
one end of a single 1-cube intersected by He . Therefore He does not self-osculate.
e that projects to a 1-cube in Rd . The G′′ -orbit of
Let σ be a 1-cube in Rd × Γ
σ is a set of 1-cubes, that all project to the same factor of Rd , since G′′ does not
permute the factors of Rd . As G′′ does not invert hyperplanes, after subdividing Rd
we can assume that the G′′ -orbit of σ is a disjoint set of 1-cubes. Therefore, after
the corresponding subdivision, we conclude that the horizontal hyperplanes in Z ′′
don’t self-osculate.
We note that the requirement in Proposition 4.8 that the immersed walls are
fortified is necessary, as the following example demonstrates.
Example 4.9. Let G = ha, b, t | [a, b] = 1, tat−1 = ai. We can decompose G as the
cyclic HNN extension of the vertex group Gv = ha, bi with stable letter t. Thus,
G is a tubular group. Let X be the corresponding tubular space with a single
vertex space Xv and edge space Xe . There is an equitable set {α1 , α2 } where α1
is a geodesic curve in Xv representing ab ∈ Gv , and α2 is a geodesic curve in Xv
−
representing ab−1 ∈ Gv . Note that each attaching map ϕ+
e and ϕe intersects each
curve in the equitable set precisely once. Therefore, we obtain a pair of embedded
immersed horizontal walls Λ1 and Λ2 , by connecting respective intersection points
−
with ϕ+
e and ϕe by an arc. A vertical wall Λe is also embedded in Xe .
e W) we can decompose W into three sets of disjoint walls: the
In the wallspace (X,
walls W1 that cover Λ1 , the walls W2 that cover Λ2 , and the walls We that cover Λe .
These walls are disjoint since the immersed walls are embedded. Furthermore, the
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
13
e W) =
walls in different sets pairwise intersect. Therefore we can conclude that C(X,
e As this is not locally finite, G\C(X,
e W) cannot be virtually special.
R2 × Γ.
5. Revisiting Equitable Sets
Although Wise proved in [11] that acting freely on a CAT(0) cube complex Ye
implied the existence of an equitable set, and thus a system of immersed walls
as in Section 2, no relationship was established between Ye and the resulting dual
e W). Proposition 5.9 gives the relationship required to reduce Theorem 5.10
C(X,
to considering cubulations obtained from equitable sets.
This section will apply the following theorem from [14]. A cubical quasiline is a
CAT(0) cube complex quasi-isomorphic to R.
Theorem 5.1. Let G be virtually Zn . Suppose G acts properly and without inversions on a CAT(0) cube complex Ye . Then G stabilizes a finite dimensional
e ⊆ Ye that is isometrically embedded in the combinatorial metric, and
subcomplex Z
Q
m
e∼
Z
= i=1 Ci , where each Ci is a cubical quasiline and m ≥ n. Moreover, StabG (Λ)
e
is a codimension-1 subgroup for each hyperplane Λ in Z.
Theorem 5.1 allows us to prove the following:
Lemma 5.2. Let G be a tubular group acting freely on a CAT(0) cube complex Ye .
ev ⊆ Ye
Let Gv be a vertex group in G, then there exists a Gv -equivariant subspace X
eve has a metric such that the intersection of a
homeomorphic to R2 . Moreover, X
ev is either empty, or a geodesic line.
hyperplane in Ye with X
Proof. By Theorem 5.1, there exists a Gv -equivariant subcomplex Yev ⊆ Ye that isoQm
metrically embeds in the combinatorial metric, and such that Yev ∼
= i=1 Ci where
each Ci is a cubical quasiline. By the flat torus theorem [2], Gv stabilizes a flat
ev ⊆ Yev , that is a convex subset in the CAT(0) metric of Yev . As the stabilizers of
X
hyperplanes in Yev are codimension-1 subgroups of Gv , the intersection of a hyperev is either empty or a geodesic line in the CAT(0) metric inherited
plane in Ye with X
from Yev .
If S is a subset of a CAT(0) cube complex Ye , then let hull(S) denote the combinatorial convex hull of S. The combinatorial convex hull of S is the minimal convex
subcomplex containing S. Equivalently, hull(S) is the intersection of all closed halfspaces containing S.
Definition 5.3. Let X be a tubular space and let Y be a nonpositively curved cube
complex. A map f : X → Y is an amicable immersion if:
(1) f∗ : π1 X → π1 Y is an isomorphism.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
14
e → Ye embeds each vertex space X
eve in Ye .
(2) The G-equivariant map fe : X
e
e
(3) Each Xve has a Euclidean metric such that if H ⊆ Y is a hyperplane, then
eve is either the empty set, or a single geodesic line in
the intersection H ∩ X
eve.
X
eee is emdedded transverse to the hyperplanes.
(4) Each edge space X
eve .
eee is contained in hull S
X
(5) Each X
e
v
e∈V Γ
eve is not the subspace metric induced from
Note that the Euclidean metric on each X
e
Y.
Lemma 5.4. Let X be a tubular space and let Y be a nonpositively curved cube
complex. Let F : π1 X → π1 Y be an isomorphism. Then there is an amicable
immersion f : X → Y such that f∗ = F .
Proof. Use F to identify G = π1 X with π1 Y . The claim is proven by constructing
e → Ye . By Lemma 5.2 for each
a G-equivariant map between the tree of spaces X
e we can Gve-equivariantly embed a Euclidean flat X
eve in Ye such that if H ⊆ Ye
ve ∈ V Γ,
eve is either the empty set, or a single
is a hyperplane, then the intersection H ∩ X
S
ee is G-equivariant. The
eve. Moreover, we can ensure that
geodesic line in X
e Xv
v
e∈V Γ
eee can then be inserted transverse to the hyperplanes in Ye so that the
edges spaces X
eee ∩ X
eve with adjacent vertex spaces is not contained in a hyperplane
intersections X
eve .
eee is contained inside hull S
X
in Ye , and X
e
v∈V Γ
e
Lemma 5.5. Let X → Y be an amicable immersion, where Y is finite dimensional.
e then hull(X
eve) embeds as a subcomplex of Rd for some d.
If ve is a vertex in Γ,
eve). If H is a hyperplane in Ye ,
Proof. Let G = π1 X. Let y be a 0-cube in hull(X
let y[H] denote the halfspace of H containing y. Each 0-cube is determined by
the halfspace containing it for each hyperplane. If H is a hyperplane that doesn’t
eve, then y[H] is the halfspace containing X
eve, and therefore is y[H] fixed
intersect X
eve).
for all 0-cubes y in hull(X
eve. Let H ∈ Hve. The intersection
Let Hve denote the hyperplanes intersecting X
e
e
Xve ∩ H is a geodesic line in Xve. Let g ∈ Gve be an isometry that stabilizes an
eve that is not parallel to X
eve ∩ H. Then {gr H}r∈Z is an infinite family of
axis in X
eve}r∈Z is a set of disjoint parallel lines in X
eve. As
hyperplanes such that {gr H ∩ X
Ye is finite dimensional, there exists an N such that H and g N H do not intersect.
Otherwise {gr H}r∈Z would be an infinite set of pairwise intersecting hyperplanes,
which would imply that there are cubes of arbitrary dimension in Ye .
Therefore, as there are only finitely many Gve-orbits of hyperplanes intersecting
eve, there exists a finite set of hyperplanes H1 , . . . , Hd ∈ Hve and g1 , . . . , gd ∈ G such
X
that Hve = {g1r H1 , . . . , gdr Hd }r∈Z , and each {gir Hi }r∈Z is a disjoint set of hyperplanes
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
15
e r∈Z is a set of disjoint geodesic lines in X
eve. Thus,
in Ye . Therefore {gir Hi ∩ E}
given a 0-cube y, there exists a unique yi ∈ Z such that y[giyi Hi ] and y[giyi +1 Hi ]
eve) → Rd by letting
properly intersect each other. Therefore, construct φ : hull(X
φ(y) = (y1 , . . . , yd ) for each 0-cube y. The map φ extends to the 1-skeleton of
eve) since adjacent 0-cubes lie on the opposite sides of precisely one hyperplane.
hull(X
eve).
Therefore φ extends to the higher dimensional cubes, and thus hull(X
e → Ye be the lift of the universal cover of an amicable immersion X → Y .
Let X
eee be an edge space adjacent to a vertex space X
eve. A hyperplane H in Ye
Let X
eve parallel to X
eee if H ∩X
eve is a geodesic line parallel to X
eee∩X
eve. Otherwise,
intersects X
eve is a geodesic line that is not parallel to X
eee ∩ X
eve, then we say H intersects
if H ∩ X
eve non-parallel to X
eee.
X
e Suppose
Lemma 5.6. Let X → Y be an amicable immersion. Let ee be an edge in Γ.
e−ee non-parallel to X
eee, then H intersects
that H ⊆ Ye is a hyperplane intersecting X
e+ee non-parallel to X
eee. Moreover, there is an arc in H ∩ X
eee joining H ∩ X
e−ee to
X
e+ee .
H ∩X
e−ee and X
eee ∩ X
e−ee are non parallel
Proof. Let G = π1 X. The geodesic lines H ∩ X
e−ee , and therefore intersect in a single point p ∈ X
eee ∩ X
e−ee . As H is two sided
in X
in Ye and X the vertex and edge spaces are transverse to H, the intersection of H
e Therefore, p is contained inside a curve in
with X is also locally two sided in X.
e
e
H ∩ Xee. As Xee is Gee-invariant and only finitely many hyperplanes separate any two
e we can deduce that p is an endpoint of a compact curve in H ∩ X
eee
points in X,
e+ee ∩ X
eee. Thus, H must also intersect X
e+ee
with its other endpoint contained in X
eee.
non-parallel to X
Lemma 5.7. Let X → Y be an amicable immersion, where Y is a finite dimensional, locally finite, nonpositively curved cube complex. If π1 X ∼
= Z2 ∗Z Z2 , then
eve and adjacent edge space X
eee there is a hyperplane H in Ye
for every vertex space X
eve parallel to X
eee.
that intersects X
e
Proof. Let G = π1 X. There are precisely two vertex orbits and one edge orbit in Γ.
e Let H denote the set of all hyperplanes in Ye intersecting
Assume that Ye = hull(X).
e Let Hve denote the set of all hyperplanes intersecting X
eve.
X.
For each vertex ve, there is precisely one Gve-orbit of adjacent edges ee. Therefore,
eve, it is non-parallel to all
if H ∈ Hve is non-parallel to an adjacent edge space to X
adjacent edge spaces, and by Lemma 5.6 it must intersect all adjacent vertex spaces
non-parallel to all adjacent edge spaces. Therefore, we deduce that any hyperplane
e will either intersect a vertex space
that doesn’t intersect every vertex space in X
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
16
parallel to its adjacent edge spaces, or its intersection will be a line contained in an
edge space.
eve such that no hyperplane in Hve
Suppose that there exists a vertex space X
1
1
eve parallel to the adjacent edge spaces. Let Ĥ = H − Hve . Every
intersects X
1
1
hyperplane in H − Hve1 must intersect each wall in Hve1 so we deduce that Ye =
eve ) × C(Ye , Ĥ) (see [4, Lem 2.5]). Furthermore, Hgev = Hve for all g ∈ G. By
hull(X
1
1
1
eve) embeds in Rd . Since X
eve is contained inside some subcomplex
Lemma 5.5, hull(X
1
eve ) × C ⊆ Ye , where C is the 0-cube determined by orienting all hyperplanes
hull(X
1
eve , we can conclude that only finitely many hyperplanes in Ĥ intersect
towards X
1
eve .
the r-neighborhood of X
1
eue be an vertex space adjacent to X
eve , and let X
eee be the edge space conLet X
1
1
eve be another vertex space adjacent to X
eue and let X
eee be the
necting them. Let X
2
2
eve and X
eve are in the same Gue -orbit. The
edge space connecting them. Note that X
1
2
eue ∩ X
eee and X
eue ∩ X
eee are parallel in X
eue . Let D ⊆ X
eue be the subgeodesic lines X
1
2
eee ∪ X
eee .
space isometric to R × [a, b] bounded by these parallel lines. Let U = D ∪ X
1
2
Finitely many hyperplanes in Ĥ intersect D.
eve that is non-parallel to
Let g1 be an isometry in Gve1 that stabilizes an axis in X
1
eve ∩ X
eee . Similarly, let g2 be an isometry in Gve that stabilizes an axis
the geodesic X
1
1
2
eve that is non-parallel to the geodesic X
eve ∩ X
eee . Note that F = hg1 , g2 i is a free
in X
2
2
2
group on two generators. Let r be such that U is contained in the r-neighbourhood
eve . As there are only finitely many hyperplanes in Ĥ intersecting Nr (X
eve ),
of X
1
1
n
e
there must exist an n such that g1 stabilizes those walls. Similarly, since Xve2 is a
eve , we can deduce that there are only finitely many hyperplanes in Ĥ
translate of X
1
eve ), and there must exist an m such that gm stabilizes those walls.
intersecting Nr (X
2
2
eve ) and Nr (X
eve ) we can deduce that the
Let F ′ = hg1n , g2m i. As U lies in both Nr (X
1
2
hyperplanes in Ĥ that intersect the F ′ -translates of U are precisely the hyperplanes
e = F ′X
eve ∪ F ′ X
eve ∪ F ′ U . Then hull(Z)
e = hull(X
eve ) × K ⊆ Ye ,
intersecting U . Let Z
1
2
1
e which is a
where K is a compact cube complex. Then F acts freely on hull(Z),
contradiction since number of 0-cubes intersecting the r-neighborhood of a 0-cube
eve ) × K grows polynomially with r, and therefore cannot permit a free
in hull(X
1
F -action.
Lemma 5.7 is a special case of the following more general statement:
Corollary 5.8. Let X → Y be an amicable immersion, where Y is a finite dimensional, locally finite, nonpositively curved cube complex. Then for every vertex space
eve and adjacent edge space X
eee there is a hyperplane H in Ye that intersects X
eve
X
eee.
parallel to X
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
17
e there is a subgroup G′ = hG−ee , G+ee i ≤ G such that
Proof. For every edge ee in Γ
G′ ∼
= Z2 ∗Z Z2 . Let Y ′ = G′ \Ye . Then there is an amicable immersion X ′ → Y ′
e±ee . Therefore, by Lemma 5.7, there is a hyperplane
e and X
e′ = X
e′ = X
such that X
±e
e
ee
e−ee parallel to X
eee, and similarly for X
e+ee .
intersecting X
The following proposition is a strengthening of one direction of Theorem 1.1 in [11].
Let f1 : A → C and f2 : B → C be maps between topological spaces A, B, C. The
fiber product A ⊗C B = {(a, b) ∈ A × B | f1 (a) = f2 (b)}. Note that there are natural
projections p1 : A ⊗C B → A and p2 : A ⊗C B → B.
Proposition 5.9. Let G be a tubular group acting freely on a CAT(0) cube complex
Ye . Then there is a tubular space X with a finite set of immersed walls such that the
e W) has the following properties:
associated wallspace (X,
e W).
(1) G acts freely on C(X,
e W) is finite dimensional if Ye is finite dimensional.
(2) C(X,
e W) is finite dimensional and locally finite if Ye is locally finite.
(3) C(X,
Proof. Let Y = G\Ye . Let X → Y be an amicable immersion. Assume that
e so every immersed hyperplane in Y intersects X. Therefore, as X
Ye = hull(X),
is compact, there are finitely many immersed hyperplanes h1 , . . . hm in Y .
Let hi → Y be an immersed hyperplane in Y . We obtain horizontal immersed
walls in X by considering the components of the fiber product X ⊗Y hi of X → Y
and h → Y . Each component Λ has a natural map into X. The components of
X ⊗Y h that have image in X contained in an edge space are ignored. Let Λp be
a component of X ⊗Y h whose image in X intersects a vertex space Xv ⊆ X. We
will show that after a minor adjustment to Λp , we obtain a horizontal immersed
wall and by considering all such components we obtain a set of horizontal walls in
X obtained from an equitable set.
Using the map Λp → X we can decompose Λp into the components of the preimages of vertex space and edge spaces. As the intersection of each hyperplane H ⊆ Ye
eve is either empty or a geodesic line, the intersection of each
with each vertex space X
hi with Xv is a set of geodesic curves, so Λp restricted to the preimage of Xv is a set
of geodesic curves. By Lemma 5.6 each hyperplane H ⊆ Ye that intersects a vertex
eee will intersect X
eee as an arc with
space Xve non-parallel to an adjacent edge space X
e−ee and X
e+ee . Thus the components of the intersection Xe ∩ hi that
endpoints in X
intersect X−e or X+e , are arcs with endpoints in both X−e and X+e . Therefore,
Λp decomposes into circles that map as local geodesics into vertex spaces, and arcs
that map into edges spaces Xe with an endpoint in each X−e and X+e .
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
18
Let {Λp1 , . . . , Λpk } be the set of all such components of X ⊗Y hi that intersect
vertex spaces. Let Svp be the set curves that map the circles in {Λpi }ki=1 to the vertex
space Xv . The elements of Svp and the attaching maps ϕ±
e of the edge spaces in X
+
−
are locally geodesic curves, and #[ϕe , Se+ ] = #[ϕe , Se− ] since both sides are equal
to the number of arcs in the walls {Λpi }ki=1 that map into Xe . As G acts freely on
eee as geodesics in at
Ye , there must be hyperplanes intersecting each vertex space X
p
least two parallelism classes. This implies that Sv contains curves generating at
least two non-commensurable cyclic subgroups of Gv , and therefore Svp generates a
finite index subgroup of Gv .
Svp is almost an equitable set: the images of the curves in Svp may not be disjoint.
Suppose that α1 , . . . αm ∈ Sv be a maximal set of curves that have identical image in
Xv . Let Nǫ (Q) denote the ǫ-neighborhood of a subset Q of either Y or Ye with respect
to the CAT(0) metric. Let ǫ ∈ (0, 13 ) be such that the neighbourhood Nǫ (α1 ) ⊆ Y
only contains the images of α1 , . . . , αm and the arcs connected to them. There is
F
F
a homotopy of ki=1 Λpi → X that is the identity outside of ki=1 Λpi ∩ Nǫ (α1 ) such
that α1 , . . . , αm are homotoped to a disjoint set of geodesic curves in Xv ∩ Nǫ (α1 )
transverse or disjoint from all the other curves in Svp . By choosing ǫ small enough
F
we can perform such a homotopy Φ : ki=1 Λpi × [0, 1] → X such that all sets of
overlapping curves in {Svp }v∈V Γ become disjoint and such that Φ is the identity map
outside of the ǫ-neighborhood of the overlapping curves. The restriction of Φ to
Λpi × {1} → X is an immersed wall that we will denote by Λi . Thus, the immersed
walls {Λi }ki=1 obtained from an equitable set {Sv }v∈V Γ . We refer to {Λpi }ki=1 as the
e as the proto-walls. Note that proto-walls
ep → X
immersed proto-walls and the lifts Λ
i
have regular and non-regular intersections in the same way that walls do.
e W) be the wallspace obtained from the immersed walls {Λi }k and
Let (X,
i=1
e → X
e covers an
adding a single vertical wall for each edge space. Each wall Λ
immersed wall Λ → X. There exists a homotopy of Λ → X to the corresponding immersed proto-wall Λp → X. This homotopy lifts to a homotopy from the
e → X
e to a unique proto-wall Λ
e p → X.
e Note that each wall is
immersed wall Λ
contained in the ǫ-neighborhood of its corresponding proto-wall. Each proto-wall
e in
corresponds to the intersection of a unique hyperplane in Ye with the image of X
Ye . Therefore, each wall in W corresponds to a unique hyperplane in Ye .
e be a wall in W, and let Λ
e p be the corresponding proto-wall. Note that Λ∩
e X
eve
Let Λ
p
e ∩X
eve are either parallel geodesic lines, or both empty intersections. Therefore,
and Λ
e
e
if Λ1 , Λ2 ∈ W are a pair of regularly intersecting walls, then they correspond to a
pair of regularly intersecting proto-walls, which correspond to a pair of intersecting
hyperplanes in Ye .
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
19
e p and Λ
e p are disjoint, then the corresponding walls in Λ
e1
If a pair of proto-walls Λ
1
2
e 2 in W are also disjoint. Moreover, since Λ
e is contained in the ǫ-neighborhood
and Λ
e p , a halfspace of Λ
e determines a halfspace of Λ
e p and therefore a halfspace of the
of Λ
e p.
hyperplane H corresponding to Λ
e W) were infinite dimensional, then by ProposiTo prove (2), suppose that C(X,
tion 2.2, there would exists an infinite set of pairwise regularly intersecting walls in
W, which implies there is an infinite set of pairwise regularly intersecting proto-walls.
Therefore, there is an infinite set of pairwise intersecting hyperplanes in Ye . This
would imply that Ye is an infinite dimensional CAT(0) cube complex. Therefore, if
e W).
Ye is finite dimensional, then so is C(X,
To prove (3) we first prove the following:
e W) is finite dimensional.
Claim. If Ye is locally finite, then C(X,
e W) is infinite dimensional, then
Proof. Suppose that Ye is locally finite. If C(X,
by Lemma 2.2 it contains an infinite cube containing a canonical 0-cube z. Let
e 1, . . . , Λ
e n , . . . be the set of infinite pairwise crossing walls corresponding to the
Λ
e p, . . . , Λ
e pn , . . . be the corresponding set of infinite pairwise crossing
infinite cube. Let Λ
1
proto-walls, and let H1 , . . . , Hn , . . . be the corresponding infinite family of pairwise
crossing hyperplanes.
Suppose that Q is a subcomplex in Ye . Let U (Q) denote the cubical neighborhood
of Q, which is the union of all cubes in U (Q) that intersect Q. As Ye is locally finite,
if Q is compact, then U (Q) is also compact. By [7, Lem 13.15], if Q is convex, then
so is U (Q). Let U n (Q) denote the cubical neighborhood of U n−1 (Q).
e be a point determining the canonical 0-cube z in C(X,
e W). Let x be
Let x ∈ X
contained in a cube C in Ye . As C is compact and convex, U n (C) is also compact
and convex, and therefore can only be intersected by finitely many Hi . There
exists an Hi such that Hi intersects U N +2 (C), but not U N +1 (C) for some N > 1.
Since Hi does not intersect U N (C) nor U N +1 (C), there must exist a hyperplane
H intersecting U N +1 (C) that separates U N (C) from Hi . Note that dYe (x, H) ≥ N
e p be the proto-wall corresponding to H and Λ
e be the
and dYe (H, Hi ) ≥ 1. Let Λ
p
p
e
e
e
corresponding wall. As Λ separates x from Λi , we can conclude Λ separates x
e i since the Λ
e and Λ
e i are respectively contained in the ǫ-neighborhoods of Λ
ep
from Λ
e p . This contradicts the fact that z is incident to a 1-cube dual to hyperplane
and Λ
i
e i.
corresponding to Λ
e W) is finite dimensional, we can apply Corollary 5.8 to each edge group
As C(X,
in G to deduce that {Λi }ki=1 are fortified. Therefore, by Proposition 4.7 we deduce
e W) is locally finite.
that C(X,
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
20
We can now prove the main theorem of this paper.
Theorem 5.10. A tubular group G acts freely on a locally finite CAT(0) cube
complex if and only if G is virtually special.
Proof. Suppose that G is virtually special. Then G embeds as the subgroup of a
finitely generated right angled Artin group, and therefore acts freely on the universal
cover of the corresponding Salvetti complex, which is necessarily locally finite.
Conversely, suppose that G acts freely on a locally finite CAT(0) cube complex.
Let X be a tubular space such that G = π1 X. By Proposition 5.9 there exists a
e W)
finite set of immersed walls such that the dual of the associated wallspace C(X,
is finite dimensional and locally finite. By Lemma 3.1 we can assume that the
immersed walls are also primitive. Therefore, by Proposition 4.8, G is virtually
special.
6. Virtual Cubical Dimension
Lemma 6.1. Let X be a tubular space and G = π1 X. Suppose there exists an
equitable set that produces primitive, non-dilated immersed walls in X. There exists
a finite index subgroup G′ 6 G such that for each vertex group G′v of the induced
splitting of G′ the natural map H1 (G′v ) → H1 (G′ ) is an injection as a summand.
Proof. Note that H1 (G), and its finite index subgroups, is a summand of two factors.
The first factor Hv1 (G) is generated by the image of the vertex groups of G, and
the second factor Hv1 (G) is generated by the stable letters in the graph of groups
presentation.
e be the Bass-Serre tree. By Proposition 4.6, since there are immersed walls
Let Γ
e such that Gve fixes the
that are primitive and non-dilated, G acts freely on Rd × Γ
e Therefore G is a subgroup of Aut(Rd × Γ)
e ∼
vertex ve in Γ.
= Zd ⋊ Aut([−1, 1]d ) ×
e There is a finite quotient Zd ⋊ Aut([−1, 1]d ) × Aut(Γ)
e → Aut([−1, 1]d ),
Aut(Γ).
so let G′′ 6 G be the finite index subgroup contained in the kernel. Note that G′′
e Let p : G′′ → Zd be the projection onto the first factor.
embeds in Zd × Aut(Γ).
Each vertex group survives in the image of p, and therefore we have embedding
H1 (G′′v ) ֒→ Hv1 (G′′ ) ֒→ Zd .
e ′′ there is a finite index subgroup Av 6 Zd such
For each vertex v in Γ′′ = Γ/G
that p(G′′v ) ∩ Av is a summand of Av . Let A = ∩v Av and G′ = p−1 (A). Each vertex
group in G′ will be a factor in A. As A is free abelian, the map G′ → A will factor
through the Hv1 (G′ ), so we can deduce that each vertex group survives as a retract
in H1 (G′ ). Therefore, each vertex group in G′ survives as a summand in the first
homology.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
21
Theorem 6.2. Let G be a tubular group acting freely on a finite dimensional
CAT(0) cube complex. Then there is a finite index subgroup G′ 6 G such that
G′ acts freely on a 3-dimensional CAT(0) cube complex.
Proof. Let X be a tubular space such that G = π1 X. By Proposition 5.9 there
exists immersed walls in X that are non-dilated, and by Lemma 3.1 we can assume
that they are also primitive. Let G′ 6 G be the finite index subgroup given by
Lemma 6.1, and let X ′ be the corresponding covering space. Let Hv1 (G′ ) ∼
= Zd be
the summand in the first homology generated by the vertex groups. By Lemma 6.1,
there is an inclusion ιv : G′v → Hv1 (G′ ), and a projection map pv : Hv1 (G′ ) → G′v .
Suppose that d = 2. Choose any pair of elements a, b ∈ Hv1 (G′ ) that generate
Hv1 (G′ ). We claim that S = {Sv = {pv (a), pv (b)} | v ∈ V (Γ)} is an equitable set. By
construction, each Sv generates G′v . The edge group G′e = hge i is adjacent to G′−e
+
and G′+e and the respective inclusions are (ϕ−
e )∗ and (ϕe )∗ . There is an isomorphism
′
+
′
A = p+e ◦ ι−e ∈ SL(2, Z), that maps (ϕ−
e )∗ (ge ) in Gu to (ϕe )∗ (ge ) in Gv . Therefore,
−
+
#[pu (a), (ϕ−
e )∗ (ge )] = #[Apu (a), A(ϕe )∗ (ge )] = #[pv (a), (ϕe )∗ (ge )],
A congruent set of equalities exist for b. These equalities also imply that the choice
of arcs for the equitable sets can be chosen so that they only join circles that are
the image of the same elements of Hv1 (G′ ). Therefore a set of embedded immersed
walls is obtained with precisely two immersed walls intersecting each vertex space:
such a set of horizontal immersed walls along with a vertical wall for each edge will
give a three dimensional dual cube complex.
If d > 2, then there exist vertex groups Gu and Gv such that they embed into
v
H1 (G′ ) as distinct summands. Let Gu = hgu , gu′ i and Gv = hgv , gv′ i. We can assume,
since they are distinct summands that gu is disjoint from the image of Gv in Hv1 (G′ ),
and that gv is disjoint from the image of Gu in Hv1 (G′ ). By attaching an edge
space in X ′ connecting Xu and Xv , that have attaching maps representing gu and
gv respectively we obtain a new graph of spaces X̂ ′ . The resulting tubular group
Ĝ′ = π1 X̂ ′ has Hv1 (Ĝ′ ) ∼
= Zd−1 , so by induction we can obtain the specified graph
of spaces X̂ ′ . Given a set of immersed walls for X̂ ′ with a dual cube complex of
dimension at most 3 we obtain immersed walls for X ′ by deleting the arcs that
map into the the edge space that was added to construct X̂ ′ . The immersed walls
obtained still give a dual cube complex with dimension at most 3.
References
[1] N. Brady and M. R. Bridson. There is only one gap in the isoperimetric spectrum. Geom.
Funct. Anal., 10(5):1053–1070, 2000.
CLASSIFYING VIRTUALLY SPECIAL TUBULAR GROUPS
22
[2] Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature, volume 319
of Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical
Sciences]. Springer-Verlag, Berlin, 1999.
[3] J.O. Button. Tubular free by cyclic groups and the strongest tits alternative.
[4] Pierre-Emmanuel Caprace and Michah Sageev. Rank rigidity for CAT(0) cube complexes.
Geom. Funct. Anal., 21(4):851–891, 2011.
[5] Christopher H. Cashen. Quasi-isometries between tubular groups. Groups Geom. Dyn.,
4(3):473–516, 2010.
[6] Frédéric Haglund and Frédéric Paulin. Simplicité de groupes d’automorphismes d’espaces à
courbure négative. In The Epstein birthday schrift, volume 1 of Geom. Topol. Monogr., pages
181–248 (electronic). Geom. Topol. Publ., Coventry, 1998.
[7] Frédéric Haglund and Daniel T. Wise. Special cube complexes. Geom. Funct. Anal., 17(5):1551–
1620, 2008.
[8] Michah Sageev. Ends of group pairs and non-positively curved cube complexes. Proc. London
Math. Soc. (3), 71(3):585–617, 1995.
[9] Daniel T. Wise. Research announcement: the structure of groups with a quasiconvex hierarchy.
Electron. Res. Announc. Math. Sci., 16:44–55, 2009.
[10] Daniel T. Wise. From riches to raags: 3-manifolds, right-angled Artin groups, and cubical geometry, volume 117 of CBMS Regional Conference Series in Mathematics. Published for the
Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 2012.
[11] Daniel T. Wise. Cubular tubular groups. Trans. Amer. Math. Soc., 366(10):5503–5521, 2014.
[12] D.T. Wise and C. Hruska. Finiteness properties of cubulated groups. Submitted for publication.
[13] D.J. Woodhouse. Classifying finite dimensional cubulations of tubular groups. Submitted for
publication.
[14] D.J. Woodhouse. A generalized axis theorem for cube complexes. Submitted for publication,
2016.
E-mail address: [email protected]
| 4math.GR
|
1
Convolutional Low-Resolution Fine-Grained
Classification
arXiv:1703.05393v3 [cs.CV] 16 Oct 2017
Dingding Cai, Ke Chen, Yanlin Qian, Joni-Kristian Kämäräinen
Abstract—Successful fine-grained image classification methods
learn subtle details between visually similar (sub-)classes, but
the problem becomes significantly more challenging if the details
are missing due to low resolution. Encouraged by the recent
success of Convolutional Neural Network (CNN) architectures
in image classification, we propose a novel resolution-aware
deep model which combines convolutional image super-resolution
and convolutional fine-grained classification into a single model
in an end-to-end manner. Extensive experiments on multiple
benchmarks demonstrate that the proposed model consistently
performs better than conventional convolutional networks on
classifying fine-grained object classes in low-resolution images.
Index Terms—Fine-Grained Image classification, Super Resolution Convoluational Neural Networks, Deep Learning
The problem of image classification is to categorise images
according to their semantic content (e.g. person, plane). Finegrained image classification further divides classes to their
“sub-categories” such as the models of cars [1], the species
of birds [2], the categories of flowers [3] and the breeds of
dogs [4]. Fine-grained categorisation is a difficult task due
to small inter-class variance between visually similar subclasses. The problem becomes even more challenging when
available images are low-resolution (LR) images where many
details are missing as compared to their high-resolution (HR)
counterparts.
Since the rise of Convolutional Neural Network (CNN)
architectures in image classification [5], the accuracy of finegrained image classification has dramatically improved and
many CNN-based extensions have been proposed [6], [7], [8],
[9], [10], [11]. However, these works assume sufficiently good
image quality and high resolution, (e.g. typically 227 × 227
for AlexNet [5]) while with low resolution images the CNN
performance quickly collapses [12], [13]. The challenge raises
from the problem of how to recover necessary texture details
from low-resolution images. Our solution is to adopt image
super-resolution (SR) techniques [14], [15], [16], [17], [18]
to enrich imagery details. In particular, inspired by the recent
work on CNN-based image SR by Deng et al. [19] we propose
a unique end-to-end deep learning framework that combines
CNN super-resolution and CNN fine-grained classification – a
resolution-aware Convolutional Neural Network (RACNN) for
fine-grained object categorisation in low-resolution images. To
our best knowledge, our work is the first end-to-end learning
model for low-resolution fine-grained object classification.
Our main principle is simple: the higher image resolution,
the easier for classification. Our research questions are: Can
computational super-resolution recover some of the important
details required for fine-grained image classification and can
Fig. 1: Owing to the introduction of the convolutional superresolution (SR) layers, the proposed deep convolutional model
(the bottom pipelines) achieves superior performance for low
resolution images.
such SR layers be added to an end-to-end deep classification
architecture? To this end, our RACNN integrates deep residual
learning for image super-resolution [20] into typical convolutional classification networks (e.g. AlexNet [5] or VGGNet [21]). On one hand, the proposed RACNN has deeper
network architecture (i.e. more network parameters) than the
straightforward solution of conventional CNN on up-scaled
images (e.g. bicubic interpolation [22]). Our RACNN learns
to refine and provide more texture details for low-resolution
images to boost fine-grained classification performance. We
conduct experiments on three benchmarks, Stanford Cars [1],
Caltech-UCSD Birds-200-2011 [2] and Oxford 102 Flower
Dataset [3]. Our results answer the aforementioned questions:
super-resolution improves fine-grained classification and SRbased fine-grained classification can be designed into a supervised end-to-end learning framework, as depicted in Figure 1
illustrating the difference between RACNN and conventional
CNN.
I. R ELATED W ORK
Fine-Grained Image Categorisation – Recent algorithms
for discriminating fine-grained classes (such as animal species
2
or plants [2], [4], [23] and man-made objects [1], [24], [25])
can be divided into two main groups. The first group of
methods utilises discriminative visual cues from local parts
obtained by detection [6], [26] or segmentation [8], [27], [28].
The second group of methods focuses on discovering interclass label dependency via pre-defined hierarchical structure of
labels [10], [29], [30], [31], [32] or manually-annotated visual
attributes [33], [34]. Significant performance improvement is
achieved by convolutional neural networks (CNNs), but this
requires a massive amount of high quality training images.
Fine-grained classification from low-resolution images is yet
challenging and unexplored. The method proposed by Peng et
al. [35] transforms detailed texture information in HR images
to LR via fine-tuning to boost the accuracy of recognizing finegrained objects in LR images. However, in [35], their strong
assumption requiring HR images available for training limits
its generalisation ability. In addtion, the same assumption also
occurs in Wang’s work [36]. Chevalier et al. [12] design
a CNN-based fine-grained object classifier with respect to
varying image resolutions, which adopts ordinary convolutional and fully-connected layers but misses considering superresolution specific layers in convolutional classification networks. On contrary, owing to the introduction of SR-specific
layers in RACNN, our method can consistently gain notable
performance improvement over conventional CNN for image
classification on fine-grained classification datasets.
Convolutional Super-Resolution Layers – Yang et al. [37]
grouped existing SR algorithms into four groups: prediction models [38], edge-based methods [39], image statistical
methods [40] and example-based methods [14], [17], [18],
[41], [42], [43], [44], [45]. Recently, Convolutional Neural Networks have been adopted for image super-resolution
achieving state-of-the-art performance. The first attempt using
convolutional neural networks for image super-resolution was
proposed by Dong et al. [18]. Their method learns a deep
mapping between low- and high- resolution patches and has
inspired a number of follow-ups [20], [46], [47]. In [47], an
additional deconvolution layer is added based on SRCNN [18]
to avoid general up-scaling of input patches for accelerating
CNN training and testing. Kim et al. [46] adopt a deep
recursive layer to avoid adding weighting layers, which does
not need to pay any price of increasing network parameters.
In [20], a convolutional deep network is proposed to learn
the mapping between LR image and its residue between
LR and HR image to speed up CNN training for very
deep network. Convolutional layers designed for image superresolution (namely SR-specific convolutional layers) have been
verified their effectiveness to improve the quality of images.
In this work, we incorporate the state-of-the-art residual CNN
layers for image super-resolution [20] into a convolutional
categorisation network for classifying fine-grained objects (i.e.
AlexNet [5], VGG-Net [21] and GoogLeNet [48]). In the
experiments, SR-specific convolutional layers are verified to
improve classification performance.
Contributions – Our contributions are two-fold:
• Our work is the first attempt to utilise super-resolution
specific convolutional layers to improve convolutional
•
fine-grained image classification.
We experimentally verify that the proposed RACNN
achieves superior performance on low-resolution images
which make ordinary CNN performance collapse.
II. R ESOLUTION -AWARE C ONVOLUTIONAL N EURAL
N ETWORKS
Given a set of N training images and corresponding class
X i , yi }, i = 1, 2, · · · , N, the goal of a conventional CNN
labels {X
X ). The typical
model is to learn a mapping function y = f (X
cross entropy (ce) loss Lce (·) on softmax classifer is adopted
X)
to measure the performance between class estimates ŷ = f (X
and ground truth class labels y :
l
Lce (ŷ, y) = − ∑ y j log(ŷ j ),
(1)
j=1
where j refers to the index of element in vectors, and l denotes
the dimension of softmax layer (i.e. the number of classes). In
this sense, CNN solves the following minimisation problem
with gradient descent back propagation:
N
min
∑ Lce ( f (XX i ), yi ).
(2)
i=1
For fine-grained categorisation in low-quality images, we propose a novel resolution-aware convolutional neural network,
which is illustrated in Fig. 2. In general, our RACNN consists
of two parts: convolutional super-resolution layers (see Sec.
II-A) and convolutional categorisation layers (see Sec. II-B).
In Sec. II-C, we describe an end-to-end training scheme for
the proposed RACNN.
A. Convolutional Super-Resolution Layers
In this section, we present convolutional super-resolution
specific layers for the resolution-aware CNN, the goal of which
is to recover texture details of low-resolution images to feed
into the following convolutional categorisation layers.
We first investigate the conventional CNN for the superresolution task. Given K training pairs of low-resolution and
X LR , X HR }i , i = 1, 2, · · · , K, a direct
high-resolution images {X
X LR ) from X LR (input obCNN-based mapping function g(X
HR
servation) to X
(output target) [18], [47] is learned by
minimising the mean square (ms) loss
X LR , X HR ) =
Lms (X
1 K
∑ kXX HR − g(XX LR )k2 .
2 i=1
(3)
Inspired by recent the state-of-the-art residual convolutional
network [20] to achieve high efficacy, we design convolutional
super-resolution layers as shown on the left hand side of Fig. 2.
Similar to [20], our convolutional super-resolution layers learn
a mapping function from LR images X LR to residual images
X HR − X LR . Object function of the proposed convolutional
super-resolution layers is as the following:
min
1 K
∑ kXX HR − X LR − g(XX LR )k2 .
2 i=1
(4)
3
Fig. 2: Pipeline of the proposed resolution-aware convolutional neural network (RACNN) for fine-grained recognition with
low-resolution images. Convolutional classification layers from AlexNet are adopted for illustrative purpose, which can be
readily replaced by those from other CNNs such as VGG-Net or GoogLeNet
.
The better performance of residual learning yields from the
fact that, since the input (LR) and output images (HR) are
largely similar, it is more meaningful to learn their residue
where similarities are removed. It is obvious that detailed
imagery information in the form of residual images is easier
for CNNs to learn than direct LR-HR CNN models [18], [47].
We utilise three typical stacked convolutional-ReLU layers with zero-padding filters as convolutional SR layers in
RACNN. Following [18], the empirical basic setting of the
layers is f1 = 9×9, n1 = 64, f2 = 5×5, n2 = 32, f3 = 5×5 and
n3 = 3, which are also illustrated in the left hand side of Fig. 2,
where fm and nm donate the size and number of the filters of
the mth layer respectively. The output of the last convolutional
SR layer is summed with the low-resolution input image
X LR to construct the full super-resolution image fed into
the remaining convolutional and fully-connected classification
layers of RACNN.
B. Categorisation Layers
The second part in our RACNN is convolutional and fullyconnected classification layers with high quality images after
super-resolution layers. A number of CNN frameworks [5],
[21], [49], [48] have been proposed for image categorisation, and in this paper we consider three popular convolutional neural networks: AlexNet [5], VGG-Net [21] and
GoogLeNet [48]. All CNNs typically consist of a number
of Convolutional-ReLU-Pool stacks followed by several fullyconnected layers. On the right-hand-side of Fig. 2, the typical
AlexNet [5] is visualised and employed as convolutional
categorisation layers in RACNN. AlexNet [5], the baseline
CNN for large-scale image classification over ImageNet [50],
consists of 5 convolutional layers (i.e. conv1, conv2, conv3,
conv4, and conv5) and 3 fully-connected layers (i.e. fc6,
fc7, and fc8). VGGNet [21] is made deeper (i.e. from 8
layers of Alexnet to 16-19 layers) and more advanced over
AlexNet by using very small (e.g. 3 × 3) convolution filters.
In our paper, we choose the VGG-Net-16 with 16 layers
for our experiments (denoted as VGG-Net in the rest of the
paper). GoogLeNet [48] comprises 22 layers but has much
less number of parameters than AlexNet and VGG-Net owing
to the smaller amount of weights of fully-connected layers.
GoogLeNet generally generates three outputs at various depths
for each input, but for simplicity only the last output (i.e.
the deepest output) is considered in our experiments. In our
experiments, all three networks are pre-trained on the Imagenet
X , y} from fine-grained data as the
data and fine-tuned with {X
baseline. For fair comparison, we fine-tune the identical pretrained CNN models as our convolutional categorisation layers
X , y} by replacing the dimension of final fully-connected
with {X
layer with the size of object classes l.
C. Network Training
The key difference between the proposed resolution-aware
CNN and conventional CNN lies in the introduction of three
convolutional super-resolution layers. Evidently, RACNN is
deeper than corresponding CNN due to the three convolutional SR layers, which can store more knowledge, i.e.
network parameters. Before learning RACNN in an end-toend fashion, we consider two weight initialization strategies
for convolutional SR layers in RACNN, i.e. standard Gaussian
weights and pre-trained weights on the ImageNet data. For fair
comparison, we adopt the identical network structure for both
initialisation schemes.
For RACNN with Gaussian initial weights, we train the
whole network to minimise cross-entropy loss (2) directly.
During training, we set learning rates 1 and weight decays
0.1 for the first two SR layers (sconv1 and sconv2) and both
learning rate and weight decay are set with 0.1 for the third
convolutional SR layer (sconv3), while learning rates and
weight decays are 0.1 and 0 for all categorisation layers except
the last fully-connected layer which uses both learning rate and
weight decay 1.
We consider an alternative initialisation strategy for better
initial weights for convolutional SR layers. To this end, we
4
Fig. 3: Low-resolution image samples after removing background from the Stanford Cars and UCSD-Caltech Birds 2002011 benchmarks.
pre-train the three convolutional SR layers by enforcing the
minimal of the mean square loss (4) on ILSVRC 2015 ImageNet object detection testing dataset [51], which consists of
11,142 high-resolution images. Given the pre-trained weights
in convolutional SR layers, RACNN is end-to-end trained by
minimising the loss function (2) for categorisation. For the
goal of direct utilisation of output of convolutional SR layers,
we train SR layers in RGB color space with all the channels,
instead of only on luminance channel Y in YCbCr color space
[18]. Specifically, we generate LR images from HR images
(e.g. 227×227 pixels) via firstly down-sampling HR images
to 50×50 pixels and then up-scaling to the original image size
by bicubic interpolation [22]. We then sample image patches
using sliding window and thus obtain thousands of pairs of LR
and HR image patches. To be consistent with the setting of
RACNN using Guassian initial weights, the super-resolution
layers are trained with image patches by setting learning rates
being 1 and weight decays being 0.1 for the first two SR
layers (sconv1 and sconv2) and both learning rate and weight
decay being 0.1 for the third SR layer (sconv3). Finally, we
jointly learn both convolutional SR and classification layers
in an end-to-end learning manner with learning rates 0.1 and
weight decays 0 for all classification layers except the last
fully-connected layer with both learning rate and weight decay
set to 1.
III. E XPERIMENTS
A. Datasets and Settings
We evaluate RACNN on three commonly-used datasets: the
Stanford Cars [1], the Caltech-UCSD Birds-200-2011 [2] and
the Oxford 102 Category Flower [3] datasets. The first one
was released by Krause et al. for fine-grained categorisation
and contains 16,185 images from 196 classes of cars and each
class is typically at the level of Brand, Model and Year. By
following the standard evaluation protocol [1], we split the
data into 8,144 images for training and 8,041 for testing.
Caltech-UCSD Birds-200-2011 is another challenging finegrained image dataset aimed at subordinate category classification by providing a comprehensive set of benchmarks
and annotation types for the domain of birds. The dataset
contains 11,788 images of 200 bird species, among which
there are 5,994 images for training and 5,794 for testing [2].
Oxford 102 Category Flower Dataset consists of 8,189 images
which commonly appear in the United Kingdom. These images
belong to 102 categories and each category contains between
40 to 258 images. In the standard evaluation protocol [3], the
whole dataset is divided into 1,020 images for training, 1,020
for validation and 6,149 for testing. In our experiments the
training and validation data are merged together to train the
networks.
Images from these datasets are first cropped with provided
bounding boxes to remove the background. Cropped images
are down-sampled to LR images of the size 50 × 50 pixels and
then up-scaled to 227 × 227 pixels by bicubic interpolation
[22] to fit the conventional CNN, which follows the settings
in [35]. Sample LR images from the both benchmarks are
illustrated in Fig. 3, which verify our motivation to mitigate the
suffering from low visual discrimination due to low-resolution.
We compare our RACNN with multiple state-of-the-art methods, the corresponding CNN model for classification (i.e.
AlexNet [5], VGG-Net [21] and GoogLeNet [48]) and StagedTraining CNN proposed by [35]. The proposed RACNN is
implemented on Caffe [52]. We adopt the average per-class
accuracy [13], [35] for the both datasets (the higher value
denotes the better performance).
In our experiments, we used a Lenovo Y900 desktop with
one Intel i7-6700K CPU and one Nvidia GTX-980 GPU. The
proposed RACNN has deeper structure than the competing
networks (i.e. AlexNet, VGGNet, GoogLeNet) which requires
longer training times as indicated in Table I.
TABLE I: Training times of RACNNs and competing CNNs
(seconds / epoch)
.
Methods
Cars [1]
Birds [2]
Flowers [3]
AlexNet
RACNNAlexNet
11
111
8
80
3
25
VGGNet
RACNNVGGNet
133
356
82
215
34
90
GoogLeNet
RACNNGoogLeNet
20
136
25
120
8
59
B. Comparative Evaluation
In Fig. 4, we compare our results with AlexNet [5] and
Staged-Training AlexNet [35] for fine-grained classification
in low-resolution images. It is evident that our RACNNAlexNet
consistently achieves the best performance on both benchmarks. Precisely, AlexNet achieves 50.4% and 51.3% ac-
5
C. Evaluation of Convolutional SR Layers
Fig. 4: Comparison to two state-of-the-art methods for classification (average per-class accuracies).
TABLE II: Evaluation on effect of convolutional SR layers
to recover high resolution details. We fix all convolutional
and fully-connected layers except the last fully-connected layer
(i.e. extracted features correspond to those with high resolution images). g-RACNN and p-RACNN denote the proposed
RACNN with weights initialized with Gaussian (g-) and pretrained weights (p-) for the convolutional SR layers.
Methods
Cars [1]
Birds [2]
Flowers [3]
AlexNet [5]
g-RACNNAlexNet
p-RACNNAlexNet
43.75%
45.77%
47.90%
44.99%
47.17%
51.23%
70.03%
71.91%
74.24%
VGG-Net [21]
g-RACNNVGG-Net
p-RACNNVGG-Net
41.49%
42.86%
44.65%
43.46%
44.72%
49.33%
67.82%
68.03%
69.17%
GoogLeNet [48]
g-RACNNGoogLeNet
p-RACNNGoogLeNet
46.85%
50.37%
50.76%
48.52%
55.16%
57.30%
69.28%
69.77%
73.51%
In this experiment, we employ all layers in the AlexNet,
VGG-Net and GoogLeNet [48] as categorisation layers in
RACNN. Note that, different from the previous experiments,
we freeze all categorisation layers by setting learning rates and
weights decays to 0 besides the last fully-connected layers
of the baseline CNNs, and our RACNN is then fine-tuned
with low-resolution data. Such setting treats categorisation
layers in RACNN as an identical classifier for evaluating
the effect of adding convolutional SR layers. RACNN with
initial Gaussian and pre-trained weights are called as gRACNN and p-RACNN respectively. Comparative results are
shown in Table II and Fig. 5. Both g-RACNN and p-RACNN
consistently outperform the baseline CNNs in all experiments.
With the same experimental setting except different initial
weights for convolutional super-resolution layers, the results
of g-RACNN and p-RACNN are reported. Test set accuracies
in Table II and Fig. 5 show that p-RACNN is superior to gRACNN. p-RACNN and g-RACNN share the same network
structure but differ only in network weights initialisation of
convolutional SR layers. In this sense, better performance of
p-RACNN is credited to the knowledge about refining lowresolution images (i.e. pre-trained weights), which verifies our
motivation to boost low-resolution image classification via
image super-resolution. It is noteworthy that since the feature
extraction layers are frozen, the networks are not fine-tuned
to low-resolution specific features, but all performance boost
are owing to recovered high-resolution details important for
classifcation by the super-resolution layers.
D. Evaluation on Varying Resolution
TABLE III: Comparison with varying resolution level (Res.
Level) on the Caltech-UCSD Birds 200-2011 Dataset.
Res. Level
25×25
50×50
100×100
curacies (collected from [35]) for the Stanford Cars and
Caltech-UCSD Birds datasets, respectively. Knowledge transfer between varying resolution images (i.e. Staged-Training
AlexNet [35]) can improve classification accuracy, that is
59.5% for the Stanford Cars and 55.3% for the Caltech-UCSD
Birds. However, the staged-training AlexNet [35] relies on the
strong assumption that high-resolution images are available
for training, which limits to its usage to other tasks. Note
that our method is more generic and transforms knowledge
of super resolution across datasets, which indicates that our
method can be readily applied to other low-resolution image
classification tasks. The proposed RACNNAlexNet significantly
beats its direct competitor AlexNet, i.e. 63.8% vs. 50.4% on
the Stanford Cars dataset and 58.1% vs. 51.3% on the CaltechUCSD Birds dataset. With the same settings and training
samples, the performance gap can only be explained by the
novel network structure of RACNN.
AlexNet [5]
g-RACNNAlexNet
p-RACNNAlexNet
31.58%
44.99%
51.01%
43.68%
47.17%
51.24%
45.06%
51.23%
52.88%
We further evaluate our proposed RACNN method with
respect to varying resolutions on the Caltech-UCSD Birds 2002011 Dataset. All low-resolution images are first up-scaled to
the input image size, i.e. 227×227, before training models.
The better performance of RACNNAlexNet over conventional
AlexNet is achieved for cross-resolution fine-grained image
classification, which is shown in Table III.
We observe that our method performs much better for lower
resolution images (e.g. 25×25) than relatively high resolution
images (e.g. 100×100). In details, p-RACNNAlexNet increases
the accuracy by above 13% for 25×25 pixel images but less
than 2% improvement on 100×100 resolution images. The
reason is that the SR layers of RACNN play a significant role
in introducing texture details especially when missing more
visual cues of object classification in lower quality images,
which further demonstrates our observation and motivation.
6
(a) AlexNet
(b) VGG-Net
(c) GoogLeNet
Fig. 5: Training process of AlexNet, VGGNet and GoogLeNet on the Caltech-UCSD Birds Dataset.
In p-RACNNAlexNet , the weights for convolutional SR layers
are pre-trained only with 50×50 resolution-level ImageNet
images, but our RACNN is applied to varying resolution levels
(i.e. 25×25 and 100×100). Further improvement on classification performance shows the generalisation of pre-trained
weights for varying resolution levels, which demonstrates the
generalisation ability of RACNN with pre-trained SR weights.
IV. C ONCLUSION
We propose and verify a simple yet effective resolutionaware convolutional neural network (RACNN) for fine-grained
image classification of low-resolution images. The results
from extensive experiments indicate that the introduction of
convolutional super-resolution layers to conventional CNNs
can indeed recover fine details for low-resolution images
and clearly boost performance in low-resolution fine-grained
classification. This result can be explained by the fact that the
super-resolution layers learn to recover high resolution details
that are important for classification when trained end-to-end
manner together with the classification layers. The concept
of our paper is generic and the existing convolutional superresolution and classification networks can be readily combined
to cope with low-resolution image classification.
R EFERENCES
[1] J. Krause, M. Stark, J. Deng, L. Fei-Fei, 3d object representations for
fine-grained categorization, in: International Conference on Computer
Vision Workshops, 2013, pp. 554–561.
[2] C. Wah, S. Branson, P. Welinder, P. Perona, S. Belongie, The caltechucsd birds-200-2011 dataset.
[3] M.-E. Nilsback, A. Zisserman, Automated flower classification over a
large number of classes, in: Indian Conference on Computer Vision,
Graphics & Image Processing, IEEE, 2008, pp. 722–729.
[4] A. Khosla, N. Jayadevaprakash, B. Yao, F.-F. Li, Novel dataset for finegrained image categorization: Stanford dogs.
[5] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with
deep convolutional neural networks, in: Advance in Neural Information
Processing Systems, 2012.
[6] N. Zhang, J. Donahue, R. Girshick, T. Darrell, Part-based R-CNNs for
fine-grained category detection, in: European Conference on Computer
Vision, 2014.
[7] T.-Y. Lin, A. RoyChowdhury, S. Maji, Bilinear cnn models for finegrained visual recognition, in: IEEE International Conference on Computer Vision, 2015, pp. 1449–1457.
[8] J. Krause, H. Jin, J. Yang, L. Fei-Fei, Fine-grained recognition without
part annotations, in: IEEE Conference on Computer Vision and Pattern
Recognition, 2015, pp. 5546–5555.
[9] K. Chen, Z. Zhang, Learning to classify fine-grained categories with
privileged visual-semantic misalignment, IEEE Transactions on Big
Data.
[10] Z. Akata, S. Reed, D. Walter, H. Lee, B. Schiele, Evaluation of output
embeddings for fine-grained image classification, in: IEEE Conference
on Computer Vision and Pattern Recognition, 2015.
[11] S. Branson, G. V. Horn, S. Belongie, P. Perona, Bird species categorization using pose normalized deep convolutional nets, in: British Machine
Vision Conference, 2014.
[12] M. Chevalier, N. Thome, M. Cord, J. Fournier, G. Henaff, E. Dusch,
LR-CNN for fine-grained classification with varying resolution, in: IEEE
International Conference of Image Processing, 2015, pp. 3101–3105.
[13] Y. Liu, Y. Qian, K. Chen, J.-K. Kämäräinen, H. Huttunen, L. Fan,
J. Saarinen, Incremental convolutional neural network training, in:
International Conference of Pattern Recognition Workshop on Deep
Learning for Pattern Recognition, 2016.
[14] R. Zeyde, M. Elad, M. Protter, On single image scale-up using sparserepresentations, in: International Conference on Curves and Surfaces,
2010, pp. 711–730.
[15] J. Yang, J. Wright, T. S. Huang, Y. Ma, Image super-resolution via sparse
representation, IEEE transactions on image processing 19 (11) (2010)
2861–2873.
[16] H. Chang, D.-Y. Yeung, Y. Xiong, Super-resolution through neighbor
embedding, in: IEEE Conference on Computer Vision and Pattern
Recognition, 2004.
[17] D. Glasner, S. Bagon, M. Irani, Super-resolution from a single image, in:
IEEE International Conference on Computer Vision, 2009, pp. 349–356.
[18] C. Dong, C. C. Loy, K. He, X. Tang, Image super-resolution using
deep convolutional networks, IEEE transactions on Pattern Analysis and
Machine Intelligence 38 (2) (2016) 295–307.
[19] D. Dai, Y. Wang, Y. Chen, L. Van Gool, Is image super-resolution helpful
for other vision tasks?, in: IEEE Winter Conference on Applications of
Computer Vision, 2016, pp. 1–9.
[20] J. Kim, J. Kwon Lee, K. Mu Lee, Accurate image super-resolution using
very deep convolutional networks, in: IEEE Conference on Computer
Vision and Pattern Recognition, 2016, pp. 1646–1654.
[21] K. Simonyan, A. Zisserman, Very deep convolutional networks for largescale image recognition.
[22] R. Keys, Cubic convolution interpolation for digital image processing,
IEEE transactions on acoustics, speech, and signal processing 29 (6)
(1981) 1153–1160.
[23] A. Angelova, S. Zhu, Efficient object detection and segmentation for
fine-grained recognition, in: IEEE Conference on Computer Vision and
Pattern Recognition, 2013, pp. 811–818.
[24] M. Stark, J. Krause, B. Pepik, D. Meger, J. J. Little, B. Schiele, D. Koller,
Fine-grained categorization for 3d scene understanding, International
Journal of Robotics Research (2011) 1543–1552.
[25] S. Maji, E. Rahtu, J. Kannala, M. Blaschko, A. Vedaldi, Fine-grained
visual classification of aircraft, arXiv preprint arXiv:1306.5151.
[26] N. Zhang, R. Farrell, F. Iandola, T. Darrell, Deformable part descriptors
for fine-grained recognition and attribute prediction, in: International
Conference on Computer Vision, 2013.
[27] Y. Chai, V. Lempitsky, A. Zisserman, Symbiotic segmentation and part
localization for fine-grained categorization, in: International Conference
on Computer Vision, 2013.
7
[28] E. Gavves, B. Fernando, C. G. Snoek, A. W. Smeulders, T. Tuytelaars,
Local alignments for fine-grained categorization, International Journal
of Computer Vision 111 (2) (2015) 191–212.
[29] J. Shotton, M. Johnson, R. Cipolla, Semantic texton forests for image
categorization and segmentation, in: IEEE Conference on Computer
Vision and Pattern Recognition, 2008.
[30] S. J. Hwang, K. Grauman, F. Sha, Semantic kernel forests from multiple
taxonomies, in: Advances in Neural Information Processing Systems,
2012.
[31] A. Mittal, M. B. Blaschko, A. Zisserman, P. H. Torr, Taxonomic multiclass prediction and person layout using efficient structured ranking, in:
European Conference on Computer Vision, 2012.
[32] J. Deng, N. Ding, Y. Jia, A. Frome, K. Murphy, S. Bengio, Y. Li,
H. Neven, H. Adam, Large-scale object classification using label relation
graphs, in: European Conference on Computer Vision, 2014.
[33] N. Zhang, M. Paluri, M. Ranzato, T. Darrell, L. Bourdev, Panda: Pose
aligned networks for deep attribute modeling, in: IEEE Conference on
Computer Vision and Pattern Recognition, 2014.
[34] Y. Fu, T. M. Hospedales, T. Xiang, Z. Fu, S. Gong, Transductive multiview embedding for zero-shot recognition and annotation, in: European
Conference on Computer Vision, 2014.
[35] X. Peng, J. Hoffman, X. Y. Stella, K. Saenko, Fine-to-coarse knowledge
transfer for low-res image classification, in: IEEE International Conference of Image Processing, 2016, pp. 3683–3687.
[36] Z. Wang, S. Chang, Y. Yang, D. Liu, T. S. Huang, Studying very low
resolution recognition using deep networks, in: Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2016, pp.
4792–4800.
[37] C.-Y. Yang, C. Ma, M.-H. Yang, Single-image super-resolution: a
benchmark, in: European Conference on Computer Vision, 2014, pp.
372–386.
[38] M. Irani, S. Peleg, Improving resolution by image registration, Computer
Vision, Graphics, and Image Processing: Graphical models and image
processing (1991) 231–239.
[39] R. Fattal, Image upsampling via imposed edge statistics, in: ACM
Transactions on Graphics, Vol. 26, 2007, p. 95.
[40] J. Huang, D. Mumford, Statistics of natural images and models, in:
IEEE Conference on Computer Vision and Pattern Recognition, 1999,
pp. 541–547.
[41] J.-B. Huang, A. Singh, N. Ahuja, Single image super-resolution from
transformed self-exemplars, in: IEEE Conference on Computer Vision
and Pattern Recognition, 2015, pp. 5197–5206.
[42] J. Yang, Z. Lin, S. Cohen, Fast image super-resolution based on inplace example regression, in: IEEE Conference on Computer Vision
and Pattern Recognition, 2013, pp. 1059–1066.
[43] G. Freedman, R. Fattal, Image and video upscaling from local selfexamples, ACM Transactions on Graphics 30 (2) (2011) 12.
[44] D. Dai, R. Timofte, L. Van Gool, Jointly optimized regressors for image
super-resolution, in: Computer Graphics Forum, Vol. 34, 2015, pp. 95–
104.
[45] S. Schulter, C. Leistner, H. Bischof, Fast and accurate image upscaling
with super-resolution forests, in: IEEE Conference on Computer Vision
and Pattern Recognition, 2015, pp. 3791–3799.
[46] J. Kim, J. Kwon Lee, K. Mu Lee, Deeply-recursive convolutional
network for image super-resolution, in: IEEE Conference on Computer
Vision and Pattern Recognition, 2016, pp. 1637–1645.
[47] C. Dong, C. C. Loy, X. Tang, Accelerating the super-resolution convolutional neural network, in: European Conference on Computer Vision,
2016, pp. 391–407.
[48] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in:
Proceedings of the IEEE conference on computer vision and pattern
recognition, 2015, pp. 1–9.
[49] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image
recognition, in: IEEE Conference on Computer Vision and Pattern
Recognition, 2016, pp. 770–778.
[50] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet:
A large-scale hierarchical image database, in: IEEE Conference on
Computer Vision and Pattern Recognition, 2009, pp. 248–255.
[51] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. FeiFei, ImageNet Large Scale Visual Recognition Challenge, International
Journal of Computer Vision 115 (3) (2015) 211–252.
[52] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick,
S. Guadarrama, T. Darrell, Caffe: Convolutional architecture for fast
feature embedding, in: ACM international conference on Multimedia,
2014, pp. 675–678.
| 1cs.CV
|
Preprint, Institute of Statistics, RWTH Aachen University
ASYMPTOTICS FOR HIGH–DIMENSIONAL COVARIANCE MATRICES AND
QUADRATIC FORMS WITH APPLICATIONS TO THE TRACE FUNCTIONAL AND
SHRINKAGE
arXiv:1711.01835v1 [math.PR] 6 Nov 2017
Ansgar Steland∗ and Rainer von Sachs
Institute of Statistics
RWTH Aachen University
Wüllnerstr. 3, D-52056 Aachen, Germany
[email protected]
and
Institut de Statistique, Biostatistique et Sciences Actuarielles (ISBA)
Université catholique de Louvain
Voie du Roman Pays 20, B-1348 Louvain-la-Neuve, Belgium
[email protected]
July 10, 2017
We establish large sample approximations for an arbitray number of bilinear
forms of the sample variance-covariance matrix of a high-dimensional vector
time series using ℓ1 -bounded weighting vectors. Estimation of the asymptotic
covariance structure is also discussed. The results hold true without any constraint on the dimension, the number of forms and the sample size or their
ratios. Concrete and potential applications are widespread and cover highdimensional data science problems such as projections onto sparse principal
components or more general spanning sets as frequently considered, e.g. in
classification and dictionary learning. As two specific applications of our results, we study in greater detail the asymptotics of the trace functional and
shrinkage estimation of the covariance matrices. In shrinkage estimation, it
turns out that the asymptotics differs for weighting vectors bounded away
from orthogonaliy and nearly orthogonal ones in the sense that their inner
product converges to 0.
AMS 2000 subject classifications: Primary 60F17; secondary 62E20 .
Keywords: Brownian motion, Linear process, Long memory, Strong approximation, Quadratic form, Trace.
1. INTRODUCTION
A large number of procedures studied to analyze high–dimensional vector time series of
dimension dn depending on the sample size n relies on projections, e.g. by projecting the
observed random vector onto a spanning set of a lower dimensional subspace of dimension
Ln . Examples include sparse principal component analysis, see e.g. [20], in order to reduce
dimensionality of data, sparse portfolio replication and index tracking as studied by [4], or
dictionary learning, see [1], where one aims at representing input data by a sparse linear
combination of the elements of a dictionary, frequently obtained as the union of several bases
and/or historical data.
b n wn ,
When studying projections, it is natural to study the associated bilinear form vn′ Σ
d
vn , wn ∈ R n , representing the dependence structure in terms of the projections’ covariances.
b n is the (uncentered) sample variance–covariance matrix.
Here and throughout the paper Σ
In order to conduct inference, large sample distributional approximations are needed. For a
vector time series model given by correlated linear processes, we established in [17] a strong
1
2
A. STELAND, R. VON SACHS
approximation by a Brownian motion for a single quadratic form, provided the weighting
vectors are uniformly bounded in the ℓ1 -norm. It turned out that the result does not require
any condition on the ratio of the dimension and the sample size, contrary to many asymptotic
results in highdimensional statistics and probability.
In the present article, we study the more general case of an increasing number of quadratic
forms as arising when projecting onto a sequence of subspaces whose dimension converges to
∞. Noting that the analysis of autocovariances of a stationary linear time series appears as a
special case of our approach, there are a few recent results related to our work: [23] established
a central limit theorem for a finite number of autocovariances, whereas in [22] the case of long
memory series has been studied. [8] has studied the asymptotic theory for detecting a change
in mean of a vector time series with growing dimension.
To treat the case of an increasing number of bilinear forms, we consider two related but
different frameworks: The first framework uses a sequence of Euclidean spaces Rdn equipped
with the usual Euclidean norm. The second framework embeds those spaces in the sequence
space ℓ2 equipped with the ℓ2 –norm. It is shown that, in both frameworks, an increasing
number of, say Ln , quadratic forms can be approximated by Brownian motions without any
constraints on Ln , dn and n apart from n → ∞. One of our main results asserts that, for the
assumed time series models, one can define, on a new probability space, equivalent versions
and a Gaussian process Gn taking values in C([0, 1], RLn ), such that
Ln
1
b ⌊nt⌋ − E Σ
b ⌊nt⌋ )wn(j)
− Gn (t) = oP (1),
vn(j) ′ (Σ
sup √
j=1
nLn
t∈[0,1]
as n → ∞, almost surely (a.s.), without any constraints on Ln , dn .
We believe that those results have many applications in diverse areas, as indicated above.
In this paper we study in some detail two direct applications: The first application considers
the trace operator, which equals the trace matrix norm k · ktr when applied to covariance
matrices. We show that the trace of the sample covariance matrix, appropriately centered,
can be approximated by a Brownian motion, a.s. on a new probability space, which also
establishes the convergence rate
b n ktr − kE Σ
b n ktr = OP (n−1/2 dn ).
kΣ
The second application elaborated in this paper is shrinkage estimation of a covariance
matrix as studied in depth for i.i.d. sequences of high–dimensional random vectors as well
as dependent vector time series, see by [12], [13] and [16] amongst others. In order to regularize the sample variance-covariance matrix, the shrinkage estimator considers a convex
combination with a well–defined target that usually corresponds to a simple regular model.
We consider the identity target, i.e. a multiple of the identity matrix In of dimension dn . To
the best of our knowledge, large sample approximations for those estimators have not yet
been studied. We show that, uniformly in the shrinkage weight for the convex combination,
a bilinear form given by the shrinkage estimator can be approximated by a Gaussian process
when it is centered at the shrunken true covariance matrix using the same shrinkage weight.
By uniformity, the result also holds for the widely used estimator of the optimal shrinkage
weight. For this estimated optimal weight the convergence rate under quite general conditions
is known. It turns out that, when comparing the matrices in terms of a natural pseudodistance induced by bilinear forms, the convergence rate carries over from the optimal weight’s
INFERENCE FOR THE TRACE
3
estimator. We also compare the shrinkage estimator using the estimated optimal weight with
an oracle estimator using the unknown optimal weight. Last, we study the case of nearly (i.e.
asymptotically) orthogonal vectors. As a consequence of the Kabatjanskii-Levenstein bound,
see [18], this property allows to place much more unit vectors on the unit sphere. It turns
out that for nearly orthogonal vectors the nonparametric part dominates in large samples,
contrary to the situation for vectors bounded away from orthogonality.
The high-dimensional time series model of the paper is as follows. At time n, n ∈ N, we
observe a d = dn dimensional mean zero vector time series
(1)
(d )
Yni = (Yni , . . . , Yni n )′ ,
1 ≤ i ≤ n,
defined on a common probability space (Ω, F, P ), whose coordinates are causal linear processes,
(1.1)
(ν)
Yni
=
∞
X
(ν)
i ∈ N0 , ν = 1, . . . , dn ,
cnj ǫi−j ,
j=0
where {ǫt } are independent mean zero error
P terms,possibly not identically distributed, such
that E|ǫk |4+δ < ∞ for some δ > 0 and n−1 ni=1 E(ǫri ), r = 2, 3, 4, converges. The coefficients
(ν)
cnj may depend on n and are therefore also allowed to depend on the dimension dn . We
impose the following growth condition.
(ν)
Assumption (A): The coefficients cnj of the linear processes (1.1) satisfy
(1.2)
(ν)
sup max |cnj |2 ≪ j −3/2−θ
n∈N 1≤ν≤dn
for some 0 < θ < 1/2.
It is well known that Assumption (A) covers common classes of weakly dependent time
series such as ARMA(p, q)-models as well as a wide range of long memory processes. We refer
to [17] for a discussion.
Define the (centered) bilinear form
(1.3)
Qn (vn , wn ) =
where
√
b n − Σn )wn ,
nvn′ (Σ
vn , wn ∈ Rdn ,
n
(1.4)
X
′
bn = 1
Σ
Yni Yni
n
and
i=1
b n.
Σn = E Σ
The class of proper (sequences of) weighting vectors,
wn = (wn1 , . . . , wndn )′ ,
n ≥ 1,
studied throughout the paper is the set W of those sequences {wn : n ≥ 1}, wn ∈ Rdn , n ≥ 1,
which have uniformly bounded ℓ1 -norm in the sense that
(1.5)
sup kwn kℓ1 = sup
n∈N
dn
X
n∈N ν=1
|wnν | < ∞.
4
A. STELAND, R. VON SACHS
ℓ1 -weighting vectors naturally arise in various applications such as sparse principal component
analysis as, see e.g. [20], or sparse financial portfolio selection as studied by [4]. For a more
detailed discussion we refer to [17].
It is worth mentioning that our results easily carry over to weighting vectors with uniformly
bounded ℓ2 -norm, provided one relies on standardized versions of the bilinear form (1.3). First
notice that conditions (1.2) and (1.5) allow us to control the linear process coefficients of a
projected time series, wn′ Yni , 1 ≤ i ≤ n, which are then O(j −3/2−θ ) and therefore decay at
the same rate as the original time series. The ℓ2 assumption
v
u dn
uX
w2 < ∞,
(1.6)
sup t
nν
n≥1
ν=1
leads to the estimate
dn
X
(ν)
wnν cnj
ν=1
v
u dn
uX (ν)
= O t
|c |2
nj
ν=1
For bounded dimension, (1.2) yields the estimate j −3/2−θ for the latter expression, but for
P
(ν) 2
growing dimension this does not hold in general. Assuming ∞
ν=1 |cnj | < ∞ for all j ≥ 0, is,
(ν)
however, not reasonable for a high-dimensional setting, since then cnj = o(1), as ν → ∞. For
(ν)
example, if cnj = ρjν , for 0 < |ρν | < 1, ν ≥ 1, the latter assumption would rule out the case
of observing dn autoregressive time series of order 1 with autoregressive parameters bounded
(ν)
away from zero. On the other hand, if wnν ≥ wmin > 0, for ν ≥ 1, and cnj ≥ cj,min > 0 for
Pn
(ν)
wnν cnj ≥ dn wmin cj,min → ∞ for each lag j ≥ 0. This can be
ν ≥ 1 and n ≥ 1, then dν=1
e n′ Yni instead of wn′ Yni , where w
e n = d−1
fixed by considering w
n wn . Then
!
d
d
n
n
X
1 X
(ν)
(ν)
j ≥ 0.
w
enν cnj = O
|cnj |2 ≪ j −3/2−θ ,
dn
ν=1
ν=1
Next observe that by Jensen’s inequality
v
u∞
dn p
X
X
1 u
1
2
t
2 .
√
|w
enν | =
wnν ≤
wnν
dn
d
n
ν=1
ν=1
ν=1
dn
X
e n : n ≥ 1} ∈ W and w
e n′ Yni is a linear time
Hence, (1.2) and the ℓ2 -condition (1.6) imply {w
series with coefficients decaying at the same rate j −3/2−θ as the original time series. Clearly,
for any sequences vn , wn of weighting vectors with uniformly bounded ℓ2 -norm, we have the
scaling property
e n ),
Qn (vn , wn ) = d2n Qn (e
vn , w
en = d−1
e n = d−1
where v
n vn and w
n wn have uniformly bounded ℓ1 -norm. But if one standardizes
2
Qn , the factor dn cancels. Hence, in this sense several of our theoretical results can be also
applied to study projection onto vectors with uniformly bounded ℓ2 -norm.
The rest of the paper is organized as follows. In Section 2, we introduce the partial sums
and partial sum processes associated to an increasing number of bilinear forms and establish
the strong and weak approximation theorems for those bilinear forms. The application to
the trace functional is discussed in Section 3. The large sample approximations for shrinkage
estimators of covariance matrices are studied in depth in Section 4.
5
INFERENCE FOR THE TRACE
2. LARGE SAMPLE APPROXIMATIONS FOR HIGH–DIMENSIONAL BILINEAR FORMS
2.1. Definitions and review
Let us define the matrix–valued partial sums
!
k
X
(ν)
(µ)
b nk =
Yi Yi
Σ
(2.1)
,
i=1
(2.2)
Σnk =
k
X
i=1
1≤ν,µ≤dn
(ν)
EYi
(µ)
Yi
!
,
1≤ν,µ≤dn
for n, k ≥ 1, and put
(2.3)
b nk − Σnk )wn ,
Dnk (vn , wn ) = vn′ (Σ
n, k ≥ 1,
for two sequences of weighting vectors {vn }, {wn } ∈ W. The associated càdlàg processes will
be denoted by
(2.4)
b n,⌊nt⌋ −Σn,⌊nt⌋ )wn ,
Dn (t; vn , wn ) = n−1/2 Dn,⌊nt⌋ (vn , wn ) = vn′ n−1/2 (Σ
t ∈ [0, 1], n ≥ 1.
Especially, we have
√
b n − Σn )wn , n ≥ 1,
Dn (1) = Dn (1; vn , wn ) = vn′ n(Σ
b n = 1 Pn E(Yni Yni )′ .
with Σn = E Σ
i=1
n
For some sequence of standard Brownian motions, {Bn (t) : t ∈ [0, ∞)}, n ≥ 1, and any
constant N > 0 we can introduce the rescaled version {N −1/2 BN (tN ) : s ∈ [0, 1]} called the
[0,1]–version of Bn . In [17] the following result on the asymptotics of a single bilinear form
for a uniformly bounded ℓ1 -projections is shown:
(2.5)
Theorem 2.1 Suppose Yni , 1 ≤ i ≤ n, n ≥ 1, is a vector time series according to
model (1.1) that satisfies Assumption (A). Let vn and wn be weighting vectors with uniformly bounded ℓ1 –norm in the sense of (1.5). Then, for each n ∈ N, there exists equivalent versions of Dnk (vn , wn ) and Dn (t; vn , wn ), t ≥ 0, again denoted by Dnk (vn , wn ) and
Dn (t; vn , wn ), and a standard Brownian motion {Wn (t) : t ≥ 0}, which depends on (vn , wn ),
i.e. Wn (t) = Wn (t; vn , wn ), both defined on some probability space (Ωn , Fn , Pn ), such that for
some λ > 0 and a constant Cn
(2.6)
|Dnt (vn , wn ) − αn (vn , wn )Wn (t)| ≤ Cn t1/2−λ ,
for all t > 0 a.s.,
where αn (vn , wn ) is defined in (A.7). If Cn n−λ = o(1), as n, t → ∞, this implies the strong
approximation
(2.7)
sup |Dn (t; vn , wn ) − αn (vn , wn )Wn (⌊nt⌋/n)| = o(1),
t∈[0,1]
as n → ∞, for the [0, 1]–version of Wn as well as the CLT
(2.8)
|Dn (1; vn , wn ) − αn (vn , wn )Wn (1)| = o(1),
as n → ∞, i.e. Dn (1) is asymptotically N (0, α2n ).
a.s.,
a.s.,
6
A. STELAND, R. VON SACHS
A multivariate version for L ∈ N bilinear forms, which approximates
L
Dn (·; vn(j) , wn(j) )
j=1
by a L–dimensional Brownian motion, has been shown in [17, Th. 4.2]. This result allows to
consider the dependence structure which arises when mapping Yn = Ynn onto the subspace
(j)
PL = span{wn : j = 1, . . . , L} spanned by L weighting vectors,
wn(j) ∈ W,
j = 1, . . . , L.
We have the canonical mapping, called projection (onto PL ) in the sequel,
(2.9)
Yn 7→ Pn Yn ,
Pn = (wn(1) , . . . , wn(L) )′ ,
(1)
(L)
which represents the orthogonal projection onto PL , if wn , . . . , wn are orthonormal. The
associated variance-covariance matrix is
.
Cov (Pn Yn ) = Cov (wn(j) Yn , wn(k) Yn )
= wn(j) ′ Σn wn(k)
1≤j,k≤L
1≤j,k≤L
(j)
If the wn ’s are eigenvectors of Σn , then Cov (Pn Yn ) is a diagonal matrix. But that property
is lost for more general spanning vectors. Given the sample Yn1 , . . . , Ynn of dn -dimensional
random vectors, the canonical nonparametric statistical estimators of Σn and Cov (Pn Yn )
b n as defined in (1.4) and
are Σ
d (Pn Yn ) = w(j) ′ Σ
b n w(k)
Cov
.
n
n
1≤j,k≤L
d (Pn Yn ) consist of bilinear forms as studied in Theorem 2.1, and for fixed
The entries of Cov
L its multivariate extension suffices to study the dependence structure of the projection onto
PL . This no longer holds, if L is allowed to grow as the sample size increases, i.e. when
studying the case L = Ln → ∞, as n → ∞. Indeed, the treatment of that high–dimensional
situation is much more involved. As we shall see, it requires a different scaling and a more
involved mathematical framework: The strong approximations we establish in this paper take
place in the Euclidean space RLn of growing dimension and the infinite-dimensional Hilbert
space ℓ2 , respectively.
Thus, to go beyond the case of a finite number of bilinear forms we now consider
b n w(j) ,
Qn (vn(j) , wn(j) ) = vn(j) ′ Σ
n
j = 1, . . . , Ln ,
Dnj = L−1/2
Dn (1; vn(j) , wn(j) ),
n
j = 1, . . . , Ln ,
(j)
(j)
∞
where ({vℓ }∞
ℓ=1 , {wℓ }ℓ=1 ) ∈ W × W, j = 1, . . . , Ln , are Ln pairs of uniformly ℓ1 -bounded
sequences of weighting vectors and Ln may tend to infinity as n → ∞. We are interested in
the joint asymptotics of the centered and scaled versions of the corresponding statistics (2.5)
given by
(2.10)
and the associated sequential càdlàg processes
(2.11)
Dnj (t) = L−1/2
Dn (t; vn(j) , wn(j) ),
n
t ∈ [0, 1],
j = 1, . . . , Ln ,
7
INFERENCE FOR THE TRACE
−1/2
cf. (2.4). The additional factor Ln
anticipates the right scaling to obtain a large sample
approximation.
Further, we are interested in studying weighted averages where averaging takes place over
all Ln forms and all sample sizes n ∈ N. Let λn be the weight for sample size n and µρ
the weight for the ρth quadratic form associated to a pair of sequences of weighting vectors
(ρ)
(ρ)
{vn , wn : n ≥ 1} for ρ = 1, . . . , Ln , n ≥ 1. Define for k ≥ 1
(2.12)
Dk ({(vn(ρ) , wn(ρ) )})
=
∞
X
λn λm
Ln X
Lm
X
ρ=1 σ=1
n,m=1
where
(2.13)
(2.14)
b nmk =
Σ
X
(σ)
b nmk − Σnmk )wm
µρ µσ vn(ρ) ′ (Σ
,
′
,
Yni Ymi
i≤k
b nmk ),
Σnmk = E(Σ
for n, m ≥ 1. Notice the relations
b nnk = Σ
b nk ,
Σ
b nn = nΣ
bn
Σ
(ρ)
(ρ)
(ρ)
(ρ)
between (1.4), (2.1) and (2.13). Dk ({(vn , wn )}) depends on all weights {vnν , wnν , λn , µρ :
1 ≤ ν ≤ dn , 1 ≤ ρ ≤ Ln , n ≥ 1} but is measurable w.r.t Gk = σ(Yni : n ∈ N, i ≤ k). Now, for
any sample size M we may consider the associated càdlàg process associated to (2.12)
(2.15)
DM (t; {(vn(ρ) , wn(ρ) )}) = M −1/2 D⌊M t⌋ ({(vn(ρ) , wn(ρ) )}),
t ∈ [0, 1].
2.2. Preliminaries
Before proceeding, recall the following facts on the Hilbert space ℓ2 and strong approximations in Hilbert spaces. We shall denote the inner product of an arbitray Hilbert space by
< ·, · >, the induced norm by| · | and the operator semi-norm of an operator T : H → H
by kT kop = supx∈H:|x|=1 |T x|. Our results take place in the Hilbert space ℓ2 of all sequences
P
f = (fj )j with j fj2 < ∞, which is a separable Hilbert space when equipped with the inner
p
P
product (f, g) = j fj gj , f = (fj )j , g = (gj )j ∈ ℓ2 , and the induced norm kf kℓ2 = (f, f ).
The associated operator norm of an operator T : ℓ2 → ℓ2 is simply denoted by kT k. For two
random variables X, Y defined on (Ω, F, P ) with E(X 2 ), E(Y 2 ) < ∞, we denote the inner
product by (X, Y )L2 , where L2 = L2 (Ω, F, P ).
Sufficient conditions for a strong approximation of partial sums of dependent random
elements taking values in a separable Hilbert space require the control of the associated
conditional covariance operator. Denote the underlying probability space by (Ω, F, P ). Let
2
X
P = (X2j )j be a random element defined on (Ω, F, P ) taking values in ℓ2 with EkXkℓ2 =
j E(Xj ) < ∞. The covariance operator CX : ℓ2 → ℓ2 associated to X is defined by
X
fj E(Xj Xk ) ,
CX (f ) = E[(f, X)X] =
j
k
f = (fj )j ∈ ℓ2 .
8
A. STELAND, R. VON SACHS
To any σ-field A we may associate the conditional covariance operator of X given A,
X
(2.16) CX (f |A) = E[(f, X)X|A] =
fj E(Xj Xk |A) ,
f = (fj )j ∈ ℓ2 .
j
k
Covariance operators are symmetric positive linear operators with operator norm
(2.17)
kCX (·|A)k =
sup
f ∈ℓ2 :kf kℓ2 =1
k(CX (f ), f )kℓ2 .
For further properties and discussion see [3].
A strong invariance principle in ℓ2 deals with the a.s. approximation of partial sums of
ℓ2 –valued random elements by a Brownian motion in ℓ2 . Recall that a random element B =
{B(t) : t ∈ [0, 1]} with values in C([0, 1], ℓ2 ) is called Brownian motion in ℓ2 , if
(i) B(0) = 0 ∈ ℓ2 ,
(ii) for all 0 ≤ t1 < · · · < tn ≤ 1 the increments B(ti+1 ) − B(ti ), i = 1, . . . , n − 1, are
independent and
(iii) for all 0 ≤ s, t ≤ 1 the increment B(t) − B(s) is Gaussian with mean 0 and covariance
operator
P min(s, t)K for some nonnegative linear and self-adjoint operator K on ℓ2 such
that ∞
i=1 (Kei , ei ) < ∞, where {ei } is some orthonormal system for ℓ2 .
If K = CX for some random element X, B is the Brownian motion generated by X. The
definition for a general separable Hilbert space is analogous.
A strong invariance principle or strong approximation for a sequence ζ1 , ζ2 , . . . of random
elements taking values in an arbitrary separable Hilbert space H with inner product < ·, · >
and induced norm | · | asserts that they can be redefined on a rich enough probability space
such that there exists a Brownian motion B with values in H and covariance operator Cζ
such that, a.s.,
(2.18)
X
j≤t
ζj − B(t) ≤ ct1/2−λ ,
for constants λ > 0 and c < ∞, if the dimension of H is finite, and,
(2.19)
X
j≤t
p
ζj − B(t) = o( t log(log t)),
as t → ∞, a.s., if H is infinite dimensional.
Throughout the paper we write, for two arrays {an′ ,m′ } and {bn′ ,m′ } of real numbers,
n′ ,m′
an′ ,m′ ≪ bn′ ,m′ , if there exists a constant c such that an′ ,m′ ≤ cbn′ ,m′ for all n′ , m′ .
2.3. Large sample approximations
We aim at showing a strong approximation for the D([0, 1]; ℓ2 )-valued processes
n
Dn = (Dnj )L
j=1 ,
n ≥ 1,
INFERENCE FOR THE TRACE
9
where the coordinate processes Dnj are given by
Dnj (t) = √
(j)
1
Dn,⌊nt⌋ (vn(j) , wn(j) ),
nLn
(j)
(j)
t ∈ [0, 1],
n ≥ 1,
(j)
b nk − Σnk )wn for j = 1, . . . , Ln , n ≥ 1, cf. (2.3) and (2.10).
with Dnk (vn , wn ) = vn ′ (Σ
The above processes can be expressed as partial sums.
Lemma 2.1
(2.20)
We have the representation
Dnk =
k
X
(n)
ξi ,
i=1
for k = 1, . . . , n, n ≥ 1, leading to
(2.21)
Dn (t) = √
⌊nt⌋
1 X (n)
ξ ,
nLn i=1 i
(n)
where the random elements ξi
t ∈ [0, 1], n ≥ 1,
are defined in (2.37).
To introduce the conditional covariance operators associated to Dn , denote the filtration
Fm = σ(ǫi : i ≤ m), m ∈ Z, and define
C (n) (f |F0 ) = E[(f, Dn (1))Dn (1) | F0 ],
f ∈ ℓ2 , n ≥ 1.
Let us also introduce the unconditional covariance operator
C (n) (f ) = E[(f, Zn )Zn ],
f ∈ ℓ2 ,
n
where Zn = (Znj )L
j=1 , n ≥ 1, with random variables Znj , 1 ≤ j ≤ Ln , satisfying E(Znj ) = 0
and
2
E(Znj Znk ) = L−1
n βn (j, k),
for j, k = 1, . . . , Ln . Here
(2.22)
βn2 (j, k) = βn2 (vn(j) , wn(j) , vn(k) , wn(k) )
are the quantities introduced in (A.9), the asymptotic covariance parameters of the bilinear
(j)
(j)
(k)
(k)
forms corresponding to the pairs (vn , wn ) and (vn , wn ), 1 ≤ j, k ≤ Ln .
The following technical but crucial result establishes the convergence of C (n) (·|F0 ) − C (n) (·)
in the operator semi–norm in expectation and provides us with a convergence rate.
Theorem 2.2
(j)
(j)
Suppose (vn , wn ), j = 1, . . . , Ln , have uniformly bounded ℓ1 -norms,
sup max max{kvn(j) kℓ1 , kwn(j) kℓ1 } ≤ C < ∞,
n≥1 1≤j≤Ln
for some constant C. Let
(2.23)
(n)
Sn′ ,m′
1
=√ ′
n Ln
′ +n′
mX
k=m′ +1
(n)
ξk ,
m′ , n′ ≥ 1, n ≥ 1,
10
A. STELAND, R. VON SACHS
(n)
with ξk
defined by (2.37). Define
(n)
(n)
(n)
Cn′ ,m′ (f |Fm′ ) = E[(f, Sn′ ,m′ )Sn′ ,m′ |Fm′ ],
for f ∈ ℓ2 . Then
n′ ,m′
(n)
EkCn′ ,m′ (·|Fm ) − C (n) (·)k ≪ (n′ )−θ/2 ,
where k · k denotes the operator norm defined in (2.17).
We are now in a position to formulate the first main result on the large sample approximations of Ln bilinear forms when Ln converges to infinity, in terms of the ℓ1 – as well as the
ℓ2 –norm. The results holds true under the weak assumption that the weighting vectors have
uniformly bounded ℓ1 norm.
Theorem 2.3 Let Yni , 1 ≤ i ≤ n, be a vector time series following model (1.1) and sat(L )
(Ln ) ∞
(1) ∞
(1)
}ℓ=1 , {wn n }∞
isfying Assumption (A). Suppose that ({vℓ }∞
ℓ=1 ) ∈
ℓ=1 , {wℓ }ℓ=1 ), . . . , ({vn
W × W, n ≥ 1, have uniformly bounded ℓ1 -norm, i.e.
sup max max{kvn(j) kℓ1 , kwn(j) kℓ1 } ≤ C,
n≥1 1≤j≤Ln
for some constant C < ∞. Then all processes can be redefined on a rich enough probability
space, such that there exists, for each n, a Brownian motion of dimension Ln ,
Bn (t) = Bn (t; {(vn(j) , wn(j) ) : 1 ≤ j ≤ Ln }),
t ≥ 0,
with coordinates Bn (t)j , j = 1, . . . , Ln , and covariance function given by
(2.24)
2
E(Bn (s)j Bn (t)k ) = min(s, t)L−1
n βn (j, k),
for 1 ≤ j, k ≤ Ln and s, t ≥ 0, such that the following assertions hold true.
(i) In the Euclidean space RLn we have the strong approximation
(2.25)
kL−1/2
Dnt
n
− Bn (t)kRLn =
t
X
(n)
L−1/2
ξi
n
i=1
− Bn (t)
RLn
≤ Cn t−1/2−λn ,
a.s, for constants Cn < ∞ and λn > 0, where λn depends only on Ln , δ and θ.
Provided
(2.26)
Cn n−λn = o(1),
as n → ∞, the following assertions hold.
(ii) With respect to the ℓ2 –norm we have
sup
Ln
X
t∈[0,1] j=1
Dnj (t) − B n
⌊nt⌋
n
2
= o(1)
j
as n → ∞, a.s., for the [0, 1]–version B n of Bn .
INFERENCE FOR THE TRACE
11
(iii) With respect to the ℓ1 –norm we have
Ln
1 X
⌊nt⌋
Dnj (t) − B n
= o(1),
sup √
n j
Ln j=1
t∈[0,1]
as n → ∞, a.s., for the [0, 1]–version B n of Bn , and with respect to the maximum norm
sup max Dnj (t) − B n
t∈[0,1] j≤Ln
⌊nt⌋
n
= o(1),
j
as n → ∞, a.s..
(iv) Let {λn : n ∈ N} and {µρ : ρ ∈ N} be ℓ1 –weights. Then there exist constants λ > 0
(ρ)
(ρ) ∞
∞
and C < ∞ and α({λn }∞
n=1 , {µρ }ρ=1 , {(vn , wn )}n,ρ=1 ) ≥ 0, such that for equivalent
versions and a standard Brownian motion B on [0, ∞), defined on a new probability
space,
(2.27)
∞
(ρ)
(ρ) ∞
1/2−λ
|Dt ({(vn(ρ) , wn(ρ) )}) − α({λn }∞
,
n=1 , {µρ }ρ=1 , {(vn , wn )}n,ρ=1 )B(t)| ≤ Ct
a.s., for all t > 0. Further, for any sample size M
(2.28)
−λ
∞
(ρ)
(ρ) ∞
,
sup |DM (t; {(vn(ρ) , wn(ρ) )})−α({λn }∞
n=1 , {µρ }ρ=1 , {(vn , wn )}n,ρ=1 )B M (t)| ≤ CM
t∈[0,1]
a.s., for the [0, 1]–version B M of B.
Remark 2.1
(2.29)
The Brownian motions can be constructed such that
⌊nt⌋ (j) (j)
⌊nt⌋
−1/2
(j)
(j)
; vn , wn
= o(1),
− Ln αn (vn , wn )Wn
sup B n
n j
n
t∈[0,1]
(j)
(j)
as n → ∞, a.s., if (2.26) holds, where Wn (·; vn , wn ) is as in Theorem 2.1, for j = 1, . . . , Ln .
Due to assertion (iv) of the above theorem we may conjecture that (2.26) holds, cf. the
discussion in [17], but we have neither a proof nor a counterexample. The following result
studies the relevant processes in the infinite–dimensional space ℓ2 and yields an approximation
in probability taking into account the additional factor log log(n).
Theorem 2.4 Suppose that the assumptions of Theorem 2.3 hold. In the Hilbert space ℓ2 we
have the strong approximation
(2.30)
kLn−1/2 Dnt
− Bn (t)kℓ2 =
t
X
k=1
(n)
L−1/2
ξk − Bn (t)
n
ℓ2
p
= o( t log(log t)),
as t → ∞, a.s. There exists a sequence {δn : n ≥ 1} ⊂ N such that with N = ⌈n log log(n)⌉
(2.31)
max
δn ≤k≤n
k
X
k
1
(n)
√
ξi − B N
N
N Ln i=1
= oP (1),
ℓ2
12
A. STELAND, R. VON SACHS
n → ∞, for the [0, 1]–version B N (t) = N −1/2 Bn (tN ). In other words,
r
k
n
k
Dn
= oP (1),
− BN
(2.32)
max
δn ≤k≤n
N
n
N
ℓ2
or equivalently
(2.33)
sup
δn
≤t≤1
n
r
n
Dn (t) − B N
N
⌊nt⌋
N
= oP (1),
ℓ2
as n → ∞.
The above result eliminates the condition (2.26), but we have no detailed information about
the sequence δn .
The question arises, whether the above results are limited to linear processes. As the main
arguments deal with approximating martingales, we have the following result, which suggests
that the class of vector time series to which the main results of this paper apply is larger.
(1)
(L )
(1)
(Ln ) ∞
}ℓ=1 )
n
∞
}∞
Theorem 2.5 Let ({vℓ }∞
ℓ=1 , {wn
ℓ=1 , {wℓ }ℓ=1 ), . . . , ({vn
projection vectors with uniformly bounded ℓ1 -norm, i.e.
∈ W × W, n ≥ 1, be
sup max max{kvn(j) kℓ1 , kwn(j) kℓ1 } ≤ C,
n≥1 1≤j≤Ln
for some constant C < ∞. Let Yni , 1 ≤ i ≤ n, be a dn -dimensional vector time series such
(µ)
(ν)
(µ)
(n) (ν)
that Dnk (vn , wn ) can be approximated by the martingales Mk (vn , wn ), k ≥ 1, defined
(ν)
in (2.38) in L2 with rate n−θ , ν = 1, . . . , dn , for certain sequences of coefficients, cnj , j ≥ 0,
ν = 1, . . . , dn , satisfying Assumption (A) and a sequence of independent mean zero random
variables {ǫk : k ≥ 0} with supk≥1 E(ǫ4k ) < ∞. If
sup sup max max{E|vn(j) ′ Ynk |4+2δ , E|wn(j) ′ Ynk |4+2δ } < ∞,
n≥1 k≥1 1≤j≤Ln
for some δ > 0, then the results of this section still hold true.
2.4. Proofs
Proof of Lemma 2.1: We argue as in the proof of Theorem 2.1 given in [17], where it was
shown that the partial sum (2.3) associated to a single bilinear form Q(vn , wn ) attains the
representation
X (n)
(2.34) Dnk (vn , wn ) =
ξi (vn , wn )
i≤k
with Gaussian random variables
(2.35)
(n)
ξi (vn , wn ) = Yni (vn )Yni (wn ) − E(Yni (vn )Yni (wn )),
i = 1, 2, . . . , n, n ≥ 1,
for linear processes
Yni (vn ) =
∞
X
j=0
(v)
cnj ǫi−j ,
Yni (wn ) =
∞
X
j=0
(w)
cnj ǫi−j ,
i ∈ Z, n ∈ N,
13
INFERENCE FOR THE TRACE
with coefficients
(2.36)
(v)
cnj
=
dn
X
(ν)
vν cnj ,
(w)
cnj
=
dn
X
(ν)
wν cnj ,
ν=1
ν=1
(j)
(j)
for j ≥ 0 and n ≥ 1. For Ln pairs of weighting vectors (vn , wn ), j = 1, . . . , Ln , we consider
the corresponding partial sum process where the summands are the Ln -dimensional vectors
(2.37)
(n)
ξi
Ln
(n)
= ξi (j)
,
j=1
(n)
(n)
ξi (j) = ξi (vn(j) , wn(j) ), j = 1, . . . , Ln ,
for i = 1, 2, . . . , which we, however, also interpret as random elements taking values in ℓ2 .
This completes the proof.
Q.E.D.
[15, Th. 1] asserts that (2.18) and (2.19), respectively, hold, if the following conditions for
Pm′ +n′
′ ′
the scaled partial sums Sn′ (m′ ) = (n′ )−1/2 k=m
′ +1 ζk , k, m , n ≥ 0, are satisfied:
2+δ
(I) supj≥1 E|ζj |
< ∞ for some δ > 0.
(II) For some ε > 0
n′ ,m′
E|E(Sn′ (m′ ))|Fm′ )| ≪ (n′ )−ε .
(III) There exists a covariance operator C such that the conditional covariance operators
Cn′ (f |Fm′ ) = E[(f, Sn′ (m′ ))Sn′ (m′ )|Fm′ ),
f ∈ H,
converge in the operator semi-norm k · k to C(f ) in expectation with rate (n′ )−θ , i.e.
n′ ,m′
EkCn′ (·|Fm′ ) − C(·)k ≪ (n′ )−θ ,
for some θ > 0.
For a discussion of this result and extensions see [24]. As shown by [5], the strong invariance
principle (2.19) also holds true for strictly stationary sequences taking values in a separable
Hilbert space, which possess a finite moment of order 2+δ′ , δ′ > 0, and are strong mixing with
′
mixing coefficients satisfying α(k) = O(k−(1+ε)(1+2/δ ) ), for some ε > 0. The above conditions
are, however, more convenient when studying linear processes. [21] has studied strong invariance principles for a univariate nonlinear time series using the physical dependence measure,
which is easy to verify for linear processes. Extensions to vector-valued time series (of fixed
dimension) have been provided by [14]. We rely on the conditions of [15], since they allow to
study time series of growing dimension and taking values in the infinite-dimensional space ℓ2
in a relatively straightforward way.
As a preparation for the proof of Theorem 2.2, we need the following lemma dealing with
the uniform convergence of unconditional and conditional covariances of the approximating
martingales defined by
)
(
∞
m′
X
X
(n)
(n)
(n)
m′ ≥ 0,
fe (ν)ǫk−l ,
fe (ν)(ǫ2 − σ 2 ) + ǫk
(2.38) M ′ (ν) =
0,0
m
k=0
k
k
l,0
l=1
(ν)
(n) (ν)
(n)
where for brevity fel,i (ν) = fel,i (vn , wn ), l, i = 0, 1, 2, . . . , ν = 1, . . . , dn .
14
A. STELAND, R. VON SACHS
Lemma 2.2
Under Assumption (A) we have
(2.39)
(n)
(n)
(n)
n≥1
n′ ,m′
(n)
sup sup E[(Mm′ +n′ (ν) − Mm′ (ν))(Mm′ +n′ (µ) − Mm′ (µ))|Fm′ ] − n′ βn (ν, µ)
1≤ν,µ
<< (n′ )1−θ ,
L1
which implies
(2.40)
(n)
(n)
sup sup (n′ )−1 sup E[(Mm′ +n′ (ν) − Mm′ (ν))2 |Fm′ ] < ∞,
n≥1
1≤ν
n′ ,m′ ≥1
a.s. Further,
(2.41)
(n)
(n)
(n)
n′ ,m′
(n)
(n′ )−1 sup sup E(Mm′ +n′ (ν) − Mm′ (ν))(Mm′ +n′ (µ) − Mm′ (µ)) − n′ βn (ν, µ) << (n′ )−θ ,
n≥1 1≤ν,µ
which implies
(2.42)
(n)
(n)
sup sup (n′ )−1 sup E[(Mm′ +n′ (ν) − Mm′ (ν))2 ] < ∞,
n≥1
1≤ν
n′ ,m′ ≥1
Proof: A direct calculation leads to
(n)
(n)
(n)
(n)
CM n (ν, µ) = E[(Mm′ +n′ (ν) − Mm′ (ν))(Mm′ +n′ (µ) − Mm′ (µ)) | Fm′ ]
(3)
(0)
= CM n (ν, µ) + · · · + CM n (ν, µ),
where
(0)
CM n (ν, µ)
′ +n′
mX
=
∞
X
(n)
(n)
fel,0 (ν)feℓ,0 (µ)E[ǫk ǫk−l (ǫ2k′ − σk2 )|Fm′ ],
k,k ′ =m′ +1 l=1
(1)
CM n (ν, µ)
′ +n′
mX
=
k=m′ +1
(2)
CM n (ν, µ)
(n)
(n)
fe0,0 (ν)fe0,0 (µ)(γk − σk4 ),
′ +n′ k−m′ +1
mX
X
=
k=m′ +1
(3)
CM n (ν, µ) =
′ +n′
mX
l=0
∞
X
(n)
(n)
2
fel,0 (ν)fel,0 (µ)σk2 σk−l
,
k=m′ +1 l,l′ =k−m′
(n)
(n)
fel,0 (ν)fel′ ,0 (µ)σk2 ǫk−l ǫk−l′ ,
(1)
(2)
for m′ , n′ ≥ 0. Let us first estimate CM n (ν, µ) + CM n (ν, µ). We have
n′ ,m′
(2)
(1)
sup sup |CM n (ν, µ) + CM n (ν, µ) − n′ βn (ν, µ)| ≪ (n′ )1−θ ,
n≥1 1≤ν,µ
see (A.10). Next, we show that
#
"
(2.43)
sup E
n≥1
(3)
sup |CM n (ν, µ)|
1≤ν,µ
n′ ,m′
≪ (n′ )1−θ .
15
INFERENCE FOR THE TRACE
Recall that c2 = supk≥1 E(ǫ2k ) < ∞ and assume w.l.o.g. c2 = 1 in what follows. The Cauchy–
Schwarz inequality yields
(3)
|(n′ )−1 CM n (ν, µ)|
′
′
m +n
∞
∞
X
X
1 X
(n)
(n)
≤ ′
fel,0 (ν)ǫk−l
fel′ ,0 (µ)σk2 ǫk−l′
n
k=m′ +1 l=k−m′
l′ =k−m′
v
!2 v
!2
u
u
′ +n′
′ +n′
mX
∞
∞
u 1 mX
X
X
u
1
(n)
(n)
≤t ′
|fel,0 (ν)|ǫk−l t ′
|fel,0 (µ)|ǫk−l′ .
n
n
′
′
′
′
k=m +1
k=m +1
l=k−m
l=k−m
Using k · kℓ1 ≤ k · kℓ4 and Jensen’s inequality, we obtain
v
u
u1
t
n′
′ +n′
mX
k=m′ +1
∞
X
l=k−m′
(n)
|fel,0 (ν)|ǫk−l
!2
v
u
u
u1
≤t ′
n
v
u
u
u1
≤t ′
n
≤
1
n′
′ +n′
mX
∞
X
k=m′ +1
l=k−m′
′ +n′
mX
∞
X
k=m′ +1
′ +n′
mX
l=k−m′
∞
X
(n)
[fel,0 (ν)]4 ǫ4k−l
!1/2
(n)
sup[fel,0 (ν)]4 ǫ4k−l
1≤ν
(n)
sup[fel,0 (ν)]4 ǫ4k−l
1≤ν
k=m′ +1 l=k−m′
!1/2
!1/4
,
where the upper bound does not depend on ν. Hence,
v
u
′ +n′
∞
X
u 1 mX
(n)
′ −1 (3)
t
E sup |(n ) CM n (ν, µ)| ≤ E
sup[fel,0 (ν)]4 ǫ4k−l
′
n
1≤ν,µ
1≤ν
k=m′ +1 l=k−m′
v
u
′ +n′
∞
X
u 1 mX
(n)
t
sup[fel,0 (ν)]4 γk−l
≤
′
n
′
′ 1≤ν
k=m +1 l=k−m
(n)
(ν)
(µ)
Using |fel,0 (ν)| = O(kvn kℓ1 kvn kℓ1 l−3/4−θ/2 ), uniformly in n, supk≥1 γk < ∞ and the elePm′ +n′ P∞
−3−2θ = O((n′ )1−2θ ), (2.43) follows. Lastly, consider
mentary fact that k=m′
l=k−m′ +1 l
(0)
CM n (ν, µ). Since the indices satisfy k − ℓ ≤ m′ , ǫk−ℓ is Fm′ -measurable, whereas ǫk , ǫk′ are
independent from Fm′ . Hence
(0)
CM n (ν, µ)
=
′ +n′
mX
∞
X
k,k ′ =m′ +1 l=1
(n)
(n)
fel,0 (ν)fel,0 (µ)ǫk−ℓ E(ǫk (ǫ2k′ − σk2′ )).
Clearly, for k 6= k′ the summands vanish, such that
(0)
CM n (ν, µ)
=
′ +n′
mX
∞
X
k=m′ +1 l=1
(n)
(n)
fel,0 (ν)fel,0 (µ)ǫk−l E(ǫ3k ).
16
A. STELAND, R. VON SACHS
(0)
If E(ǫ3k ) = 0 for all k, CM n (ν, µ) = 0. Otherwise, put c3 = supk≥1 E|ǫk |3 . We have the estimate
′
′ −1
(n )
(0)
E|CM n (ν, µ)|
′
m +n ∞
1 X X e(n)
(n)
≤ c3 ′
|fl,0 (ν)fel,0 (µ)|E|ǫk−l |
n
k=m′ l=1
v
∞
m′ +n′ uX
1 X u
t [fe(n) (ν)]2 [fe(n) (µ)]2
≤ c3 ′
l,0
l,0
n
l=1
k=m′
v
′ +n′ u ∞
mX
uX
n′ ,m′
1
t
l−3−2θ ≪ (n′ )−θ .
= O ′
n
′
l=1
k=m
Hence (2.39) follows, The above arguments also imply that
n′ ,m′
sup sup E(n′ )−1 |CM n (ν, µ)| ≪ 1,
n≥1 1≤ν,µ
since
|(n′ )−1 CM n (ν, µ)| ≤ (n′ )−1 |CM n (ν, µ) − n′ βn (ν, µ)| + (n′ )−1 |βn (ν, µ)|,
where the first term is a.s. finite, since its L1 -norm is ≪ (n′ )−θ , and the second one is ≪ 1,
uniformly in 1 ≤ ν, µ, such that
sup sup E|(n′ )−1 CM n (ν, µ)| ≤ (n′ )−1 |CM n (ν, µ) − n′ βn (ν, µ)| + (n′ )−1 |βn (ν, µ)| ≪ 1.
n≥1 1≤ν,µ
which in turn implies (2.40). To verify (2.41) one first conditions on Fm′ and then argues simi(3)
larly in order to estimate ECM n (ν, µ). Observe that with c24 = max{supk≥1 (σk2 )3 , supk≥1 γk4 supk≥1 σk2 }
′ −1
E(n )
(3)
CM n (ν, µ)
2
m′ +n′ ∞
1 X X
(n)
≤ c24 ′
sup |fl,0 (ν)|
n
1≤ν
k=m′ +1 l=0
v
u
′ +n′
4
∞
X
u 1 mX
(n)
sup |fl,0 (ν)|
≤ c24 t ′
n
1≤ν
′
k=m +1 l=0
n′ ,m′
≪ (n′ )−θ ,
using k · kℓ1 ≤ k · kℓ4 and Jensen’s inequality, which verifies (2.41) and in turn (2.42). Q.E.D.
(n)
(ν)
(ν)
Introduce for m′ , n′ ≥ 1 and each coordinate 1 ≤ ν ≤ Ln the partial sums Dn′ ,m′ (vn , wn ) =
Pm′ +n′ (n) (ν) (ν)
i=m′ +1 ξi (vn , wn ) and denote the appropriately scaled versions by
(n)
(2.44)
(n)
Tn′ ,m′ (ν)
(ν)
(ν)
Dn′ ,m′ (vn , wn )
√
=
,
Ln n ′
m′ , n′ ≥ 1,
for 1 ≤ ν ≤ Ln . The corresponding martingale approximations are given by
(n)
(2.45)
(n)
Mn′ ,m′ (ν) =
(ν)
(ν)
(n)
(ν)
(ν)
Mn′ +m′ (vn , wn ) − Mm′ (vn , wn )
√
,
Ln n ′
m′ , n′ ≥ 1.
17
INFERENCE FOR THE TRACE
We need to study the approximation error,
(n)
(n)
(n)
n′ , m′ ≥ 1.
Rn′ ,m′ (ν) = Tn′ ,m′ (ν) − Mn′ ,m′ (ν),
The next result improves upon [Lemma 2][11] by showing that, firstly, the error is of order
(n′ )−θ in terms of the L1 –norm when conditioning on the past, and, secondly, that the result
is uniform over ℓ1 –bounded weighting vectors.
Lemma 2.3
We have
(n)
sup E (Rn′ ,m′ (ν))2 | Fm′
1≤ν
n′ ,m′
1
′ −θ
≪ L−1
n (n ) .
Proof: Consider, as in [11], the decomposition
(n)
(n)
(n)
(n)
Rn′ ,m′ (ν) = Qn′ ,m′ (ν) + Pn′ ,m′ (ν) + On′ ,m′ (ν),
where
′
′
−i−1
nX
−1 n X
1
(n)
(σ 2 ′ ′ 1{l=0} − ǫm′ +n′ −i ǫm′ +n′ −l ),
fe
=√
Ln n′ i=0 l=0 l,i+1 m −n −i
(2.46)
(n)
Qn′ ,m′ (ν)
(2.47)
(n)
Pn′ ,m′ (ν) = √
(2.48)
(n)
On′ ,m′ (ν)
∞ ∞
1 X X e(n)
(n)
2
(ν) − fel,i+n′ +1 (ν))(ǫm′ −i ǫm′ −i−l − σm
(f
′ −l 1{l=0} ),
Ln n′ i=0 l=0 l,i+1
′
(n)
nX
−1 X
∞
1
(n)
fe
(ν)ǫm′ +n′ −k ǫm′ +n′ −i .
= −√
Ln n′ i=0 k=n′ k−i,i+1
(n)
Pn′ ,m′ (ν) is the projection of Rn′ ,m′ (ν) onto the subspace spanned by {ǫr ǫs −σr2 1{r=s} : −∞ <
2
r, s ≤ m′ } and therefore Fm′ –measurable. Hence with em′ ,i,l = ǫm′ −i ǫm′ −i−l − σm
′ −l 1{l=0}
(n)
Ln n′ E (Pn′ ,m′ (ν))2 | Fm′ =
≤
∞
∞ X
X
(n)
(n)
(fel,i+1 (ν) − fel,i+n′ +1 (ν))em′ ,i,l
i=0 l=0
∞
∞ X
X
i=1 l=0
!2
!2
(n)
(n)
|fel,i+1 (ν) − fel,i+n′ +1 (ν)||em′ ,i,l |
∞
∞ X
X
(n)
(n)
(fel,i+1 (ν) − fel,i+n′ +1 (ν))2 e2m′ ,i,l
≤
i=1 l=0
≤
∞
∞ X
X
(n)
2
(fel,i+1 (ν)
sup γk sup
1≤ν i=0
k≥1
l=0
a.s., such that due to (A.4)
m′ ,n′
(n)
′ −θ
E sup E (Pn′ ,m′ (ν))2 | Fm′ << L−1
n (n ) , a.s.
(n)
− fel,i+n′ +1 (ν))2 ,
1≤ν
(n)
(n)
Qn′ ,m′ (ν) is the projection of Rn′ ,m′ (ν) onto the subspace spanned by {ǫr ǫs − σr2 1{r=s} :
(n)
m′ < r, s ≤ m′ + n′ } and thus independent from Fm′ , such that E (Qn′ ,m′ (ν))2 | Fm′ =
18
A. STELAND, R. VON SACHS
m′ ,n′
(n)
E(Qn′ ,m′ (ν))2 << supk≥1 γk (Ln n′ )−1
Last, by Fatou
(n)
(Ln n′ )E (On′ ,m′ (ν))2 | Fm′
≤ lim
N →∞
≤
′ −1
nX
N
X
i,i′ =0 k,k ′ =n′
(n)
(n)
fek−i,i+1 (ν)fek−i′ ,i′ +1 (ν)ǫm′ +n′ −k ǫm′ +n′ −k′ E(ǫm′ +n′ −i ǫm′ +n′ −i′ )
′ −1
nX
N
X
(n)
(n)
2
sup σk lim
|fek−i,i+1 (ν)fek′ −i,i+1 (ν)ǫm′ +n′ −k ǫm′ +n′ −k′ |
N
→∞
k≥1
i=0 k,k ′ =n′
≤ sup σk2 lim
k≥1
N →∞
v
′ −1 u
nX
u
i=0
n′ −1
≤
m′ ,n′
Pn′ Pn′ −i e(n)
2 << (n′ )−θ , uniformly in 1 ≤ ν.
(ν))
(
f
i=1
l=0
l,i
t
N
X
k,k ′ =n′
(n)
(n)
[fek−i,i+1 (ν)]2 [fek′ −i,i+1 (ν)]2 ǫ2m′ +n′ −k ǫ2m′ +n′ −k′
N
X X
(n)
2
[fek−i,i+1 (ν)]2 ǫ2m′ +n′ −k
sup σk lim
N
→∞
k≥1
i=0 k=n′
a.s., where we estimated the ℓ1 –norm by the ℓ2 –norm. Hence,
′ −1 ∞
nX
X
(n)
(n)
2
′ −1
2 2
sup[fel,i (ν)]2
E sup E (On′ ,m′ (ν)) | Fm′ ≤ (Ln n ) sup(σk )
1≤ν
i=0 l=1 1≤ν
k≥1
m′ ,n′
′ −θ
≪ L−1
n (n ) ,
by virtue of (A.6). This completes the proof.
Q.E.D.
Proof of Theorem 2.2: For a sequence of conditional covariance operators Cn (·|A) =
E[(·, Xn )Xn | A] with Xn = (Xnj )j , E(Xn ) = 0 , n ≥ 1, say, we have convergence in the
operator semi-norm, defined as kT k = supf :kf k=1 |(f, T f )| for an operator T acting on ℓ2 ,
to some unconditional covariance operator C(·) = E[(·, Z)Z], Z = (Zj )j , E(Z) = 0, in
expectation, if
E
sup
f ∈ℓ2 :kf k=1
|(f, Cn (f |A) − C(f ))| = E
sup
∞
X
f ∈ℓ2 :kf k=1 j,k=1
fj fk [E(Xnj Xnk |A) − E(Zj Zk )]
converges to 0, as n → ∞. Define the ℓ2 –valued random elements
∞
∞
(n)
(n)
(n)
(n)
Mn′ ,m′ = Mn′ ,m′ (ν)
,
Tn′ ,m′ = Tn′ ,m′ (ν)
,
ν=1
(n)
(n)
ν=1
(n)
where Tn′ ,m′ (ν) = 0 and Mn′ ,m′ (ν) = Tn′ ,m′ (ν), for ν > Ln . Recall that
(n)
(n)
(n)
Cn′ ,m′ (f |Fm ) = E[(f, Tn′ ,m′ )Tn′ ,m′ |Fm′ ],
f ∈ ℓ2 ,
and let
(n)
(n)
(n)
C n′ ,m′ (f |Fm ) = E[(f, Mn′ ,m′ )Mn′ ,m′ |Fm′ ],
f ∈ ℓ2 ,
19
INFERENCE FOR THE TRACE
be the conditional covariance operator associated to the martingale approximations. Obviously,
(n)
sup EkCn′ ,m′ (·|Fm′ ) − C (n) (·)k ≤ Ξn′ ,m′ + Ψn′ ,m′
n≥1
where
(n)
(n)
Ξn′ ,m′ = sup EkCn′ ,m′ (·|Fm′ ) − C n′ ,m′ (·|Fm′ )k,
n≥1
(n)
Ψn′ ,m′ = sup EkC n′ ,m′ (·|Fm′ ) − C (n) (·)k.
n≥1
We shall estimate both terms separately. To simplify notation, let
(n)
(n)
CT n (ν, µ) = E Tn′ ,m′ (ν)Tn′ ,m′ (µ) | Fm′ ,
(n)
(n)
CM n (ν, µ) = E Mn′ ,m′ (ν)Mn′ ,m′ (µ) | Fm′ ,
′ −θ/2 ),
for 1 ≤ ν, µ ≤ Ln . To estimate Ξn′ ,m′ , we shall show that |CT n (ν, µ)−CM n (ν, µ)| is O(L−1
n (n )
uniformly in n, ν, µ. By an application of the Cauchy-Schwarz inequality, we have
E sup |CT n (ν, µ) − CM n (ν, µ)|
1≤ν,µ
(n)
(n)
(n)
(n)
= E sup E Tn′ ,m′ (ν)Tn′ ,m′ (µ) | Fm′ − E Mn′ ,m′ (ν)Mn′ ,m′ (µ) | Fm′
1≤ν,µ
(n)
(n)
(n)
≤ E sup E |Tn′ ,m′ (ν) − Mn′ ,m′ (ν)||Tn′ ,m′ (µ)| Fm′
1≤ν,µ
(n)
(n)
(n)
+ E sup E |Mn′ ,m′ (ν)||Tn′ ,m′ (µ) − Mn′ ,m′ (µ)| Fm′ ,
1≤ν,µ
where
(n)
(n)
(n)
E sup E |Tn′ ,m′ (ν) − Mn′ ,m′ (ν)| Tn′ ,m′ (µ)| | Fm′
1≤ν,µ
r
r
(n)
(n)
(n)
≤ E sup E (Tn′ ,m′ (ν) − Mn′ ,m′ (ν))2 | Fm′
sup E|Tn′ ,m′ (ν)|2
1≤ν
1≤ν
n′ ,m′
′ −θ/2
≪ L−1
,
n (n )
(n)
(n)
(n)
since Tn′ ,m′ (µ) is independent from Fm′ and the decomposition Tn′ ,m′ (µ) = Mn′ ,m′ (µ) +
(n)
(n)
(n)
(Tn′ ,m′ (µ) − Mn′ ,m′ (µ)) leads to supn≥1 sup1≤ν E|Tn′ ,m′ (ν)|2 < ∞, by virtue of (2.42) and
Lemma 2.3. Further,
(n)
(n)
(n)
E |Mn′ ,m′ (ν)||Tn′ ,m′ (µ) − Mn′ ,m′ (µ)| Fm′
r
r
(n)
(n)
(n)
2
≤ E (Tn′ ,m′ (µ) − Mn′ ,m′ (µ)) | Fm′
E |Mn′ ,m′ (ν)|2 | Fm′
n′ ,m′
′ −θ/2
≪ L−1
,
n (n )
20
A. STELAND, R. VON SACHS
by (2.40) and Lemma 2.3. uniformly in ν, µ = 1, . . . , Ln . Consequently,
n′ ,m′
sup Ln E sup |CT n (ν, µ) − CM n (ν, µ)| ≪ (n′ )−θ/2
n≥1
1≤ν,µ
Hence, using the inequality
P Ln
1/2
ν=1 |fν | ≤ Ln
(n)
(n)
P
Ln
2
ν=1 fν
1/2
, we obtain
Ξn′ ,m′ ≤ sup EkCn′ ,m′ (·|Fm′ ) − C n′ ,m′ (·|Fm′ )k
n≥1
= sup E
n≥1
Ln X
Ln
X
sup
T
M
fν fµ (Cνµ
− Cνµ
)
f :kf kℓ2 =1 ν=1 µ=1
T
M
≤ sup E sup |Cνµ
− Cνµ
|
n≥1
1≤ν,µ
n′ ,m′
′ −θ/2
≪ L−1
n (n )
Ln
X
ν=1
Ln X
Ln
X
sup
f :kf kℓ2 =1 ν=1 µ=1
!2
|fν |
|fν fµ |
n′ ,m′
≪ (n′ )−θ/2 ,
(n)
By Lemma 2.2, see (2.39), and the scaling of the martingale approximations, Mn′ ,m′ (ν),
1 ≤ ν ≤ Ln , by the factor (Ln n′ )−1/2 ,
sup Ln
n≥1
max
1≤ν,µ≤Ln
(n)
n′ ,m′
(n)
2
′ −θ
kE[Mn′ ,m′ (ν)Mn′ ,m′ (µ)|Fm′ ] − L−1
n βn (ν, µ)kL1 ≪ (n ) .
Therefore
(n)
Ψn′ ,m′ = sup EkC n′ ,m′ (·|Fm ) − C (n) (·)k
n≥1
≤ sup E sup
n≥1
kf kℓ2 =1 ν=1 µ=1
≤ sup E sup
n≥1
Ln X
Ln
X
sup
kf kℓ2 =1 1≤ν,µ
M
2
fν fµ (Cνµ
− L−1
n βn (ν, µ))
M
|Cνµ
−
n′ ,m′
2
L−1
n βn (ν, µ)|
′ −θ
≪ sup L−1
sup Ln
n (n )
n≥1
kf kℓ2 =1
Ln
X
ν=1
Ln
X
ν=1
!2
|fν |
n′ ,m′
fν2 ≪ (n′ )−θ .
Q.E.D.
Proof of Theorem 2.3: By virtue of Lemma 2.1, Equation (2.20), we have the representations
X (n) Ln
X (n) Ln
1
Dnt =
ξk (j)
, Dn (t) = √
ξk (j)
,
j=1
j=1
nLn
k≤t
k≤⌊nt⌋
−1/2 (n)
and therefore we check conditions (I) – (III) of [15] discussed above for ζk = Ln ξk , cf.
(2.37). The summands can be seen as attaining values in the Euclidean space RLn of finite (but
21
INFERENCE FOR THE TRACE
increasing in n) dimension Ln or as random elements taking values in the infinite dimensional
Hilbert space ℓ2 .
To show (I) observe that by the Cr –inequality, for each 1 ≤ j ≤ Ln ,
(n)
E|ξk (j)|2+δ ≤ E(|Ynk (vn(j) )Ynk (wn(j) )| + E|Ynk (vn(j) )Ynk (wn(j) )|)2+δ
≤ 23+δ E|Ynk (vn(j) )Ynk (wn(j) )|2+δ ,
such that
(2.49)
(n)
E|ξk (j)|2+δ
3+δ
≤2
q
(j)
E|Ynk (vn )|4+2δ
q
(j)
E|Ynk (wn )|4+2δ .
Repeating the arguments of [11, p. 343], we obtain for δ′ ∈ (0, 2) with χ = δ′ /2
′
E|Ynk (vn(j) )|4+δ ≤ sup E|ǫk′ |4+δ
k ′ ≥0
′
∞
X
l=0
(v )
|cnℓj |2(2+χ)
X
1+χ X
1+χ
∞
∞
(v )
(v )
|cnℓj |2
,
|cnℓj |2
+ sup E(ǫ2k′ ) sup E(ǫ2k′ )
k ′ ≥1
k ′ ≥0
(v )
ℓ=0
(j)
ℓ′ =0
(ν)
(ν)
uniformly in k ≥ 1. But sup1≤j |cnℓj | ≤ kvn kℓ1 sup1≤ν |cnℓ | and sup1≤ν |cnℓ | ≤ (l∨1)−3/4−θ/2 ,
P
P∞ (vj ) 2(2+χ)
(vj ) 2
due to Assumption (A), imply sup1≤j ∞
< ∞.
ℓ=0 |cnℓ | < ∞ and, in turn,
l=0 |cnℓ |
Noting that the above bounds hold uniformly in k and n, we obtain
(n)
sup sup max E|ξk (j)|2+δ < ∞.
n≥1 k≥1 1≤j≤Ln
By virtue of Jensen’s inequality, we may now conclude that
(n)
EkLn−1/2 ξk k2+δ
ℓ2
1+δ/2
Ln
Ln
X
X
1
(n)
(n)
=E
[ξk (j)]2
E|ξk (j)|2+δ < ∞,
≤ L−1
n
Ln
j=1
j=1
which establishes (I).
Introduce the partial sums
(2.50)
(n)
Sn′ ,m′
1
=√ ′
n Ln
′ +n′
mX
(n)
ξk ,
k=m′ +1
n′ , m′ ≥ 1.
(n)
(n)
Condition (II) can be shown as follows. Denote the coordinates of Sn′ ,m′ by Sn′ ,m′ (j) and
Pm′ +n′
−1/2 (n)
(n)
ξk (j). Denote the corresponding
notice that they are given by Sn′ ,m′ (j) = k=m
′ +1 Ln
(n)
(n)
(n)
(n)
martingale approximations by Mn′ ,m′ and Mn′ ,m′ (j), respectively, and let Rn′ ,m′ = Sn′ ,m′ −
(n)
(n)
Mn′ ,m′ be the remainder with coordinates Rn′ ,m′ (j), 1 ≤ j ≤ Ln , cf. the preparations above.
(n)
(n)
Clearly, the martingale property implies E(Sn′ ,m′ (j)|Fm′ ) = E(Rn′ ,m′ (j)|Fm′ ), 1 ≤ j ≤ Ln .
Lemma 2.3 asserts that
"
#
n′ ,m′
(n)
2
sup Ln E sup E (Rn′ ,m′ (j)) | Fm′
≪ (n′ )−θ ,
n≥1
1≤j
22
A. STELAND, R. VON SACHS
such that two applications of Jensen’s inequality lead to
v
u Ln h
i2
uX
(n)
(n)
sup E E(Sn′ ,m′ | Fm′ ) ≤ sup t
E E(Rn′ ,m′ (j) | Fm′ )
ℓ2
n≥1
n≥1
j=1
v
u Ln h
i
uX
(n)
E E (Rn′ ,m′ (j))2 | Fm′
≤ sup t
n≥1
j=1
v
"
#
u
u
(n)
t
≤ sup Ln E sup E (Rn′ ,m′ (j))2 | Fm′
n≥1
1≤j
n′ ,m′
≪ (n′ )−θ/2 ,
which shows (II). Condition (III) follows from Theorem 2.2.
Consequently, we may conclude that we may redefine all processes on a rich enough probability space where a Brownian motion Bn (t) = (Bn (t)j )j with covariance operator C (n) , i.e.
2
with covariances E(Bn (t)j Bn (t)k ) = L−1
n βn (j, k), exists, such that for constants λn > 0 and
Cn < ∞
kLn−1/2 Dn,t (vn , wn ) − Bn (t)kRLn ≤ Cn t1/2−λn ,
for all t ≥ 0,
a.s.. Therefore
sup kDn (t) − B n (⌊nt⌋/n)kRLn ≤ Cn n−λn ,
t∈[0,1]
a.s., for the [0, 1]–version B n of Bn , which implies assertions (i) and (ii).
−1/2
To show (iii) recall that the vector 1–norm of RLn can be bounded by Ln k · kRLn n , such
that
Ln
X
j=1
1/2
|Dnj (t) − B n (⌊nt⌋/n)j | ≤ L1/2
n kDn (t) − B n (⌊nt⌋/n)kℓ2 = o(Ln ),
as n → ∞, a.s. Further, using |xj | ≤
qP
j
x2j , we have
sup |Dn (t; vn(j) , wn(j) ) − B n (⌊nt⌋/n)j | ≤ sup kDn (t) − B n (⌊nt⌋/n)kℓ2 = o(1),
t∈[0,1]
t∈[0,1]
as n → ∞, a.s., for j = 1, . . . , Ln .
It remains to prove (iv). We may argue as in [17] to obtain
X
(σ) ′
(σ) ′
b nmk − Σnmk )w(σ) =
vn(ρ) ′ (Σ
[(vn(ρ) ′ Yni )(wm
Ymi ) − E((vn(ρ) ′ Yni )(wm
Ymi ))]
m
i≤k
=
X
i≤k
(σ)
(σ)
[Yni (vn(ρ) )′ Ymi (wm
) − E(Yni (vn(ρ) )′ Ymi (wm
))]
where
Yni (vn(ρ) ) =
dn
∞ X
X
j=0 ν=1
(ν)
(ρ)
vnν
cnj ǫi−j
and
(σ)
Ymi (wm
)=
dn
∞ X
X
j=0 µ=1
(µ)
(σ)
wmµ
cmj ǫi−j ,
23
INFERENCE FOR THE TRACE
for ρ = 1, . . . , Ln and σ = 1, . . . , Lm , n, m ≥ 1. Therefore, for k ≥ 1, we obtain the representation
Dk ({(vn(ρ) , wn(ρ) )})
XXX
(σ)
(σ)
λn λm µρ µσ [Yni (vn(ρ) )′ Ymi (wm
) − E(Yni (vn(ρ) )′ Ymi (wm
))]
=
i≤k n,m ρ,σ
X
=
[Yi ({cj })Yi ({dj }) − E(Yi ({cj })Yi ({dj }))]
i≤k
for the linear processes Yi ({cj }) =
cj =
∞
X
λn
dj =
µρ
λn
Ln
X
µρ
ρ=1
n=1
dn
X
j=0 cj ǫi−j
(ν)
(ρ)
vnν
cnj ,
dn
X
and Yi ({dj }) =
P∞
j=0 dj ǫi−j
with coefficients
j ≥ 0,
ν=1
ρ=1
n=1
∞
X
Ln
X
P∞
(ν)
(ρ)
wnν
cnj ,
j ≥ 0.
ν=1
Hence the result follows from [11].
Proof of Remark 2.1:
ability space,
Q.E.D.
By Theorem 2.1, we may and will assume that, on the same prob-
|Dn (t; vn(j) , wn(j) ) − αn (vn(j) , wn(j) )Wn (⌊nt⌋/n; vn(j) , wn(j) )| = o(1),
(j)
(j)
as n → ∞, a.s., for Ln standard Brownian motions Wn (⌊nt⌋/n; vn , wn ), by virtue of The−1/2
(j)
(j)
orem 2.1. But then, since Dnj (t) = Ln Dn (t; vn , wn ),
αn (vn(j) , wn(j) )Wn (⌊nt⌋/n; vn(j) , wn(j) )
sup B n (⌊nt⌋/n)j − L−1/2
n
t∈[0,1]
≤ sup |B n (⌊nt⌋/n)j − Dnj (t)|
t∈[0,1]
+ sup L−1/2
|Dn (t; vn(j) , wn(j) ) − αn (vn(j) , wn(j) )Wn (⌊nt⌋/n; vn(j) , wn(j) )|
n
t∈[0,1]
= o(1),
as n → ∞, a.s., for each j = 1, . . . , Ln , which verifies the remark.
Q.E.D.
Proof of Theorem 2.4: Observe that the conditions (I)-(III) of [15, Theorem 1] hold in
the Hilbert space ℓ2 as well, since for any x ∈ Rdn the Euclidean vector norm coincides with
the ℓ2 –norm. Therefore, we obtain the a.s. strong approximation
k
X
(n)
L−1/2
ξi
n
i=1
− Bn (k)
= ǫnk
ℓ2
p
k log log k,
as k → ∞, for sequences εnk = o(1), k → ∞, a.s., n ≥ 1. Put N = ⌈n log log n⌉. Let ηn ↓ 0 be
given. Then for each n ∈ N we may find δn ∈ N such that P (maxδn ≤k′ ǫnk′ > ε) ≤ ηn . Hence
maxδn ≤k′ ǫnk′ = oP (1), as n → ∞. Now we may conclude that for δn ≤ k ≤ n
k
X
i=1
(n)
L−1/2
ξi
n
√
− Bn (k) ≤ max′ εnk′ N ,
δn ≤k
24
A. STELAND, R. VON SACHS
such that
k
1 X −1/2 (n)
Ln ξi − Bn (k) = oP (1),
max √
δn ≤k≤n
N i=1
as n → ∞, which verifies (2.31)–(2.33).
Q.E.D.
3. ASYMPTOTICS FOR THE TRACE NORM
The trace plays an important role in multivariate analysis and also arises when studying
shrinkage estimation. Before providing the large sample approximation by a Brownian motion,
we shall briefly review its relation to several matrix norms.
3.1. The trace and related matrix norms
There are various matrix norms that can be used to measure the size of (covariance) matrices. Here we shall use the trace norm defined as the ℓ1 -norm of the eigenvalues λi (A) of a
dn -dimensional matrix A,
X
kAktr =
|λi (A)|.
i
Also notice that the trace norm is a linear mapping on the subspace of non-negative definite
matrices and satisfies kAktr = tr(A) for any covariance matrix A. It induces the Frobenius
norm via kAk2F = tr(AA′ ). Further, it is worth mentioning that the trace norm is also related
to the Frobenius norm via the fact
X
λi (A) = kAktr .
kA1/2 k2F =
i
In this way, our results formulated in terms of (scaled) trace norms can be interpreted in
terms of (scaled) squared Frobenius norms of square roots, too.
There is a third interesting direct link to another family of norms, namely the Schatten-p
norms kAkS,p , p ≥ 0, of a n × m matrix A of rank r, which is defined as the ℓp -norm of its
(non-negative) singular values σ1 (A) ≥ · · · ≥ σr (A) > σr+1 (A) = · · · = σn (A) = 0, i.e. of
the eigenvalues of |A| = (AA′ )1/2 , i.e.
X
kAkpS,p =
σi (A)p .
i
b n = AA′ , if A = Yn /√n,
The Schatten-1 norm kAkS,1 is also called nuclear norm. Since Σ
b n ) = σi (Yn /√n)2 , we have the identity
such that λi (Σ
X
√
b n ktr =
b n ) = kYn / nk2 .
kΣ
λi (Σ
S,2
i
between the trace norm of the sample covariance matrix and the Schatten-2 norm of the
scaled data matrix.
For a sequence {An } of matrices of growing dimension dn × dn it makes sense to attach a
scalar weight depending on the dimension to a given norm, such that simple matrices such as
INFERENCE FOR THE TRACE
25
the identity matrix receive bounded norms. Having in mind that the squared Frobenius norm
of An is the trace of An A′n , it is natural to attach a scalar weight f (dn ) to the trace operator
leading to the scalar weight f (dn )1/2 for the Frobenius norm. As proposed by [13], one may
select f (dn ) such that tr(A∗ )f (dn ) = 1 for some simple benchmark matrix A∗ such as the
dn –dimensional identity matrix In . Since tr(Idn ) = dn , we choose f (dn ) = d−1
n and therefore
define the scaled trace operator by
∗
tr (A) =
d−1
n
dn
X
λi (A)
i=1
for a square matrix A = (aij )i,j of dimension dn × dn . The scaled trace operator induces the
scaled trace norm
kAk∗tr = d−1
n kAktr
for a square matrix A, which is given by kAk∗tr = d−1
n tr(A) for a covariance matrix and
averages the (modulus) of the eigenvalues, and the scaled Frobenius matrix norm given by
(3.1)
2
∗
′
−1
kAk∗2
F = tr (AA ) = dn kAkF .
3.2. Trace asymptotics
Let us now turn to the trace asymptotics. If the dimension is fixed, it is well known that the
eigenvalues of a sample covariance matrix, and thus their sum as well, have convergence rate
OP (n−1/2 ) and are asymptotically normal, see [9] and [10]. For the high-dimensional case,
the situation is more involved. The sample covariance matrix is not consistent w.r.t. to the
Frobenius norm, even in the presence of a dimension reducing factor model, see Remark 3.1.
The following result provides a large sample normal approximation for the scaled trace
b n for arbitrarily growing dimension dn when properly normalized. The result also
norm of Σ
shows that the trace norm has convergence rate
b n ktr − kΣn ktr = OP (n−1/2 dn ),
kΣ
as n → ∞.
Introduce for t ∈ [0, 1]
⌊nt⌋
(3.2)
X
′
b n (t) = 1
Yni Yni
,
Σ
n
i=1
and notice that
bn = Σ
b n (1)
Σ
and
b n (t)),
Σn (t) = E(Σ
Σn = Σn (1).
We are interested in studying the scaled trace norm process
(3.3)
Tn (t) =
√
b n (t)k∗ − kΣn (t)k∗ ,
n kΣ
tr
tr
t ∈ [0, 1].
26
A. STELAND, R. VON SACHS
Theorem 3.1 Let Yni , i = 1, . . . , n, be a vector time series following model (1.1) and
satisfying Assumption (A). If (2.26) holds, then under the construction of Theorem 2.3,
sup Tn (t) − d−1/2
n
t∈[0,1]
dn
X
B n (⌊nt⌋/n)j = o(1),
j=1
as n → ∞, a.s. Here B n denotes the [0, 1]–version of the Brownian motion Bn arising in
(j)
(j)
Theorem 2.3, when choosing the dn pairs (vn , wn ) = (ej , ej ), j = 1, . . . , dn , where ej
denotes the jth unit vector, and satisfies properties (ii) and (iii) of Theorem 2.3.
Suppose that, in addition to the assumptions of Theorem 3.1, Yn1 , . . . , Ynn is strictly
stationary. Since the weighting vectors used in Theorem 3.1 are the first dn unit vectors, the
covariance of B n (1)i and B n (1)j , which is associated to the asymptotic covariance of Dni (1)i
2
2
2
and Dnj (1)j , is given by d−1
n βn (i, j) where βn (i, j) = βn (ei , ei , ej , ej ), i, j = 1, . . . , dn , cf.
(A.9). We have the asymptotic representations
√
√
b n − Σn )ei , ne′ (Σ
b n − Σn )ej + o(1)
ne′i (Σ
βn2 (i, j) = Cov
j
!
n
n
X
1 X (i) 2
1
(j)
= Cov √
[(Ynk ) − σ 2 ], √
[(Ynk )2 − σ 2 ] + o(1)
n
n
k=1
k=1
n
1 X
(j)
(i)
=
E (Ynk )2 − σ 2 (Ynk′ )2 − σ 2 + o(1).
n ′
k,k =1
Therefore, up to negligible terms, we may express βn2 (i, j) as a long–run variance parameter,
(3.4)
βn2 (i, j) = γn(i,j) (0) + 2
n−1
X
τ =1
where
(3.5)
n − τ (i,j)
γn (τ ) + o(1),
n
(i)
(j) 2
γn(i,j) (τ ) = Cov (Yn0 )2 , (Ynτ
) ,
τ = 0, . . . , n − 1,
(i)
(j)
is the lag τ cross–covariance of the series {(Ynk )2 : k ≥ 0} and {(Ynk )2 : k ≥ 0}, i, j =
1, . . . , dn . Those cross-covariances can be estimated by
γn(i,j) (τ )
b
n−τ
1X 2
2
(ej ) − µ
bn (j)]
bn (i)][Yk+τ
[Yk (ei ) − µ
=
n
k=1
Pn
−1
2
′
with µ
bn (i) = n
k=1 Yk (ei ), where Yk (ei ) = ei Ynk , for k = 1, . . . , n, i = 1, . . . , dn , n ≥ 1.
The associated estimator for βn2 (i, j) is then given by
b n(i,j) (0) + 2
βbn2 (i, j) = γ
m
X
τ =1
b n(i,j) (τ ) ,
wmτ γ
where m = mn is a sequence of lag truncation constants and {wmh } a sequence of window
weights typically defined by a kernel function w (e.g. a Bartlett kernel) via wmτ = w(τ /bm ),
for some bandwidth parameter b = bm .
INFERENCE FOR THE TRACE
27
By Theorem 3.1,
2
Var (Tn (1)) = σtr,n
+ o(1),
a.s., with
(3.6)
2
σtr,n
dn
1 X
=
Cov (B n (1)j , B n (1)k ),
dn
j,k=1
and, using the canonical estimator
(3.7)
2
σ
btr,n
=
dn
1 X
βbn2 (i, j),
d2n
j,k=1
an asymptotic confidence interval with nominal coverage probability 1 − α, α ∈ (0, 1), for
kΣn k∗tr is given by
√
b n k∗tr ± Φ−1 (1 − α/2)b
kΣ
σtr,n / n.
Lemma 3.1 Assume wmτ → 1 as m → ∞, for all τ ∈ Z, and 0 ≤ wmτ ≤ W < ∞, for some
(ν)
(ν)
constant W , for all m ≥ 1, τ ∈ Z. Further, suppose that cnj = cj , ν ∈ N, satisfy the decay
condition
(ν)
sup |cj | ≪ (j ∨ 1)−(1+δ)
1≤ν
for some δ > 0, and ǫk are i.i.d. with E(ǫ81 ) < ∞. If m = mn → ∞ with m2 /n = o(1), as
n → ∞, then
2
2
| = 0.
lim E|b
σn,tr
− σn,tr
n→∞
Remark 3.1 It is worth comparing our result with the following result obtained by [6] for a
dimension-reducing factor model: Suppose that the generic random vector Yn = (Y1 , . . . , Ydn )′
satisfies a factor model
Yn = Bn f + ǫ
with K = K(dn ) ≤ dn observable factors f = (f1 , . . . , fK )′ , errors ǫ = (ǫ1 , . . . , ǫdn )′ and
a dn × K factor loading matrix Bn . Then the sample covariance matrix of an i.i.d. sample
(Y1 , f1 ), . . . , (Yn , fn ) has the convergence rate OP (n−1/2 dn K),
b n − Σn kF = OP (n−1/2 dn K),
kΣ
if EkYn k2F , maxi E(fi4 ) and maxi E(ǫ4i ) are bounded, see [6, Theorem 1]. This means, compared to the rate for fixed dimension, the Frobenius norm is inflated by the factor dn K.
28
A. STELAND, R. VON SACHS
3.3. Proofs
b
b
Proof of Theorem 3.1: Clearly,
P ′ Σk (1) is p.s.d. for all k ∈ N and thus Σn (t) =
as well. The fact that tr(A) = i ei Aei leads to
b n (t)ktr − kΣn (t)ktr = tr(Σ
b n (t)) − tr(Σn (t)) =
kΣ
dn
X
j=1
−1/2
⌊nt⌋ b
n Σ⌊nt⌋
b n (t) − Σn (t))ej .
e′j (Σ
n
Let Dn = (Dnj )dj=1
with Dnj (t) = dn Dn (t; ej , ej ) for j = 1, . . . , dn . We shall apply Theorem 2.3 with Ln = dn . Therefore, when redefining all processes on a new probability space
together with a dn -dimensional Brownian motion B n with covariances as described in Theorem 2.3, we may argue as follows. Since k · k∗tr = d−1
n tr(·), we have
dn
1 X
Tn (t) = √
Dnj (t).
dn j=1
Now we can conclude that the process
En (t) =
satisfies
√
b n (t)k∗ − kΣn (t)k∗ ) −
n(kΣ
tr
tr
dn
X
d−1/2
B n (⌊nt⌋/n)j
n
j=1
dn
1 X
Dnj (t) − B n (⌊nt⌋/n)j
|En (t)| = √
dn j=1
dn
1 X
≤√
Dnj (t) − B n (⌊nt⌋/n)j
dn i=1
= o(1),
as n → ∞, a.s., by Theorem 2.3 (iv).
Q.E.D.
Proof of Lemma 3.1: The proof follows easily from [17, Theorem 4.4] by noting that the
covariances of the coordinates of the Brownian motion are given by d−1
n βn (i, j), for 1 ≤ i, j ≤
dn .
Q.E.D.
4. SHRINKAGE ESTIMATION
Shrinkage is a well-established approach to regularize the sample variance–covariance matrix
and we shall review in Section 4.1 the results obtained for high–dimensional settings. When
shrinking towards the identity matrix in terms of a convex combination with the sample
variance–covariance matrix, the optimal weight depends on the trace of the true variance–
covariance matrix, which can be estimated canonically by the trace of the sample variance–
covariance matrix. As a consequence, we can apply the results obtained in the previous section
to obtain large sample approximations for shrinkage variance–covariance matrix estimators.
Recall that the approximations deal with the norm of the difference between partial sums and
INFERENCE FOR THE TRACE
29
a Brownian motion, both attaining values in a vector space. In order to compare covariance
matrices, we shall work with the following pseudometric: Define
(4.1)
∆n (An , Bn ; vn , wn ) = |vn′ (An − Bn )wn |,
for (sequences of) matrices An and Bn of dimension dn ×dn . Indeed, for fixed vn , wn the mapping (An , Bn ) 7→ ∆n (An , Bn ) = ∆n (An , Bn ; vn , wn ) is symmetric, non-negative, semidefinite
(i.e. An = Bn implies ∆n (An , Bn ) = 0) and satisfies the triangle inequality. Hence, (4.1) defines a pseudometric on the space of (dn × dn )–dimensional matrices, for each n.
We establish three main results: For regular weighting vectors vn , wn that are bounded
away from orthogonality we establish a large sample approximation, which holds uniformly
in the shrinkage weight and therefore also when using the common estimator for the optimal
weight. Further, we compare the shrinkage estimator using the estimated optimal weight
with an oracle estimator using the unknown optimal weight. In both cases, it turns out that
the convergence rate of the estimated optimal shrinkage weight carries over to the shrinkage
covariance estimator.
Lastly, we study the case of orthogonal and nearly orthogonal vectors. The latter case
is of particular interest, since then one may place more unit vectors on the unit sphere
corresponding to overcomplete bases as studied in areas such as dictionary learning.
4.1. Shrinkage of covariance matrix estimators
The results of the previous chapters show that, under general conditions, inference relying on
ℓ1 -bounded inner products of high-dimensional series can be based on the sample covariance
b n is singular. However, from a statistical point of view, the
matrix, even if dn > n such that Σ
use of this classical estimator is not recommended in such situations of high dimensionality:
important criteria such as its mean-squared error or its condition number (defined to be the
ratio of the largest to the smallest eigenvalue) deteriorate, and it is advisable to regularise
b n in order to improve its performance, both asymptotically and for finite sample sizes, with
Σ
respect to these criteria. Obviously, a particular interest lies in maximum-likelihood based
approaches using an invertible estimator of Σn (as, e.g. in the semi-parametric approach
of [7] on shrinkage estimation in multivariate hidden Markov models). One well-established
b n without needing to impose any structural assumptions on Σn ,
possibility to regularise Σ
in particular avoiding sparsity, is the following approach of shrinkage: ([13], [16]) consider a
b n with a well-conditioned
shrinkage estimator defined by a linear (or convex) combination of Σ
”target” matrix Tn ,
(4.2)
b n + Wn T n ,
Σsn = Σsn (Wn ) = (1 − Wn )Σ
where Wn are the ”shrinkage weights” of this convex combination, to be chosen in an optimal
way to minimise the mean-square error between Σsn and Σn (see below). The role of the target
Tn is, similar to ridge regression, to reduce a potentially large condition number of the highdimensional variance-covariance matrix Σn , by adding a highly regular (”well conditioned”)
matrix. A popular choice for the target is to take a multiple of the dn −dimensional identity
matrix In , i.e.
(4.3)
T n = µ n In ,
with µn = d1n tr Σn , in order to respect the scale of both matrices in the convex combination
(4.2). This choice of the target reduces the dispersion of the eigenvalues of Σn around its
30
A. STELAND, R. VON SACHS
”grand mean” µn = d1n tr Σn , as large eigenvalues are pulled down towards µn and small
eigenvalues are lifted up to µn (and in particular lifted up away from zero). Although a bias is
b n , the gain in variance reduction, in parintroduced in estimating Σn by Σsn compared to Σ
ticular in high-dimensions, helps to considerably reduce the mean-square error in estimating
Σn .
In order to develop the correct asymptotic framework of the behaviour of large covariance
matrices, the authors of [13] propose to use the scaled Frobenius norm given by (3.1) to
measure the distance between two matrices of asymptotically growing dimension dn , to be
used also and in particular to define the mean-square error between Σsn and Σn to become
the expected normalised Frobenius loss E[kΣsn − Σn k∗2
F ].
Furthermore, with this scaling,
µn =
1
tr(Σn )
dn
is the appropriate choice of the factor in front of the identity matrix In in the definition of
the target Tn in equation (4.3).
b n , i.e. by
In practice µn needs to be estimated from the trace of Σ
µ
bn =
1
b n) .
tr(Σ
dn
Similarly, the theoretical shrinkage weight 0 ≤ Wn ≤ 1 need to be replaced by its sample
cn . Thus, the fully data-driven expression for the shrinkage estimator of Σ writes as
analog W
follows:
bn + W
cn ) = (1 − W
cn )Σ
cn µ
Σsn (W
bn In ,
which shrinks the sample covariance matrix towards the (estimated) shrinkage target µ
bn In . It
cn ) with
remains to optimally choose the shrinkage weights Wn (and its data-driven analogue W
the purpose of balancing between a good fit and good regularisation. For this a prominent
possibility is indeed to choose the shrinkage weights Wn such that the mean-squared error
(MSE) between Σsn and Σn , is minimised:
Wn∗ = argminWn ∈[0,1] E[kΣsn (Wn ) − Σn k∗2
F ] ,
which leads to the MSE-optimally shrunken matrix Σ∗n = Σsn (Wn∗ ). A closed form solution
([13] or [16], Proposition 1) can be derived as
Wn∗ =
b n − Σn k∗2 ]
E[kΣ
F
b
E[kµn In − Σn k∗2 ]
F
This choice leads to the interesting property that
∗2
b
E[kΣ∗n − Σn k∗2
F < E[kΣn − Σn kF ,
showing the actual relative gain of the shrunken estimator compared to the classical unshrunken sample covariance, in terms of the mean-squared error. Moreover, it can be shown
that this property continues to hold even if one replaces the in practice yet unknown optimal
c ∗ which is constructed by replacing the population quantities
weights Wn∗ by an estimator W
n
31
INFERENCE FOR THE TRACE
in numerator and denominator of Wn∗ by sample analogs. Whereas the denominator can be
b n k∗2 , it is slightly less straightforward to estimate the nuessentially estimated by kb
µ n In − Σ
F
b n − Σn k2 : one possibility suggested by [16], and further developed by [17] for
merator E[kΣ
F
b n.
our set-up, is based on the estimation of the long-run variance αn of Σ
Note that
dn
√
1 X
∗2
b n (i, j)),
b
EkΣn − Σn kF =
Var ( nΣ
ndn
i,j=1
where, under stationarity, for 1 ≤ i, j ≤ dn ,
σn2 (i, j)
n
X (i) (j)
(i) (j)
b n (i, j) = 1
Σ
[Ynk Ynk − E(Ynk Ynk )].
n
k=1
n−1
X
√
(i,j)
b
(n − τ )Γn(i,j) (τ )
= Var ( nΣn (i, j)) = Γn (0) + 2
τ =1
(i,j)
Γn (τ )
with
of the
(i) (j)
(i)
(j)
= Cov (Ynk Ynk , Yn,k+τ Yn,k+τ ), τ = 0, . . . , n − 1, n
optimal weights Wn∗ can now be obtained as follows. Let
≥ 1. A consistent estimator
n−τ
X (i) (j)
(j)
(i)
b n(i,j) (τ ) = 1
Γ
bn (i, j)] ,
bn (i, j)] [Yn,t+τ Yn,t+τ − κ
[Ynt Ynt − κ
n
t=1
Pn
(i)
(j)
where κ
bn (i, j) = n1 t=1 Ynt Ynt , 1 ≤ i, j ≤ dn . Then, similar as in the previous section, the
long-run variances σn2 (i, j) can be estimated by
b (i,j) (0) + 2
σ
bn2 (i, j) = Γ
n
m
X
τ =1
b (i,j) (τ ) ,
wmτ Γ
n
1 ≤ i, j ≤ dn . Consistency of a more general version has been shown in [17, Equation (4.22)],
under similar assumptions as stated in Lemma 3.1, for dn → ∞, as n → ∞. We are led to the
estimator
P
bn2 (i, j)
(ndn )−1 1≤i,j≤dn σ
∗
c =
W
n
b n k∗2
kb
µ n In − Σ
F
for Wn∗ also studied in depth in [16]. A rate of consistency in an asymptotic framework with
growing dimensionality dn can be achieved again following [16] for the specific shrinkage target
µn In , also considered in [7]. Let 0 < γ ≤ 2 be such that d2−γ
n /n → 0 (i.e. the larger γ, the
faster is dn allowed to grow with n), and that, as n → ∞,
∗2
d1−γ
n kµn In − Σn kF → c > 0 .
Recalling that µn = d−1
n tr(Σn ), we observe that γ measures the closeness of the target to the
true covariance matrix Σn . Then [16] (Theorem 1) and ([7] (Theorem 1) show that
n c∗
∗
(4.4)
2−γ |Wn − Wn | = oP (1) .
dn
In order to apply the results of the previous sections onto the fully data-driven shrinkage
cn∗ ), one needs to study the convergence of the estimated shrinkage weight
estimator Σ∗n (W
normalised by n1/2 , as will become clear from the proof of Theorem 4.1 to be stated below.
We already observe here that (4.4) implies
(4.5)
c ∗ − W ∗ | = oP (1),
n1/2 |W
n
n
if d4−2γ /n = O(1). Thus, for γ close to 2, the dimension dn may even grow faster than n.
32
A. STELAND, R. VON SACHS
4.2. Asymptotics for regular projections
Our interest is now in deriving the asymptotics for bilinear forms based on the shrinkage
estimator of the covariance matrix. We can and will assume that the uniformly ℓ1 –bounded
weighting vectors vn and wn are ℓ2 –normed. It turns out that, due to the shrinkage target,
the inner product vn′ wn , i.e. the angle between the vectors vn and wn , appears in the approximating Brownian functional. The inner product is bounded but may converge to 0, as n
tends to ∞. The latter case requires special treatment and will be studied separately. We shall
call a pair (vn , wn ) of projections regular, if it has uniformly bounded ℓ1 –norm and satisfies
(4.6)
vn′ wn ≥ c > 0,
for all n,
for some constant c, i.e., if it is, in addition, bounded away from orthogonality.
Let 0 < W ≤ 1 be an arbitrary shrinkage weight and consider the associated shrinkage
estimator
b s (W ) = (1 − W )Σ
bn + Wµ
Σ
b n In .
n
b s (W ) estimates the unobservable shrunken variance matrix
Notice that Σ
n
Σsn0 (W ) = (1 − W )Σn + W tr(Σn )d−1
n In .
Define for 0 < W ≤ 1
(4.7)
An (W ) =
√ ′
b s (W ) − Σs (W ))wn .
nvn (Σ
n
n0
We shall apply the trace asymptotics obtained in Theorem 3.1, i.e.
dn
X
√
b n (t)k∗ − kΣn k∗ (t)) = d−1/2
n(kΣ
B n (⌊nt⌋/n)j + o(1),
tr
tr
n
j=1
as n → ∞, a.s. The variance of the approximating linear functional of the dn –dimensional
Brownian motion is given by
2
σtr,n
dn
dn
1 X
1 X
βn2 (i, j),
Cov (B n (1)i , B n (1)j ) = 2
=
dn
dn
i,j=1
i,j=1
where βn2 (i, j) are long–run–variance parameters, see (3.4). Since, typically, long–run–variance
parameters have positive limits, it is natural to assume that
(4.8)
2
inf σtr,n
≥ σ 2tr > 0.
n≥1
Theorem 4.1 Let (vn , wn ) be a regular pair of projections. Under the assumptions of Theorem 3.1 and condition (4.8) there exists on a new probability space, which carries an equivalent
n
on [0, 1], such that
version of the vector time series, a Brownian motion B n (t) = (B n (t)j )dj=0
(4.9)
sup |An (W ) − Bn (W )| = o(1),
0<W ≤1
33
INFERENCE FOR THE TRACE
as n → ∞, a.s., where
(4.10)
Bn (W ) = (1 − W )B n (1)0 +
W vn′ wn d−1/2
n
dn
X
B n (1)j ,
j=1
The covariance structure of B n (t) is given by
Var (B n (1)0 ) = (dn +1)−1 α2n (vn , wn ),
Cov (B n (1)0 , B n (1)i ) = (dn +1)−1 βn2 (vn , wn , ei , ei ),
for i = 1, . . . , dn , and
Cov (B n (1)i , B n (1)j ) = (dn + 1)−1 βn2 (i, j),
for 1 ≤ i, j ≤ dn . Especially, for any deterministic or random sequence of shrinkage weights
Wn we have the large sample approximation for the corresponding shrinkage estimator
√
b s (Wn ) − Σs (Wn ))wn − Bn (Wn )| = o(1),
| nvn′ (Σ
n
n0
as n → ∞, a.s..
Notice that
Var (Bn (W )) =
dn
W 2 (vn′ wn )2 X
(1 − W )2 2
αn (vn , wn ) +
βn2 (i, j)
dn + 1
dn (dn + 1)
i,j=1
dn
(1 − W )W vn′ wn X
βn2 (vn , wn , ej , ej ) + o(1),
+2 p
dn (dn + 1) j=1
as n → ∞. Hence, under assumption (4.6), the variance of the approximating Wiener process
adressing the nonparametric part of the shrinkage estimator is of the order O(d−1
n ), whereas
the variance of the term approximating the target is of the order O(1). This is due to the fact
that we need a (dn + 1)-dimensional Brownian motion (from which dn coordinates are used
to approximate the estimated target). This requires to scale all coordinates (dn + 1)−1/2 , cf.
Theorem 2.3.
The following theorem resolves that issue by approximating the shrinkage estimator by
two Brownian motions, one in dimension 1 for the nonparametric part and one in dimension
dn for the target. Those Brownian motions are constructed separately, such that, a priori,
nothing can be said about their exact covariance structure. It turns out, however, that the
covariances converge properly. We shall see that for this alternative construction the terms of
the resulting variance-covariance decomposition are of the same order.
Theorem 4.2 Let (vn , wn ) be a regular pair of projections. Suppose that the underlying
probability space (Ω, F, P ) is rich enough to carry, in addition to the vector time series {Yni :
1 ≤ i ≤ n, n ≥ 1}, a uniform random variable U1 . Then there exist, on (Ω, F, P ), a univariate
′
Brownian motion {B n (t)0 : t ∈ [0, 1]} with mean zero and
′
′
Cov (B n (s)0 , B n (t)0 ) = min(s, t)α2n (vn ; wn ),
34
A. STELAND, R. VON SACHS
′
n
for s, t ∈ [0, 1], and a mean zero Brownian motion {(B n (t)j )dj=1
: t ∈ [0, 1]} in dimension dn
with covariance function
′
′
2
Cov (B n (s)i , B n (t)j ) = min(s, t)d−1
n βn (i, j),
for s, t ∈ [0, 1] and 1 ≤ i, j ≤ dn , such that
√
b s (W ) − Σs (W ))wn − Bn (W ) = o(1),
nvn′ (Σ
n
n0
′
′
−1/2 Pdn
j=1 B n (1)j .
as n → ∞, a.s., with Bn′ (W ) = (1 − W )B n (1)0 + W vn′ wn dn
(4.11)
′
′
Further,
max |Cov (B n (1)0 , B n (1)j ) − d−1/2
βn2 (vn , wn , ej , ej ) = o(1),
n
1≤j≤dn
as n → ∞.
Observe that
Var (Bn′ (W ))
= (1 −
W )2 α2n (vn , wn )
+2
dn
W 2 (vn′ wn )2 X
βn2 (i, j)
+
d2n
i,j=1
(1 − W )W vn′ wn
dn
dn
X
βn2 (vn , wn , ej , ej ) + o(1),
j=1
as n → ∞, a.s., where all three terms are O(1).
The above result shows that the nonparametric part, namely the sample covariance matrix
b n , as well as the shrinkage target µ
Σ
bn In contribute to the asymptotics. In this sense, shrinking
with respect to the chosen scaled norms provides us with a large sample approximation that
mimics the finite sample situation.
4.3. Comparisons with oracle estimators
Recall that an oracle estimator is an estimator that depends on quantities unknown to us
such as the optimal shrinkage weight Wn∗ . Of course, it is of interest to study the distance
b s (W
c ∗ ) with estimated optimal weight and the associated
between the shrinkage estimator Σ
n
n
∗
oracle using Wn . In particular, the question arises how the rate of convergence (4.4) affects
the difference between the fully data adaptive estimator and an oracle.
b sn = Σ
b sn (W
cn∗ ) that uses the estimated
The next theorem compares the shrinkage estimator Σ
c ∗ and the oracle estimator
optimal shrinkage weight W
n
b s (W ∗ ) = (1 − W ∗ )Σ
b n + W ∗µ
Σ
n
n
n
n b n In ,
which shrinks the sample covariance matrix towards the target using the optimal shrinkage
weight Wn∗ , in terms of the pseudometric ∆n (·, ·; vn , wn ) and thus considers the quantity
b s (W
b s (W ∗ ); vn , wn ) = v′ (Σ
b s (W
b s (W ∗ ))wn .
cn∗ ), Σ
cn∗ ) − Σ
∆n (Σ
n
n
n
n
n
n
n
The following result shows that even now the rate of convergence is equal to the rate of
c ∗.
convergence of the estimator W
n
35
INFERENCE FOR THE TRACE
Theorem 4.3 Under the assumptions of Theorem 4.1 and the construction described there
we have, on the new probability space,
b s (W
c ∗ ), Σ
b s (W ∗ ); vn , wn )
∆n (Σ
n
n
n
n
!
dn
X
∗
∗
−1/2
−1/2
−1/2 −1/2
c
= |Wn − Wn | O(1) + O(B n (1)0 n
),
)+O n
dn
B n (1)i + o(n
i=1
as n → ∞, a.s..
The next result investigates the difference between the shrinkage estimator and the oracle
type estimator
Σsn0 (Wn∗ ) = (1 − Wn∗ )Σn + Wn∗ tr(Σn )d−1
n In
using the oracle shrinkage weight and assuming knowledge of Σn , in terms of the pseudodistance
c ∗ ) − Σs (W ∗ )]wn .
b s (W
c ∗ ), Σs (W ∗ ); vn , wn ) = v′ [Σ
b s (W
∆n (Σ
n
n0
n
n
n
n
n
n
n0
Theorem 4.4 Under the assumptions of Theorem 4.1 and the construction described there
we have, on the new probability space,
s
b n (W
cn∗ ), Σsn0 (Wn∗ ); vn , wn ) = n−1/2 Bn (W
cn∗ ) + (Wn∗ − W
cn∗ )O(1) + o(n−1/2 ),
∆n (Σ
as n → ∞, a.s..
The above result is remarkable in that it shows that it is optimal in the sense that
c ∗ of
b s (W
c ∗ ), Σs (W ∗ ); vn , wn ) inherits the rate of convergence from the estimator W
∆n (Σ
n
n
n
n0
n
∗
the optimal shrinkage weight Wn , cf. (4.4).
4.4. Nearly orthogonal projections
(i)
Let vn , i = 1, . . . , Ln , be unit vectors in Rdn on which we may project Yn , e.g. in order
to determine the best approximating direction. Recall that the true covariance between two
(i)
(j)
projections vn ′ Yn and vn ′ Yn is
Cov (vn(i) ′ Yn , vn(j) ′ Yn ) = vn(i) ′ Σn vn(j) ,
and the corresponding shrinkage estimator is
d (v(i) ′ Yn , v(j) ′ Yn ) = v(i) ′ Σ
b s (W
c ∗ )v(j) .
Cov
n
n
n
n
n n
(i)
b s (W
cn∗ ), as
Clearly, those covariances vanish for i 6= j, if the vn are chosen as eigenvectors of Σ
n
in a classical principal component analysis (PCA) applied to the shrinkage covariance matrix
estimator. But when analyzing high–dimensional data it is common to rely on procedures
such as sparse PCA, see [19], which yield sparse principal components. Then analyzing the
(i)
covariances of the projections vn ′ Yn is of interest.
36
A. STELAND, R. VON SACHS
(j)
If OLn = {vn : 1 ≤ j ≤ Ln } is an orthogonal system and Ln < dn , then OLn spans a
Ln –dimensional subspace of Rdn . Of course, there are at most Ln orthogonal vectors, i.e. Ln
cannot be larger than dn . However, if one relaxes the orthogonality condition
vn(i) ′ vn(j) = 0,
i 6= j,
then one can place much more unit vectors in the Euclidean space Rdn in such a way that their
pairwise angles are small. Indeed, [18] provides an elegant proof of the following KabatjanskiiLevenstein bound.
Theorem 4.5 (Cheap version of the Kabatjanskii-Levenstein bound, Tao 2013).
√
−1/2
for some 1/2 ≤ A ≤ 12 dn .
Let x1 , . . . , xm be unit vectors in Rdn such that |x′i xj | ≤ Adn
2
n CA
for some universal constant C.
Then we have m ≤ Cd
A2
Theorem 4.5 motivates to study the case of nearly orthogonal weighting vectors, defined as
a pair (vn , wn ) ∈ W × W satisfying
(4.12)
vn′ wn = o(1),
n → ∞,
Now the asymptotics of the shrinkage estimator is as follows:
Theorem 4.6 Let vn and wn be unit vectors satisfying the nearly orthogonal condition
(4.12) and suppose that the conditions of Theorem 4.1 hold. Then
√
b s (W ) − Σ
b s (W ))wn − Bn (W )| = oP (1),
| nv′ (Σ
n
n
n0
as n → ∞, a.s., where
Bn (W ) = (1 − W )B n (1)0 +
W vn′ wn d−1/2
n
dn
X
B n (1)i .
i=1
−1/2 Pdn
Observe that for asymptotically orthogonal weighting vectors the term W vn′ wn dn
i=1 B n (1)i
corresponding to the (parametric) shrinkage target is oP (1) and thus vanishes asymptotically.
In this situation, the nonparametric part dominates in large samples.
4.5. Proofs
Proof of Theorem 4.1: First notice that (4.6) ensures that the second term in (4.10)
does not converge to 0 in probability, since (vn , wn ) is a regular projection and condition
−1/2 Pdn
(4.8) ensures that the Gaussian random variable Vn = dn
j=1 B n (1)j is not oP (1), since
inf P (|Vn | > δ) ≥ inf P (|Vn /σtr,n | > δ/σ) > 0
n≥1
n≥1
for any δ > 0. Hence W vn′ wn Vn = oP (1) iff. vn′ wn = o(1), which is excluded by (4.6).
n
We argue similarly as in the proof of Theorem 3.1: Put Dn = (Dnj )dj=0
with Dn0 (t) =
−1/2
−1/2
Ln Dn (t; vn , wn ) and Dnj (t) = Ln Dn (t, ej , ej ), j = 1, . . . , dn , where Ln = dn + 1. Since
the weighting vectors are uniformly ℓ1 -bounded, Theorem 2.3 yields, on a new probability
space where a process equivalent to Dn can be defined and will be denoted again by Dn , the
37
INFERENCE FOR THE TRACE
n
existence of a Brownian motion {(B n (t))dj=0
: t ≥ 0} as characterized in Theorem 4.1, such
that
√
b n wn − v′ Σn wn ) − B n (1)0 | = o(1),
| n(vn′ Σ
n
and
√
b n k∗ − kΣn k∗ ) − d−1/2
n(kΣ
tr
tr
n
dn
X
B n (1)i = o(1),
i=1
b n ktr d−1 and the fact that |v′ wn | ≤ kvn kℓ kwn kℓ =
as n → ∞, a.s. Using these results, µ
bn = kΣ
n
n
1
1
O(1), we have for any 0 < W ≤ 1
|An (W ) − Bn (W )|
√
b n − Σn )wn − (1 − W )B n (1)0
= nvn′ (1 − W )(Σ
dn
X
√
b n k∗ − kΣn k∗ )wn − W v′ wn d−1/2
B n (1)i
+ W nvn′ (kΣ
tr
tr
n
n
i=1
√
b n − Σn )wn − B n (1)0 |
≤ (1 − W )| nvn′ (Σ
+W
√
∗
b k∗ − kΣn k∗ ) − v′ wn d−1/2
nvn′ wn (kΣ
n tr
tr
n
n
√
b n − Σn )wn − B n (1)0 |
≤ (1 − W )| nvn′ (Σ
+
W |vn′ wn |
= o(1),
dn
X
B n (1)i
i=1
dn
X
√
∗ ∗
∗
−1/2
b
n(kΣn ktr − kΣn ktr ) − dn
B n (1)i
i=1
as n → ∞, a.s., which shows (4.9).
Q.E.D.
Proof of Theorem 4.2: By Theorem 2.1 there exist, on a new probability space (Ω′ , F ′ , P ′ ),
d
e ni : i ≥ 1} =
en (t)0 : t ∈
an equivalent process {Y
{Yni : i ≥ 1} and a Brownian motion {B
[0, 1]} on [0, 1], such that
√
e n (t) − Σn (t))wn − B
en (t)0 | = o(1),
(4.13)
sup | nvn′ (Σ
t∈[0,1]
e n (t) = n−1 P⌊nt⌋ Y
e ni Y
e ′ . By Billingsley’s lemma, [2, Section 21,
as n → ∞, P ′ -a.s., where Σ
ni
i=1
′
d
en (t)0 : t ∈ [0, 1]}n defined
Lemma 2], there exist Brownian motions {B n (t)0 : t ∈ [0, 1]}n = {B
on the original probability space (Ω, F, P ), such that
(4.14)
√
b n (t) − Σn (t))wn − B ′n (t)0 | = o(1),
sup | nvn′ (Σ
t∈[0,1]
as n → ∞, a.s. Indeed, recall that the infinite product of a complete and separabe metric
space is complete and separable; in our case (D[0, 1])∞ equipped with the usual metric, see
[2, p. 241], induced by th Skorohod metric making D[0, 1] separable and complete. Then
38
A. STELAND, R. VON SACHS
√
e n (·) − Σn (·))wn , {B
en (·)0 }n≥1 ) and σ =
apply [2, Sec. 21, Lemma 1] with ν = L({ nvn′ (Σ
√ ′ e
′
{ nvn (Σn (·) − Σn (·))wn } to conclude the existence of τ =: {B n (·)0 }, a function of σ and
U1 , such that (4.14) holds, where the convergence w.r.t. the supnorm follows from the a.s.
′
continuity of B n (·)0 .
Further, by Theorem 3.1, there exist, on a new probability space (Ω′′ , F ′′ , P ′′ ), an equivalent
d
en (t)j )dn : t ∈
vector time series {Y̌ni : i ≥ 1} = {Yni : i ≥ 1}, and Brownian motions {(B
j=1
[0, 1]} on [0, 1] in dimension dn characterized as in the theorem, such that
√
(4.15)
n(kΣ̌n (t)k∗tr − kΣn (t)k∗tr ) − d−1/2
n
as n → ∞, P ′′ -a.s., where Σ̌n (t) = n−1
dn
X
j=1
P⌊nt⌋
en (1)j = o(1),
B
′ . Again, an application of Billingsley’s
Y̌ni Y̌ni
′
d
n
en (t)j )dn :
: t ∈ [0, 1]}n = {(B
lemma shows the existence of Brownian motions {(B n (t)j )dj=1
j=1
t ∈ [0, 1]}n , such that
√
(4.16)
i=1
b n (t)k∗ − kΣn (t)k∗ ) − d−1/2
n(kΣ
tr
tr
n
dn
X
′
B n (1)j = o(1),
j=1
as n → ∞, a.s., i.e. on the original probability space.
A priori, we have no information on the exact second order structure of the two Brownian
′
(n)
motions, but B n (1)j is close to the associated process Dnj (t) = Tn,0 (j), cf. (2.44), and to the
(n)
corresponding martingale approximation M⌊nt⌋,0 (j) defined in (2.45), which allows us to study
′
′
′
′
the convergence of the covariances Cov (B n (1)0 , B n (1)j ) = (B n (1)0 , B n (1)j )L2 , j = 1, . . . , dn .
First observe that
′
(n)
max kB n (1)j − Mn,0 (j)kL2 = o(1),
1≤j≤dn
′
n
see [11, Lemma 2], [17], and Lemma 2.2, because of (4.13) and since (B n (t)j )dj=1
satisfies
Pdn
′
2
supt∈[0,1] j=1 |B n (t)j − Dnj (t)| = o(1), as n → ∞, a.s., see Theorem 2.3 (ii). Also notice
′
(n)
that kB n (1)j kL2 and kMn,0 (j)kL2 are O(1), uniformly in j ≥ 0. Now use the decomposition
(X, Y )L2 = (X ′ , Y ′ )L2 + (X ′ , Y − Y ′ )L2 + (X − X ′ , Y )L2
for X, Y, X ′ , Y ′ ∈ L2 to conclude that
′
′
(n)
(n)
′
(n)
(n)
Cov (B n (1)0 , B n (1)j )) = (Mn,0 (0), Mn,0 (j))L2 + (Mn,0 (0), B n (1)j − Mn,0 (j))L2
′
(n)
(n)
+ ((B n (1)0 − Mn,0 (0), Mn,0 (j))L2 ,
where the last two terms are o(1), uniformly in j. E.g.,
′
(n)
′
(n)
(n)
(n)
max |(B n (1)0 − Mn,0 (0), Mn,0 (j))L2 | ≤ sup kB n (1)0 − Mn,0 (0)kL2 kMn,0 (j)kL2 = o(1),
1≤j≤dn
1≤j
as n → ∞. Combining these estimates with Lemma 2.2 yields
′
′
(n)
(n)
max |Cov (B n (1)0 , B n (1)j ) − Cov (Mn,0 (0), Mn,
(j))| ≪ d−1
n + o(1),
1≤j≤dn
39
INFERENCE FOR THE TRACE
−1/2
which establishes (4.11), since the covariances of the approximating martingales equal dn βn2 (vn , wn , ej , ej )+
−1/2
o(1), as n → ∞; the factor dn
is due to the additional scaling of Dnj (t) to approximate dn
bilinear forms by Theorem 2.3.
Q.E.D.
Proof of Theorem 4.3: Recall that, since Σn = Cn ΛC′n and (1.2), the elements of Σn
are uniformly bounded (in n), such that |vn′ Σn wn | ≤ kvn kℓ1 kwn kℓ1 = O(1). This in turn
implies that λmax (Σn ) = O(1) and
(4.17)
tr(Σn )d−1
n = O(1).
Put
(4.18)
c ∗ )wn − v′ Σs (W ∗ )wn
c ∗ , W ∗ ) = v′ Σs (W
Rn (W
n
n n0
n
n
n n0
n
and notice that
′
cn∗ , Wn∗ ) = (Wn∗ − W
cn∗ )vn′ Σn wn + (W
cn∗ − Wn∗ )tr(Σn )d−1
Rn (W
n vn wn .
Using (4.17) we obtain the bound
(4.19)
c ∗ , W ∗ ) = (W
c ∗ − W ∗ )O(1).
Rn (W
n
n
n
n
Observe that
b s (W
b s (W ∗ ) = (W ∗ − W
b n + (W
b n )d−1 In
cn∗ ) − Σ
cn∗ )Σ
cn∗ − Wn∗ )tr(Σ
Σ
n
n
n
n
n
is equal to the difference
cn∗ ) − Σsn0 (Wn∗ ) = (Wn∗ − W
cn∗ )Σn + (W
cn∗ − Wn∗ )tr(Σn )d−1
Σsn0 (W
n In
b n . We have
when replacing Σn by Σ
b s (W
c∗) − Σ
b s (W ∗ )
Σ
n
n
n
n
∗
c
cn∗ − Wn∗ )tr(Σn )d−1
= (Wn − Wn∗ )Σn + (W
n In
cn∗ )(Σ
b n − Σn ) + (W
cn∗ − Wn∗ )(tr(Σ
b n ) − tr(Σn ))d−1
+ (Wn∗ − W
n In
Using (4.19), we therefore obtain for the associated bilinear form
b n (W
c∗) − Σ
b s (W ∗ ))wn
vn′ (Σ
n
n
n
∗
∗
c
= Rn (Wn , Wn )
cn∗ )vn′ (Σ
b n − Σn )wn + (Wn∗ − W
cn∗ )vn′ wn (kΣ
b n k∗tr − kΣn k∗tr )
+ (Wn∗ − W
c ∗ − W ∗ )O(1)
= (W
n
n
∗
cn∗ ) B n (1)0 n−1/2 + o(n−1/2 )
+ (Wn − W
!
dn
X
−1/2
∗
∗
−1/2
−1/2
c −W ) n
B n (1)i + o(n
) v ′ wn ,
+ (W
d
n
n
n
n
i=1
as n → ∞, a.s..
Q.E.D.
40
A. STELAND, R. VON SACHS
Proof of Theorem 4.4:
Recall that An (W ) =
√
b s (W ) − Σs (W ))wn . We have
nvn′ (Σ
n
n0
b sn (W
cn∗ ) − Σsn0 (Wn∗ ))wn
vn′ (Σ
c ∗ ) − Σs (W
c ∗ ))wn + v′ (Σs (W
c ∗ ) − Σs (W ∗ ))wn
b s (W
= vn′ (Σ
n
n
n0
n
n
n0
n
n0
n
−1/2
∗
∗
∗
c
c
=n
An (W ) + Rn (W , W )
n
n
n
cn∗ , Wn∗ ) = (W
cn∗ − Wn∗ )O(1). Further, using
where again Rn (W
√ ′
b s (W
cn∗ ) − Σsn0 (W
cn∗ ))wn = Bn (W
cn∗ ) + o(1),
nvn (Σ
n
a.s., we arrive at
b s (W
c ∗ ) − Σs (W ∗ ))wn
vn′ (Σ
n
n
n0
n
−1/2
∗
c
c∗, W ∗)
=n
An (W ) + Rn (W
n
=n
−1/2
n
n
cn∗ ) + o(1)) + O(W
cn∗ − Wn∗ ),
(Bn (W
as n → ∞, a.s., which completes the proof.
Proof of Theorem 4.6:
Theorem 4.1, we obtain
Q.E.D.
Let An (W ) be defined as in (4.7). Arguing as in the proof of
√
b n − Σn )wn − αn (vn , wn )B n (1)0 |
|An (W ) − Bn (W )| ≤ (1 − W )| nvn′ (Σ
dn
X
√
b n k∗tr − kΣn k∗tr ) − vn′ wn d−1/2
B n (1)i
+ W vn′ wn n(kΣ
n
i=1
The first summand is o(1), a.s., by Theorem 2.1. Under assumption (4.12), the second term
can be bounded by
Rn =
√
b n k∗
vn′ wn n(kΣ
tr
= W |vn′ wn |
= o(1),
√
b n k∗ ) − v′ wn
− kΣ
tr
n
dn
X
B n (1)i
i=1
b n k∗tr − kΣ
b n k∗tr ) − d−1/2
n(kΣ
n
dn
X
B n (1)i
i=1
as n → ∞, a.s., which completes the proof.
Q.E.D.
APPENDIX A: NOTATION AND FORMULAS
We denote
σk2 = E(ǫ2k ), γk = E(ǫ4k ).
The approximating martingales used to obtain the strong approximations require to control
the following quantities. For the reader’s convenience, we reproduce them here from [17] as
well as some related formulas and results. Let
(A.1)
(n)
(n)
f0,j = f0,j (vn , wn ) =
dn
X
ν,µ=1
(ν) (µ)
vν wµ cj cj ,
j = 0, 1, . . . ,
41
INFERENCE FOR THE TRACE
(A.2)
(n)
(n)
fl,j = fl,j (vn , wn ) =
dn
X
(ν) (µ)
(µ) (ν)
vν wµ [cj cj+l + cj cj+l ],
l = 1, 2, . . . ; j = 0, 1, . . . ,
ν,µ=1
and
(A.3)
(n)
(n)
fel,i = fel,i (vn , wn ) =
∞
X
(n)
fl,j =
dn
∞ X
X
(ν) (µ)
(µ) (ν)
vν wµ [cj cj+l + cj cj+l ],
l, i = 0, 1, . . . .
j=i ν,µ=1
j=i
Lemma and Definition A.1 Suppose that vn , wn have uniformly bounded ℓ1 -norm in the
sense of equation (1.5). Then Assumption (A) implies
(A.4)
sup
∞
∞ X
X
(n)
(n)
(fel,i − fel,i+n′ )2 ≤ C(n′ )1−θ ,
for all n′ = 1, 2, . . . ,
n∈N i=1 l=0
′
(A.5)
n X
∞
X
(n)
(fer+k,0 )2 ≤ C(n′ )1−θ ,
sup
for all n′ = 1, 2, . . . ,
n∈N k=1 r=0
′
(A.6)
n X
∞
X
(n)
sup
(fel,k )2 ≤ C(n′ )1−θ ,
for all n′ = 1, 2, . . . ,
n∈N k=1 l=0
and there exist
(A.7)
α2n = α2n (vn , wn ) ≥ 0,
n ≥ 1,
such that
j−1
n X
n
X
X
(n)
′ 2
′ 1−θ
2
2
4
,
(fej−l,0 )2 σm
)
+
(γm′ +j − σm
′ +j σm′ +l − n αn ≤ C(n )
′ +j
′
′
(A.8)
(n)
(fe00 )2
j=1 l=1
j=1
for all n′ , m′ = 0, 1, · · · .
en , w
e n , n ≥ 1, have uniformly bounded ℓ1 -norms, then there exist
Further, if vn , wn , v
(A.9)
with
en , w
e n ),
βn2 = βn2 (vn , wn , v
n ≥ 1,
j−1
n X
n
X
X
(n)
(n)
2
2
4
e n )σm
vn , w
fej−l,0 (vn , wn )fej−l,0 (e
(γm′ +j − σm′ +j ) +
′ +j σm′ +l
′
′
(A.10)
(n)
(n)
e n)
fe0,0 (vn , wn )fe0,0 (e
vn , w
j=1
j=1 l=1
n′ ,m′
en , w
e n ) << (n′ )1−θ .
− n′ βn2 (vn , wn , v
ACKNOWLEDGMENTS
Part of this work has been supported by a grant of the first author from Deutsche Forschungsgemeinschaft (DFG), grant No. STE 1034/11-1, which he gratefully acknowledges. Rainer
von Sachs gratefully acknowledges funding by contract ”Projet d’Actions de Recherche Concertées” No. 12/17-045 of the ,,Communauté française de Belgique” and by IAP research
network Grant P7/06 of the Belgian government (Belgian Science Policy).
42
A. STELAND, R. VON SACHS
REFERENCES
[1] Andrew R. Barron, Albert Cohen, Wolfgang Dahmen, and Ronald A. DeVore. Approximation and learning by greedy algorithms. Ann. Statist., 36(1):64–94, 2008.
[2] Patrick Billingsley. Convergence of probability measures. Wiley Series in Probability and
Statistics: Probability and Statistics. John Wiley & Sons, Inc., New York, second edition,
1999. A Wiley-Interscience Publication.
[3] D. Bosq. Linear processes in function spaces, volume 149 of Lecture Notes in Statistics.
Springer-Verlag, New York, 2000. Theory and applications.
[4] Joshua Brodie, Ingrid Daubechies, Christine De Mol, Domenico Giannone, and Ignace
Loris. Sparse and stable Markowitz portfolios. Proceedings the National Academy of
Sciences of the United States of America, 106(30):12267–12272, 2009.
[5] Herold Dehling and Walter Philipp. Almost sure invariance principles for weakly dependent vector-valued random variables. Ann. Probab., 10(3):689–701, 1982.
[6] Jianqing Fan, Yingying Fan, and Jinchi Lv. High dimensional covariance matrix estimation using a factor model. J. Econometrics, 147(1):186–197, 2008.
[7] Mark Fiecas, Jürgen Franke, Rainer von Sachs, and Joseph Tadjuidje. Shrinkage estimation for multivariate hidden Markov models. J. Amer. Statist. Assoc., to appear,
2016.
[8] Moritz Jirak. Change-point analysis in increasing dimension. J. Multivariate Anal.,
111:136–159, 2012.
[9] Tōnu Kollo and Heinz Neudecker. Asymptotics of eigenvalues and unit-length eigenvectors of sample variance and correlation matrices. J. Multivariate Anal., 47(2):283–300,
1993.
[10] Tōnu Kollo and Heinz Neudecker. Corrigendum: “Asymptotics of eigenvalues and unitlength eigenvectors of sample variance and correlation matrices” [J. Multivariate Anal.
47 (1993), no. 2, 283–300; MR1247379 (94j:62040)]. J. Multivariate Anal., 51(1):210,
1994.
[11] Michael A. Kouritzin. Strong approximation for cross-covariances of linear variables with
long-range dependence. Stochastic Process. Appl., 60(2):343–353, 1995.
[12] Oliver Ledoit and Michael Wolf. Improved estimation of the covariance matrix of stock
returns with an application to portfolio selection. Journal of Empirical Finance, 10:603–
621, 2003.
[13] Olivier Ledoit and Michael Wolf. A well-conditioned estimator for large-dimensional
covariance matrices. J. Multivariate Anal., 88(2):365–411, 2004.
[14] Weidong Liu and Zhengyan Lin. Strong approximation for a class of stationary processes.
Stochastic Process. Appl., 119(1):249–280, 2009.
[15] Walter Philipp. A note on the almost sure approximation of weakly dependent random
variables. Monatsh. Math., 102(3):227–236, 1986.
[16] Alessio Sancetta. Sample covariance shrinkage for high dimensional dependent data. J.
Multivariate Anal., 99(5):949–967, 2008.
[17] A. Steland and R. von Sachs. Large-sample approximations for variance-covariance matrices of high-dimensional time series. Bernoulli, in press, 2016.
[18] Terence Tao. A cheap version of the Kabatjanski-Levenstein bound for almost orthogonal
vectors.
[19] Daniela M. Witten and Robert Tibshirani. Testing significance of features by lassoed
principal components. Ann. Appl. Stat., 2(3):986–1012, 2008.
INFERENCE FOR THE TRACE
43
[20] Daniela M. Witten, Robert Tibshirani, and Trevor Hastie. A penalized decomposition,
with applications to sparse principal components and canonical correlation analysis. Biostatistics, 10:515–534, 2009.
[21] Wei Biao Wu. Strong invariance principles for dependent random variables. Ann. Probab.,
35(6):2294–2320, 2007.
[22] Wei Biao Wu, Yinxiao Huang, and Wei Zheng. Covariances estimation for long-memory
processes. Adv. in Appl. Probab., 42(1):137–157, 2010.
[23] Wei Biao Wu and Wanli Min. On linear processes with dependent innovations. Stochastic
Process. Appl., 115(6):939–958, 2005.
[24] Li-xin Zhang. Strong approximations of martingale vectors and their applications in
Markov-chain adaptive designs. Acta Math. Appl. Sin. Engl. Ser., 20(2):337–352, 2004.
| 10math.ST
|
Genetic Algorithm for Solving Simple Mathematical Equality Problem
Denny Hermawanto
Indonesian Institute of Sciences (LIPI), INDONESIA
Mail: [email protected]
Abstract
This paper explains genetic algorithm for novice in this field. Basic philosophy of genetic
algorithm and its flowchart are described. Step by step numerical computation of genetic
algorithm for solving simple mathematical equality problem will be briefly explained.
Basic Philosophy
Genetic algorithm developed by Goldberg was inspired by Darwin's theory of evolution
which states that the survival of an organism is affected by rule "the strongest species that
survives". Darwin also stated that the survival of an organism can be maintained through the
process of reproduction, crossover and mutation. Darwin's concept of evolution is then
adapted to computational algorithm to find solution to a problem called objective function in
natural fashion. A solution generated by genetic algorithm is called a chromosome, while
collection of chromosome is referred as a population. A chromosome is composed from
genes and its value can be either numerical, binary, symbols or characters depending on the
problem want to be solved. These chromosomes will undergo a process called fitness
function to measure the suitability of solution generated by GA with problem. Some
chromosomes in population will mate through process called crossover thus producing new
chromosomes named offspring which its genes composition are the combination of their
parent. In a generation, a few chromosomes will also mutation in their gene. The number of
chromosomes which will undergo crossover and mutation is controlled by crossover rate and
mutation rate value. Chromosome in the population that will maintain for the next generation
will be selected based on Darwinian evolution rule, the chromosome which has higher fitness
value will have greater probability of being selected again in the next generation. After
several generations, the chromosome value will converges to a certain value which is the best
solution for the problem.
The Algorithm
In the genetic algorithm process is as follows [1]:
Step 1. Determine the number of chromosomes, generation, and mutation rate and crossover
rate value
Step 2. Generate chromosome-chromosome number of the population, and the initialization
value of the genes chromosome-chromosome with a random value
Step 3. Process steps 4-7 until the number of generations is met
Step 4. Evaluation of fitness value of chromosomes by calculating objective function
Step 5. Chromosomes selection
Step 6. Crossover
Step 7. Mutation
Step 8. Solution (Best Chromosomes)
The flowchart of algorithm can be seen in Figure 1
ith Population
Chromosome
Chromosome
Solutions
encoding
Chromosome
Chromosome
Evaluation
Selection
Next generation
Roulette
\’ wheel
Crossover
Mutation
End ?
N
i=i+1
Y
Best Chromosome
Decoding
Best Solution
Figure 1. Genetic algorithm flowchart
Numerical Example
Here are examples of applications that use genetic algorithms to solve the problem of
combination. Suppose there is equality a + 2b + 3c + 4d = 30, genetic algorithm will be used
to find the value of a, b, c, and d that satisfy the above equation. First we should formulate
the objective function, for this problem the objective is minimizing the value of function f(x)
where f(x) = ((a + 2b + 3c + 4d) - 30). Since there are four variables in the equation, namely
a, b, c, and d, we can compose the chromosome as follow: To speed up the computation, we
can restrict that the values of variables a, b, c, and d are integers between 0 and 30.
a
b
c
d
Step 1. Initialization
For example we define the number of chromosomes in population are 6, then we generate
random value of gene a, b, c, d for 6 chromosomes
Chromosome[1] = [a;b;c;d] = [12;05;23;08]
Chromosome[2] = [a;b;c;d] = [02;21;18;03]
Chromosome[3] = [a;b;c;d] = [10;04;13;14]
Chromosome[4] = [a;b;c;d] = [20;01;10;06]
Chromosome[5] = [a;b;c;d] = [01;04;13;19]
Chromosome[6] = [a;b;c;d] = [20;05;17;01]
Step 2. Evaluation
We compute the objective function value for each chromosome produced in initialization
step:
F_obj[1] = Abs(( 12 + 2*05 + 3*23 + 4*08 ) - 30)
= Abs((12 + 10 + 69 + 32 ) - 30)
= Abs(123 - 30)
= 93
F_obj[2] = Abs((02 + 2*21 + 3*18 + 4*03) - 30)
= Abs((02 + 42 + 54 + 12) - 30)
= Abs(110 - 30)
= 80
F_obj[3] = Abs((10 + 2*04 + 3*13 + 4*14) - 30)
= Abs((10 + 08 + 39 + 56) - 30)
= Abs(113 - 30)
= 83
F_obj[4] = Abs((20 + 2*01 + 3*10 + 4*06) - 30)
= Abs((20 + 02 + 30 + 24) - 30)
= Abs(76 - 30)
= 46
F_obj[5] = Abs((01 + 2*04 + 3*13 + 4*19) - 30)
= Abs((01 + 08 + 39 + 76) - 30)
= Abs(124 - 30)
= 94
F_obj[6] = Abs((20 + 2*05 + 3*17 + 4*01) - 30)
= Abs((20 + 10 + 51 + 04) - 30)
= Abs(85 - 30)
= 55
Step 3. Selection
1. The fittest chromosomes have higher probability to be selected for the next
generation. To compute fitness probability we must compute the fitness of each
chromosome. To avoid divide by zero problem, the value of F_obj is added by 1.
Fitness[1] = 1 / (1+F_obj[1])
= 1 / 94
= 0.0106
Fitness[2] = 1 / (1+F_obj[2])
= 1 / 81
= 0.0123
Fitness[3] = 1 / (1+F_obj[3])
= 1 / 84
= 0.0119
Fitness[4] = 1 / (1+F_obj[4])
= 1 / 47
= 0.0213
Fitness[5] = 1 / (1+F_obj[5])
= 1 / 95
= 0.0105
Fitness[6] = 1 / (1+F_obj[6])
= 1 / 56
= 0.0179
Total = 0.0106 + 0.0123 + 0.0119 + 0.0213 + 0.0105 + 0.0179
= 0.0845
The probability for each chromosomes is formulated by: P[i] = Fitness[i] / Total
P[1] = 0.0106 / 0.0845
= 0.1254
P[2] = 0.0123 / 0.0845
= 0.1456
P[3] = 0.0119 / 0.0845
= 0.1408
P[4] = 0.0213 / 0.0845
= 0.2521
P[5] = 0.0105 / 0.0845
= 0.1243
P[6] = 0.0179 / 0.0845
= 0.2118
From the probabilities above we can see that Chromosome 4 that has the highest fitness, this
chromosome has highest probability to be selected for next generation chromosomes. For the
selection process we use roulette wheel, for that we should compute the cumulative
probability values:
C[1] = 0.1254
C[2] = 0.1254 + 0.1456
= 0.2710
C[3] = 0.1254 + 0.1456 + 0.1408
= 0.4118
C[4] = 0.1254 + 0.1456 + 0.1408 + 0.2521
= 0.6639
C[5] = 0.1254 + 0.1456 + 0.1408 + 0.2521 + 0.1243
= 0.7882
C[6] = 0.1254 + 0.1456 + 0.1408 + 0.2521 + 0.1243 + 0.2118
= 1.0
Having calculated the cumulative probability of selection process using roulette-wheel can be
done. The process is to generate random number R in the range 0-1 as follows.
R[1] = 0.201
R[2] = 0.284
R[3] = 0.099
R[4] = 0.822
R[5] = 0.398
R[6] = 0.501
If random number R[1] is greater than C[1] and smaller than C[2] then select
Chromosome[2] as a chromosome in the new population for next generation:
NewChromosome[1] = Chromosome[2]
NewChromosome[2] = Chromosome[3]
NewChromosome[3] = Chromosome[1]
NewChromosome[4] = Chromosome[6]
NewChromosome[5] = Chromosome[3]
NewChromosome[6] = Chromosome[4]
Chromosomes in the population thus became:
Chromosome[1] = [02;21;18;03]
Chromosome[2] = [10;04;13;14]
Chromosome[3] = [12;05;23;08]
Chromosome[4] = [20;05;17;01]
Chromosome[5] = [10;04;13;14]
Chromosome[6] = [20;01;10;06]
In this example, we use one-cut point, i.e. randomly select a position in the parent
chromosome then exchanging sub-chromosome. Parent chromosome which will mate is
randomly selected and the number of mate Chromosomes is controlled using crossover_rate
(ρc) parameters. Pseudo-code for the crossover process is as follows:
begin
k← 0;
while(k<population) do
R[k] = random(0-1);
if(R[k]< ρc) then
select Chromosome[k] as parent;
end;
k = k + 1;
end;
end;
Chromosome k will be selected as a parent if R[k]<ρc. Suppose we set that the crossover rate
is 25%, then Chromosome number k will be selected for crossover if random generated value
for Chromosome k below 0.25. The process is as follows: First we generate a random number
R as the number of population.
R[1] = 0.191
R[2] = 0.259
R[3] = 0.760
R[4] = 0.006
R[5] = 0.159
R[6] = 0.340
For random number R above, parents are Chromosome[1], Chromosome[4] and
Chromosome[5] will be selected for crossover.
Chromosome[1] >< Chromosome[4]
Chromosome[4] >< Chromosome[5]
Chromosome[5] >< Chromosome[1]
After chromosome selection, the next process is determining the position of the crossover
point. This is done by generating random numbers between 1 to (length of Chromosome – 1).
In this case, generated random numbers should be between 1 and 3. After we get the
crossover point, parents Chromosome will be cut at crossover point and its gens will be
interchanged. For example we generated 3 random number and we get:
C[1] = 1
C[2] = 1
C[3] = 2
Then for first crossover, second crossover and third crossover, parent’s gens will be cut at
gen number 1, gen number 1 and gen number 3 respectively, e.g.
Chromosome[1] = Chromosome[1] >< Chromosome[4]
= [02;21;18;03] >< [20;05;17;01]
= [02;05;17;01]
Chromosome[4] = Chromosome[4] >< Chromosome[5]
= [20;05;17;01] >< [10;04;13;14]
= [20;04;13;14]
Chromosome[5] = Chromosome[5] >< Chromosome[1]
= [10;04;13;14] >< [02;21;18;03]
= [10;04;18;03]
Thus Chromosome population after experiencing a crossover process:
Chromosome[1] = [02;05;17;01]
Chromosome[2] = [10;04;13;14]
Chromosome[3] = [12;05;23;08]
Chromosome[4] = [20;04;13;14]
Chromosome[5] = [10;04;18;03]
Chromosome[6] = [20;01;10;06]
Step 5. Mutation
Number of chromosomes that have mutations in a population is determined by the
mutation_rate parameter. Mutation process is done by replacing the gen at random position
with a new value. The process is as follows. First we must calculate the total length of gen in
the population. In this case the total length of gen is total_gen =
number_of_gen_in_Chromosome * number of population
=4*6
= 24
Mutation process is done by generating a random integer between 1 and total_gen (1 to 24).
If generated random number is smaller than mutation_rate(ρm) variable then marked the
position of gen in chromosomes. Suppose we define ρm 10%, it is expected that 10% (0.1) of
total_gen in the population that will be mutated:
number of mutations = 0.1 * 24
= 2.4
≈2
Suppose generation of random number yield 12 and 18 then the chromosome which have
mutation are Chromosome number 3 gen number 4 and Chromosome 5 gen number 2. The
value of mutated gens at mutation point is replaced by random number between 0-30.
Suppose generated random number are 2 and 5 then Chromosome composition after mutation
are:
Chromosome[1] = [02;05;17;01]
Chromosome[2] = [10;04;13;14]
Chromosome[3] = [12;05;23;02]
Chromosome[4] = [20;04;13;14]
Chromosome[5] = [10;05;18;03]
Chromosome[6] = [20;01;10;06]
Finishing mutation process then we have one iteration or one generation of the genetic
algorithm. We can now evaluate the objective function after one generation:
Chromosome[1] = [02;05;17;01]
F_obj[1] = Abs(( 02 + 2*05 + 3*17 + 4*01 ) - 30)
= Abs((2 + 10 + 51 + 4 ) - 30)
= Abs(67 - 30)
= 37
Chromosome[2] = [10;04;13;14]
F_obj[2] = Abs(( 10 + 2*04 + 3*13 + 4*14 ) - 30)
= Abs((10 + 8 + 33 + 56 ) - 30)
= Abs(107 - 30)
= 77
Chromosome[3] = [12;05;23;02]
F_obj[3] = Abs(( 12 + 2*05 + 3*23 + 4*02 ) - 30)
= Abs((12 + 10 + 69 + 8 ) - 30)
= Abs(87 - 30)
= 47
Chromosome[4] = [20;04;13;14]
F_obj[4] = Abs(( 20 + 2*04 + 3*13 + 4*14 ) - 30)
= Abs((20 + 8 + 39 + 56 ) - 30)
= Abs(123 - 30)
= 93
Chromosome[5] = [10;05;18;03]
F_obj[5] = Abs(( 10 + 2*05 + 3*18 + 4*03 ) - 30)
= Abs((10 + 10 + 54 + 12 ) - 30)
= Abs(86 - 30)
= 56
Chromosome[6] = [20;01;10;06]
F_obj[6] = Abs(( 20 + 2*01 + 3*10 + 4*06 ) - 30)
= Abs((20 + 2 + 30 + 24 ) - 30)
= Abs(76 - 30)
= 46
From the evaluation of new Chromosome we can see that the objective function is decreasing,
this means that we have better Chromosome or solution compared with previous
Chromosome generation. New Chromosomes for next iteration are:
Chromosome[1] = [02;05;17;01]
Chromosome[2] = [10;04;13;14]
Chromosome[3] = [12;05;23;02]
Chromosome[4] = [20;04;13;14]
Chromosome[5] = [10;05;18;03]
Chromosome[6] = [20;01;10;06]
These new Chromosomes will undergo the same process as the previous generation of
Chromosomes such as evaluation, selection, crossover and mutation and at the end it produce
new generation of Chromosome for the next iteration. This process will be repeated until a
predetermined number of generations. For this example, after running 50 generations, best
chromosome is obtained:
Chromosome = [07; 05; 03; 01]
This means that: a = 7, b = 5, c = 3, d = 1
If we use the number in the problem equation:
a + 2b + 3c + 4d = 30
7 + (2 * 5) + (3 * 3) + (4 * 1) = 30
We can see that the value of variable a, b, c and d generated by genetic algorithm can satisfy
that equality.
REFERENCE
[1] Mitsuo Gen, Runwei Cheng, “Genetic Algorithms And Engineering Design”, John Wiley
& Sons, 1997.
| 9cs.NE
|
arXiv:1609.04288v1 [math.GR] 14 Sep 2016
A FAMILY OF TWO GENERATOR NON-HOPFIAN GROUPS
DONGHI LEE AND MAKOTO SAKUMA
Abstract. We construct 2-generator non-Hopfian groups Gm , m = 3, 4, 5, . . . ,
where each Gm has a specific presentation Gm = ha, b | urm,0 = urm,1 = urm,2 =
· · · = 1i which satisfies small cancellation conditions C(4) and T (4). Here, urm,i
is the single relator of the upper presentation of the 2-bridge link group of slope
rm,i , where rm,0 = [m + 1, m, m] and rm,i = [m + 1, m − 1, (i − 1)hmi, m + 1, m]
in continued fraction expansion for every integer i ≥ 1.
1. Introduction
Recall that a group G is called Hopfian if every epimorphism G → G is an automorphism. The non-Hopfian property of finitely generated groups has a close
connection with the non-residual finiteness. In fact, the classical work due to
Mal’cev [12] shows that every finitely generated non-Hopfian group is non-residually
finite. One of the hardest open problems about hyperbolic groups is whether or not
every hyperbolic group is residually finite. An important progress on this problem
was given by Sela [19] asserting that every torsion-free hyperbolic group is Hopfian.
In 2007, Osin [14] proved that this problem is equivalent to the question on whether
or not a group G is residually finite if G is hyperbolic relative to a finite collection of residually finite subgroups. The notion of relatively hyperbolic groups is an
important generalization of hyperbolic groups in geometric group theory originally
introduced by Gromov [5] (cf. [3], [4], [15]). Motivating examples for this generalization include the fundamental groups of non-compact hyperbolic manifolds of
finite volume. In particular, every 2-bridge link complement except for a torus link
is a hyperbolic manifold with cusps, so its fundamental group, that is, the 2-bridge
link group, is hyperbolic relative to its peripheral subgroups although it is not a
hyperbolic group. It is known by Groves [6] that a finitely generated torsion-free
group is Hopfian, if it is hyperbolic relative to free abelian subgroups. It is also
proved by Reinfeldt and Weidmann [16, 17] that every hyperbolic group, possibly
with torsion, is Hopfian. In addition, based on this result, Coulon and Guirardel [2]
2010 Mathematics Subject Classification. Primary 20F06, 57M25
The first author was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(2014R1A1A2054890). The second author was supported by JSPS Grants-in-Aid 15H03620.
1
proved that every lacunary hyperbolic group, which is characterized as a direct limit
of hyperbolic groups with a certain radii condition, is also Hopfian.
As for small cancellation groups, it is known that if a group has a finite presentation which satisfies small cancellation conditions either C ′ (1/6) or both C ′ (1/4)
and T (4), then it is hyperbolic (see [20]). Wise [22] also proved that every finite
C ′ (1/6)-small cancellation presentation defines a residually finite group.
Historically, not many have been known examples of finitely generated nonHopfian groups with specific presentations. The earliest such example was found
by Neumann [13] in 1950 as follows: ha, b | e2 = e3 = · · · = 1i, where ei =
a−1 b−1 ab−i ab−1 a−1 bi a−1 bab−i aba−1 bi for every integer i ≥ 2. Soon after, the first
non-Hopfian group with finite presentation was discovered by Higman [7] as follows:
ha, s, t | as = a2 , at = a2 i. Also a non-Hopfian group with the simplest presentation
up to now was produced by Baumslag and Solitar [1] as follows: ha, t | (a2 )t = a3 i.
Many other non-Hopfian groups with specific finite presentations have been obtained
by generalizing Higman’s group or Baumslag-Solitar’s group (see, for instance, [18],
[21]). Another notable non-Hopfian group was obtained by Ivanov and Storozhev [8]
in 2005. They constructed a family of finitely generated, but not finitely presented,
non-Hopfian relatively free groups with direct limits of hyperbolic groups, although
the defining relations of their group presentations are not explicitly described in
terms of generators.
Motivated by this background, we construct non-Hopfian groups by using hyperbolic 2-bridge link groups. In more detail, we construct a family of 2-generator
non-Hopfian groups each of which has the form G = ha, b | ur0 = ur1 = ur2 = · · · = 1i
satisfying small cancellation conditions C(4) and T (4), where uri is the single relator
of the upper presentation of the link group of the 2-bridge link of slope ri for every
i = 0, 1, 2, . . . . Here, the rational numbers ri may be parametrized by i ≥ 0, and
there is an explicit formula to express uri in terms of a and b. To parametrize the
rational numbers ri , we express ri in continued fraction expansion. Note that every
rational number 0 < s ≤ 1 has a unique continued fraction expansion such that
1
s=
m1 +
=: [m1 , m2 , . . . , mk ],
1
1
m2 + . . .
+
mk
where k ≥ 1, (m1 , . . . , mk ) ∈ (Z+ )k and mk ≥ 2 unless k = 1.
The main result of the present paper is the following, whose proof is contained in
in Section 3.
Theorem 1.1. Let r0 = [4, 3, 3], and let ri = [4, 2, (i − 1)h3i, 4, 3] for every integer
i ≥ 1. Then the group presentation G = ha, b | ur0 = ur1 = ur2 = · · · = 1i satisfies
small cancellation conditions C(4) and T (4), and G is non-Hopfian.
2
Here, the symbol “(i − 1)h3i” represents i − 1 successive 3’s if i − 1 ≥ 1, whereas
“0h3i” means that 3 does not occur in that place, so that r1 = [4, 2, 0h3i, 4, 3] =
[4, 2, 4, 3].
Remark 1.2. (1) Once we allow the components of a continued fraction expansion
to be “−”, meaning that the two integers immediately before and after − are added
to form one component, ri ’s in Theorem 1.1 can be parametrized including i = 0 as
ri = [4, ih2, 1, −i, 3, 3] for every i = 0, 1, 2, . . . .
(2) If we express the rational number ri in Theorem 1.1 as qi /pi , where pi and qi
are relatively prime positive integers, then |uri | = 2pi (see Section 2.1). A simple
computation shows that the inequality 3 < pi+1 /pi < 4 holds for every i = 0, 1, 2, . . . ,
so that the length |uri | of the word uri satisfies the inequality c · 3i < |uri | < c · 4i
for every integer i ≥ 1, where c = |ur0 |.
By looking at the proof of Theorem 1.1 in Section 3, it is not hard to see that a
similar result holds not only for r0 = [4, 3, 3] but also for r0 = [m + 1, m, m] with
m being any integer greater than 3. Thus we only state its general form without a
detailed proof.
Theorem 1.3. Suppose that m is an integer with m ≥ 3. Let r0 = [m + 1, m, m],
and let ri = [m + 1, m − 1, (i − 1)hmi, m + 1, m] for every integer i ≥ 1. Then the
group presentation G = ha, b | ur0 = ur1 = ur2 = · · · = 1i satisfies small cancellation
conditions C(4) and T (4), and G is non-Hopfian.
The present paper is organized as follows. In Section 2, we recall the upper
presentation of a 2-bridge link group, and basic facts established in [9] concerning
the upper presentations. We also recall key facts from [9] obtained by applying small
cancellation theory to the upper presentations. Section 3 is devoted to the proof of
the main result (Theorem 1.1).
2. Preliminaries
2.1. Upper presentations of 2-bridge link groups
We recall some notation in [9]. The Conway sphere S is the 4-times punctured
sphere which is obtained as the quotient of R2 − Z2 by the group generated by the
π-rotations around the points in Z2 . For each s ∈ Q̂ := Q ∪ {∞}, let αs be the
simple loop in S obtained as the projection of a line in R2 − Z2 of slope s. We call
s the slope of the simple loop αs .
For each r ∈ Q̂, the 2-bridge link K(r) of slope r is the sum of the rational tangle
(B 3 , t(∞)) of slope ∞ and the rational tangle (B 3 , t(r)) of slope r. Recall that
∂(B 3 − t(∞)) and ∂(B 3 − t(r)) are identified with S so that α∞ and αr bound disks
in B 3 − t(∞) and B 3 − t(r), respectively. By van-Kampen’s theorem, the link group
G(K(r)) = π1 (S 3 − K(r)) is obtained as follows:
G(K(r)) = π1 (S 3 − K(r)) ∼
= π1 (S)/hhα∞ , αr ii ∼
= π1 (B 3 − t(∞))/hhαr ii.
3
Let {a, b} be the standard meridian generator pair of π1 (B 3 − t(∞), x0 ) as described
in [9, Section 3]. Then π1 (B 3 − t(∞)) is identified with the free group F (a, b) with
basis {a, b}. For the rational number r = q/p, where p and q are relatively prime
positive integers, let ur be the word in {a, b} obtained as follows. Set ǫi = (−1)⌊iq/p⌋ ,
where ⌊x⌋ is the greatest integer not exceeding x.
(1) If p is odd, then
q
uq/p = aûq/p b(−1) û−1
q/p ,
where ûq/p = bǫ1 aǫ2 · · · bǫp−2 aǫp−1 .
(2) If p is even, then
uq/p = aûq/p a−1 û−1
q/p ,
where ûq/p = bǫ1 aǫ2 · · · aǫp−2 bǫp−1 .
∼ π1 (B 3 − t(∞)) is represented by the simple loop αr , and we
Then ur ∈ F (a, b) =
obtain the following two-generator and one-relator presentation of a 2-bridge link
groups:
G(K(r)) ∼
= π1 (B 3 − t(∞))/hhαr ii ∼
= ha, b | ur i.
This presentation is called the upper presentation of a 2-bridge link group.
2.2. Basic facts concerning the upper presentations
Throughout this paper, a cyclic word is defined to be the set of all cyclic permutations of a cyclically reduced word. By (v) we denote the cyclic word associated
with a cyclically reduced word v. Also the symbol “≡” denotes the letter-by-letter
equality between two words or between two cyclic words. Now we recall definitions
and basic facts from [9] which are needed in the proof of Theorem 1.1 in Section 3.
Definition 2.1. (1) Let v be a reduced word in {a, b}. Decompose v into
v ≡ v1 v2 · · · vt ,
where, for each i = 1, . . . , t − 1, vi is a positive (resp., negative) subword (that is, all
letters in vi have positive (resp., negative) exponents), and vi+1 is a negative (resp.,
positive) subword. Then the sequence of positive integers S(v) := (|v1 |, |v2 |, . . . , |vt |)
is called the S-sequence of v.
(2) Let v be a cyclically reduced word in {a, b}. Decompose the cyclic word (v)
into
(v) ≡ (v1 v2 · · · vt ),
where vi is a positive (resp., negative) subword, and vi+1 is a negative (resp., positive) subword (taking subindices modulo t). Then the cyclic sequence of positive
integers CS(v) := ((|v1 |, |v2 |, . . . , |vt |)) is called the CS-sequence of (v). Here the
double parentheses denote that the sequence is considered modulo cyclic permutations.
4
Definition 2.2. For a rational number r with 0 < r ≤ 1, let ur be the word defined
in the beginning of this section. Then the symbol CS(r) denotes the CS-sequence
CS(ur ) of (ur ), which is called the CS-sequence of slope r.
A reduced word w in {a, b} is said to be alternating if a±1 and b±1 appear in w
alternately, to be precise, neither a±2 nor b±2 appears in w. Also a cyclically reduced
word w in {a, b} is said to be cyclically alternating, i.e., all the cyclic permutations
of w are alternating. In particular, ur is a cyclically alternating word in {a, b}. Note
that every alternating word w in {a, b} is determined by the sequence S(w) and the
initial letter (with exponent) of w. Note also that if w is a cyclically alternating
word in {a, b} such that CS(w) = CS(r), then either (w) ≡ (ur ) or (w) ≡ (u−1
r ) as
cyclic words.
In the remainder of this section, we suppose that r is a rational number with
0 < r ≤ 1, and write r as a continued fraction expansion r = [m1 , m2 , . . . , mk ],
where k ≥ 1, (m1 , . . . , mk ) ∈ (Z+ )k and mk ≥ 2 unless k = 1. Note from [9] that
if k ≥ 2, then some properties of CS(r) differ according to m2 = 1 or m2 ≥ 2. For
brevity, we write m for m1 .
Lemma 2.3 ([9, Proposition 4.3]). For the rational number r = [m1 , m2 , . . . , mk ]
satisfying that m2 ≥ 2 if k ≥ 2, the following hold.
(1) Suppose k = 1, i.e., r = 1/m. Then CS(r) = ((m, m)).
(2) Suppose k ≥ 2. Then each term of CS(r) is either m or m + 1. Moreover,
no two consecutive terms of CS(r) can be (m + 1, m + 1), so there is a cyclic
sequence of positive integers ((t1 , t2 , . . . , ts )) such that
CS(r) = ((m + 1, t1 hmi, m + 1, t2 hmi, . . . , m + 1, ts hmi)).
Here, the symbol “ti hmi” represents ti successive m’s.
Definition 2.4. If k ≥ 2, the symbol CT (r) denotes the cyclic sequence ((t1 , t2 , . . . , ts ))
in Lemma 2.3, which is called the CT -sequence of slope r.
Lemma 2.5 ([9, Proposition 4.4 and Corollary 4.6]). For the rational number r =
[m1 , m2 , . . . , mk ] with k ≥ 2 and m2 ≥ 2, let r ′ be the rational number defined as
r ′ = [m2 − 1, m3 , . . . , mk ].
Then we have CT (r) = CS(r ′ ).
Lemma 2.6 ([9, Proposition 4.5]). For the rational number r = [m1 , m2 , . . . , mk ],
the cyclic sequence CS(r) has a decomposition ((S1 , S2 , S1 , S2 )) which satisfies the
following.
(1) Each Si is symmetric, i.e., the sequence obtained from Si by reversing the
order is equal to Si . (Here, S1 is empty if k = 1.)
(2) Each Si occurs only twice in the cyclic sequence CS(r).
(3) The subsequence S1 begins and ends with m + 1.
5
(4) The subsequence S2 begins and ends with m.
Lemma 2.7 ([9, Proof of Proposition 4.5]). For the rational number r = [m1 , m2 , . . . , mk ]
with k ≥ 2 and m2 ≥ 2, let r ′ be the rational number defined as in Lemma 2.5. Also
let CS(r ′ ) = ((T1 , T2 , T1 , T2 )) and CS(r) = ((S1 , S2 , S1 , S2 )) be the decompositions
described in Lemma 2.6. Then the following hold.
(1) If k = 2, then T1 = ∅, T2 = (m2 − 1), and S1 = (m + 1), S2 = ((m2 − 1)hmi).
(2) If k ≥ 3, then T1 = (t1 , . . . , ts1 ), T2 = (ts1 +1 , . . . , ts2 ), and
S1 = (m + 1, ts1 +1 hmi, m + 1, . . . , m + 1, ts2 hmi, m + 1),
S2 = (t1 hmi, m + 1, t2 hmi, . . . , ts1 −1 hmi, m + 1, ts1 hmi).
The following lemma is useful in the proof of Lemma 3.6.
Lemma 2.8. For two distinct rational numbers r = [m1 , m2 , . . . , mk ] and s =
[m1 , l2 , . . . , lt ], assume that
(i) m is a positive integer;
(ii) mi and lj are integers greater than 1 for every i ≥ 2 and j ≥ 2;
(iii) k, t ≥ 3 and k 6= t; and
(iv) if k < t, then m2 ≥ l2 , while if k > t, then m2 ≤ l2 .
Let r ′ and s′ be the rational numbers defined as in Lemma 2.5. Also let CS(r) =
((S1 , S2 , S1 , S2 )) and CS(r ′ ) = ((T1 , T2 , T1 , T2 )) be the decompositions described in
Lemma 2.6. Suppose that CS(s) contains S1 or S2 as a subsequence. Then CS(s′ )
contains T1 or T2 as a subsequence.
In the above lemma (and throughout this paper), we mean by a subsequence a
subsequence without leap. Namely a sequence (a1 , a2 , . . . , al ) is called a subsequence
of a cyclic sequence, if there is a sequence (b1 , b2 , . . . , bn ) representing the cyclic
sequence such that l ≤ n and ai = bi for 1 ≤ i ≤ l.
Proof. First suppose that CS(s) contains S1 as a subsequence. By Lemma 2.7(2),
CS(s) contains (m + 1, ts1 +1 hmi, m + 1, . . . , m + 1, ts2 hmi, m + 1) as a subsequence,
where T2 = (ts1 +1 , . . . , ts2 ). Then clearly CS(s′ ) = CT (s) contains (ts1 +1 , . . . , ts2 ),
that is, T2 , as a subsequence. So we are done.
Next suppose that CS(s) contains S2 as a subsequence. Again by Lemma 2.7(2),
CS(s) contains (t1 hmi, m + 1, t2 hmi, . . . , ts1 −1 hmi, m + 1, ts1 hmi) as a subsequence,
where T1 = (t1 , . . . , ts1 ). Then CS(s′ ) = CT (s) contains (d1 + t1 , t2 , . . . , ts1 −1 , ts1 +
d2 ) as a subsequence, where d1 , d2 ≥ 0. In the reminder of the proof, we show
that d1 = d2 = 0, so that CS(s′ ) = CT (s) contains T1 = (t1 , t2 , . . . , ts1 ) as a
subsequence. To this end, note that since r ′ = [m2 − 1, m3 , . . . , mk ], t1 = ts1 = m2
by Lemma 2.6(3). Also since s′ = [l2 − 1, l3 , . . . , lt ], CS(s′ ) = CT (s) consists of l2 − 1
and l2 by Lemma 2.3(2). Hence each of d1 + t1 = d1 + m2 and ts1 + d2 = m2 + d2
is either l2 − 1 or l2 . Suppose first that k < t. Then m2 ≥ l2 by the assumption
(iv), and thus the only possibility is m2 = l2 . Thus we have d1 = d2 = 0. Suppose
6
next that k > t. Then m2 ≤ l2 again by the assumption (iv). Note that k ≥ 4,
because t ≥ 3 by the assumption (iii). Thus we can see, by using Lemma 2.7 and
the assumption mi ≥ 2 for every i ≥ 2, that S2 contains (m + 1, (m2 − 1)hmi, m + 1)
as a subsequence. This implies that CS(s′ ) = CT (s) contains a term m2 − 1. Since
m2 − 1 ≤ l2 − 1, the only possibility m2 = l2 . Thus we again have d1 = d2 = 0,
completing the proof of Lemma 2.8.
2.3. Small cancellation theory applied to the upper presentations
A subset R of the free group F (a, b) is called symmetrized, if all elements of R are
cyclically reduced and, for each w ∈ R, all cyclic permutations of w and w−1 also
belong to R.
Definition 2.9. Suppose that R is a symmetrized subset of F (a, b). A nonempty
word v is called a piece (with respect to R) if there exist distinct w1 , w2 ∈ R such
that w1 ≡ vc1 and w2 ≡ vc2 . The small cancellation conditions C(p) and T (q),
where p and q are integers such that p ≥ 2 and q ≥ 3, are defined as follows (see
[11]).
(1) Condition C(p): If w ∈ R is a product of n pieces, then n ≥ p.
(2) Condition T (q): For w1 , . . . , wn ∈ R with no successive elements wi , wi+1 an
inverse pair (i mod n), if n < q, then at least one of the products w1 w2 , . . . ,
wn−1 wn , wn w1 is freely reduced without cancellation.
The following proposition enables us to apply small cancellation theory to the
upper presentation ha, b | ur i of G(K(r)).
Proposition 2.10 ([9, Theorem 5.1]). Let r be a rational number such that 0 <
r < 1, and let R be the symmetrized subset of F (a, b) generated by the single relator
ur of the group presentation G(K(r)) = ha, b | ur i. Then R satisfies C(4) and T (4).
This proposition follows from the following characterization of pieces, which in
turn is proved by using Lemma 2.6.
Lemma 2.11 ([10, Corollary 3.25]). Let r and R be as in Proposition 2.10. Then a
subword w of the cyclic word (u±1
r ) is a piece with respect to R if and only if S(w)
contains neither S1 nor (ℓ1 , S2 , ℓ2 ) with ℓ1 , ℓ2 ∈ Z+ as a subsequence.
3. Proof of Theorem 1.1
In this section, for brevity of notation, we sometimes write x̄ for x−1 for a letter
or a word x. For a quotient group H of the free group F (a, b) and two elements w1
and w2 of F (a, b), the symbol w1 =H w2 means the equality in the group H.
For r0 = [4, 3, 3], we have by using Lemma 2.7
CS(ur0 ) = CS(r0 ) = ((5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4)).
7
Let G0 = ha, b | ur0 = 1i. Also let X ≡ a · · · ā be the alternating word in {a, b} such
that S(X) = (4, 4, 4, 5, 4, 4), and let f : F (a, b) → F (a, b) be the homomorphism
defined by f (a) = X̄ and f (b) = b̄.
Lemma 3.1. Under the foregoing notation, let f˜ : F (a, b) → G0 be the composition
of f and the canonical surjection F (a, b) → G0 . Then f˜ is onto.
Proof. Since f˜(b) = b̄, it suffices to show that a ∈ G0 is contained in the image of f˜.
Let w ≡ a · · · ā be the alternating word in {a, b} such that S(w) = (3, 3, 3, 4, 3, 3).
Then
f (w) = X̄ b̄X̄bXbX̄ b̄X̄bXbX b̄X̄ b̄XbX.
Here, since X ≡ a · · · ā and X̄ ≡ a · · · ā are alternating words in {a, b}, we see that
f (w) ≡ a · · · ā is also an alternating word in {a, b} with
S(f (w)) = (S(X̄ b̄), S(X̄), S(bX), S(bX̄ b̄), S(X̄), S(bX), S(bX b̄), S(X̄ b̄), S(X), S(bX)).
Since S(X) = (4, 4, 4, 5, 4, 4) and S(X̄) = (4, 4, 5, 4, 4, 4), we have S(X̄ b̄) = (4, 4, 5, 4, 4, 5),
S(bX) = (5, 4, 4, 5, 4, 4), S(bX̄ b̄) = (5, 4, 5, 4, 4, 5) and S(bX b̄) = (5, 4, 4, 5, 4, 5), so
that
S(f (w)) = (4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4,
5, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4).
Letting v1 ≡ a · · · b̄, v2 ≡ a · · · b̄, and v3 ≡ b · · · ā be the cyclically alternating words
in {a, b} such that
S(v1 ) = (4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4),
S(v2 ) = (5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4),
S(v3 ) = (4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4),
we see that f (w) ≡ v1 v2 āv3 . Moreover, for each i = 1, 2, 3, since CS(vi ) = CS(r0 ),
(vi ) ≡ (u±1
r0 ) as cyclic words by [9, Lemma 3.2], which implies that vi =G0 1. Hence
f (w) =G0 ā, and thus a ∈ G0 is contained in the image of f˜, as required.
At this point, we set up the following notation which will be used at the end of
the proofs of Lemmas 3.3 and 3.4.
Notation 3.2. (1) Suppose that v is an alternating word in {a, b} such that there
is a sequence (t1 , t2 , . . . , ts ) of positive integers satisfying
S(v) = (ǫ1 h5i, t1 h4i, 5, t2 h4i, . . . , 5, ts h4i, ǫ2 h5i),
where ǫi is 0 or 1 for i = 1, 2. Then the symbol T (v) denotes the sequence
(t1 , t2 , . . . , ts ).
(2) Suppose that v is a cyclically alternating word in {a, b} such that there is a
cyclic sequence ((t1 , t2 , . . . , ts )) of positive integers satisfying
CS(v) = ((5, t1 h4i, 5, t2 h4i, . . . , 5, ts h4i)).
8
Then the symbol CT (v) denotes the cyclic sequence ((t1 , t2 , . . . , ts )). In particular,
by Lemma 2.7, if v ≡ ur for some r = [4, m2 , . . . , mk ] with m2 ≥ 2, then CT (ur ) =
CT (r) = CS(r ′ ), where r ′ = [m2 − 1, . . . , mk ].
(3) Suppose that v is an alternating word in {a, b} such that there is a sequence
(h1 , h2 , . . . , hp ) of positive integers satisfying
T (v) = (ǫ1 h2i, h1 h1i, 2, h2 h1i, . . . , 2, hp h1i, ǫ2 h2i),
where T (v) is defined as in (1) and ǫi is 0 or 1 for i = 1, 2. Then the symbol V (v)
denotes the sequence (h1 , h2 , . . . , hp ).
(4) Suppose that v is a cyclically alternating word in {a, b} such that there is a
cyclic sequence ((h1 , h2 , . . . , hp )) of positive integers satisfying
CT (v) = ((2, h1 h1i, 2, h2 h1i, . . . , 2, hp h1i)),
where CT (v) is defined as in (2). Then the symbol CV (v) denotes the cyclic sequence ((h1 , h2 , . . . , hp )). In particular, by Lemma 2.7, if v ≡ ur for some r =
[4, 2, m3 , . . . , mk ] with m3 ≥ 2, then CV (ur ) = CT (r ′ ) = CS(r ′′ ), where r ′ =
[1, m3 , . . . , mk ] and r ′′ = [m3 − 1, . . . , mk ].
Lemma 3.3. Under the foregoing notation, (f (ur0 )) =G0 (u±1
r1 ).
Proof. Recall that
CS(ur0 ) = CS(r0 ) = ((5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4)).
Clearly the cyclic word (ur0 ) has six positive or negative subwords of length 5.
Cutting in the middle of such subwords, we may write the cyclic word (ur0 ) as a
product (v1 · · · v6 ), where
v1 ≡ abab̄āb̄ābabab̄āb̄āb,
v2 ≡ ababāb̄āb̄ababāb̄ā,
v3 ≡ b̄ābabab̄āb̄āb,
v4 ≡ ababāb̄āb̄ababāb̄āb̄ab,
v5 ≡ v2−1 ≡ abab̄āb̄ābabab̄āb̄ā,
v6 ≡ v3−1 ≡ b̄ababāb̄āb̄ab.
Put wn :≡ f (vn ) for every n = 1, . . . , 6, namely
w1 :≡ X̄ b̄X̄bXbX b̄X̄ b̄X̄bXbX b̄,
w2 :≡ X̄ b̄X̄ b̄XbXbX̄ b̄X̄ b̄XbX,
w3 :≡ bX b̄X̄ b̄X̄bXbX b̄,
w4 :≡ X̄ b̄X̄ b̄XbXbX̄ b̄X̄ b̄XbXbX̄ b̄,
w5 :≡ w2−1
and w6 :≡ w3−1 .
9
It then follows that
(f (ur0 )) = (f (v1 · · · v6 )) = (w1 · · · w6 ).
Claim 1. X̄ b̄X̄bXbX b̄ =G0 z1 , where z1 ≡ a · · · b̄ is an alternating word in {a, b}
with S(z1 ) = (4, 5, 4, 5).
Proof of Claim 1. Recall that X ≡ a · · · ā and X̄ ≡ a · · · ā are alternating words in
{a, b} such that S(X) = (4, 4, 4, 5, 4, 4) and S(X̄) = (4, 4, 5, 4, 4, 4). It is not hard to
see that
S(X̄ b̄X̄bXbX b̄) = (S(X̄ b̄), S(X̄), S(bX), S(bX b̄))
= ((4, 4, 5, 4, 4, 5), (4, 4, 5, 4, 4, 4), (5, 4, 4, 5, 4, 4), (5, 4, 4, 5, 4, 5))
= (4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 5).
Letting y1 ≡ a · · · b̄ and z1 ≡ a · · · b̄ be alternating words in {a, b} such that
S(y1 ) = (4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4) and S(z1 ) = (4, 5, 4, 5), clearly
X̄ b̄X̄bXbX b̄ ≡ y1 z1 . Here, since CS(y1 ) = CS(r0 ) and so y1 =G0 1, we finally have
X̄ b̄X̄bXbX b̄ ≡ y1 z1 =G0 z1 , as required.
Claim 2. X̄ b̄X̄ b̄XbX =G0 z2 , where z2 ≡ a · · · ā is the alternating word in {a, b}
with S(z2 ) = (4, 4, 5, 4).
Proof of Claim 2. As in the proof of Claim 1, we have
S(X̄ b̄X̄ b̄XbX) = (S(X̄ b̄), S(X̄ b̄), S(X), S(bX))
= ((4, 4, 5, 4, 4, 5), (4, 4, 5, 4, 4, 5), (4, 4, 4, 5, 4, 4), (5, 4, 4, 5, 4, 4))
= (4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4).
Letting z2 ≡ a · · · ā and y2 ≡ b · · · ā be alternating words in {a, b} such that
S(z2 ) = (4, 4, 5, 4) and S(y2 ) = (4, 5, 4, 4, 5, 4, 4, 5, 4, 4, 4, 5, 4, 4, 5, 4, 4, 5, 4, 4), clearly
X̄ b̄X̄ b̄XbX ≡ z2 y2 . Here, since CS(y2 ) = CS(r0 ) and so y2 =G0 1, we finally have
X̄ b̄X̄ b̄XbX ≡ z2 y2 =G0 z2 , as required.
By Claims 1 and 2, it follows that
w1 =G0 z12 ≡: w1′ ,
w2 =G0 z2 z1−1 ≡: w2′ ,
w3 =G0 bX b̄z1 ≡: w3′ ,
w4 =G0 z2 z1−1 bX̄ b̄ ≡: w4′ ,
w5 = w2−1 =G0 (w2′ )−1 ≡: w5′ ,
w6 = w3−1 =G0 (w3′ )−1 ≡: w6′ ,
so that
(f (ur0 )) = (w1 · · · w6 ) =G0 (w1′ · · · w6′ ).
10
Moreover, we see that w1′ ≡ a · · · b̄, w2′ ≡ a · · · ā, w3′ ≡ b · · · b̄, w4′ ≡ a · · · b̄, w5′ ≡
a · · · ā and w6′ ≡ b · · · b̄ are alternating words in {a, b} such that
S(w1′ ) = (S(z1 ), S(z1 )) = (4, 5, 4, 5, 4, 5, 4, 5),
S(w2′ ) = (S(z2 ), S(z1−1 )) = (4, 4, 5, 4, 5, 4, 5, 4),
S(w3′ ) = (S(bX b̄), S(z1 )) = (5, 4, 4, 5, 4, 5, 4, 5, 4, 5),
S(w4′ ) = (S(z2 ), S(z1−1 ), S(bX̄ b̄)) = (4, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 4, 5),
S(w5′ ) = S((w2′ )−1 ) = (4, 5, 4, 5, 4, 5, 4, 4),
S(w6′ ) = S((w3′ )−1 ) = (5, 4, 5, 4, 5, 4, 5, 4, 4, 5).
This implies that
CS(w1′ · · · w6′ ) = ((S(w1′ ), . . . , S(w6′ ))).
Following Notation 3.2, we also have
T (w1′ ) = (1, 1, 1, 1),
T (w2′ ) = (2, 1, 1, 1),
T (w3′ ) = (2, 1, 1, 1),
T (w4′ ) = (2, 1, 1, 1, 1, 2),
T (w5′ ) = (1, 1, 1, 2),
T (w6′ ) = (1, 1, 1, 2),
and that
CT (w1′ · · · w6′ ) = ((T (w1′ ), . . . , T (w6′ ))).
We furthermore have
V (w1′ ) = (4),
V (w2′ ) = (3),
V (w3′ ) = (3),
V (w4′ ) = (4),
V (w5′ ) = (3),
V (w6′ ) = (3),
and
CV (w1′ · · · w6′ ) = ((V (w1′ ), . . . , V (w6′ ))) = ((4, 3, 3, 4, 3, 3)).
Since ((4, 3, 3, 4, 3, 3)) is the CS-sequence corresponding to the rational number
[3, 3], we see that
(w1′ · · · w6′ ) ≡ (u±1
r )
′′
for some rational number r with r = [3, 3]. For this rational number r, since
CS(r ′ ) = CT (r) = CT (w1′ · · · w6′ ) consists of 1 and 2, we have r ′ = [1, 4, 3]. Furthermore since CS(r) = CS(w1′ · · · w6′ ) consists of 4 and 5, we finally have r = [4, 2, 4, 3]
which equals r1 in the statement of the theorem. This completes the proof of
Lemma 3.3.
Lemma 3.4. Under the foregoing notation, (f (uri )) =G0 (u±1
ri+1 ) for every i ≥ 1.
Proof. Fix i ≥ 1. Then ri = [4, 2, m3 , . . . , mk ] with m3 ≥ 3. By Lemma 2.3(2),
CS(ri ) consists of 4 and 5 without (5, 5). Moreover, since ri′ = [1, m3 , . . . , mk ], by
Lemmas 2.3(2) and 2.5, CT (ri ) = CS(ri′ ) consists of 1 and 2, which implies that
the number of occurrences of 4’s between any two 5’s is one or two.
11
Claim. By cutting the cyclic word (uri ) in the middle of each positive or negative
subwords of length 5, we may write (uri ) as a product (vi,1 · · · vi,ki ), where each vi,j
is one of the following:
v1 ≡ b̄ababāb̄āb̄ab,
v2 ≡ v1−1 ≡ b̄ābabab̄āb̄āb,
v3 ≡ ababāb̄āb̄ababāb̄ā,
v4 ≡ v3−1 ≡ abab̄āb̄ābabab̄āb̄ā,
v5 ≡ b̄ababāb̄ā,
v6 ≡ b̄ābabab̄āb̄ā,
v7 ≡ v6−1 ≡ ababāb̄āb̄ab,
v8 ≡ v5−1 ≡ abab̄āb̄āb.
Proof of Claim. Note that for every n = 1, . . . , 8, vn is an alternating word in {a, b}
such that S(vn ) = (kn , tn h4i, ℓn ), where tn ∈ {1, 2} and kn , ℓn ∈ {1, 2, 3, 4}. Consider
the graph as in Figure 1, where the vertex set is equal to {v1 , . . . , v8 } and each
edge is endowed with one or two orientations. Observe that if vn and vm are the
initial and terminal vertices, respectively, of an oriented edge of the graph, then
the word vn vm is an alternating word such that S(vn vm ) = (kn , tn h4i, 5, tm h4i, ℓm ),
namely, the terminal subword of vn , corresponding to the last component ℓn of
S(vn ), and the initial subword of vm , corresponding to the first component km of
S(vm ), are amalgamated into a maximal positive or negative alternating subword
of vn vm , of length 5. Moreover, the weight tn (resp. tm ) is 1 or 2 according to
whether the vertex vn (resp. vm ) has valence 3 or 2. Thus, if vn1 , vn2 , . . . , vnp , where
vnj ∈ {v1 , . . . , v8 }, is a closed edge path in the graph which is compatible with
the specified edge orientations (a compatible closed edge path, in brief), namely,
if vnj and vnj+1 are the initial and terminal vertices of an oriented edge of the
graph for each j = 1, 2, . . . , p, where the indices are considered modulo p, then
the cyclically reduced word vn1 vn2 · · · vnp is a cyclically alternating word with CSsequence ((5, tn1 h4i, 5, tn2 h4i, 5, . . . , tnp h4i)).
Since the weight tnj is 1 or 2 according to whether the vertex vnj has valence 3 or 2,
we see that for any compatible closed edge path, the CT -sequence ((tn1 , tn1 , . . . , tnp ))
of the corresponding cyclically alternating word consists of only 1 and 2 and that it
has isolated 2’s. Moreover, for any such cyclic sequence, we can construct a compatible closed edge path such that the CT -sequence of the corresponding cyclically alternating word is equal to the given cyclic sequence. In particular, we can find a compatible closed edge path such that the CT -sequence of the corresponding cyclically
alternating word, (w), is equal to CT (uri ). This implies that CS(w) = CS(uri ).
12
v1
v8
v3
v6
v2
v7
v4
v5
Figure 1. Proof of Claim in the proof of Lemma 3.4
Hence (w) ≡ (u±1
ri ) as cyclic words by [9, Lemma 3.2]. This completes the proof of
Claim.
Putting
w1 :≡ bX̄ b̄X̄ b̄XbXbX̄ b̄,
w3 :≡ X̄ b̄X̄ b̄XbXbX̄ b̄X̄ b̄XbX,
w5 :≡ bX̄ b̄X̄ b̄XbX,
w6 :≡ bX b̄X̄ b̄X̄bXbX,
w2 :≡ w1−1 ,
w4 :≡ w2−1 ,
w7 :≡ w6−1
and
w8 :≡ w5−1 ,
we obviously have f (vn ) = wn for every n = 1, 2, . . . , 8, so that
(f (uri )) = (f (vi,1 · · · vi,ki )) = (wi,1 · · · wi,ki ),
where each wi,j ∈ {w1 , . . . , w8 }.
Recall from Claims 1 and 2 in the proof of Lemma 3.3 that X̄ b̄X̄bXbX b̄ =G0 z1 ,
where z1 ≡ a · · · b̄ is the alternating word in {a, b} with S(z1 ) = (4, 5, 4, 5), and
that X̄ b̄X̄ b̄XbX =G0 z2 , where z2 ≡ a · · · ā is the alternating word in {a, b} with
S(z2 ) = (4, 4, 5, 4). It follows that
w1 =G0 z1−1 bX̄ b̄ ≡: w1′ ,
w3 =G0 z2 z1−1 ≡: w3′ ,
w5 =G0 z1−1 ≡: w5′ ,
w6 =G0 bX b̄z2−1 ≡: w6′ ,
w2 = w1−1 =G0 (w1′ )−1 ≡: w2′ ,
w4 = w3−1 =G0 (w3′ )−1 ≡: w4′
w7 = w6−1 =G0 (w6′ )−1 ≡: w7′
w8 = w5−1 =G0 (w5′ )−1 ≡: w8′ .
13
Then we have
′
′
(f (uri )) =G0 (wi,1
· · · wi,k
),
i
′ ∈ {w ′ , . . . , w ′ }. Moreover
where each wi,j
1
8
w1′ ≡ b · · · b̄,
w2′ ≡ b · · · b̄,
w3′ ≡ a · · · ā,
w4′ ≡ a · · · ā,
w5′ ≡ b · · · ā,
w6′ ≡ b · · · ā,
w7′ ≡ a · · · b̄,
w8′ ≡ a · · · b̄
are alternating words in {a, b} such that
S(w1′ ) = (S(z1−1 ), S(bX̄ b̄)) = (5, 4, 5, 4, 5, 4, 5, 4, 4, 5)
S(w3′ ) = (S(z2 ), S(z1−1 )) = (4, 4, 5, 4, 5, 4, 5, 4),
S(w5′ ) = S(z1−1 ) = (5, 4, 5, 4),
S(w6′ ) = (S(bX b̄), S(z2−1 )) = (5, 4, 4, 5, 4, 5, 4, 5, 4, 4),
and S(w2′ ) = S((w1′ )−1 ) = (5, 4, 4, 5, 4, 5, 4, 5, 4, 5), S(w4′ ) = S((w3′ )−1 ) = (4, 5, 4, 5, 4, 5, 4, 4),
S(w7′ ) = S((w6′ )−1 ) = (4, 4, 5, 4, 5, 4, 5, 4, 4, 5), and S(w8′ ) = S((w5′ )−1 ) = (4, 5, 4, 5).
Observe in the graph in Figure 1 that if vn and vm are the initial and terminal
′ is an alternating word such
vertices, respectively, of an oriented edge, then wn′ wm
′
′
′
′
that S(wn wm ) = (S(wn ), S(wm )), which consists of 4 and 5, and moreover the
components 5 are isolated. This observation yields that
′
′
′
′
))),
) = ((S(wi,1
), . . . , S(wi,k
CS(wi,1
· · · wi,k
i
i
′
′
′
′
CT (wi,1
· · · wi,k
) = ((T (wi,1
), . . . , T (wi,k
))).
i
i
Here
T (w1′ ) = (1, 1, 1, 2),
T (w2′ ) = (2, 1, 1, 1),
T (w3′ ) = (2, 1, 1, 1),
T (w4′ ) = (1, 1, 1, 2),
T (w5′ ) = (1, 1),
T (w6′ ) = (2, 1, 1, 2),
T (w7′ ) = (2, 1, 1, 2),
T (w8′ ) = (1, 1).
This also yields that
′
′
′
′
CV (wi,1
· · · wi,k
) = ((V (wi,1
), . . . , V (wi,k
))),
i
i
where V (wn′ ) = (3) if n = 1, 2, 3, 4, and V (wn′ ) = (2) otherwise.
Define N (vn ) to be the number of positive or negative proper subwords of vn of
length 4 for each n = 1, . . . , 8. Here, by a proper subword of vn , we mean a subword
which lies in the interior of vn . Then we see that V (wn′ ) = (N (vn ) + 1) for each
n = 1, . . . , 8. Since (vi,1 · · · vi,ki ) is a product being cut in the middle of each positive
or negative subwords of length 5, we also see that
((N (vi,1 ), . . . , N (vi,ki ))) = CT (ri ) = CS(ri′ )
14
′ ) = (N (v ) + 1) for each j = 1, . . . , k ,
with ri′ = [1, m3 , . . . , mk ]. Since V (wi,j
i,j
i
′ · · · w ′ ) = ((N (v )+1, . . . , N (v
CV (wi,1
i,1
i,ki )+1)) is the CS-sequence corresponding
i,ki
to the rational number [2, m3 , . . . , mk ]. Hence
′
′
) ≡ (u±1
(f (uri )) =G0 (wi,1
· · · wi,k
r )
i
for some rational number r with r ′′ = [2, m3 , . . . , mk ]. For this rational number
′ · · · w ′ ) consists of 1 and 2, we have r ′ =
r, since CS(r ′ ) = CT (r) = CT (wi,1
i,ki
′ · · · w ′ ) consists of 4 and
[1, 3, m3 , . . . , mk ]. Furthermore, since CS(r) = CS(wi,1
i,ki
5, we finally have r = [4, 2, 3, m3 , . . . , mk ] which equals ri+1 in the statement of the
theorem. This completes the proof of Lemma 3.4.
Since G = ha, b | ur0 = ur1 = ur2 = · · · = 1i, Lemmas 3.1–3.4 imply that f
descends to an epimorphism fˆ : G → G. Now we show that fˆ is not an isomorphism.
Let s = [3, 3, 4]. Then
CS(us ) = CS(s) = ((3, 3, 3, 4, 3, 3, 4, 3, 3, 4, 3, 3, 4, 3, 3, 3, 4, 3, 3, 4, 3, 3, 4, 3, 3, 4)),
so that
(us ) ≡ (abab̄āb̄abab̄āb̄ābabāb̄ābabab̄āb̄abab̄āb̄ābabāb̄ābaba
b̄āb̄abab̄āb̄ababāb̄ābabāb̄āb̄abab̄āb̄ababāb̄ābabāb̄āb̄).
As in the proof of Lemma 3.1, letting w ≡ a · · · ā be an alternating word in {a, b}
such that S(w) = (3, 3, 3, 4, 3, 3), we have
(us ) ≡ (wbwbabab̄āb̄w−1 b̄w−1 b̄).
Lemma 3.5. We have fˆ(us ) = 1.
Proof. Clearly
(f (us )) = (f (w)b̄f (w)b̄X̄ b̄X̄bXbf (w−1 )bf (w−1 )b).
Here, since fˆ(w) = ā from the proof of Lemma 3.1, we have
(f (us )) =G (āb̄āb̄X̄ b̄X̄bXbabab),
where (āb̄āb̄X̄ b̄X̄bXbabab) is a cyclically alternating word in {a, b} such that
CS(āb̄āb̄X̄ b̄X̄bXbabab) = ((S(āb̄āb̄), S(X̄ b̄), S(X̄), S(bX), S(babab))),
= ((4, (4, 4, 5, 4, 4, 5), (4, 4, 5, 4, 4, 4), (5, 4, 4, 5, 4, 4), 5))
which equals CS(r0 ). This implies that (f (us )) =G (āb̄āb̄X̄ b̄X̄bXbabab) =G 1,
namely fˆ(us ) = 1, as required.
Lemma 3.6. Under the foregoing notation, let R be the symmetrized subset of
F (a, b) generated by the set of relators {uri | i ≥ 0} of the upper presentation G =
ha, b | ur0 = ur1 = ur2 = · · · = 1i. Then R satisfies C(4) and T (4).
15
Proof. Since every element in R is cyclically alternating, R clearly satisfies T (4).
To show that R satisfies C(4), we begin by setting some notation. Recall from
Lemma 2.6 that for every rational number r with 0 < r ≤ 1, CS(r) has a decomposition ((S1 , S2 , S1 , S2 )) depending on r. For clarity, we write ((S1 (r), S2 (r), S1 (r), S2 (r)))
for this decomposition. On the other hand, if r is a rational number with r =
[m1 , . . . , mk ] with k ≥ 2 and (m2 , . . . , mk ) ∈ (Z≥2 )k−1 , then the symbol r (n) denotes
the rational number with continued fraction expansion [mn+1 − 1, mn+2 , . . . , mk ] for
each n = 1, . . . , k − 1.
Claim 1. For any two integers i, j ≥ 0 with i 6= j, the cyclic word (urj ) does not
contain a subword corresponding to (S1 (ri )) or (ℓ1 , S2 (ri ), ℓ2 ) with ℓ1 , ℓ2 ∈ Z+ .
Proof of Claim 1. Suppose on the contrary that there are some i 6= j such that the
cyclic word (urj ) contains a subword corresponding to S1 (ri ) or (ℓ1 , S2 (ri ), ℓ2 ). We
first show that this assumption implies that CS(rj ) contains S1 (ri ) or S2 (ri ) as
a subsequence. If (urj ) contains a subword corresponding to (ℓ1 , S2 (ri ), ℓ2 ), then
clearly CS(rj ) contains S2 (ri ) as a subsequence. So assume that (urj ) contains a
subword corresponding to S1 (ri ). Then CS(rj ) contains (d1 + s1 , s2 , . . . , st−1 , st +
d2 ) as a subsequence, where S1 (ri ) = (s1 , s2 , . . . , st ). Since the continued fraction
expansions of both ri and rj begin with 4, we see that S1 (ri ) begins and ends with 5
by Lemma 2.6(3) and that CS(rj ) also consists of 4 and 5 by Lemma 2.3(2). Hence,
we must have d1 = d2 = 0 and therefore CS(rj ) contains S1 (ri ) as a subsequence.
Thus we have proved that CS(rj ) contains S1 (ri ) or S2 (ri ) as a subsequence.
Note that the lengths of the continued fraction expansions of ri and rj are i+3 and
(n)
j + 3, respectively. Hence we can apply Lemma 2.8 successively to see that CS(rj )
(n)
(n)
contains S1 (ri ) or S2 (ri ) as a subsequence for every n = 1, . . . , min{i + 1, j + 1}.
Since i 6= j, there are two cases.
Case 1. j > i ≥ 0. Recall that ri is equal to [4, 3, 3] or ri = [4, 2, (i − 1)h3i, 4, 3]
(i+1)
according to whether i = 0 or i ≥ 1. So we have ri
= [m, 3]. Here, m = 2 if
(i+1)
i = 0, and m = 3 otherwise. Since j > i, we can observe that rj
has a continued
fraction expansion of the form [m − 1, n1 , . . . , nk ], where k ≥ 2 and each nt is 3 or 4.
(i+1)
(i+1)
) consists of m−1 and m, the cyclic sequence
) = (m+1) and CS(rj
Since S1 (ri
(i+1)
CS(rj
(i+1)
) cannot contain S1 (ri
must contain
(i+1)
S2 (ri
)
(i+1)
) = (m + 1) as a subsequence. Hence CS(rj
as a subsequence. But since
n1 ≥ 3, (m, m) does not occur in
(i+1)
CS(rj
)
(m, m) by Lemma 2.7(1), this implies that
contradiction.
(i+1)
rj
= [m − 1, n1 , . . . , nk ] with
(i+1)
by Lemma 2.3(2). Since S2 (ri
(i+1)
S2 (ri
)
cannot occur in
(j+1)
Case 2. i > j ≥ 0. As in Case 1, we can observe that rj
if j = 0, and m = 3 otherwise, and that
(j+1)
ri
16
)
)=
(i+1)
CS(rj
),
a
= [m, 3], where m = 2
has a continued fraction expansion of
(j+1)
the form [m − 1, n1 , . . . , nk ], where k ≥ 2 and each nt is 3 or 4. Then both S1 (ri
)
(j+1)
(j+1)
and S2 (ri
) contain a term m−1 by Lemma 2.7(2). But since CS(rj
) consists
of only m and m + 1, this is impossible.
By Claim 1, we see that the assertion in Lemma 2.11 holds even if (u±1
r ) is
)
for
any
i
≥
0
and
the
symmetrized
subset
R
in
the
lemma
is
replaced by (u±1
ri
enlarged to be the set in the current setting, namely, R is the symmetrized subset
of F (a, b) generated by the set of relators {uri | i ≥ 0} of the group presentation
G = ha, b | ur0 = ur1 = ur2 = · · · = 1i. To be precise, the following hold.
Claim 2. For each i ≥ 0, a subword w of the cyclic word (u±1
ri ) is a piece with
respect to the symmetrized subset R in Lemma 3.6 if and only if S(w) contains
neither S1 (ri ) nor (ℓ1 , S2 (ri ), ℓ2 ) with ℓ1 , ℓ2 ∈ Z+ as a subsequence.
By using Claim 2, we can see, as in [9, Proof of Corollary 5.4], that each cyclic
word (u±1
ri ) is not a product of less than 4 pieces with respect to R. Hence R satisfies
C(4).
Lemma 3.7. Under the foregoing notation, us 6=G 1.
Proof. Suppose on the contrary that us =G 1. Then there is a reduced van Kampen
diagram ∆ over G = ha, b | ur0 = ur1 = ur2 = · · · = 1i such that (φ(∂∆)) ≡ (us ) (see
[11]). Since ∆ is a [4, 4]-map by Lemma 3.6, (φ(∂∆)) contains a subword of some
(u±1
ri ) which is a product of 3 pieces with respect to the symmetrized subset R in
Lemma 3.6 (see [9, Section 6]). This implies that CS(φ(∂∆)) must contain a term
5, which is a contradiction to the fact CS(φ(∂∆)) = CS(us ) = CS(s) consists of
only 3 and 4.
Lemma 3.7 together with Lemma 3.5 shows that fˆ is an epimorphism of G,
but not an isomorphism of G. Consequently, G is non-Hopfian, and the proof of
Theorem 1.1 is now completed.
References
[1] G. Baumslag and D. Solitar, Some two-generator one-relator non-Hopfian groups, Bull. Amer.
Math. Soc. 68 (1962), 199–201.
[2] R. Coulon and V. Guirardel, Automorphisms and endomorphisms of lacunary hyperbolic groups,
arXiv:1606.00679.
[3] B. H. Bowditch, Relatively hyperbolic groups, Int. J. Algebra Comput. 22 (2012), 1250016, 66pp.
[4] B. Farb, Relatively hyperbolic groups, Geom. Funct. Anal. 8 (1998), 810–840.
[5] M. Gromov, Hyperbolic groups, Essays in Group Theory (G. M. Gersten, ed.), MSRI Publ. 8,
Springer, 1987, 75–263.
[6] D. Groves, Limit groups for relatively hyperbolic groups, II: Makanin-Razborov diagrams, Geom.
Topol. 9 (2005), 2319–2358.
[7] G. Higman, A finitely related group with an isomorphic proper factor group, J. London Math.
Soc. 26 (1951), 59–61.
17
[8] S. V. Ivanov and A. M. Storozhev, Non-Hopfian relatively free groups, Geom. Dedicata 114
(2005), 209–228.
[9] D. Lee and M. Sakuma, Epimorphisms between 2-bridge link groups: homotopically trivial simple
loops on 2-bridge spheres, Proc. London Math. Soc. 104 (2012), 359–386.
[10] D. Lee and M. Sakuma, Homotopically equivalent simple loops on 2-bridge spheres in 2-bridge
link complements (I), Geom. Dedicata 171 (2014), 1–28.
[11] R. C. Lyndon and P. E. Schupp, Combinatorial group theory, Springer-Verlag, Berlin, 1977.
[12] A. I. Mal’cev, On the faithful representation of infinite groups by matrices, Mat. Sb. 8(50)
(1940), 405–422.
[13] B. H. Neumann, A two-generator group isomorphic to a proper factor group, J. London Math.
Soc. 25 (1950), 247–248.
[14] D. Osin, Peripheral fillings of relatively hyperbolic groups, Invent. Math. 167 (2007), no. 2,
295–326.
[15] D. Osin, Relatively hyperbolic groups: Intrinsic geometry, algebraic properties, and algorithmic
problems, Memoirs Amer. Math. Soc. 179 (2006), no. 843, vi+100 pp.
[16] C. Reinfeldt, Limit groups and Makanin-Razborov diagrams for hyperbolic groups, PhD thesis,
Heriot-Watt University, 2010.
[17] C. Reinfeldt and R. Weidmann, Makanin-Razborov diagrams for hyperbolic groups, Preprint,
2010, updated 2014.
[18] M. Sapir and D. T. Wise, Ascending HNN extensions of residually finite groups can be nonHopfian and can have very few finite quotients, J. Pure Appl. Algebra 166 (2002), 191–202.
[19] Z. Sela, Endomorphisms of hyperbolic groups. I, The Hopf property, Topology 38 (1999), 301–
321.
[20] R. Strebel, Appendix. Small cancellation groups, in “Sur les groupes hyperboliques d’après
Mikhael Gromov”, Papers from the Swiss Seminar on Hyperbolic Groups held in Bern, 1988,
E. Ghys and P. de la Harpe (editors), Progr. Math., vol. 83, Birkhäuser Boston Inc., Boston,
MA, 1990.
[21] D. T. Wise, A non-Hopfian automatic group, J. Algebra 180 (1996), 845–847.
[22] D. T. Wise, Research announcement: the structure of groups with a quasiconvex hierarchy,
Electron. Res. Announc. Math. Sci. 16 (2009), 44–55.
Department of Mathematics, Pusan National University, San-30 Jangjeon-Dong,
Geumjung-Gu, Pusan, 609-735, Korea
E-mail address: [email protected]
Department of Mathematics, Graduate School of Science, Hiroshima University,
Higashi-Hiroshima, 739-8526, Japan
E-mail address: [email protected]
18
| 4math.GR
|
1
Energy-Efficient Transponder Configuration for
FMF-based Elastic Optical Networks
arXiv:1705.10198v1 [cs.IT] 26 May 2017
Mohammad Hadi, Member, IEEE, and Mohammad Reza Pakravan, Member, IEEE
Abstract—We propose an energy-efficient procedure for
transponder configuration in FMF-based elastic optical networks
in which quality of service and physical constraints are guaranteed and joint optimization of transmit optical power, temporal,
spatial and spectral variables are addressed. We use geometric
convexification techniques to provide convex representations for
quality of service, transponder power consumption and transponder configuration problem. Simulation results show that our
convex formulation is considerably faster than its mixed-integer
nonlinear counterpart and its ability to optimize transmit optical
power reduces total transponder power consumption up to 32%.
We also analyze the effect of mode coupling and number of
available modes on power consumption of different network
elements.
Keywords—Convex optimization, Green communication, Elastic
optical networks, Few-mode fibers, Mode coupling.
I. I NTRODUCTION
D temporally, spectrally and spatially Elastic Optical Network (EON) has been widely acknowledged as the next
generation high capacity transport system and the optical
society has focused on its architecture and network resource allocation techniques [1]. EONs can provide an energy-efficient
network configuration by adaptive 3D resource allocation
according to the communication demands and physical conditions. Higher energy efficiency of Orthogonal Frequency
Division Multiplex (OFDM) signaling has been reported in [2]
which nominates OFDM as the main technology for resource
provisioning over 2D resources of time and spectrum. On the
other hand, enabling technologies such as Few-Mode Fibers
(FMFs) and Multi-Core Fibers (MCFs) have been used to
increase network capacity and efficiency through resource allocation over spatial dimension [1], [3]. Although many variants
of algorithms have been proposed for resource allocation in
1D/2D EONs [4], joint assignment of temporal, spectral and
spatial resources in 3D EONs needs much research and study.
Among the available works on 3D EONs, a few of them
have focused on energy-efficiency which is a fundamental
requirement of the future optical networks [5]–[7]. Moreover,
the available energy-efficient 3D approaches do not consider
transmit optical power as an optimization variable which
results in inefficient network provisioning [8].
Flexible resource allocation is an NP-hard problem and it
is usually decomposed into several sub-problems with lower
complexity [9]. Following this approach, we decompose the
resource allocation problem into 1) Routing and Ordering
Sub-problem (ROS) and 2) Transponder Configuration Subproblem (TCS) and mainly focus on TCS which is more complex and time-consuming [10], [11]. We consider FMF because
3
it has simple amplifier structure, easier fusion process, lower
nonlinear effects and lower manufacturing cost compared to
other Space-Division Multiplexed (SDM) optical fibers [3],
[12]. In TCS, we optimally configure transponder parameters
such as modulation level, number of sub-carriers, coding rate,
transmit optical power, number of active modes and central
frequency such that total transponders power consumption is
minimized while Quality of Service (QoS) and physical constraints are met. Unlike the conventional approach, we provide
convex expressions for transponder power consumption and
Optical Signal to Noise Ratio (OSNR), as an indicator of
QoS. We then use the results to formulate TCS as a convex
optimization problem which can efficiently be solved using fast
convex optimization algorithms. We consider transmit optical
power as an optimization variable and show that it has an
important impact on total transponder power consumption.
Simulation results show that our convex formulation can be
solved almost 20 times faster than its Mixed-Integer NonLinear
Program (MINLP) counterpart. Optimizing transmit optical
power also improves total transponder power consumption
by a factor of 32% for European Cost239 optical network
with aggregate traffic 60 Tbps. We analyze the effect of
mode coupling on power consumption of the different network
elements. As simulation results show, total network power
consumption can be reduced more than 50% using stronglycoupled FMFs rather than weakly-coupled ones. Numerical
outcomes also demonstrate that increasing the number of
available modes in FMFs provides a trade-off between FFT
and DSP power consumption such that the overall transponder
power consumption is a descending function of the number of
available modes.
II. S YSTEM M ODEL
Consider a coherent optical communication network characterized by topology graph G(V, L) where V and L are the
sets of optical nodes and directional optical strongly-coupled
FMF links, respectively. The optical FMFs have M modes
and gridless bandwidth B. Q is the set of connection requests
and Ql shows the set of requests sharing FMF l on their
routes. Each request q is assigned a contiguous bandwidth
∆q around carrier frequency ωq and modulates mq modes of
its available M modes. The assigned contiguous bandwidth
includes 2bq OFDM sub-carriers with sub-carrier space of
F so, ∆q = 2bq F. To have a feasible MIMO processing,
the remaining unused modes of a request cannot be shared
among others [12]. We assume that the assigned bandwidths
are continuous over their routes to remove the high cost of
spectrum/mode conversion [13]. Request q passes Nq fiber
spans along its path and has Nq,i shared spans with request
i 6= q. Each FMF span has fixed length of Lspn and an
Mode DEMUX
MIMO
PD
ADC
S/P
FFT
+
EQ
SM
P/S
DEC
PD
ADC
S/P
FFT
+
EQ
SM
P/S
DEC
PD
ADC
S/P
FFT
+
EQ
SM
P/S
DEC
Mode MUX
LO
Optical Mixer
LO
Optical Mixer
2
LD
DAC
P/S
IFFT
SM
S/P
ENC
LD
DAC
P/S
IFFT
SM
S/P
ENC
LD
DAC
P/S
IFFT
SM
S/P
ENC
Fig. 1: Block diagram of a pair of OFDM-based transmit and receive transponders with
3 available modes.
optical amplifier to compensate for its attenuation. There are
pre-defined modulation levels c and coding rates r where
each pair of (c, r) requires minimum OSNR Θ(c, r) to get
a pre-FEC BER value of 1 × 10−4 [10]. Each transponder
is given modulation level cq , coding rate rq and injects
optical power pq /mq to each active mode of each polarization.
Chromatic dispersion and mode coupling signal broadenings
p
are respectively proportional to Nq 2bq and m−0.78
q
p Nq with
coefficients σ = 2π|β2 | FLspn and % = 5∆β1 L
√sec Lspn
where β2 is chromatic dispersion factor and ∆β1 Lsec is
the product of rms uncoupled group delay spread and section
length [14]. Transponders add a sufficient cyclic prefix to each
OFMD symbol to resolve the signal broadening induced by
mode coupling and chromatic dispersion. Transponders have
maximum information bit rate C. There is also a guard band
G between any two adjacent requests on a link. Considering
the architecture of Fig. 1, the power consumption of each pair
of transmit and receive transponders Pq can be calculated as
follows:
Pq = Ptrb + 2Pedc mq rq−1 + 2mq 2bq bq Pf f t + 2m2q 2bq Pdsp
(1)
where Ptrb is transmit and receive transponder bias term,
Pedc is the scaling coefficient of encoder and decoder power
consumptions, Pf f t denotes the power consumption for a
two point FFT operation and Pdsp is the power consumption
scaling coefficient of the receiver DSP and MIMO operations
[2], [12].
To have a green EON, we need a resource allocation
algorithm to determine the values of system model variables
such that the transponders consume the minimum power while
physical constrains are satisfied and desired levels of OSNR are
guaranteed. In general, such a problem is modeled as an NPhard MINLP optimization problem [10]. To simplify the problem and provide a fast-achievable near-optimum solution, the
resource allocation problem is usually decomposed into two
sub-problems: ROS, where the routing and ordering of requests
on each link are defined, and 2) TCS, where transponders
are configured. Usually the search for a near optimal solution
involves iterations between these two sub-problems. To save
this iteration time, it is of great interest to hold the running
time of each sub-problem at its minimum value. In this work,
we mainly focus on TCS which is the most time-consuming
sub-problem and formulate it as a convex problem to benefit
from fast convex optimization algorithms. For a complete study
of ROS, one can refer to [8], [11].
III.
T RANSPONDER C ONFIGUARTION P ROBLEM
A MINLP formulation for TCS is as follows:
min
c,b,r,p,m,ω
X
Pq
(2)
q∈Q
s.t. Ψq > Θq , ∀q ∈ Q
∆Υl,j + ∆Υl,j+1
|Q |
+ G 6 ωΥl,j+1 , ∀l ∈ L, ∀j ∈ M1 l
ωΥl,j +
2
∆q
∆q
6 ωq 6 B −
, ∀q ∈ Q
2
2
−1
2F mq rq cq ∆q
p , ∀q ∈ Q
Rq 6
F −1 + σNq 2bq + %m−0.78
Nq
q
(3)
(4)
(5)
(6)
where c, b, r, p, m and ω are variable vectors of transponder
configuration parameters i.e. modulation level, number of subcarriers, coding rate, transmit optical power, number of active
modes and central frequency. Mba shows the set of integer
numbers {a, a+1, · · · , b−1}. The goal is to minimize the total
transponder power consumption where Pq is obtained using
(1). Constraint (3) is the QoS constraint that forces OSNR Ψq
to be greater than its required minimum threshold Θq . Ψq is a
nonlinear function of b, p, m and ω while the value of Θq is
related to rq and cq [7], [11]. Constraint (4) is nonoverlappingguard constraint that prevents two requests from sharing the
same frequency spectrum. Υl,j is a function that shows which
request occupies j-th assigned spectrum bandwidth on link
l and its values are determined by solving ROS [7], [11].
Constraint (5) holds all assigned central frequencies within
the acceptable range of the fiber spectrum. The last constraint
guarantees that the transponder can convey the input traffic
rate Rq in which wasted cyclic prefix times are considered.
Generally, this problem is a complex MINLP which is
NP-hard and cannot easily be solved in a reasonable time.
Therefore, we use geometric convexification techniques to
convert this MINLP to a mixed-integer convex optimization
problem and then use relaxation method to solve it. To have
a convex problem, we first provide a generalized posynomial
expression [15] for the optimization and then define a variable
change to convexify the problem. A posynomial expression for
OSNR of a request in 2D EONs has been proposed in [11]. We
simply consider each active mode as an independent source of
nonlinearity and incoherently add all the interferences [10].
Therefore the extended version of the posynomial OSNR
expression is:
Ψq =
p
ζNq ∆q + κ1 ς mqq
pq /mq
P
, ∀q ∈ Q (7)
pi 2
) Nq,i /∆i /dq,i
mi ( m
i
i∈Q,q6=i
2
3γ
.
where κ1 = 0.4343, ζ = (eαLspn − 1)hνnsp and ς = 2απ|β
2|
nsp is the spontaneous emission factor, ν is the light frequency,
h is Planck’s constant, α is attenuation coefficient, β2 is
dispersion factor and γ is nonlinear constant. Furthermore,
dq,i is the distance between carrier frequencies ωq and ωi and
equals to dq,i = ωq − ωi . We use Θq ≈ rqκ2 (1 + κ3 cq )κ4
for posynomial curve fitting of OSNR threshold values where
κ2 = 3.37, κ3 = 0.21, κ4 = 5.73 [8]. Following the same
approach as [11], we arrive at this new representation of the
3
optimization problem:
X
min
c,b,r,p,m,ω,t,d
Pq + K
X
d−1
q,i
(8)
q,i∈Q
q6=i,Nq,i 6=0
q∈Q
h
bq
s.t. rqκ2 tκq 4 ζF Nq mq p−1
+ κ1 ςF −1
q 2
X
−bi
Nq,i p2i m−1
i 2
i∈Q,i6=q
i
d−1
q,i 6 1, ∀q ∈ Q
ωΥl,j + 0.5F2
bΥ
l,j
(9)
bΥ
+ G + 0.5F2
l,j+1
6 ωΥl,j+1 , ∀l ∈ L
|Q |
∈ M1 l
bq
, ∀j
0.5F2
(10)
+ ωq 6 B, ∀q ∈ Q
(11)
0.5F2bq 6 ωq , ∀q ∈ Q
(12)
−1 −bq
−1 −1
0.5Rq F −1 rq−1 c−1
+ 0.5σNq Rq m−1
q mq 2
q rq c q
p
−1 −1 −1.78 −bq
+ 0.5% Nq Rq rq cq mq
2
6 1, ∀q ∈ Q
(13)
(14)
1 + κ3 cq 6 tq , ∀q ∈ Q
|Ql |
dΥl,i ,Υl,j + ωΥl,j 6 ωΥl,i , ∀l ∈ L, ∀j ∈ M1
dΥl,i ,Υl,j + ωΥl,i 6 ωΥl,j , ∀l ∈ L, ∀j ∈
|Q |+1
(15)
Mj1
(16)
l
, ∀i ∈ Mj+1
|Q |+1
M2 l , ∀i
∈
Ignoring constraints (14), (15), (16) and the penalty term of
the goal function (8), the above formulation is equivalent geometric program of the previous MINLP in which expressions
(7) and the mentioned posynomial curve fitting have been
used for QoS constraint (9). Constraints (14) and (15) and
the penalty term are added to guarantee the implicit equality
of dq,i = ωq − ωi [11]. Constraint (14) is also needed to
convert the generalized posynomial QoS constraint to a valid
geometric expression, as explained in [15]. Now, consider the
following variable change:
/b
x = eX : x ∈ R>0 −→ X ∈ R, ∀x ∈
(17)
Applying this variable change to the goal function (which is
the most difficult part of the variable change), we have:
X
[Ptrb + 2Pedc emq −rq + 5.36e0.82bq +mq Pf f t + 2e2mq 2bq Pdsp ]
q∈Q
+K
X
e−dq,i
(18)
q,i∈Q,q6=i,Nq,i 6=0
Clearly, e−dq,i , emq −rq and e2mq 2bq are convex over variable
domain. We use expression 5.36e0.82bq +mq to provide a convex
approximation for the remaining term 2emq bq 2bq . The approximation relative error is less than 3% for practical values of
mq > 1 and 4 6 bq 6 11. Consequently, function (18) which
is a nonnegative weighted sum of convex functions is also convex. The same statement (without any approximation) can be
applied to show the convexity of the constraints under variable
change of (17) (for some constraints, we need to apply an extra
log to both sides of the inequality). To solve this problem,
a relaxed continuous version of the proposed mixed-integer
convex formulation is iteratively optimized in a loop [15]. At
each epoch, the continuous convex optimization is solved and
obtained values for relaxed integer variables are rounded by a
given precision. Then, we fix the acceptable rounded variables
and solve the relaxed continuous convex problem again. The
loop continues untill all the integer variables have valid values.
The number of iterations is at most equal to (in practice, is
usually less than) the number of integer variables. Furthermore,
a simpler problem should be solved as the number of iteration
increases because some of the integer variables are fixed during
each loop.
IV. N UMERICAL R ESULTS
In this section, we use simulation results to demonstrate the
performance of the convex formulation for TCS. The European
Cost239 optical network is considered with the topology and
traffic matrix given in [2]. Simulation constant parameters are
|β2 | = 20393 fs2 /m , α = 0.22 dB/km, Lspn = 80 km,
ν = 193.55 THz, nsp = 1.58, γ = 1.3 1/W/km, F = 80
MHz, % = 113 ps, σ = 14 fs, G = 20 GHz, B = 2 THz,
Ptrb = 36 W, Pedc = 3.2 W, Pf f t = 4 mW, Pdsp = 3
mW [2], [14]. We use MATLAB, YALMIP and CVX software
packages for programming, modeling and optimization. The
total power consumption of different network elements in
terms of aggregate traffic with and without adaptive transmit
optical power assignment has been reported in Fig. 2. We
have used the proposed approach of [16] for fixed assignment
of transmit optical power. Clearly, for all the elements, the
total power consumption is approximately a linear function
of aggregate traffic but the slope of the lines are lower
when transmit optical powers are adaptively assigned. As an
example, adaptive transmit optical power assignment improves
total transponder power consumption by a factor of 32%
for aggregate traffic of 60 Tbps. Fig. 3 shows total power
consumption of different network elements versus number of
available modes M in FMFs. The power consumption values
are normalized to their corresponding values for the scenario
with single mode fibers i.e. M = 1. As M increases, the
amount of transponder power consumption decreases but there
is no considerable gain for M > 5. Moreover, there is a
tradeoff between DSP and FFT power consumption such that
the overall transponder power consumption is a decreasing
function of the number of available modes. Fig. 4 shows power
consumption of different network elements in terms of aggregate traffic for strongly- and weakly-coupled FMFs. Obviously,
total transponder power consumption is considerably reduced
for strongly-coupled FMFs (in which group delay spread is
proportional to square root of path lengths) in comparison
to weakly-coupled FMFs (in which group delay spread is
proportional to path lengths). This is the same as the results
published in [12]. As an example, improvement can be more
than 50% for aggregate traffic of 60 Tbps. Numerical outcomes
also show that our convex formulation can be more than 20
times faster than its mixed-integer nonlinear counterpart which
is compatible with the results reported in [8].
V. C ONCLUSION
Energy-efficient resource allocation and quality of service
provisioning is the fundamental problem of green 3D FMFbased elastic optical networks. In this paper, we decompose
the resource allocation problem into two sub-problems for
routing and traffic ordering, and transponder configuration.
We mainly focus on transponder configuration sub-problem
and provide a convex formulation in which joint optimization
4
50
Total
Total
Total
Total
Total
Total
Total
Total
45
40
35
of temporal, spectral and spatial resources along with optical
transmit power are considered. Simulation results show that
our formulation is considerably faster than its mixed-integer
nonlinear counterpart and its ability to optimize transmit optical power can improve total transponder power consumption
up to 32%. We demonstrate that there is a tradeoff between
DSP and FFT power consumptions as the number of modes in
FMFs increases but the overall transponder power consumption
is a descending function of the number of available modes.
We also calculate the power consumption of different network
elements and show that strongly-coupled FMFs reduce the
power consumption of these elements.
Encoder+Decoder Power Consumption (Fixed Transmit Power)
FFT+IFFT Power Consumption (Fixed Transmit Power)
DSP Power Consumption (Fixed Transmit Power)
Transmit+Receive Transponder Power Consumption (Fixed Transmit Power)
Encoder+Decoder Power Consumption (Adaptive Transmit Power)
FFT+IFFT Power Consumption (Adaptive Transmit Power)
DSP Power Consumption (Adaptive Transmit Power)
Transmit+Receive Transponder Power Consumption (Adaptive Transmit Power)
kW
30
25
20
15
10
5
R EFERENCES
0
10
15
20
25
30
35
40
45
[1]
50
Aggregate Tra/c (Tbps)
Fig. 2: Total power consumption of different network elements in terms of aggregate
traffic with and without adaptive transmit optical power assignment.
[2]
1.8
[3]
1.7
1.6
[4]
1.5
1.4
[5]
1.3
[6]
1.2
Normalized
Normalized
Normalized
Normalized
1.1
1
Total
Total
Total
Total
Transmit+Receive Transponder Power Consumption
IFFT+FFT Power Consumption
Encoder+Decoder Power Consumption
DSP Power Consumption
[7]
0.9
[8]
0.8
1
2
3
4
5
6
Number of Available Modes in FMFs
Fig. 3: Normalized total power consumption of different network elements in terms of
the number of available modes in FMFs.
[9]
30
25
20
Total
Total
Total
Total
Total
Total
Total
Total
Encoder+Decoder Power Consumption (Strongly-Coupled FMF)
FFT+IFFT Power Consumption (Strongly-Coupled FMF)
DSP Power Consumption (Strongly-Coupled FMF)
Transmit+Receive Transponder Power Consumption (Strongly-Coupled FMF)
Encoder+Decoder Power Consumption (Weakly-Coupled FMF)
FFT+IFFT Power Consumption (Weakly-Coupled FMF)
DSP Power Consumption (Weakly-Coupled FMF)
Transmit+Receive Transponder Power Consumption (Weakly-Coupled FMF)
[10]
[11]
kW
[12]
15
[13]
10
[14]
5
[15]
0
10
15
20
25
30
35
40
45
50
Aggregate Tra/c (Tbps)
Fig. 4: Total power consumption of different network elements in terms of aggregate
traffic for strongly- and weakly-coupled FMFs.
[16]
R. Proietti, et al., “3d elastic optical networking in the temporal, spectral, and spatial domains,” IEEE Communications Magazine, vol. 53,
no. 2, pp. 79–87, 2015.
H. Khodakarami, et al., “Flexible optical networks: An energy efficiency
perspective,” Journal of Lightwave Technology, vol. 32, no. 21, pp.
3356–3367, 2014.
G. M. Saridis, et al., “Survey and evaluation of space division multiplexing: From technologies to optical networks,” IEEE Communications
Surveys & Tutorials, vol. 17, no. 4, pp. 2136–2156, 2015.
B. C. Chatterjee, N. Sarma, and E. Oki, “Routing and spectrum
allocation in elastic optical networks: A tutorial,” IEEE Communications
Surveys & Tutorials, vol. 17, no. 3, pp. 1776–1800, 2015.
A. Muhammad, et al., “Resource allocation for space-division multiplexing: optical white box versus optical black box networking,” Journal
of Lightwave Technology, vol. 33, no. 23, pp. 4928–4941, 2015.
P. J. Winzer, “Energy-efficient optical transport capacity scaling through
spatial multiplexing,” IEEE Photonics Technology Letters, vol. 23,
no. 13, pp. 851–853, 2011.
L. Yan, et al., “Joint assignment of power, routing, and spectrum in
static flexible-grid networks,” Journal of Lightwave Technology, no. 99,
2017.
M. Hadi and M. R. Pakravan, “Energy-efficient resource allocation
for elastic optical networks using convex optimization,” arXiv preprint
arXiv:1705.06891, 2017.
H. Khodakarami, B. S. G. Pillai, and W. Shieh, “Quality of service
provisioning and energy minimized scheduling in software defined
flexible optical networks,” Journal of Optical Communications and
Networking, vol. 8, no. 2, pp. 118–128, 2016.
L. Yan, et al., “Resource allocation for flexible-grid optical networks
with nonlinear channel model,” Journal of Optical Communications and
Networking, vol. 7, no. 11, pp. B101–B108, 2015.
M. Hadi and M. R. Pakravan, “Resource allocation for elastic optical networks using geometric optimization,” arXiv preprint
arXiv:1705.01191, 2017.
K. Ho and J. M. Kahn, “Mode coupling and its impact on spatially
multiplexed systems,” Optical Fiber Telecommunications VI, vol. 17,
pp. 1386–1392, 2013.
M. Hadi and M. R. Pakravan, “Spectrum-convertible bvwxc placement
in ofdm-based elastic optical networks,” IEEE Photonics Journal,
no. 99, pp. 1–7, 2017.
S. Ö. Arık, D. Askarov, and J. M. Kahn, “Adaptive frequency-domain
equalization in mode-division multiplexing systems,” Journal of Lightwave Technology, vol. 32, no. 10, pp. 1841–1852, 2014.
S. Boyd, et al., “A tutorial on geometric programming,” Optimization
and engineering, vol. 8, no. 1, p. 67, 2007.
G. Gao, et al., “Analytical expressions for nonlinear transmission
performance of coherent optical ofdm systems with frequency guard
band,” Journal of Lightwave Technology, vol. 30, no. 15, pp. 2447–
2454, 2012.
| 7cs.IT
|
arXiv:1407.4546v7 [math.ST] 28 Sep 2016
The Annals of Statistics
2016, Vol. 44, No. 5, 1931–1956
DOI: 10.1214/15-AOS1375
c Institute of Mathematical Statistics, 2016
CRAMÉR-TYPE MODERATE DEVIATIONS FOR STUDENTIZED
TWO-SAMPLE U -STATISTICS WITH APPLICATIONS
By Jinyuan Chang1,∗,† , Qi-Man Shao2,‡ and Wen-Xin Zhou3,§,†
Southwestern University of Finance and Economics,∗ University of
Melbourne,† Chinese University of Hong Kong‡ and Princeton University§
Two-sample U -statistics are widely used in a broad range of applications, including those in the fields of biostatistics and econometrics. In this paper, we establish sharp Cramér-type moderate deviation theorems for Studentized two-sample U -statistics in a general framework, including the two-sample t-statistic and Studentized
Mann–Whitney test statistic as prototypical examples. In particular,
a refined moderate deviation theorem with second-order accuracy is
established for the two-sample t-statistic. These results extend the
applicability of the existing statistical methodologies from the onesample t-statistic to more general nonlinear statistics. Applications to
Tribute: Peter was a brilliant and prolific researcher, who has made enormously influential contributions to mathematical statistics and probability theory. Peter had extraordinary knowledge of analytic techniques that he often applied with ingenious simplicity
to tackle complex statistical problems. His work and service have had a profound impact
on statistics and the statistical community. Peter was a generous mentor and friend with
a warm heart and keen to help the young generation. Jinyuan Chang and Wen-Xin Zhou
are extremely grateful for the opportunity to learn from and work with Peter in the last
two years at the University of Melbourne. Even in his final year, he had afforded time to
guide us. We will always treasure the time we spent with him. Qi-Man Shao is so grateful
for all the helps and supports that Peter had provided during the various stages of his
career. Peter will be dearly missed and forever remembered as our mentor and friend.
Received June 2015.
1
Supported in part by the Fundamental Research Funds for the Central Universities
(Grant No. JBK160159, JBK150501), NSFC (Grant No. 11501462), the Center of Statistical Research at SWUFE and the Australian Research Council.
2
Supported by Hong Kong Research Grants Council GRF 603710 and 403513.
3
Supported by NIH Grant R01-GM100474-4 and a grant from the Australian Research
Council.
AMS 2000 subject classifications. Primary 60F10, 62E17; secondary 62E20, 62F40,
62H15.
Key words and phrases. Bootstrap, false discovery rate, Mann–Whitney U test, multiple hypothesis testing, self-normalized moderate deviation, Studentized statistics, twosample t-statistic, two-sample U -statistics.
This is an electronic reprint of the original article published by the
Institute of Mathematical Statistics in The Annals of Statistics,
2016, Vol. 44, No. 5, 1931–1956. This reprint differs from the original in
pagination and typographic detail.
1
2
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
two-sample large-scale multiple testing problems with false discovery
rate control and the regularized bootstrap method are also discussed.
1. Introduction. The U -statistic is one of the most commonly used nonlinear and nonparametric statistics, and its asymptotic theory has been well
studied since the seminal paper of Hoeffding (1948). U -statistics extend the
scope of parametric estimation to more complex nonparametric problems
and provide a general theoretical framework for statistical inference. We refer to Koroljuk and Borovskich (1994) for a systematic presentation of the
theory of U -statistics, and to Kowalski and Tu (2007) for more recently
discovered methods and contemporary applications of U -statistics.
Applications of U -statistics can also be found in high dimensional statistical inference and estimation, including the simultaneous testing of many
different hypotheses, feature selection and ranking, the estimation of high
dimensional graphical models and sparse, high dimensional signal detection. In the context of high dimensional hypothesis testing, for example,
several new methods based on U -statistics have been proposed and studied
in Chen and Qin (2010), Chen, Zhang and Zhong (2010) and Zhong and
Chen (2011). Moreover, Li et al. (2012) and Li, Zhong and Zhu (2012) employed U -statistics to construct independence feature screening procedures
for analyzing ultrahigh dimensional data.
Due to heteroscedasticity, the measurements across disparate subjects
may differ significantly in scale for each feature. To standardize for scale,
unknown nuisance parameters are always involved and a natural approach
is to use Studentized, or self-normalized statistics. The noteworthy advantage of Studentization is that compared to standardized statistics, Studentized ratios take heteroscedasticity into account and are more robust against
heavy-tailed data. The theoretical and numerical studies in Delaigle, Hall
and Jin (2011) and Chang, Tang and Wu (2013, 2016) evidence the importance of using Studentized statistics in high dimensional data analysis. As
noted in Delaigle, Hall and Jin (2011), a careful study of the moderate
deviations in the Studentized ratios is indispensable to understanding the
common statistical procedures used in analyzing high dimensional data.
Further, it is now known that the theory of Cramér-type moderate deviations for Studentized statistics quantifies the accuracy of the estimated
p-values, which is crucial in the study of large-scale multiple tests for controlling the false discovery rate [Fan, Hall and Yao (2007), Liu and Shao
(2010)]. In particular, Cramér-type moderate deviation results can be used
to investigate the robustness and accuracy properties of p-values and critical
values in multiple testing procedures. However, thus far, most applications
have been confined to t-statistics [Fan, Hall and Yao (2007), Wang and Hall
(2009), Delaigle, Hall and Jin (2011), Cao and Kosorok (2011)]. It is conjectured in Fan, Hall and Yao (2007) that analogues of the theoretical properties
STUDENTIZED TWO-SAMPLE U -STATISTICS
3
of these statistical methodologies remain valid for other resampling methods based on Studentized statistics. Motivated by the above applications,
we are attempting to develop a unified theory on moderate deviations for
more general Studentized nonlinear statistics, in particular, for two-sample
U -statistics.
The asymptotic properties of the standardized U -statistics are extensively
studied in the literature, whereas significant developments are achieved in
the past decade for one-sample Studentized U -statistics. We refer to Wang,
Jing and Zhao (2000) and the references therein for Berry–Esseen-type
bounds and Edgeworth expansions. The results for moderate deviations can
be found in Vandemaele and Veraverbeke (1985), Lai, Shao and Wang (2011)
and Shao and Zhou (2016). The results in Shao and Zhou (2016) paved the
way for further applications of statistical methodologies using Studentized
U -statistics in high dimensional data analysis.
Two-sample U -statistics are also commonly used to compare the different (treatment) effects of two groups, such as an experimental group and
a control group, in scientifically controlled experiments. However, due to
the structural complexities, the theoretical properties of the two-sample U statistics have not been well studied. In this paper, we establish a Cramértype moderate deviation theorem in a general framework for Studentized
two-sample U -statistics, especially the two-sample t-statistic and the Studentized Mann–Whitney test. In particular, a refined moderate deviation
theorem with second-order accuracy is established for the two-sample tstatistic.
The paper is organized as follows. In Section 2, we present the main
results on Cramér-type moderate deviations for Studentized two-sample U statistics as well as a refined result for the two-sample t-statistic. In Section 3, we investigate statistical applications of our theoretical results to the
problem of simultaneously testing many different hypotheses, based particularly on the two-sample t-statistics and Studentized Mann–Whitney tests.
Section 4 shows numerical studies. A discussion is given in Section 5. All the
proofs are relegated to the supplementary material [Chang, Shao and Zhou
(2016)].
2. Moderate deviations for Studentized U -statistics. We use the following notation throughout this paper. For two sequences of real numbers an
and bn , we write an ≍ bn if there exist two positive constants c1 , c2 such that
c1 ≤ an /bn ≤ c2 for all n ≥ 1, we write an = O(bn ) if there is a constant C
such that |an | ≤ C|bn | holds for all sufficiently large n, and we write an ∼ bn
and an = o(bn ), respectively, if limn→∞ an /bn = 1 and limn→∞ an /bn = 0.
Moreover, for two real numbers a and b, we write for ease of presentation
that a ∨ b = max(a, b) and a ∧ b = min(a, b).
4
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
2.1. A review of Studentized one-sample U -statistics. We start with a
brief review of Cramér-type moderate deviation for Studentized one-sample
U -statistics. For an integer s ≥ 2 and for n > 2s, let X1 , . . . , Xn be independent and identically distributed (i.i.d.) random variables taking values in a
metric space (X, G), and let h : Xd 7→ R be a symmetric Borel measurable
function. Hoeffding’s U -statistic with a kernel h of degree s is defined as
Un =
X
1
n
s
h(Xi1 , . . . , Xis ),
1≤i1 <···<is ≤n
which is an unbiased estimate of θ = E{h(X1 , . . . , Xs )}. In particular, we
focus on the case where X is the Euclidean space Rr for some integer r ≥ 1.
When r ≥ 2, write Xi = (Xi1 , . . . , Xir )T for i = 1, . . . , n.
Let
h1 (x) = E{h(X1 , . . . , Xs )|X1 = x}
for any x = (x1 , . . . , xr )T ∈ Rr
and
σ 2 = var{h1 (X1 )},
vh2 = var{h(X1 , X2 , . . . , Xs )}.
Assume that 0 < σ 2 < ∞, then the standardized nondegenerate U -statistic
is given by
Zn =
n1/2
(Un − θ).
sσ
Because σ is usually unknown, we are interested in the following Studentized U -statistic:
1/2
bn = n (Un − θ),
U
sb
σ
(2.1)
where σ
b2 denotes the leave-one-out jackknife estimator of σ 2 given by
n
(n − 1) X
(qi − Un )2
σ
b =
(n − s)2
2
with
i=1
qi =
1
n−1
s−1
X
h(Xi , Xℓ1 , . . . , Xℓs−1 ).
1≤ℓ1 <···<ℓs−1 ≤n
ℓj 6=i for each j=1,...,s−1
Shao and Zhou (2016) established a general Cramér-type moderate deviation theorem for Studentized nonlinear statistics, in particular for Studentized U -statistics.
5
STUDENTIZED TWO-SAMPLE U -STATISTICS
Theorem 2.1. Assume that vp := [E{|h1 (X1 ) − θ|p }]1/p < ∞ for some
2 < p ≤ 3. Suppose that there are constants c0 ≥ 1 and κ ≥ 0 such that for
all x1 , . . . , xs ∈ R,
#
"
s
X
2
2
2
(2.2)
{h1 (xi ) − θ} .
{h(x1 , . . . , xs ) − θ} ≤ c0 κσ +
i=1
Then there exist constants C, c > 0 depending only on d such that
bn ≥ x)
P(U
= 1 + O(1){(vp /σ)p (1 + x)p n1−p/2 + (as1/2 + vh /σ)(1 + x)3 n−1/2 }
1 − Φ(x)
holds uniformly for 0 ≤ x ≤ c min{(σ/vp )n1/2−1/p , (n/as )1/6 }, where |O(1)| ≤
C and as = max(c0 κ, c0 + s). In particular, we have
bn ≥ x)
P(U
→1
1 − Φ(x)
holds uniformly in x ∈ [0, o(n1/2−1/p )).
Condition (2.2) is satisfied for a large class of U -statistics. Below are some
examples.
Statistic
t-statistic
Sample variance
Gini’s mean difference
One-sample Wilcoxon’s statistic
Kendall’s τ
Kernel function
h(x1 , x2 ) = 0.5(x1 + x2 )
h(x1 , x2 ) = 0.5(x1 − x2 )2
h(x1 , x2 ) = |x1 − x2 |
h(x1 , x2 ) = I{x1 + x2 ≤ 0}
h(x1 , x2 ) = 2I{(x22 − x21 )(x12 − x11 ) > 0}
c0
2
10
8
1
1
κ
0
(θ/σ)2
(θ/σ)2
σ −2
σ −2
2.2. Studentized two-sample U -statistics. Let X = {X1 , . . . , Xn1 } and Y =
{Y1 , . . . , Yn2 } be two independent random samples, where X is drawn from
a probability distribution P and Y is drawn from another probability distribution Q. With s1 and s2 being two positive integers, let
h(x1 , . . . , xs1 ; y1 , . . . , ys2 )
be a kernel function of order (s1 , s2 ) which is real and symmetric both in its
first s1 variates and in its last s2 variates. It is known that a nonsymmetric kernel can always be replaced with a symmetrized version by averaging
across all possible rearrangements of the indices.
Set θ := E{h(X1 , . . . , Xs1 ; Y1 , . . . , Ys2 )}, and let
Un̄ =
1
X
X
h(Xi1 , . . . , Xis1 ; Yj1 , . . . , Yjs2 ),
n1 n2
s1 s2 1≤i1 <···<is ≤n1 1≤j1 <···<js ≤n2
2
1
6
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
be the two-sample U -statistic, where n̄ = (n1 , n2 ). To lighten the notation,
we write Xi1 ,...,iℓ = (Xi1 , . . . , Xiℓ ), Yj1 ,...,jk = (Yj1 , . . . , Yjk ) such that
h(Xi1 ,...,iℓ ; Yj1 ,...,jk ) = h(Xi1 , . . . , Xiℓ ; Yj1 , . . . , Yjk ),
and define
h1 (x) = E{h(X1,...,s1 ; Y1,...,s2 )|X1 = x},
(2.3)
h2 (y) = E{h(X1,...,s1 ; Y1,...,s2 )|Y1 = y}.
Also let vh2 = var{h(X1,...,s1 ; Y1,...,s2 )}, σ12 = var{h1 (Xi )}, σ22 = var{h2 (Yj )}
and
σ 2 = σ12 + σ22 ,
(2.4)
2 2 −1
σn̄2 = s21 σ12 n−1
1 + s 2 σ2 n 2 .
For the standardized two-sample U -statistic of the form σn̄−1 (Un̄ − θ), a
uniform Berry–Esseen bound of order O{(n1 ∧ n2 )−1/2 } was obtained by
Helmers and Janssen (1982) and Borovskich (1983). Using a concentration
inequality approach, Chen and Shao (2007) proved a refined uniform bound
and also established an optimal nonuniform Berry–Esseen bound. For large
deviation asymptotics of two-sample U -statistics, we refer to Nikitin and
Ponikarov (2006) and the references therein.
Here, we are interested in the following Studentized two-sample U -statistic:
(2.5)
where
bn̄ = σ
U
bn̄−1 (Un̄ − θ)
n
n
i=1
i=1
1
1
1 X
1 X
σ
b12 =
qi
qi −
n1 − 1
n1
and
2 2 −1
with σ
bn̄2 = s21 σ
b12 n−1
b2 n2 ,
1 + s2 σ
!2
,
n
n
j=1
j=1
2
2
1 X
1 X
pj
pj −
σ
b22 =
n2 − 1
n2
1
X
X
1
X
X
!2
qi =
h(Xi,i2 ,...,is1 ; Yj1 ,...,js2 ),
n1 −1 n2
s1 −1 s2 1≤i2 <···<is1 ≤n1 1≤j1 <···<js2 ≤n2
iℓ 6=i,ℓ=2,...,s1
pj =
h(Xi1 ,...,is1 ; Yj,j2,...,js2 ).
n1 n2 −1
s1 s2 −1 1≤i1 <···<is1 ≤n1 1≤j2 <···<js2 ≤n2
jk 6=j,k=2,...,s2
Note that σ
b12 and σ
b22 are leave-one-out jackknife estimators of σ12 and σ22 ,
respectively.
7
STUDENTIZED TWO-SAMPLE U -STATISTICS
bn̄ . For p > 2, let
2.2.1. Moderate deviations for U
(2.6) v1,p = [E{|h1 (X1 ) − θ|p }]1/p
and v2,p = [E{|h2 (Y1 ) − θ|p }]1/p .
Moreover, put
s = s1 ∨ s2 ,
n̄ = (n1 , n2 ),
n = n1 ∧ n2
and
λn̄ = vh
n1 + n2
σ12 n2 + σ22 n1
1/2
with vh2 = var{h(X1,...,s1 ; Y1,...,s2 )}.
bn̄ given
The following result gives a Cramér-type moderate deviation for U
in (2.5) under mild assumptions. A self-contained proof can be found in the
supplementary material [Chang, Shao and Zhou (2016)].
Assume that there are constants c0 ≥ 1 and κ ≥ 0 such
Theorem 2.2.
that
2
(2.7) {h(x; y) − θ} ≤ c0
"
s2
s1
X
X
2
2
{h2 (yj ) − θ}2
{h1 (xi ) − θ} +
κσ +
j=1
i=1
#
for all x = (x1 , . . . , xs1 ) and y = (y1 , . . . , ys2 ), where σ 2 is given in (2.4).
Assume that v1,p and v2,p are finite for some 2 < p ≤ 3. Then there exist
constants C, c > 0 independent of n1 and n2 such that
bn̄ ≥ x)
P(U
1 − Φ(x)
(2.8)
(
= 1 + O(1)
holds uniformly for
p
2
X
vℓ,p
(1 + x)p
p/2−1
ℓ=1
σℓp nℓ
p/2−1
0 ≤ x ≤ c min[(σ1 /v1,p )n1
1/2
+ (ad
3
+ λn̄ )(1 + x)
p/2−1
, (σ2 /v2,p )n2
n1 + n2
n1 n2
1/2 )
, a−1/6
{n1 n2 /(n1 + n2 )}1/6 ],
s
where |O(1)| ≤ C and as = max(c0 κ, c0 + s). In particular, as n → ∞,
(2.9)
bn̄ ≥ x)
P(U
→1
1 − Φ(x)
holds uniformly in x ∈ [0, o(n1/2−1/p )).
Theorem 2.2 exhibits the dependence between the range of uniform convergence of the relative error in the central limit theorem and the optimal moment conditions. In particular, if p = 3, the region becomes 0 ≤ x ≤
8
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
O(n1/6 ). See Theorem 2.3 in Jing, Shao and Wang (2003) for similar results
on self-normalized sums. Under higher order moment conditions, it is not
clear if our technique can be adapted to provide a better approximation for
bn̄ ≥ x) for x lying between n1/6 and n1/2 in order.
the tail probability P(U
It is also worth noticing that many commonly used kernels in nonparametric statistics turn out to be linear combinations of the indicator functions
and, therefore, satisfy condition (2.7) immediately.
2.2.2. Two-sample t-statistic. As a prototypical example of two-sample
U -statistics, the two-sample t-statistic is of significant interest due to its
wide applicability. The advantage of using t-tests, either one-sample or twosample, is their high degree of robustness against heavy-tailed data in which
the sampling distribution has only a finite third or fourth moment. The robustness of the t-statistic is useful in high dimensional data analysis under
the sparsity assumption on the signal of interest. When dealing with two
experimental groups, which are typically independent, in scientifically controlled experiments, the two-sample t-statistic is one of the most commonly
used statistics for hypothesis testing and constructing confidence intervals
for the difference between the means of the two groups.
Let X = {X1 , . . . , Xn1 } be a random sample from a one-dimensional population with mean µ1 and variance σ12 , and let Y = {Y1 , . . . , Yn2 } be a random
sample from another one-dimensional population with mean µ2 and variance
σ22 independent of X . The two-sample t-statistic is defined as
Tbn̄ = q
X̄ − Ȳ
,
2 n−1
σ
b12 n−1
+
σ
b
1
2 2
Pn1
−1 Pn2
where n̄ = (n1 , n2 ), X̄ = n−1
1
j=1 Yj and
i=1 Xi , Ȳ = n2
n
σ
b12
1
1 X
(Xi − X̄)2 ,
=
n1 − 1
i=1
n
σ
b22
2
1 X
=
(Yj − Ȳ )2 .
n2 − 1
j=1
The following result is a direct consequence of Theorem 2.2.
Theorem 2.3. Assume that µ1 = µ2 , and E(|X1 |p ) < ∞, E(|Y1 |p ) < ∞
for some 2 < p ≤ 3. Then there exist absolute constants C, c > 0 such that
2
X
P(Tbn̄ ≥ x)
1−p/2
= 1 + O(1)(1 + x)p
(vℓ,p /σℓ )p nℓ
1 − Φ(x)
ℓ=1
1/2−1/p
holds uniformly for 0 ≤ x ≤ c minℓ=1,2 {(σℓ /vℓ,p )nℓ
}, where |O(1)| ≤ C
p
1/p
p
1/p
and v1,p = {E(|X1 − µ1 | )} , v2,p = {E(|Y1 − µ2 | )} .
9
STUDENTIZED TWO-SAMPLE U -STATISTICS
Motivated by a series of recent studies on the effectiveness and accuracy
of multiple-hypothesis testing using t-tests, we investigate whether a higher
order expansion of the relative error, as in Theorem 1.2 of Wang (2005) for
self-normalized sums, holds for the two-sample t-statistic, so that one can
use bootstrap calibration to correct skewness [Fan, Hall and Yao (2007),
Delaigle, Hall and Jin (2011)] or study power properties against sparse alternatives [Wang and Hall (2009)]. The following theorem gives a refined
Cramér-type moderate deviation result for Tbn̄ , whose proof is placed in the
supplementary material [Chang, Shao and Zhou (2016)].
Theorem 2.4. Assume that µ1 = µ2 . Let γ1 = E{(X1 − µ1 )3 } and γ2 =
E{(Y1 − µ2 )3 } be the third central moment of X1 and Y1 , respectively. Moreover, assume that E(|X1 |p ) < ∞, E(|Y1 |p ) < ∞ for some 3 < p ≤ 4. Then
−2
P(Tbn̄ ≥ x)
γ1 n−2
3
1 − γ2 n2
x
= exp −
2 −1 3/2
1 − Φ(x)
3(σ12 n−1
1 + σ2 n 2 )
(2.10)
"
#
p
2 3
X
vℓ,3 (1 + x) vℓ,p (1 + x)p
+ p p/2−1
× 1 + O(1)
1/2
σℓ3 nℓ
σℓ n ℓ
ℓ=1
holds uniformly for
(2.11)
1/2
1/2−1/p
0 ≤ x ≤ c min min{(σℓ /vℓ,3 )3 nℓ , (σℓ /vℓ,p )nℓ
ℓ=1,2
},
where |O(1)| ≤ C and for every q ≥ 1, v1,q = {E(|X1 − µ1 |q )}1/q , v2,q =
{E(|Y1 − µ2 |q )}1/q .
A refined Cramér-type moderate deviation theorem for the one-sample tstatistic was established in Wang (2011), which to our knowledge, is the best
result for the t-statistic known up to date, or equivalently, self-normalized
sums.
2.2.3. More examples of two-sample U -statistics. Beyond the two-sample
t-statistic, we enumerate three more well-known two-sample U -statistics
and refer to Nikitin and Ponikarov (2006) for more examples. Let X =
{X1 , . . . , Xn1 } and Y = {Y1 , . . . , Yn2 } be two independent random samples
from population distributions P and Q, respectively.
Example 2.1 (The Mann–Whitney test statistic).
order (s1 , s2 ) = (1, 1), defined as
h(x; y) = I{x ≤ y} − 1/2
The kernel h is of
with θ = P(X1 ≤ Y1 ) − 1/2,
10
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
and in view of (2.3),
h1 (x) = 1/2 − G(x),
h2 (y) = F (y) − 1/2.
In particular, if F ≡ G, we have σ12 = σ22 = 1/12.
Example 2.2 (The Lehmann statistic).
(2, 2), defined as
The kernel h is of order (s1 , s2 ) =
h(x1 , x2 ; y1 , y2 ) = I{|x1 − x2 | ≤ |y1 − y2 |} − 1/2
with θ = P(|X1 − X2 | ≤ |Y1 − Y2 |) − 1/2. Then under H0 : θ = 0, E{h(X1 , X2 ;
Y1 , Y2 )} = 0, and
h1 (x) = G(x){1 − G(x)} − 1/6,
h2 (y) = F (y){F (y) − 1} + 1/6.
In particular, if F ≡ G, then σ12 = σ22 = 1/180.
Example 2.3 (The Kochar statistic). The Kochar statistic was constructed by Kochar (1979) to test if the two hazard failure rates are different.
Denote by F the class of all absolutely continuous cumulative distribution
functions (CDF) F (·) satisfying F (0) = 0. For two arbitrary CDF’s F, G ∈ F ,
and let f = F ′ , g = G′ be their densities. Thus, the hazard failure rates are
defined by
rF (t) =
f (t)
,
1 − F (t)
rG (t) =
g(t)
,
1 − G(t)
as long as both 1 − F (t) and 1 − G(t) are positive. Kochar (1979) considered the problem of testing the null hypothesis H0 : rF (t) = rG (t) against
the alternative H1 : rF (t) ≤ rG (t), t ≥ 0 with strict inequality over a set of
nonzero measures. Observe that H1 holds if and only if δ(s, t) = F̄ (s)Ḡ(t) −
F̄ (t)Ḡ(s) ≥ 0 for s ≥ t ≥ 0 with strict inequality over a set of nonzero measures, where F̄ (·) := 1 − F (·) for any F ∈ F .
Recall that X1 , . . . , Xn1 and Y1 , . . . , Yn2 are two independent samples
drawn respectively from F and G. Following Nikitin and Ponikarov (2006),
we see that
η(F ; G) = E{δ(X ∨ Y, X ∧ Y )}
= P(Y1 ≤ Y2 ≤ X1 ≤ X2 ) + P(X1 ≤ Y2 ≤ Y2 ≤ X2 )
− P(X1 ≤ X2 ≤ Y1 ≤ Y2 ) − P(Y1 ≤ X1 ≤ X2 ≤ Y2 ).
Under H0 , η(F ; G) = 0 while under H1 , η(F ; G) > 0. The U -statistic with
the kernel of order (s1 , s2 ) = (2, 2) is given by
h(x1 , x2 ; y1 , y2 ) = I{yyxx or xyyx} − I{xxyy or yxxy}.
STUDENTIZED TWO-SAMPLE U -STATISTICS
11
Here, the term “yyxx” refers to y1 ≤ y2 ≤ x1 ≤ x2 and similar treatments
apply to xyyx, xxyy and yxxy. Under H0 : rF (t) = rG (t), we have
h1 (x) = −4G3 (x)/3 + 4G2 (x) − 2G(x),
h2 (y) = 4F 3 (y)/3 − 4F 2 (y) + 2F (y).
In particular, if F ≡ G, then σ12 = σ22 = 8/105.
3. Multiple testing via Studentized two-sample tests. Multiple-hypothesis
testing occurs in a wide range of applications including DNA microarray
experiments, functional magnetic resonance imaging analysis (fMRI) and
astronomical surveys. We refer to Dudoit and van der Laan (2008) for a systematic study of the existing multiple testing procedures. In this section, we
consider multiple-hypothesis testing based on Studentized two-sample tests
and show how the theoretical results in the previous section can be applied
to these problems.
3.1. Two-sample t-test. A typical application of multiple-hypothesis testing in high dimensions is the analysis of gene expression microarray data.
To see whether each gene in isolation behaves differently in a control group
versus an experimental group, we can apply the two-sample t-test. Assume
that the statistical model is given by
Xi,k = µ1k + εi,k ,
i = 1, . . . , n1 ,
(3.1)
Yj,k = µ2k + ωj,k ,
j = 1, . . . , n2 ,
for k = 1, . . . , m, where index k denotes the kth gene, i and j indicate the
ith and jth array, and the constants µ1k and µ2k , respectively, represent the
mean effects for the kth gene from the first and the second groups. For each k,
ε1,k , . . . , εn1 ,k (resp., ω1,k , . . . , ωn2 ,k ) are independent random variables with
2 > 0 (resp., σ 2 > 0). For the kth marginal test,
mean zero and variance σ1k
2k
2 and σ 2 are unequal, the two-sample twhen the population variances σ1k
2k
statistic is most commonly used to carry out hypothesis testing for the null
H0k : µ1k = µ2k against the alternative H1k : µ1k 6= µ2k .
Since the seminal work of Benjamini and Hochberg (1995), the Benjamini
and Hochberg (B–H) procedure has become a popular technique in microarray data analysis for gene selection, which along with many other procedures
depend on p-values that often need to be estimated. To control certain simultaneous errors, it has been shown that using approximated p-values is
asymptotically equivalent to using the true p-values for controlling the kfamilywise error rate (k-FWER) and false discovery rate (FDR). See, for
example, Kosorok and Ma (2007), Fan, Hall and Yao (2007) and Liu and
Shao (2010) for one-sample tests. Cao and Kosorok (2011) proposed an alternative method to control k-FWER and FDR in both large-scale one-
12
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
and two-sample t-tests. A common thread among the aforementioned literature is that theoretically for the methods to work in controlling FDR at a
given level, the number of features m and the sample size n should satisfy
log m = o(n1/3 ).
Recently, Liu and Shao (2014) proposed a regularized bootstrap correction
method for multiple one-sample t-tests so that the constraint on m may
be relaxed to log m = o(n1/2 ) under less stringent moment conditions as
assumed in Fan, Hall and Yao (2007) and Delaigle, Hall and Jin (2011). Using
Theorem 2.4, we show that the constraint on m in large scale two-sample
t-tests can be relaxed to log m = o(n1/2 ) as well. This provides theoretical
justification of the effectiveness of the bootstrap method which is frequently
used for skewness correction.
To illustrate the main idea, here we restrict our attention to the special
case in which the observations are independent. Indeed, when test statistics are correlated, false discovery control becomes very challenging under
arbitrary dependence. Various dependence structures have been considered
in the literature. See, for example, Benjamini and Yekutieli (2001), Storey,
Taylor and Siegmund (2004), Ferreira and Zwinderman (2006), Leek and
Storey (2008), Friguet, Kloareg and Causeur (2009) and Fan, Han and Gu
(2012), among others. For completeness, we generalize the results to the
dependent case in Section 3.1.3.
3.1.1. Normal calibration and phase transition. Consider the large-scale
significance testing problem:
H0k : µ1k = µ2k
versus
H1k : µ1k 6= µ2k ,
1 ≤ k ≤ m.
Let V and R denote, respectively, the number of false rejections and the
number of total rejections. The well-known false discovery proportion (FDP)
is defined as the ratio FDP = V / max(1, R), and FDR is the expected FDP,
that is, E{V / max(1, R)}. Benjamini and Hochberg (1995) proposed a
distribution-free method for choosing a p-value threshold that controls the
FDR at a prespecified level where 0 < α < 1. For k = 1, . . . , m, let pk be
the marginal p-value of the kth test, and let p(1) ≤ · · · ≤ p(m) be the order
statistics of p1 , . . . , pm . For a predetermined control level α ∈ (0, 1), the B–H
procedure rejects hypotheses for which pk ≤ p(k̂) , where
αk
k̂ = max 0 ≤ k ≤ m : p(k) ≤
(3.2)
m
with p(0) = 0.
In microarray analysis, two-sample t-tests are often used to identify differentially expressed genes between two groups. Let
Tk = q
X̄k − Ȳk
2 n−1 + σ
2 n−1
σ
b1k
b2k
1
2
,
k = 1, . . . , m,
STUDENTIZED TWO-SAMPLE U -STATISTICS
where X̄k = n−1
1
Pn 1
i=1 Xi,k ,
Ȳk = n−1
2
n
2
=
σ
b1k
1
1 X
(Xi,k − X̄k )2 ,
n1 − 1
i=1
Pn 2
j=1 Yj,k
2
=
σ
b2k
13
and
n
2
1 X
(Yj,k − Ȳk )2 .
n2 − 1
j=1
1
2
Here and below, {Xi,1 , . . . , Xi,m }ni=1
and {Yj,1 , . . . , Yj,m}nj=1
are independent
random samples from {X1 , . . . , Xm } and {Y1 , . . . , Ym }, respectively, generated according to model (3.1), which are usually non-Gaussian in practice.
Moreover, assume that the sample sizes of the two samples are of the same
order, that is, n1 ≍ n2 .
Before stating the main results, we first introduce a number of notation.
Set H0 = {1 ≤ k ≤ m : µ1k = µ2k }, let m0 = #H0 denote the number of true
null hypotheses and m1 = m − m0 . Both m = m(n1 , n2 ) and m0 = m0 (n1 , n2 )
are allowed to grow as n = n1 ∧ n2 increases. We assume that
m0
= π0 ∈ (0, 1].
lim
n→∞ m
In line with the notation used in Section 2, set
2
= var(Xk ),
σ1k
2
= var(Yk ),
σ2k
γ1k = E{(Xk − µ1k )3 },
γ2k = E{(Yk − µ2k )3 }
2 = σ 2 n−1 + σ 2 n−1 . Throughout this subsection, we focus on the
and σn̄,k
2k 2
1k 1
normal calibration and let pbk = 2 − 2Φ(|Tk |), where Φ(·) is the standard
normal distribution function. Indeed, the exact null distribution of Tk and
thus the true p-values are unknown without the normality assumption.
Theorem 3.1. Assume that {X1 , . . . , Xm , Y1 , . . . , Ym } are independent
nondegenerate random variables; n1 ≍ n2 , m = m(n1 , n2 ) → ∞ and log m =
o(n1/2 ) as n = n1 ∧ n2 → ∞. For independent random samples {Xi,1 , . . . ,
1
2
Xi,m }ni=1
and {Yj,1 , . . . , Yj,m}nj=1
, suppose that
max max{E(ξk4 ), E(ηk4 )} ≤ C < ∞
(3.3) min min(σ1k , σ2k ) ≥ c > 0,
1≤k≤m
1≤k≤m
−1
−1
for some constants C and c, where ξk = σ1k
(Xk − µ1k ) and ηk = σ2k
(Yk −
µ2k ). Moreover, assume that
(3.4)
#{1 ≤ k ≤ m : |µ1k − µ2k | ≥ 4(log m)1/2 σn̄,k } → ∞
as n → ∞, and let
(3.5)
c0 = lim inf
n,m→∞
n1/2 X −3
−2
−2
σn̄,k |γ1k n1 − γ2k n2 | .
m0
k∈H0
14
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
(i) Suppose that log m = o(n1/3 ). Then as n → ∞, FDPΦ →P απ0 and
FDRΦ → απ0 .
(ii) Suppose that c0 > 0, log m ≥ c1 n1/3 for some c1 > 0 and that log m1 =
1/3
o(n ). Then there exists some constant β ∈ (α, 1] such that
lim P(FDPΦ ≥ β) = 1
n→∞
and
lim inf FDRΦ ≥ β.
n→∞
(log m)/n1/3
(iii) Suppose that c0 > 0,
→ ∞ and log m1 = o(n1/3 ). Then
as n → ∞, FDPΦ →P 1 and FDRΦ → 1.
Here, FDRΦ and FDPΦ denote, respectively, the FDR and the FDP of the
B–H procedure with pk replaced by pbk in (3.2).
Together, conclusions (i) and (ii) of Theorem 3.1 indicate that the number
of simultaneous tests can be as large as exp{o(n1/3 )} before the normal calibration becomes inaccurate. In particular, when n1 = n2 = n, the skewness
parameter c0 given in (3.5) reduces to
1 X |γ1k − γ2k |
.
c0 = lim inf
2 + σ 2 )3/2
m→∞
m0
(σ1k
2k
k∈H
0
As noted in Liu and Shao (2014), the limiting behavior of the FDRΦ varies
in different regimes and exhibits interesting phase transition phenomena as
the dimension m grows as a function of (n1 , n2 ). The average of skewness c0
plays a crucial role. It is also worth noting that conclusions (ii) and (iii) hold
under the scenario π0 = 1, that is, m1 = o(m). This corresponds to the sparse
settings in applications such as gene detections. Under finite 4th moments of
Xk and Yk , the robustness of two-sample t-tests and the accuracy of normal
calibration in the FDR/FDP control have been investigated in Cao and
Kosorok (2011) when m1 /m → π1 ∈ (0, 1). This corresponds to the relatively
dense setting, and the sparse case that we considered above is not covered.
3.1.2. Bootstrap calibration and regularized bootstrap correction. In this
subsection, we first use the conventional bootstrap calibration to improve
the accuracy of FDR control based on the fact that the bootstrap approximation removes the skewness term that determines first-order inaccuracies
of the standard normal approximation. However, the validity of bootstrap
approximation requires the underlying distribution to be very light tailed,
which does not seem realistic in real data applications. As pointed in the
literature of gene study, many gene data are commonly recognized to have
heavy tails which violates the assumption on underlying distribution used to
make conventional bootstrap approximation work. Recently, Liu and Shao
(2014) proposed a regularized bootstrap method that is shown to be more
robust against the heavy tailedness of the underlying distribution and the
dimension m is allowed to be as large as exp{o(n1/2 )}.
15
STUDENTIZED TWO-SAMPLE U -STATISTICS
†
†
†
†
Let Xk,b
= {X1,k,b
, . . . , Xn† 1 ,k,b }, Yk,b
= {Y1,k,b
, . . . , Yn†2 ,k,b }, b = 1, . . . , B,
denote bootstrap samples drawn independently and uniformly, with replacement, from Xk = {X1,k , . . . , Xn1 ,k } and Yk = {Y1,k , . . . , Yn2 ,k }, respectively.
†
†
Let Tk,b
be the two-sample t-statistic constructed from {X1,k,b
− X̄k , . . . ,
†
Xn† 1 ,k,b − X̄k } and {Y1,k,b
− Ȳk , . . . , Yn†2 ,k,b − Ȳk }. Following Liu and Shao
(2014), we use the following empirical distribution:
m
†
Fm,B
(t) =
B
1 XX
†
I{|Tk,b
| ≥ t}
mB
k=1 b=1
to approximate the null distribution, and thus the estimated p-values are
†
given by pbk,B = Fm,B
(|Tk |). Respectively, FDPB and FDRB denote the FDP
and the FDR of the B–H procedure with pk replaced by pbk,B in (3.2).
The following result shows that the bootstrap calibration is accurate provided log m increases at a strictly slower rate than (n1 ∧ n2 )1/2 , and the
underlying distribution has sub-Gaussian tails.
Theorem 3.2.
Assume the conditions in Theorem 3.1 hold and that
2
2
max max{E(et0 ξk ), E(et0 ηk )} ≤ C < ∞
1≤k≤m
for some constants t0 , C > 0.
(i) Suppose that log m = o(n1/3 ). Then as n → ∞, FDPB →P απ0 and
FDRB → απ0 .
(ii) Suppose that log m = o(n1/2 ) and m1 ≤ mρ for some ρ ∈ (0, 1). Then
as n → ∞, FDPB →P α and FDRB → α.
The sub-Gaussian condition in Theorem 3.2 is quite stringent in practice,
whereas it can hardly be weakened in general when the bootstrap method
is applied. In the context of family-wise error rate control, Fan, Hall and
Yao (2007) proved that the bootstrap calibration is accurate if the observed
data are bounded and log m = o(n1/2 ). The regularized bootstrap method,
however, adopts the very similar idea of the trimmed estimators and is a twostep procedure that combines the truncation technique and the bootstrap
method.
First, define the trimmed samples
bi,k = Xi,k I{|Xi,k | ≤ λ1k },
X
Ybj,k = Yi,k I{|Yj,k | ≤ λ2k }
for i = 1, . . . , n1 , j = 1, . . . , n2 , where λ1k and λ2k are regularized parameters
†
b† , . . . , X
b†
b†
b†
b†
to be specified. Let Xbk,b
= {X
1,k,b
n1 ,k,b } and Yk,b = {Y1,k,b , . . . , Yn2 ,k,b },
b = 1, . . . , B, be the corresponding bootstrap samples drawn by sampling
randomly, with replacement, from
b1,k , . . . , X
bn ,k } and Ybk = {Yb1,k , . . . , Ybn ,k },
Xbk = {X
1
2
16
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
†
respectively. Next, let Tbk,b
be the two-sample t-test statistic constructed from
P
†
n
−1
−1 Pn1 b
−1
1
b†
b
b†
b
{X
i=1 Xi,k } and {Y1,k,b − n2 ×
i=1 Xi,k , . . . , Xn1 ,k,b − n1
1,k,b − n1
Pn 2 b
−1 Pn2 b
b†
j=1 Yj,k }. As in the previous procedure, define
j=1 Yj,k , . . . , Yn2 ,k,b − n2
the estimated p-values by
m
†
pbk,RB = Fbm,RB
(|Tk |)
B
1 XX
†
†
with Fbm,RB
(t) =
I{|Tbk,b
| ≥ t}.
mB
k=1 b=1
Let FDPRB and FDRRB denote the FDP and the FDR, respectively, of the
B–H procedure with pk replaced by pbk,RB in (3.2).
Theorem 3.3.
(3.6)
Assume the conditions in Theorem 3.1 hold and that
max max{E(|Xk |6 ), E(|Yk |6 )} ≤ C < ∞.
1≤k≤m
The regularized parameters (λ1k , λ2k ) are such that
1/6
1/6
n1
n2
λ1k ≍
(3.7)
and λ2k ≍
.
log m
log m
(i) Suppose that log m = o(n1/3 ). Then as n → ∞, FDPRB →P απ0 and
FDRRB → απ0 .
(ii) Suppose that log m = o(n1/2 ) and m1 ≤ mρ for some ρ ∈ (0, 1). Then
as n → ∞, FDPRB →P α and FDRRB → α.
In view of Theorem 3.3, the regularized bootstrap approximation is valid
under mild moment conditions that are significantly weaker than those required for the bootstrap method to work theoretically. The numerical performance will be investigated in Section 4. To highlight the main idea, a
self-contained proof of Theorem 3.1 is given in the supplementary material
[Chang, Shao and Zhou (2016)]. The proofs of Theorems 3.2 and 3.3 are
based on straightforward extensions of Theorems 2.2 and 3.1 in Liu and
Shao (2014), and thus are omitted.
3.1.3. FDR control under dependence. In this section, we generalize the
results in previous sections to the dependence case. Write ̺ = n1 /n2 . For
2 + ̺σ 2 and define
every k, ℓ = 1, . . . , m, let σk2 = σ1k
2k
(3.8)
rkℓ = (σk σℓ )−1 {cov(Xk , Xℓ ) + ̺ cov(Yk , Yℓ )},
which characterizes the dependence between (Xk , Yk ) and (Xℓ , Yℓ ). Partic2 = σ 2 , we see that r = 1 {corr(X , X ) +
ularly, when n1 = n2 and σ1k
kℓ
k
ℓ
2k
2
corr(Yk , Yℓ )}. In this subsection, we impose the following conditions on the
dependence structure of X = (X1 , . . . , Xm )T and Y = (Y1 , . . . , Ym )T .
STUDENTIZED TWO-SAMPLE U -STATISTICS
17
(D1) There exist constants 0 < r < 1, 0 < ρ < (1 − r)/(1 + r) and b1 > 0
such that
max
1≤k6=ℓ≤m
|rkℓ | ≤ r
and
max sk (m) ≤ b1 mρ ,
1≤k≤m
where for k = 1, . . . , m,
sk (m) = {1 ≤ ℓ ≤ m : corr(Xk , Xℓ ) ≥ (log m)−2−γ
or corr(Yk , Yℓ ) ≥ (log m)−2−γ }
for some γ > 0.
(D2) There exist constants 0 < r < 1, 0 < ρ < (1 − r)/(1 + r) and b1 > 0
such that max1≤k6=ℓ≤m |rkℓ | ≤ r and for each Xk , the number of variables Xℓ
that are dependent of Xk is less than b1 mρ .
The assumption max1≤k6=ℓ≤m |rkℓ | ≤ r for some 0 < r < 1 imposes a constraint on the magnitudes of the correlations, which is natural in the sense
that the correlation matrix R = (rkℓ )1≤k,ℓ≤m is singular if max1≤k6=ℓ≤m |rkℓ | =
1. Under condition (D1), each (Xk , Yk ) is allowed to be “moderately” correlated with at most as many as O(mρ ) other vectors. Condition (D2) enforces
a local dependence structure on the data, saying that each vector is dependent with at most as many as O(mρ ) other random vectors and independent
of the remaining ones. The following theorem extends the results in previous sections to the dependence case. Its proof is placed in the supplementary
material [Chang, Shao and Zhou (2016)].
Theorem 3.4. Assume that either condition (D1) holds with log m =
O(n1/8 ) or condition (D2) holds with log m = o(n1/3 ).
(i) Suppose that (3.3) and (3.4) are satisfied. Then as n → ∞, FDPΦ →P
απ0 and FDRΦ → απ0 .
(ii) Suppose that (3.3), (3.6) and (3.7) are satisfied. Then as n → ∞,
FDPRB →P απ0 and FDRRB → απ0 .
In particular, assume that condition (D2) holds with log m = o(n1/2 ) and
m1 ≤ mc for some 0 < c < 1. Then as n → ∞, FDPRB →P απ0 and FDRRB →
απ0 .
3.2. Studentized Mann–Whitney test. Let X = {X1 , . . . , Xn1 } and Y =
{Y1 , . . . , Yn2 } be two independent random samples from distributions F and
G, respectively. Let θ = P(X ≤ Y ) − 1/2. Consider the null hypothesis H0 :
θ = 0 against the one-sided alternative H1 : θ > 0. This problem arises in
many applications including testing whether the physiological performance
of an active drug is better than that under the control treatment, and testing
the effects of a policy, such as unemployment insurance or a vocational
training program, on the level of unemployment.
18
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
The Mann–Whitney (M–W) test [Mann and Whitney (1947)], also known
as the two-sample Wilcoxon test [Wilcoxon (1945)], is prevalently used for
testing equality of means or medians, and serves as a nonparametric alternative to the two-sample t-test. The corresponding test statistic is given
by
n
(3.9)
n
2
1 X
1 X
Un̄ =
I{Xi ≤ Yj },
n1 n2
n̄ = (n1 , n2 ).
i=1 j=1
The M–W test is widely used in a wide range of fields including statistics, economics and biomedicine, due to its good efficiency and robustness
against parametric assumptions. Over one-third of the articles published
in Experimental Economics use the Mann–Whitney test and Okeh (2009)
reported that thirty percent of the articles in five biomedical journals published in 2004 used the Mann–Whitney test. For example, using the M–W U
test, Charness and Gneezy (2009) developed an experiment to test the conjecture that financial incentives help to foster good habits. They recorded
seven biometric measures (weight, body fat percentage, waist size, etc.) of
each participant before and after the experiment to assess the improvements
across treatments. Although the M–W test was originally introduced as a
rank statistic to test if the distributions of two related samples are identical, it has been prevalently used for testing equality of medians or means,
sometimes as an alternative to the two-sample t-test.
It was argued and formally examined recently in Chung and Romano
(2016) that the M–W test has generally been misused across disciplines. In
fact, the M–W test is only valid if the underlying distributions of the two
groups are identical. Nevertheless, when the purpose is to test the equality of
distributions, it is recommended to use a statistic, such as the Kolmogorov–
Smirnov or the Cramér–von Mises statistic, that captures the discrepancies
of the entire distributions rather than an individual parameter. More specifically, because the M–W test only recognizes deviation from θ = 0, it does
not have much power in detecting overall distributional discrepancies. Alternatively, the M–W test is frequently used to test the equality of medians.
However, Chung and Romano (2013) presented evidence that this is another
improper application of the M–W test and suggested to use the Studentized
median test.
Even when the M–W test is appropriately applied for testing H0 : θ = 0,
the asymptotic variance depends on the underlying distributions, unless
the two population distributions are identical. As Hall and Wilson (1991)
pointed out, the application of resampling to pivotal statistics has better
asymptotic properties in the sense that the rate of convergence of the actual
significance level to the nominal significance level is more rapid when the
19
STUDENTIZED TWO-SAMPLE U -STATISTICS
pivotal statistics are resampled. Therefore, it is natural to use the Studentized Mann–Whitney test, which is asymptotic pivotal.
Let
bn̄ = σ
U
bn̄−1 (Un̄ − 1/2)
(3.10)
bn̄2 = σ
b12 n−1
denote the Studentized test statistic for Un̄ as in (3.9), where σ
1 +
−1
2
σ
b2 n2 ,
!2
!2
n1
n2
n1
n2
1 X
1 X
1 X
1 X
2
2
qi −
pj −
qi ,
pj
σ
b2 =
σ
b1 =
n1 − 1
n1
n2 − 1
n2
i=1
j=1
i=1
j=1
−1 Pn1
−1 Pn2
with qi = n2
i=1 I{Xi ≤ Yj }.
j=1 I{Yj < Xi } and pj = n1
When dealing with samples from a large number of geographical regions (suburbs, states, health service areas, etc.), one may need to make
many statistical inferences simultaneously. Suppose we observe a family
of paired groups, that is, for k = 1, . . . , m, Xk = {X1,k , . . . , Xn1 ,k }, Yk =
{Y1,k , . . . , Yn2 ,k }, where the index k denotes the kth site. Assume that Xk is
drawn from Fk , and independently, Yk is drawn from Gk .
For each k = 1, . . . , m, we test the null hypothesis H0k : θk = P(X1,k ≤
Y1,k )−1/2 = 0 against the one-sided alternative H1k : θk > 0. If H0k is rejected,
we conclude that the treatment effect (of a drug or a policy) is acting within
the kth area. Define the test statistic
−1
bn̄,k = σ
U
bn̄,k
(Un̄,k − 1/2),
bn̄,k is constructed from the kth paired samples according to (3.10).
where U
Let
bn̄,k ≤ t|H k )
Fn̄,k (t) = P(U
0
and Φ(t) = P(Z ≤ t),
where Z is the standard normal random variable. Then the true p-values
bn̄,k ), and pbk = 1 − Φ(U
bn̄,k ) denote the estimated p-values
are pk = 1 − Fn̄,k (U
based on normal calibration.
To identify areas where the treatment effect is acting, we can use the
B–H method to control the FDR at α level by rejecting the null hypotheses
indexed by S = {1 ≤ k ≤ m : pbk ≤ pb(k̂) }, where k̂ = max{1 ≤ k ≤ m : pb(k) ≤
αk/m}, and {b
p(k) } denote the ordered values of {b
pk }. As before, let FDRΦ
be the FDR of the B–H method based on normal calibration.
Alternative to normal calibration, we can also consider bootstrap calibra†
†
†
†
tion. Recall that Xk,b
= {X1,k,b
, . . . , Xn† 1 ,k,b } and Yk,b
= {Y1,k,b
, . . . , Yn†2 ,k,b },
b = 1, . . . , B, are two bootstrap samples drawn independently and uniformly,
with replacement, from Xk = {X1,k , . . . , Xn1 ,k } and Yk = {Y1,k , . . . , Yn2 ,k }, reb†
spectively. For each k = 1, . . . , m, let U
n̄,k,b be the bootstrapped test statistic
20
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
†
†
constructed from Xk,b
and Yk,b
, that is,
#
"
n2
n1 X
X
1
†
−1
b
I{Xi,k ≤ Yj,k } ,
U
bn̄,k,b Un̄,k,b −
n̄,k,b = σ
n1 n2
i=1 j=1
where Un̄,k,b and σ
bn̄,k,b are the analogues of Un̄ given in (3.9) and σ
bn̄ specified
†
†
below (3.10) via replacing Xi and Yj by Xi,k,b
and Yj,k,b
, respectively. Using
the empirical distribution function
m
B
XX
b † | ≤ t},
b† (t) = 1
I{|U
G
n̄,k,b
m,B
mB
k=1 b=1
b† (U
b † ). For a predewe estimate the unknown p-values by pbk,B = 1 − G
m,B
n̄,k,b
termined α ∈ (0, 1), the null hypotheses indexed by SB = {1 ≤ k ≤ m : pbk,B ≤
pb(k̂B ),B } are rejected, where k̂B = max{0 ≤ k ≤ m : pbk,B ≤ αk/m}. Denote by
FDRB the FDR of the B–H method based on bootstrap calibration.
Applying the general moderate deviation result (2.9) to Studentized Mann–
bn̄,k leads to the following result. The proof is based on
Whitney statistics U
a straightforward adaptation of the arguments we used in the proof of Theorem 3.1, and hence is omitted.
Theorem 3.5. Assume that {X1 , . . . , Xm , Y1 , . . . , Ym } are independent
random variables with continuous distribution functions Xk ∼ Fk and Yj ∼
Gk . The triplet (n1 , n2 , m) is such that n1 ≍ n2 , m = m(n1 , n2 ) → ∞, log m =
o(n1/3 ) and m−1 #{k = 1, . . . , m : θk = 1/2} → π0 ∈ (0, 1] as n = n1 ∧ n2 →
1
2
∞. For independent samples {Xi,1 , . . . , Xi,m }ni=1
and {Yj,1 , . . . , Yj,m}nj=1
, suppose that min1≤k≤m min(σ1k , σ2k ) ≥ c > 0 for some constant c > 0 and as
n → ∞,
#{1 ≤ k ≤ m : |θk − 1/2| ≥ 4(log m)1/2 σn̄,k } → ∞,
2 = var{G (X )}, σ 2 = var{F (Y )} and σ 2 = σ 2 n−1 + σ 2 n−1 .
where σ1k
k k
k
k
2k 2
1k 1
n̄,k
2k
Then as n → ∞, FDPΦ , FDPB →P απ0 and FDRΦ , FDRB → απ0 .
Attractive properties of the bootstrap for multiple-hypothesis testing were
first noted by Hall (1990) in the case of the mean rather than its Studentized
counterpart. Now it has been rigorously proved that bootstrap methods
are particularly effective in relieving skewness in the extreme tails which
leads to second-order accuracy [Fan, Hall and Yao (2007), Delaigle, Hall
and Jin (2011)]. It is interesting and challenging to investigate whether these
advantages of the bootstrap can be inherited by multiple U -testing in either
the standardized or the Studentized case.
21
STUDENTIZED TWO-SAMPLE U -STATISTICS
4. Numerical study. In this section, we present numerical investigations
for various calibration methods described in Section 3 when they are applied to two-sample large-scale multiple testing problems. We refer to the
simulation for two-sample t-test and Studentized Mann–Whitney test as
Sim1 and Sim2 , respectively. Assume that we observe two groups of m1
2
dimensional gene expression data {Xi }ni=1
and {Yj }nj=1
, where X1 , . . . , Xn1
and Y1 , . . . , Yn2 are independent random samples drawn from the distributions of X and Y, respectively.
For Sim1 , let X and Y be such that
(4.1)
X = µ1 + {ε1 − E(ε1 )} and
Y = µ2 + {ε2 − E(ε2 )},
where ε1 = (ε1,1 , . . . , ε1,m )T and ε2 = (ε2,1 , . . . , ε2,m )T are two sets of i.i.d.
random variables. The i.i.d. components of noise vectors ε1 and ε2 follow two
types of distributions: (i) the exponential distribution Exp(λ) with density
function λ−1 e−x/λ ; (ii) Student t-distribution t(k) with k degrees of freedom.
The exponential distribution has nonzero skewness, while the t-distribution
is symmetric and heavy-tailed. For each type of error distribution, both cases
of homogeneity and heteroscedasticity were considered. Detailed settings for
the error distributions are specified in Table 1.
For Sim2 , we assume that X and Y satisfy
(4.2)
X = µ1 + ε1
and
Y = µ2 + ε2 ,
where ε1 = (ε1,1 , . . . , ε1,m )T and ε2 = (ε2,1 , . . . , ε2,m )T are two sets of i.i.d.
random variables. We consider several distributions for the error terms ε1,k
and ε2,k : standard normal distribution N (0, 1), t-distribution t(k), uniform
distribution U (a, b) and Beta distribution Beta(a, b). Table 2 reports four
settings of (ε1,k , ε2,k ) used in our simulation. In either setting, we know
P(ε1,k ≤ ε2,k ) = 1/2 holds. Hence, the power against the null hypothesis
H0k : P(Xk ≤ Yk ) = 1/2 will generate from the magnitude of the difference
between the kth components of µ1 and µ2 .
In both Sim1 and Sim2 , we set µ1 = 0, and assume that the first m1 =
1/2 and
2 −1
⌊1.6m1/2 ⌋ components of µ2 are equal to c{(σ12 n−1
1 + σ2 n2 ) log m}
2
2
the rest are zero. Here, σ1 and σ2 denote the variance of ε1,k and ε2,k , and
Table 1
Distribution settings in Sim1
Exponential distributions
Student t-distributions
Homogeneous case
Heteroscedastic case
ε1,k ∼ Exp(2)
ε2,k ∼ Exp(2)
ε1,k ∼ Exp(2)
ε2,k ∼ Exp(1)
ε1,k ∼ t(4)
ε2,k ∼ t(4)
ε1,k ∼ t(4)
ε2,k ∼ t(3)
22
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
Table 2
Distribution settings in Sim2
Identical distributions
Nonidentical distributions
Case 1
ε1,k ∼ N (0, 1)
ε2,k ∼ N (0, 1)
ε1,k ∼ N (0, 1)
ε2,k ∼ t(3)
Case 2
ε1,k ∼ U (0, 1)
ε2,k ∼ U (0, 1)
ε1,k ∼ U (0, 1)
ε2,k ∼ Beta(10, 10)
c is a parameter employed to characterize the location discrepancy between
the distributions of X and Y. The sample size (n1 , n2 ) was set to be (50, 30)
and (100, 60), and the discrepancy parameter c took values in {1, 1.5}. The
significance level α in the B–H procedure was specified as 0.05, 0.1, 0.2 and
0.3, and the dimension m was set to be 1000 and 2000. In Sim1 , we compared
three different methods to calculate the p-values in the B–H procedure: normal calibration given in Section 3.1.1, bootstrap calibration and regularized
bootstrap calibration proposed in Section 3.1.2. For regularized bootstrap
calibration, we used a cross-validation approach as in Section 3 of Liu and
Shao (2014) to choose regularized parameters λ1k and λ2k . In Sim2 , we
compared the performance of normal calibration and bootstrap calibration
proposed in Section 3.2. For each compared method, we evaluated its performance via two indices: the empirical FDR and the proportion among the
true alternative hypotheses was rejected. We call the latter correct rejection
proportion. If the empirical FDR is low, the proposed procedure has good
FDR control; if the correct rejection proportion is high, the proposed procedure has fairly good performance in identifying the true signals. For ease
of exposition, we only report the simulation results for (n1 , n2 ) = (50, 30)
and m = 1000 in Figures 1 and 2. The results for (n1 , n2 ) = (100, 60) and
m = 2000 are similar, which can be found in the supplementary material
[Chang, Shao and Zhou (2016)]. Each curve corresponds to the performance
of a certain method and the line types are specified in the caption below.
The horizontal ordinates of the four points on each curve depict the empirical FDR of the specified method when the pre-specified level α in the B–H
procedure was taken to be 0.05, 0.1, 0.2 and 0.3, respectively, and the vertical
ordinates indicate the corresponding empirical correct rejection proportion.
We say that a method has good FDR control if the horizontal ordinates
of the four points on its performance curve are less than the prescribed α
levels.
In general, as shown in Figures 1 and 2, the B–H procedure based on
(regularized) bootstrap calibration has better FDR control than that based
on normal calibration. In Sim1 where the errors are symmetric (e.g., ε1,k and
ε2,k follow the Student t-distributions), the panels in the first row of Figure 1 show that the B–H procedures using all the three calibration methods
STUDENTIZED TWO-SAMPLE U -STATISTICS
23
Fig. 1. Performance comparison of B–H procedures based on three calibration methods in Sim1 with (n1 , n2 ) = (50, 30) and m = 1000. The first and second rows show the
results when the components of noise vectors ε1 and ε2 follow t-distributions and exponential distributions, respectively; left and right panels show the results for homogeneous
and heteroscedastic cases, respectively; horizontal and vertical axes depict empirical false
discovery rate and empirical correct rejection proportion, respectively; and the prescribed
levels α = 0.05, 0.1, 0.2 and 0.3 are indicated by unbroken horizontal black lines. In each
panel, dashed lines and unbroken lines represent the results for the discrepancy parameter
c = 1 and 1.5, respectively, and different colors express different methods employed to calculate p-values in the B–H procedure, where blue line, green line and red line correspond
to the procedures based on normal, conventional and regularized bootstrap calibrations,
respectively.
are able to control or approximately control the FDR at given levels, while
the procedures based on bootstrap and regularized bootstrap calibrations
outperform that based on normal calibration in controlling the FDR. When
the errors are asymmetric in Sim1 , the performances of the three B–H procedures are different from those in the symmetric cases. From the second row
of Figure 1, we see that the B–H procedure based on normal calibration is
distorted in controlling the FDR while the procedure based on (regularized)
bootstrap calibration is still able to control the FDR at given levels. This
24
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
Fig. 2. Performance comparison of B–H procedures based on two different calibration
methods in Sim2 with (n1 , n2 ) = (50, 30) and m = 1000. The first and second rows show
the results when the components of noise vectors ε1 and ε2 follow the distributions specified in cases 1 and 2 of Table 2, respectively; left and right panels show the results for
the cases of identical distributions and nonidentical distributions, respectively; horizontal
and vertical axes depict empirical false discovery rate and empirical correct rejection proportion, respectively; and the prescribed levels α = 0.05, 0.1, 0.2 and 0.3 are indicated by
unbroken horizontal black lines. In each panel, dashed lines and unbroken lines represent
the results for the discrepancy parameter c = 1 and 1.5, respectively, and different colors
express different methods employed to calculate p-values in the B–H procedure, where blue
line and red line correspond to the procedures based on normal and bootstrap calibrations,
respectively.
phenomenon is further evidenced by Figure 2 for Sim2 . Comparing the B–H
procedures based on conventional and regularized bootstrap calibrations, we
find that the former approach is uniformly more conservative than the latter
in controlling the FDR. In other words, the B–H procedure based on regularized bootstrap can identify more true alternative hypotheses than that
using conventional bootstrap calibration. This phenomenon is also revealed
in the heteroscedastic case. As the discrepancy parameter c gets larger so
that the signal is stronger, the correct rejection proportion of the B–H pro-
STUDENTIZED TWO-SAMPLE U -STATISTICS
25
cedures based on all the three calibrations increase and the empirical FDR
is closer to the prescribed level.
5. Discussion. In this paper, we established Cramér-type moderate deviations for two-sample Studentized U -statistics of arbitrary order in a general framework where the kernel is not necessarily bounded. Two-sample
U -statistics, typified by the two-sample Mann–Whitney test statistic, have
been widely used in a broad range of scientific research. Many of these applications rely on a misunderstanding of what is being tested and the implicit
underlying assumptions, that were not explicitly considered until relatively
recently by Chung and Romano (2016). More importantly, they provided
evidence for the advantage of using the Studentized statistics both theoretically and empirically.
Unlike the conventional (one- and two-sample) U -statistics, the asymptotic behavior of their Studentized counterparts has barely been studied
in the literature, particularly in the two-sample case. Recently, Shao and
Zhou (2016) proved a Cramér-type moderate deviation theorem for general
Studentized nonlinear statistics, which leads to a sharp moderate deviation
result for Studentized one-sample U -statistics. However, extension from onesample to two-sample in the Studentized case is totally nonstraightforward,
and requires a more delicate analysis on the Studentizing quantities. Further,
for the two-sample t-statistic, we proved moderate deviation with secondorder accuracy under a finite 4th moment condition (see Theorem 2.4),
which is of independent interest. In contrast to the one-sample case, the
two-sample t-statistic cannot be reduced to a self-normalized sum of independent random variables, and thus the existing results on self-normalized
ratios [Jing, Shao and Wang (2003), Wang (2005, 2011)] cannot be directly
applied. Instead, we modify Theorem 2.1 in Shao and Zhou (2016) to obtain
a more precise expansion that can be used to derive a refined result for the
two-sample t-statistic.
Finally, we show that the obtained moderate deviation theorems provide
theoretical guarantees for the validity, including robustness and accuracy, of
normal, conventional bootstrap and regularized bootstrap calibration methods in multiple testing with FDR/FDP control. The dependence case is also
covered. These results represent a useful complement to those obtained by
Fan, Hall and Yao (2007), Delaigle, Hall and Jin (2011) and Liu and Shao
(2014) in the one-sample case.
Acknowledgements. The authors would like to thank Peter Hall and Aurore Delaigle for helpful discussions and encouragement. The authors sincerely thank the Editor, Associate Editor and three referees for their very
constructive suggestions and comments that led to substantial improvement
of the paper.
26
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
SUPPLEMENTARY MATERIAL
Supplement to “Cramér-type moderate deviations for Studentized twosample U -statistics with applications” (DOI: 10.1214/15-AOS1375SUPP;
.pdf). This supplemental material contains proofs for all the theoretical results in the main text, including Theorems 2.2, 2.4, 3.1 and 3.4, and additional numerical results.
REFERENCES
Benjamini, Y. and Hochberg, Y. (1995). Controlling the false discovery rate: A practical
and powerful approach to multiple testing. J. R. Stat. Soc. Ser. B. Stat. Methodol. 57
289–300. MR1325392
Benjamini, Y. and Yekutieli, D. (2001). The control of the false discovery rate in
multiple testing under dependency. Ann. Statist. 29 1165–1188. MR1869245
Borovskich, Y. V. (1983). Asymptotics of U -statistics and Von Mises’ functionals. Soviet
Math. Dokl. 27 303–308.
Cao, H. and Kosorok, M. R. (2011). Simultaneous critical values for t-tests in very high
dimensions. Bernoulli 17 347–394. MR2797995
Chang, J., Shao, Q. and Zhou, W.-X. (2016). Supplement to “Cramér-type
moderate deviations for Studentized two-sample U -statistics with applications.”
DOI:10.1214/15-AOS1375SUPP.
Chang, J., Tang, C. Y. and Wu, Y. (2013). Marginal empirical likelihood and sure
independence feature screening. Ann. Statist. 41 2123–2148. MR3127860
Chang, J., Tang, C. Y. and Wu, Y. (2016). Local independence feature screening
for nonparametric and semiparametric models by marginal empirical likelihood. Ann.
Statist. 44 515–539. MR3476608
Charness, G. and Gneezy, U. (2009). Incentives to exercise. Econometrica 77 909–931.
Chen, S. X. and Qin, Y.-L. (2010). A two-sample test for high-dimensional data with
applications to gene-set testing. Ann. Statist. 38 808–835. MR2604697
Chen, L. H. Y. and Shao, Q.-M. (2007). Normal approximation for nonlinear statistics
using a concentration inequality approach. Bernoulli 13 581–599. MR2331265
Chen, S. X., Zhang, L.-X. and Zhong, P.-S. (2010). Tests for high-dimensional covariance matrices. J. Amer. Statist. Assoc. 105 810–819. MR2724863
Chung, E. and Romano, J. P. (2013). Exact and asymptotically robust permutation
tests. Ann. Statist. 41 484–507. MR3099111
Chung, E. and Romano, J. (2016). Asymptotically valid and exact permutation tests
based on two-sample U -statistics. J. Statist. Plann. Inference. 168 97–105. MR3412224
Delaigle, A., Hall, P. and Jin, J. (2011). Robustness and accuracy of methods for high
dimensional data analysis based on Student’s t-statistic. J. R. Stat. Soc. Ser. B. Stat.
Methodol. 73 283–301. MR2815777
Dudoit, S. and van der Laan, M. J. (2008). Multiple Testing Procedures with Applications to Genomics. Springer, New York. MR2373771
Fan, J., Hall, P. and Yao, Q. (2007). To how many simultaneous hypothesis tests can
normal, Student’s t or bootstrap calibration be applied? J. Amer. Statist. Assoc. 102
1282–1288. MR2372536
Fan, J., Han, X. and Gu, W. (2012). Estimating false discovery proportion under arbitrary covariance dependence. J. Amer. Statist. Assoc. 107 1019–1035. MR3010887
Ferreira, J. A. and Zwinderman, A. H. (2006). On the Benjamini–Hochberg method.
Ann. Statist. 34 1827–1849. MR2283719
STUDENTIZED TWO-SAMPLE U -STATISTICS
27
Friguet, C., Kloareg, M. and Causeur, D. (2009). A factor model approach to multiple testing under dependence. J. Amer. Statist. Assoc. 104 1406–1415. MR2750571
Hall, P. (1990). On the relative performance of bootstrap and Edgeworth approximations
of a distribution function. J. Multivariate Anal. 35 108–129. MR1084945
Hall, P. and Wilson, S. R. (1991). Two guidelines for bootstrap hypothesis testing.
Biometrics 47 757–762. MR1132543
Helmers, R. and Janssen, P. (1982). On the Berry–Esseen theorem for multivariate
U -statistics. In Math. Cent. Rep. SW 90 1–22. Mathematisch Centrum, Amsterdam.
Hoeffding, W. (1948). A class of statistics with asymptotically normal distribution. Ann.
Math. Statistics 19 293–325. MR0026294
Jing, B.-Y., Shao, Q.-M. and Wang, Q. (2003). Self-normalized Cramér-type large
deviations for independent random variables. Ann. Probab. 31 2167–2215. MR2016616
Kochar, S. C. (1979). Distribution-free comparison of two probability distributions with
reference to their hazard rates. Biometrika 66 437–441. MR0556731
Koroljuk, V. S. and Borovskich, Y. V. (1994). Theory of U -Statistics. Mathematics
and Its Applications 273. Kluwer Academic, Dordrecht. MR1472486
Kosorok, M. R. and Ma, S. (2007). Marginal asymptotics for the “large p, small
n” paradigm: With applications to microarray data. Ann. Statist. 35 1456–1486.
MR2351093
Kowalski, J. and Tu, X. M. (2007). Modern Applied U -Statistics. Wiley, Hoboken, NJ.
MR2368050
Lai, T. L., Shao, Q.-M. and Wang, Q. (2011). Cramér type moderate deviations for
Studentized U -statistics. ESAIM Probab. Stat. 15 168–179. MR2870510
Leek, J. T. and Storey, J. D. (2008). A general framework for multiple testing dependence. Proc. Natl. Acad. Sci. USA 105 18718–18723.
Li, R., Zhong, W. and Zhu, L. (2012). Feature screening via distance correlation learning.
J. Amer. Statist. Assoc. 107 1129–1139. MR3010900
Li, G., Peng, H., Zhang, J. and Zhu, L. (2012). Robust rank correlation based screening.
Ann. Statist. 40 1846–1877. MR3015046
Liu, W. and Shao, Q.-M. (2010). Cramér-type moderate deviation for the maximum of
the periodogram with application to simultaneous tests in gene expression time series.
Ann. Statist. 38 1913–1935. MR2662363
Liu, W. and Shao, Q.-M. (2014). Phase transition and regularized bootstrap in largescale t-tests with false discovery rate control. Ann. Statist. 42 2003–2025. MR3262475
Mann, H. B. and Whitney, D. R. (1947). On a test of whether one of two random variables is stochastically larger than the other. Ann. Math. Statistics 18 50–60. MR0022058
Nikitin, Y. and Ponikarov, E. (2006). On large deviations of nondegenerate two-sample
U - and V -statistics with applications to Bahadur efficiency. Math. Methods Statist. 15
103–122. MR2225432
Okeh, U. M. (2009). Statistical analysis of the application of Wilcoxon and Mann–
Whitney U test in medical research studies. Biotechnol. Molec. Biol. Rev. 4 128–131.
Shao, Q.-M. and Zhou, W.-X. (2016). Cramér type moderate deviation theorems for
self-normalized processes. Bernoulli 22 2029–2079.
Storey, J. D., Taylor, J. E. and Siegmund, D. (2004). Strong control, conservative
point estimation and simultaneous conservative consistency of false discovery rates: A
unified approach. J. R. Stat. Soc. Ser. B. Stat. Methodol. 66 187–205. MR2035766
Vandemaele, M. and Veraverbeke, N. (1985). Cramér type large deviations for Studentized U -statistics. Metrika 32 165–179. MR0824452
Wang, Q. (2005). Limit theorems for self-normalized large deviation. Electron. J. Probab.
10 1260–1285 (electronic). MR2176384
28
J. CHANG, Q.-M. SHAO AND W.-X. ZHOU
Wang, Q. (2011). Refined self-normalized large deviations for independent random variables. J. Theoret. Probab. 24 307–329. MR2795041
Wang, Q. and Hall, P. (2009). Relative errors in central limit theorems for Student’s t
statistic, with applications. Statist. Sinica 19 343–354. MR2487894
Wang, Q., Jing, B.-Y. and Zhao, L. (2000). The Berry–Esseen bound for Studentized
statistics. Ann. Probab. 28 511–535. MR1756015
Wilcoxon, F. (1945). Individual comparisons by ranking methods. Biometrics 1 80–83.
Zhong, P.-S. and Chen, S. X. (2011). Tests for high-dimensional regression coefficients
with factorial designs. J. Amer. Statist. Assoc. 106 260–274. MR2816719
J. Chang
School of Statistics
Southwestern University of Finance
and Economics
Chengdu, Sichuan 611130
China
and
School of Mathematics and Statistics
University of Melbourne
Parkville, Victoria 3010
Australia
E-mail: [email protected]
Q.-M. Shao
Department of Statistics
Chinese University of Hong Kong
Shatin, NT
Hong Kong
E-mail: [email protected]
W.-X. Zhou
Department of Operations Research
and Financial Engineering
Princeton University
Princeton, New Jersey 08544
USA
and
School of Mathematics and Statistics
University of Melbourne
Parkville, Victoria 3010
Australia
E-mail: [email protected]
| 10math.ST
|
A Unified Method for First and Third Person Action
Recognition
Ali Javidani
Ahmad Mahmoudi-Aznaveh
Department of Computer Science and Engineering
Shahid Beheshti University
Tehran, Iran
[email protected]
Cyberspace Research Center
Shahid Beheshti University
Tehran, Iran
[email protected]
Keywords—Video
Classification;
Human
Action
Recognition; Deep Learning; Convolutional Neural Network
(CNN); Optical Flow
motion and recognizing activities in them is highly
challenging [3]. Due to the fact that circumstances of camera
and recording actions are completely different from each
other in the first and third person videos, there exist two
main approaches for classifying each group and to the best of
our knowledge, there does not exist a unified method that
works perfectly for both.
The main motivation of our work is to provide a unified
framework which can classify both first and third-person
videos. Toward this goal, two complementary streams are
designed to capture motion and appearance features of video
data. The motion stream is based on calculating optical flow
images to estimate short-term motion in video and by
following them over time, using PoT representation method
with different pooling operators, long-term motion dynamics
are extracted efficiently. The appearance stream is obtained
via describing the middle frame of the input video utilizing
pre-trained networks. Our method is evaluated on two
different datasets UCF11 (third-person) [1] and DogCentric
(first-person) [2]. It is demonstrated that the proposed
method achieves high accuracy for both datasets.
I. INTRODUCTION
II. RELATED WORKS
Video recognition is one of the popular fields in artificial
intelligence that aims to detect and recognize ongoing events
from videos. This can help humans to inject vision to robots
in order to assist them in different situations. For instance,
one of the most prominent applications of video
classification is driver-less cars which are going to become
available in the market.
Totally, there are two main categories of videos that
researchers conduct their experiments on them: third-person
and first-person videos. In third-person videos, most of the
times, camera is located in a specific place without any
movement or scarcely with a slight movement and records
the actions of humans [1]; while, in first-person videos the
person wears the camera and involves directly in events [2].
This is the reason that first-person videos are full of ego-
In general, there are two major approaches for
classifying videos: Traditional and Modern approaches.
Traditional ones are based on descriptors which try to detect
different aspects of each action. At the first step, features of
video segments are extracted. These features can be interest
points or dense points obtained from raw input frames[4, 5].
Harris3D is one of the ways to obtain 3D corner points from
video[6]. Then feature points are described by handcrafted
descriptors such as HOG[7], HOF[8] and MBH[9]. To
describe features more effectively, some of these descriptors
have been extended to 3 dimensions to incorporate temporal
information in their calculations. HOG3D[10] and
SIFT3D[11] are two of the most popular ones.
Abstract—In this paper, a new video classification
methodology is proposed which can be applied in both first and
third person videos. The main idea behind the proposed
strategy is to capture complementary information of
appearance and motion efficiently by performing two
independent streams on the videos. The first stream is aimed to
capture long-term motions from shorter ones by keeping track
of how elements in optical flow images have changed over time.
Optical flow images are described by pre-trained networks
that have been trained on large scale image datasets. A set of
multi-channel time series are obtained by aligning descriptions
beside each other. For extracting motion features from these
time series, PoT representation method plus a novel pooling
operator is followed due to several advantages. The second
stream is accomplished to extract appearance features which
are vital in the case of video classification. The proposed
method has been evaluated on both first and third-person
datasets and results present that the proposed methodology
reaches the state of the art successfully.
Raw Frames
Optical Flow
Motion Feature
Vector
Description
Description
Final Feature
Vector
PoT (Max-SumHistogram of time
series gradientVariance)
Multi-Class
SVM
Description
Middle
Frame
Description
Appearance
Feature Vector
Figure 1:General pipeline of the proposed methodology. Our framework has two streams for obtaining motion and appearance features. The top stream
extracts motion, while the bottom extracts appearance features.
Following feature extraction and description phases, in
order to obtain a constant-length feature vector and
becoming independent of some variables, such as number of
frames for each video or number of interest points, an
encoding step is required. For so doing, encoding methods
like Bag of Visual Words (BoVW) [12], Fisher kernel[13]
have been used till now. The experiments illustrate that
fisher kernel is more accurate than the former one.
However, recently PoT encoding method was proposed by
Ryoo et al. and it could reach state-of-the-art results in the
case of first-person videos[3].
Modern approaches are mostly based on deep learning.
Convolutional Neural Networks (CNNs) could succeed in
giving the best results on image recognition, image
segmentation and so forth. Although one problem in the
case of video domain is that these networks are designed for
two-dimensional input images. To address this problem,
some researches have been conducted. As a case in point,
Karpathy et al. introduced four models of two-dimensional
but multi-channel CNNs and in their models time dimension
was incorporated in different channels of the network[14].
Zisserman et al. proposed two stream CNN model to classify
videos. However, their method suffers from the problem that
the number of stacked optical flow frames given to CNN is
limited due to the problem of overfitting of the network[15].
Furthermore for better estimation of motions in video, a
3-dimensional convolutional neural network (C3D) was
devised[16]. All convolution and pooling layers in C3D
operate 3-d and the depth of time dimension convolution is a
small number due to the vast amount of convolution
calculations. Hence, it can only capture the short-term
motion dynamics and longer ones would be lost by using this
network. A recent work used stacked auto-encoders to obtain
tractable dimensionality for the case of first-person
videos[17]. Also another work designed a deep fusion
framework in which by the aid of LSTMs, representative
temporal information have been extracted and could reach
state-of-the-art results on three widely used datasets[18].
III. PROPOSED METHOD
In this section, we introduce our proposed method
pipeline. Generally videos, either third-person or first-person,
consist of two different aspects: appearance and motion.
Appearance is related to detecting and recognizing existing
objects in each frame, while motion is their following over
time. Hence, motion information is highly correlated with
temporal dimension. As it is depicted in Fig. 1, in order to
capture two mentioned aspects in video, our proposed
framework has two streams independent of each other. The
upper stream is for extracting motion and the bottom is for
appearance features.
In the following, motion feature extraction is explained in
more detail. Firstly the images of optical flow between
consecutive frames are calculated. This helps to estimate
short-term motion through nearby frames for each video.
However, estimating long-term motion is strictly challenging
and is an open area of research.
Here, the idea is to keeping track of how short-term
motion elements vary over time to estimate longer changes.
Therefore, optical flow images should be described by a
specific set of features to be pursued over the time
dimension. We found that the best way for doing so is
utilizing pre-trained networks that already have been trained
on large scale image datasets. By doing this, not only
training of network, which is a highly time consuming
process, is not needed but also a strong representation would
be obtained since these networks could reach state-of-the-art
results for image recognition.
Aligning representation of sequential frames beside each
other leads to obtaining a set of multi-channel time series.
There are various ways to extract features from time series.
For doing so, PoT representation plus a novel pooling
operator is chosen due to several prominent reasons. Firstly,
thanks to the temporal filters, time series are break down to
sub-intervals which assists to represent high-level activity
from lower levels. Furthermore, PoT is benefited from
extracting different features from time series that each of
them can represent different aspects of data. The resulted
time series, especially those coming from first person videos
which are full of ego-motion, are more sophisticated than a
specific feature (e.g. max) can represent them. As a result,
PoT framework can be beneficial for extracting motion
features from time series.
PoT representation method extracts different features
max, sum and histogram of time series gradient from time
series. The final feature vector representation, which is
designed to be motion features in our framework, is the
concatenation of all of these features together for each time
series. We add variance as another pooling operator to the
pooling set and demonstrate that this feature can also extracts
useful information. Below is the definition of each pooling
operator in the time domain [ts , te]. fi(t)is the value of ith time
series in time t. The max and sum pooling operators are
defined as:
[
]
[
(1)
]
∑
(2)
And the histograms of time series gradient pooling
operators are defined as:
[
]
[
∑
]
∑
(3)
where
{
{
(
)
(4)
(
)
(5)
We proposed variance as a new pooling operator as
follows:
(
)̅
(6)
[
] ∑
̅
Figure 2: Some sample frames from different classes of two different
datasets UCF11 (left) and DogCentric (right).
vectors that are expected to represent complementary
information. As the last step, a SVM is trained on the final
feature vector.
IV. EXPERIMENTAL RESULTS
We conducted our experiments on two public datasets:
UCF11 and DogCentric. Fig. 2 represents some sample
frames of them. UCF11 is a third-person type which contains
11 different human activities e.g. playing basketball,
volleyball and it consists of about 1600 human activity
videos. In this dataset camera is not usually located in a
specific place and it has large amounts of movement. To
evaluate our method, we performed Leave One Out Cross
Validation (LOOCV) on this dataset as in the original
work[1].
DogCentric is a first-person activity dataset in which a
camera is located on the back of dogs. Thus, it has large
amounts of ego-motion which makes it highly challenging
[2]. It has 10 class activities consisting activities of dogs
such as walking and drinking as well as interacting with
humans. The total number of videos in this dataset is 209.
Like other previous methods, half of the total number of
videos per class is used as training and the other half is used
as testing. For classes with odd number of clips, the number
TABLE 1: COMPARISON OF THE ENCODING METHODS ON
THE DOGCENTRIC DATASET. F1-SCORES PER CLASS AND FINAL
CLASSIFICATION ACCURACY.
Activity Class
∑
(7)
Besides, in order to concentrate better on the resulted
time series, temporal pyramid filters are applied on each
series. Hence, the resultant time series are: the whole time
domain, half-parts of the whole time domain, one-fourth
parts and so forth.
Motion features are extracted as explained above by
applying pooling operators on each level of the resulted time
series after exploiting temporal pyramid filters. On the other
hand, appearance features have substantial role in classifying
videos. In our pipeline, we used middle frame of each video
and feed it to a pre-trained network to obtain appearance
feature vector. The final representation for each video is
acquired from concatenating motion and appearance feature
Ball Play
Car
Drink
Feed
Turn Head(left)
Turn Head(right)
Pet
Body Shake
Sniff
Walk
Final accuracy
BoVW
48.29
73.23
12.60
50.62
29.73
50.44
42.85
41.33
54.93
44.69
47.39
Method Accuracy (%)
IFV
PoT
Proposed
61.46
81.54
45.23
57.64
31.13
31.82
74.66
89.54
64.02
67.65
46.54
64.23
44.91
55.68
78.34
72.70
75.51
56.42
65.00
59.41
29.15
35.57
74.73
89.60
64.34
57.35
32.96
36.85
77.23
72.76
76.10
56.87
67.12
of test instances is one more than number of training. We
performed our algorithm 100 times with different
permutation for train and test sets. The mean classification
accuracy is reported in Table 1. It is clear that the proposed
method has been achieved a significant improvement in
terms of classification accuracy in compare with two
traditional representation methods Bag of Visual Words
(BoVW) and Improved Fisher Vector (IFV). In addition, the
proposed method could outperform the baseline PoT method
in most classes of the DogCentric dataset and also in the
final accuracy.
For obtaining optical flow images of consecutive
frames, we used popular method of HORN & SCHUNCK
[19]. To convert it to colorful images, in order to be fed to
pre-trained networks, flow visualization code of Baker et al.
[20] was followed. In our implementation, this method was
applied to all frames of each video and we did not sample
the existing frames. GoogleNet as a pre-trained network is
utilized to describe either optical flow images (in motion
stream) or middle frame (in appearance stream). This was
feasible by omitting softmax and fully-connected layers of
TABLE 2: COMPARISON OF CLASSIFICATION ACCURACY OF
THE DOGCENTRIC DATASET ACCORDING TO TEMPORAL
PYRAMID LEVELS.
Number of temporal
pyramids
1
2
3
4
5
Classification
accuracy
58.25%
59.66%
66.33%
67.12%
66.85%
TABLE 3: COMPARISON OF OUR RESULTS TO THE STATE-OFTHE-ART APPROACHES ON UCF11 DATASET.
Method
Accuracy
Hasan et al. [21]
Liu et al. [1]
Ikizler-Cinbis et al. [22]
Dense Trajectories[23]
Soft attention[24]
Cho et al. [25]
54.5%
71.2%
75.2%
84.2%
84.9%
88.0%
Snippets[26]
Two Stream LSTM(conv-L) [18]
Two Stream LSTM(fc-L) [18]
89.5%
89.2%
93.7%
Proposed method(Linear SVM)
Two Stream LSTM(fu1) [18]
Two Stream LSTM(fu2) [18]
93.75%
94.2%
94.6%
Proposed method(chi-squared SVM)
95.73%
the network. GoogleNet has 1024 neurons in this layer.
Furthermore in the case of number of temporal
pyramids, different experiments have been conducted and
results are illustrated in the Table 2. It can be seen that by
increasing the number of temporal pyramids up to 4 levels,
the classification accuracy has been improved. While by
increasing it to five levels, it has decreased in compare with
four levels. We believe this phenomenon is due to the fact
that by increasing the temporal pyramid levels, the number
of dimensionality would increase dramatically. On the other
hand, there do not exist enough training data for learning the
classifier. This is the reason that increasing number of
temporal pyramids cannot always improve the performance
of the system.
The proposed method is also evaluated on third-person
video dataset UCF11. The number of temporal pyramids
used for this dataset is 3 and sampling between frames was
not performed. Comparison of our method to the state-ofthe-art results on this dataset is reported in Table 3. As can
be seen, the proposed method with chi-squared SVM could
reach the best results on this dataset.
In all our experiments SVM classifier with linear and
chi-squared kernel is used and the later one showed better
performances.
V. CONCLUSION
In this paper, a new approach for video classification
was proposed which has the capability of employing for two
different categories of first and third-person videos. Shortterm/long-term motion changes are calculated by extracting
discriminant features from motion time series following PoT
representation method with a novel pooling operator. Final
feature vector is resulted from concatenating two
complementary feature vectors of appearance and motion to
perform the classification. By evaluating the proposed
method on two different types of datasets and comparing the
obtained results to the state of the art, it is concluded that the
proposed method not only works perfectly for both groups
but also increases the accuracy.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
J. Liu, J. Luo, and M. Shah, "Recognizing realistic actions from
videos “in the wild”," in Computer vision and pattern
recognition, 2009. CVPR 2009. IEEE conference on, 2009, pp.
1996-2003.
Y. Iwashita, A. Takamine, R. Kurazume, and M. S. Ryoo,
"First-person animal activity recognition from egocentric
videos," in Pattern Recognition (ICPR), 2014 22nd
International Conference on, 2014, pp. 4310-4315.
M. S. Ryoo, B. Rothrock, and L. Matthies, "Pooled motion
features for first-person videos," in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2015,
pp. 896-904.
L. Liu, L. Shao, F. Zheng, and X. Li, "Realistic action
recognition via sparsely-constructed Gaussian processes,"
Pattern Recognition, vol. 47, pp. 3819-3827, 2014.
J. R. Uijlings, I. Duta, N. Rostamzadeh, and N. Sebe, "Realtime
video classification using dense HOF/HOG," in Proceedings of
International Conference on Multimedia Retrieval, 2014, p. 145.
I. Laptev, "On space-time interest points," International Journal
of Computer Vision, vol. 64, pp. 107-123, 2005.
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
N. Dalal and B. Triggs, "Histograms of oriented gradients for
human detection," in 2005 IEEE Computer Society Conference
on Computer Vision and Pattern Recognition (CVPR'05), 2005,
pp. 886-893.
N. Dalal, B. Triggs, and C. Schmid, "Human detection using
oriented histograms of flow and appearance," in European
conference on computer vision, 2006, pp. 428-441.
H. Wang and C. Schmid, "Action recognition with improved
trajectories," in Proceedings of the IEEE International
Conference on Computer Vision, 2013, pp. 3551-3558.
A. Klaser, M. Marszałek, and C. Schmid, "A spatio-temporal
descriptor based on 3d-gradients," in BMVC 2008-19th British
Machine Vision Conference, 2008, pp. 275: 1-10.
P. Scovanner, S. Ali, and M. Shah, "A 3-dimensional sift
descriptor and its application to action recognition," in
Proceedings of the 15th ACM international conference on
Multimedia, 2007, pp. 357-360.
G. Csurka, C. Dance, L. Fan, J. Willamowski, and C. Bray,
"Visual categorization with bags of keypoints," in Workshop on
statistical learning in computer vision, ECCV, 2004, pp. 1-2.
G. Csurka and F. Perronnin, "Fisher vectors: Beyond bag-ofvisual-words image representations," in International
Conference on Computer Vision, Imaging and Computer
Graphics, 2010, pp. 28-42.
A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar,
and L. Fei-Fei, "Large-scale video classification with
convolutional neural networks," in Proceedings of the IEEE
conference on Computer Vision and Pattern Recognition, 2014,
pp. 1725-1732.
K. Simonyan and A. Zisserman, "Two-stream convolutional
networks for action recognition in videos," in Advances in
Neural Information Processing Systems, 2014, pp. 568-576.
D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri,
"C3D: generic features for video analysis," CoRR,
abs/1412.0767, vol. 2, p. 7, 2014.
X. Wang, L. Gao, J. Song, X. Zhen, N. Sebe, and H. T. Shen,
"Deep appearance and motion learning for egocentric activity
recognition," Neurocomputing, vol. 275, pp. 438-447, 2018.
H. Gammulle, S. Denman, S. Sridharan, and C. Fookes, "Two
Stream LSTM: A Deep Fusion Framework for Human Action
Recognition," in Applications of Computer Vision (WACV),
2017 IEEE Winter Conference on, 2017, pp. 177-186.
B. K. Horn and B. G. Schunck, "Determining optical flow,"
Artificial intelligence, vol. 17, pp. 185-203, 1981.
S. Baker, D. Scharstein, J. Lewis, S. Roth, M. J. Black, and R.
Szeliski, "A database and evaluation methodology for optical
flow," International Journal of Computer Vision, vol. 92, pp. 131, 2011.
M. Hasan and A. K. Roy-Chowdhury, "Incremental activity
modeling and recognition in streaming videos," in Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, 2014, pp. 796-803.
N. Ikizler-Cinbis and S. Sclaroff, "Object, scene and actions:
Combining multiple features for human action recognition,"
Computer Vision–ECCV 2010, pp. 494-507, 2010.
H. Wang, A. Kläser, C. Schmid, and C.-L. Liu, "Action
recognition by dense trajectories," in Computer Vision and
Pattern Recognition (CVPR), 2011 IEEE Conference on, 2011,
pp. 3169-3176.
S. Sharma, R. Kiros, and R. Salakhutdinov, "Action recognition
using visual attention," arXiv preprint arXiv:1511.04119, 2015.
J. Cho, M. Lee, H. J. Chang, and S. Oh, "Robust action
recognition using local motion and group sparsity," Pattern
Recognition, vol. 47, pp. 1813-1825, 2014.
J. Yue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O.
Vinyals, R. Monga, and G. Toderici, "Beyond short snippets:
Deep networks for video classification," in Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition,
2015, pp. 4694-4702.
| 1cs.CV
|
Cyber-Physical Attacks on UAS NetworksChallenges and Open Research Problems
Vahid Behzadan
arXiv:1702.01251v1 [cs.CR] 4 Feb 2017
Dept. of Computer Science and Engineering, University of Nevada, Reno, USA
[email protected]
Abstract—Assignment of critical missions to unmanned aerial
vehicles (UAV) is bound to widen the grounds for adversarial intentions in the cyber domain, potentially ranging from
disruption of command and control links to capture and use
of airborne nodes for kinetic attacks. Ensuring the security
of electronic and communications in multi-UAV systems is of
paramount importance for their safe and reliable integration
with military and civilian airspaces. Over the past decade, this
active field of research has produced many notable studies and
novel proposals for attacks and mitigation techniques in UAV
networks. Yet, the generic modeling of such networks as typical
MANETs and isolated systems has left various vulnerabilities out
of the investigative focus of the research community. This paper
aims to emphasize on some of the critical challenges in securing
UAV networks against attacks targeting vulnerabilities specific to
such systems and their cyber-physical aspects.
Index Terms—UAV, Cyber-Physical Security, Vulnerabilities
I. I NTRODUCTION
The 21st century is scene to a rapid revolution in our
civilization’s approach to interactions. Advancement of communication technologies, combined with an unprecedentedly
increasing trust and interest in autonomy, are pushing mankind
through an evolutionary jump towards delegation of challenging tasks to non-human agents. From mars rovers to search and
rescue robots, we have witnessed this trend of overcoming the
limitations inherent to us, through replacement of personnel
with cyber-physical systems capable of performing tasks that
are risky, repetitive, physically difficult or simply economically
infeasible for human actors.
Unmanned Aerial Vehicles, or UAVs, are notable examples of this revolution. Since the early 2000s, military and
intelligence theaters have seen an explosive growth in the
deployment of tactical UAVs for surveillance, transport and
combat operations. In the meantime, civilian use of UAVs has
gained traction as the manufacturing and operations costs of
small and mid-sized UAVs are undergoing a steady decline.
The cheaper cost of such UAVs has also led to a growing
interest in collaborative deployment of multiple UAVs to
perform specific tasks, such as monitoring the conditions
of farms and patrolling national borders. Yet, there are a
multitude of challenges associated with this vision, solving
which are crucial for safe and reliable employment of such
systems in civilian and military scenarios. One such challenge
is ensuring the security of systems that comprise UAVs,
as their remote operational conditions leave the burden of
command and control reliant on the onboard cyber-physical
GNSS, Telemetry
Satellite Relay
UAV1
UAV3
UAV2
Mobile
Ground Unit
Inter-UAV Link
Satellite Link
ATG Link
Ground
Control
Station
Fig. 1: Communication Links in a UAS Network
components. The body of literature on this issue has seen an
accelerated growth in recent years [1], which is partly due to
major cyber attacks on UAVs [2]. The overwhelming number
of potential vulnerabilities in UAVs indicates the need for
vigorous standards and frameworks for assurance of reliability
and resilience to malicious manipulations in all aspects of
UAVs, from the mechanical components to the information
processing units and communications systems.
In multi-UAV operations, Inter-UAV links are necessary
for exchange of situational and operational commands, which
are the basis of essential functions such as formation control
and task optimization. As for the architecture of these UAV
networks, the current consensus in the research community is
biased towards decentralized and ad hoc solutions, which allow
dynamic deployment of Unmanned Aerial Systems (UAS)
with minimal time and financial expenditure on pre-mission
preparations.
Structure of a typical UAS network is shown in Figure 1.
By considering the various types of links and interfaces
depicted in this figure, it can be deduced that such networks
are inherently of a complex nature. Integration of multiple
subsystems not only aggregates their individual vulnerabilities,
but may result in new ones that are rooted in the interactions
between those subsystems. Hence, UAS present the research
community with a novel interdisciplinary challenge. The aim
of this paper is to emphasize on some of the critical vulnerabilities specific to network and communications aspects of
UAVs, and provide the research community with a list of open
problems in ensuring the safety and security of this growing
technology.
II. U NIQUENESS OF UAS N ETWORKS
Accurate analysis of vulnerabilities in UAS networks necessitates an understanding of how an airborne network differs
from traditional computer networks. Much of recent studies in
this area compare UAS networks to Mobile Ad hoc Networks
(MANETs) and Wireless Sensor Networks (WSN), as UAS
communications and protocols may initially seem similar to
those of generic distributed and mobile networks. Yet differences in mobility and mechanical degrees of freedom, as well
as their operational conditions, build the grounds for separate
classification of UAS networks. One such distinguishing factor
is the velocity of airborne vehicles, which may range up to
several hundreds of miles per hour. The high mobility of
airborne platforms increases the complexity of requirements
for the communications subsystem and many aspects of the
UAS network. In the link layer, management of links and
adaptation of access control has to be fast enough to accommodate tasks such as neighbor discovery and resource allocation
in an extremely dynamic environment. Likewise, the network
layer must be able to provide fast route discovery and path
calculation while preserving the reliability of the information
flow.
In the physical layer, not only communications, but the
kinetic aspects of the UAS give rise to unique requirements.
As the span of a UAS network may vary from close-by clusters
to far and sparse distributions, the transmission power of UAV
radios must be adjustable for efficient power consumption and
sustained communications. Also, since the geography and environment of the mission may vary rapidly, channel availability
in UAS links is subject to change. A potential solution is
for the UAS to be equipped with Dynamic Spectrum Access
(DSA) and adaptive radios to provide the required agility. Furthermore, the conventional antenna arrangement on airborne
platforms is such that changes in orientation and attitude of
the aircraft affect the gain of onboard radios. This problem is
further intensified in unmanned aircraft, as the elimination of
risk to human pilot allows longer unconventional maneuvers.
These considerations clarify the demand for a fresh vantage
point for analyzing the problem of security in UAS networks.
The reliability of today’s mission-critical UAVs need to be
studied with models that adopt a more inclusive view of such
systems and the impact of seemingly benign deficiencies on
the overall vulnerability of UAVs.
III. A NATOMY OF A UAV
UAVs are cyber-physical systems, meaning that their operations are reliant on the interaction between physical and
computational elements of the system. Consequently, security
of a UAV is dependent not only on the computation and
communications elements and protocols, but also on the physical components of the system [3]. This heavy entanglement
of traditionally independent components requires a thorough
framework for analysis of security issues in UAVs to be
inclusive of the entire airframe. One obstacle in developing
such a framework is the variety of UAV architectures and
capabilities which makes the design of a generic model
IFF
Antenna
SATCOM
Antenna
GNSS+IMU NAV
Optical and
IR Sensors
Multiband
Data Link
Radar Antenna Antenna
UHF Antenna
Fig. 2: Sensing and Communication Components of a UAV
difficult. Yet, the similarity of fundamental requirements of
such systems allows for generation of a high level system
model for conventional types of UAVs. Figure 2 depicts
a breakdown of components in a conventional UAV. Most
UAVs contain multiple communication antennas, including
air to ground (ATG), air to air (ATA), satellite data link
and navigation antennas, along with a set of sensors. The
positioning and navigation of a UAV is typically consisted
of a Global Navigation Satellite System (GNSS) receiver for
accurate positioning, and an Inertial Measurement Unit (IMU)
for relative positioning based on readings from kinetic sensors.
This subsystem can be further extended to include air traffic
monitors such as ADS-B and collision avoidance systems.
Inside the fuselage, one or more processors supervise the
operation and navigation of the UAV, using the output of
various radios and sensors for adjustment of electronic and
mechanical parameters. This process is performed by adaptive
control mechanisms, many of which are dependent on feedback loops. Each of the elements mentioned in this section
may become the subject of malicious exploitation, leading the
UAV into undesirable states and critical malfunctions.
IV. OVERVIEW OF P OTENTIAL ATTACKS
Table I lists some of the uninvestigated attacks on UAS
networks, categorized according to both network functionalities and cyber-physical factors. The table emphasizes on the
criticality of the security problem, as the potential for vulnerability exists in every major component, ranging from the
outer fuselage and antennas to network layers and application
stack. This section provides an overview on the attacks listed in
Table I, and presents preliminary ideas on potential mitigating
approaches and areas of research.
A. Sensors and Navigation
Absence of a human pilot from the airframe of UAVs
puts the burden of observing the environment on the set of
sensors onboard the aircraft. Whether autonomous or remotely
piloted, sensors are the “eyes and ears” of the flight controller
and provide the environmental measurements necessary for
safe and successful completion of the mission. However,
malicious exploitation of sensors in critical cyber-physical
systems is widely neglected in vulnerability assessment of
TABLE I: Cyber-Physical Attacks on UAS Networks
Component
Attacks
Sensors
Visual Navigation
and Spoofing
Physical
Layer
Adaptive radios: deceptive attacks on spectrum sensing
Jamming
Antennas: Disruption and
Deception of Direction of
Arrival estimator, Beamnullinduced Jamming
Orientation: Self-disruption by
Induction of Defensive Maneuvers
Link Layer
Topology Inference, Topological Vulnerability of Formation
to Adaptive Jamming, Routing
attacks
Network
Layer
Traffic Analysis, Disruption of
Convergence
Air Traffic
Control
ADS-B Spoofing,
Induced Collisions
Fault
Handling
Manipulation of fault detection
TCAS-
such systems. An attacker may manipulate or misuse sensory
input or functions to trigger or transfer malware, misguide the
processes dependent on such sensors, or simply disable them
to cause denial of service attacks and trigger undesired failsafe mechanisms [4].
For navigational measurements, GNSS and IMU units are
traditionally used in tandem to provide accurate positioning of
the aircraft. It is well-known that GNSS signals, such as GPS,
are highly susceptible to spoofing attacks. The report in [5]
demonstrates that UAVs that only rely on commercial GPS
receivers for positioning are vulnerable to relatively simple
jamming and spoofing attacks, which may lead to crash or
capture of the UAV by adversaries. Since the establishment of
GPS, various countermeasures against GNSS spoofing have
been proposed, ranging from exploitation of direction and
polarization of the received GPS signal for attack detection
to beamforming and statistical signal processing methods for
elimination of spoofing signals [6]. However, the speed and
spatial freedom of UAVs render many of the basic assumptions
and criteria of such techniques inapplicable. In [7], the authors
propose the cross-examination of variations in IMU and GPS
readings for detection of spoofing attacks from anomalies in
fused measurements. While theoretically attractive, practical
deployment of this technique requires highly reliable IMUs
and adaptive threshold control for an efficient performance,
which are economically undesirable for the small UAVs industry. Such practical limitations in accuracy and implementation
leave this detection technique ineffective to advanced spoofing
attacks, demonstrating the insufficiency of current civilian
GNSS technology for mission-critical applications.
Fusion of IMU and GNSS systems with other sensors, such
as video camera, may lessen the possibility of spoofing. Yet,
vision-based navigation is also subject to attacks, the simplest
of which is blinding the camera by saturating its receptive
sensors with high intensity laser beams. A more sophisticated
attack may aim for deception of the visual navigation system:
In smaller areas, homogenizing or periodically modifying the
texture of the terrain beneath a camera-equipped UAV may
cause miscalculations of movement and orientation. Investigating the effect of such attacks on the control loop of a
fused positioning system may determine the feasibility of such
attacks and potential mitigation techniques.
Detection of attacks on the navigation subsystem is the basis
of reactive countermeasures, such as triggering of hovering or
return-to-base mechanisms. However, as the following section
demonstrates, fault-handling mechanisms are also potential
subjects to malicious manipulation. Robustness of the sensory
and navigational subsystem against spoofing attacks may be
further improved by implementation of proactive mechanisms
through elimination of spoofing signals, applicability of which
to UAVs is yet to be investigated.
B. Fault Handling Mechanisms
Even with the stringent reliability requirements of UAVs,
mechanical and electronic subsystems of UAVs remain prone
to faults due to physical damage and unpredicted state transitions. Therefore, critical UAV systems must consider the
possibility of faults and implement Fault Handling mechanisms to reduce the impact of such events on the system.
Typical examples of fault handling mechanisms are entering
a hovering pattern when temporary faults occur, return-tobase for persistent faults and self-destruction in the event
of fatal faults such as capture or crash. In remotely operated systems, fault handling mechanisms may be triggered
automatically once a certain fault is detected. This process
adds yet another attack surface to UAS networks, as the fault
detection mechanisms may be subject to manipulation [8].
For instance, if a temporary disruption of communications
triggers the hovering pattern of a UAV, an adversary can jam
the link to bind the motion of the aircraft, thus simplify its
kinetic destruction or physical capture. A more severe case is
when sensory manipulation allows the induction of capture
conditions on a tactical UAV, thereby triggering its autodestruction mechanism.
C. Air Traffic Control (ATC) and Collision Avoidance
Integration of unmanned vehicles with national and international airspaces requires guarantees on safety and reliability
of UAV operations. One major consideration in the safety
D. Physical Layer
Typical UAVs require multiple radio interfaces to retain
continuous connectivity with essential links to satellite relays,
ground control stations and other UAVs. This degree of complexity, along with the physical and mechanical characteristics
of UAVs, widen the scope of potential vulnerabilities and
enable multiple attacks that are specific to UAS networks. This
section presents a discussion on some of such attacks on the
physical layer of UAV nodes.
1. Adaptive Radios: As the operational environment of UAS
𝑡1 -- Detect UAV1
UAV1
𝑡2 -- TCAS Advisory: Descend
𝑡2 -- TCAS Advisory: Climb
UAV2
𝑡1 -- Detect UAV2
Altitude
of all airborne operations -manned and unmanned- is situation awareness and collision avoidance. Modern manned
aircraft in the major civilian airspaces are equipped with
secondary surveillance technologies such as Automatic Dependent Surveillance - Broadcast (ADS-B), which allow each
aircraft to monitor the air traffic in their vicinity. This information, along with other available means of traffic monitoring,
provide situation awareness to the Traffic advisory and Collision Avoidance System (TCAS), which monitors the risk of
collision with other aircraft and generates advisories on how
to prevent collisions.
With the growing interest in large-scale deployment of
UAVs, implementation of similar technologies in UAS is
crucial. Recent literature contain several proposals on TCAS
and ATC solutions for UAVs, many of which are based on
adaptation of ADS-B and commercial TCAS protocols. From
a security point of view, this approach suffers from several
critical vulnerabilities, rendering it unfeasible for missioncritical UAS applications. Firstly, ADS-B is an insecure protocol by design [1]. Lack of authentication and the unencrypted
broadcast nature of this protocol make room for relatively
simple attacks, ranging from eavesdropping to manipulation
of air traffic data by jamming or injection of false data.
Consequently, a TCAS system relying on ADS-B can produce
erroneous results and advisories, leading to unwanted changes
in the flight path or in the worst scenario, collisions.
Also, TCAS is shown to be susceptible to a flaw known as
“TCAS-Induced Collisions” [9]. Common implementations of
TCAS are not equipped with prediction capabilities to foresee
the longer-term effect of an advisory that they produce. In
dense traffic conditions, certain scenarios may cause the TCAS
to generate advisories that lead to a state where avoidance
of collision is not possible. Hence, an adversary capable
of manipulating the traffic data can intentionally orchestrate
conditions leading to TCAS-induced collisions. Authors of [9]
provide an example of this flaw for a 4 airplane scenario,
as illustrated in Figure 3. In this scenario, UAV1 and UAV2
are initially in a collision path, hence the TCAS in each
generates a collision avoidance advisory to descend and climb,
respectively. At a lower altitude, the same situation holds for
UAV3 and UAV4, causing UAV4 to climb, which puts UAV1
and UAV4 on a collision path. Even though that TCAS does
not fail to generate new correction advisories in both UAVs,
but the advisory is no longer practical as there is not enough
time before the collision to implement the new path.
𝑡2 -- TCAS Advisory: Descend
𝑡1 --Detect UAV4
UAV3
𝑡1 -- Detect UAV3
𝑡2 -- TCAS Advisory: Climb
UAV4
Fig. 3: Example of TCAS-Induced Collision in a 4 Airplane
Scenario
networks is highly dynamic, sustained and reliable communications necessitates the employment of radios that are capable
of adjusting to changes in propagation and links conditions.
Depending on the operational requirements, this adaptability
may apply to any of the physical layer parameters such
as transmit power, frequency, modulation, and configuration
of antennas. The procedure responsible for controlling these
parameters must essentially rely on environmental inputs,
which can be manipulated by adversaries to result in undesirable configurations. This issue is analogous to deceptive
attacks on the spectrum sensing process of cognitive radio
networks, for which various mitigation techniques have been
proposed based on anomaly detection and fusion of distributed
measurements [10]. However, the rapid variation of conditions
in a UAS network may lead to situations where determination
of a baseline for anomaly detection is not practical. The
same consideration also develops a necessity for rapid adjustments, which limits the acceptable amounts of redundancy and
overhead. Similarly, deployment of airborne nodes in hostile
environments further reduces the feasibility of relying on
real-time collaboration between distributed sensors. Therefore,
such countermeasures will not be sufficient for agile UAS
radios and novel solutions must be tailored according to the
unique requirements of airborne networks.
2. Antennas: The current trend in antenna selection for UAV
radios is favored towards omnidirectional antennas, defined
by their relatively homogeneous reception and transmission
in all directions of the horizontal or vertical planes. This
feature simplifies communications in mobile nodes, as the
homogeneity of gain eliminates the need for considering the
direction of transmissions. On the other hand, the indiscriminate nature of omnidirectional antennas extends the attack
surface for eavesdroppers and jammers, since they also do
not need to tune towards the exact direction of radios to
implement their attacks. A countermeasure against this class of
attack is the utilization of directional antennas, which can only
communicate in certain directions and are “blind” to others.
Besides their higher security, other advantages of directional
antennas include longer transmission ranges and spatial reuse,
thus providing a higher network capacity. One downside
associated with this approach is the inevitable escalation of
overhead. Maintaining directional communications in highly
mobile networks is a complex and costly task, as it requires
real-time knowledge of other nodes’ positions, as well as
employment of antennas capable of reconfiguring their beam
patterns.
To overcome the disadvantages of these two approaches, a
midway solution combining the simplicity of omnidirectional
radios and spatial selectivity of directional antennas can be
actualized in the form of beamforming antenna arrays. Such
antennas are capable of detecting the Direction of Arrival
(DoA) of individual signals. This measurement, along with
other system parameters, are then used to electronically reconfigure the radiation pattern and directionality of the antenna
array. Beamforming has been studied as a mitigation technique
against jamming attacks, as it allows spatial filtering of the
jammer’s signals by adjusting the antenna pattern such that a
null is placed towards the direction of the jammer [11]. The
accuracy and efficiency of this technique depends on correct
detection of the jamming signal, as well as the resolution of
beamformer’s DoA estimations. An adversary may attack the
DoA estimator by shaping its jamming signals to mimic waveforms of a nearby legitimate node, thus avoiding detection or
causing false detections.
Another attack scenario exploits the process of beamnulling
itself. In an ad hoc UAS network, beamnulling must be
implemented in a distributed fashion to allow targeted nodes to
retain or regain connectivity with the network independently.
Due to lack of coordination, nulls created by one node
towards a jammer may also null the direction of legitimate
signals. Depending on the mobility model and formation of the
network, an adversary may deploy multiple mobile jammers
with strategically controlled trajectories to manipulate the DoA
measurements, and eventually cause the network to null more
of its legitimate links than is necessary. In certain conditions,
the adversary can maximize the efficiency of jamming attacks by persistently manipulating the distributed beamnulling
mechanism in such a way that its solution converges towards
a maximally disconnected state. Analytical studies into feasibility criteria of this attack may produce insights into possible
countermeasures and mitigation techniques.
3. Orientation: As depicted in figure 2, a conventional
UAV employs multiple fixed antennas on different sides, each
of which is dedicated to a certain application. Consider the
ATG antenna which is placed on the lower side of the UAV.
As discussed previously, if the UAV performs a half-roll
maneuver or ascends with a steep climb angle, the ATG
antenna is no longer capable of communicating with the
ground antenna and therefore the ATG link is lost. This issue
can be exploited for jamming in UAS Networks that employ
the spatial retreat as a mitigation technique. By observing the
reaction of the nodes to jamming attacks, an adversary may
infer their reformation strategy, and adapt its attack such that
the defensive reformation of certain nodes leads to the loss of
some links due to the new orientation of antennas.
E. Link Layer and Formation
Similar to generic multihop wireless networks, the topology
of a UAS network is determined based on the location of
UAVs relative to each other: UAVs closer than a threshold
can directly communicate with each other, while those that
are farther must utilize relay nodes to reach their destination.
Knowledge of the topology of a network allows adversaries to
optimize attacks by analyzing the structure of their target and
determine the most vulnerable regions by identifying nodes
whose disconnection incur the maximum loss of connectivity
in the network. Even though the effect of topology on the
resilience of the network is widely studied, the proposed mitigation techniques fail to provide practical solutions for UAS
networks. A class of such solutions are based on a security
by obscurity approach, suggesting the employment of covert
communications between nodes to hide the topology of the
network from adversaries. Besides the undesirable overhead
of this approach in terms of decreased network throughput
and increased processing costs, it has been shown that the
topology of such networks can be estimated with a high degree
of accuracy via timing analysis attacks [12]. Therefore, hiding
the topology may not serve as a reliable solution in mission
critical scenarios.
An alternative mitigation technique is adaptive control of the
topology [13]. In this approach, detection of a jamming attack
triggers a reformation process during which the nodes of a
UAS network change their positions to retain connectivity. A
fundamental assumption of this approach is the ability of the
nodes to detect and localize attacks, which may not always be
practical. A promising area of further investigation is the problem of minimizing the topological vulnerability to targeted
jamming attacks. Development of real-time and distributed
formation control techniques that consider this optimization
problem may lead to highly efficient techniques for ensuring
dynamic resilience of mission-critical UAS networks.
A mitigation technique against topology inference attacks
is randomization of transmission delays. It is expected that
introducing randomness in forwarding delays weakens the
observed correlation between connected hops, and therefore
reduces the accuracy of timing analysis attacks. However,
the high mobility of UAS networks and the consequent requirement for minimal latency limit the maximum amount of
delay permissible in such networks. This constraint limits the
randomness of the forwarding delays, which may neutralize
the effect of mitigation technique. A potential alternative for
delay randomization is transmission of decoy signals to perturb
the adversary’s correlation analysis. This proposal may be
extended by incorporating it in topology control, such that
the resultant formation is optimized for decoy transmissions
in a way that spatial distribution of traffic in the network
appears homogeneous to an outside observer, thereby inducing
an artificial correlation between all nodes in the network. To
the extent of authors’ knowledge, the feasibility, overhead
and optimal implementation of this approach are yet to be
analytically and experimentally studied.
F. Network Layer
The impact of high mobility in UAS networks is greatly
accentuated in the network layer. Speed and frequency of
changes in the topology of a UAS network give rise to
many challenges that are still active subjects of research. Yet,
studies on security of routing mechanisms tend to follow the
tradition of equating UAS networks with MANETs. Indeed,
the unique features of unmanned airborne networks generate
a set of challenges in the network layer that do not match the
criteria of conventional MANETs. The highly dynamic nature
of UAS networks, as well as stringent requirements on latency,
necessitate novel routing mechanisms capable of calculating
paths in rapidly changing topologies. A survey of the state of
the art in this area is presented in [14]. The proposed methods
may be prone to potential vulnerabilities, and the demand for a
detailed technical analysis and comparison of these proposals
in terms of their security is yet to be fulfilled.
Similar to the link layer, the routing layer of UAS networks
is also vulnerable to traffic analysis attacks, aiming to infer
individual flows, as well as source-destination pairs of end-toend connections. Various mitigation techniques against such
attacks have been proposed [15], many of which rely on traditional approaches such as mixing and decoy transmissions. As
such techniques require addition of redundancies and overhead
to the UAS networks, a comprehensive feasibility analysis and
optimal design of the corresponding defense strategies is vital,
but not yet available to the research community.
Mobile routing in UAS networks is a surface for attacks
on convergence of the network. As discussed, the topology
of unmanned airborne networks is subject to manipulation by
adversarial actions such as exploitation of adaptive formation
control and jamming attacks. Also, many of the recently
proposed routing mechanisms for airborne networks rely on
global knowledge of the geographical positions of every node
in the network, which may also be prone to manipulation.
A sophisticated adversary may be able to design a strategic
combination of topological perturbation and sensor manipulations to prevent or slow the convergence of routing in the
network. Investigation of this attack in terms of feasibility, as
well as potential countermeasures may prove to be valuable
for efficient protection of UAS networks operating in hostile
environments.
V. C ONCLUSIONS
The cyber-physical nature of UAVs demand an extension to
the scope of ordinary vulnerability analysis for such systems.
In addition to threats in the electronic and computational
components, a largely overlooked class of vulnerabilities is
fostered by the interactions between the mechanical elements
and the computational subsystems. Pondering on the list of
critical attacks presented in this paper, an alarming conclusion
can be drawn: serious threats still remain unmitigated not
only in every networking component of UAS communications, but also in the interdependency of the network and
other components, including sensors and physical elements
of UAVs. Considering the seriousness of open issues in the
cyber-physical aspects of UAVs, a successful move towards the
age of mainstream unmanned aviation cannot be envisioned
without remedying the void of effective solutions for such
critical challenges.
R EFERENCES
[1] A. Kim, B. Wampler, J. Goppert, I. Hwang, and H. Aldridge, “Cyber
attack vulnerabilities analysis for unmanned aerial vehicles,” Infotech@
Aerospace, 2012.
[2] A. Y. Javaid, W. Sun, V. K. Devabhaktuni, and M. Alam, “Cyber security
threat analysis and modeling of an unmanned aerial vehicle system,” in
Homeland Security (HST), 2012 IEEE Conference on Technologies for,
pp. 585–590, IEEE, 2012.
[3] A. Banerjee, K. K. Venkatasubramanian, T. Mukherjee, and S. K. S.
Gupta, “Ensuring safety, security, and sustainability of mission-critical
cyber–physical systems,” Proceedings of the IEEE, vol. 100, no. 1,
pp. 283–299, 2012.
[4] V. Subramanian, R. Beyah, et al., “Sensory channel threats to cyber
physical systems: A wake-up call,” in Communications and Network
Security (CNS), 2014 IEEE Conference on, pp. 301–309, IEEE, 2014.
[5] K. Wesson and T. Humphreys, “Hacking drones,” Scientific American,
vol. 309, no. 5, pp. 54–59, 2013.
[6] A. Jafarnia-Jahromi, A. Broumandan, J. Nielsen, and G. Lachapelle,
“Gps vulnerability to spoofing threats and a review of antispoofing techniques,” International Journal of Navigation and Observation, vol. 2012,
2012.
[7] T. E. Humphreys, B. M. Ledvina, M. L. Psiaki, B. W. OHanlon,
and P. M. Kintner Jr, “Assessing the spoofing threat: Development
of a portable gps civilian spoofer,” in Proceedings of the ION GNSS
international technical meeting of the satellite division, vol. 55, p. 56,
2008.
[8] K. Hartmann and C. Steup, “The vulnerability of uavs to cyber attacksan approach to the risk assessment,” in Cyber Conflict (CyCon), 2013
5th International Conference on, pp. 1–23, IEEE, 2013.
[9] J. Tang, Causal models for analysis of TCAS-induced collisions. PhD
thesis, Universitat Autònoma de Barcelona, 2015.
[10] S. Bhattacharjee, S. Sengupta, and M. Chatterjee, “Vulnerabilities
in cognitive radio networks: A survey,” Computer Communications,
vol. 36, no. 13, pp. 1387–1398, 2013.
[11] S. Bhunia, V. Behzadan, P. A. Regis, and S. Sengupta, “Performance
of adaptive beam nulling in multihop ad hoc networks under jamming,”
IEEE International Symposium on Cyberspace Safety and Security (CSS
2015), New York.
[12] V. Behzadan and S. Sengupta, “Real-time inference of topological
structure and vulnerabilities for adaptive jamming against tactical ad hoc
networks,” Under Review in Elsevier Journal of Computer and System
Sciences, 2016.
[13] M. Zhu and S. Martı́nez, “Attack-resilient distributed formation control
via online adaptation,” in Decision and Control and European Control
Conference (CDC-ECC), 2011 50th IEEE Conference on, pp. 6624–
6629, IEEE, 2011.
[14] I. Bekmezci, O. K. Sahingoz, and Ş. Temel, “Flying ad-hoc networks
(fanets): A survey,” Ad Hoc Networks, vol. 11, no. 3, pp. 1254–1270,
2013.
[15] J. Kong, X. Hong, and M. Gerla, “An identity-free and on-demand
routing scheme against anonymity threats in mobile ad hoc networks,”
Mobile Computing, IEEE Transactions on, vol. 6, no. 8, pp. 888–902,
2007.
| 3cs.SY
|
A construction of q-ary linear codes with two weights✩
Ziling Henga,b,c , Qin Yuea,b,c
a Department
of Mathematics, Nanjing University of Aeronautics and Astronautics, Nanjing, 211106, PR China
b State Key Laboratory of Cryptology, P. O. Box 5159, Beijing, 100878, PR China
c State Key Laboratory of Information Security, Institute of Information Engineering, Chinese Academy of Sciences, Beijing,
100093, PR China
Abstract
arXiv:1605.04063v2 [cs.IT] 24 Jul 2017
Linear codes with a few weights are very important in coding theory and have attracted a lot of attention. In
this paper, we present a construction of q-ary linear codes from trace and norm functions over finite fields. The
weight distributions of the linear codes are determined in some cases based on Gauss sums. It is interesting
that our construction can produce optimal or almost optimal codes. Furthermore, we show that our codes
can be used to construct secret sharing schemes with interesting access structures and strongly regular graphs
with new parameters.
Keywords: linear codes, secret sharing schemes, strongly regular graphs, Gauss sums
2010 MSC: 94B05, 94B60, 11L05
1. Introduction
Let Fq denote the finite field with q elements. An [n, k, d] linear code C over Fq is a k-dimensional subspace
of Fnq with minimum Hamming distance d. An [n, k, d] code is called optimal if no [n, k, d + 1] code exists. Let
Ai denote the number of codewords with Hamming weight i in a code C with length n. The weight enumerator
5
of C is defined by 1 + A1 z + · · · + An z n . The sequence (1, A1 , · · · , An ) is called the weight distribution of C.
The code C is said to be t-weight if the number of nonzero Aj , 1 ≤ j ≤ n, in the sequence (1, A1 , · · · , An )
equals t. The weight distribution is an interesting topic which was investigated in [5, 9, 12, 21, 22, 30] and
many other papers. In particular, a survey of three-weight cyclic codes and their weight distributions were
provided in [11]. Weight distribution gives the minimum distance and the error correcting capability of a code.
10
In addition, it contains important information on the computation of the probability of error detection and
correction with respect to some error detection and correction algorithms [20].
Recently, Ding et al. proposed a very effective construction of linear codes in [9, 12] as follows. Let
D = {d1 , d2 , · · · , dn } ⊆ Fr , where r is a power of q. A linear code of length n over Fq is defined by
CD = (Trr/q (xd1 ), Trr/q (xd2 ), · · · , Trr/q (xdn )) : x ∈ Fr ,
where Trr/q (x) = x + xq + · · · + xq
s−1
denotes the trace function from Fr to Fq and r = q s . The set D is
called the defining set of C. If the set D is well chosen, the code C may have good parameters. By using this
✩ The
paper is supported by Foundation of Science and Technology on Information Assurance Laboratory (No. KJ-15-009).
Email addresses: [email protected] (Ziling Heng), [email protected] (Qin Yue)
Preprint submitted to Journal of LATEX Templates
July 25, 2017
construction and selecting proper defining sets, many good codes were found in [9, 12, 16, 17, 26, 28, 29]. Let
f be a function over Fr . Then this construction can be equivalently written as
CD = (Trr/q (xf (d1 )), Trr/q (xf (d2 )), · · · , Trr/q (xf (dn ))) : x ∈ Fr .
Let m, m1 , m2 be positive integers such that m1 |m, m2 |m and gcd(m1 , m2 ) = e. Let Trqmi /q be the trace
function from Fqmi to Fq , i = 1, 2. Let Nqm /qmi be the norm function from Fqm to Fqmi , then for x ∈ Fqm ,
Nqm /qmi (x) = x1+q
mi
m −1
+···+(qmi ) mi
qm −1
= x qmi −1 , i = 1, 2.
In this paper, we present a construction of a q-ary linear code as
CD =
Trqm1 /q (xNqm /qm1 (d1 )), · · · , Trqm1 /q (xNqm /qm1 (dn )) : x ∈ Fqm1 ,
(1.1)
where the defining set D is given as
D = {x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) + a = 0}
for a ∈ Fq . Since the norm function Nqm /qm2 : F∗qm → F∗qm2 is surjective, there exists an element c ∈ F∗qm such
that Nqm /qm2 (c) =
1
a
for a ∈ F∗q . If a ∈ F∗q , then
D
= {x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) + a = 0}
= {x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (cx)) + 1 = 0}
= {x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) + 1 = 0}.
15
This implies that we only need to consider a = 0, 1. We remark that this construction is a further generalization
of that in [18]. When m = m1 , in [18], the authors determined a lower bound of the minimum Hamming
distance of CD and gave its weight distributions for m2 = 2, a = 0 and m2 = 1, 2, a = 1, respectively.
The purpose of this paper is to determine the weight distribution of CD defined in Equation (1.1) in some
cases. Our main mathematical tools used in this paper are Gauss sums. Consequently, we obtain four classes
20
of two-weight linear codes with very flexible parameters. Examples given by us show that some codes are
optimal or almost optimal. As some applications, our codes are used to construct secret sharing schemes with
interesting access structures and strongly regular graphs with new parameters.
The following notations will be used in this paper:
25
χ, χ1 , χ2
canonical additive characters of Fq , Fqm1 , Fqm2 , respectively;
λ, λ1 , λ2
generators of multiplicative character groups of Fq , Fqm1 , Fqm2 , respectively;
G(λ), G(λ1 ), G(λ2 )
Gauss sums over Fq , Fqm1 , Fqm2 , respectively;
α
primitive element of F∗qm ;
e
e = gcd(m1 , m2 );
l
l = gcd( me1 , q − 1).
2. Gauss sums
In this section, we recall some basic results of Gauss sums which are important tools in this paper.
2
Let Fq be a finite field with q elements, where q is a power of a prime p. The canonical additive character
of Fq is defined as follows:
Trq/p (x)
where ζp = e
√
2π −1
p
χ : Fq −→ C∗ , χ(x) = ζp
,
denotes the p-th primitive root of complex unity and Trq/p is the trace function from Fq
to Fp . The orthogonal property of additive characters (see [23]) is given by:
q, if a = 0,
X
χ(ax) =
0 otherwise.
x∈Fq
i
Let ψ : F∗q −→ C∗ be a multiplicative character of F∗q . For F∗q = hβi, ψ(β) = ζq−1
for some 0 ≤ i ≤ q − 2. The
trivial multiplicative character ψ0 is defined by ψ0 (x) = 1 for all x ∈ F∗q . It is known from [23] that all the
b∗ , which is isomorphic to F∗ . The orthogonal property
multiplicative characters form a multiplication group F
q
q
of a multiplicative character ψ (see [23]) is given by:
q − 1, if ψ = ψ ,
X
0
ψ(x) =
0,
otherwise.
x∈F∗
q
The Gauss sum over Fq is defined by
G(ψ) =
X
ψ(x)χ(x).
x∈F∗
q
It is easy to see that G(ψ0 ) = −1 and G(ψ̄) = ψ(−1)G(ψ). If ψ 6= ψ0 , we have |G(ψ)| =
√
q. Gauss sums
can be viewed as the Fourier coefficients in the Fourier expansion of the restriction of χ to F∗q in terms of the
multiplicative characters of Fq , i.e.
χ(x) =
1 X
G(ψ̄)ψ(x), x ∈ F∗q .
q−1
(2.1)
ψ∈b
F∗
q
30
In this paper, Gauss sum is an important tool to compute exponential sums. In general, the explicit determination of Gauss sums is a difficult problem. In some cases, Gauss sums are explicitly determined in
[6, 23].
In the following, we state the Gauss sums in the so-called semi-primitive case.
Lemma 2.1. [6, Semi-primitive case Gauss sums] Let φ be a multiplicative character of order N of F∗r .
35
Assume that N 6= 2 and there exists a least positive integer j such that pj ≡ −1 (mod N ). Let r = p2jγ for
some integer γ. Then the Gauss sums of order N over Fr are given by
(−1)γ−1 √r,
if p = 2,
G(φ) =
γ(pj +1) √
(−1)γ−1+ N
r, if p ≥ 3.
Furthermore, for 1 ≤ s ≤ N − 1, the Gauss sums G(φs ) are given by
(−1)s √r,
if N is even, p, γ and
s
G(φ ) =
√
(−1)γ−1 r, otherwise.
3
pj +1
N
are odd,
The well-known quadratic Gauss sums are the following.
Lemma 2.2. [23, Theorem 5.15] Suppose that q = pt and η is the quadratic multiplicative character of Fq ,
where p is an odd prime. Then
(−1)t−1 √q,
if p ≡ 1 (mod 4),
√
G(η) = (−1)t−1 ( p∗ )t =
√
√
(−1)t−1 ( −1)t q, if p ≡ 3 (mod 4),
where p∗ = (−1)
40
p−1
2
p.
3. Exponential sums
In this section, we investigate two exponential sums which will be used to calculate the weight distribution
of CD .
Let χ be the canonical additive character of Fq . Let χi be the canonical additive character of Fqmi , i = 1, 2,
respectively. Denote
Ω(b) =
X
X
x∈F∗
y,z∈F∗
q
qm
and
∆(b) =
X
qm −1
qm −1
χ1 (ybx qm1 −1 )χ2 (zx qm2 −1 )χ(z), b ∈ F∗qm1 ,
X
qm −1
x∈F∗
y,z∈F∗
q
qm
qm −1
χ1 (ybx qm1 −1 )χ2 (zx qm2 −1 ), b ∈ F∗qm1 .
Firstly, we begin to compute the exponential sum Ω(b).
Lemma 3.1. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m and gcd( me1 , q − 1) = l, where e =
b∗ = hλi, F
b ∗mi = hλi i, ti =
gcd(m1 , m2 ). Let F
q
q
Ω(b) =
qmi −1
qe −1 ,
i = 1, 2. For b ∈ F∗qm1 , we have
m2
(q m − 1)(q − 1) X
G(λ̄t11 s )G(λt22 s )λt11 s (b)G(λ̄ e s ),
(q m1 − 1)(q m2 − 1)
s∈S
e
q −1
where S = { q−1
l j : j = 0, 1, . . . , q−1 l − 1}.
qm −1
45
qm −1
Proof. For F∗qm = hαi, let α1 = α qm1 −1 and α2 = α qm2 −1 . Then we have F∗qm1 = hα1 i and F∗qm2 = hα2 i. This
implies that
m
Ω(b) =
−2
X qX
qm −1
i
qm −1
i
χ1 (ybα qm1 −1 )χ2 (zα qm2 −1 )χ(z)
y,z∈F∗
q i=0
m
=
−2
X qX
y,z∈F∗
q
χ1 (ybαi1 )χ2 (zαi2 )χ(z).
i=0
Using the Fourier expansion of additive characters (see Equation (2.1)), we have
m
Ω(b) =
(q m1
qX
−2
X
1
χ(z)
− 1)(q m2 − 1)
∗
i=0
y,z∈Fq
X
ψ1 ∈b
F∗
q m1
4
G(ψ̄1 )ψ1 (ybαi1 )
X
ψ2 ∈b
F∗
q m2
G(ψ̄2 )ψ2 (zαi2 )
=
(q m1
X
1
χ(z)
m
2
− 1)(q − 1)
∗
y,z∈Fq
X
G(ψ̄1 )G(ψ̄2 )ψ1 (yb)ψ2 (z)
m
qX
−2
ψ1 (αi1 )ψ2 (αi2 ).
i=0
ψj ∈b
F∗m
q j
j=1,2
Since mi |m, we obtain ord(ψi )|(q m − 1), where i = 1, 2. Therefore, we have
(ψ1 (α1 )ψ2 (α2 ))q
m
−1
=1
and
m
qX
−2
ψ1 (αi1 )ψ2 (αi2 )
=
(ψ1 (α1 )ψ2 (α2 ))i
i=0
i=0
=
50
m
qX
−2
q m − 1, if ψ (α )ψ (α ) = 1,
1
1
2
2
0
otherwise.
b ∗mi = hλi i such that λi (αi ) = ζqmi −1 , where i = 1, 2. Assume that ψ1 = λu and ψ2 = λv for 0 ≤ u ≤
Let F
q
1
2
q m1 − 2 and 0 ≤ v ≤ q m2 − 2. If ψ1 (α1 )ψ2 (α2 ) = 1, then ζqum1 −1 ζqvm2 −1 = 1 which is equivalent to
(q m2 − 1)u + (q m1 − 1)v ≡ 0
(mod (q m1 − 1)(q m2 − 1)).
(3.1)
This implies that (q m2 − 1)u + (q m1 − 1)v ≡ 0 (mod q mi − 1), i = 1, 2. Therefore, (q m2 − 1)u ≡ 0 (mod q m1 − 1)
and (q m1 − 1)v ≡ 0 (mod q m2 − 1). It is known that
gcd(q m1 − 1, q m2 − 1) = q gcd(m1 ,m2 ) − 1 = q e − 1.
Then we have u ≡ 0 (mod
e
q − 2, where t1 =
qm1 −1
qe −1 )
qm1 −1
qe −1 ,
t2 =
and v ≡ 0 (mod
qm2 −1
qe −1 .
qm2 −1
qe −1 ).
Denote u = t1 s1 and v = t2 s2 for 0 ≤ s1 , s2 ≤
Substituting u = t1 s1 , v = t2 s2 into Equation (3.1), we have
s1 + s2 = q e − 1. Hence,
e
Ω(b) =
(q m1
qX
−2
X
qm − 1
G(λ̄1t1 s1 )G(λt22 s1 )λ1t1 s1 (yb)λ̄2t2 s1 (z)
χ(z)
− 1)(q m2 − 1)
∗
s =0
y,z∈Fq
1
e
=
(q m1
qX
−2
X
X
qm − 1
G(λ̄1t1 s1 )G(λt22 s1 )λ1t1 s1 (b)
χ(z)λ̄2t2 s1 (z)
λ1t1 s1 (y).
m
2
− 1)(q − 1) s =0
∗
∗
z∈Fq
1
y∈Fq
qm2 −1
qm1 −1
Assume that F∗q = hβi, where β = α1 q−1 = α2 q−1 . Hence, λ1 (β) = λ2 (β) = ζq−1 . Since gcd( me1 , q − 1) = l,
55
we have
X
λ1t1 s1 (y)
y∈F∗
q
Denote S = {s1 : s1 ≡ 0 (mod
that
Ω(b) =
q−1
l ), 0
=
q−2
X
λ1t1 s1 (β i ) =
q−2
X
m1 s 1
e
(ζq−1
)i
i=0
i=0
q − 1, if s ≡ 0 (mod
1
=
0
otherwise.
q−1
l ),
b ∗ = hλi. Since λ2 (z) = λ(z) for z ∈ Fq , we have
≤ s1 ≤ q e − 2}. Let F
q
X
(q m − 1)(q − 1) X
G(λ̄t11 s1 )G(λ2t2 s1 )λ1t1 s1 (b)
χ(z)λ̄2t2 s1 (z)
m
m
2
1
(q − 1)(q − 1)
∗
s1 ∈S
z∈Fq
5
=
X
m2
(q m − 1)(q − 1) X
G(λ̄t11 s1 )G(λ2t2 s1 )λ1t1 s1 (b)
χ(z)λ̄ e s1 (z)
m
m
2
1
(q − 1)(q − 1)
∗
s1 ∈S
z∈Fq
X
G(λ̄t11 s1 )G(λ2t2 s1 )λ1t1 s1 (b)G(λ̄
m
=
(q − 1)(q − 1)
(q m1 − 1)(q m2 − 1)
m2
e
s1
).
s1 ∈S
The proof is completed.
We remark that the Fourier expansion of additive characters used in Lemma 3.1 is an effective technique
60
in computing exponential sums. It was also employed in [21] to determine the weight distribution of cyclic
codes by Li and Yue. By Lemma 3.1, we know that the value distribution of Ω(b) can be determined if the
Gauss sums are known. In the following, we mainly consider some special cases to give the value distribution
of Ω(b).
Lemma 3.2. Let l = 1 and other notations and hypothesises be the same as those of Lemma 3.1. Then the
65
value distribution of Ω(b), b ∈ F∗qm1 , is the following.
(1) If e = 1, then
Ω(b) =
−(q m − 1)(q − 1)
, b ∈ F∗qm1 .
(q m1 − 1)(q m2 − 1)
(2) If e = 2, then
Ω(b) =
−(qm −1)(q−1)
(qm1 −1)(qm2 −1) (1
−(qm −1)(q−1)
(qm1 −1)(qm2 −1) (1
+ (−1)
m1 +m2
2
q
+ (−1)
m1 +m2
2
+1
Proof. If l = 1, by Lemma 3.1, we have that G(λ̄
Ω(b) =
m2
e
s
m1 +m2
2
q
+1
m1 +m2
2
),
),
qm1 −1
q+1 times
q(qm1 −1)
times.
q+1
) = −1 and
−(q m − 1)(q − 1) X
G(λ̄t11 s )G(λt22 s )λt11 s (b),
(q m1 − 1)(q m2 − 1)
s∈S
e
−1
where S = {(q − 1)j : j = 0, 1, . . . , qq−1
− 1}. In the following, we discuss the value distribution of the
exponential sum Ω(b) for e = 1, 2, respectively.
(1)Assume that e = 1. It is clear that S = {0}. Then
Ω(b) =
−(q m − 1)(q − 1)
, b ∈ F∗qm1 .
(q m1 − 1)(q m2 − 1)
(2) Assume that e = 2. Then we have S = {(q − 1)j : j = 0, 1, . . . , q}. Hence,
q
Ω(b)
qm2 −1
70
=
q 2 −1
q 1 −1
q 1 −1
−(q m − 1)(q − 1) X
q+1 j
q+1 j
q+1 j
)G(λ
)λ
(b)
G(
λ̄
1
2
1
m
m
2
1
(q − 1)(q − 1) j=0
=
q
qm1 −1
qm2 −1
qm1 −1
X
−(q m − 1)(q − 1)
q+1 j
q+1 j
q+1 j
G(
λ̄
(1
+
)G(λ
)λ
(b)).
1
2
1
(q m1 − 1)(q m2 − 1)
j=1
m
m
m
qm1 −1
Note that ord(λ2 q+1 ) = ord(λ1 q+1 ) = q + 1. Now we give the value distribution of Ω(b) in several cases.
• If q is even, by Lemma 2.1 we have
qm1 −1
G(λ̄1 q+1
j
qm2 −1
)G(λ2 q+1
j
) = (−1)
6
m1 +m2
2
q
m1 +m2
2
, 1 ≤ j ≤ q.
Then
Ω(b) =
q
qm1 −1
m1 +m2 X
m1 +m2
−(q m − 1)(q − 1)
q+1 j
2
2
λ
q
(1
+
(−1)
(b)).
1
(q m1 − 1)(q m2 − 1)
j=1
Let b = αs1 , 0 ≤ s ≤ q m1 − 2. Then we have
q
X
qm1 −1
q+1 j
λ1
(b) =
q
X
js
ζq+1
j=1
j=1
q,
if s ≡ 0 (mod q + 1)
=
−1, otherwise.
Hence, the value distribution of Ω(b) is
m
m1 +m2
m1 +m2
−1)(q−1)
−(q
2
q 2 +1 ),
(qm1 −1)(qm2 −1) (1 + (−1)
Ω(b) =
m
m1 +m2
m1 +m2
−1)(q−1)
−(q
+1
2
q 2 ),
(qm1 −1)(qm2 −1) (1 + (−1)
qm1 −1
q+1 times,
q(qm1 −1)
times.
q+1
• If q is odd and m2 ≡ 0 (mod 4), we have m1 ≡ 2 (mod 4) due to gcd( m21 , q − 1) = 1. Since
and
m2
2
is even, by Lemma 2.1 we have
qm1 −1
G(λ̄1 q+1
j
qm2 −1
)G(λ2 q+1
j
) = (−1)j+1 q
m1 +m2
2
, 1 ≤ j ≤ q.
For b = αs1 , 0 ≤ s ≤ q m1 − 2,
Ω(b) =
=
q
qm1 −1
m1 +m2 X
−(q m − 1)(q − 1)
q+1 j
j+1
2
(−1)
λ
(1
+
q
(b))
1
(q m1 − 1)(q m2 − 1)
j=1
q
m1 +m2 X
−(q m − 1)(q − 1)
js
2
(−1)j+1 ζq+1
).
(1
+
q
(q m1 − 1)(q m2 − 1)
j=1
For s ≡ 0 (mod q + 1), we have
q
X
m1 +m2
−(q m − 1)(q − 1)
js
(−1)j+1 ζq+1
= 1 and Ω(b) = m1
(1 + q 2 ).
m
2
(q − 1)(q − 1)
j=1
For s ≡
q+1
2
(mod q + 1), we have
q
X
j=1
js
(−1)j+1 ζq+1
= −q and Ω(b) =
m1 +m2
−(q m − 1)(q − 1)
(1 − q 2 +1 ).
(q m1 − 1)(q m2 − 1)
(mod q + 1), one can see that
For s 6≡ 0, q+1
2
qs
s
3s
5s
ζq+1
+ ζq+1
+ ζq+1
+ · · · + ζq+1
=0
and
(q−1)s
2s
4s
6s
ζq+1
+ ζq+1
+ ζq+1
+ · · · + ζq+1
= −1.
This implies that
q
X
m1 +m2
−(q m − 1)(q − 1)
js
(−1)j+1 ζq+1
= 1 and Ω(b) = m1
(1 + q 2 ).
m
2
(q − 1)(q − 1)
j=1
Hence, the value distribution of Ω(b) is
m1 +m2
(qm −1)(q−1)
m
+1
2
− 1),
(q 1 −1)(qm2 −1) (q
Ω(b) =
m
m1 +m2
−(q
−1)(q−1)
m
(1 + q 2 ),
m
(q
1 −1)(q
2 −1)
7
qm1 −1
q+1 times,
q(qm1 −1)
times.
q+1
m1
2
is odd
75
• If q is odd and m2 ≡ 2 (mod 4), we have m1 ≡ 2 (mod 4) due to gcd( m21 , q − 1) = 1. In this case,
the value distribution of Ω(b) can be obtained in a similar way. We omit the details here. The value
distribution of Ω(b) is given as
m
m1 +m2
−1)(q−1)
−(q
+1
2
),
(qm1 −1)(qm2 −1) (1 + q
Ω(b) =
m
m1 +m2
−1)(q−1)
−(q
(1 − q 2 ),
m
m
(q
1 −1)(q
2 −1)
qm1 −1
q+1 times,
q(qm1 −1)
times.
q+1
Note that the value distribution of Ω(b) can be represented in a unified form for e = 2. The proof is completed.
80
Lemma 3.3. Let l = 2, e = 1and other notations and hypothesises be the same as those of Lemma 3.1. Then
the value distribution of Ω(b), b ∈ F∗qm1 , is given as follows.
m1 +m2 +1
(qm −1)(q−1)
m
2
),
(q 1 −1)(qm2 −1) (−1 − q
Ω(b) =
m
m1 +m2 +1
(q
−1)(q−1)
m
2
),
(q 1 −1)(qm2 −1) (−1 + q
qm1 −1
2
qm1 −1
2
times,
times.
Proof. Since l = 2, e = 1, by Lemma 3.1 we have that
Ω(b) =
(q m − 1)(q − 1) X
G(λ̄t11 s )G(λt22 s )λt11 s (b)G(λ̄m2 s ), b ∈ F∗qm1 ,
(q m1 − 1)(q m2 − 1)
s∈S
where S = {0, q−1
2 }. It is clear that m1 is even and m2 is odd. Hence, by Lemma 2.2,
Ω(b) =
=
=
=
qm1 −1
qm2 −1
qm1 −1
q−1
(q m − 1)(q − 1)
2
2
2
(−1
+
G(
λ̄
)G(λ
)λ
(b)G(λ̄ 2 m2 ))
1
2
1
(q m1 − 1)(q m2 − 1)
qm2 −1
qm1 −1
qm1 −1
q−1
(q m − 1)(q − 1)
2
2
2
)G(λ
)λ
(b)G(λ̄ 2 ))
(−1
+
G(
λ̄
1
2
1
m
m
(q 1 − 1)(q 2 − 1)
qm1 −1
m1 +m2 +1
(p−1)(m1 +m2 +1)t
(q m − 1)(q − 1)
2
4
2
q
λ
(b))
(−1
−
(−1)
1
m
m
(q 1 − 1)(q 2 − 1)
m1
m1 +m2 +1
(qm −1)(q−1)
m
2
), q 2−1 times,
(q 1 −1)(qm2 −1) (−1 − q
m1
m1 +m2 +1
(qm −1)(q−1)
m
2
(−1 + q
), q −1 times.
m
(q
1 −1)(q
2
2 −1)
For l = e = 2, the value distribution of Ω(b) can’t be given because the Gauss sums of order 2(q + 1) are
85
unknown in general. However, for e = 1 and l = 3, 4, we can easily obtain the value distributions of Ω(b)
because the cubic and quartic Gauss sums are known. We omit the details here.
In the following, we begin to investigate the exponential sum ∆(b), b ∈ F∗qm1 .
Lemma 3.4. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m. Denote e = gcd(m1 , m2 ). Let ti =
qmi −1
qe −1 ,
i = 1, 2. For b ∈ F∗qm1 , we have
∆(b) =
(q m − 1)(q − 1)2 X
G(λ̄t11 s )G(λt22 s )λt11 s (b),
(q m1 − 1)(q m2 − 1)
s∈S
e
−1
b∗mi = hλi i, i = 1, 2.
− 1} and F
where S = {(q − 1)j : j = 0, 1, . . . , qq−1
q
8
qm −1
qm −1
Proof. For F∗qm = hαi, let α1 = α qm1 −1 and α2 = α qm2 −1 . Then we have F∗qm1 = hα1 i and F∗qm2 = hα2 i. This
90
implies that
∆(b) =
X
m
qX
−2
X
m
qX
−2
qm −1
qm −1
i
i
χ1 (ybα qm1 −1 )χ2 (zα qm2 −1 )
y,z∈F∗
q i=0
=
y,z∈F∗
q
χ1 (ybαi1 )χ2 (zαi2 ).
i=0
Using the Fourier expansion of additive characters (see Equation (2.1)), we have
m
∆(b) =
=
Let ti =
qmi −1
qe −1 ,
(q m1
(q m1
−2
X qX
1
− 1)(q m2 − 1)
∗ i=0
y,z∈Fq
X
1
− 1)(q m2 − 1)
∗
X
G(ψ̄1 )ψ1 (ybαi1 )
G(ψ̄2 )ψ2 (zαi2 )
ψ2 ∈b
F∗
q m2
ψ1 ∈b
F∗
q m1
X
X
G(ψ̄1 )G(ψ̄2 )ψ1 (yb)ψ2 (z)
m
qX
−2
ψ1 (αi1 )ψ2 (αi2 ).
i=0
y,z∈Fq ψj ∈bF∗mj
q
j=1,2
i = 1, 2. From the proof of Lemma 3.1, we know that
m
qX
−2
q m − 1, if ψ (α )ψ (α ) = 1,
1
1
2
2
ψ1 (αi1 )ψ2 (αi2 ) =
0
otherwise,
i=0
and ψ1 (α1 )ψ2 (α2 ) = 1 if and only if ψ1 = λ1t1 s1 and ψ2 = λ2t2 s2 , where s2 = q e − 1 − s1 and 0 ≤ s1 ≤ q e − 2.
Hence,
e
∆(b)
=
(q m1
−2
X qX
qm − 1
G(λ̄t11 s1 )G(λ2t2 s1 )λt11 s1 (yb)λ̄2t2 s1 (z)
− 1)(q m2 − 1)
∗ s =0
y,z∈Fq
1
e
=
(q m1
qX
−2
X
X
qm − 1
G(λ̄1t1 s1 )G(λt22 s1 )λ1t1 s1 (b)
λ̄2t2 s1 (z).
λ1t1 s1 (y)
m
2
− 1)(q − 1) s =0
∗
∗
y∈Fq
1
qm1 −1
95
qm2 −1
z∈Fq
Assume that F∗q = hβi, where β = α1 q−1 = α2 q−1 . Hence, λ1 (β) = λ2 (β) = ζq−1 . This implies that
X
λ1t1 s1 (y)
q−2
X
λ1t1 s1 (β i ) =
q−2
X
m1 s 1
e
)i
(ζq−1
i=0
i=0
y∈F∗
q
q − 1, if m1 s1 ≡ 0
e
=
0
otherwise,
and
X
=
λ̄2t1 s1 (z) =
λ̄2t2 s1 (β i ) =
=
q−2
X
−
m2 s 1
(ζq−1 e )i
i=0
i=0
z∈F∗
q
Since gcd( me1 , me2 ) = 1, the system
q−2
X
(mod q − 1),
q − 1, if m2 s1 ≡ 0 (mod q − 1),
e
0
otherwise.
m1 s1
e
m2 s1
e
≡0
(mod q − 1),
≡0
(mod q − 1),
9
is equivalent to s1 ≡ 0 (mod q − 1), where 0 ≤ s1 ≤ q e − 2. Denote S = {s1 : s1 ≡ 0 (mod q − 1), 0 ≤ s1 ≤
q e − 2}. Then we have that
∆(b) =
(q m − 1)(q − 1)2 X
G(λ̄t11 s1 )G(λ2t2 s1 )λ1t1 s1 (b).
(q m1 − 1)(q m2 − 1)
s1 ∈S
100
For e = 1, 2, the value distribution of ∆(b) can be given as follows.
Lemma 3.5. Let the notations be the same as those of Lemma 3.3. Then the value distribution of ∆(b), b ∈
F∗qm1 , is given as follows.
(1) If e = 1, then ∆(b) =
105
(qm −1)(q−1)2
(qm1 −1)(qm2 −1)
for all b ∈ F∗qm1 .
(2) If e = 2, then
∆(b) =
(qm −1)(q−1)2
(qm1 −1)(qm2 −1) (1
(qm −1)(q−1)2
(qm1 −1)(qm2 −1) (1
+ (−1)
+ (−1)
m1 +m2
2
q
m1 +m2
2
+1
m1 +m2
2
q
+1
m1 +m2
2
),
),
qm1 −1
q+1 times
q(qm1 −1)
times.
q+1
Proof. The proof is similar to that of Lemma 3.2. We omit the details here.
4. The weight distribution of CD
In this section, we give the weight distribution of CD defined in Equation (1.1) in some special cases. The
well-known Griesmer bound of linear codes is the following.
Lemma 4.1. [24, Griesmer bound] For an [n, k, d] code over Fq , we have
n≥
110
k−1
X
i=0
⌈d/q i ⌉,
where ⌈x⌉ denotes the smallest integer which is larger than or equal to x.
4.1. The case a = 0
In the following, we determine the weight distribution of CD for a = 0.
Denote n = |D| = |{x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) = 0}|. Since the norm function
qm −1
Nqm /qm2 : F∗qm −→ F∗qm2 , x 7−→ x qm2 −1 ,
is an epimorphism of two multiplicative groups and the trace function Trqm2 /q : Fqm2 −→ Fq is an epimorphism
of two additive groups, we have
n = | ker(Nqm /qm2 )| · (| ker(Trqm2 /q )| − 1) =
q m − 1 m2 −1
(q
− 1).
q m2 − 1
(4.1)
Note that n = 0 when m2 = 1. Hence, we always assume that m2 > 1 in this section. For b ∈ F∗qm1 , we denote
N (b) = |{x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) = 0 and Trqm1 /q (bNqm /qm1 (x)) = 0}|.
10
115
By the basic facts of additive characters, we have that
N (b) =
1 X
q2
∗
X
χ(y Trqm1 /q (bNqm /qm1 (x)))χ(z Trqm2 /q (Nqm /qm2 (x)))
1 X
q2
∗
X
χ1 (ybx qm1 −1 )χ2 (zx qm2 −1 )
x∈Fqm y,z∈Fq
=
qm −1
qm −1
x∈Fqm y,z∈Fq
=
q
qm −1
qm −1
−1
1 X X
1 X X
qm1 −1 ) +
qm2 −1 )
+
χ
(ybx
χ
(zx
1
2
q2
q2
q2
∗
∗
∗
∗
m
x∈Fqm y∈Fq
1 X
+ 2
q
∗
X
χ1 (ybx
x∈Fqm z∈Fq
qm −1
qm1 −1
)χ2 (zx
qm −1
qm2 −1
)
x∈Fqm y,z∈F∗
q
=
qm −1
qm −1
1 X X
1
qm − 1
1 X X
qm1 −1 ) +
qm2 −1 ) +
+
∆(b).
χ
(ybx
χ
(zx
1
2
2
2
q2
q2
q
q
∗
∗
∗
∗
x∈Fqm y∈Fq
x∈Fqm z∈Fq
Note that the norm function Nqm /qm2 is an epimorphism. Hence,
X X
qm −1
χ2 (zx qm2 −1 )
=
qm − 1 X X
χ2 (zx)
q m2 − 1
∗
∗
=
(q − 1)(1 − q m )
q −1 X X
.
χ
(zx)
=
2
q m2 − 1
q m2 − 1
∗
∗
z∈F∗
x∈F∗
q
qm
x∈Fqm z∈Fq
m
z∈Fq x∈Fqm
Similarly,
X X
qm −1
χ1 (ybx qm1 −1 )
=
qm − 1 X X
χ1 (ybx)
q m1 − 1
∗
∗
=
(q − 1)(1 − q m )
q −1 X X
.
χ
(ybx)
=
1
q m1 − 1
q m1 − 1
∗
∗
y∈F∗
x∈F∗
q
qm
x∈Fqm y∈Fq
m
y∈Fq x∈Fqm
From the discussions above, we obtain that
N (b) =
qm − 1
q−1
q−1
1
(1 − m2
−
) + 2 ∆(b).
q2
q − 1 q m1 − 1
q
(4.2)
For any b ∈ F∗qm1 , the weight of a codeword
c(b) = (Trqm1 /q (bNqm /qm1 (d1 )), · · · , Trqm1 /q (bNqm /qm1 (dn )))
equals
wH (c(b)) = n − N (b) =
(q − 1)(q m − 1)(q m1 +m2 − q m1 +1 + q − 1)
1
− 2 ∆(b)
q 2 (q m1 − 1)(q m2 − 1)
q
(4.3)
by Equations (4.1) and (4.2). Hence, by Lemma 3.5, the parameters of CD for e = 1 are
[
120
q m − 1 m2 −1
q m1 −1 (q − 1)(q m − 1)(q m2 −1 − 1)
].
(q
−
1),
m
,
1
m
q 2 −1
(q m1 − 1)(q m2 − 1)
Then CD is an optimal one-weight linear code with respect to the Griesmer bound. However, any one-weight
linear code is not new because it is equivalent to a concatenated version of a simplex code. For e = 2, the
weight distribution of CD is given in the following.
11
Theorem 4.2. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m, and m2 > 1. Denote gcd(m1 , m2 ) =
e. Let CD be the linear code defined in Equation (1.1) for a = 0. If e = 2 and (m1 , m2 ) 6= (2, 2), then CD is a
m
125
m2 −1
two-weight linear code with parameters [ qqm2−1
− 1), m1 ] and its weight enumerator is given by Table I.
−1 (q
Table I. Weight distribution of the code in Theorem 4.2.
Weight
Frequency
0
1
m1 +m2
2
q m1 −1 (q−1)(q m −1)(q m2 −1 −1−(−1)
(q m1 −1)(q m2 −1)
m2 −m1
2
)
(q−1)q
m2 −m1
−1
2
)
q m1 −1
q+1
m1 +m2
2
q m1 −1 (q−1)(q m −1)(q m2 −1 −1−(−1)
(q m1 −1)(q m2 −1)
(q−1)q
+1
q(q m1 −1)
q+1
Proof. For e = 2, the weight distributions of CD can be obtained by Lemma 3.5 and Equation (4.3). It is easy
to verify that wH (cb ) > 0 for all b ∈ F∗qm1 if (m1 , m2 ) 6= (2, 2), then the dimension equals m1 .
Example 4.3. Let m1 = 2, m2 = 4, m = 4. If q = 2, then CD in Theorem 4.2 is an optimal [7, 2, 4] two-weight
linear code according to the Griesmer bound and has weight enumerator 1 + 2z 4 + z 6 . If q = 3, then CD in
130
Theorem 4.2 is an almost optimal [26, 2, 18] two-weight linear code according to the Griesmer bound and has
weight enumerator 1 + 6z 18 + 2z 24 .
Example 4.4. Let m1 = 4, m2 = 6, m = 12 and q = 2. Then CD in Theorem 4.2 is a [2015, 4, 1040] two-weight
linear code. Its weight enumerator is given by 1 + 10z 1040 + 5z 1144 .
4.2. The case a = 1
135
In the following, we determine the weight distribution of CD for a = 1.
Denote n = |D| = |{x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) + 1 = 0}|. It is clear that
n = | ker(Nqm /qm2 )| · | ker(Trqm2 /q )| =
q m2 −1 (q m − 1)
.
q m2 − 1
For b ∈ F∗qm1 , we denote
N (b) = |{x ∈ F∗qm : Trqm2 /q (Nqm /qm2 (x)) + 1 = 0 and Trqm1 /q (bNqm /qm1 (x)) = 0}|.
By the basic facts of additive characters, we have that
N (b) =
1 X
q2
∗
X
χ(y Trqm1 /q (bNqm /qm1 (x)))χ(z Trqm2 /q (Nqm /qm2 (x)) + z)
1 X
q2
∗
X
χ1 (ybx qm1 −1 )χ2 (zx qm2 −1 )χ(z)
x∈Fqm y,z∈Fq
=
qm −1
qm −1
x∈Fqm y,z∈Fq
=
q
m
q2
+
=
−1
1
q2
+
qm −1
qm −1
1 X X
1 X X
qm1 −1 ) +
qm2 −1 )χ(z)
χ
(ybx
χ
(zx
1
2
q2
q2
∗
∗
∗
∗
x∈Fqm y∈Fq
X
X
χ1 (ybx
x∈Fqm z∈Fq
qm −1
qm1 −1
)χ2 (zx
qm −1
qm2 −1
)χ(z)
y,z∈F∗
x∈F∗
q
qm
qm −1
qm −1
1 X X
1 X X
Ω(b)
qm − 1
qm1 −1 ) +
qm2 −1 )χ(z) +
+
.
χ
(ybx
χ
(zx
1
2
2
2
2
q
q
q
q2
∗
∗
∗
∗
x∈Fqm y∈Fq
x∈Fqm z∈Fq
12
(4.4)
Note that
X X
qm − 1 X X
χ2 (zx)χ(z)
q m2 − 1
∗
∗
qm −1
χ2 (zx qm2 −1 )χ(z) =
x∈F∗
z∈F∗
q
qm
x∈Fqm z∈Fq
X
q −1 X
qm − 1
.
χ(z)
χ
(zx)
=
2
q m2 − 1
q m2 − 1
∗
∗
m
=
z∈Fq
x∈Fqm
From Section 4.1 above, we have
X X
qm −1
χ1 (ybx qm1 −1 ) =
y∈F∗
x∈F∗
q
qm
140
(q − 1)(1 − q m )
.
q m1 − 1
From the discussions above, we obtain that
N (b) =
qm − 1
1
q−1
1
(1 + m2
− m1
) + 2 Ω(b).
2
q
q −1 q −1
q
(4.5)
For any b ∈ F∗qm1 , the weight of a codeword
c(b) = (Trqm1 /q (bNqm /qm1 (d1 )), · · · , Trqm1 /q (bNqm /qm1 (dn )))
equals
wH (c(b)) = n − N (b) =
(q − 1)(q m − 1)(q m1 +m2 − 1)
1
− 2 Ω(b)
m
2
m
2
1
q (q − 1)(q − 1)
q
(4.6)
by Equations (4.4) and (4.5). Hence, by Lemma 3.2, the parameters of CD for e = l = 1 is
[
(q − 1)(q m − 1)q m1 +m2 −2
q m2 −1 (q m − 1)
].
,
m
,
1
q m2 − 1
(q m1 − 1)(q m2 − 1)
Then CD is an optimal one-weight linear code with respect to the Griesmer bound and is not new as mentioned
above. For e = 2 and l = 1, the weight distribution of CD is given in the following.
Theorem 4.5. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m, and gcd( me1 , q − 1) = l = 1, where
145
gcd(m1 , m2 ) = e. Let CD be the linear code defined in Equation (1.1) for a = 1. If e = 2, then CD is a
two-weight linear code with parameters [ q
m2 −1
(qm −1)
, m1 ]
qm2 −1
and its weight enumerator is given by Table II.
Table II. Weight distribution of the code in Theorem 4.5.
Weight
Frequency
0
1
m1 +m2
2
(q−1)(q m −1)(q m1 +m2 +(−1)
q 2 (q m1 −1)(q m2 −1)
m1 +m2
2
(q−1)(q m −1)(q m1 +m2 +(−1)
q 2 (q m1 −1)(q m2 −1)
q
m1 +m2
+1
2
)
+1
q
m1 +m2
2
)
q m1 −1
q+1
q(q m1 −1)
q+1
Proof. For e = 2, the weight distributions of CD can be obtained by Lemma 3.2 and Equation (4.6). Note that
wH (cb ) > 0 for all b ∈ F∗qm1 . Then the dimension equals m1 .
Example 4.6. Let m1 = 2, m2 = 4, m = 4. If q = 2, then CD in Theorem 4.5 is an almost optimal [8, 2, 4]
150
linear code according to the Griesmer bound and has weight enumerator 1 + z 4 + 2z 6 . If q = 3, then CD
in Theorem 4.5 is an nearly optimal [27, 2, 18] linear code, while the corresponding optimal linear codes have
parameters [27, 2, 20].
13
Example 4.7. Let m1 = 4, m2 = 6, m = 12 and q = 2. Then CD in Theorem 4.5 is a [2080, 4, 1040] two-weight
linear code. Its weight enumerator is given by 1 + 5z 1040 + 10z 1144 .
Theorem 4.8. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m, and gcd( me1 , q − 1) = l = 2, where
gcd(m1 , m2 ) = e. Let CD be the linear code defined in Equation (1.1) for a = 1. If e = 1, then CD is a
two-weight linear code with parameters
[
155
m1 +m2 +1
2
(q − 1)(q m − 1)(q m1 +m2 − q
q m2 −1 (q m − 1)
,
m
,
1
m
q 2 −1
q 2 (q m1 − 1)(q m2 − 1)
)
]
and its weight distribution is given in Table III.
Table III. Weight distribution of the code in Theorem 4.8.
Weight
Frequency
0
1
m1 +m2 +1
2
)
(q−1)(q m −1)(q m1 +m2 −q
q 2 (q m1 −1)(q m2 −1)
q m1 −1
2
(q−1)(q m −1)(q m1 +m2 +q
q 2 (q m1 −1)(q m2 −1)
q m1 −1
2
m1 +m2 +1
2
)
Proof. The proof is completed by Lemma 3.3 and Equation (4.6).
Example 4.9. Let m1 = 2, m2 = 3, m = 6 and q = 3. Then CD in Theorem 4.8 is a [252, 2, 168] two-weight
linear code. Its weight enumerator is given by 1 + 4z 168 + 4z 210 . Its dual is a near-MDS code with parameters
[252, 250, 2].
160
4.3. Shortened linear codes of CD
It is observed that the weights of the code in Theorems 4.2 have a common divisor q − 1. This indicates
that the code CD may be punctured into a shorter one.
Assume that a = 0. Note that x ∈ D implies that ux ∈ D for any u ∈ Fq . Hence, the defining set of CD in
Equation (1.1) can be expressed as
D = (F∗q )D1 = {uv : u ∈ F∗q and v ∈ D1 },
165
(4.7)
where di /dj 6∈ F∗q for every pair of distinct elements di , dj in D1 . Then we obtain a shortened linear code CD1
of CD . By Theorem 4.2, we directly obtain the following result.
Corollary 4.10. Let m, m1 , m2 be positive integers such that m1 |m, m2 |m, and m2 > 1. Denote gcd(m1 , m2 ) =
e. Let CD1 be the linear code and its defining set is given in Equation (4.7) for a = 0. If e = 2 and
m
m2 −1
−1)
, m1 ] and its weight
(m1 , m2 ) 6= (2, 2), then CD1 is a two-weight linear code with parameters [ (q (q−1)(q
m2 −1)(q−1)
170
enumerator is given by Table IV.
Table IV. Weight distribution of the code in Corollary 4.10.
Weight
Frequency
0
1
m1 +m2
m2 −m1
2
2
q
(q −1)(q
−1−(−1)
(q−1)q
)
m
m
(q 1 −1)(q 2 −1)
m2 −m1
m1 +m2
+1
−1
2
2
(q−1)q
)
q m1 −1 (q m −1)(q m2 −1 −1−(−1)
(q m1 −1)(q m2 −1)
m1 −1
m
m2 −1
14
q
m1
−1
q+1
q(q m1 −1)
q+1
Example 4.11. Let m1 = 2, m2 = 4, m = 4. If q = 3, then CD1 in Corollary 4.10 is an optimal [13, 2, 9]
two-weight linear code according to the Griesmer bound and has weight enumerator 1 + 6z 9 + 2z 12 . Its dual
has parameters [13, 11, 2] which is optimal according to [15].
5. Applications
175
In this section, we apply our linear codes to construct secret sharing schemes and strongly regular graphs.
We denote by C ⊥ the dual code of a code C.
5.1. Secret sharing schemes from linear codes
Secret sharing schemes were introduced by Shamir and Blakley for the first time in 1979 [3, 25]. Secret
sharing schemes are used in banking systems, cryptographic protocols, electronic voting systems, and the
180
control of nuclear weapons.
It was shown in [2, 27] that any linear code over Fq can be employed to construct secret sharing schemes.
In order to describe the secret sharing scheme of a linear code (see [1, 12]), we need to introduce the covering
problem of linear codes. The support of a vector c = {c0 , c1 , . . . , cn−1 } ∈ Fnq is defined as {0 ≤ i ≤ n − 1 :
ci 6= 0}. A codeword c1 covers a codeword c2 if the support of c1 contains that of c2 . A minimal codeword
185
of a linear code C is a nonzero codeword that does not cover any other nonzero codeword of C. The covering
problem of a linear code is to determine all the minimal codewords of C. From [12, Theorem 12], we know that
secret sharing scheme with interesting access structure can be derived from C ⊥ provided that each nonzero
codeword of a linear code C is minimal.
If the weights of a linear code C are close enough to each other, then all nonzero codewords of C are minimal,
190
as described as follows.
Lemma 5.1. [1] Let wmin and wmax denote the minimum and maximum nonzero Hamming weights of a q-ary
linear code C, respectively. If wmin /wmax >
q−1
q ,
then every nonzero codeword of C is minimal.
For the codes in Theorem 4.2 and Corollary 4.10, we have
m2 −m1
q−1
q m2 −1 − 1 − (q − 1)q 2
wmin
>
=
m2 −m1
−1
m
−1
wmax
q
2
q
− 1 + (q − 1)q 2
if m1 + m2 ≡ 0 (mod 4), and
m2 −m1
q m2 −1 − 1 − (q − 1)q 2 −1
wmin
q−1
=
>
m2 −m1
wmax
q
q m2 −1 − 1 + (q − 1)q 2
if m1 + m2 ≡ 2 (mod 4).
For the code in Theorem 4.5, we have
m1 +m2
q−1
wmin
q m1 +m2 − q 2
>
=
m1 +m2
+1
wmax
q
q m1 +m2 + q 2
if m1 + m2 ≡ 0 (mod 4), and
m1 +m2
q−1
q m1 +m2 − q 2 +1
wmin
>
=
m1 +m2
wmax
q
q m1 +m2 + q 2
15
if m1 + m2 ≡ 2 (mod 4).
For the code in Theorem 4.8, we have
m1 +m2 +1
2
2
q−1
wmin
q m1 +m2 − q
=
m1 +m2 +1 = 1 −
m1 +m2 −1 >
wmax
q
2
2
q m1 +m2 + q
1+q
195
if m1 + m2 > 3.
From the discussions above, the linear codes obtained in this paper can be used to construct secret sharing
schemes with interesting access structures using the framework in [27].
5.2. Strongly regular graphs from linear codes
A connected graph with N vertices is called a strongly regular graph with parameters (N, K, λ, µ) if it is
200
regular of valency K and the number of vertices joined to two given vertices is λ or µ according as the two
given vertices are adjacent or non-adjacent. The theory of strongly regular graphs was introduced by Bose in
1963 for the first time [4].
A code C is said to be projective if the minimum distance of its dual code C ⊥ is at least 3.
The following lemma gives a connection between projective two-weight linear codes and strongly regular
205
graphs.
Lemma 5.2. [7] If C is a projective two-weight [n, k] linear code over Fq with two nonzero weights w1 , w2 ,
then it is equivalent to a strongly regular graph with the following parameters:
N
=
qk ,
K
=
n(q − 1),
λ =
µ
=
K 2 + 3K − q(w1 + w2 ) − Kq(w1 + w2 ) + q 2 w1 w2 ,
q 2 w1 w2
.
qk
Due to Lemma 5.2, new projective two-weight linear codes yield new strongly regular graphs. Examples
in Section 4 show that our codes are not always projective. In particular, we find two classes of projective
210
two-weight codes in the following.
Lemma 5.3. Let m = m1 , m2 = 2 and other notations be the same as those of Theorem 4.5. Then the linear
m
−1)
code CD in Theorem 4.5 is a projective two-weight [ q(qq2 −1
, m] linear code with weight enumerator
m
2
q m − 1 qm −(−1)
q+1
1+
z
q+1
m
q 2
m
m
−1
2 q 2
q(q m − 1) qm +(−1)
q+1
+
.
z
q+1
Proof. The weight enumerator can be directly obtained by Theorem 4.5. We now prove that CD is projective.
⊥
Let Ai , Bi denote the numbers of codewords with Hamming weight i in CD and CD
, respectively. Denote
m
w1 =
m
m
m
q m − (−1) 2 q 2
qm − 1
q(q m − 1)
q m + (−1) 2 q 2 −1
, w2 =
, Aw1 =
, Aw2 =
.
q+1
q+1
q+1
q+1
By the first three Pless Power Moments (see [19]), we have
A + Aw2 = q m − 1,
w1
w1 Aw1 + w2 Aw2 = (n(q − 1) − B1 )q m−1 ,
w2 A + w2 A = (n(q − 1) (n(q − 1) + 1) − B (q + 2(n − 1)(q − 1)) + 2B ) q m−2 .
1
2
2 w2
1 w1
16
Note that n =
q(qm −1)
q2 −1 .
⊥
Solving the above system, we have B1 = B2 = 0. Hence, the minimum distance of CD
is at least 3. The proof is completed.
Lemma 5.4. Let m = m1 , m2 = 2 and other notations be the same as those of Corollary 4.10. Then the
m
−1
linear code CD1 in Corollary 4.10 is a projective two-weight [ qq2 −1
, m] linear code with weight enumerator
m
m
m
2 q 2
2
q(q m − 1) qm−1 −(−1)
q m − 1 qm−1 +(−1)
q+1
q+1
z
z
+
1+
q+1
q+1
m −1
q 2
.
Proof. The weight enumerator can be directly obtained by Corollary 4.10. We now prove that CD1 is projective.
⊥
Let Ai , Bi denote the numbers of codewords with Hamming weight i in CD1 and CD
, respectively. Denote
1
m
m
m
m
q m−1 + (−1) 2 q 2
qm − 1
q(q m − 1)
q m−1 − (−1) 2 q 2 −1
w1 =
, w2 =
, Aw1 =
, Aw2 =
.
q+1
q+1
q+1
q+1
By the first three Pless Power Moments (see [19]), we have
A + Aw2 = q m − 1,
w1
w1 Aw1 + w2 Aw2 = (n(q − 1) − B1 )q m−1 ,
w2 A + w2 A = (n(q − 1) (n(q − 1) + 1) − B (q + 2(n − 1)(q − 1)) + 2B ) q m−2 .
1
2
2 w2
1 w1
215
Note that n =
qm −1
q2 −1 .
⊥
Solving the above system, we have B1 = B2 = 0. Hence, the minimum distance of CD
1
is at least 3. The proof is completed.
Lemmas 5.2 and 5.3 yield the following theorem.
Theorem 5.5. Let gcd( m
2 , q − 1) = 1 and 2|m. Then there exists a strongly regular graph with the following
parameters:
= qm ,
q(q m − 1)
K =
,
q+1
m
m
q m+2 − 2q 2 − 3q − (−1) 2 q 2 (1 − q)
,
λ =
(q + 1)2
m
m
m
m
q(q 2 − (−1) 2 )(q 2 +1 + (−1) 2 )
µ =
.
(q + 1)2
N
220
The following theorem can be directly obtained by Lemmas 5.2 and 5.4.
Theorem 5.6. Let 2|m and m ≥ 4. Then there exists a strongly regular graph with the following parameters:
N
=
K
=
λ =
µ
=
qm ,
qm − 1
,
q+1
m
m
q m − 3q − 2 − (−1) 2 q 2 +1 (q − 1)
,
(q + 1)2
q(q
m
2
m
m
m
− (−1) 2 )(q 2 −1 + (−1) 2 )
.
(q + 1)2
We remark that the parameters of the strongly regular graphs in Theorems 5.5 and 5.6 are probably new
after comparing with known ones in the literature.
17
6. Concluding remarks
225
In this paper, we presented a construction of q-ary linear codes and determined the weight distributions in
some cases based on Gauss sums. Four classes of two-weight linear codes were obtained. Note that these linear
codes have very flexible parameters and are probably new after comparing with known two-weight linear codes
in the literature (see [7, 9, 12, 13, 17, 29] for some known two-weight linear codes). It is interesting that our
construction can produce optimal or almost optimal codes. What’s more, our codes can be used to construct
230
secret sharing schemes with interesting access structures and strongly regular graphs.
Acknowledgments
The authors are very grateful to the reviewers and the Editor for their valuable comments that improved the
quality of this paper. Special thanks go to one of the reviewers for pointing out some knowledge of one-weight
linear codes.
235
References
References
[1] A. Ashikhmin, A. Barg, Minimal vectors in linear codes, IEEE Trans. Inf. Theory 44 (5) (1998) 2010-2017.
[2] R. Anderson, C. Ding, T. Helleseth, and T. Kløve, How to build robust shared control systems, Des.
Codes Cryptogr. 15 (2) (1998) 111-124.
240
[3] G. R. Blakley, Safeguarding cryptographic keys, Proc. Nat. Comput. Conf. 48 (1979) 313-317.
[4] R. C. Bose, Strongly regular graphs, partial geometries and partially balanced designs, Pacific J. Math.
13 (1963) 389-419.
[5] L. D. Baumert and R. J. McEliece, Weights of irreducible cyclic codes, Inf. Contr. 20 (2) (1972) 158-175.
[6] B. C. Berndt, R. J. Evans, K. S. Williams, Gauss and Jacobi Sums, J. Wiley and Sons Company, New
245
York, 1997.
[7] A. Calderbank, W. Kantor, The geometry of two-weight codes, Bull. London Math. Soc. 18 (1986) 97-122.
[8] C. Carlet, C. Ding, and J. Yuan, Linear codes from perfect nonlinear mappings and their secret sharing
schemes, IEEE Trans. Inf. Theory 51 (6) (2005) 2089-2102.
[9] C. Ding, Linear codes from some 2-designs, IEEE Trans. Inf. Theory 61 (6) (2015) 3265-3275.
250
[10] C. Ding, X. Wang, A coding theory construction of new systematic authentication codes, Theoretical
Computer Science 330 (1) (2005) 81-99.
[11] C. Ding, C. Li, N. Li, and Z. Zhou, Three-weight cyclic codes and their weight distributions, Discrete
Math. 339 (2) (2016) 415-427.
18
[12] K. Ding, C. Ding, A class of two-weight and three-weight codes and their applications in secret sharing,
255
IEEE Trans. Inf. Theory 61 (11) (2015) 5835-5842.
[13] F. De Clerk, M. Delanote, Two-weight codes, partial geometries and Steiner systems, Des. Codes Cryptogr.
21 (2000) 87-98.
[14] P. Delsarte, Weights of linear codes and strongly regular normed spaces, Discrete Math. 3 (1972) 47-64.
[15] M. Grassl, Bounds on the parameters of various types of codes. Avaliable at http://www.codetables.de/.
260
[16] Z. Heng, Q. Yue, A class of binary linear codes with at most three weights, IEEE Commun. Letters 19
(9) (2015) 1488-1491.
[17] Z. Heng, Q. Yue, Two classes of two-weight linear codes, Finite Fields Appli. 38 (2016) 72-92.
[18] Z. Heng, Q. Yue, Evaluation of the Hamming weights of a class of linear codes based on Gauss sums, Des.
Codes Cryptogr. 83 (2) (2017) 307-326.
265
[19] W. C. Huffman, V. Pless, Fundamentals of Error-Correcting Codes. Cambridge, Cambridge Univ. Press,
2003.
[20] T. Kløve, Codes for Error Detection, Singapore, World Scientific, 2007.
[21] C. Li, Q. Yue, A class of cyclic codes from two distinct finite fields, Finite Fields Appli. 34 (2015) 305-316.
[22] C. Li, Q. Yue, and F. Li, Hamming weights of the duals of cyclic codes with two zeros, IEEE Trans.
270
Inform. Theory 56 (6) (2014) 2568-2570.
[23] R. Lidl, H. Niederreiter, Finite Fields, Cambridge Univ. Press, Cambridge, 1984.
[24] F. J. MacWilliams and N. J. A. Sloane, The Theory of Error Correcting Codes. II, Amsterdam, The
Netherlands: North-Holland, 1977, vol.16.
[25] A. Shamir, How to share a secret, Commun. Assoc. Comp. Mach. 22 (11) (1979) 612-613.
275
[26] G. Xu, X. Cao, S. Xu, Two classes of p-ary bent functions and linear codes with three or four weights,
Cryptogr. Commun. 9 (2017) 117-131.
[27] J. Yuan, C. Ding, Secret sharing schemes from three classes of linear codes, IEEE Trans. Inf. Theory 52
(1) (2016) 206-212.
[28] S. Yang, Z. Yao, Complete weight enumerators of a family of three-weight linear codes, Des. Codes
280
Cryptogr. 82 (2017) 663-674.
[29] Z. Zhou, N. Li, C. Fan, et al., Linear codes with two or three weights from quadratic bent functions, Des.
Codes Cryptogr. 81 (2016) 283-295.
[30] Z. Zhou, C, Ding, A class of three-weight cyclic codes, Finite Fields Appli. 25 (2014) 79-93.
19
| 7cs.IT
|
An efficient counting method for the colored triad census
Jeffrey Lienerta,b,c , Laura Koehlya , Felix Reed-Tsochasb , Christopher Steven
Marcuma,c
arXiv:1802.01481v1 [cs.DS] 5 Feb 2018
a
National Human Genome Research Institute, National Institutes of Health
b
Saı̈d Business School, University of Oxford
c
Corresponding Author
Abstract
The triad census is an important approach to understand local structure in network
science, providing comprehensive assessments of the observed relational configurations between triples of actors in a network. However, researchers are often interested
in combinations of relational and categorical nodal attributes. In this case, it is desirable to account for the label, or color, of the nodes in the triad census. In this paper,
we describe an efficient algorithm for constructing the colored triad census, based, in
part, on existing methods for the classic triad census. We evaluate the performance
of the algorithm using empirical and simulated data for both undirected and directed
graphs. The results of the simulation demonstrate that the proposed algorithm reduces computational time by approximately 17,400% over the naı̈ve approach. We
also apply the colored triad census to the Zachary karate club network dataset. We
simultaneously show the efficiency of the algorithm, and a way to conduct a statistical test on the census by forming a null distribution from 1, 000 realizations of a
mixing-matrix conditioned graph and comparing the observed colored triad counts
to the expected. From this, we demonstrate the method’s utility in our discussion
of results about homophily, heterophily, and bridging, simultaneously gained via the
colored triad census. In sum, the proposed algorithm for the colored triad census
brings novel utility to social network analysis in an efficient package.
Keywords: triad census, labeled graphs, simulation
1. Introduction
The triad census is an important approach towards understanding local network
structure. ? ] first presented the 16 isomorphism classes of structurally unique triads
Preprint submitted to XXX
February 6, 2018
possible in a directed network. To conduct a triad census, one simply counts each
occurrence of these structures, without respect to the labeling of the nodes (here we
use node label, color, characteristic, and attribute interchangeably). This is useful
insofar as specific triads, or combinations thereof, may relate to underlying social
processes giving rise to an observed network. For example, bridges (triads with one
null dyad and two non-null dyads) may be important in navigating social networks
[? ], and certain triads may be more or less favorable based on structural balance
theory (e.g. the 300 is balanced but the 201 is not, see Figure 1) [? ]. Moreover, a
variant of the triad census, motif analysis, investigates the statistics of various triad
configurations (motifs), and has found wide application in biology [? ].
Also important to network structure are nodal characteristics and how they relate
to tie formation or dissolution. This has been the subject of research on homophily
(individuals having similar attributes with those to whom they are connected) [? ].
However, homophily is an observed phenomenon, not a process. The processes giving
rise to homophily are varied, often confound the relationship between networks and
outcomes, and are difficult to tease apart [? ]. Methodological advances, such as
stochastic actor-oriented models can disentangle these effects to some some extent [?
]. Other analyses have attempted to disentangle the processes leading to homophily
from structural processes, such as triadic closure [? ]. Additionally, the coloring of
nodes in a network has been an important question for many graph theorists and
indeed represents a major topic in this field [? ].
Although nodal characteristics and the triad census are important, they have
rarely been examined fully in conjunction. Yet, there are a few cases where specific
2
colored triads have been studied. For example, ? ] study brokerage based on triad
structure and group membership simultaneously. This same approach has been used
to study brokerage in dynamic networks [? ]. As well, a study by ? ] examined specific colored triads based on generational membership within families; in this work
the authors showed that inter-generational ties were observed in different quantities
than expected based on the underlying null model. None of the past research evaluated the full census of colored triads, rather, researchers have focused instead on
specific colored triads that were a priori expected to be relevant to the processes at
hand. As a result, these foundational works were not exhaustive with respect to all
alternatives. In other words, previous research examining a subset of colored triads
likely had an amount of false negatives due to not examining every colored triad;
this could be addressed by censusing the colored triads.
The examination of node characteristics together with local structure is important as it provides opportunity to simultaneously study the occurrence of triadic
structure, nodal attributes, and the interactions between them. For instance, certain
colored triads may be impermissible, such as three-cycles between strict heterosexuals in mixed-orientation sexual contact networks [? ]. Impermissible triads would
be categorized the same as those that were not observed due to chance in a triad
census, potentially missing important social processes or constraints at play in this
type of network. Only by incorporating node coloring into the triad census can this
pattern be fully elucidated.
Based on this methodological gap in the literature, we develop a method to census
the colored triads for any one-mode binary network with arbitrary number of colors.
3
Due to the large numbers of unique isomorphism classes as the number of colors
increases, this method requires computational efficiency in addition to mathematical
accuracy. As well, one is often interested in forming a null distribution with which to
compare observed colored triad counts. If the null distribution cannot be analytically
solved, one would likely census the colored triads of many simulated networks, further
increasing the need for the algorithm to be computationally efficient.
Current efficient methods for the triad census exploit the sparseness of networks
[? ], and scale sub-quadratically (as the number of edges increases the time to run
the algorithm is faster than the number of edges squared). However, methods that
exploit network sparseness by inferring the number of null triads do not work in
the colored case because they do not explicitly interrogate every triad, and there
are variations within the null triads due to the coloring. Therefore, we extend the
methodology of ? ], which is based on matrix algebra and interrogates every triad;
his method scales sub-quadratically with the number of nodes.
This paper (1) presents the colored triad census and its computational complexity,
(2) shows that this approach can be used on large networks (tested for up to 10, 000
nodes) with up to 10 colors in relatively efficient time, and (3) uses the method many
times to create null distributions of colored triad censuses to form the basis of conditional uniform graph tests. We illustrate the benefits of an analysis incorporating
the colored triad census using a well-known dataset, Zachary’s Karate Club [? ].
4
2. Algorithm
Since the original appearance of the triad census in 1976, a number of papers
have explored how to compute the triad census of a network in an efficient manner.
Although sub-quadratic methods exist for calculating the triad census (e.g. ? ]), we
use the quadratic algorithm presented by ? ] here. This is because the more efficient
methods avoid interrogating null triads directly by taking advantage of the sparseness
of graphs, the subsequent large number of null (003) triads, and the known number
of total triads. Instead, they interrogate all triads with at least one edge, and then
subtract that count from the total number of triads in the network to arrive at the
number of null triads. This is insufficient in the colored triad census as there are
differently-colored null triads, and their number cannot be algebraically determined.
Moody’s algorithm does not employ this limiting shortcut, and we therefore use it as
a basis for our colored triad census algorithm. Additionally, because many networks
are sparse, we can leverage computational techniques for increasing the efficiency of
sparse matrix operations [? ], further reducing the computational complexity of our
method.
? ] showed that the count of each of the 16 triad isomorphism classes could
be derived by using matrix algebra on the adjacency matrix of the graph and its
derivatives. To review, let A be the adjacency matrix of a network, and Aij = 1
when a tie exists from node i to node j. Let E be the symmetrized matrix A, formed
by making any edge in A reciprocal via Eij = max(Aij , Aji ). The complement of
E, Ē, is formed by subtracting the complete network adjacency matrix from E,
so that Eij = 1 if and only if there is neither a tie from i to j nor a tie from j
5
to i. Next, we have M , the mutual matrix of A, and is made by removing any
asymmetric edges from A, or Mij = Mji = 1 ⇐⇒ Aij = Aji = 1. Finally, C is
the matrix of only asymmetric edges, and is calculated by C = A − M . Therefore,
Cij = 1 ⇐⇒ Aij = 1 & Aji = 0. Based on these matrices, Moody demonstrates how
to calculate the number of each of the 16 isomorphism classes for the case of unlabeled
graphs (or, equivalently, for a graph consisting of nodes of the same single color).
Generally, this was done by multiplying (either through dot-product or element-wise
multiplication) the three matrices corresponding to the relevant edges in the triad of
interest. There were two triads (111U and 111D) that were not directly amenable to
this process and were calculated via addition and subtraction of other triad types,
respectively.
To extend this work to the case of multiple colors, we introduce the out-coloring
0
and in-coloring matrices, K r and K r , respectively, where r is the focal color of
matrix K. Here, the in-coloring matrix is the transpose of the out-coloring matrix.
The out-coloring matrix is calculated by evaluating the color of the nodes row-wise,
such that rows indexing nodes of the focal color are composed in the following way:
r
Ki•
=
1 if R(i) = r
(1)
0 if R(i) 6= r
Where R(i) is a function returning the color of node i. As above, the in-coloring
matrix is the transpose of the out-coloring matrix in Eq. 1.
Our algorithm works by using the in- and out-coloring matrices to evaluate and
“switch on” edges that have nodes of the focal colors at the ends (or tails) of edges
6
in the adjacency matrix A of the network. We adapt the triad census nomenclature
of ? ] by appending the colors after the name of the triad. The colors are ordered
from the top node proceeding clockwise in Figure 1. We have arbitrarily adapted
the orientation of the triads from the triad census figure in ? ] for computational
reasons. The orientation is important here because triads with the same orientation
may no longer be isomorphic when color is introduced. Figure 1 makes it possible
to count unambiguously and name only unique colored triads. Therefore, T102−123
is the triad consisting of 1 symmetric dyad and 2 null dyads, where the top node is
of color 1, the bottom-right node is of color 2 and the bottom-left node is of color
3. This is distinct from the T102−312 triad because the coloring of the nodes is not
identical from the previous triad.
Following this, the general formula for an arbitrary triad “T” with an arbitrary
coloring triplet is:
0
0
0
T = T r (K 1 × H(T, 1, 2) × K 2 )(K 2 × H(T, 2, 3) × K 3 )(K 3 × H(T, 3, 1) × K 1 ) (2)
In the above, “×” refers to element-wise multiplication, and “Tr” is the trace
function. For an arbitrary triad, “T” has a color triplet r1 , r2 , r3 . H(T, i, j) is a
function returning the matrix specific to the type of edge between nodes i and j
in triad “T ”. For example, in a 102 triad, the first edge from the top node going
clockwise is a symmetric edge from node one to node two (Figure 1). H(T102 , 1, 2)
in this case would be the matrix E for the symmetric matrix, and the sandwiching
color matrices would turn the proper edges on and off if nodes one and two were of
7
the specified colors. If the edge is an asymmetric one, and the direction of the edge
in Figure 1 is counter-clockwise, then C 0 is used instead of C to force the edge to go
in the proper direction.
At this point, there are redundant triads due to certain colored triads being
isomorphic. For instance, the T003−122 is isomorphic with T003−221 and T003−212 , and
would be triple-counted. These are removed by checking for isomorphisms based
on matrix row and column permutations of the triad. If two colored matrices are
identical after such row and/or column permutations, then they are isomorphic, and
one is removed. We arbitrarily decide to discard the triad whose coloring triplet
name comes second alphanumerically. It should be noted that removing in this way
is computationally expensive, particularly as the number of colors and nodes grows
large. We therefore shorten this process by performing it once for 1 to 10 colors and
storing the unique isomorphism classes. This leaves only unique isomorphism classes
of colored triads, which can then be accessed in linear time.
The number of unique isomophism classes for a given number of colors can be
shown for each of the 16 ismorphism classes in the triad census. The 16 classes
separate into four types of colored triads, depending on how many structurallydistinct positions there are in the triad (e.g. the two ends of the edge in a 102 triad
are not structurally-distinct from one another, but are distinct from the node with
no edges). The calculation for the number of each isomorphism class for arbitrary
number of colors (k) is shown in Table 1. Each combinatoric term in each row
(together with their respective leading permutation coefficients) counts the number
of colored triads when there are three, two, or one unique color(s), respectively. For
8
example, in a network with three colors,the ‘300’ and ‘003’ classes have only one
accessible permutation when there are three colors present in the triad (i.e. 33 ), six
ways when there are two colors (i.e. 2 32 ), and one way when there is one color in
the triad (i.e. 31 ).
Isomorphism classes
Number of colored
triads
k
k
k
300 and 003
+ 2 2 + 1
3
k
030C
2 3 + 2 k2 + k1
102 021D 021U 201 120D and 120U
3 k3 + 4 k2 + k1
012 021C 111D 111U 030T 120C and 210
6 k3 + 6 k2 + k1
Table 1: Expression for the number of isomorphism classes within a triad class. k is the number of
colors
If these numbers are summed over the 16 isomorphism classes, the total number
of colored isomorphism classes of triads for k colors is returned. Similarly, the same
can be done for undirected triads, solely summing over the 4 triads observed in the
undirected case. Table 2 reports the total number of colored triads for undirected
and directed networks over a range of k. Clearly, the number of isomorphism classes
grows quite quickly as k increases.
The algorithm implemented as an R package is publicly available and is linked
to this paper via github: https://github.com/jlienert/ColoredTriadCensus.
3. Algorithmic Performance
Theoretically, if basic matrix multiplication is used, this algorithm runs with
computational complexity O(N 2 ∗ 3k ). It scales with the number of nodes squared
(N 2 ) because of the matrix multiplication involved in the algorithm. The scaling
with the number of colors cubed comes from the number of distinct colored triads
9
Number of
colors
Number of directed
colored triads
Number of undirected
colored triads
1
2
3
4
5
6
7
8
9
10
16
104
328
752
1440
2456
3864
5728
8112
11080
4
20
56
120
220
364
560
816
1140
1540
Table 2: The number of colored triad isomorphism classes for directed and undirected networks for
k ranging from 1 to 10.
the algorithm needs to evaluate. By taking advantage of methods for matrix multiplication using sparse matrices (as appropriate due to the sparse nature of most
social networks), this complexity is reduced to something closer to O(N log(N ) ∗ 3k ).
To test the efficiency of the algorithm, we apply it to networks ranging in size
from n = 10 to n = 10, 000 with the number of colors ranging from k = 1 to k = 10,
all holding the average density constant at 0.01 by creating Erdös-Rényi graphs with
those parameters. The runtime of the algorithm with these parameters can be seen
in Figure 2. In general, increasing K results in constant increases in log(runtime),
which is what we expect based on the theoretical computational complexity. As
expected, we also observe a super linear increase in log(runtime) as N increases.
Although it is super linear, it is still well below the linear curve that would exist if
we used matrix multiplication not optimized for sparse matrices. Finally, we observe
changes in the rank-order and decreases in runtime going from 10 to 100 nodes. This
is also due to the computational time involved in initializing the sparse matrices and
10
storing and operating on sparse matrices, and as such is not unexpected. To be perfectly optimized, therefore, the algorithm would use standard matrix multiplication
for small networks, and switch to sparse methods for larger networks. However, the
gains would be minimal, generally under 10 seconds, and would require additional
logical steps to check for network size, further minimizing the gain. We therefore use
sparse matrix methods for all network sizes.
4. Empirical Use and Example
To show the empirical value of this algorithm, we use the Zachary karate club
social network [? ]. This is a well-known historical network that describes the social
relationships between 34 members of a university karate club. Ties exist between
members if they overlapped in at least one of eight contexts representing undirected
relations. These relations varied in terms of likely strength of the association. Likely
at the weak end of the spectrum is being enrolled in the same class at the university,
while likely at the strong end is being a student-teacher at the studio. Additionally,
three ties are specific to activities with a part-time instructor. Member “factions”
were identified as a node attribute, taking one of five mutually exclusive values:
strongly associated with the president, weakly associated with the president, neutral,
weakly associated with the part-time instructor, or strongly associated with the parttime instructor. These are labeled ”Zs”, ”Zw”, ”N”, ”Hw”, and ”Hs”, respectively.
These labels can be placed on an ordinal scale from -2 (Zs) to 2 (Hs) to quantify
members’ direction and strength of alignment. This undirected network with five
colors represents a case that is rich in the number of colored triads (220) for detailed
11
conclusions to be drawn using the proposed algorithm (which is general to both
undirected and directed networks).
We initially ran the colored triad census on the social network using the faction
as the nodal attribute. This gave our empirical observed colored triad census. To
determine whether these triads were observed more or less often than expected by
chance, we construct a null model. As the choice of null model can have important
ramifications for the null distribution of triads, we chose a model where edge formation is a function of the probability of ties between nodes of specific attributes [? ].
The null model is a mixing-matrix conditioned uniform random graph distribution
based on probabilities of edges between nodes of particular color combinations [?
]. This matrix comprises empirical probabilities of ties between groups, with the
diagonal representing within-group tie probabilities. This means that significantly
over- or under-represented colored triads are observed as such due to network effects
beyond homophily and heterophily. Networks are then generated from this matrix
via a Bernoulli random graph process á la ? ]. Here, then, this null model therefore
conditions on graph size, the distribution of node factions, and the probability of
ties within and between factions. By generating networks from the null model, we
can observe whether colored triad counts deviate from that expected based on the
marginal distribution of faction mixing. Because we condition on the above parameters, if we observe statistical deviations in our colored triad census, it indicates that
the structure of the network is dependent on parameters other than those on which
we conditioned.
Moreover, for any triad, the expected number and variance can be calculated
12
assuming each tie follows a Binomial distribution (which is a reasonable assumption
for most binary social network data). The observed number can then be compared
to these numerical results and a p-value extracted from an exact Binomial test. This
equates to the following probability, expectation, and variance for an example colored
triad:
P (T ) =P (Aij = 1|R(i) = r1 , R(j) = r2 ) × P (Aij = 1|R(i) = r2 , R(j) = r3 )
(3)
×P (Aij = 1|R(i) = r3 , R(j) = r1 )
L(T ) P
E(T ) = P (T ) ×
Y
r=1
r
K•1
S(T, r)
V (T ) = E(T ) × (1 − P (T ))
(4)
(5)
The probability of T , P (T ) in Equation 3 is based on the mixing-matrix of the
three colors (r) involved in the triad T . As is standard for the mixing-matrix approach, this continues to assume that all edges in the graph are independent. For
the expected value of a specific triad, we multiply the probability of a single one
of those triads by the total number of colored triplets that exist in the graph. In
Equation 4, the expectation of the triad, L(T ) returns the number of unique colors in
P r
T and
K•1 is the number of nodes of color r in the graph. Also, we take the nodes
one, two, or three at a time depending on how many times that color repeats in T ,
represented by S(T, r). This expectation therefore follows a binomial distribution,
and it’s variance follows accordingly in Equation 5.
13
However, to show that this method also works for null distributions that are not
analytically solvable, we construct a null distribution based on simulated draws from
the null model. As the number of trials increases, the simulated null distribution
of the colored triad census should asymptotically approach the analytical solution
shown above. For each of 1, 000 trials, we draw random networks from the null
distribution, and run the triad census on all these networks. Comparing our observed
count to the null distribution then allows us to get an approximate p-value for a
conditional uniform graph test, and test the over- or under-representation of each
colored triad. We now turn to these results.
4.1. Results
Figure 3 is a heatmap of the approximate p-values associated with each binomial
exact test against the null for each triad, clustered by the triad and the colored
triplet as returned by the proposed algorithm. We use a clustering algorithm to
group color triplets with similar profiles across the types of triads. This assists
with identifying trends across different colored triads, leading to conclusions that
would likely be missed if all the colored triads were individually examined. We find
particular importance in three branch cutpoints in the clustering algorithm on the
color triplets. The first branch in the clustering algorithm (A in Figure 3) separates
four color triplets, comprising 16 colored triads, with a pattern of over-observed 003
and 102 triads, and under-observed 201 and 300 triads. These results show that these
color triplets are those that are less clustered than expected by chance. The color
triplets all contain nodes of two factions with the first two nodes being Hs, that is,
those strongly aligned with the part-time instructor. This indicates that those who
14
are so aligned are likely to form ties to one another, but not to members of other
factions. The only exception in this group is that two Hs nodes are more likely to
form a tie from one of the Hs members to a Hw member, but even in this case the
complete triad (003) is still observed less than expected by chance. This particular
result is, perhaps, unsurprising, since Hs and Hw members are close in alignment,
more so than with those aligning with the president. Therefore, given the tendency
towards homophily they are likely to overlap, though less strongly than members of
the same faction; hence, the under-observed T003−HsHsHw .
15
003
012
102
021D
021U
021C
111D
111U
030T
030C
201
120D
120U
120C
210
300
Figure 1: The 16 isomorphism classes of triads and their orientation used here with respect to the
color numbering. When colors are added to these triads, they are labeled starting from the top
node and proceeding clockwise.
16
Algorithmic runtime
3
2
0
1
Log(running time)
4
5
3 Colors
4 Colors
5 Colors
6 Colors
7 Colors
8 Colors
9 Colors
10 Colors
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Log(# nodes)
Figure 2: Runtime of the algorithm on networks ranging from size 10 to 10, 000 nodes in orders
of magnitude, and from one to ten colors. These runtimes were generated using a PC running
Windows 7, with an Intel i7 2.80 GHz Chip and 16GB of RAM.
17
18
80
Count
40
A
0 0.2
B
Value
0.6
1
C
Color Key
and Histogram
003
102
201
300
H sZsZs
ZsN N
ZsZsZs
ZsN Zs
ZsH w Zs
ZsZw Zs
ZsN H w
H sH sH s
N H sH w
H sZw N
H sZw H w
H sN N
H w ZsZs
ZsZw H w
H sZw Zw
ZsZw N
H sN H w
N ZsZs
H sH w Zs
ZsH sN
ZsZw Zw
N H sH s
ZsH sH s
ZsH sH w
Zw ZsZs
N H sZs
H w H sH s
H w H sZs
ZsH sZw
ZsH sZs
Zw H sZs
H sN Zs
Zw H sH s
H sH sN
H sH sZs
H sH sZw
H sH sH w
H w H sH w
Zw H sZw
N H sN
H sH w H w
H sZw Zs
ZsH w H w
ZwZwN
ZwZwZw
ZwZwHw
ZwNN
ZwNHw
ZwHwHw
NNN
NNHw
NHwHw
HwHwHw
NZwZw
NZwN
NZwHw
NZwZs
HwZwZw
HwZwN
HwZwHw
HwZwZs
HwNN
HwNHw
HwNZs
ZwNZs
ZwZwZs
ZwHwZs
NNZs
NHwZs
HwHwZs
Zw H sH w
Zw H sN
N H sZw
H w H sZw
H w H sN
Figure 3: Heatmap of colored triads and their corresponding p-value of how often they were observed in the empirical networks relative to the null distribution.
The columns separate triads based on the MAN configuration, and the rows separate triads based on the triplet of colors. Standard clustering algorithms were used
to create the dendrograms. White space indicates redundant isomorphism classes. Gray boxes are either those with 0 triads observed in the network or in any of
the networks of the null distribution, and therefore have an undefined pseudo p-value, or those with a pseudo p-value of 0.5. The three labels correspond to three
breakpoints in the clustering that separate meaningful groups. (A) is a group of four color triplets exhibiting homophily between Hs nodes. (B) is a group of 21
colored triplets exhibiting low clustering between heterogeneous nodes. (C) is a group of 6 colored triplets that show potential significant amounts of bridging.
0
The second branching point in the clustering (B in Figure 3) separates the group
of color triplets that are over-observed for the 003 triad, under-observed for the 102
and 201 triads, and observed about as much as expected for the 300 triads. All the
triplets in question have nodes of different factions in the first and second position.
Because the edge in the 102 triad is between the first and second node in the triplet
(Figure 1), this means that these are all triplets where the first edge is less likely
than expected by chance, and the lack of formation of the first edge subsequently
hampers the formation of the edge between the second and third nodes in the triplet
(201 triad). The first two nodes of these triplets are often (e.g., 16 out of 21) two
factions at least a distance of two away (e.g. N and Hs), indicating members of
a faction are not likely to overlap with members who are too disparate from their
faction. Put another way, this pattern of triads shows a lack of faction heterophily.
The third branch point (unlabeled) is primarily singling out the group of color
triplets that were not observed in the network, and we cannot draw conclusions about
their prevalence. The fourth branch point (C in Figure 3), however, distinguishes a
group of five triplets that are under-observed for the 102 triad and over-observed for
the 201 triad. This means that the edge between the first two nodes is less likely
than expected by chance, but once that edge does occur, the second edge occurs
more often than expected by chance. All these triplets begin with a Zs member,
and the 201 triad in this case is effectively a bridging tie between it and another.
Interestingly, the bridging node is anything other than an Hs (whom are primarily
consigned to this role in branch A, as discussed above). The third node was another
Zs member in four of five triplets. This indicates that Zs members of the karate
19
club did not often overlap members of other factions, but when they did, provided
it was not with an Hs, that second person also often overlapped with another Zs.
Although the above examples show homophily and bridging, analyzing the full
colored triad census allows us to draw further conclusions by looking at other colored
triads. In particular, the homophily has mostly been a story of the Hs nodes, and the
bridging primarily about the Zs nodes. The 300 triad of both of these factions, when
comprising three nodes of the same faction, are observed more often than expected
by chance in both cases, which has different implications on the previously-noted
results. For the Hs nodes, homophily is strengthened, as not only do Hs nodes not
often overlap with members of other factions, they also very strongly overlap with
one another. This may partially be an artifact of the types of overlap, as stated
before, three of the overlap activities involve direct participation in the part-time
instructor’s studio, but there are no corresponding groups for the president. This
means that those who are Hs or Hw may have more opportunity to overlap with
one another due solely to the structure of the data. On the other hand, the triplet
of all Zs members also has an over-observed 300 triad. Although there are other
triads that seem to indicate bridging between Zs members (C in Figure 3), given
that Zs members are also densely connected to one-another, the practical effect of
these potential bridging ties is reduced. Observing this joint effect of homophily and
bridging ties was possible only through the complete colored triad census. Neither
a standard triad census nor a brokerage analysis would have revealed the intricacies
of these results.
In sum, it is clear from these results that the colored triad census allows one
20
to examine multiple trends simultaneously that are often done in isolated analyses,
including homophily, heterophily, and brokerage. Importantly, it also allows for
generalizations based on the clustering of various triads or color triplets, as well as
specific results based on individual triads. In this manner, the colored triad census
can yield results on multiple structural levels simultaneously, all while examining
local structure, nodal attributes, and their interaction—that is, net of all alternatives
involving mixtures of node coloring and triadic configurations.
5. Limitations
There are some limitations to this method. First, it is only computationally
efficient relative to existing methods (including brute force counting). Networks of
10, 000 nodes or more will take over a day to run using the proposed algorithm for the
colored triad census. However, this is an easily paralellizable process (by partitioning
the separate algebraic steps, for example), and so the real time necessary to run the
analysis can be greatly reduced by taking advantage of this feature. The time needed
for the parallelized colored triad census is approximately inversely proportional to
the number of computational cores used in the calculation (plus some overhead).
Second, the complete explication of all colored triads has both benefits and potential
pitfalls. Examining all of the triads simultaneously eliminates the possibility of
missing interesting results because a specific colored triad was excluded. However,
the sheer number of colored triads means that making complete sense of results can
be difficult, due to information overload. Even if the results are carefully examined
for all colored triads, it is conceivable that one might miss an important result out of
21
the 11, 080 colored triads in a directed, 10-color network, no matter how meticulous
the examiner’s eye. However, use of standard clustering algorithms and heatmaps
(as here) may help to ease interpretation of the results at both a coarse (general
groups of triads) or fine-grained (individual colored triads) perspective.
6. Conclusions
In this paper, we have extended the matrix algebra methods of ? ] to calculate
the colored triad census for any network, directed or undirected, with an arbitrary
number of colors in a relatively computationally efficient manner. We have shown a
number of mathematical results regarding the colored triad census, including a generalized equation for an arbitrary colored triad, the number of isomorphism classes
for arbitrary numbers of colors, and the expectation and variances for colored triads.
We analyzed an empirical social network using our algorithm, and calculated approximate p-values for each colored triad, based on an analytic exact binomial test
for less complex null distributions, or approximately through simulation for more
complex null distributions. We have also shown the type of conclusions that can be
drawn from these results, observing results that would not be feasible with many
other currently available methods.
One additional benefit of this method is that it can be directly used as a counting
tool for sufficient statistics in network inference models, such as exponential random
graphs (ERGM). The colored triad census essentially allows one to simultaneously
evaluate the effect of local structure and node attribute on network structure in an
ERGM, building off previous work where researchers explicated the ERGMs capacity
22
for including the triad census [? ]. We believe that the colored triad census is a useful
technique with an efficient implementation that can be widely-applicable in social
networks research, showing the continued importance of the triad census even in this
era of stochastic models for complex networks.
7. Acknowledgements
8. References
23
Appendix A. Variable and Functional Definitions
Variable or function notation
A
E
Ē
M
C
Kr
R(i)
H(T, i, j)
T
L(T )
S(T, r)
P (T )
E(T )
V (T )
Description of variable or function
Adjacency matrix
Symmetrized adjacency matrix
Complement of symmetrized adjacency matrix
Adjacency matrix including only mutual ties
Adjacency matrix including only asymmetric ties
Coloring matrix for color r
Function returning the color of node i
Function returning the matrix of the edge in triad T
between nodes i and j
An arbitrary colored triad, with a MAN configuration,
and colored triplet r1 , r2 , r3
A function returning the number of unique colors for
a given colored triad
Function returning the number of times color r
appears in colored triad T
The probability of observing triad T
The expectation of triad T under a binomial model
The variance of triad T under a binomial model
Table A.3: List of variables, constants, and functions defined in this manuscript.
24
| 8cs.DS
|
Compressive Sampling of Ensembles of Correlated Signals
Ali Ahmed and Justin Romberg∗
arXiv:1308.5146v3 [cs.IT] 15 Dec 2017
DRAFT: 2:46am, December 18, 2017
Abstract
We propose several sampling architectures for the efficient acquisition of an ensemble of correlated
signals. We show that without prior knowledge of the correlation structure, each of our architectures
(under different sets of assumptions) can acquire the ensemble at a sub-Nyquist rate. Prior to sampling,
the analog signals are diversified using simple, implementable components. The diversification is achieved
by injecting types of “structured randomness” into the ensemble, the result of which is subsampled.
For reconstruction, the ensemble is modeled as a low-rank matrix that we have observed through an
(undetermined) set of linear equations. Our main results show that this matrix can be recovered using
a convex program when the total number of samples is on the order of the intrinsic degree of freedom of
the ensemble — the more heavily correlated the ensemble, the fewer samples are needed.
To motivate this study, we discuss how such ensembles arise in the context of array processing.
1
Introduction
This paper considers the exact reconstruction of correlated signals from the samples collected at a subNyquist rate. We propose several implementable architectures, and derive a sampling theorem that relates
the bandwidth and the (a priori unknown) correlation structure to the sufficient sampling rate for successful
signal reconstruction.
We consider ensembles of signals output from M sensors, each of which is bandlimited to frequencies below
W/2 (see Figure 1). The entire ensemble can be acquired by taking W uniformly spaced samples per second
in each channel, leading to a combined sampling rate of M W . We will show that if the signals are correlated,
meaning that the ensemble can be written as (or closely approximated by) distinct linear combinations of
R M latent signals, then this net sampling rate can be reduced to approximately RW using coded
acquisition. The sampling architectures, we propose are blind to the correlation structure of the signals; this
structure is discovered as the signals are reconstructed.
Each architecture involves a different type of analog diversification which ensures that the signals are sufficiently “spread out” so each point sample captures information about the ensemble. Ultimately, what is
measured are not actual samples of the individual signals, but rather are different linear combinations that
combine multiple signals and capture information over an interval of time. Later, we will show that these
samples can be expressed as linear measurements of a low-rank matrix. Over the course of one second, we aim
to acquire an M × W matrix comprised of samples of the ensemble taken at the Nyquist rate. The proposed
sampling architecture produces a series of linear combinations of entries of this matrix. Conditions under
which a low-rank matrix can be effectively recovered from an under-determined set of linear measurements
have been the object of intense study in the recent literature [1–4]; the mathematical contributions in this
paper show how these conditions are met by systems with clear implementation potential.
∗ A. A. and J. R. are with the School of Electrical and Computer Engineering at Georgia Tech in Atlanta, Georgia. Email:
[email protected], [email protected]. This work was supported by NSF grant CNS-0910592, ONR grant N00014-11-1-0459,
and a grant from the Packard Foundation.
1
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Our motivation for studying these architectures comes from classical problems in array signal processing. In
these applications, one or more “narrowband” signals are measured at multiple sensors at different spatial
locations. While narrowband signals can have significant bandwidth, they are modulated up to a high carrier
frequency, making them very heavily spatially correlated as they arrive at the array. This correlation, which
we review in more detail in Section 1.4, can be systematically exploited for spatial filtering (beamforming),
interference removal, direction-of-arrival estimation, and multiple source separation. These activities all
depend on estimates of the inter-sensor correlation matrix, and the rank of this matrix can typically be
related to the number of sources that are present.
Compressive sampling has been used in array processing in the past: sparse regularization was used for
direction of arrival estimation [5–7] long before any of the “sub-Nyquist” sampling theorems started to make
the theoretical guarantees concrete [8–10]. These results (along with more recent works including [11–13]),
show how exploiting the structure of the array response in free space (for narrowband signals, this consists
of samples of a superposition of a small number of sinusoids) can be used to either super-resolve the DOA
estimate or reduce the number of array elements required to locate a certain number of sources. A single
sample is associated with each sensor, and the acquisition complexity scales with the number of array
elements.
In this paper, we exploit this structure in a different way. Our goal is to completely reconstruct the timevarying signals at all the array elements. The structure imposed on this ensemble is more general than the
spatial spectral sparsity in the previous work; we ask that the signals are correlated in some a priori unknown
manner. Our ensemble sampling theorems remain applicable even when the array response depends on the
position of the source in a complicated way. Moreover, our reconstruction algorithms are indifferent to what
the spatial array response actually is, as long as the narrowband signals remain sufficiently correlated.
The paper is organized as follows. In Sections 1.2 and 1.4 we describe the signal model and its motivation
from problems in array processing. In Section 1.5, we introduce the components (and their corresponding
mathematical models) that we will use in our sampling architectures. In Section 2, we present the sampling
architectures, show how the measurements taken correspond to generalized measurements of a low-rank
matrix, and state the relevant sampling theorems. Numerical simulations, illustrating our theoretical results,
are presented in Section 3. Finally, Section 5, and Section 7 provide the derivation of the theoretical results.
1.1
Notation
We use upper, and lower case bold letters for matrices and vectors, respectively. Scalars are represented
by upper, and lower case, non-bold letters. The notation x∗ denotes a row vector formed by taking the
hermitian transpose of a column vector x. Linear operators, and sets are represented using script letters.
We use [N ] to denote the set {1, 2, 3, . . . , N }. The notation I W denotes a W × W . For a set B` ⊂ [W ],
I B` denotes a W × W matrix with ones at diagonal positions indexed by B` , and zeros elsewhere. Given
two matrices A, and B, we denote by A B, the rank-1 matrix: [vec(A)][vec(B)]T , where vec(A), and
vec(B) are the column vectors formed by stretching the columns of A, and B, respectively, and T denotes
the transpose. We will use A ⊗ B is the usual Kronecker product of A, and B. We will use 1P to denote
a P × 1 vector of all ones. Lastly, the operator E refers to the expectation operator, and P represents the
probability measure.
1.2
Signal model
Our signal model is illustrated in Figure 1. We denote a continuous-time signal ensemble by X c (t): a set
of M individual signals x1 (t), . . . , xM (t). Conceptually, we may think of X c (t) as a “matrix” with a finite
number M of rows, with each row containing a bandlimited signal. Our underlying assumption is that every
signal in the ensemble can be approximated as the linear combination of underlying R independent signals
in a smaller ensemble S c (t). We write
X c (t) = AS c (t),
(1)
2
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
X c (t)
A
X
S c (t)
A
S
M
W
(a)
(b)
Figure 1: (a) Our model is that an ensemble of continuous-time signals are correlated, meaning the M signals can
be closely approximated by a linear combination of R underlying signals. We can write the M signals in X c (t) as
a tall matrix (capturing the correlation structure) multiplied by an ensemble of R latent signals. (b) The matrix of
TexPoint)fonts)used)in)EMF.))
samples inherits the low-rank structure of the continuous-time
ensemble.
Read)the)TexPoint)manual)before)you)delete)this)box.:)A)
where A is an M × R matrix with entries A[m, r]. We will use the convention that fixed matrices operating
to the left of the signal ensembles simply “mix” the signals point-by-point, and so (1) is equivalent to
xm (t) =
R
X
A[m, r]sr (t).
r=1
The only structure, we impose on the individual signals is that they are real-valued, and bandlimited. To
keep the mathematics clean, we take the signals to be periodic for now, however, the results can be extended
to non-periodic signals as will be discussed shortly. We begin with a natural way to discretize the problem;
that is, what exists in X c (t) for t ∈ [0, 1) is all there is to know, and each signal can be captured exactly
with W equally-spaced samples. Each bandlimited, periodic signal in the ensemble can be written as
xm (t) =
B
X
αm [ω] eι2πωt ,
(2)
ω=−B
where the αm [ω] are complex but are symmetric, αm [−ω] = αm [ω]∗ , to ensure that xm (t) is real. We can
capture xm (t) perfectly by taking W = 2B + 1 equally spaced samples per row. We will call this the M × W
matrix of samples X; knowing every entry in this matrix is the same as knowing the entire signal ensemble.
We can write
X = CF ∗ ,
(3)
where C is an M × W matrix whose rows contain Fourier series coefficients for the signals in X c (t), and F
is a W × W normalized discrete Fourier matrix with entries
1
F [ω, `] = √ e−ι2πω`/W ,
W
ω = −B, −B + 1, . . . , B, ` = 0, 1, 2, . . . , W − 1.
Observe that both X, and hence C inherit the correlation structure of the ensemble X c (t). Before moving
on, observe that (1), and (2) impose an R, and B dimensional subspace structure on X, where rank(X) =
min(R, B + 1). If R ≥ B + 1 then we can take R = B + 1 with the underlying independent signals in S c (t)
being the known sinusoids at frequencies ω = 0, 1, 2, . . . , B. However, we are interested in the more pertinent
and challenging case of R < B + 1. In this case, the underlying independent signals in S c (t) are not known in
advance, and the main contribution of this paper is to leverage this unknown correlation structure in X c (t)
to reduce the sampling rate. Lastly, in the interest of readability of our technical results, we assume without
loss of generality that W ≥ M , that is, the bandwidth of the signals is greater than the number of signals.
Same correlated signal model was considered in [14] for compressive sampling of multiplexed signals. Two
multiplexing architectures were proposed and for each, a sampling theorem was proved that dictated minimum number of samples for exact recovery of the signal ensemble. This paper presents sampling architectures, where we use a separate ADC for each channel and rigorously prove that ADCs can operate at roughly
3
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
the optimal sampling rate to guarantee signal recovery. Other types of correlated signal models have been
exploited previously to achieve gains in the sampling rate. For example, [15] shows that two signals related
by a sparse convolution kernel can be reconstructed jointly at a reduced sampling rate. The signal model
in [16] considers multiple signals residing in a fixed subspace spanned by a subset of the basis functions of a
known basis, and shows that the sampling rate to successfully recover the signals scales with the number of
basis functions used in the construction of the signals. In this paper, we also show that the sampling rate
scales with the number of independent latent signals but we do this without the knowledge of the basis. For
a more applied treatment of the results with similar flavor as in [16], we refer the reader to [17–19].
As will be shown later, we observe the signal ensemble X c (t) through a limited set of random projections,
and signal recovery is achieved by a nuclear norm minimization program. A related work [20] considers the
case when given a few random projections of a signal, we find out the subspace to which it belongs by solving
a series of least-squares programs.
1.3
Extension to non-periodic signals
We end this section by noting that their are many ways this problem might be discretized. Using Fourier
series is convenient in two ways: we can easily tie together the notion of a signal being bandlimited with
having a limited support in Fourier space, and our sampling operators have representations in Fourier space
that make them more straightforward to analyze. In practice, however, the recovery technique can be
extended to non-periodic signals by windowing the input, and representing each finite interval using any one
of a number of basis expansions — the low rank structure is preserved under any linear representation. It
is also possible that we are interested in performing the ensemble recovery over multiple time frames, and
would like the recovery to transition smoothly between these frames. For this we might consider a windowed
Fourier series representations (e.g. the lapped orthogonal transform in [21]) that are carefully designed so
that the basis functions are tapered sinusoids (so we again get something close to bandlimited signals by
truncating the representation to a certain depth) but remain orthonormal. It is also possible to adjust our
recovery techniques to allow for measurements which span consecutive frames, yielding another natural way
to tie the reconstructions together.
A framework similar to this for sparse recovery is described in detail in [22].
1.4
Applications in array signal processing
One application area where low-rank ensembles of signals play a central role is array processing of narrowband
signals. In this section, we briefly review how these low-rank ensembles arise. The central idea is that
sampling a wavefront at multiple locations in space (as well as in time) leads to redundancies which can be
exploited for spatial processing. These concepts are very general, and are common to applications as diverse
as surveillance radars, underwater acoustic source localization and imaging, seismic exploration, wireless
communications.
The essential scenario is that multiple signals are emitted from different locations, each of the signals occupies
the same bandwidth of size W which has been modulated up to a carrier frequency ωc . The signals observed
by receivers in the array are, to a rough approximation, complex multiples of one another. To a very
close approximation, the observed signals lie in a subspace with dimension close to one — this subspace is
determined by the location of the source. This redundancy between the observations at the array elements
is precisely what causes the ensemble of signals to be low rank; the rank of the ensemble is determined by
the number of emitters. The only conceptual departure from the discussion in previous sections, as we will
see below, is that each emitter may be responsible for a subspace spanned by a number of latent “signals”
that is greater than one (but still small).
Having an array with a large number of appropriately spaced elements can be very advantageous even when
there only a relatively small number of emitters present. Observing multiple delayed versions of a signal
4
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
allows us to perform spatial processing, we can beamform to enhance or null out emitters at certain angles,
and separate signals coming from different emitters. The resolution to which we can perform this spatial
processing depends on the number of elements in the array (and their spacing).
The main results of this paper do not give any guarantees about how well these spatial processing tasks can
be performed. Rather, they say that the same correlation structure that makes these tasks possible can be
used to lower the net sampling rate over time. The entire signal ensemble can be reconstructed from this
reduced set of samples, and spatial processing can follow.
We now discuss in more detail how these low rank ensembles come about. For simplicity, this discussion
will center on linear arrays in free space. As we just need the signal ensemble to lie in a low dimensional
subspace, and do not need to know what this subspace may be beforehand, the essential aspects of the model
extend to general array geometries channel responses and frequency-selective/multipath channels.
Suppose that a signal is incident on the array (as a plane wave) at an angle θ. Each array element observes
a different shift of this signal — if we denote what is seen at the array center (the origin in Figure 2(a))
by s(t), then an element m at distance dm from the center sees xm (t) = s(t − (dm /c) sin θ). If the signal
consists of a single complex sinusoid, s(t) = eι2πωt , then these delays translated into different (complex)
linear multiples of the same signal,
xm (t) = e−ι2πωrm sin(θ)/c eι2πωt .
(4)
In this case, the signal ensemble has rank1 1; we can write X c (t) = a(θ, ω)eι2πωt , where a(θ, ω) is an
M -dimensional steering vector of complex weights given above.
This decomposition of the signal ensemble makes it clear how spatial information is coded into the array
observations. For instance, standard techniques [23,24] for estimating the direction of arrival involve forming
the spatial correlation matrix by averaging in time,
L
Rxx =
1X
X(t` )X(t` )∗ .
L
`=1
As the column space of Rxx should be a(θ, ω), we can correlate the steering vector for every direction to see
which one comes closest to matching the principal eigenvector of Rxx .
The ensemble remains low rank when the emitter has a small amount of bandwidth relative to a larger carrier
frequency. If we take s(t) = sb (t) eι2πωc t , where sb (t) is bandlimited to W/2, then when W ωc , the a(θ, ω)
for ω ∈ [ωc − W/2, ωc + W/2] will be very closely correlated with one another. In the standard scenario where
the array elements are uniformly spaced c/(2ωc ) along a line, we can make this statement more precise using
classical results on spectral concentration [25,26]. In this case, the steering vectors a(θ, ω) for ω ∈ [ωc ±W/2]
are equivalent to integer spaced samples of a signal whose (continuous-time) Fourier transform is bandlimited
to frequencies in (1 ± W/(2ωc ))(sin θ)/2, for a bandwidth less than W/(2ωc ). Thus the dimension of the
subspace spanned by {a(θ, ω), ω ∈ [ωc ± W/2]} is, to within a very good approximation, ≈ M W/ωc + 1.
Figure 2(b) illustrates a particular example. The plot shows the (normalized) eigenvalues of the matrix
Z
ωc +W/2
Raa =
a(θ, ω)a(θ, ω)∗ dω,
(5)
ωc −W/2
for the fixed values of ωc = 5 GHz, W = 100 MHz, c equals the speed of light, M = 101, and θ = π/4. We
have M W/ωc + 1 = 3.02, and only 3 of the eigenvalues are within a factor of 104 of the largest one.
It is fair, then, to say that the rank of the signal ensemble is a small constant times the number of narrow
band emitters.
1 We are using complex numbers here to make the discussion go smoothly; the real part of the signal ensemble is rank 2,
having a cos(2πωt) and a sin(2πωt) term.
5
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
log10 (kth largest eigenvalue)
0
dm sin ✓
✓
dm
-1
-2
-3
-4
-5
-6
-7
1
2
(a)
3
k→
(b)
4
5
Figure 2: (a) A plane wave impinges on a linear array in free space. When the wave is a pure tone in time, then the
responses at each element will simply be phase shifts of one another. (b) Eigenvalues for Raa , on a log10 scale and
normalized so that the largest eigenvalue is 1, defined in (5) for an electromagnetic signal with a bandwidth of 100
MHz and a carrier frequency of 5 GHz; the array elements are spaced half a carrier-wavelength apart. Even when the
signal has an appreciable bandwidth, the signals at each of the array elements are heavily correlated — the effective
dimension in this case is R = 3 or 4.
1.5
Architectural components
In addition to analog-to-digital converters, our proposed architectures will use three standard components:
analog vector-matrix multipliers, modulators, and linear time-invariant filters. The signal ensemble is passed
through these devices, and the result is sampled using an analog-to-digital converter (ADC) taking either
uniformly or non-uniformly spaced samples — these samples are the final outputs of our acquisition architectures.
dc (t)
x(t)
AVMM
x(t)dc (t)
(b)
x(t)
LTI filter
hc (t)
x(t) ⇤ hc (t)
(c)
M ⇥N
x(t)
(a)
ADC
x(tk )
(d)
Figure 3: (a) The analog vector-matrix multiplier (AVMM) takes random linear combinations of M input signals to
produce N output signals. The action of AVMM can be thought of as the left multiplication of random matrix A to
ensemble X c (t). Intuitively, this operation amounts to distributing energy in the ensemble equally across channels.
(b) Modulators multiply a signal in analog with a random binary waveform that disperses energy in the Fourier
transform of the signal. (c) Random LTI filters randomize the phase information in the Fourier transform of a given
signal by convolving it with hc (t) in analog, which distributes energy in time. (d) Finally, ADCs convert an analog
stream of information in discrete form. We use both uniform and non-uniform sampling devices in our architectures.
The analog vector-matrix multiplier (AVMM) produces an output signal ensemble AX c (t) when we input
it with signal ensemble X c (t), where A is an N × M matrix whose elements are fixed. Since the matrix
operates pointwise on the ensemble of signals, sampling output AX c (t) is the same as applying A to matrix
6
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
X of the samples (i.e., sampling commutes with the application of A). Recently, AVMM blocks have been
built with hundreds of inputs and outputs and with bandwidths in the tens-to-hundreds of megahertz [27,28].
We will use the AVMM block to ensure that energy disperses more or less evenly throughout the channels. If
A is a random orthogonal transform, it is highly probable that each signal in AX c (t) will contain about the
same amount of energy regardless of how the energy is distributed among the signals in X c (t) (formalized
in Lemma 1 below), allowing us to deploy equal sampling resources in each channel while ensuring that
resources on “quiet” channels are not being wasted.
The second component of the proposed architecture is the modulators, which simply take a single signal
x(t) and multiply it by fixed and known signal dc (t). We will take dc (t) to be a binary ±1 waveform that
is constant over time intervals of a certain length 1/W . That is, the waveform alternates at the Nyquist
sampling rate. If we take W samples of dc (t)x(t) on [0, 1), then we can write the vector of samples y as
y = Dx,
(6)
where x is the W -vector containing the Nyquist-rate samples of x(t) on [0, 1), and D is an W × W diagonal
matrix whose non-zero entries are samples d ∈ RW of dc (t). We will choose a binary sequence that randomly
generates dc (t), which amounts to D being a random matrix of the following form:
d[0]
d[1]
D=
(7)
where d[n] = ±1 with probability 1/2,
.
..
d[W − 1]
and the d[n] are independent. Conceptually, the modulator disperses the information in the entire band of
x(t) — this allows us to acquire the information at a smaller rate by filtering a sub-band as will be shown
in Section 2.
Compressive sampling architectures based on the random modulator have been analyzed previously in the
literature [18,29]. The principal finding is that if the input signal is spectrally sparse (meaning the total size
of the support of its Fourier transform is a small percentage of the entire band), then the modulator can be
followed by a low-pass filter and an ADC that takes samples at a rate comparable to the size of the active
band. This architecture has been implemented in hardware in multiple applications [17, 30–33].
The third type of component we will use to preprocess the signal ensemble is a linear time-invariant (LTI)
filter that takes an input x(t) and convolves it with a fixed and known impulse response hc (t). We will
assume that we have complete control over hc (t), even though this brushes aside admittedly important
implementation questions. Because x(t) is periodic and bandlimited, we can write the action of the LTI
filter as a W × W circular matrix H operating on samples x (the first row of H consists of samples h of
hc (t)), that is, y = Hx, where y is the vector of W samples in t ∈ [0, 1) of the signal obtained at the output
of the filter. We will make repeated use of the fact that H is diagonalized by the discrete Fourier transform:
H = F ∗ ĤF ,
(8)
where F is the W√× W normalized discrete Fourier matrix with entries, and Ĥ is a diagonal matrix whose
entries are Ĥ = W F H. The vector Ĥ is a scaled version of the non-zero Fourier series coefficients of
hc (t).
To generate the impulse response, we will use a random unit-magnitude sequence in the Fourier domain. In
particular, we will take
ĥ(0)
ĥ(1)
Ĥ =
(9)
..
.
ĥ(W − 1)
7
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
where
ω=0
±1, with prob. 1/2,
ĥ(ω) = ejθω , with θω ∼ Uniform([0, 2π]), 1 ≤ ω ≤ (W − 1)/2
.
∗
ĥ(W − ω) ,
(W + 1)/2 ≤ ω ≤ W − 1
(10)
These symmetry constraints are imposed so that h (and hence, hc (t)) is real-valued. Conceptually, convolution with hc (t) disperses a signal over time while maintaining fixed energy (note that H is an orthonormal
matrix).
Convolution with a random pulse followed by sub-sampling has also been analyzed in the compressed sensing
literature [34–37]. If the random filter is created in the Fourier domain as above, then following the filter with
an ADC that samples at random locations produces a universally efficient compressive sampling architecture
— the number of samples that we need to recover a signal with only S active terms at unknown locations
in any fixed basis scales linearly in S and logarithmically in ambient-dimension W .
2
Main Results: Sampling Architectures
The main contribution of the paper is a design and theoretical analysis of a sampling architecture in Section
2.3 that enables the sub-Nyquist acquisition of correlated signals. We state a Sampling Theorem 1 that
claims exact reconstruction of the signal ensemble using a much fewer samples compared to those dictated by
Shannon-Nyquist sampling theorem. The proof of the theorem involves the construction of a dual certificate
via golfing scheme to show that nuclear-norm minimization recovers the signal ensemble. Theorem 1 is also
of an independent interest as it is low-rank matrix recovery result form structured-random measurement
ensemble.
We begin with a straightforward architecture in Section 2.1 that minimizes the sample rate when the correlation structure is known. We then combine our components from the last section in a specific way to create
architectures that are provably effective under different assumptions on the signal ensemble. The main sampling architecture in Section 2.3 uses random modulators prior to the ADCs; this architecture is effective
when the energy in the ensemble is approximately uniformly dispersed across time. Moreover, we expect
the signal energy to be dispersed across array elements when the AVMM upfront does not mix the signals.
In Section 2.5, we present a variation of the above architecture in which ensembles are not required to be
dispersed a priori instead the ensemble is preprocessed with LTI filters, and AVMM to ensure dispersion of
energy across time, and array elements.
2.1
Fixed projections for known correlation structure
If the mixing matrix A for ensemble X c (t) is known, then a straightforward way exists to sample the
ensemble efficiently. Let A = U ΣV ∗ be the singular value decomposition of A, where U is M × R matrix
with orthogonal columns, Σ is R ×R diagonal matrix, and V is W ×R with orthogonal columns. An efficient
way is to whiten ensemble A with U ∗ and sample the resulting R signals (each at rate W ). This scheme is
shown in Figure 4. X can be written as a multiplication of matrix U and R × W matrix Y containing the
Nyquist samples of signals x1 (t), . . . , xR (t) respectively in its R rows. The discretized signal ensemble X is
then simply
X = UY .
Knowing the correlation structure U , the ensemble X and hence X c (t) (using sinc interpolation of samples
in X) can be recovered using only the RW samples in Y . Observe that RW is the optimal sampling rate
as it only scales linearly with R, and not with M .
In many interesting applications, the correlation structure of the ensemble X c (t) is not known at the time of
acquisition. In this paper, we design sampling strategies that are blind to the correlation structure U but are
able achieve signal reconstruction at a near optimal sampling rate nonetheless by introducing AVMMs, filters,
8
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
ADC
AVMM
ADC
…
Figure 4: Known correlation structure U : Optimal sampling strategy is to whiten the ensemble X c (t) with U ∗ and
then sample and then sample each of the resultant R signal at rate W . Total RW samples per second is optimal as
it is the actual number of degrees of freedom in underlying R independent signals each bandlimited to W/2.
and modulators. Intuitively, the randomness introduced by these components disperses limited information
in the correlated ensemble across time and array elements; resultantly, the ADCs collect more generalized
samples that in turn enable the reconstruction algorithm to operate successfully in the sub-Nyquist regime.
2.2
Architecture 1: Random sampling of time-dispersed correlated signals
The architecture presented in this section, shown in Figure 5, consists of one non-uniform sampling (nus)
ADC per channel. Each ADC takes samples at randomly selected locations, and these locations are chosen
independently from channel to channel. Over the time interval t ∈ [0, 1), a nus ADC takes input signal xm (t)
and returns the samples {xm (tk ) : tk ∈ Tm ⊂ {0, 1/W, . . . , 1 − 1/W }. The average sampling rate in each
channel is |Tm | = Ω. Collectively, M nus ADCs return M Ω random samples of the input signal ensemble on
a uniform grid.
avg-rate
nus
ADC
nus
ADC
…
…
avg-rate
avg-rate
nus
ADC
Figure 5: M signals recorded by the sensors are sampled separately by the independent random sampling ADCs,
each of which samples on a uniform grid at an average rate of Ω samples per second. This sampling scheme takes on
the average a total of M Ω samples per second and is equivalent to observing M Ω entries of the matrix of samples X
in (3) at random.
Sampling model is equivalent to observing M Ω randomly chosen entries of the matrix of samples X, defined
in (3). This problem is exactly the same as the matrix-completion problem [3], where given a few randomly
chosen entries of a low-rank matrix enable us to fill in the missing entries under some incoherence assumptions
on the matrix X. Since X is rank-R, its SVD is
X = U ΣV ∗ ,
9
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(11)
where U ∈ RM ×R , Σ ∈ RR×R , and V ∈ RW ×R . The coherence is then defined as
M
∗
∗
∗ 2
2
2 W
2 MW
κ (U , V ) := max
max kU em k2 ,
max kV e` k2 ,
kU V k∞ .
R m∈[M ]
R `∈[W ]
R
(12)
For brevity, we will sometime drop the dependence on U , and V in κ2 (U , V ). In the interest of readability,
we assume without loss of generality here and in the rest of the write up that bandwidth W of the signal is
larger or at least equal to their number M , that is, W ≥ M . Now the matrix-completion result [38] in the
noiseless case asserts that if
Ω ≥ Cκ2 (R/M )W log2 W,
(13)
then the solution of the nuclear-norm minimization in (21) (with A : RM ×W → RM Ω such that A maps
X to randomly chosen entries of X) exactly equals X with high probability. The result indicates that the
sampling rate scales (within some log factors) with the number R of independent signals rather than with
the total number M of signals in the ensemble. When the measurements y are contaminated with additive
measurement noise as in (22) then the result in [39] suggest that the solution X̂ to a modified nuclear-norm
minimization (23) satisfies
kX̂ − Xk2F ≤ Cκ M δ 2 ,
where Cκ is a constant that depends on the coherence κ, defined in (12).
As discussed before, the number of samples for matrix completion scale linearly with κ2 . The coherence
parameter κ2 quantifies the distribution of energy across the entries of X, and κ2 is small for matrices with
even distribution of energy among their entries; see, [3] for details. In the signal reconstruction application
under investigation here, this means that for successful recovery, a smaller sampling rate would suffice if
the signals are well-dispersed across time and array elements. One can avoid this dispersion requirement by
preprocessing the signals with AVMM, and filters. We will adopt this strategy in the construction of the
main sampling architecture of this paper.
2.3
Architecture 2: The random modulator for correlated signals
To efficiently acquire the correlated signal ensemble, the architecture shown in Figure 6, follows a two-step
approach:
In the first step, the AVMM takes M input to produce N output signals, where N/M = P > 1 meaning
that the output signals are more than the inputs. For now, we take N signals at the output to be just P
replicas of M input signals without any mixing.2 This amounts to an N × M mixing matrix
A = IM ⊗
√1
P
1P .
(14)
The normalization by P ensures that A∗ A = I M . We will take a more general random orthogonal A in our
next sampling architecture.
In the second step, each of the N output signals x̃1 (t), x̃2 (t), . . . , x̃N (t) undergo analog preprocessing, which
involves modulation, and low-pass filtering. The modulator takes an input signal x̃n (t) and multiplies it by
a fixed and known dn (t). We will take dn (t) to be a binary ±1 waveform that is constant over an interval
of length 1/W . Intuitively, the modulation results in the diversification of the signal information over the
frequency band of width W . The diversified analog signals are then processed by an analog-low-pass filter;
implemented using an integrator, see [29] for details. Each of the resultant signals is then acquired using
Ω < W uniformly spaced samples per second.
2 Our sampling theorem will show later that it suffices to take the ratio P of the number of output to the input to be
reasonably small. However, as will be suggested by our simulations, it seems P = 1 is always enough, and we believe P > 1
merely a technical requirement arising due to the proof method.
10
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Our main sampling result in Theorem 1 shows that exact signal reconstruction is achieved in the sub-Nyquist
regime Ω < W , in particular, we only roughly require Ω to be a factor R/M 1 of the Nyquist rate W .
Intuitively, the sub-Nyquist acquisition is possible as the signals are diversified across frequency using random
demodulators, and therefore, every sample provides a generalized or global information.
LTI
low pass
LTI
low pass
ADC
ADC
…
…
LTI
low pass
ADC
Figure 6: Architecture 2: Randomly modulated sampling: M correlated signals in X c (t) are replicated P times to
produce N output signals, this amounts to choosing A = I M ⊗ √1P 1P ∈ RN ×M as the mixing matrix. In practice,
P = 1 suffices. Signals are then preprocessed in analog using a bank of modulators, and low-pass filters. The
resultant signal is then sampled uniformly by an ADC in each channel operating at rate Ω samples per second. The
net sampling rate is ΩN samples per second.
2.3.1
System Model
This section models3 the measured samples as the linear measurements of an unknown low-rank matrix.
We will show that signal reconstruction in t ∈ [0, 1) from samples in the sub-Nyquist regime corresponds to
recovering a M × W approximately rank-R matrix from an under-determined set of linear equations.
The input signal ensemble X c (t) is mixed using AVMM to produce an ensemble of N signals AX c (t). Let
us denote the individual N signals at the output of AVMM by x̃1 (t), x̃2 (t), . . . , x̃N (t). Since mixing is a
linear operation, every signal in the ensemble AX c (t) is bandlimited just as was the case with X c (t) in (3);
therefore, the DFT coefficients of the mixed signals are simply
e = AC.
C
(15)
Each signal x̃n (t) at the output of AVMM is then multiplied by a corresponding binary sequence dn (t)
alternating at rate W . Each of the binary sequences d1 (t), d2 (t), . . . , dN (t) will be generated randomly, and
independently. The output after modulation in the nth channel is
yn (t) = x̃n (t) · dn (t),
n ∈ [N ], and t ∈ [0, 1).
The modulated outputs yn (t) are then low-pass filtered using an integrator, which integrates yn (t) over an
interval of width 1/Ω and the result is then sampled at rate Ω using an ADC. The `th sample acquired by
the ADC in the nth channel is
Z
`/Ω
yn [`] =
yn (t)dt,
(`−1)/Ω
` ∈ [Ω].
The integration operation commutes with the modulation process; hence, we can equivalently integrate the
signals zn (t) over the interval of width 1/W , and treat them as samples Z 0 ∈ RM ×W of the ensemble Z c (t).
3 Some of the initial development in this section may resemble with [29], but it is to be noted that compared to [29] the
signal structure to be exploited here is correlations among the signals and not the sparsity. This leads to a completely different
development towards the end of this section.
11
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
The entries Z0 [n, `] of the matrix Z 0 are
Z
`/W
Z0 [n, `] =
x̃n (t)dt =
(`−1)/W
B
X
ω=−B
ι2πω/W
e
− 1 −ι2πω`/W
e
e
C[n, ω]
,
ι2πω
(16)
e defined in (15), and the bracketed term representing the
e ω] are the entries of the matrix C
where C[n,
low-pass filter
ι2πω/W
e
−1
L[ω] =
, ω = −B, −B + 1, . . . , B,
ι2πω
where W = 2B + 1, as defined in (2). We will denote by L as a diagonal matrix containing L[ω] along the
diagonal. It is important to note that L is invertible as L[ω] does not vanish on any ω = −B, −B + 1, . . . , B.
In view of (16), it is clear that
∗
e
Z 0 = CLF
= ACLF ∗ = AX 0
(17)
where X 0 = CLF ∗ inherits its low-rank structure from X c (t).
Since we have already carried out integration over intervals of length 1/W , the action of modulator followed
by integration over 1/Ω now simply reduces to randomly, and independently flipping every entry of Z 0 and
adding consecutive W/Ω such entries in a given row to produce the value of the sample acquired by the
ADC. Mathematically, we can write this concisely by defining a vector dn` supported on an index set
B` := {(` − 1)W/Ω + 1 : `W/Ω},
` ∈ [Ω],
(18)
of size |B` | = W/Ω, where we are assuming4 for simplicity that Ω is a factor of W . On the support set B` ,
the entries of the vector dn` are independent binary ±1 random variables, and are zeros on B`c . Moreover,
assume that a∗n are the rows of A. With these notations in place, we can concisely write the `th sample in
t ∈ [0, 1) in the nth branch as
yn [`] := a∗n X 0 dn` , (n, `) ∈ [N ] × [Ω].
(19)
All this shows is that the samples taken by the ADC in the sampling architecture in Figure 6 are linear
measurements of an underlying low-rank matrix X 0 ∈ RM ×W , defined in (17). The rank of X 0 does not
exceed R — recalling from Section 1.2 that R constitutes the number of linearly independent signals in the
ensemble X c (t). Our objective is to recover X 0 from a a few linear measurements yn [`], which amounts to
reconstructing X c (t) at a sub-Nyquist rate sampling.
2.4
Matrix recovery
Define a linear map A : X 0 → y, where y is a length N Ω vector containing linear measurements yn [`] in
(19) as its entries. Formally,
A(X 0 ) = {a∗n X 0 dn` : (n, `) ∈ [N ] × [Ω]}.
(20)
We are mainly interested in the scenario when the linear map A is under determined, that is, the number of
measurements N Ω is much smaller than the number of unknowns M W . Therefore, to uniquely determine
the true solution X 0 , we solve a nuclear-norm penalized optimization program:
argmin kXk∗ subject to y = A(X),
(21)
X
where kXk∗ is the nuclear norm (the sum of the singular values of X). The nuclear norm penalty encourages
the solution to be low rank [1], and has concrete performance guarantees when the linear map A obeys certain
properties [2]. In case of noisy measurements
y = A(X 0 ) + ξ
4a
slight modification of this can result in an argument when Ω is not a factor of W ; for details, see [29]
12
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(22)
with bounded noise kξk2 ≤ δ, we solve the following quadratically constrained convex optimization program
argmin kXk∗ subject to ky − A(X)k2 ≤ δ.
(23)
X
This optimization program is also provably effective; see, for example, [39, 40] for suitable A.
2.4.1
Sampling theorem: Exact and stable recovery
The unknown matrix X 0 in (17) is at most rank-R, and assume X 0 = U ΣV ∗ is its reduced form SVD,
where U ∈ RM ×R , and V ∈ RW ×R are the matrices of left and right singular vectors, respectively; and
Σ ∈ RR×R is a diagonal matrix of singular values. Define coherences of X 0 as
µ2 (U ) :=
M
Ω
max kU ∗ em k22 , ρ2 (V ) := max kV ∗ I B` k2F ,
R m∈[M ]
R `∈[Ω]
M
Ω
max max ke∗ U V ∗ I B` k22 ,
ν 2 (U , V ) :=
R m∈[M ] `∈[Ω] m
and
(24)
where I B` is a diagonal W × W matrix containing ones at the diagonal positions indexed by B` . We may
sometime just work with notations µ2 , ρ2 , and ν 2 and drop the dependence on U , and V when it is clear
from the context. It can easily be verified that 1 ≤ µ2 ≤ M/R. In a similar manner, one can show that
1 ≤ ρ2 ≤ W/R. To see this, notice that
Ω
ρ2 ≥
Ω
Ω
ΩX
kV ∗ I B` k2F =
kV ∗ I B` k2F =⇒ Ωρ2 ≥
· R,
R
R
R
`=1
that is, ρ2 ≥ 1. Using the fact that kV ∗ I B` kF ≤ kV k2 kI B` k2F = W/Ω, the upper bound also follows. Finally,
similar techniques also show that 1 ≤ ν 2 ≤ M Ω/R. One can attach meaning to the values of coherences in
the context of sampling application under consideration. For example, the smallest value of ρ2 is achieved the
energy of X 0 is roughly equally distributed among the columns indexed by B1 , B2 , . . . , BΩ . In the context of
the sampling problem, this means that the energy in the signal ensemble X c (t) should be dispersed equally
across time. Similarly, the coherence µ2 quantifies the spread of signal energy across array elements, and ν 2
measures the dispersion of energy across both the time and array elements. Let us define
ϑ2 (U , V ) := max(µ2 (U ), ρ2 (V ), ν 2 (U , V )).
(25)
We are now ready to state our main result that dictates the minimum sampling rate Ω at which each ADCs
needs to be operated to guarantee the reconstruction of signal ensemble X c (t).
Theorem 1. Correlated signal ensemble X c (t) in (1) can be acquired using the sampling architecture in
Figure 6 by operating each of the ADCs at a rate
Ω ≥ C β ϑ2
R
W log2 W,
M
(26)
where Cβ is a universal constant that only depends on the fixed parameter β ≥ 1. In addition, the ratio of
the number of output to the input signals in AVMM must satisfy N/M ≥ C log W , where C is a numerical
constant. The exact signal reconstruction can be achieved with probability at least 1 − O(W 1−β ) by solving
the nuclear-norm minimization program in (21).
The result indicates that M well spread out (ϑ2 ≈ 1) correlated signals can be acquired by operating each
ADC in Figure 6 at a rate of R/M 1 times the Nyquist-rate W to within log factors. Moreover, we also
require the number N of output signals at the AVMM to be larger than number M of input signals by a
log factor. However, we believe this is merely an artifact of the proof technique and our experiments also
corroborate that successful recovery is always obtained for Ω satisfying (26) even when N = M or P = 1.
13
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Also note that the result in Theorem 1 assumes without loss of generality that W ≥ M . In the other case,
when M ≥ W , the sufficient sampling rate at each ACC can be obtained by replacing W in (26) with M .
Another important observation is that the sampling rate scales linearly with coherence ϑ2 implying that
the sampling architecture is not as effective for correlated signals concentrated across time. To remedy this
shortcoming, a preprocessing step using random filters, a mixing AVMM can be added to ensure signals are
well-dispersed across time, and array elements.
2.4.2
Stable recovery
In a realistic scenario, the measurements (19) are almost always contaminated with noise ξn [`]
yn [`] = a∗n X 0 dn` + ξn [`],
(n, `) ∈ [N ] × [Ω]
(27)
compactly
expressed using the vector equality in (22). In the case, when the noise is bounded, i.e.,
P
2
|ξ
[`]|
≤ δ 2 , then following the template of the proof in [39], it can be shown that under the conn
n,`
c of (23) obeys
ditions of Theorem 1, the solution X
√
c − X 0 kF ≤ C M δ
kX
(28)
with high probability; for more details on this,
p see a similar stability result in Theorem 2 in [41]. The upper
bound above is suboptimal by a factor of min(W, M ). In theory, we can improve this suboptimal result
and show the effectiveness of the nuclear norm penalty by analyzing a different estimator:
argmin kX − A∗ (y)k2F + λkXk∗ .
(29)
X
This estimator was proposed in [42], and can be theoretically shown to obey essentially optimal stable
c is the minimizer of (29) if and only if 0 ∈ ∂ kX −A∗ (y)k2 +λkXk∗ ,
recovery results. Using the fact that X
F
c is a simple soft thresholding of the singular values of the matrix
one can show [42] that the estimate X
A∗ (y) ∈ RM ×W
X
λ
c=
X
σr (A∗ (y)) −
ur (A∗ (y))v ∗r (A∗ (y)),
(30)
2
+
r
where x+ = max(x, 0); in addition, ur (A∗ (y)), and v r (A∗ (y)) are the left and right singular vectors of the
matrix A∗ (y), respectively; and σr (A∗ (y)) is the corresponding singular value.
In comparison to the estimator (29), the matrix Lasso in (23) does not use the knowledge of the known
distribution of A and instead minimizes the empirical risk ky − A(X)k2 = kyk22 − 2hy, A(X)i + kA(X)k22 .
Knowing the distribution, and the fact that E A∗ A = I holds in our case, we replace kA(X)k22 , by its
expected value E kA(X)k22 = kXk2F in the empirical risk to obtain the estimator in (29) by completing the
square. Although the KLT estimator is easier to analyze, and will be shown to give optimal stable recovery
results in theory, but it does not empirically perform as well as matrix Lasso in (23).
We quantify the strength of the noise vector ξ ∈ RN Ω through its Orlicz-2 norm. For a random vector z, we
define
2
2
kzkψ2 := inf{u > 0 : E ekzk2 /u ≤ 2 },
and for scaler random variables, we simply take z ∈ R1 in the above definition. The Orlicz-2 norm is finite
if the entries of z are subgaussain, and is proportional to variance, if the entries are Gaussian. We assume
that the entries of the noise vector ξ obey
kξn [`]kψ2 ≤ cδ N1Ω , and kξkψ2 ≤ cδ.
With this the following result is in order.
14
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(31)
Theorem 2. Fix β ≥ 1. Given measurements y of X 0 in (27) contaminated with additive noise ξ with
c to (29) obeys
statistics in (31), the solution X
c − X 0 k2 ≤ C max(δ, δ 2 )
kX
F
(32)
with probability at least 1−O(W −β ) whenever Ω ≥ Cβ ν 2 (R/M )W log3/2 W , where Cβ is a universal constant
depending only on β.
Roughly speaking, the stable recovery theorem states that the nuclear norm penalized estimators are stable
in the presence of additive measurement noise. The results in Theorem 2 are derived assuming that ξn [`] are
random with statistics in (31). In contrast, the stable recovery results in the compressed sensing literature
only assume that the noise is bounded, i.e., kξk2 ≤ δ, where ξ is the noise vector introduced earlier. Here,
we give a brief comparison of Theorem (2) with the stable recovery results in [39, 43].
Compare
√ the result in (28) with (32), it follows that our results improve upon the results in [39] by a factor
of 1/ M . We will also compare our stable recovery results against the stable recovery results derived in [43].
The result roughly states if the linear operator A satisfies the matrix RIP [2], and kξk2 ≤ δ, then the solution
c to (23) obeys
X
c − X 0 kF ≤ Cδ.
kX
(33)
The above result is essentially optimal stable recovery result. In comparison to (33), the result in (32) is
also optimal, however, we prove it for a different estimator and under a statistical bound on the noise term
kξkψ2 ≤ δ. In addition, we also donot require the matrix RIP for A, which is generally required to prove
optimal results of the form of (33).
2.5
Architecture 3: Uniform sampling architecture
The discussion in Section 2.2, and the result in Theorem 1 suggest that sampling rate sufficient for exact
recovery using the architecture 1 and 2 scales linearly with the coherence parameter µ2 , and ϑ2 , respectively.
As discussed earlier, the coherence parameters quantify the energy dispersion in the correlated signal ensemble X c (t) across time and array elements. Ideally, we would like the sampling rate Ω to only scale with factor
of W , and be independent of signal characteristics (coherences). To achieve this, signals are preprocessed
with random filters and AVMM so that signal energy is evenly distributed across time, and array elements.
The resultant signals are the randomly modulated, low-pass filtered, and sampled uniformly at a rate Ω.
The modified sampling architectures are depicted in Figure 7 and 8.
avg-rate
nus
ADC
nus
ADC
avg-rate
…
…
…
avg-rate
nus
ADC
Figure 7: Architecture 3: Analog vector-matrix multiplier (AVMM) takes random linear combinations of M input
signals to produce M output signals. This equalizes energy across channels. The random LTI filters convolve the
signals with a diverse waveform that results in dispersion of signals across time. The resultant signals are then
sampled, at locations selected randomly on a uniform grid, at an average rate Ω, using a non-uniform sampling (nus)
ADC in each channel.
15
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
LTI
low pass
ADC
LTI
ADC
low pass
…
…
…
LTI
low pass
ADC
Figure 8: Architecture 4: Random LTI filters disperse each of the M signal across time. An Analog vector-matrix
multiplier (AVMM) takes random linear combinations of M input signals to produce N output signals. This amounts
to choosing [Φ ⊗ √1P 1P ] = AΦ as the mixing matrix, where A is as in (14), and Φ is an M × M dense, randomorthogonal matrix. The well dispersed signals across time and array elements are now randomly modulated, low-pass
filtered, and sampled at rate Ω.
Recall that random LTI filters are all pass, and convolve the signals with a diverse impulse response hc (t),
which disperses signal energy over time w.h.p. (see Lemma 1). We will use the same random LTI filter
hc (t) in each channel. The action of the random convolution [34] of hc (t) with each signal in the ensemble
can be modeled by the right multiplication of a circulant random orthogonal matrix H ∈ RW ×W with the
underlying low-rank X 0 in (17). The AVMM takes the random linear combination of M input signals to
produce N output signals, which then equalizes w.h.p., the signal energy across array elements regardless of
the initial energy distribution.
As discussed earlier that the action of AVMM is left multiplication of A with the low-rank ensemble X c (t).
In Architecture 3, the AVMM is M -in-M -out, and to ensure mixing of signals across array elements, we
take the mixing matrix to be an M × M random orthonormal matrix Φ. Thus, the samples collected in
Architecture 3 are not the subset of the entries of X, defined in (3) but of
X̃ = ΦXH.
In Architecture 4, the AVMM is modified from A = [I M ⊗
[Φ ⊗
√1
P
√1
P
(34)
1P ] ∈ RN ×M in (14) to
1P ] = AΦ,
(35)
where Φ ∈ RM ×M is a random orthonormal matrix. This implies that unlike the samples y = A(X 0 ) , the
Architecture 4 collects y = A(X̃ 0 ), where
X̃ 0 = ΦX 0 H,
(36)
and A is same as defined in (20).
Both (34), and (36) multiply the matrix of samples X, and X 0 with random orthogonal matrices on the left,
and right. This multiplication results in modifying the singular vectors U ∈ RM ×R , and V ∈ RW ×R of the
e = ΦU ∈ RM ×R and Ve = HV ∈ RW ×R . Note that matrix X̃,
matrix of samples (either X, or X 0 ) to U
and X̃ 0 are an isometry with, and have the same rank as X, and X 0 , respectively. The new left and right
16
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
singular vectors Ũ and Ve of X̃, or X̃ 0 are in some sense random orthogonal matrices and hence, incoherent
w.h.p.
The following lemma shows the incoherence of matrix Ũ , and Ṽ .
Lemma 1. Fix matrices U ∈ RM ×R and V ∈ RW ×R of the left and right singular vectors, respectively.
e = ΦU , and Ve = HV , and the
Create random orthonormal matrices Φ ∈ RM ×M and H ∈ RW ×W . Let U
2 e e
2 e e
coherences κ (U , V ), and ϑ (U , V ) be as defined in (12), (25). Then for a β ≥ 1, the following conclusions
1
log M ),
R
(37)
e , Ve ) ≤ Cβ log W max (1, 1 log M )
ϑ2 ( U
R
(38)
e , Ve ) ≤ Cβ log W max (1,
κ2 (U
and
each holding with probability exceeding 1 − O(W −β ).
Proof of Lemma 1 is presented in Section 4. In light of (34), it is clear that samples collected using
Architecture 3 are randomly selected subset of the entries of X̃, and using the result in (37), and (13), the
sufficient sampling rate for the successful reconstruction of signals becomes
Ω≥C
1
max(R, log M )W log3 W.
M
In light of (36), it is clear that the samples collected using Architecture 4 are the same as in (19) with X 0
f0 . With this observation, combining the bound on ν 2 in Lemma 1 with Theorem 1 immediately
replaced by X
provides with the following corollary that dictates the sampling rate sufficient for exact recovery using the
uniform sampling architecture in Figure 8.
Corollary 1. Fix β ≥ 1. The correlated ensemble X c (t) in (1) can be exactly reconstructed using the
optimization program in (21) with probability at least 1 − O(W −β ) from the samples collected by each of the
ADC in Figure 8 at a sub-Nyquist rate:
Ω ≥ Cβ
1
max(R, log M )W log4 W,
M
(39)
where Cβ is a universal constant depending only on β. In addition, the ratio of the number of output to the
input signals in AVMM must satisfy N/M ≥ C log W for a sufficiently large constant C.
3
Numerical Experiments
In this section, we study the performance of the proposed sampling architectures with some numerical
experiments. We mainly show that a correlated ensemble X c (t) in (1) can be acquired by only paying a
small factor on top of the optimal sampling rate of roughly RW . We then show the distributed nature of the
sampling architecture in Figure 6 by showing that increasing the number of ADCs (or the array elements),
the sampling burden on each of the ADC can be reduced as the net sampling rate is shared evenly among
the ADCs. Finally, we show that the reconstruction algorithm is robust to additive noise.
3.1
Sampling performance
In all of the experiments in this section, we generate the unknown rank-R matrix X 0 synthetically by
multiplying tall M × R, and fat R × W Gaussian matrices. Our objective is to recover a batch of M = 100
signals, with W = 1024 samples taken in a given window of time using the sampling architecture in Figure
6. We take P = 1 or N = M for all these experiments, and the results hint that P > 1 or N > M in
17
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Theorem 1 is only a technical requirement due to the proof technique. We will use the following parameters
to evaluate the performance of the sampling architecture:
Oversampling factor : η =
MΩ
,
R(W + M − R)
where the oversampling factor is the ratio between the cumulative sampling rate, and the inherent unknowns
in X 0 . The successful reconstruction is declared when the relative error obeys
Relative error :=
c − X 0 kF
kX
≤ 10−2 .
kX 0 kF
The first experiment shows a graph, in Figure 9(a), between η, and R. Each point, marked with a black
dot, represents the minimum sampling rate required for the successful reconstruction of an X c (t) with a
specific R. The probability of success for each point is 0.99, and is computed empirically by averaging over
100 independent iterations. The blue line shows the least-squares fit of the black dots. It is clear from the
plot that the for reasonably large values of R, the sampling rate is within a small constant of the optimal
rate R(W + M − R).
14
700
12
600
10
Sampling rate Ω
Oversampling η
In context of the application, and under the narrow-band assumption described in Section 1.4, the graph
in Figure 9(b) shows that for a fixed number of sources R = 10, the sufficient sampling rate Ω is inversely
proportional to number M of the receiver array elements. Each black dot represents the minimum sampling
rate required for the successful reconstruction with probability 0.99. The blue line is the least-squares fit
of these marked points. In other words, Figure 9(b) illustrates the relationship between the number M of
ADCs and the sampling rate Ω for a fixed number of sources R = 10. Importantly, an increase in the receiver
array elements reduces the sampling burden on each of the ADCs.
8
6
4
500
400
300
200
2
0
0
Number of source: R = 10
5
10
Rank R
15
100
0
20
50
100
150
200
250
300
350
M: Number of ADCs
(a)
(b)
Figure 9: Performance of Sampling Architecture 2. In these experiments, we take an ensemble of 100 signals, each
bandlimited to 512Hz. The probability of success is computed over 100 iterations. (a) Oversampling factor η as a
function of the number R of underlying independent signals in X c (t). The blue line is the least-squares fit of the
data points. (b) Sampling rate Ω versus the number M of recieving antennas. The blue line is the least-sqaures fit
of the data points.
3.2
Stable recovery
In the second set of experiments, we study the performance of the the recovery algorithm when the measurements are contaminated with additive measurement noise ξ as in (22). We generate noise using the standard
18
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
√
Gaussian model: ξ ∼ N (0, σ 2 I). We select δ ≤ σ(L + L)1/2 ; a natural choice as the condition kξk2 ≤ δ
holds with high probability. For the experiments in Figure 10, we solve the optimization program in (23).
The plot in Figure 10(a) shows the relationship between the signal-to-noise ratio (SNR):
kX 0 k2F
,
SNR(dB) = 10 log
kξk22
and the realtive error(dB):
c − X 0 k2
kX
F
kX 0 k2F
Relative error (dB) = 10 log
!
for a fixed oversampling factor η = 3.5. The result shows that the relative error degrades gracefully with
decreasing SNR. In the Figure 10(b), the plot depicts relative error as a function of the oversampling factor
for a fixed SNR = 40dB. The relative error decrease with increasing sampling rate.
−3
0
12
x 10
Relative error
Relative error (dB)
11
−20
−40
−60
10
9
8
7
6
−80
0
20
40
SNR (dB)
(a)
60
5
2
80
2.5
3
3.5
Oversampling η
4
4.5
(b)
Figure 10: Recovery using matix Lasso in the presence of noise. The input ensemble to the simulated random
demodulator consists of 100 signals, each bandlimited to 512Hz with number R = 15 of latent independent signals.(a)
The SNR in dB versus the relative error in dB. The oversampling factor η = 3.5. (b) Relative error as a function of
the sampling rate. The SNR is fixed at 40dB.
4
Proof of Lemma 1
We start with the proof of Lemma 1
e = AU , where we are taking A ∈ RM ×M to be a random orthogonal matrix, and
Proof. Recall that U
Ve = HV , where H was defined in (8). Let em denote the standard basis vectors in RM . We begin the
proof by noting a standard result; see, [44], that reads
∗
e em k2 ≤
max kU
2
m∈[M ]
1
Cβ max (R, log M ),
M
(40)
with probability at least 1 − O(M −β ). Before proving the lemma, we prove an intermediate result:
∗
1
max kVe ek k22 ≤
Cβ max (R, log W ),
W
k∈[W ]
19
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(41)
where ek are standard basis vectors in RW . Assuming W is even; it will be clear how to extend the argument
to W odd. We can write H = W Q∗ , where
z
√1
√0 ,
ω=0
ω=0
W
W
√2 cos 2πωn + θ ω = [1, W − 1]
√2 cos 2πωn ω = [1, W − 1]
ω
W
2
W
2
W
W
W [n, ω] = zW/2
.
Q[n, ω] =
1
W
W
k−1
k−1
√
√
(−1)
ω
=
(−1)
,
ω
=
W
W
2
2
√2 sin 2πωn + θ , ω = [ W + 1, W − 1]
√2 sin 2πωn ω = [ W + 1, W − 1]
W
W
2
W
W
ω
2
(42)
and z0 , zW/2 = ±1 with equal probability and θω for ω = 1, . . . , W/2 − 1 are uniform on [0, 2π] and all
W/2 + 1 of these random variables are independent.
It is a fact that in (42) for fixed a and θ ∼ Uniform([0, 2π]), the random variables sign(cos(a + θ)) and
sign(sin(a + θ)) are independent of one another. Thus H has the same probability distribution as W ZQ∗ ,
where Z = diag(z) and the entries of z are iid ±1 random variables. In light of this, we will replace H with
W ZQ∗ . For a fixed k, we can write
∗
e ∗ Zwk =
Ve ek = V ∗ H ∗ ek = Q
W
X
z[ω]wk [ω]q ω
ω=1
e = Q∗ V and wk = W ∗ ek and q ω is the ωth column of Q
e ∗ . We will apply the following concentration
where Q
inequality.
Theorem 3. [45] Let η ∈ Rn be a vector whose entries η[i] are independent random variables with |η[i]| < 1,
and let S be a fixed m × n matrix. Then for every t ≥ 0
2
2
P kSηk2 ≥ E kSηk2 + t ≤ 2e−t /16kSk , where E kSηk2 ≤ kSkF .
∗
e W k , where W k = diag(wk ), and η = z. In this case, we have
We can apply the above theorem with S = Q
e ∗W k
Q
2
F
=
W
X
ω=1
∗
e W kk ≤
and kQ
q
2
W
e =
kQk
q
2
W
2
2
|wk [ω]| kq ω k2 ≤
W
2R
2 X
2
kq ω k2 ≤
,
W ω=1
W
q
q
∗
2
2R
2
. Thus, P kVe ek k2 >
+
t
≤ 2e−t /16 , and using the union
W
W
bound
r
r
∗
2
2R
2
e
P max kV ek k2 >
+t
≤ 2W e−t /16 .
W
W
k∈[W ]
√
−β
We can make this probability less than W
by taking t ≥ C log W , and (41) follows.
Now to prove (37), and (38), we can write H = W ZQ∗ . Let wk be the kth column of W ∗ and let ũ∗m be
e . For a fixed row index m and column index k, we can write an entry of U
e Ve ∗ as
the mth row of U
h
i
h
i
h
i
e Ve ∗
e (W ZQ∗ V )∗
eQ
e ∗ ZW ∗
e ∗ Zwk
U
= U
= U
= ũ∗m Q
m,k
m,k
m,k
∗
e = Q∗ V is a tall orthonormal matrix. Let p∗ = ũ∗ Q
e . Since the z[ω] are iid random variables, a
where Q
m
m
standard applications of the Hoeffding inequality tells us that
P
2
2
e Ve ∗
U
> λ ≤ 2e−λ /2σ ,
m,k
2
σ 2 = E wk Z ∗ pm p∗m Zwk ≤
where
2
2 kũm k2
2 kpm k2
=
.
W
W
Thus, with probability exceeding 1 − 2W −β
h
i
e Ve ∗
U
m,k
2
≤
4β log W
2
kũm k2 .
W
20
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(43)
Taking the maximum over m ∈ [M ], and k ∈ [W ] on both sides, and plugging in the bound in (40) shows
that
MW
max max
R m∈[M ] k∈[W ]
e Ve ∗ ek
e∗m U
2
≤ Cβ log W max(1,
1
log M )
R
(44)
holds with probability at least 1 − O(W −β + M −β ) = 1 − O(W −β ), where the equality follows from the fact
that W ≥ M . This proves the first claim (37) in Lemma 1.
Similarly, (43) implies that
X
e Ve ∗
U
m,k
2
k∈B`
≤ |B` |
4β log W
4β log W
2
kũm k2 =
kũm k22 ,
W
Ω
(45)
where B` is defined in (18), and the last equality follows from the fact that |B` | = W/Ω. Finally, evaluating
the maximum over m ∈ [M ], and ` ∈ [Ω] on both sides, and using the bound in (40) shows that
MΩ
e Ve ∗ I B k2 ≤ Cβ log W max(1, 1 log M ),
max max ke∗m U
` 2
R m∈[M ] `∈[Ω]
R
which proves the second claim (38) in Lemma 1.
5
5.1
Proof of Theorem 1
Preliminaries
Recall from (19), we obtain measurements yn [`] of an unknown low-rank matrix X 0 through a random
M × W rank-1 measurement ensemble an d∗n` , (n, `) ∈ [N ] × [Ω], where a∗n ∈ RM denote the rows of mixing
matrix A ∈ RN ×M , and dn` ∈ RW are random binary on support set B` , and zero elsewhere. In addition,
the vectors`n` are independently generated for every n, and `. In Theorem 1, the AVMM simply replicates
(without mixing) P copies of M input signals to produce N output signals. This amounts to choosing
A=
√1
P
[I M ⊗ 1P ] =
[I M , . . . , I M ]∗ ,
√1
P
(46)
where P = N/M . From this construction, we have kan k22 = 1/P , and A∗ A = I M . Also recall that using
the definition of linear map A in (20), the measurements are compactly expressed as y = A(X 0 ). Moreover,
the adjoint operator is
X X
X X
A∗ (y) =
yn [`]an d∗n` =
an a∗n X 0 dn` d∗n` ,
n∈[N ] `∈[Ω]
n∈[N ] `∈[Ω]
where the second equality is the result of (19). It will also be useful to visualize the linear operator A∗ A in
a matrix form:
X X
X X
A∗ A =
an d∗n` an d∗n` =
an a∗n ⊗ dn` d∗n` ,
(47)
n∈[N ] `∈[Ω]
n∈[N ] `∈[Ω]
where denotes the tensor product. In general, the tensor product of rank-1 matrices x1 y ∗1 , x2 y ∗2 with
xi ∈ RM , and y i ∈ RW is given by the big matrix
x1 [1]x2 [1]y 1 y ∗2 x1 [1]x2 [2]y 1 y ∗2 · · · x1 [1]x2 [M ]y 1 y ∗2
x1 [2]x2 [1]y 1 y ∗2 x1 [2]x2 [2]y 1 y ∗2 · · · x1 [2]x2 [M ]y 1 y ∗2
∗
∗
x1 y 1 x2 y 2 =
.
..
..
.
.
x1 [M ]x2 [1]y 1 y ∗2
···
···
x1 [M ]x2 [M ]y 1 y ∗2
21
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
With this definition it is easy to visualize that E A∗ A = I. Let {u∗k }k∈[M ] , and {v ∗k }k∈[W ] denote the rows
of the matrices U , and V , respectively.
We begin by defining a subspace T ⊂ RM ×W associated with X 0 with singular-value decomposition given
by X 0 = U ΣV ∗
T = {X : X = U Z ∗1 + Z 2 V ∗ , Z 1 ∈ RW ×R , Z 2 ∈ RM ×R }.
The orthogonal projections onto T , and its orthogonal complement T ⊥ are defined as PT (Z) = U U ∗ Z +
ZV V ∗ − U U ∗ ZV V ∗ , and PT ⊥ (Z) = Z − PT (Z), respectively. In the proofs later, we repeatedly make use
of the following calculation
2
kPT (an d∗n` )kF = hPT (an d∗n` ), an d∗n` i
= hU ∗ an d∗n` , U ∗ an d∗n` i + han d∗n` V , an d∗n` V i − hU ∗ an d∗n` V , U ∗ an d∗n` V i
= kU ∗ an d∗n` k2F + kan d∗n` V k2F − kU ∗ an d∗n` V k2F ≤ kU ∗ an d∗n` k2F + kan d∗n` V k2F .
Observe that
kU ∗ an d∗n` k2F = kU ∗ an k22 kd∗n` k22 =
W
1
kU ∗ an k22 , and kan d∗n` V k2F = kan k22 kd∗n` V k22 = kd∗n` V k22 .
Ω
P
This leads us to
1
W
kU ∗ an k22 + kd∗n` V k22 .
(48)
Ω
P
Finally, we will also require a bound on the operator norm of the linear map A. To this end, note that the
measurement matrices an d∗n` are orthogonal for every ` ∈ [Ω] in the standard Hilbert-Schmidt inner product,
that is, han d∗n` , an d∗n`0 i = 0 whenever ` 6= `0 . This directly implies a following bound on the operator of A:
r
sX
MW
∗ 2
kAk ≤
kan dn` kF =
≤ W,
(49)
Ω
kPT (an d∗n` )k2F ≤
n∈[N ]
where in the last inequality we used the fact that M ≤ W , and Ω ≥ 1. Although a much tighter bound can
be achieved using results from random matrix theory, the loose bound is sufficient for our purposes.
5.2
Sufficient condition for the uniqueness
Uniqueness of the minimizer to (21) can be guaranteed by the sufficient condition [3, 4] given below.
Proposition 1. The matrix X is the unique minimizer to (21) if ∃Y ∈ Range(A∗ ) such that ∀Z ∈ Null(A)
(1 − kPT ⊥ (Y )k) kPT ⊥ (Z)k∗ − kU V ∗ − PT (Y )kF kPT (Z)kF > 0.
In light of the proposition, it is sufficient to show that ∃Y ∈ Range(A∗ ), such that
kPT (Y ) − U V ∗ kF ≤ 1/3W,
and for every Z ∈ Null(A),
kPT ⊥ (Y )k ≤ 1/2,
√
kPT ⊥ (Z)kF ≥ (1/ 2W )kPT (Z)kF
(50)
(51)
holds. This can be immediately shown as follows 0 = kA(Z)kF ≥ kA(PT (Z))kF − kA(PT ⊥ (Z))kF ≥
kA(PT (Z))kF − W kPT⊥ (Z)kF . In addition, for an arbitrary Z, we have
kA(PT (Z))k2F = hA(PT (Z)), A(PT (Z))i = hZ, PT A∗ APT (Z)i
1
≥ (1 − kPT A∗ APT − PT k)kPT (Z)k2F ≥ kPT (Z)k2F ,
2
where the last inequality is obtained by plugging in kPT A∗ APT − PT k ≤ 1/2, which will be shown to be
true under appropriate choice of Ω with probability at least 1 − O(W −β ) in Corollary 2. Combining the last
two inequalities gives us the result in (51).
22
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
5.3
Golfing scheme for the random modulator
For technical reasons, we will work with partial linear maps Ap : RM ×W → RM Ω , p ∈ [P ] modified from the
linear map A in (20). Define P partitions {Γp }P
p=1 of the index set [N ] as Γp := {(p − 1)M + 1, . . . , pM } for
5
every p ∈ [P ]. Clearly, Γp ∩ Γp0 = ∅, and ∪P
p=1 Γp = [N ]. We will take the number of partitions P = N/M .
The partial linear maps Ap are defined as
Ap (X) := {a∗n Xdn` , n ∈ Γp , ` ∈ [Ω]}.
(52)
Using the definition of A in (46), it is clear that
{an : n ∈ Γp } = { √1P em : m ∈ [M ]} for every p ∈ [P ] =⇒
X
an a∗n =
1
P
IM .
(53)
n
The corresponding adjoint operator maps a vector z ∈ RM Ω to an M × W matrix
X X
A∗p (z) =
zn [`]an d∗n` .
n∈Γp `∈[Ω]
It will also be useful to make a note of the following versions of the above definition
X X
X X
A∗p Ap (X) =
an a∗n Xdn` d∗n` , and A∗p Ap =
an d∗n` an d∗n` ,
n∈Γp `∈[Ω]
(54)
n∈Γp `∈[Ω]
where the second definition just emphasizes the fact that the linear map A∗p Ap can be thought of as a big
M W × M W matrix that operates on a vectorized X.
With the linear operators defined on the subsets {Γp }P
p=1 above, we write the iterative construction of the
dual certificate:
Y p = Y p−1 − A∗p Ap PT (Y p−1 ) − U V ∗ , where Y p ∈ Range(A∗ ),
(55)
where we take Y 0 = 0. Projecting onto the subspace T on both sides results in PT (Y p ) = PT (Y p−1 ) −
PT A∗p Ap (PT (Y p−1 ) − U V ∗ ). Define
W p : = PT (Y p ) − U V ∗ ,
(56)
the iteration takes the equivalent form W p = W p−1 − PT A∗p Ap PT (W p−1 ). We will take Y = Y P to be our
candidate for the dual certificate and the rest of this section concerns showing that Y P obeys the conditions
in (50). Let’s start by showing that kPT (Y P ) − U V ∗ kF ≤ (3W )−1 holds. To this end, note that from the
iterative construction above, the following bound immediately follows
kW p kF ≤ kPT A∗p Ap PT − PT kkW p−1 kF .
From Lemma 2, we have kPT A∗p Ap PT − PT k ≤ 1/2 for every p ∈ [P ]. This means that kW p kF cuts after
every iteration giving us the following bound on the Frobenius norm of the final iterate W P
√
(57)
kW P kF ≤ 2−P kU V ∗ kF = 2−P R ≤ (3W )−1 when P ≥ 2 log2 (3W ).
Using the union bound over p ∈ [P ], the bound on kW P kF holds with probability at least 1 − O(P W −β ) ≥
1 − O(W 1−β ). This proves that the candidate dual certificate Y P obeys the first condition in (50).
Since P = N/M , this implies that the number N of output channels from the analog-vector-matrix multiplier
in Figure 6 must be a factor of roughly log W compared to the input channels, i.e.,
N ≥ CM log W.
5 We
assume that N/M is an integer–this can be ensured in the worst case by doubling N .
23
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(58)
However, we believe this requirement is merely an artifact of using golfing scheme as the proof strategy for
Theorem 1. In practice, all our simulations point to N = M , that is, the number of channels at the output
of the AVMM are equal to the input channels.
PP
From the iterative construction (55), it is clear that Y P = − p=1 A∗p Ap (W p−1 ). We will now converge on
showing that Y P satisfies the second condition in (50). Begin with
kPT ⊥ (Y P )k ≤
P
X
p=1
PT ⊥ (A∗p Ap (W p−1 )) =
P
X
p=1
PT ⊥ A∗p Ap (W p−1 ) − W p−1
,
where the last equality follows from the fact that W p−1 ∈ T . Since kPT ⊥ k ≤ 1, we have
P
X
p=1
PT ⊥ A∗p Ap (W p−1 ) − W p−1
≤
P
X
A∗p Ap (W p−1 ) − W p−1 ≤
p=1
P
X
2−p−1 < 1/2,
p=1
the second last inequality above requires kA∗p Ap (W p−1 ) − W p−1 k ≤ 2−p−1 , for every p ∈ [P ], which
using Lemma 4 is only true when Ω ≥ Cβ ν 2 (R/M )W log2 W with probability at least 1 − O(P W −β ) ≥
1 − O(W −β+1 ), where the factor P comes from the union bound over every p ∈ [P ]. Lemma 4
Combining sample complexities in (59), and (62), and using the definition of ϑ2 in (25) gives us the proof of
Theorem 1.
5.4
Key Lemmas
We now state the key lemmas to prove Theorem 1.
Lemma 2. Fix β ≥ 1. Assume that
Ω ≥ Cβ max(µ2 , ρ2 )
R
W log2 W,
M
(59)
where Cβ is a universal constant only depending on β. Then the linear operator Ap obeys
PT A∗p Ap PT − PT ≤
1
2
with probability at least 1 − O(W −β ).
Proof of this lemma will be presented in Section 6.2.
R
Corollary 2. Fix β ≥ 1. Assume Ω ≥ Cβ max(µ2 , ρ2 ) M
W log2 W , where Cβ is a universal constant that
only depends on β. Then the linear operator A defined in (20) obeys kPT A∗ APT −PT k ≤ 2−1 with probability
at least 1 − O(W −β ).
Proof. Proof of this corollary follows exactly the same steps as the proof of Lemma 2 with only difference
being that we take P = N/M = 1.
Lemma 3. Define coherence νp2 of the iterates W p as
νp2 :=
MΩ
max max ke∗ W p I B` k2F .
R m∈[M ] `∈[Ω] m
(60)
Then under the same conditions as in (59), we have νp ≤ (1/2)νp−1 with probability at least 1 − O(W −β ).
Proof. The proof of this lemma follows similar techniques and matrix Bernstein inequality as used in Lemma
2. Similar results can be found in [41]. We skip the proof due to space constraints.
24
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Using the definition of ν 2 in (12), and the fact that W 0 = −U V ∗ , we can see that ν02 = ν 2 . Invoking Lemma
3 for every p ∈ [P ], we can iteratively conclude that
νP ≤ 2−P ν
(61)
with probability at least 1 − O(P W −β ) ≥ 1 − O(W 1−β ).
Lemma 4. Fix β ≥ 1. Take
Ω ≥ Cβν 2
R
W log2 W
M
(62)
for a sufficiently large constant C. Let W p−1 be a fixed M × W matrix defined in (56) then
A∗p Ap (W p−1 ) − W p−1 ≤ 2−p−1
with probability at least 1 − O(W −β ).
Proof of this lemma will be presented in Section 6.3
References
[1] M. Fazel, “Matrix rank minimization with applications,” Ph.D. dissertation, Stanford University, March
2002.
[2] B. Recht, M. Fazel, and P. Parrilo, “Guaranteed minimum-rank solutions of linear matrix equations via
nuclear norm minimization,” SIAM Review, vol. 52, no. 3, pp. 471–501, 2010.
[3] E. Candès and B. Recht, “Exact matrix completion via convex optimization,” Found. Comput. Math.,
vol. 9, no. 6, pp. 717–772, 2009.
[4] D. Gross, “Recovering low-rank matrices from few coefficients in any basis,” IEEE Trans. Inform.
Theory, vol. 57, no. 3, pp. 1548–1566, 2011.
[5] I. F. Gorodnitsky and B. D. Rao, “Sparse signal reconstruction from limited data using FOCUSS: a
re-weighted minimum norm algorithm,” IEEE Trans. Sig. Process., vol. 45, no. 3, pp. 600–616, 1997.
[6] J. J. Fuchs, “Multipath time-delay detection and estimation,” IEEE Trans. Sig. Process., vol. 47, pp.
237–243, 1999.
[7] ——, “On the application of the global matched filter to doa estimation with uniform circular arrays,”
IEEE Trans. Signal Process., vol. 49, no. 4, pp. 702–709, April 2001.
[8] E. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from
highly incomplete frequency information,” IEEE Trans. Inform. Theory, vol. 52, no. 2, pp. 489–509,
February 2006.
[9] S. Kunis and H. Rauhut, “Random sampling of sparse trigonometric polynomials,” Appl. Comp. Harmon. Analysis, vol. 22, 2006.
[10] M. Rudelson and R. Vershynin, “On sparse reconstruction from Fourier and Gaussian measurements,”
Comm. Pure Appl. Math., vol. 61, no. 8, pp. 1025–1045, 2008.
[11] M. F. Duarte and R. G. Baraniuk, “Spectral compressive sensing,” Appl. Comp. Harm. Analysis, vol. 35,
no. 1, pp. 111–129, July 2013.
[12] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht, “Compressed sensing off the grid,” IEEE Trans. Inform.
Theory, vol. 59, no. 11, pp. 7465–7490, 2013.
25
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
[13] E. J. Candès and C. Fernandez-Granda, “Towards a mathematical theory of super-resolution,” Comm.
Pure Appl. Math., vol. 67, no. 6, pp. 906–956, June 2014.
[14] Ali Ahmed and Justin Romberg, “Compressive multiplexing of correlated signals,” IEEE Trans. Inform.
Theory, vol. 1, pp. 479–498, 2015.
[15] A. Hormati, O. Roy, Y. M. Lu, and M. Vetterli, “Distributed sampling of signals linked by sparse
filtering: Theory and applications,” IEEE Trans. Sig. Process., vol. 58, no. 3, pp. 1095–1109, 2010.
[16] D. Baron, M. F. Duarte, M. B. Wakin, S. Sarvotham, and R. G. Baraniuk, “Distributed compressive
sensing,” arXiv preprint arXiv:0901.3403, 2009.
[17] M. Mishali and Y. Eldar and O. Dounaevsky and E. Shoshan, “Xampling: Analog to digital at subNyquist rates,” IET Circuits Devices Syst., vol. 5, no. 1, pp. 8–20, 2011.
[18] M. Mishali and Y. Eldar, “Blind multiband signal reconstruction: Compressed sensing for analog signals,” IEEE Trans. Sig. Process., vol. 57, no. 3, pp. 993–1009, 2009.
[19] M. Mishali, Y. C. Eldar, and A. J. Elron, “Xampling: Signal acquisition and processing in union of
subspaces,” IEEE Trans. Sig. Process., vol. 59, no. 10, pp. 4719–4734, 2011.
[20] W. Mantzel and J. Romberg, “Compressed subspace matching on the continuum,” arXiv preprint
arXiv:1407.5234, 2014.
[21] H. Malvar and D. Staelin, “The LOT: Transform coding without blocking effects,” IEEE Trans. Acoust.,
Speech, Signal Process., vol. 37, pp. 553–559, April 1989.
[22] M. S. Asif and J. Romberg, “Sparse recovery of streaming signals using `1 -homotopy,” IEEE Trans.
Sig. Process., vol. 62, no. 16, pp. 4209–4223, 2014.
[23] R. Schmidt, “Multiple emitter location and signal parameter estimation,” IEEE Trans. Antennas
Propag., vol. 34, no. 3, pp. 276–280, 1986.
[24] R. Roy and T. Kailath, “Esprit-estimation of signal parameters via rotational invariance techniques,”
IEEE Trans. Acoust., Speech, Signal Process., vol. 37, no. 7, pp. 984–995, 1989.
[25] D. Slepian, “On bandwidth,” Proceedings of the IEEE, vol. 64, no. 3, pp. 292–300, March 1976.
[26] ——, “Prolate spheroidal wave functions, fourier analysis and uncertainty v — the discete case,” Bell
Systems Tech. Journal, vol. 57, pp. 1371–1430, 1978.
[27] Schlottmann, C.R. and Hasler, P.E., “A highly dense, low power, programmable analog vector-matrix
multiplier: The fpaa implementation,” IEEE J. Emerg. Sel. Topic Circuits Sys., vol. 1, no. 3, pp.
403–411, 2011.
[28] Chawla, R. and Bandyopadhyay, A. and Srinivasan, V. and Hasler, P., “A 531 nw/mhz, 128× 32 currentmode programmable analog vector-matrix multiplier with over two decades of linearity,” in Proc. IEEE
Conf. Custom Integr. Circuits., 2004, pp. 651–654.
[29] J. Tropp and J. Laska and M. Duarte and J. Romberg, and R. Baraniuk, “Beyond Nyquist: Efficient
sampling of sparse bandlimited signals,” IEEE Trans. Inform. Theory, vol. 56, no. 1, pp. 520–544, 2010.
[30] J. Laska and S. Kirilos and M. Duarte and T. Raghed and R. Baraniuk and Y. Massoud, “Theory and
implementation of an analog-to-information converter using random demodulation,” in Proc. IEEE Int.
Symp. Circuits Syst., 2007, pp. 1959–1962.
[31] J. Yoo, S. Becker, M. Loh, M. Monge, and E. Candès, “A 100MHz-2GHz 12.5x sub-Nyquist rate receiver
in 90nm CMOS,” in Proc. IEEE Radio Freq. Integr. Circuits Symp. (RFIC), 2012.
26
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
[32] J. Yoo, C. Turnes, E. Nakamura, C. Le, S. Becker, E. Sovero, M. Wakin, M. Grant, J. Romberg,
A. Emami-Neyestanak, and E. Candès, “A compressed sensing parameter extraction platform for radar
pulse signal acquisition,” Submitted to IEEE J. Emerg. Sel. Topics Circuits Syst., February 2012.
[33] T. Murray, P. Pouliquen, A. Andreou, and K. Lauritzen, “Design of a CMOS A2I data converter: Theory,
architecture and implementation,” in Proc. IEEE Annu. Conf. Inform. Sci. Syst. (CISS), Baltimore,
MD, 2011, pp. 1–6.
[34] J. Romberg, “Compressive sensing by random convolution,” SIAM J. Imag. Sci., vol. 2, no. 4, pp.
1098–1128, 2009.
[35] J. Haupt and W. Bajwa and G. Raz and R. Nowak, “Toeplitz compressed sensing matrices with applications to sparse channel estimation,” IEEE Trans. Inform. Theory, vol. 56, no. 11, pp. 5862–5875,
2010.
[36] H. Rauhut and J. Romberg and J. Tropp, “Restricted isometries for partial random circulant matrices,”
Appl. Comput. Harmonic Anal., vol. 32, no. 2, pp. 242–254, 2012.
[37] J. Tropp and M. Wakin and M. Duarte and D. Baron and R. Baraniuk, “Random filters for compressive
sampling and reconstruction,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP),
Toulouse, France, 2006.
[38] B. Recht, “A simpler approach to matrix completion,” J. Mach. Learn. Res., vol. 12, pp. 3413–3430,
2011.
[39] E. Candès and Y. Plan, “Matrix completion with noise,” Proc. IEEE, vol. 98, no. 6, pp. 925–936, 2010.
[40] K. Mohan and M. Fazel, “New restricted isometry results for noisy low-rank recovery,” in Proc. IEEE
Int. Symp. Inform. Theory (ISIT), Austin, Texas, June 2010.
[41] A. Ahmed and B. Recht and J. Romberg, “Blind deconvolution using convex programming,” IEEE
Trans. Inform. Theory, vol. 60, no. 3, pp. 1711–1732, 2014.
[42] V. Koltchinskii, K. Lounici, and A. Tsybakov, “Nuclear-norm penalization and optimal rates for noisy
low-rank matrix completion,” Ann. Stat., vol. 39, no. 5, pp. 2302–2329, 2011.
[43] M. Fazel and E. Candès and B. Recht and P. Parrilo, “Compressed sensing and robust recovery of low
rank matrices,” in Proc. 42nd IEEE Asilomar Conf. Signals, Syst. Comput., Pacific Grove, CA, 2008,
pp. 1043–1047.
[44] B. Laurent and P. Massart, “Adaptive estimation of a quadratic functional by model selection,” Ann.
Stat., pp. 1302–1338, 2000.
[45] M. Ledoux, The concentration of measure phenomenon.
AMS, 2001, vol. 89.
[46] J. Tropp, “User-friendly tail bounds for sums of random matrices,” Found. Comput. Math., vol. 12,
no. 4, pp. 389–434, 2012.
[47] Y. C. Eldar and G. Kutyniok, Compressed sensing: theory and applications.
Press, 2012.
27
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
Cambridge University
6
Appendix: Proof of Key Lemmas
Proof of all the key lemmas mainly relies on using matrix Bernstein inequality to control the operator norms
of sums of random matrices.
6.1
Matrix Bernstein-type inequality
We will use a specialized version of the matrix Bernstein-type inequality [42, 46] that depends on the Orlicz
norms. The Orlicz norm of a random matrix Z is defined as
kZkψα = inf{u > 0 : E exp(
kZkα
) ≤ 2},
uα
α ≥ 1.
(63)
Suppose that, for some constant Uα > 0, kZ q kψα ≤ U(α) , q = 1, . . . , Q then the following proposition holds.
Proposition 2. Let Z 1 , Z 2 , . . . , Z Q be iid random matrices with dimensions M × N that satisfy E(Z q ) = 0.
Suppose that kZkψα < ∞ for some α ≥ 1. Define
1/2
1/2
Q
Q
X
X
(64)
,
(E Z ∗q Z q )
σZ = max
(E Z q Z ∗q )
q=1
q=1
Then ∃ a constant C > 0 such that , for all t > 0, with probability at least 1 − e−t
p
QUα2
kZ 1 + · · · + Z Q k ≤ C max σZ t + log(M + N ), Uα log1/α
(t
+
log(M
+
N
))
2
σZ
6.2
(65)
Proof of Lemma 2
We start by writing PT A∗p Ap PT as a sum of independent random matrices using (54) to obtain
PT A∗p Ap PT =
Using (53), and the fact that
P
`∈[Ω]
X X
n∈Γp `∈[Ω]
P [PT (an d∗n` ) PT (an d∗n` )] .
E dn` d∗n` = I W , the expectation of the quantity above evaluates to
X X
X
X
E PT A∗p Ap PT = PT P
E(an a∗n ⊗ dn` d∗n` ) PT = PT P
an a∗n ⊗
E dn` d∗n` PT = PT .
n∈Γp `∈[Ω]
n∈Γp
`∈[Ω]
The quantity PT A∗p Ap PT − PT can therefore be expressed as a sum of independent zero mean random
matrices in the following form
X X
PT A∗p Ap PT − PT =
P [PT (an d∗n` ) PT (an d∗n` ) − E PT (an d∗n` ) PT (an d∗n` )] .
n∈Γp `∈[Ω]
We will employ matrix Bernstein inequality to control the operator norm of the above sum. To proceed
define the operator Zn` which maps Z to hPT (an d∗n` ), Zi PT (an d∗n` ), i.e., Zn` = [PT (an d∗n` ) PT (an d∗n` )].
2
∗
This operator
is rank one, therefore,
P
P Pthe operator norm kZn` k = kPT (an dn` )kF . To ease the notation, we
will use
as a shorthand for
. We begin by computing the variance in (64) as follows
n∈Γp `∈[Ω]
n,`
σ2 = P 2
X
n,`
E(Zn` − E Zn` )2 = P 2
X
n,`
2
[E Zn`
− (E Zn` )2 ] ≤ P 2
28
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
X
n,`
2
,
E Zn,`
2
where the last inequality follows from the fact that E(Zn`
), and (E Zn` )2 are symmetric, and positive2
semidefinite matrices. The square of the rank-1 matrices Zn` is simply given by Zn`
= kPT (an d∗n` )k2F Zn` .
Now we develop the operator norm of the result simplified expression using (48)
X W
X
1
2
∗
∗
∗
2
2
E
kPT (an dn` )kF Zn` ≤
kU an k2 + kV dn` k2 Zn` .
Ω
P
n,`
n,`
Using the definition in (12), and (53), we can bound kU ∗ an k22 ≤ µ2 R/P M . Using this fact, we have
X
X
X
∗
∗
2
2 RW
2
2 RW
2
E kV dn` k2 Zn` ≤ µ
σ ≤ Pµ
E Zn` + P
E kV dn` k2 Zn` .
+P
MΩ
MΩ
n,`
n,`
(66)
n,`
The second term in (66) can be simplified as
X
X
E kV ∗ dn` k22 (an d∗n` an d∗n` ) PT
E kV ∗ dn` k22 Zn` = PT
n,`
n,`
≤
X
n,`
E kV ∗ dn` k22 (an d∗n` an d∗n` )
,
(67)
where the last inequality follows form the fact that kPT k ≤ 1. Since an d∗n` an d∗n` = an a∗n ⊗ dn,` d∗n,` , and
a simple calculation reveals the expectation
E kV ∗ dn` k22 (dn` d∗n` ) = kV ∗ I B` k2F I B` + 2I B` V V ∗ I B` − 2diag(I B` V V ∗ I B` ) 4 3kV ∗ I B` k2F
where for diag(X) is the diagonal matrix obtained by setting the off-diagonal entries of X to zero, and I B`
denotes the W × W identity matrix with ones only at the diagonal positions indexed by B` . This directly
implies that
X
X
X
∗
∗
∗
∗
2
∗
2
3kV I B` kF I B` ⊗
an an
E kV dn` k2 (an dn` an dn` ) ≤
n,`
`∈[Ω]
n∈Γp
R
3
≤ max kV ∗ I B` k2F = 3ρ2
,
P `∈[Ω]
PΩ
(68)
where the last equality follows from the definition of the coherence in (12). Plugging (68) in (66), we have
the bound
2
2 RW
2R
σ ≤C µ
+ρ
.
(69)
MΩ
Ω
Finally, we calculate the Orlicz norm, the last ingredient to obtain the Bernstein bound. First, it is important
to see that
P kZn` − E Zn` k ≤ 2P kZn` k = 2P kZn` kF = 2P kPT (an d∗n` )k2F ,
where the second-last equality follows form the fact that Zn` is the rank-1 operator. Using the last equation,
and (48), we have
W
1
kU ∗ a∗n k22 + kV ∗ dn` k22
Ω
P
1
∗
2
2 RW
2R
2 R W
≤ CP µ
+ max kV I B` kF ≤ C µ
+ρ
PM Ω
P `∈[Ω]
MΩ
Ω
U1 = max max 2P kPT (an d∗n` )k2F
n∈[N ] `∈[Ω]
ψ1
≤ 2P max max
n∈[N ] `∈[Ω]
ψ1
(70)
Moreover, a simple calculation, and using the facts that µ2 ≤ M/R, and ρ2 ≤ W/R shows that log(M Ω ·
U12 /σ 2 ) ≤ C log(W M ). Using this together with (69), and (70), and using t = β log(W M ) in the Bernstein’s
inequality in Proposition 2, we have
s
RW
R p
RW
R
∗
kPT Ap Ap PT − PT k ≤ C max
+ ρ2
β log(W M ),
+ ρ2
β log2 (W M )
MΩ
Ω
MΩ
Ω
We can conclude now that choosing Ω ≥ Cβ (µ2 R(W/M )+ρ2 R) log2 (W M ) ensures that kPT A∗p Ap PT −PT k ≤
1/2, which proves the lemma after using the fact that W ≥ M .
29
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
6.3
Proof of Lemma 4
Just as in the proof of Lemma 2, we will start with writing the A∗p Ap (W p−1 ) as a sum of independent
random matrices using (47) as follows
X X
A∗p Ap (W p−1 ) =
P [an a∗n W p−1 dn` d∗n` ] .
n∈Γp `∈[Ω]
Recall that dn` are random binary defined earlier. Then the expectation of the random quantity above is
X X
X X
E A∗p Ap (W p−1 ) =
P E [an a∗n W p−1 d`n d∗`n ] =
P [an a∗n W p−1 I B` ] = W p−1 ,
n∈Γp `∈[Ω]
n∈Γp `∈[Ω]
where the last two equalities follow from the fact that
X
an a∗n =
n∈Γp
1
IM ,
P
X
and
E dn` d∗n` =
`∈[Ω]
X
I B` = I W .
(71)
`∈[Ω]
We bound the operator norm kA∗p Ap (W p−1 ) − W p−1 k. In light of discussion above, A∗p Ap (W p−1 ) − W p−1
can be expressed as a following sum of independent, and zero mean random matrices
X
A∗p Ap (W p−1 ) − W p−1 =
P [an a∗n W p−1 dn` d∗n` − E an a∗n W p−1 dn` d∗n` ] ,
n,`
where
P
n,`
P P
is a shorthand for
n∈Γp `∈[Ω]
. Define Z n` := P [an a∗n W p−1 dn` d∗n` − E an a∗n W p−1 dn` d∗n` ] . To
compute the variance in (64), we start with
X
X
∗
∗
∗
2
∗
∗
W
E Z n` Z ∗n` = P 2
E an a∗n W p−1 dn` d∗n` ,
Ω E (an W p−1 dn` ) an an − E an an W p−1 dn` dn`
n,`
n,`
where we used the fact that kdn` k22 = W/Ω. Since E Z n` Z ∗n` is a symmetric positive-semidefinite matrix,
that is, E Z n` Z ∗n` < 0, this together with definition of Z n` implies that
X
∗
X
∗
2
∗
W
<
E an a∗n W p−1 dn` d∗n` E an a∗n W p−1 dn` d∗n` ,
Ω E (an W p−1 dn` ) an an
n,`
n,`
and, therefore,
X
n,`
E Z n` Z ∗n` ≤ P 2
W X
W X ∗
E (a∗n W p−1 dn` )2 an a∗n ≤ P 2
kan W p−1 I B` k22 an a∗n
Ω
Ω
≤ PW
n,`
n,`
max max ka∗n W p−1 I B` k22
n∈[N ] `∈[Ω]
= W max max ke∗m W p−1 I B` k22
m∈[M ] `∈[Ω]
W/M
2
≤ νp−1
R
,
Ω
(72)
2
where the inequalities follow by using (71), the definition of coherence νp−1
in (60), and (53).
For the second variance term in (64), we skip through similar step as for the first term and land directly at
X X
n∈Γp `∈[Ω]
E Z n` Z ∗n` ≤
X X
n∈Γp `∈[Ω]
=P
P 2 kan k22 E (a∗n W p−1 dn` )2 dn` d∗n`
X X
E (a∗n W p−1 dn` )2 dn` d∗n` ,
n∈Γp `∈[Ω]
30
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
(73)
where the last equality is the result of (53). One can show that for a fixed vector x ∈ RW , and the fact
that dn` is a vector with independent Rademacher random variables at locations indexed by B` , and zero
elsewhere, the following
E (x∗ dn` )2 dn` d∗n` ≤ kxB` k22 I B` + 2xB` x∗B` 4 3kxB` k22 I B`
(74)
holds, where xB` is equal to x on B` , and zero elsewhere. Moreover, I B` is a diagonal matrix with ones at
B` , and zero elsewhere. Using (74) with x∗ = a∗n W p−1 , we have
X X
n∈Γp `∈[Ω]
≤ 3P
E Z n` Z ∗n` ≤ 3P
X
n∈Γp
X X
n∈Γp `∈[Ω]
ka∗n W p−1 I B` k22 I B`
2
max ka∗n W p−1 I B` k22 ≤ 3P M max max ka∗n W p−1 I B` k22 ≤ 3νp−1
n∈Γp `∈[Ω]
`∈[Ω]
R
,
Ω
(75)
2
where in the last inequality, we use the definition of νp−1
in (61) combined with (53). In light of (64), the
maximum of (72), and (75) accounts for the variance σ 2
2
σ 2 ≤ 3νp−1
R
(W/M )
(W/M ) + 1
2
≤ 6νp−1
R
,
Ω
Ω
(76)
where in the last inequality follows from our assumption that W ≥ M . Finally, we need to compute an
upper bound on the Orlicz norm of the random variable kZ n` k. Begin by using similar simple facts above
that
r
1
W ∗
∗
∗
∗
kZ n` k ≤ 2P kan an W p−1 dn` dn` k = 2P kan k2 kdn` k2 |an W p−1 dn` | = 2P √
|a W p−1 dn` |.
(77)
Ω n
P
Using standard calculations; see, for example, [47], we can compute the following finite bound on the Orlicz-1
norm of the random variable |a∗n W p−1 dn` |
max max ka∗n W p−1 dn` kψ2 = max max
n∈Γp `∈[Ω]
m∈[M ] `∈[Ω]
1 ∗
ke W p−1 dn` kψ2
P m
C
max max ke∗ W p−1 I B` k2 ≤ Cνp−1
≤
P n∈Γp `∈[Ω] m
r
R
,
PMΩ
where the last inequality follows from (60). Using (77), P = N/M , and (63) then directly gives us
2
U22 := max max kZ n` k2ψ2 ≤ Cνp−1
R
n∈Γp `∈[Ω]
W/M
.
Ω2
(78)
2
Moreover using a loose bound on variance σ 2 ≤ 3νp−1
R((W/M ) + 1)/Ω, it is easy to see that
log
ΩM · U22
≤ C log M.
σ2
The results in (72), (75), and (78) can be plugged in Proposition 2 to obtain
(r
)
r
W/M p
W/M
3/2
∗
2
2
kAp Ap (W p−1 ) − W p−1 kF ≤ C max
νp−1 R
β log(W M ), νp−1 R
β log (W M )
Ω
Ω2
(79)
with t = (β − 1) log(W M ), which holds with probability at least 1 − O(W M )−β . Recall that W ≥ M . The
2
lemma now follows by using the bound on νp−1
in (61), and choosing Ω ≥ Cβ Rν 2 (W/M ) log3/2 W for a
universal constant Cβ that only depends on a fixed parameter β ≥ 1.
31
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
7
Proof of Theorem 2
The first step in the proof is the following Oracle inequality in [42] that gives an upper bound on the deviation
c in (30) from the true solution X 0 in the mean squared sense.
of X
Theorem 4 (Oracle inequlaity in [42]). Suppose we observe the noisy measurements y in (22) of X 0 with
c
rank(X 0 ) ≤ R, and it is given that kA∗ (y) − E A∗ (y)k ≤ λ/2 fro some scalar λ ≥ 0. Then the solution X
2
2
c
of the nuclear norm penalized estimator in (29) obeys kX − X 0 kF ≤ min 2λkX 0 k∗ , 1.46λ R .
All that is required is to bound the spectral norm:
kA∗ (y) − E A∗ (y)k ≤ kA∗ A(X 0 ) − X 0 k + kA∗ (ξ)k
(80)
We begin by bounding the first term above kA∗ A(X 0 ) − X 0 k using a corollary to Lemma 4 stated as follows.
Corollary 3. Let X 0 be a fixed M × W matrix defined in 17 then
(r
)
r
W/M p
W/M
3/2
∗
2
2
ν
β log W , ν
β log W .
kA A(X 0 ) − X 0 k ≤ CkX 0 kF max
Ω
Ω2
(81)
with probability at least 1 − O(W −β ).
Proof. The proof of the corollary is very similar to the proof of Lemma 4— the main difference is that
the number of partitions is P = N/M √
= 1. Moreover, we have X 0 in place of W p−1 , and in the proof
development replace kW p−1 kF ≤ 2−p+1 R with kX 0 kF to obtain bound in (81), which is understandably
similar to (79)
Lemma 5. Fix β ≥ 1. The for a sufficiently large constant C, the following bound
r
W/M p
∗
kA (ξ)k ≤ Ckξkψ2
β log W
Ω
holds with probability at least 1 − O(W −β ).
Using Corollary 3, and Lemma 5, we can bound (80), and obtain
s
q
(W/M )(ν 2 kX 0 k2F + kξk2ψ2 )
β log3/2 W
λ≥C
Ω
with probability at least 1−O(W −β ). Taking kX 0 kF = 1 without loss of generality, and Ω ≥ Cβ ν 2 R W
log3/2 W ,
qM
where Cβ is a universal constant that depends on a fixed parameter β ≥ 1, allows us to choose λ ≥ kξk2ψ2 /R.
With this, an application of Theorem 4 proves Theorem 2.
7.1
Proof of Lemma 5
The proof of this lemma requires the
P use of matrix Bernstein’s inequality 2. As it is required to bound the
spectral norm of the sum A∗ (ξ) = n,` ξn [`]an d∗n` , we start with the summands Z n` = ξn [`]an d∗n` . Because
variables ξn` are zero mean, it follows that E Z n` = 0. We start by computing the variance
X
n,`
E Z n` Z ∗n` =
W X
W
E |ξn [`]|2 · an a∗n =
Ω
Ω
n,`
max
n∈[N ]
X
`∈[Ω]
E |ξn [`]|2 ·
X
n∈[N ]
an a∗n ≤
32
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
W
kξk2ψ2 , (82)
NΩ
P
where the last inequality follows from the facts that n∈[N ] an a∗n = I M , and that ξn [`] for (n, `) ∈ [N ] × [Ω]
P
are independent and identically distributed implying maxn∈[N ] `∈[Ω] E |ξn [`]|2 ≤ Ckξk2ψ2 /N .
Similarly arguments lead to
X
n,`
E Z n` Z ∗n` =
X
n,`
kan k22 E |ξn [`]|2 E dn` d∗n` =
M X
M
E |ξn [`]|2 I B` ≤
kξk2ψ2 .
N
NΩ
(83)
n,`
Combining (82) and (83) and using (64) gives σ 2 ≤ kξk2ψ2 P W/M Ω, where P = N/M , and we assume that
W ≥ M . The final quantity required is the Orlicz norm of Z n` , which is simply
kZ n` k2ψ2 = kξn [`]k2ψ2 kan d∗n` k2 = C
then
kZ n` kψ2 log
1/2
M Ω · kZ n` k2ψ2
σ2
1
MW
1 W
kξk2ψ2 ·
= Ckξk2ψ2 2
NΩ
NΩ
P M Ω2
!
r
≤C
kξk2ψ2
W/M
log1/2 (M W ).
P 2 Ω2
At the end, using t = β log W in the Bernstein’s bound (65), we have
(
)
r
r
W/M p
W/M
3/2
∗
kA (ξ)k ≤ C max kξkψ2
β log(M W ), kξkψ2
(β log (M W )) ,
Ω
P 2 Ω2
and using the fact that P = O(log W ), and M ≤ W from (57) proves the result.
33
Draft by A. Ahmed and J. Romberg – December 18, 2017 – 2:46
| 7cs.IT
|
White matter fiber segmentation using
functional varifolds
Kuldeep Kumar1,2 , Pietro Gori2,4 , Benjamin Charlier2,5 , Stanley Durrleman2 ,
Olivier Colliot2,3 and Christian Desrosiers1
arXiv:1709.06144v1 [cs.CV] 18 Sep 2017
1
LIVIA, École de technologie supérieure (ÉTS), Montreal, Canada
Aramis project-team, Inria Paris, Sorbonne Universités, UPMC Univ Paris 06,
Inserm, CNRS, Institut du cerveau et la moelle (ICM) - Hôpital Pitié-Salpêtrière,
Boulevard de l0 hôpital, F-75013, Paris, France
3
AP-HP, Departments of Neurology and Neuroradiology, Hôpital Pitié-Salpêtrière,
75013, Paris, France
4
LTCI lab - IMAGES group, Télécom ParisTech, Paris, France
5
Université de Montpellier, France
2
Abstract. The extraction of fibers from dMRI data typically produces a
large number of fibers, it is common to group fibers into bundles. To this
end, many specialized distance measures, such as MCP, have been used
for fiber similarity. However, these distance based approaches require
point-wise correspondence and focus only on the geometry of the fibers.
Recent publications have highlighted that using microstructure measures
along fibers improves tractography analysis. Also, many neurodegenerative diseases impacting white matter require the study of microstructure
measures as well as the white matter geometry. Motivated by these, we
propose to use a novel computational model for fibers, called functional
varifolds, characterized by a metric that considers both the geometry
and microstructure measure (e.g. GFA) along the fiber pathway. We use
it to cluster fibers with a dictionary learning and sparse coding-based
framework, and present a preliminary analysis using HCP data.
1
Introduction
Recent advances in diffusion magnetic resonance imaging (dMRI) analysis have
led to the development of powerful techniques for the non-invasive investigation
of white matter connectivity in the human brain. By measuring the diffusion
of water molecules along white matter fibers, dMRI can help identify connection pathways in the brain and better understand neurological diseases related
to white matter [6]. Since the extraction of fibers from dMRI data, known as
tractography, typically produces a large number of fibers, it is common to group
these fibers into larger clusters called bundles. Clustering fibers is also essential
for the creation of white matter atlases, visualization, and statistical analysis of
microstructure measures along tracts [12].
Most fiber clustering methods use specialized distance measures, such as
Mean Closest Points (MCP) distance [4,11]. However, these distance-based approaches require point-wise correspondence between fibers and only consider
fiber geometry. Another important aspect for white matter characterization
is the statistical analysis of microstructure measures. As highlighted in recent
publications, using microstructure measures along fibers improves tractographic
analysis [3,10,12,15,16,17]. Motivated by these, we propose to use a novel computational model for fibers, called functional varifolds, characterized by a metric
that considers both the geometry and microstructure measure (e.g. generalized
fractional anisotropy) along fiber pathways.
Motivation for this work comes from the fact that the integrity of white
matter is an important factor underlying many cognitive and neurological disorders. In vivo, tissue properties may vary along each tract for several reasons:
different populations of axons enter and exit the tract, and disease can strike at
local positions within the tract. Hence, understanding diffusion measures along
each fiber tract (i.e., tract profile) may reveal new insights into white matter
organization, function, and disease that are not obvious from mean measures of
that tract or from the tract geometry alone [3,17]. Recently, many approaches
have been proposed for tract based morphometry [12], which perform statistical analysis of microstructure measures along major tracts after establishing
fiber correspondences. While studies highlight the importance of microstructure
measures, most approaches either consider the geometry or signal along tracts,
but not both. The intuitive approach would be to consider microstructure signal
during clustering also. However, this has been elusive due to lack of appropriate
framework.
As a potential solution, we explore a novel computational model for fibers,
called functional varifolds [1], which is a generalization of the varifolds framework
[2]. The advantages of using functional varifolds are as follows. First, functional
varifolds can model the fiber geometry as well as signal along the fibers. Also,
it does not require pointwise correspondences between fibers. Lastly, fibers do
not need to have the same orientation as in the framework of currents [5]. We
test the impact of this new computational model on a fiber clustering task, and
compare its performance against existing approaches for this task.
As clustering method, we reformulate the dictionary learning and sparse coding based framework proposed in [8,7,9]. This choice of framework is driven by
its ability to describe the entire data-set of fibers in a compact dictionary of
prototypes. Bundles are encoded as sparse non-negative combinations of multiple dictionary prototypes. This alleviates the need for explicit representation
of a bundle centroid, which may not be defined or may not represent an actual
object. Also, sparse coding allows assigning single fibers to multiple bundles,
thus providing a soft clustering.
The contributions of this paper are threefold: 1) a novel computational model
for modeling both fiber geometry and signal along fibers, 2) a generalized clustering framework, based on dictionary learning and sparse coding, adapted to the
computational models, and 3) a comprehensive comparison of fully-unsupervised
models for clustering fibers.
2
2.1
White matter fiber segmentation using functional
varifolds
Modeling fibers using functional varifolds
In the framework of functional varifolds [1,2], a fiber X is assumed to be a
polygonal line of P segments described by their center point xp ∈ R3 and tangent
vector βp ∈ R3 centered at xp and of length cp (respectively, yq ∈ R3 , γq ∈ R3
and dq for a fiber Y with Q segments). Let fp and gp be the signal values
at center points xp and yq respectively, and ω the vector field belonging to a
reproducing kernel Hilbert space (RKHS) W ∗ . Then the fibers X and Y can be
PP
modeled based on functional varifolds as: V(X,f ) (ω) ≈ p=1 ω(xp , βp , fp )cp and
PQ
V(Y,g) (ω) ≈ q=1 ω(yq , γq , gp )dq . More details can be found in [1].
The inner product metric between X and Y is defined as:
hV(X,f ) , V(Y,g) iW ∗ =
Q
P
X
X
κf (fp , gq )κx (xp , yq )κβ (βp , γq )cp dq
(1)
p=1 q=1
where κf and κx are Gaussian kernels and κβ is a Cauchy-Binet kernel. This can
be re-written as:
Q
P
−kx − y k2 β T γ 2
−kf − g k2
X
X
p
q
p
q
p
q
exp
cp dq
hV(X,f ) , V(Y,g) iW ∗ =
exp
2
2
λ
λ
c
p dq
M
W
p=1 q=1
(2)
where λM and λW are kernel bandwidth parameters. For varifolds [2], a computational model using only fiber geometry and used for comparison in the
experiments, we drop the signal values at center
PPpoints. Thus, the varifoldsbased representation of fibers will be: VX (ω) ≈ p=1 ω(xp , βp )cp and VY (ω) ≈
PQ
ω(yq , γq )dq . Hence, the inner product is defined as:
q=1
hVX , VY iW ∗ =
Q
P
X
X
p=1 q=1
2.2
exp
−kx − y k2 β T γ 2
p
q
p
q
cp dq .
λ2W
cp dq
(3)
Fiber Clustering using Dictionary learning and sparse coding
For fiber clustering, we extend the dictionary learning and sparse coding based
framework presented in [8,7,9]. Let VT be the set of n fibers modeled using
be the atom matrix representing the dictionary
functional varifolds, A ∈ Rn×m
+
coefficients for each fiber belonging to one of the m bundles, and W ∈ Rm×n
+
be the cluster membership matrix containing the sparse codes for each fiber.
Instead of explicitly representing bundle prototypes, each bundle is expressed as
a linear combination of all fibers. The dictionary is then defined as D = VT A.
Since this operation is linear, it is defined for functional varifolds.
The problem of dictionary learning using sparse coding [8,7] can be expressed
as finding the matrix A of m bundle prototypes and the fiber-to-bundle assignment matrix W that minimize the following cost function:
arg min
A,W
1
||VT − VT AW ||2W ∗ ,
2
subject to: ||wi ||0 ≤ Smax .
(4)
Parameter Smax defines the maximum number of non-zero elements in wi (i.e.,
the sparsity level), and is provided by the user as input to the clustering method.
An important advantage of using the above formulation is that the reconstruction error term only requires inner product between the varifolds. Let
Q ∈ Rn×n be the Gram matrix denoting inner product between all pairs of
training fibers, i.e., Qij = hVXi ,fi , VXj ,fj iW ∗ . Matrix Q can be calculated once
and stored for further computations. The problem then reduces to linear algebra
operations involving matrix multiplications. The solution of Eq. (4) is obtained
by alternating between sparse coding and dictionary update [8]. The sparse codes
of each fiber can be updated independently by solving the following sub-problem:
arg min
wi ∈Rm
+
1
||VXi − VT Awi ||2W ∗ ,
2
subject to: ||wi ||0 ≤ Smax .
which can be re-written as:
1
>
Q(i, i) + w>
arg min
i A QAw i − 2Q(i, :)Aw i ,
m
2
wi ∈R+
(5)
s.t.: ||wi ||0 ≤ Smax . (6)
The non-negative weights wi can be obtained using the kernelized Orthogonal
Matching Pursuit (kOMP) approach proposed in [8], where the most positively
correlated atom is selected at each iteration, and the sparse weights ws are
obtained by solving a non-negative regression problem. Note that, since the size
of ws is bounded by Smax , it can be otained rapidly. Also, in case of a large
number of fibers, the Nystrom method can be used for approximating the Gram
matrix [7]. For dictionary update, A is recomputed by applying the following
update scheme, until convergence:
Aij ← Aij
3
(QW > )ij
,
(QAW W > )ij
i = 1, . . . , n,
j = 1, . . . , m.
(7)
Experiments
Data: We evaluate different computational models on the dMRI data of 10 unrelated subjects (6 females and 4 males, age 22-35) from the Human Connectome
Project (HCP) [14]. DSI Studio [18] was used for the signal reconstruction (in
MNI space, 1mm), and streamline tracking employed to generate 50, 000 fibers
per subject (minimum length 50 mm, maximum length 300 mm). Generalized
Fractional Anisotropy (GFA), which extends standard fractional anisotropy to
orientation distribution functions, was considered as along-tract measure of microstructure. While we report results obtained with GFA, any other along-tract
measure may have been used.
Parameter impact: We performed k-means clustering and manually selected
pairs of fibers from clusters most similar to major bundles. We then modeled
these fibers using different computational models, and analyzed the impact of
varying the kernel bandwidth parameters. The range of these parameters were
estimated by observing the values of distance between centers of fiber segments
and difference between along tract GFA values for selected multiple pairs of
fibers. Figure 1 (top left) shows GFA color-coded fibers for 3 pairs corresponding
to a) right Corticospinal tract – CST (R), b) Corpus Callosum – CC, and c) right
Inferior Fronto-Occipital Fasciculus – IFOF (R). Cosine similarity (in degrees) is
reported for the fiber pairs modeled using varifolds (Var) and functional varifolds
(fVar), for λW = 7 mm and λM = 0.01.
Figure 1 (top left) shows GFA color-coded fiber pairs. The color-coded visualization reflect the variation of fiber geometry, microstructure measure (i.e.
GFA) along fiber, and difference in GFA along fiber for the select fiber pairs.
This visualization of variation and difference in GFA values along fibers support
our hypothesis that modeling along tract signal along with geometry provides
additional information. The change in cosine similarity for CC from 45.8 degrees
Fig. 1: Along-fiber GFA visualization and cosine similarity between pairs of fibers
from three prominent bundles: a) CST (R), b) CC, c) IFOF (R), using framework of varifolds (Var) and functional varifolds (fVar) (top left), and Comparing
variation of cosine similarity for the select fiber pairs over kernel bandwidth
parameters λW and λM for the framework of functional varifolds (top right:
CST (R), middle left: CC, middle right: IFOF (R)); Impact of λM on clustering
consistency (measured using Average Silhouette) for m = 100, 125, 150 for functional Varifolds vs Varifolds (bottom left), and functional Varifolds vs GFA only
(bottom right)
(using varifolds) to 66.3 degrees (using functional varifolds) while for CST (R)
from 45.6 degrees to 72.4 degrees, reflect more drop in cosine similarity if along
tract signal profiles are not similar. This shows that functional varifolds imposes
penalty for different along fiber signal profiles.
Figure 1 also compares the impact of varying the kernel bandwidth parameters for functional varifolds using similarity angle between pairs of these selected
fibers (top right: CST (R), bottom left: CC, bottom right: IFOF (R)). We show
variation over λW = 3, 5, 7, 9 and 11 (mm) and λM = 0.001, 0.005, 0.01, 0.05,
and 0.1.
Comparing the parameter variation images in Figure 1 we observe that the
cosine similarity values over the parameter space show similar trends for all 3
pairs of fibers. This observation allows us to select a single pair of parameter
Model m=100 m=125 m=150
fVar
Var
GFA
MCP
0.3624
0.3356
-0.0579
0.3240
0.3451
0.3089
-0.0584
0.2888
0.3314
0.2905
-0.0610
0.2619
Fig. 2: Mean silhouette obtained with Varifolds, Varifolds, GFA, and MCP, computed for varying a number of clusters, over 10 subjects and 3 seed values (left).
Detailed results obtained for 10 subjects using m=100 (right).
values for our experiments. We have used λW = 7 mm and λM = 0.01 for
our experiments based on the cosine similarity values in Figure 1. The smaller
values for λW (< 7mm) and λM (< 0.01mm) will make the current fiber pairs
orthogonal while for larger values we lose the discriminative power as all fiber
pairs will have very high similarity.
Quantitative analysis: We report a quantitative evaluation of clusterings
obtained using as functional varifolds (fVar), varifolds (var), MCP and GFA
computational model. The same dictionary learning and sparse coding framework is applied for all computational models. For each of the 10 HCP subjects,
we compute the Gramian matrix using 5, 000 fibers randomly sampled over the
full brain for 3 seed values. The MCP distance dij is calculated between each
fiber pair (i, j), as described in [4], and the Gramian matrix
obtained using a
radial basis function (RBF) kernel: kij = exp − γ · d2ij . Parameter γ was set
empirically to 0.007 in our experiments.
Since our evaluation is performed in an unsupervised setting, we use the
silhouette measure [11,13] to assess and comparing clustering consistency. Silhouette values, which range from −1 to 1, measure how similar an object is to its
own cluster (cohesion) compared to other clusters (separation). Figure 1 (bottom row) shows impact of λM on clustering consistency for functional Varifolds
w.r.t Varifolds and GFA only. Figure 2 (right) gives the average silhouette for m
= 100, 125, and 150 clusters, computed over 10 subjects and 3 seed values. The
impact of using both geometry and microstructure measures along fibers is evaluated quantitatively by comparing clusterings based on functional varifolds with
those obtained using only geometry (i.e., varifolds, MCP), and only along-fiber
signal (i.e., GFA). As can be seen, using GFA alone leads to poor clusterings, as
reflected by the negative silhouette values. Comparing functional varifolds with
varifolds and GFA, we observe a consistently improved performance for different
numbers of clusters. To further validate this hypothesis, we also report the average silhouette (over 3 seed values) obtained for 10 subjects using m = 100. These
results demonstrate that functional varifolds give consistently better clustering,
compared to other computational models using the same framework1 .
Qualitative visualization: Figure 3 (top row) shows the dictionary learned for
a single subject (m = 100) using functional varifolds (fVar), varifolds (Var), and
MCP distance. For visualization purposes, each fiber is assigned to a single cluster, which is represented using a unique color. The second and third rows of the
1
Silhouette analyzes only clustering consistency, not the along-fiber signal profile.
fVar
Var
MCP
Fig. 3: Full clustering visualization (m = 100, top row), single cluster visualization (mid row), and GFA based color coded visualization of the selected single
cluster (bottom row). Using following computational models for fibers: functional
varifolds (left column), varifolds (middle column), and MCP distance (right column). Superior axial views. Note: (top row) each figure has a unique color code.
figure depict a specific cluster and its corresponding GFA color-coded profiles.
We observe that all three computational models produce plausible clusterings.
From the GFA profiles of the selected cluster (with correspondence across computational models), we observe that functional varifolds enforce both geometric
as well as along-tract signal profile similarity. Moreover, the clustering produced
with varifolds or MCP (i.e., using only geometric properties of fibers), are similar
to one another and noticeably different from that of functional varifolds.
4
Conclusion
A novel computational model, called functional varifolds, was proposed to model
both geometry and microstructure measure along fibers. We considered the task
of fiber clustering and integrated our functional varifolds model within framework based on dictionary learning and sparse coding. The driving hypothesis that
combining along-fiber signal with fiber geometry helps tractography analysis was
validated quantitatively and qualitatively using data from Human Connectome
Project. Results show functional varifolds to yield more consistent clusterings
than GFA, varifolds and MCP. While this study considered a fully unsupervised
setting, further investigation would be required to assess whether functional varifolds augment or aid the reproducibility of results.
Acknowledgements: Data were provided by the Human Connectome
Project.
References
1. Charlier, B., Charon, N., Trouvé, A.: The fshape framework for the variability
analysis of functional shapes. Foundations of Computational Mathematics pp. 1–
71 (2014)
2. Charon, N., Trouvé, A.: The varifold representation of nonoriented shapes for diffeomorphic registration. SIAM Journal on Imaging Sciences 6(4), 2547–2580 (2013)
3. Colby, J.B., Soderberg, L., Lebel, C., Dinov, I.D., Thompson, P.M., Sowell, E.R.:
Along-tract statistics allow for enhanced tractography analysis. Neuroimage 59(4),
3227–3242 (2012)
4. Corouge, I., Gouttard, S., Gerig, G.: Towards a shape model of white matter fiber
bundles using diffusion tensor MRI. In: ISBI 2004. pp. 344–347. IEEE (2004)
5. Gori, P., Colliot, O., Marrakchi-Kacem, L., Worbe, Y., Fallani, F.D.V., Chavez, M.,
Lecomte, S., Poupon, C., Hartmann, A., Ayache, N., et al.: A prototype representation to approximate white matter bundles with weighted currents. In: MICCAI.
pp. 289–296. Springer (2014)
6. Hagmann, P., Jonasson, L., Maeder, P., Thiran, J.P., Wedeen, V.J., Meuli, R.: Understanding diffusion mr imaging techniques: From scalar diffusion-weighted imaging to diffusion tensor imaging and beyond 1. Radiographics 26(suppl 1), S205–
S223 (2006)
7. Kumar, K., Desrosiers, C.: A sparse coding approach for the efficient representation
and segmentation of white matter fibers. In: ISBI, 2016. pp. 915–919. IEEE (2016)
8. Kumar, K., Desrosiers, C., Siddiqi, K.: Brain fiber clustering using non-negative
kernelized matching pursuit. In: Machine Learning in Medical Imaging, LNCS, vol.
9352, pp. 144–152 (2015)
9. Kumar, K., Desrosiers, C., Siddiqi, K., Colliot, O., Toews, M.: Fiberprint: A subject
fingerprint based on sparse code pooling for white matter fiber analysis. NeuroImage 158, 242 – 259 (2017)
10. Maddah, M., Grimson, W.E.L., Warfield, S.K., Wells, W.M.: A unified framework
for clustering and quantitative analysis of white matter fiber tracts. Medical image
analysis 12(2), 191–202 (2008)
11. Moberts, B., Vilanova, A., van Wijk, J.J.: Evaluation of fiber clustering methods
for diffusion tensor imaging. In: VIS 2005. pp. 65–72. IEEE (2005)
12. O’Donnell, L.J., Westin, C.F., Golby, A.J.: Tract-based morphometry for white
matter group analysis. Neuroimage 45(3), 832–844 (2009)
13. Siless, V., Medina, S., Varoquaux, G., Thirion, B.: A comparison of metrics and
algorithms for fiber clustering. In: PRNI, 2013. pp. 190–193. IEEE (2013)
14. Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E., Yacoub, E., Ugurbil,
K., Consortium, W.M.H., et al.: The wu-minn human connectome project: an
overview. Neuroimage 80, 62–79 (2013)
15. Wang, Q., Yap, P.T., Wu, G., Shen, D.: Application of neuroanatomical features
to tractography clustering. Human brain mapping 34(9), 2089–2102 (2013)
16. Wassermann, D., Bloy, L., Kanterakis, E., Verma, R., Deriche, R.: Unsupervised
white matter fiber clustering and tract probability map generation: Applications
of a Gaussian process framework for white matter fibers. NeuroImage 51(1) (2010)
17. Yeatman, J.D., Dougherty, R.F., Myall, N.J., Wandell, B.A., Feldman, H.M.: Tract
profiles of white matter properties: automating fiber-tract quantification. PloS one
7(11), e49790 (2012)
18. Yeh, F.C., Tseng, W.Y.I.: Ntu-90: a high angular resolution brain atlas constructed
by q-space diffeomorphic reconstruction. Neuroimage 58(1), 91–99 (2011)
| 1cs.CV
|
arXiv:1506.06796v2 [nlin.AO] 19 Oct 2015
When slower is faster
Carlos Gershenson1,2,3,4,5 and Dirk Helbing6
1
Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas,
Universidad Nacional Autónoma de México.
[email protected] http://turing.iimas.unam.mx/~cgg
2
Centro de Ciencias de la Complejidad, UNAM, México.
3
SENSEable City Lab, Massachusetts Institute of Technology, USA.
4
MoBS Lab, Northeastern University, USA.
5
ITMO University, St. Petersburg, Russian Federation.
6
Department of Humanities, Social and Political Sciences (GESS)
ETH Zürich
[email protected] http://www.coss.ethz.ch
October 21, 2015
Abstract
The slower is faster (SIF) effect occurs when a system performs worse as its components try to do better. Thus, a moderate individual efficiency actually leads to a
better systemic performance. The SIF effect takes place in a variety of phenomena.
We review studies and examples of the SIF effect in pedestrian dynamics, vehicle traffic, traffic light control, logistics, public transport, social dynamics, ecological systems,
and adaptation. Drawing on these examples, we generalize common features of the SIF
effect and suggest possible future lines of research.
1
Introduction
How fast should an athlete run a race? If she goes too fast, she will burn out and become
tired before finishing. If she runs conservatively, she will not get tired, but will not make
her best time. To minimize her race time, she has to go as fast as possible but without
burning out. If she goes faster, she will actually race more slowly. This is an example of
the “slower-is-faster” (SIF) effect: in order to run faster, sometimes it is necessary to run
slower, not to burn out. It is not trivial to calculate the running speed which will lead to the
best race, as this depends on the athlete, race distance, track, temperature, humidity, and
daily performance: Running 100m dash should be done as fast as you can, while running a
marathon demands a carefully paced race. How fast would an athlete run a marathon if she
started with a speed for a 100m? To finish the marathon successfully, she would obviously
have to run more slowly.
1
There are several other examples of the SIF effect, which will be described in the next
section. We then generalize the common features of these phenomena to discuss potential
causes and promising lines of research towards a unified explanation of the SIF effect.
2
2.1
Examples
Pedestrian evacuation
Perhaps the first formal study of the SIF effect was related to pedestrian flows (Helbing et al.,
2000a). Modelling crowds like self-driven particles with “social forces” interacting among
them (Helbing and Molnár, 1995; Helbing et al., 2000b), it has been shown that when individuals try to evacuate a room too quickly, they lead to intermittent clogging and a reduced
outflow as compared to a calmer evacuation. In this context, the SIF effect is also known as
“freezing by heating” (Stanley, 2000). Trying to exit fast makes pedestrians slower, while
calmer people manage to exit faster. This has led people to suggest obstacles close to exits,
precisely to reduce friction (Helbing et al., 2005; Heliövaara et al., 2013). Counterintuitively,
a slowdown of the evacuation can increase the outflow. Also, in a related study of aircraft
evaluation, it was found that there is a critical door width which determines whether “competitive” evacuation will increase or decrease evacuation time (Kirchner et al., 2003). In
other words, pushy people will evacuate slower if there are narrow doors (SIF), but will
evacuate faster if the doors are wide enough (FIF, faster-is-faster).
2.2
Pedestrians crossing a road
Another example concerns mixed pedestrian and vehicle traffic. Imagine pedestrians are
trying to cross a road at a location where there is no traffic light and no pedestrian crossing
is marked. This is a typical situation along speed-reduced roads (e.g. with a speed limit
of 30km/h) or in shared spaces for multi-modal use. Pedestrians would cross when the
gap between two successive vehicles exceeds a certain critical separation that ensures a safe
crossing of the road. However, there are two types of pedestrians: patient and pushy ones.
Pushy pedestrians might force a vehicle to slow down while patient pedestrians would not do
this, i.e. they would wait for a larger gap. Surprisingly, if all pedestrians were of the patient
type, on average they would have to wait for a shorter time period (Jiang et al., 2006).
How does this SIF effect come about? When a pushy pedestrian has slowed a vehicle
down, other arriving pedestrians will pass the road, too, and it takes a long time until no
further pedestrians arrive, and the stopped cars can accelerate again. During the waiting
time, however, a long vehicle queue has formed, such that no large enough gap to cross the
road occurs between vehicles until the entire vehicle queue has dissolved. As a consequence,
pedestrians will have to wait for a long time until they can cross again. Altogether, it is
better if pedestrians wait for large enough gaps such that they don’t force vehicles to slow
down.
2
2.3
Vehicle traffic
SIF effects are also known from vehicle traffic (Helbing and Huberman, 1998; Helbing and Treiber,
1998; Helbing, 2001; Helbing and Nagel, 2004). Surprisingly, speed limits can sometimes
reduce travel times. This is the case, when the traffic density enters the multi-stable (“metastable”) regime. Then, traffic flow is sensitive to disruptions and may break down, which
causes largely increased travel times. A speed limit can delay the breakdown of fluid traffic
flows, because it reduces the variability of vehicle speeds. This homogenization avoids disturbances in the flow, which are big enough to trigger a breakdown (i.e. have a “super-critical”
amplitude).
If vehicles go fast, the safety distance between vehicles must be increased. Thus, less
vehicles will be able to use a road. For example, at 80km/h, a maximum capacity of about
22 vehicles per km per lane is reached before free traffic flow breaks down. At 120km/h, this
capacity is reduced to about 15 vehicles per km per lane. Once vehicles slow down due to
an increased density, traffic jams will propagate, as a following car tends to brake more than
the vehicle ahead. This phase transition of “stable” to “unstable” flow in traffic depends on
the desired speed. Thus, to maximize flow, the optimal speed of a highway will depend on
the current density. However, the maximum flow lies at the tipping point, and thus a small
perturbation can trigger stop-and-go waves which can reduce the highway capacity by 30%.
A similar consideration applies to over-taking maneuvers (Kesting et al., 2007). Pushy
drivers might force cars in the neighboring lane to slow down when changing lanes to overtake
another car, while patient drivers would not do this. As a consequence, pushy drivers may
cause a disruption of metastable traffic flow, which may trigger a breakdown (“capacity
drop”). Consequently, patient drivers will avoid or delay a breakdown of traffic flow, thereby
managing to progress faster on average. One may also formulate this in game theoretical
terms. When traffic flow is metastable, drivers are faced with a social dilemma situation:
choosing a patient behavior will be beneficial for everyone, while pushy behavior will produce
small individual advantages at the cost of other drivers. As a consequence, a “tragedy of
the commons” results: pushy drivers undermine the stability of the metastable traffic flow,
causing congestion that forces everyone to spend more time on travel.
A complementary phenomenon is observed in Braess’s paradox (Braess et al., 2005; Steinberg and Zangw
1983), where adding roads can reduce the flow capacity of a road network.
2.4
Traffic light control
The SIF effect is also found in further systems such as urban traffic light control (Helbing and Mazloumian,
2009). Here, a first-come-first-serve approach works only well at low traffic volumes. Otherwise, forcing vehicles to wait for some time can speed up their overall progress. The
reason is that this will produce vehicle platoons, such that a green light will efficiently serve
many vehicles in a short time period (Gershenson, 2005; Gershenson and Rosenblueth, 2012;
Zubillaga et al., 2014). Similarly, it may be better to switch traffic lights less frequently, because switching reduces service times (due to time lost on amber lights). A “green wave”,
i.e., a coordination of vehicle flows such that several successive traffic lights can be passed
without stopping, is another good example demonstrating that waiting (at a red light) may
be rewarding altogether.
3
Similarly interesting observations can be made for self-organized traffic light control
(“self-control”), which based on decentralized flow control (“distributed control”) (Lämmer et al.,
2006; Lämmer and Helbing, 2008, 2010; Helbing, 2013). If each intersection strictly minimizes the travel times of all vehicles approaching it, according to the principle of a “homo
economicus”, this can create efficient traffic flows, when these are low or moderate (“invisible
hand phenomenon”). However, vehicle queues might get out of hand when the intersection
utilization increases. Therefore, it is beneficial to interrupt travel time minimization in order
to clear a vehicle queue when it has grown beyond a certain “critical” limit. This avoids
spillover effects, which would block other intersections and cause a quick spreading of congestion over large parts of a city. Consequently, waiting for a long queue to be cleared can
speed up traffic altogether. Putting it differently, other-regarding self-organization can beat
the selfish optimization à la “homo economicus”, who strictly does the best, but neglects a
coordination with neighbors.
2.5
Logistics and supply chains
Similar phenomena as in urban traffic flows are found in logistic systems and supply chains (Helbing and Läm
2005; Helbing et al., 2006; Seidel et al., 2008; Peters et al., 2008). We have studied, for example, a case of harbor logistics using automated guided vehicles for container transport.
Our proposal was to reduce the speed of these vehicles. This reduced the required safety
distances between vehicles, such that less conflicts of movement occurred, and the automatic
guided vehicles had to wait less. In this way, transportation times could be overall reduced,
even though movement times obviously increased.
We made a similar observation in semiconductor production. So-called “wet benches” are
used to etch structures into silicium wavers, using particular chemical solutions. To achieve
good results, the wavers should stay in the chemical baths longer than a minimum and shorter
than a maximum time period. Therefore, it might happen that several silicium wavers need
to be moved around at about the same time, while a moving gripper, the “handler”, must
make sure to stay within the minimum and maximum times. It turns out that slightly
extending the exposure time in the chemical bathes enables much better coordination of the
movement processes, thereby reaching a 30 percent higher throughput.
In a third logistics project, the throughput of a packaging plant had to be increased.
One of the central production machines of this plant frequently broke down, such that it
was operated at full speed whenever it was operating well. However, this filled the buffer
of the production plant to an extent that made its operation inefficient. This effect can be
understood with queuing theory, according to which cycle times can dramatically increase
as the capacity of a buffer is approached.
2.6
Public transport
In public transportation systems, it is desirable to have equal headways between vehicles
such as buses, i.e., to reach regular time separations between vehicles. However, the equal
headway configuration is unstable (Gershenson and Pineda, 2009). Forcing equal headways
minimizes waiting times at stations. Nevertheless, the travel time is not independent of the
waiting time, as equal headways imply idling or leaving some passengers at stations. This is
4
because there is a different demand for each vehicle at each station. Still, self-organization
can be used to regulate the headways adaptively (Gershenson, 2011a). Considering only
local information, vehicles are able to respond adaptively to the immediate demand of each
station. With this method, there is also a SIF effect, as passengers wait more time at a
station, but reach their destination faster once they board a vehicle because there is no
idling necessary to maintain equal headways.
2.7
Social dynamics
Axelrod (Axelrod, 1997) proposed an interesting model of opinion formation. In this model,
agents may change their opinion depending on the opinion of their neighbors. Eventually,
the opinions converge to a stable state. However, if agents switch their opinion too fast,
this might delay convergence (Stark et al., 2008a,b). Thus, there is a SIF effect because
the fastest convergence will not necessarily be obtained with the fastest opinion change. In
this model, there is also a phase transition which is probably related to the optimal opinion
change rate (Vilone et al., 2002). There is also experimental evidence of the SIF effect in
group decisions. While designing new buildings, slowing down the deliberative process of
teams accelerates the design and construction of buildings (Cross et al., 2015).
Extrapolating these results, one may speculate that high-frequency financial trading (Narang,
2013) may also produce a SIF effect, in the sense that trading at the microseconds scale generates price and information fluctuations which could generate market instabilities leading
to crashes and slower economic growth (Easley et al., 2011).
In combinatorial game theory (Siegel, 2013), sometimes the best possible move (e.g. taking a queen in chess) is not necessarily the best move in the long term. In other words,
having the highest possible gain at each move does not give necessarily the best game result (Russell and Norvig, 2003, pp. 163–171).
2.8
Ecology
If a predator consumes its prey too fast, there will be no prey to consume and the predator population will decline. Thus, a “prudent predator” (Slobodkin, 1961; Goodnight et al.,
2008) will actually spread faster than a greedy one. A similar SIF effect applies to parasite/host relationships, where parasites taking too many resources from their host are causing
their own demise (Dunne et al., 2013). Over long timescales, evolution will favor symbiotic
over parasitic relationships, promoting mechanisms for cooperation which can regulate the
interaction between different individuals (Sachs et al., 2004; Virgo et al., 2013).
We can see that the same principle applies to natural resource management, such as
fisheries (Pauly et al., 1998). If catches are excessive, there will not be enough fishes left
to maintain their numbers, and subsequent catches will be poor. It is estimated that apart
from its ecological impact, overfishing has left a void of US$32 billion per year due to reduced catches (Toppe et al., 2012). However, regulating how much fish is caught per year
is complicated. The maximum sustainable yield varies from species to species (Maunder,
2002), so the calculation of the “optimal” yields per year is not at all a trivial task.
5
2.9
Adaptation
Evolution, development, and learning can be seen as different types of adaptation, acting
at different timescales (Aguilar et al., 2014). Also, adaptation can be seen as a type of
search (Downing, 2015). In computational searches, it is known that there needs to be
a balance between “exploration” and “exploitation” (Blum and Roli, 2003). An algorithm
can explore different possible solutions or exploit solutions similar to those already found.
Too much exploration or too much exploitation will lead to longer search times. Too much
breadth (exploration) will only explore slightly different types of solution, while too much
depth (exploitation) might lead to local optima and data overfitting. A key problem is
that the precise balance between exploration (diversification) and exploitation (intensification) depends the precise search space (Wolpert and Macready, 1995, 1997) and timescale
(Gershenson, 2010; Watson et al., 2011).
An example of the SIF was described in biological evolution (Sellis et al., 2011). Haploid
species (with a single copy of their genome, such as bacteria) can adapt faster than diploid
species (with two copies of their genome, such as most plants and animals). Still, in a fastchanging environment, haploids adapt “too fast”, i.e. the population loses genome variation,
while diploids can maintain a diversity. Having such a diversity, diploids can adapt faster to
changes in their environment, as they can begin an evolutionary search from many different
states at once.
In principle, it would be desirable to find a solution as fast as possible, exploiting current
solutions. Still, as mentioned, this might lead to suboptimality (SIF) in evolving new features, optimizing a multidimensional function, or training a neural network. To be efficient,
search should eventually “slow down”, as it is known from “simulated annealing”. As too
much exploration would be suboptimal also, the critical question is how to find the precise
balance to speed up search as much as possible. Computationally, it seems that this question is not reducible (Wolfram, 2002), so we can only know a posteriori the precise balance
for a given problem. Still, finding this balance would be necessary for adiabatic quantum
computation (Farhi et al., 2000; Aharonov et al., 2008), as if the system evolves too fast, the
information is destroyed.
3
Generalization
What do all the above examples have in common? They can be described as complex
dynamical systems composed of many non-linearly interacting components. In the above
cases, the system can have at least two different states: an efficient and an inefficient one.
Unfortunately, the efficient state can be unstable, such that the system will tend to end up
in the inefficient state. In the case of freeway traffic, for example, it is well known that the
most efficient state (with the highest throughput) is unstable, thereby causing the traffic
flow to break down sooner or later (“capacity drop”). To avoid the undesired outcome, the
system components must stay sufficiently away from the instability point, which requires
them to be somewhat slower than they could be, but as a reward they will be able to sustain
a relatively high speed for a long time. If they go faster, the efficient state will break down
and trigger another one that is typically slower. This situation might be characterized as a
6
“tragedy of the commons” (Hardin, 1968).
Even though it might be counterintuitive, the SIF effect occurs in a broad variety of systems. For practical purposes, many systems have a monotonic relation between “inputs” and
“outputs”. This is true for systems that “break” (Ashby, 1947). For example, if temperature
is increased in a constrained gas with a constant volume, pressure rises. Yet, if temperature
increases too much, then the gas container will break, leading to a pressure reduction. Still,
without breaking, many physical and non-physical systems have thresholds, where they become unstable and a phase transition to a different systems state occurs. A typical situation
of many-particle systems is that they may get overloaded and turn dysfunctional through a
cascading effect.
To reduce the SIF effect we can seek to adjust the interactions which cause a reduction
in the system performance (Gershenson, 2011b). The vehicle traffic case offers an interesting
example: when vehicles go too fast (and their density crosses a critical density), their changes
in speed will affect other vehicles, generating an amplification of oscillations, which lead
to stop-and-go traffic and, as a consequence, to a reduced average speed. If vehicles go
slower, then such oscillations can be avoided and the average speed will be higher. The
key here is that the critical speed where traffic flow changes from “laminar” (where FIF)
to “unstable” (where SIF) changes with the density. However, suitably designed adaptive
systems, such as driver assistance systems, can be used to drive systems towards their best
possible performance in their respective context (Gershenson, 2007; Helbing, 2015).
4
Discussion
It could be argued that the SIF effect is overly simplistic, as there is only the requirement
of having two dynamical phases, where one comes with a reduced efficiency after crossing
the phase transition point. Still, as we have presented, the SIF effect shows up in a variety of interesting phenomena at different scales. Thus, we can say that having a better
understanding of the SIF effect can be useful and potentially have a broad impact. A challenge lies in characterizing the nature of the different types of interactions which will reduce
efficiency (Gershenson, 2013).
We can identify the following necessary conditions for the SIF effect:
1. There is an instability (internal or external) in the system.
2. The instability is amplified, sometimes through cascading effects.
3. There is a transition from the unstable to a new stable state which leads to inefficiency.
4. Such a state can be characterized as “overloaded”.
It is worth noting that in some cases, single variables may be stable to perturbations,
but their interactions are the ones that trigger instability. This implies that the SIF in these
cases has to be studied at two scales: the scale of the components and the scale of the system,
as studying components in isolation will not provide enough information to reproduce the
SIF effect.
7
Whether all phenomena with a SIF effect can be described with the same mathematical
framework remains to be seen. We believe this is an avenue of research worth pursuing and
with relevant implications for the understanding of complex systems.
Acknowledgments
We should like to thank Luis Álvarez de Icaza, Jeni Cross, Tom Froese, Marios Kyriazis, Gleb
Oshanin, Sui Phang, Frank Schweitzer, Diamantis Sellis, Simone Severini, Thomas Wisdom,
Héctor Zenil, and two anonymous referees for useful comments. C.G. was supported by
CONACYT projects 212802, 221341, and SNI membership 47907. D.H. was supported by
ERC Advanced Grant MOMENTUM 324247.
References
Aguilar, W., Santamarı́a Bonfil, G., Froese, T., and Gershenson, C. (2014).
The past, present, and future of artificial life. Frontiers in Robotics and AI 1 (8). URL
http://dx.doi.org/10.3389/frobt.2014.00008.
Aharonov, D., Van Dam, W., Kempe, J., Landau, Z., Lloyd, S., and Regev, O.
(2008). Adiabatic quantum computation is equivalent to standard quantum computation.
SIAM review 50 (4): 755–787. URL http://dx.doi.org/10.1137/080734479.
Ashby, W. R. (1947). The nervous system as physical machine: With special reference to the origin of adaptive behavior. Mind 56 (221) (January): 44–59. URL
http://tinyurl.com/aqcmdy.
Axelrod, R. (1997). The dissemination of culture: A model with local convergence
and global polarization. Journal of Conflict Resolution 41 (2): 203–226. URL
http://jcr.sagepub.com/content/41/2/203.abstract.
Blum, C. and Roli, A. (2003). Metaheuristics in combinatorial optimization: Overview
and conceptual comparison. ACM Comput. Surv. 35 (3) (Sept.): 268–308. URL
http://doi.acm.org/10.1145/937503.937505.
Braess, D., Nagurney, A., and Wakolbinger, T. (2005). On a paradox of traffic
planning. Transportation Science 39 (4) (November): 446–450. Translated from the original German: Braess, Dietrich. 1968. Über ein Paradoxon aus der Verkehrsplanung. Unternehmensforschung 12 258–268. URL http://dx.doi.org/10.1287/trsc.1050.0127.
Cross, J. E., Barr, S. K., Putnam, R., Dunbar, B. H., and Plaut, J. M. (2015).
The social network of integrative design. Tech. rep., Institute for the Built Environment,
Colorado State University, Fort Collins, CO, USA.
Downing, K. L. (2015). Intelligence Emerging: Adaptivity and Search in Evolving Neural
Systems. MIT Press, Cambridge, MA, USA.
8
Dunne, J. A., Lafferty, K. D., Dobson, A. P., Hechinger, R. F., Kuris,
A. M., Martinez, N. D., McLaughlin, J. P., Mouritsen, K. N., Poulin,
R., Reise, K., Stouffer, D. B., Thieltges, D. W., Williams, R. J., and
Zander, C. D. (2013).
Parasites affect food web structure primarily through
increased diversity and complexity.
PLoS Biol 11 (6) (06): e1001579.
URL
http://dx.doi.org/10.1371/journal.pbio.1001579.
Easley, D., de Prado, M. M. L., and O’Hara, M. (2011). The microstructure of the “flash crash”: Flow toxicity, liquidity crashes, and the probability of informed trading. The Journal of Portfolio Management 37 (2) (Winter): 118–128. URL
http://dx.doi.org/10.3905/jpm.2011.37.2.118.
Farhi, E., Goldstone, J., Gutmann, S., and Sipser, M. (2000).
tum computation by adiabatic evolution.
Tech. Rep. CTP-2936, MIT.
http://arxiv.org/abs/quant-ph/0001106.
QuanURL
Gershenson, C. (2005). Self-organizing traffic lights. Complex Systems 16 (1): 29–53.
URL http://www.complex-systems.com/pdf/16-1-2.pdf.
Gershenson, C. (2007). Design and Control of Self-organizing Systems. CopIt Arxives,
Mexico. http://tinyurl.com/DCSOS2007. URL http://tinyurl.com/DCSOS2007.
Gershenson, C. (2010). Computing networks: A general framework to contrast neural
and swarm cognitions. Paladyn, Journal of Behavioral Robotics 1 (2): 147–153. URL
http://dx.doi.org/10.2478/s13230-010-0015-z.
Gershenson, C. (2011a).
Self-organization leads to supraoptimal performance in public transportation systems.
PLoS ONE 6 (6): e21469.
URL
http://dx.doi.org/10.1371/journal.pone.0021469.
Gershenson, C. (2011b).
The sigma profile: A formal tool to study organization and its evolution at multiple scales.
Complexity 16 (5): 37–44.
URL
http://arxiv.org/abs/0809.0504.
Gershenson, C. (2013). The implications of interactions for science and philosophy. Foundations of Science 18 (4): 781–790. URL http://arxiv.org/abs/1105.2827.
Gershenson, C. and Pineda, L. A. (2009). Why does public transport not arrive on
time? The pervasiveness of equal headway instability. PLoS ONE 4 (10): e7292. URL
http://dx.doi.org/10.1371/journal.pone.0007292.
Gershenson, C. and Rosenblueth, D. A. (2012).
Self-organizing traffic lights at multiple-street intersections.
Complexity 17 (4): 23–39.
URL
http://dx.doi.org/10.1002/cplx.20392.
Goodnight, C., Rauch, E., Sayama, H., De Aguiar, M. A. M., Baranger,
M., and Bar-yam, Y. (2008). Evolution in spatial predator–prey models and
the “prudent predator”: The inadequacy of steady-state organism fitness and the
9
concept of individual and group selection.
http://dx.doi.org/10.1002/cplx.20209.
Complexity 13 (5):
23–44.
URL
Hardin, G. (1968). The tragedy of the commons. Science 162 (3859): 1243–1248. URL
http://www.sciencemag.org/content/162/3859/1243.abstract.
Helbing, D. (2001). Traffic and related self-driven many-particle systems. Reviews of
modern physics 73 (4): 1067.
Helbing, D. (2013). Economics 2.0: the natural step towards a self-regulating, participatory
market society. Evolutionary and Institutional Economics Review 10 (1): 3–41.
Helbing, D. (2015). Thinking Ahead-Essays on Big Data, Digital Revolution, and Participatory Market Society. Springer.
Helbing, D., Buzna, L., Johansson, A., and Werner, T. (2005). Self-organized
pedestrian crowd dynamics: Experiments, simulations, and design solutions. Transportation science 39 (1): 1–24.
Helbing, D., Farkas, I., and Vicsek, T. (2000a). Simulating dynamical features of
escape panic. Nature 407 (6803): 487–490.
Helbing, D., Farkas, I. J., and Vicsek, T. (2000b).
ing in a driven mesoscopic system.
Phys. Rev. Lett. 84:
http://link.aps.org/doi/10.1103/PhysRevLett.84.1240.
Freezing by heat1240–1243.
URL
Helbing, D. and Huberman, B. A. (1998). Coherent moving states in highway traffic.
Nature 396: 738–740. URL http://dx.doi.org/10.1038/25499.
Helbing, D. and Lämmer, S. (2005). Supply and production networks: from the bullwhip
effect to business cycles. In Networks of Interacting Machines: Production Organization
in Complex Industrial Systems and Biological Cells, D. Armbruster, A. S. Mikhailov, and
K. Kaneko, (Eds.). World Scientific, Singapore, 33–66.
Helbing, D. and Mazloumian, A. (2009).
Operation regimes and sloweris-faster effect in the controlof traffic intersections.
The European Physical
Journal B - Condensed Matter and Complex Systems 70:
257–274.
URL
http://dx.doi.org/10.1140/epjb/e2009-00213-5.
Helbing, D. and Molnár, P. (1995). Social force model for pedestrian dynamics. Physical
review E 51 (5): 4282.
Helbing, D. and Nagel, K. (2004).
The physics of traffic
gional development.
Contemporary Physics 45 (5):
405–426.
http://www.tandfonline.com/doi/abs/10.1080/00107510410001715944.
and
reURL
Helbing, D., Seidel, T., Lämmer, S., and Peters, K. (2006). Self-organization principles in supply networks and production systems. In Econophysics and Sociophysics, B. K.
Chakrabarti, A. Chakraborti, and A. Chatterjee, (Eds.). Wiley, Weinheim, 535–559. URL
http://dx.doi.org/10.1002/9783527610006.ch19.
10
Helbing, D. and Treiber, M. (1998). Jams, waves, and clusters. Science 282 (5396):
2001–2003. URL http://arxiv.org/abs/cond-mat/9812299.
Heliövaara, S., Ehtamo, H., Helbing, D., and Korhonen, T. (2013). Patient and
impatient pedestrians in a spatial game for egress congestion. Phys. Rev. E 87: 012802.
URL http://link.aps.org/doi/10.1103/PhysRevE.87.012802.
Jiang, R., Helbing, D., Shukla, P. K., and Wu, Q.-S. (2006).
Inefficient emergent oscillations in intersecting driven many-particle flows.
Physica A: Statistical Mechanics and its Applications 368 (2): 567 – 574.
URL
http://www.sciencedirect.com/science/article/pii/S0378437106000070.
Kesting, A., Treiber, M., and Helbing, D. (2007). General lane-changing model
MOBIL for car-following models. Transportation Research Record: Journal of the Transportation Research Board 1999 (1): 86–94.
Kirchner,
A., Klüpfel,
H., Nishinari,
K., Schadschneider,
A.,
and Schreckenberg, M. (2003).
Simulation of competitive egress behavior:
comparison with aircraft evacuation data.
Physica A: Statistical Mechanics and its Applications 324 (3–4):
689 – 697.
URL
http://www.sciencedirect.com/science/article/pii/S0378437103000761.
Lämmer, S. and Helbing, D. (2008).
Self-control of traffic lights and vehicle flows in urban road networks.
J. Stat. Mech. 2008: P04019.
URL
http://dx.doi.org/10.1088/1742-5468/2008/04/P04019.
Lämmer, S. and Helbing, D. (2010). Self-stabilizing decentralized signal control of realistic, saturated network traffic. Tech. Rep. 10-09-019, Santa Fe Institute.
Lämmer, S., Kori, H., Peters, K., and Helbing, D. (2006). Decentralised control
of material or traffic flows in networks using phase-synchronisation. Physica A 363 (1)
(April): 39–47. URL http://dx.doi.org/10.1016/j.physa.2006.01.047.
Maunder, M. N. (2002). The relationship between fishing methods, fisheries management
and the estimation of maximum sustainable yield. Fish and Fisheries 3 (4): 251–260.
URL http://dx.doi.org/10.1046/j.1467-2979.2002.00089.x.
Narang, R. K. (2013). High-speed trading. In Inside the Black Box: A Simple Guide
to Quantitative and High-Frequency Trading, 2nd ed. John Wiley & Sons, Inc., Hoboken,
NJ, USA, 243–264. URL http://dx.doi.org/10.1002/9781118662717.ch14.
Pauly, D., Christensen, V., Dalsgaard, J., Froese, R., and Torres, F.
(1998). Fishing down marine food webs. Science 279 (5352): 860–863. URL
http://dx.doi.org/10.1126/science.279.5352.860.
Peters, K., Seidel, T., Lämmer, S., and Helbing, D. (2008). Logistics networks: Coping with nonlinearity and complexity. In Managing Complexity: Insights,
Concepts, Applications, D. Helbing, (Ed.). Springer Berlin Heidelberg, 119–136. URL
http://dx.doi.org/10.1007/978-3-540-75261-5_6.
11
Russell, S. and Norvig, P. (2003). Artificial intelligence: a modern approach, 2nd ed.
Prentice Hall, Upper Saddle River, New Jersey.
Sachs, J. L., Mueller, U. G., Wilcox, T. P., and Bull, J. J. (2004). The evolution of cooperation. The Quarterly Review of Biology 79 (2): pp. 135–160. URL
http://www.jstor.org/stable/10.1086/383541.
Seidel, T., Hartwig, J., Sanders, R. L., and Helbing, D. (2008). An agentbased approach to self-organized production. In Swarm Intelligence: Introduction
and Applications, C. Blum and D. Merkle, (Eds.). Springer, Berlin, 219–252. URL
http://dx.doi.org/10.1007/978-3-540-74089-6_7.
Sellis, D., Callahan, B. J., Petrov, D. A., and Messer, P. W. (2011).
Heterozygote advantage as a natural consequence of adaptation in diploids. Proceedings of the National Academy of Sciences 108 (51): 20666–20671.
URL
http://www.pnas.org/content/108/51/20666.abstract.
Siegel, A. N. (2013). Combinatorial Game Theory. American Mathematical Society.
Slobodkin, L. B. (1961). Growth and Regulation of Animal Populations. Holt, Reinhart
and Winston, New York.
Stanley, H. E. (2000). Non-equilibrium physics: Freezing by heating. Nature 404 (6779)
(04): 718–719. URL http://dx.doi.org/10.1038/35008188.
Stark, H.-U., Tessone, C. J., and Schweitzer, F. (2008a). Decelerating microdynamics can accelerate macrodynamics in the voter model. Phys. Rev. Lett. 101: 018701.
URL http://link.aps.org/doi/10.1103/PhysRevLett.101.018701.
Stark, H.-U., Tessone, C. J., and Schweitzer, F. (2008b). Slower is faster: Fostering
consensus formation by heterogeneous inertia. Advances in Complex Systems 11 (04):
551–563.
Steinberg, R. and Zangwill, W. I. (1983). The prevalence of Braess’ paradox. Transportation Science 17 (3): 301–318. URL http://dx.doi.org/10.1287/trsc.17.3.301.
Toppe, J., Bondad-Reantaso, M., Hasan, M., Josupeit, H., Subasinghe, R., Halwart, M., and James, D. (2012). Aquatic biodiversity for sustainable diets: the role of
aquatic foods in food and nutrition security. In Sustainable diets and biodiversity: directions and solutions for policy, research and action, B. Burlingame and S. Dernini, (Eds.).
FAO, Rome, 94–101. URL http://www.fao.org/docrep/016/i3004e/i3004e.pdf.
Vilone, D., Vespignani, A., and Castellano, C. (2002).
Ordering phase
transition in the one-dimensional axelrod model.
The European Physical Journal B - Condensed Matter and Complex Systems 30 (3):
399–406.
URL
http://dx.doi.org/10.1140/epjb/e2002-00395-2.
12
Virgo, N., Froese, T., and Ikegami, T. (2013). The positive role of parasites in the
origins of life. In Artificial Life (ALIFE), 2013 IEEE Symposium on. IEEE, pp. 1–4. URL
http://dx.doi.org/10.1109/ALIFE.2013.6602424.
Watson, R. A., Mills, R., and Buckley, C. L. (2011). Global adaptation in networks of selfish components: Emergent associative memory at the system scale. Artificial
Life 17 (3) (2015/06/01): 147–166. URL http://dx.doi.org/10.1162/artl_a_00029.
Wolfram, S. (2002).
A New Kind of Sciene.
http://www.wolframscience.com/.
Wolfram Media.
URL
Wolpert, D. H. and Macready, W. G. (1995). No free lunch theorems for search.
Tech. Rep. SFI-WP-95-02-010, Santa Fe Institute. URL http://tinyurl.com/yz274ej.
Wolpert, D. H. and Macready, W. G. (1997). No Free Lunch Theorems for Optimization. IEEE Transactions on Evolutionary Computation 1 (1): 67–82.
Zubillaga, D., Cruz, G., Aguilar, L. D., Zapotécatl, J., Fernández, N.,
Aguilar, J., Rosenblueth, D. A., and Gershenson, C. (2014). Measuring
the complexity of self-organizing traffic lights. Entropy 16 (5): 2384–2407. URL
http://dx.doi.org/10.3390/e16052384.
13
| 9cs.NE
|
Rank 1 deformations of non-cocompact hyperbolic lattices
arXiv:1702.00508v1 [math.GT] 2 Feb 2017
Samuel A. Ballas, Julien Paupert, Pierre Will
February 3, 2017
Abstract
Let X be a negatively curved symmetric space and Γ a non-cocompact lattice in Isom(X). We show that
small, parabolic-preserving deformations of Γ into the isometry group of any negatively curved symmetric
space containing X remain discrete and faithful (the cocompact case is due to Guichard). This applies
in particular to a version of Johnson-Millson bending deformations, providing for all n infnitely many noncocompact lattices in SO(n, 1) which admit discrete and faithful deformations into SU(n, 1). We also produce
deformations of the figure-8 knot group into SU(3, 1), not of bending type, to which the result applies.
1
Introduction
This paper concerns an aspect of the deformation theory of discrete subgroups of Lie groups, namely that of
non-cocompact lattices in rank 1 semisimple Lie groups. More specifically, we consider the following questions,
given a discrete subgroup Γ of a rank 1 Lie group H:
(1) Does Γ admit any deformations in H?
(2) If so, do these deformations have any nice properties (e.g. remain discrete and faithful)?
(3) What if we replace H with a larger Lie group G?
Here we call deformation of Γ in H any continuous 1-parameter family of representations ρt : Γ −→ H (for
t in some interval (−ε, ε)) satisfying ρ0 = ι (the inclusion of Γ in H), and ρt not conjugate to ρt′ for any
t 6= t′ ∈ (−ε, ε). We say that Γ is locally rigid in H if it does not admit any deformations into H.
When H is a semisimple real Lie group without compact factors there are a variety of general local rigidity
results which we now outline. Weil proved in [W] that Γ is locally rigid in H if H/Γ is compact and H not locally
isomorphic to SL(2, R). Garland and Raghunathan extended this result to the case where Γ is a non-cocompact
lattice in a rank 1 semisimple group H not locally isomorphic to SL(2, R) or SL(2, C) (Theorem 7.2 of [GR]).
The exclusion of SL(2, R) and SL(2, C) is necessary. Generically, lattices in H = SL(2, R) admit many
deformations in H. The identification of PSL(2, R) with Isom+ (H2R ) allow us to relate lattices in SL(2, R) with
hyperbolic structures which are in turn parameterized by the classical Teichmüller space when Γ is a surface
group. The case of quasi-Fuchsian deformations of a discrete subgroup Γ of H = SL(2, R) into G = SL(2, C)
is also classical, well-studied, and well understood by the Bers simultaneous uniformization theorem [Bers]. In
this setting, PSL(2, C) can be identified with Isom+ (H3R ) and the discrete group Γ ⊂ SL(2, C) gives rise to a
hyperbolic structure on the manifold M ∼
= Σ × R, where Σ is a hyperbolic surface. Deforming Γ in SL(2, C)
corresponds to deforming the hyperbolic structure on M . Such deformations are abundant and according the
Bers simultaneous uniformization can be parameterized by a cartesian product of two copies of the classical
Teichmüller space of the surface Σ. Notice that the existence of deformations into G does not violate Weil’s
result as G/Γ is not compact. This situation can be generalized to the case where H = SO0 (n, 1) ∼
= Isom+ (HnR )
+
0
n+1
∼
and G = SO (n + 1, 1) = Isom (HR ). In this setting, a lattice Γ in H gives rise to a hyperbolic structure on
M = HnR /Γ. Again, regarding Γ as a subgroup of SO0 (n + 1, 1) gives rise to a hyperbolic structure on M × R and
deformations of Γ into G correspond to deforming this hyperbolic structure. In this more general setting there
is no general theorem that guarantees the existence of deformations of Γ into G. However, when n = 3 this
deformation problem has been studied by Scannell [Sc], Bart–Scannell [BSc], and Kapovich [Kap] who prove
some rigidity results.
1
Returning to the 3-dimensional case, many non-cocompact lattices Γ in H = SL(2, C) ∼
= SO0 (3, 1) are known
to admit deformations into H. In particular, when Γ is torsion-free Thurston showed that for each cusp there
exists a (real) 2-dimensional family of deformations of Γ into H, called Dehn surgery deformations (see Section
5.8 of [T]). Geometrically, in each of these families the commuting pair of parabolic isometries generating the
correspdonding cusp group is deformed to a pair of loxodromic isometries sharing a common axis. In particular
these deformations are all non-discrete or non-faithful. If H3R /Γ is an orbifold then the existence of deformations
depends more subtly on the topology of the cusp cross-sections.
Another case of interest in the context of deformations of geometric structures is that of projective deformations of hyperbolic lattices, i.e. deformations of lattices Γ of H = SO(n, 1) into G = SL(n + 1, R). When Γ is a
torsion-free cocompact lattice in SO(n, 1) such that the hyperbolic manifold M = HnR /Γ contains an embedded
totally geodesic hypersurface Σ, Johnson and Millson showed in [JM] that Γ admits a 1-parameter family of
deformations into SL(n + 1, R). They obtained these deformations, called bending deformations of M along
Σ, by introducing an algebraic version of Thurston’s bending deformations of a hyperbolic 3-manifold along a
totally geodesic surface. This algebraic version is very versatile, and can be generalized in a variety of ways. For
example, the hypothesis that M is compact may be dropped, see [BM]. Furthermore, the construction can be
applied to the setting of other Lie groups and will provide us with a rich source of examples that are discussed
in Section 4. In addition to deformations constructed via bending there are also instances of projective deformations that do not arise via the previously mentioned bending technique (see [B1, B2, BDL]). On the other hand,
despite the existence of these bending examples, empirical evidence complied by Cooper–Long–Thistlethwaite
[CLTII] suggests that the existence of deformations into SL(4, R) is quite rare for closed hyperbolic 3-manifolds.
In another direction, complex hyperbolic quasi-Fuchsian deformations of Fuchsian groups have also been
extensively studied (see e.g. [S2], the survey [PP] and references therein). With the above notation, this concerns
deformations of discrete subgroups Γ of H into G, with (H, G) = (SO(2, 1), SU(2, 1)) or (SU(1, 1), SU(2, 1)).
(Recall that the Lie groups SL(2, R), SO(2, 1), SU(1, 1) are all isomorphic, up to index 2). It turns out that for
any n > 2, by work of Cooper–Long–Thistlethwaite [CLT] there is an intricate relationship between projective
deformations and complex hyperbolic deformations of finitely generated subgroups Γ of SO(n, 1), based on the
fact that the Lie algebras of SL(n + 1, R) and SU(n, 1) are isomorphic as modules over the SO(n, 1) group ring.
Specifically, they prove:
Theorem 1.1 ([CLT]) Let Γ be a finitely generated group, and let ρ : Γ −→ SO0 (n, 1) be a smooth point of
the representation variety Hom(Γ, SL(n + 1, R)). Then ρ is also a smooth point of Hom(Γ, SU(n, 1)), and near
ρ the real dimensions of Hom(Γ, SL(n + 1, R)) and Hom(Γ, SU(n, 1)) are equal.
The primary motivation for this article is to construct examples of complex hyperbolic deformations of real
hyperbolic lattices that have nice algebraic and geometric properties. Our main result can be roughly described
as providing a sufficient condition for a deformation of a lattice Γ in H into G to continue to be faithful and
have discrete image. In what follows, the condition of being parabolic-preserving roughly means that parabolic
elements remain parabolic, see Definition 2.1 for a more precise statement.
Theorem 1.2 Let X be a negatively curved symmetric space, S a totally geodesic subspace of X and denote
G = Isom(X), H = StabG (S). Let Γ be a non-cocompact lattice in H, and let ι denote the inclusion of Γ into
G. Then any parabolic-preserving representation ρ : Γ −→ G sufficiently close to ι is discrete and faithful.
Remarks:
1. The Dehn surgery deformations of non-cocompact lattices in SO(3, 1) described above are either indiscrete
or non-faithful, showing the necessity of the parabolic-preserving assumption in general.
2. It was pointed out to us by Elisha Falbel that Theorem 1.2 still holds, with the same proof, under the
weaker hypothesis that Γ is a subgroup of a non-cocompact lattice in H, with no global fixed point in S.
If Γ is not itself a lattice, this is equivalent in this context (see [CG]) to saying that Γ is a thin subgroup
of that lattice, i.e. an infinite-index subgroup with the same Zariski-closure as the lattice.
3. When Γ is a cocompact lattice in H the result is a consequence of the following result of Guichard, as Γ
is then convex-cocompact in G.
2
Theorem 1.3 ([Gui]) Let G be a semisimple Lie group with finite center, H a rank 1 subgroup of G, Γ a
finitely generated discrete subgroup of H and denote ι : Γ −→ G the inclusion map. If Γ is convex-cocompact
then ι has a neighborhood in Hom(Γ, G) consisting entirely of discrete and faithful representations.
We prove Theorem 1.2 in section 2, then apply it in Section 3 to a family of deformations of the figure-8
knot group Γ8 < SO(3, 1) into SU(3, 1). Denoting Γ8 = π1 (S 3 \ K8 ) (where K8 is the figure-8 knot), and
ρhyp : Γ8 −→ SO(3, 1) its hyperbolic representation, i.e. the holonomy representation of the complete hyperbolic
structure on S 3 \ K8 , we obtain:
Theorem 1.4 Let Γ8 be the figure-8 knot group and ρhyp : Γ8 −→ SO(3, 1) its hyperbolic representation. Then
there exists a 1-parameter family of discrete, faithful deformations of ρhyp into SU(3, 1).
In Section 4 we apply Theorem 1.2 to a variation of the Johnson-Millson bending deformations, to obtain
the following result. As above, given a finite-volume hyperbolic manifold M = HnR /Γ, we call hyperbolic representation of Γ = π1 (M ) the holonomy representation into SO(n, 1) of the complete hyperbolic structure on M .
(This is well-defined up to conjugation by Mostow rigidity).
Theorem 1.5 For any n > 3 there exist infinitely many non-commensurable cusped hyperbolic n-manifolds
whose corresponding hyperbolic representation admits a 1-parameter family of discrete, faithful deformations
into SU(n, 1).
Here, two groups Γ, Γ′ ⊂ H are commensurable (in the wide sense) if Γ ∩ gΓ′ g −1 has finite index in both Γ
and gΓ′ g −1 for some g ∈ H. The incommensurability conclusion ensures that in each dimension n the manifolds
in Theorem 1.5 are quite distinct in the sense that they are not obtained by taking covering spaces of a single
example.
2
Discreteness and faithfulness of parabolic-preserving deformations
In this section we prove Theorem 1.2, stated in the introduction. Our strategy of proof in the non-compact
case is to use invariant horospheres, more precisely a variation of what Schwartz ([S1]) called ρ(Γ)-invariant
neutered space, see Definition 2.2 below.
From E. Cartan’s classification of real semisimple Lie groups, any negatively curved symmetric space is a
hyperbolic space HnK , with K = R, C, H or O (and n > 2 if K = R, n = 2 if K = O). We refer the reader to
[CG] for general properties of these spaces and their isometry groups. In particular isometries of such spaces
are roughly classified into the following 3 types: elliptic (having a fixed point in X), parabolic (having no fixed
point in X and exactly one on ∂∞ X) or loxodromic (having no fixed point in X and exactly two on ∂∞ X). For
our purposes we will need to distinguish between elliptic isometries with an isolated fixed point in X, which we
call single-point elliptic and elliptic isometries having boundary fixed points, which we call boundary elliptic.
Definition 2.1 Let X be a negatively curved symmetric space, G = Isom(X), and Γ a subgroup of G. A
representation ρ : Γ −→ G is called parabolic-preserving if for every parabolic (resp. boundary elliptic) element
γ ∈ Γ, ρ(γ) is again parabolic (resp. boundary elliptic).
Remark 2.1 If Γ∞ is a parabolic subgroup of Γ (i.e. a subgroup fixing a point on ∂∞ X), then any parabolicpreserving representation of Γ is faithful on Γ∞ . Indeed, all elements of Γ∞ \ {Id} are parabolic or boundary
elliptic.
Lemma 2.1 Let Γ be a discrete subgroup of G containing a parabolic element P ; denote q∞ = Fix(P ) ∈ ∂∞ X
and Γ∞ = StabΓ (q∞ ). Then, for any parabolic-preserving representation ρ : Γ −→ G, ρ(Γ∞ ) preserves each
horosphere based at Fix(ρ(P )).
Proof. It is well known that first, parabolic and boundary elliptic isometries with fixed point q∞ ∈ ∂∞ X
preserve each horosphere based at q∞ and secondly, in a discrete group of hyperbolic isometries, loxodromic
and parabolic elements cannot have a common fixed point. Therefore Γ∞ consists of parabolic and possibly
boundary elliptic isometries, and likewise for ρ(Γ∞ ) if ρ is parabolic-preserving. The only thing that remains
3
to be seen is that for any Q ∈ Γ∞ and parabolic-preserving representation ρ : Γ −→ G, ρ(Q) fixes Fix(ρ(P )).
This follows from the fact that pairs of isometries having a common fixed boundary point can be characterized
algebraically. Namely, by the assumption that P is parabolic, P and Q have a common fixed boundary point if
and only if the group hP, Qi is virtually nilpotent. This property is preserved by any representation of Γ.
Definition 2.2 Let X be a negatively curved symmetric space, G = Isom(X), Γ a subgroup of G and Γ∞ a
subgroup of Γ. We say that a (closed) horoball H∞ in X is (Γ, Γ∞ )-consistent if the following conditions hold:
(1) γH∞ = H∞ for all γ ∈ Γ∞ , and
(2) γH∞ ∩ H∞ = ∅ for all γ ∈ Γ \ Γ∞ .
Definition 2.3 Given two disjoint horoballs H1 , H2 in X, we call orthogeodesic for the pair {H1 , H2 } the
unique geodesic segment with endpoints in the boundary horospheres ∂H1 , ∂H2 and perpendicular to these horospheres; note that it is is the unique distance-minimizing geodesic segment between H1 and H2 . We will call the
set of points of ∂H1 which are endpoints of a geodesic ray perpendicular to ∂H1 and intersecting H2 the shadow
of H2 on H1 .
Remark 2.2 Since the geodesics othogonal to ∂H1 are exactly those geodesics having the vertex v1 of H1 as
an endpoint, the shadow of H2 on H1 is the intersection with ∂H1 of the geodesic cone over H2 from v1 .
Lemma 2.2 Given two disjoint horoballs H1 , H2 with orthogeodesic [x1 , x2 ], the shadow of H2 on H1 is the
intersection with ∂H1 of a closed ball centered at x1 .
Proof. Note that any isometry fixing the geodesic (x1 , x2 ) pointwise preserves H1 and H2 , hence the shadow
of H2 on H1 has rotational symmetry around x1 . The statement follows by observing that this shadow is closed,
bounded, and has non-empty-interior, which is clear in the upper half-space model of HnR , and the related Siegel
domain models of the other hyperbolic spaces, where horospheres based at the special point ∞ are horizontal
slices of the domain and geodesics through ∞ are vertical lines (see [Go] for the complex case and [KP] for the
quaternionic case).
Proposition 2.1 Let X be a negatively curved symmetric space, G = Isom(X), Γ a discrete subgroup of G
and Γ∞ a subgroup of Γ. Assume that there exists a (Γ, Γ∞ )-consistent horoball H∞ in X such that Γ∞ acts
cocompactly on the horosphere ∂H∞ . Then the set of lengths of orthogeodesics for pairs {H∞ , γH∞ } (with
γ ∈ Γ) is discrete, and each of its values is attained only finitely many times modulo the action of Γ∞ .
Proof. First note that for γ∞ ∈ Γ∞ and γ ∈ Γ\Γ∞ , d(H∞ , γH∞ ) = d(H∞ , γ∞ γH∞ ), and let K be a compact
subset of ∂H∞ whose Γ∞ -orbit covers ∂H∞ . The orbit ΓH∞ is closed, since the single cusp neighborhood
p(H∞ ) = p(ΓH∞ ) is closed in X/Γ (denoting p the projection map X −→ X/Γ), hence so is ΓH∞ \ H∞ . Then
the distance between the compact set K and the closed set ΓH∞ \ H∞ is positive and attained, say by some
point x0 ∈ K and some point in the horoball γ0 H∞ .
We claim that only finitely many Γ∞ -orbits of horospheres in ΓH∞ \ H∞ realize this minimum. To see this,
it suffices to show that any horosphere based at ∞ intersects finitely many Γ∞ -orbits of horospheres in ΓH∞ .
′
′
Fix a horosphere H∞
based at ∞. Consider a horosphere H in ΓH∞ \ H∞ that intersects H∞
, and let BH
′
′
be its shadow on H∞ . By Lemma 2.2, BH is (the intersection with H∞ of) a closed ball. There exists r > 0,
′
depending only on H∞
, such that the radius of BH is at least r. (Indeed r is the radius of the shadow of any
′
horosphere tangent to H∞
.)
Now consider two such horospheres H1 and H2 and assume they are disjoint. Call x1 and x2 the centers
′
of their shadows B1 and B2 on H∞
. We claim that the distance between x1 and x2 is at least r. If not, then
x1 ∈ B2 and x2 ∈ B1 . Consider the geodesic σ1 connecting ∞ to x1 . Since x1 is the center of B1 the intersection
σ1 ∩ H1 is the geodesic ray connecting the highest point on H1 to the endpoint of σ1 which isn’t ∞. As H1 and
H2 are disjoint, σ1 ∩ H2 is a compact geodesic segment contained in σ1 \ (σ1 ∩ H1 ). Permuting the roles of x1
and x2 gives the opposite situation on the geodesic σ2 connecting ∞ to x2 . Now if we move continuously from
4
x1 to x2 along a curve x(t) and consider the associated pencil of geodesics σt connecting ∞ to x(t), we see that
there must be a value of t for which σt ∩ H1 and σt ∩ H2 intersect, contradicting disjointness of H1 and H2 .
′
′
Finally, since K was a compact subset of H∞
whose Γ∞ -translates cover H∞
, for any horosphere H in
′
Γ · H∞ \ H∞ , we can apply an element of Γ∞ that maps the center of BH to a point of K. If H∞
meets
an infinite number of classes in Γ · H∞ we obtain in this way a sequence of distinct points in K, which must
accumulate by compactness of K. But by consistency the corresponding horospheres are disjoint, and the
previous discussion tells us that the distance between the centers of their shadows is uniformly bounded from
below, a contradiction.
The result follows inductively, repeating the argument after removing the first layer of closest horoballs.
Remark 2.3 the hypothesis that the cusp stabilizer acts cocompactly on any horosphere based at the cusp
holds for any non-cocompact lattice. In fact, it holds more generally for any discrete group with a maximal
rank parabolic subgroup.
Proposition 2.2 Let X be a negatively curved symmetric space, S a totally geodesic subspace of X and denote
G = Isom(X), H = StabG (S). If Γ is a non-cocompact lattice in H then for any parabolic-preserving representation ρ : Γ −→ G sufficiently close to the inclusion ι : Γ −→ G, there exists a (ρ(Γ), ρ(Γ∞ ))-consistent horoball,
where Γ∞ is any cusp stabilizer in Γ.
Proof. Since Γ is non-cocompact, it contains a parabolic isometry P . Let as above q∞ = Fix(P ) ∈ ∂∞ S
S
and Γ∞ = StabΓ (q∞ ). Then there exists a horoball H∞
in S, based at q∞ , which is (Γ, Γ∞ )-consistent (this
can be seen by lifting to S an embedded horoball neighborhood of the image of q∞ in the quotient S/Γ, see
S
e.g. Lemma 2.1 of [S1]). Since S is totally geodesic, H∞
is the intersection of S with a horoball H∞ in X
which is (ι(Γ), ι(Γ∞ ))-consistent. Now for any parabolic-preserving representation ρ : Γ −→ G and any γ ∈ Γ∞ ,
ρ(γ)H∞ = H∞ by Lemma 2.1, which is condition (1) of Definition 2.2 for the pair (ρ(Γ), ρ(Γ∞ )). It follows from
Proposition 2.1 that for ρ sufficiently close to ι, the horoballs γH∞ (with γ ∈ Γ \ Γ∞ ) stay disjoint from H∞
as long as some finite subcollection of them do (note that since S is totally geodesic and the horoballs convex,
the distance beween 2 horoballs H∞ and γH∞ based at points of ∂S is given by their distance in S). Hence
condition (2) of Definition 2.2 for the pair (ρ(Γ), ρ(Γ∞ )) holds for ρ sufficiently close to ι.
Proposition 2.3 Let X be a negatively curved symmetric space, denote G = Isom(X) and let Γ be a subgroup
of G without a global fixed point in X. If there exists a (Γ, Γ∞ )-consistent horoball in X for some subgroup Γ∞
of Γ, then Γ is discrete.
Proof. First assume for simplicity that Γ does not preserve any proper totally geodesic subspace of X. Then
Γ is either discrete or dense in G (Corollary 4.5.1 of [CG]). If Γ is dense in G then the orbit of any point of X
is dense in X. But if H∞ is a (Γ, Γ∞ )-consistent horoball, the orbit of any point of H∞ is entirely contained in
ΓH∞ , in which case it cannot be dense in X as ΓH∞ and X \ ΓH∞ both have nonempty interior. Therefore Γ
must be discrete.
Now if Γ does preserve a strict totally geodesic subspace of X, and if S is the minimal such subspace then by
the same argument either Γ is discrete or every orbit of a point of S is dense in S. But the consistent horoball
must be based at a point of ∂∞ S (since it is preserved by all elements of Γ∞ ), hence it intersects S along a
horoball of S and we conclude as before.
Lemma 2.3 Let X be a negatively curved symmetric space, denote G = Isom(X), let Γ be a subgroup of G
and Γ∞ a subgroup of Γ. If ρ : Γ −→ G is a parabolic-preserving representation such that there exists a
(ρ(Γ), ρ(Γ∞ ))-consistent horoball, then ρ is faithful.
Proof. Let γ ∈ Γ \ {Id}. If γ ∈ Γ∞ then ρ(γ) 6= Id by parabolic-preservation, and if γ ∈ Γ \ Γ∞ then
ρ(γ) 6= Id by condition (2) of the definition of (ρ(Γ), ρ(Γ∞ ))-consistent horoball.
Now Theorem 1.2 follows immediately from Propositions 2.2 and 2.3 and Lemma 2.3.
5
3
Deformations of the figure-8 knot group into SU(3, 1)
In this section we construct a family of parabolic-preserving deformations of the hyperbolic representation
of the figure-8 knot group into SU(3, 1). Consider Γ8 = π1 (S 3 \ K8 ) where K8 is the figure-8 knot, and denote
ρhyp : Γ8 −→ SO(3, 1) the holonomy of the complete hyperbolic structure on S 3 \ K8 .
Recall that in the presence of a smoothness hypothesis on the relevant representation varieties, Theorem 1.1
implies that the existence of deformations of ρhyp into SL(4, R) guarantees the existence of deformations of ρhyp
into SU(3, 1). Work of Ballas–Danciger–Lee [BDL] shows that the smoothness hypothesis is guaranteed in the
presence of a cohomological condition. Specifically, they prove the following.
Theorem 3.1 ([BDL]) Let M be an orientable complete finite volume hyperbolic manifold with fundamental
group Γ, and let ρhyp : Γ −→ SO(3, 1) be the holonomy representation of the complete hyperbolic structure. If
M is infinitesimally projectively rigid rel boundary, then ρhyp is a smooth point of Hom(Γ, SL(4, R)) and its
conjugacy class is a smooth point of χ(Γ, SL(4, R)).
Roughly speaking, infinitesimally projectively rigid rel boundary is a cohomological condition that says that a
certain induced map from the twisted cohomology of M into the twisted cohomology of ∂M is an injection. For a
more precise definition, see [HP]. By work of Heusener–Porti [HP], it is known that the figure-8 knot complement
is infinitesimally rigid rel boundary, and so we can apply Theorems 3.1 and 1.1 to produce deformations of ρhyp
into SU(3, 1). However, there is no reason why these representations should be parabolic-preserving, and in
many cases the deformations will not have this property.
Fortunately, work of the first author (see [B1, B2]) provides a family of deformations of ρhyp into SL(4, R)
whose corresponding deformations into SU(3, 1) are parabolic preserving.
Theorem 3.2 ([B1],[B2]) Let Γ8 be the figure-8 knot group. Then there exists a 1-parameter family of discrete,
faithful deformations of ρhyp into SL(4, R).
The construction of this 1-parameter family can be found in [B1] and ultimately constructs a curve ρt of
representations of Γ8 into SL(4, R) containing the hyperbolic representation ρhyp at t = 1/2. In fact, allowing
the parameter t to take complex values gives a 1-complex parameter family of representations into SL(4, C).
Moreover, it turns out that taking 2t to be a unit complex number u gives a 1-parameter family of representations
into SU(3, 1). (The reason for this choice of value of the parameter is that the eigenvalues of one of the peripheral
elements in ρt (Γ8 ) are 1 and a power of 2t, see Section 6 of [B2]).
We now give explicit matrices for the generators and Hermitian form for this family, using the presentation
and notation of Section 6 of [B2]. There, the following presentation of Γ8 was used:
Γ8 = hm, n | mw = wni, where w = [n, m−1 ].
The family of representations ρu : Γ8 −→ SL(4, C) is defined by ρu (m) = Mu and
1 0 1 u/2 − 1
1
0 0
0 1 1
2(1 + ū) 1 0
u/2
Mu =
and Nu =
0 0 1 (u + 1)/2
2
1 1
0 0 0
1
1
1 0
(3.1)
ρu (n) = Nu , where:
0
0
(3.2)
0
1
When |u| = 1, the group ρu (Γ8 ) preserves the Hermitian form Hu on C4 given by H(X, Y ) = X T Ju Ȳ ,
where:
1 + (u + ū)/2
−1 − (u + ū)/2
1+u
−3 − 2(u + ū) − ū2
−1 − (u + ū)/2
1 + (u + ū)/2
−1 − u
1+u
Ju =
(3.3)
1 + ū
−1 − ū
4 + 2(u + ū)
−4 − 2(u + ū)
−3 − 2(u + ū) − u2
1 + ū
−4 − 2(u + ū)
4 + 2(u + ū)
Lemma 3.1 The form Hu has signature (3,1) for all u = eiα with |α| < 2π/3, and signature (2,2) when
α ∈ ±(2π/3, π).
6
Proof. Computing the determinant of Ju gives:
det Ju
= −96 − 83(u + ū) − 53(u2 + ū2 ) − 24(u3 + ū3 ) − 7(u4 + ū4 ) − (u5 + ū5 )
= −96 − 166cos(α) − 106cos(2α) − 48cos(3α) − 14cos(4α) − 2cos(5α)
= −4(cos(α) + 1)2 (2cos(α) + 1)3 .
The latter function of α is negative for |α| < 2π/3, and positive for α ∈ ±(2π/3, π); the result then follows by
noting that Hu has signature (3,1) when u = 1 (corresponding to the hyperbolic representation), and (2,2) for
e.g. u = ±3π/4.
Lemma 3.2 The representations ρu are pairwise non-conjugate in SL(4, C).
Proof. A straightforward computation gives: Tr Mu Nu = 6 + u.
Lemma 3.3 The representations ρu are parabolic-preserving.
Proof. The peripheral subgroup Γ∞ of Γ8 is generated by m and l = wwop = nm−1 n−1 m2 n−1 m−1 n, with
the notation of the presentation (3.1) (see [B1]). Now Mu = ρu (m) is unipotent for all u, and a straightforward
computation (using eg Maple) shows that Lu = ρu (l) is non-diagonalizable (with eigenvalues (u, u, u, ū3)) for
all u, hence parabolic. Since Γ∞ ≃ Z2 , all elements of ρu (Γ∞ ) = hρu (m), ρu (l)i will also remain parabolic for u
in a neighborhood of 1.
The previous result along with Theorem 1.2 has the following immediate corollary
Corollary 3.3 The representations ρu are discrete and faithful for u in some neighborhood of 1 in U(1).
It would be interesting to know how far u can get from 1 before discreteness or faithfulness is lost.
4
Bending deformations
In this section we construct additional examples in arbitrary dimensions, proving Theorem 1.5 stated in the
introduction. We start with a cusped hyperbolic manifold M = HnR /Γ, and ρhyp : Γ −→ SO(n, 1) the hyperbolic
representation of Γ = π1 (M ), i.e. the holonomy representation of the complete hyperbolic structure on M . We
will construct a one-parameter family of representations ρθ : Γ −→ SU(n, 1), (θ ∈ S 1 ) such that ρ0 = ρhyp , using
the bending procedure described by Johnson–Millson ([JM]). Their construction is quite general and allows one
to deform representations in a variety of Lie groups. We briefly outline how to use bending to produce families
of representations in the complex hyperbolic setting.
Define a Hermitian form H on Cn+1 via the formula H(X, Y ) = X T Jn+1 Ȳ where Jn+1 is the diagonal
matrix Diag(1, ..., 1, −1), with signature (n, 1). Using this form we produce a projective model for HnC given by
HnC = {[V ] ∈ CP n | H(V, V ) < 0}
n−1
Using the splitting Cn+1 = C × Cn we can embed HC
into the CP n−1 corresponding to the second factor.
n−1
We will refer to this copy of HC as H0 . Using this embedding we can identify U(n − 1, 1) with the intersection
of SU(n, 1) and the stabilizer of the second factor, and we will refer to this subgroup as U0 (n − 1, 1). It is well
n−1
known that all other copies of HC
inside HnC are isometric to H0 , and similarly, all copies of U(n − 1, 1) inside
SU(n, 1) are conjugate to U0 (n − 1, 1).
Let Cn−1 denote the identity component of the centralizer of U0 (n − 1, 1) in SU(n, 1). Cn−1 is a onedimensional Lie group isomorphic to S 1 and can be written explicitly in block form as
iθ
e
0
Cn−1 = Mθ =
|
θ
∈
R/2πZ
0 e−iθ/n In
7
Let Γ be a lattice in SO(n, 1) ⊂ SU(n, 1); then M = HnR /Γ is a finite volume hyperbolic n-orbifold. For
simplicity, we will assume that Γ is torsion-free and thus M will be a hyperbolic manifold. Suppose that M
contains an embedded orientable totally geodesic hypersurface Σ. By applying a conjugacy of SO(n, 1) we
n−1
n−1
can assume that Σ = HR
/∆ where HR
is thought of as the set of real points of H0 and ∆ is a lattice in
U0 (n − 1, 1) ∩ SO(n, 1).
The hypersurface Σ provides a decomposition of Γ into either an amalgamated free product or an HNN
extension, depending on whether or not Σ is separating. Using this decomposition we can construct a family
ρθ : Γ → SU(n, 1) such that ρ0 = ι, where ι is the inclusion of Γ into SU(n, 1), as follows.
If Σ is separating, then M \Σ consists of two connected components M1 and M2 , with fundamental groups
Γ1 and Γ2 respectively. In this case Γ = Γ1 ∗∆ Γ2 . The group Γ is generated by Γ1 ∪ Γ2 and we define
ι(γ)
γ ∈ Γ1
ρθ (γ) =
Mθ ι(γ)Mθ−1 γ ∈ Γ2
on this generating set. Since Mθ centralizes ∆ we see that the relations coming from the amalgamated
product decomposition are satisfied, and so ρθ : Γ → SU(n, 1) is well defined.
If Σ is non-separating, then M ′ = M \Σ is connected. If we let Γ′ be the fundamental group of M ′ then we
can arrive at the decomposition Γ = Γ′ ∗t . In this case Γ is generated by Γ′ ∪ {t}, where t is a free letter and we
define ρθ on generators as
ι(γ)
γ ∈ Γ′
ρθ (γ) =
Mθ ι(γ) γ = t
Again, since Mθ centralizes ∆ we see that the relations for the HNN extension are satisfied and so ρθ : Γ →
SU(n, 1) is well defined. The representations constructed above are called bending deformations of Γ along ∆,
or just bending deformations if Γ and ∆ are clear from context. By work of Johnson–Millson [JM] this path of
representations is in fact a deformation of ρhyp (i.e. the ρθ are pairwise non-conjugate for small values of θ).
Proof. (proof of Theorem 1.5) We proceed by constructing infinitely many commensurability classes of
cusped hyperbolic manifolds containing totally geodesic hypersurfaces. This is done via a well known arithmetic
construction (see [Ber]). The rough idea is to look at the group, Γ, of integer points of the orthogonal groups
of various carefully selected quadratic forms of signature (n, 1). The quotient M = HnR /Γ will be a cusped
hyperbolic n-orbifold containing a totally geodesic hypersurface. After passing to a carefully selected cover we
can produce our parabolic preserving representations via the bending construction.
We now discuss the details for a specific form and observe that the proof is essentially unchanged if one
ˆ = Γ̂ ∩ U0 (n − 1, 1). The group Γ̂ clearly
selects a different form. Let Γ̂ = SL(n + 1, Z) ∩ SO(n, 1) and let ∆
contains unipotent elements and so we see that M̂ = HnR /Γ̂ is a cusped hyperbolic n-orbifold, which contains
n−1 ˆ
an immersed totally geodesic codimension-1 suborbifold isomorphic to Σ̂ = HR
/∆. By combining work of
Bergeron (Théorème 1 of [Ber]) and McReynolds–Reid–Stover (Proposition 3.1 of [MRS]) we can find finite
ˆ and corresponding manifolds M = Hn /Γ and Σ = Hn−1 /∆ with the
index subgroups Γ ⊂ Γ̂ and ∆ ⊂ ∆
R
R
following properties.
• Γ is torsion-free
• Σ is embedded in M
• M has only torus cusps.
Each M contains the totally geodesic hypersurface Σ along which we can bend to produce a family ρθ of
representations from Γ into SU(n, 1). We now show that the representatons ρθ are parabolic-preserving; then
by Theorem 1.2 the ρθ are discrete and faithful for small values of θ.
Lemma 4.1 The representations ρθ : Γ −→ SU(n, 1) obtained by bending Γ along ∆ are parabolic-preserving.
Proof. By construction, we have arranged that Γ is torsion-free, and so there are no elliptic elements to
consider. Furthermore, the only parabolic elements of Γ correspond to loops in M that are freely homotopic to
one of the torus cusps. We now discuss how such an element is modified when one bends. Let γ be a parabolic
8
element of Γ and let q∞ be its fixed point on ∂∞ HnC . There is a foliation of HnC by horospheres centered at q∞ and
γ preserves this foliation leafwise. Furthermore, leafwise preservation of this foliation characterizes parabolic
isometries of HnC that fix p∞ . Thus it suffices to show that ρθ (γ) preserves this foliation.
Regard γ as a loop in M based at x0 ∈ Σ and lift γ to a path γ̃ in M̃ ⊂ HnC based at x̃0 . Let Σ̃ be the
lift of Σ that contains x̃0 . Each time γ̃ intersects a lift of Σ to HnC (counted with orientation) the holonomy
is modified by composing with a Heisenberg rotation of angle εi θ centered at q∞ that acts as the identity on
Σ̃i . Each of these modifications is by an element of SU(n, 1) that leafwise preserves the foliation of horospheres
centered at q∞ , and so ρθ (γ) also preserves this foliation leafwise, and is thus parabolic. More specifically, if we
Pk
let ε = i=1 εi , then there are two cases. If ε = 0 then ρθ (γ) is a unipotent parabolic which is conjugate to γ.
If ε 6= 0 then ρθ (γ) is an ellipto-parabolic isometry, whose angle of rotation is εθ.
See Apanasov [Ap, §4] for a detailed description in the n = 2 case.
Remark 4.1 It is well known, see [T] or more generally [HT], that the complement in S 3 of the figure-8 knot
does not contain an embedded totally geodesic hypersurface. Therefore, the deformations produced in Theorem
1.4 are distinct from those produced by Theorem 1.5.
References
[Ap]
B. Apanasov; Bending deformations of complex hyperbolic surfaces. J. Reine Angew. Math. 492
(1997), 75–91.
[B1]
S.A. Ballas; Deformations of noncompact projective manifolds. Algebr. Geom. Topol. 14 (2014), no. 5,
2595–2625.
[B2]
S.A. Ballas; Finite volume properly convex deformations of the figure-eight knot. Geom. Dedicata 178
(2015), 49–73.
[BDL] S.A. Ballas, J. Danciger, G.-S. Lee; Convex projective structures on non-hyperbolic three-manifolds.
Preprint 2015, arXiv 1508.0479.
[BL]
S.A. Ballas, D.D. Long; Constructing thin subgroups commensurable with the figure-eight knot group.
Algebr. Geom. Topol. 15 (2015), no. 5, 3011–3024.
[BM]
S.A. Ballas, L. Marquis; Properly convex bending of hyperbolic manifolds. Preprint 2016, arXiv
1609.03046.
[BSc]
A. Bart and K. Scannell; The generalized cuspidal cohomology problem. Canad. J. Math. 58 (2006),
no. 4, 673–690.
[Ber]
N. Bergeron; Premier nombre de Betti et spectre du laplacien de certaines variétés hyperboliques,
Enseign. Math. (2). (2000), 109–137.
[Bers] L. Bers; Simultaneous uniformization, Bull. Amer. Math. Soc. (66). (1960), 94–97.
[CG]
S. Chen, L. Greenberg; Hyperbolic spaces, in Contributions to Analysis. Academic Press, New York
(1974), 49–87.
[CLT] D. Cooper, D.D. Long and M.B. Thistlethwaite; Flexing closed hyperbolic manifolds. Geom. Topol. 11
(2007), 2413–2440.
[CLTII] D. Cooper, D.D. Long and M.B. Thistlethwaite; Computing varieties of representations of hyperbolic
3-manifolds into SL(4, R).. J. of Experimental Math. vol. 15 (2006) pp. 291 – 305.
[GR]
H. Garland, M.S. Raghunathan; Fundamental domains for lattices in (R-)rank 1 semisimple Lie
groups, Ann. of Math. (2) 92 (1970), 279–326.
9
[Go]
W.M. Goldman; Complex Hyperbolic Geometry. Oxford Mathematical Monographs. Oxford University
Press (1999).
[GPS] M. Gromov, I. Piatetski-Shapiro; Non-arithmetic groups in Lobachevsky spaces. Publ. Math. IHES 66
(1987), 93–103.
[Gui]
O. Guichard; Groupes plongés quasi-isométriquement dans un goupe de Lie. Math. Ann. 330 (2004),
no. 2, 331–351.
[HT]
A. Hatcher and W. Thurston; Incompressible surfaces in 2-bridge knot complements. Invent. Math. 79
(1985), no. 2, 225–246.
[HP]
M. Heusener and J. Porti; Infinitesimal projective rigidity under Dehn filling. Geom. Topol. 15 (2011),
no. 4, 2017–2071.
[JM]
D. Johnson and J. Millson; Deformation Spaces Associated to Compact Hyperbolic Manifolds, in
Papers in Honor of G.D. Mostow on His Sixtieth Birthday (Roger Howe, ed.). Birkhäuser, Progress in
Mathematics 67 (1986), 48–106.
[Kap] M. Kapovich; Deformations of representations of discrete subgroups of SO(3, 1). Math. Annalen. 299
(1994), 341–354.
[KP]
I. Kim, J. R. Parker; Geometry of quaternionic hyperbolic manifolds. Math. Proc. Camb. Phil. Soc.
135 (2003), 291–320.
[PP]
J.R. Parker, I.D. Platis; Complex hyperbolic quasi-Fuchsian groups, in: Geometry of Riemann
Surfaces. London Mathematical Society Lecture Notes 368 (2010) 309–355.
[MRS] D.B. McReynolds. and A.W. Reid, and M. Stover; Collisions at infinity in hyperbolic manifolds, Math.
Proc. Cambridge Philos. Soc. (2013), 459–463.
[Sc]
K. Scannell; Local rigidity of hyperbolic 3-manifolds after Dehn surgery. Duke Math. J. 114 (2002),
no. 1, 1–14.
[S1]
R.E. Schwartz; The quasi-isometry classification of rank one lattices. Publ Math. IHES 82 (1995),
133–168.
[S2]
R.E. Schwartz; Complex hyperbolic triangle groups, from: Proceedings of the International Congress of
Mathematicians, Vol. II (Beijing, 2002), Higher Ed. Press, Beijing (2002) 339–349.
[T]
W.P. Thurston; The Geometry and Topology of Three-Manifolds. Electronic version 1.1 (2002).
Available at: http://www.msri.org/publications/books/gt3m/
[W]
A. Weil; Discrete subgroups of Lie groups (II), Ann. of Math. (2) 75 (1962), 578–602.
Samuel Ballas
Department of Mathematics, Florida State University
[email protected]
Julien Paupert
School of Mathematical and Statistical Sciences, Arizona State University
[email protected]
Pierre Will
Institut Fourier, Université de Grenoble I
[email protected]
10
| 4math.GR
|
Published as a conference paper at ICLR 2018
ON
THE IMPORTANCE OF SINGLE DIRECTIONS FOR
GENERALIZATION
arXiv:1803.06959v1 [stat.ML] 19 Mar 2018
Ari S. Morcos1 , David G.T. Barrett, Neil C. Rabinowitz, & Matthew Botvinick
DeepMind
London, UK
{arimorcos,barrettdavid,ncr,botvinick}@google.com
A BSTRACT
Despite their ability to memorize large datasets, deep neural networks often
achieve good generalization performance. However, the differences between the
learned solutions of networks which generalize and those which do not remain
unclear. Additionally, the tuning properties of single directions (defined as the
activation of a single unit or some linear combination of units in response to some
input) have been highlighted, but their importance has not been evaluated. Here,
we connect these lines of inquiry to demonstrate that a network’s reliance on single
directions is a good predictor of its generalization performance, across networks
trained on datasets with different fractions of corrupted labels, across ensembles
of networks trained on datasets with unmodified labels, across different hyperparameters, and over the course of training. While dropout only regularizes this
quantity up to a point, batch normalization implicitly discourages single direction
reliance, in part by decreasing the class selectivity of individual units. Finally,
we find that class selectivity is a poor predictor of task importance, suggesting
not only that networks which generalize well minimize their dependence on individual units by reducing their selectivity, but also that individually selective units
may not be necessary for strong network performance.
1
I NTRODUCTION
Recent work has demonstrated that deep neural networks (DNNs) are capable of memorizing extremely large datasets such as ImageNet (Zhang et al., 2017). Despite this capability, DNNs in practice achieve low generalization error on tasks ranging from image classification (He et al., 2015)
to language translation (Wu et al., 2016). These observations raise a key question: why do some
networks generalize while others do not?
Answers to these questions have taken a variety of forms. A variety of studies have related generalization performance to the flatness of minima and PAC-Bayes bounds (Hochreiter & Schmidhuber,
1997, Keskar et al., 2017, Neyshabur et al., 2017, Dziugaite & Roy, 2017), though recent work
has demonstrated that sharp minima can also generalize (Dinh et al., 2017). Others have focused
on the information content stored in network weights (Achille & Soatto, 2017), while still others
have demonstrated that stochastic gradient descent itself encourages generalization (Bousquet &
Elisseeff, 2002, Smith & Le, 2017, Wilson et al., 2017).
Here, we use ablation analyses to measure the reliance of trained networks on single directions. We
define a single direction in activation space as the activation of a single unit or feature map or some
linear combination of units in response to some input. We find that networks which memorize the
training set are substantially more dependent on single directions than those which do not, and that
this difference is preserved even across sets of networks with identical topology trained on identical
data, but with different generalization performance. Moreover, we found that as networks begin to
overfit, they become more reliant on single directions, suggesting that this metric could be used as a
signal for early stopping.
1
Corresponding author: [email protected]
1
Published as a conference paper at ICLR 2018
We also show that networks trained with batch normalization are more robust to cumulative ablations than networks trained without batch normalization and that batch normalization decreases the
class selectivity of individual feature maps, suggesting an alternative mechanism by which batch
normalization may encourage good generalization performance. Finally, we show that, despite the
focus on selective single units in the analysis of DNNs (and in neuroscience; Le et al., 2011, Zhou
et al., 2014, Radford et al., 2017, Britten et al., 1992), the class selectivity of single units is a poor
predictor of their importance to the network’s output.
2
A PPROACH
In this study, we will use a set of perturbation analyses to examine the relationship between a network’s generalization performance and its reliance upon single directions in activation space. We
will then use a neuroscience-inspired measure of class selectivity to compare the selectivity of individual directions across networks with variable generalization performance and examine the relationship between class selectivity and importance.
2.1
S UMMARY OF MODELS AND DATASETS ANALYZED
We analyzed three models: a 2-hidden layer MLP trained on MNIST, an 11-layer convolutional
network trained on CIFAR-10, and a 50-layer residual network trained on ImageNet. In all experiments, ReLU nonlinearities were applied to all layers but the output. Unless otherwise noted, batch
normalization was used for all convolutional networks (Ioffe & Szegedy, 2015). For the ImageNet
ResNet, top-5 accuracy was used in all cases.
Partially corrupted labels As in Zhang et al. (2017), we used datasets with differing fractions
of randomized labels to ensure varying degrees of memorization. To create these datasets, a given
fraction of labels was randomly shuffled and assigned to images, such that the distribution of labels
was maintained, but any true patterns were broken.
2.2
P ERTURBATION ANALYSES
Ablations We measured the importance of a single direction to the network’s computation by
asking how the network’s performance degrades once the influence of that direction was removed.
To remove a coordinate-aligned single direction , we clamped the activity of that direction to a fixed
value (i.e., ablating the direction). Ablations were performed either on single units in MLPs or an
entire feature map in convolutional networks. For brevity, we will refer to both of these as ‘units.’
Critically, all ablations were performed in activation space, rather than weight space.
More generally, to evaluate a network’s reliance upon sets of single directions, we asked how the network’s performance degrades as the influence of increasing subsets of single directions was removed
by clamping them to a fixed value (analogous to removing increasingly large subspaces within activation space). This analysis generates curves of accuracy as a function of the number of directions
ablated: the more reliant a network is on low-dimensional activation subspaces, the more quickly
the accuracy will drop as single directions are ablated.
Interestingly, we found that clamping the activation of a unit to the empirical mean activation across
the training or testing set was more damaging to the network’s performance than clamping the activation to zero (see Appendix A.1). We therefore clamped activity to zero for all ablation experiments.
Addition of noise As the above analyses perturb units individually, they only measure the influence
of coordinate-aligned single directions. To test networks’ reliance upon random single directions,
we added Gaussian noise to all units with zero mean and progressively increasing variance. To
scale the variance appropriately for each unit, the variance of the noise added was normalized by the
empirical variance of the unit’s activations across the training set.
2.3
Q UANTIFYING CLASS SELECTIVITY
To quantify the class selectivity of individual units, we used a metric inspired by the selectivity indices commonly used in systems neuroscience (De Valois et al., 1982, Britten et al., 1992, Freedman
2
Published as a conference paper at ICLR 2018
Figure 1: Memorizing networks are more sensitive to cumulative ablations. Networks were
trained on MNIST (2-hidden layer MLP, a), CIFAR-10 (11-layer convolutional network, b), and
ImageNet (50-layer ResNet, c). In a, all units in all layers were ablated, while in b and c, only
feature maps in the last three layers were ablated. Error bars represent standard deviation across 10
random orderings of units to ablate.
& Assad, 2006). The class-conditional mean activity was first calculated across the test set, and the
selectivity index was then calculated as follows:
µmax − µ−max
selectivity =
(1)
µmax + µ−max
with µmax representing the highest class-conditional mean activity and µ−max representing the
mean activity across all other classes. For convolutional feature maps, activity was first averaged
across all elements of the feature map. This metric varies from 0 to 1, with 0 meaning that a unit’s
average activity was identical for all classes, and 1 meaning that a unit was only active for inputs of
a single class.
We note that this metric is not a perfect measure of information content in single units; for example,
a unit with a little information about every class would have a low class selectivity index. However,
it does measure the discriminability of classes along a given direction. The selectivity index also
identifies units with the same class tuning properties which have been highlighted in the analysis
of DNNs (Le et al., 2011, Zeiler & Fergus, 2014, Coates et al., 2012, Zhou et al., 2014, Radford
et al., 2017). However, in addition to class selectivity, we replicate all of our results using mutual
information, which, in contrast to class selectivity, should highlight units with information about
multiple classes, and we find qualitively similar outcomes (Appendix A.5). We also note that while
a class can be viewed as a highly abstract feature, implying that our results may generalize to feature
selectivity, we do not examine feature selectivity in this work.
3
E XPERIMENTS
3.1
G ENERALIZATION
Here, we provide a rough intuition for why a network’s reliance upon single directions might be
related to generalization performance. Consider two networks trained on a large, labeled dataset
with some underlying structure. One of the networks simply memorizes the labels for each input
example and will, by definition, generalize poorly (‘memorizing network’) while the other learns
the structure present in the data and generalizes well (‘structure-finding network’). The minimal
description length of the model should be larger for the memorizing network than for the structurefinding network. As a result, the memorizing network should use more of its capacity than the
structure-finding network, and by extension, more single directions. Therefore, if a random single
direction is perturbed, the probability that this perturbation will interfere with the representation of
the data should be higher for the memorizing network than for the structure-finding network2 .
2
Assuming that the memorizing network uses a non-negligible fraction of its capacity.
3
Published as a conference paper at ICLR 2018
a
b
Figure 2: Memorizing networks are more sensitive to random noise. Networks were trained
on MNIST (2-hidden layer MLP, a), and CIFAR-10 (11-layer convolutional network, b). Noise
was scaled by the empirical variance of each unit on the training set. Error bars represent standard
deviation across 10 runs. X-axis is on a log scale.
a
b
Figure 3: Networks which generalize poorly are more reliant on single directions. 200 networks with identical topology were trained on unmodified CIFAR-10. a, Cumulative ablation curves
for the best and worst 5 networks by generalization error. Error bars represent standard deviation
across 5 models and 10 random orderings of feature maps per model. b, Area under cumulative
ablation curve (normalized) as a function of generalization error.
To test whether memorization leads to greater reliance on single directions, we trained a variety of
network types on datasets with differing fractions of randomized labels and evaluated their performance as progressively larger fractions of units were ablated (see Sections 2.2 and 2.1). By definition, these curves must begin at the network’s training accuracy (approximately 1 for all networks
tested) and fall to chance levels when all directions have been ablated. To rule out variance due to
the specific order of unit ablation, all experiments were performed with mutliple random ablation
orderings of units. As many of the models were trained on datasets with corrupted labels and, by
definition, cannot generalize, training accuracy was used to evaluate model performance. Consistent with our intuition, we found that networks trained on varying fractions of corrupted labels were
significantly more sensitive to cumulative ablations than those trained on datasets comprised of true
labels, though curves were not always perfectly ordered by the fraction of corrupted labels (Fig. 1).
We next asked whether this effect was present if networks were perturbed along random bases. To
test this, we added noise to each unit (see Section 2.2). Again, we found that networks trained on
corrupted labels were substantially and consistently more sensitive to noise added along random
bases than those trained on true labels (Fig. 2).
The above results apply to networks which are forced to memorize at least a portion of the training
set – there is no other way to solve the task. However, it is unclear whether these results would
apply to networks trained on uncorrupted data. In other words, do the solutions found by networks
with the same topology and data, but different generalization performance exhibit differing reliance
upon single directions? To test this, we trained 200 networks on CIFAR-10, and evaluated their
generalization error and reliance on single directions. All networks had the same topology and
4
Published as a conference paper at ICLR 2018
were trained on the same dataset (unmodified CIFAR-10). Individual networks only differed in their
random initialization (drawn from identical distributions) and the data order used during training.
We found that the 5 networks with the best generalization performance were more robust to the
ablation of single directions than the 5 networks with the worst generalization performance (Fig. 3a).
To quantify this further, we measured the area under the ablation curve for each of the 200 networks
and plotted it as a function of generalization error (Fig. 3b). Interestingly, networks appeared to
undergo a discrete regime shift in their reliance upon single directions; however, this effect might
have been caused by degeneracy in the set of solutions found by the optimization procedure, and we
note that there was also a negative correlation present within clusters (e.g., top left cluster). These
results demonstrate that the relationship between generalization performance and single direction
reliance is not merely a side-effect of training with corrupted labels, but is instead present even
among sets networks with identical training data.
3.2
R ELIANCE ON SINGLE DIRECTIONS AS A SIGNAL FOR MODEL SELECTION
This relationship raises an intriguing question: can single direction reliance be used to estimate
generalization performance without the need for a held-out test set? And if so, might it be used
as a signal for early stopping or hyperpameter selection? As a proof-of-principle experiment for
early stopping, we trained an MLP on MNIST and measured the area under the cumulative ablation
curve (AUC) over the course of training along with the train and test loss. Interestingly, we found
that the point in training at which the AUC began to drop was the same point that the train and test
loss started to diverge (Fig. 4a). Furthermore, we found that AUC and test loss were negatively
correlated (Spearman’s correlation: -0.728; Fig. 4b).
As a proof-of-principle experiment for hyperparameter selection, we trained 192 CIFAR-10 models
with different hyperparemeter settings (96 hyperparameters with 2 repeats each; see Appendix A.2).
We found that AUC and test accuracy were highly correlated (Spearman’s correlation: 0.914; Fig.
4c), and by performing random subselections of 48 hyperparameter settings, AUC selected one of
the top 1, 5, and 10 settings 13%, 83%, and 98% of the time, respectively, with an average difference
in test accuracy between the best model selected by AUC and the optimal model of only 1 ± 1.1%
(mean ± std). These results suggest that single direction reliance may serve as a good proxy for
hyperparameter selection and early stopping, but further work will be necessary to evaluate whether
these results hold in more complicated datasets.
3.3
R ELATIONSHIP TO DROPOUT AND BATCH NORMALIZATION
Dropout Our experiments are reminiscent of using dropout at training time, and upon first inspection, dropout may appear to discourage networks’ reliance on single directions (Srivastava et al.,
2014). However, while dropout encourages networks to be robust to cumulative ablations up until
the dropout fraction used in training, it should not discourage reliance on single directions past that
point. Given enough capacity, a memorizing network could effectively guard against dropout by
merely copying the information stored in a given direction to several other directions. However, the
network will only be encouraged to make the minimum number of copies necessary to guard against
the dropout fraction used in training, and no more. In such a case, the network would be robust
to dropout so long as all redundant directions were not simultaneously removed, yet still be highly
reliant on single directions past the dropout fraction used in training.
To test whether this intuition holds, we trained MLPs on MNIST with dropout probabilities
∈ {0.1, 0.2, 0.3} on both corrupted and unmodified labels. Consistent with the observation in Arpit
et al. (2017), we found that networks with dropout trained on randomized labels required more
epochs to converge and converged to worse solutions at higher dropout probabilities, suggesting that
dropout does indeed discourage memorization. However, while networks trained on both corrupted
and unmodified labels exhibited minimal loss in training accuracy as single directions were removed
up to the dropout fraction used in training, past this point, networks trained on randomized labels
were much more sensitive to cumulative ablations than those trained on unmodified labels (Fig.
5a). Interestingly, networks trained on unmodified labels with different dropout fractions were all
similarly robust to cumulative ablations. These results suggest that while dropout may serve as an
effective regularizer to prevent memorization of randomized labels, it does not prevent over-reliance
on single directions past the dropout fraction used in training.
5
Published as a conference paper at ICLR 2018
a
b
c
Figure 4: Single direction reliance as a signal for hyperparameter selection and early stopping.
a, Train (blue) and test (purple) loss, along with the normalized area under the cumulative ablation
curve (AUC; green) over the course of training for an MNIST MLP. Loss y-axis has been cropped
to make train/test divergence visible. b, AUC and test loss for a CIFAR-10 ConvNet are negatively
correlated over the course of training. c, AUC and test accuracy are positively corrleated across a
hyperparameter sweep (96 hyperparameters with 2 repeats for each).
a
b
Figure 5: Impact of regularizers on networks’ reliance upon single directions. a, Cumulative
ablation curves for MLPs trained on unmodified and fully corrupted MNIST with dropout fractions
∈ {0.1, 0.2, 0.3}. Colored dashed lines indicate number of units ablated equivalent to the dropout
fraction used in training. Note that curves for networks trained on corrupted MNIST begin to drop
soon past the dropout fraction with which they were trained. b, Cumulative ablation curves for
networks trained on CIFAR-10 with and without batch normalization. Error bars represent standard
deviation across 4 model instances and 10 random orderings of feature maps per model.
6
Published as a conference paper at ICLR 2018
a
b
Figure 6: Batch normalization decreases class selectivity and increases mutual information.
Distributions of class selectivity (a) and mutual information (b) for networks trained with (blue)
and without batch normalization (purple). Each distribution comprises 4 model instances trained on
uncorrupted CIFAR-10.
Batch normalization In contrast to dropout, batch normalization does appear to discourage reliance upon single directions. To test this, we trained convolutional networks on CIFAR-10 with and
without batch normalization and measured their robustness to cumulative ablation of single directions. Networks trained with batch normalization were consistently and substantially more robust to
these ablations than those trained without batch normalization (Fig. 5b). This result suggests that in
addition to reducing covariate shift, as has been proposed previously (Ioffe & Szegedy, 2015), batch
normalization also implicitly discourages reliance upon single directions.
3.4
R ELATIONSHIP BETWEEN CLASS SELECTIVITY AND IMPORTANCE
Our results thus far suggest that networks which are less reliant on single directions exhibit better
generalization performance. This result may appear counter-intuitive in light of extensive past work
in both neuroscience and deep learning which highlights single units or feature maps which are
selective for particular features or classes (Le et al., 2011, Zeiler & Fergus, 2014, Coates et al.,
2012, Zhou et al., 2014, Radford et al., 2017). Here, we will test whether the class selectivity of
single directions is related to the importance of these directions to the network’s output.
First, we asked whether batch normalization, which we found to discourage reliance on single directions, also influences the distribution of information about class across single directions. We
used the selectivity index described above (see Section 2.3) to quantify the discriminability between
classes based on the activations of single feature maps across networks trained with and without
batch normalization. Interestingly, we found that while networks trained without batch normalization exhibited a large fraction of feature maps with high class selectivity3 , the class selectivity of
feature maps in networks trained with batch normalization was substantially lower (Fig. 6a). In contrast, we found that batch normalization increases the mutual information present in feature maps
(Fig. 6b). These results suggest that batch normalization actually discourages the presence of feature maps with concentrated class information and rather encourages the presence of feature maps
with information about multiple classes, raising the question of whether or not such highly selective
feature maps are actually beneficial.
We next asked whether the class selectivity of a given unit was predictive of the impact on the
network’s loss of ablating said unit. Since these experiments were performed on networks trained
on unmodified labels, test loss was used to measure network impact. For MLPs trained on MNIST,
we found that there was a slight, but minor correlation (Spearman’s correlation: 0.095) between
a unit’s class selectivity and the impact of its ablation, and that many highly selective units had
minimal impact when ablated (Fig. 7a). By analyzing convolutional networks trained on CIFAR-10
and ImageNet, we again found that, across layers, the ablation of highly selective feature maps was
no more impactful than the ablation of non-selective feature maps (Figs. 7b and 7d). In fact, in the
CIFAR-10 networks, there was actually a negative correlation between class selectivity and feature
map importance (Spearman’s correlation: -0.428, Fig. 7b). To test whether this relationship was
depth-dependent, we calculated the correlation between class selectivity and importance separately
for each layer, and found that the vast majority of the negative correlation was driven by early
3
And dead feature maps. Feature maps with no activity would have a selectivity index of 0.
7
Published as a conference paper at ICLR 2018
CIFAR-10 ConvNet
MNIST MLP
a
b
c
ImageNet ResNet
d
e
Figure 7: Selective and non-selective directions are similarly important. Impact of ablation
as a function of class selectivity for MNIST MLP (a), CIFAR-10 convolutional network (b-c), and
ImageNet ResNet (d-e). c and e show regression lines for each layer separately.
layers, while later layers exhibited no relationship between class selectivity and importance (Figs.
7c and 7e). Interestingly, in all three networks, ablations in early layers were more impactful than
ablations in later layers, consistent with theoretical observations (Raghu et al., 2016). Additionally,
we performed all of the above experiments with mutual information in place of class selectivity, and
found qualitatively similar results (Appendix A.5).
As a final test, we compared the class selectivity to the L1 -norm of the filter weights, a metric
which has been found to be a successful predictor of feature map importance in the model pruning
literature (Li et al., 2017). Consistent with our previous observations, we found that class selectivity
was largely unrelated to the L1 -norm of the filter weights, and if anything, the two were negatively
correlated (Fig. A3, see Appendix A.4 for details). Taken together, these results suggest that class
selectivity is not a good predictor of importance, and imply that class selectivity may actually be
detrimental to network performance. Further work will be necessary to examine whether class
and/or feature selectivity is harmful or helpful to network performance.
4
R ELATED WORK
Much of this work was directly inspired by Zhang et al. (2017), and we replicate their results using
partially corrupted labels on CIFAR-10 and ImageNet. By demonstrating that memorizing networks
are more reliant on single directions, we also provide an answer to one of the questions they posed:
is there an empirical difference between networks which memorize and those which generalize?
Our work is also related to work linking generalization and the sharpness of minima (Hochreiter
& Schmidhuber, 1997, Keskar et al., 2017, Neyshabur et al., 2017). These studies argue that flat
minima generalize better than sharp minima (though Dinh et al. (2017) recently found that sharp
minima can also generalize well). This is consistent with our work, as flat minima should correspond
to solutions in which perturbations along single directions have little impact on the network output.
Another approach to generalization has been to contextualize it in information theory. For example, Achille & Soatto (2017) demonstrated that networks trained on randomized labels store more
8
Published as a conference paper at ICLR 2018
information in their weights than those trained on unmodfied labels. This notion is also related to
Shwartz-Ziv & Tishby (2017), which argues that during training, networks proceed first through a
loss minimization phase followed by a compression phase. Here again, our work is consistent, as
networks with more information stored in their weights (i.e., less compressed networks) should be
more reliant upon single directions than compressed networks.
More recently, Arpit et al. (2017) analyzed a variety of properties of networks trained on partially
corrupted labels, relating performance and time-to-convergence to capacity. They also demonstrated
that dropout, when properly tuned, can serve as an effective regularizer to prevent memorization.
However, we found that while dropout may discourage memorization, it does not discourage reliance
on single directions past the dropout probability.
We found that class selectivity is a poor predictor of unit importance. This observation is consistent
with a variety of recent studies in neuroscience. In one line of work, the benefits of neural systems
which are robust to coordinate-aligned noise have been explored (Barrett et al. (2016), Montijn
et al. (2016)). Another set of studies have demonstrated the presence of neurons with multiplexed
information about many stimuli and have shown that task information can be decoded with high
accuracy from populations of these neurons with low individual class selectivity (Averbeck et al.
(2006), Rigotti et al. (2013), Mante et al. (2013), Raposo et al. (2014), Morcos & Harvey (2016),
Zylberberg (2017)).
Perturbation analyses have been performed for a variety of purposes. In the model pruning literature, many studies have removed units with the goal of generating smaller models with similar
performance (Li et al., 2017, Anwar et al., 2015, Molchanov et al., 2017), and recent work has explored methods for discovering maximally important directions (Raghu et al. (2017)). A variety of
studies within deep learning have highlighted single units which are selective for features or classes
(Le et al., 2011, Zeiler & Fergus, 2014, Coates et al., 2012, Zhou et al., 2014, Radford et al., 2017,
Agrawal et al., 2014). Additionally, Agrawal et al. (2014) analyzed the minimum number of sufficient feature maps (sorted by a measure of selectivity) to achieve a given accuracy. However, none of
the above studies has tested the relationship between a unit’s class selectivity or information content
and its necessity to the network’s output.
Bau et al. (2017) have quantified a related metric, concept selectivity, across layers and networks,
finding that units get more concept-selective with depth, which is consistent with our own observations regarding class selectivity (see Appendix A.3). However, they also observed a correlation
between the number of concept-selective units and performance on the action40 dataset across networks and architectures. It is difficult to compare these results directly, as the data used are substantially different as is the method of evaluating selectivity. Nevertheless, we note that Bau et al.
(2017) measured the absolute number of concept-selective units across networks with different total
numbers of units and depths. The relationship between the number of concept-selective units and
network performance may therefore arise as a result of a larger number of total units (if a fixed fraction of units is concept-selective) and increased depth (we both observed that selectivity increases
with depth).
5
D ISCUSSION AND FUTURE WORK
In this work, we have taken an empirical approach to understand what differentiates neural networks which generalize from those which do not. Our experiments demonstrate that generalization
capability is related to a network’s reliance on single directions, both in networks trained on corrupted and uncorrupted data, and over the course of training for a single network. They also show
that batch normalization, a highly successful regularizer, seems to implicitly discourage reliance on
single directions.
One clear extension of this work is to use these observations to construct a regularizer which more
directly penalizes reliance on single directions. As it happens, the most obvious candidate to regularize single direction reliance is dropout (or its variants), which, as we have shown, does not appear
to regularize for single direction reliance past the dropout fraction used in training (Section 3.3).
Interestingly, these results suggest that one is able to predict a network’s generalization performance
without inspecting a held-out validation or test set. This observation could be used in several interesting ways. First, in situations where labeled training data is sparse, testing networks’ reliance on
single directions may provide a mechanism to assess generalization performance without sacrific9
Published as a conference paper at ICLR 2018
ing training data to be used as a validation set. Second, by using computationally cheap empirical
measures of single direction reliance, such as evaluating performance at a single ablation point or
sparsely sampling the ablation curve, this metric could be used as a signal for early-stopping or
hyperparameter selection. We have shown that this metric is viable in simple datasets (Section 3.2),
but further work will be necessary to evaluate viability in more complicated datasets.
Another interesting direction for further research would be to evaluate the relationship between single direction reliance and generalization performance across different generalization regimes. In this
work, we evaluate generalization in which train and test data are drawn from the same distribution,
but a more stringent form of generalization is one in which the test set is drawn from a unique, but
overlapping distribution with the train set. The extent to which single direction reliance depends on
the overlap between the train and test distributions is also worth exploring in future research.
This work makes a potentially surprising observation about the role of individually selective units in
DNNs. We found not only that the class selectivity of single directions is largely uncorrelated with
their ultimate importance to the network’s output, but also that batch normalization decreases the
class selectivity of individual feature maps. This result suggests that highly class selective units may
actually be harmful to network performance. In addition, it implies than methods for understanding
neural networks based on analyzing highly selective single units, or finding optimal inputs for single
units, such as activation maximization (Erhan et al., 2009) may be misleading. Importantly, as we
have not measured feature selectivity, it is unclear whether these results will generalize to featureselective directions. Further work will be necessary to clarify all of these points.
ACKNOWLEDGMENTS
We would like to thank Chiyuan Zhang, Ben Poole, Sam Ritter, Avraham Ruderman, and Adam
Santoro for critical feedback and helpful discussions.
R EFERENCES
Alessandro Achille and Stefano Soatto. On the Emergence of Invariance and Disentangling in Deep
Representations. pp. 1–17, 2017. URL http://arxiv.org/abs/1706.01350.
Pulkit Agrawal, Ross B Girshick, and Jitendra Malik. Analyzing the Performance of Multilayer
Neural Networks for Object Recognition. Eccv, 2014.
Guillaume Alain and Yoshua Bengio. Understanding intermediate layers using linear classifier
probes. (2003):1–11, 2016. URL http://arxiv.org/abs/1610.01644.
Sajid Anwar, Kyuyeon Hwang, and Wonyong Sung. Structured Pruning of Deep Convolutional
Neural Networks. 2015.
Devansh Arpit, Stanisław Jastrzȩbski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S. Kanwal, Tegan Maharaj, Asja Fischer, Aaron Courville, Yoshua Bengio, and Simon
Lacoste-Julien. A Closer Look at Memorization in Deep Networks. 2017. ISSN 1938-7228.
URL http://arxiv.org/abs/1706.05394.
Bruno B Averbeck, Peter E Latham, and Alexandre Pouget. Neural correlations, population coding
and computation. Nature reviews. Neuroscience, 7(5):358–366, May 2006. ISSN 1471-003X.
doi: 10.1038/nrn1888. URL http://dx.doi.org/10.1038/nrn1888.
David G.T. Barrett, Sophie Denève, and Christian K. Machens. Optimal compensation for neuron
loss. eLife, 5(DECEMBER2016), 2016. ISSN 2050084X. doi: 10.7554/eLife.12454.
David Bau, Bolei Zhou, Aditya Khosla, Aude Oliva, and Antonio Torralba. Network Dissection:
Quantifying Interpretability of Deep Visual Representations. 2017. doi: 10.1109/CVPR.2017.
354. URL http://arxiv.org/abs/1704.05796.
Olivier Bousquet and André Elisseeff. Stability and generalization. Journal of machine learning
research: JMLR, 2(Mar):499–526, 2002. ISSN 1532-4435. URL http://www.jmlr.org/
papers/volume2/bousquet02a/bousquet02a.pdf.
10
Published as a conference paper at ICLR 2018
Kenneth H Britten, Michael N Shadlen, William T Newsome, and J Anthony Movshon. The analysis of visual motion: a comparison of neuronal and psychophysical performance. Journal of
Neuroscience, 12(12):4745–4765, 1992.
Adam Coates, Andrej Karpathy, and Andrew Y Ng. Emergence of Object-Selective Features
in Unsupervised Feature Learning. Nips’2012, pp. 1–9, 2012. ISSN 10495258. doi:
10.1002/elps.200900118. URL http://www.stanford.edu/{˜}acoates/papers/
coateskarpathyng{_}nips2012.pdf.
Russell L De Valois, E William Yund, and Norva Hepler. The orientation and direction selectivity
of cells in macaque visual cortex. Vision research, 22(5):531–544, 1982.
Laurent Dinh, Razvan Pascanu, Samy Bengio, and Yoshua Bengio. Sharp Minima Can Generalize
For Deep Nets. 2017.
Gintare Karolina Dziugaite and Daniel M. Roy. Computing Nonvacuous Generalization Bounds for
Deep (Stochastic) Neural Networks with Many More Parameters than Training Data. 2017. URL
http://arxiv.org/abs/1703.11008.
Dumitru Erhan, Yoshua Bengio, Aaron Courville, and Pascal Vincent. Visualizing Higher-Layer
Features of a Deep Network Technical Report. pp. 1–13, 2009.
David J Freedman and John a Assad. Experience-dependent representation of visual categories in
parietal cortex. Nature, 443(7107):85–8, sep 2006. ISSN 1476-4687. doi: 10.1038/nature05078.
URL http://www.ncbi.nlm.nih.gov/pubmed/16936716.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image
Recognition. Arxiv.Org, 7(3):171–180, 2015. ISSN 1664-1078. doi: 10.3389/fpsyg.2013.00124.
Sepp Hochreiter and Jürgen Schmidhuber. Flat Minima. Neural Comput, 9(1), 1997.
Sergey Ioffe and Christian Szegedy. Batch Normalization: Accelerating Deep Network Training
by Reducing Internal Covariate Shift. Arxiv, 2015. URL http://arxiv.org/abs/1502.
03167.
Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikahail Smelyanskiy, and Ping Tak Peter Tang. On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minima.
In ICLR, pp. 1–16, 2017.
Quoc V Le, Marc’Aurelio Ranzato, Rajat Monga, Matthieu Devin, Kai Chen, Greg S Corrado, Jeff
Dean, and Andrew Y Ng. Building high-level features using large scale unsupervised learning.
International Conference in Machine Learning, 2011. ISSN 10535888. doi: 10.1109/MSP.2011.
940881.
Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for
efficient ConvNets. (2016):1–10, 2017.
Valerio Mante, David Sussillo, Krishna V Shenoy, and William T Newsome. Context-dependent
computation by recurrent dynamics in prefrontal cortex. Nature, 503(7474):78–84, November
2013. ISSN 0028-0836, 1476-4687. doi: 10.1038/nature12742. URL http://dx.doi.org/
10.1038/nature12742.
Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning Convolutional
Neural Networks for Resource Efficient Inference. ICLR, (2015):1–17, 2017.
Jorrit S. Montijn, Guido T. Meijer, Carien S. Lansink, and Cyriel M A Pennartz. Population-Level
Neural Codes Are Robust to Single-Neuron Variability from a Multidimensional Coding Perspective. Cell Reports, 16(9):2486–2498, 2016. ISSN 22111247. doi: 10.1016/j.celrep.2016.07.065.
URL http://dx.doi.org/10.1016/j.celrep.2016.07.065.
Ari S Morcos and Christopher D Harvey. History-dependent variability in population dynamics
during evidence accumulation in cortex. Nature neuroscience, 19(12):nn.4403, October 2016.
ISSN 1097-6256. doi: 10.1038/nn.4403. URL http://www.nature.com/articles/
nn.4403.
11
Published as a conference paper at ICLR 2018
Behnam Neyshabur, Srinadh Bhojanapalli, David McAllester, and Nathan Srebro. Exploring Generalization in Deep Learning. 2017. URL https://arxiv.org/abs/1706.08947.
Alec Radford, Rafal Jozefowicz, and Ilya Sutskever. Learning to Generate Reviews and Discovering
Sentiment. 2017. URL http://arxiv.org/abs/1704.01444.
Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the Expressive Power of Deep Neural Networks. 2016. URL http://arxiv.org/abs/1606.
05336.
Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. SVCCA: Singular Vector
Canonical Correlation Analysis for Deep Understanding and Improvement. pp. 1–16, 2017. URL
http://arxiv.org/abs/1706.05806.
David Raposo, Matthew T Kaufman, and Anne K Churchland. A category-free neural population
supports evolving demands during decision-making. Nature neuroscience, 17(12):1784–1792,
November 2014. ISSN 1097-6256. doi: 10.1038/nn.3865. URL http://www.nature.
com/doifinder/10.1038/nn.3865.
Mattia Rigotti, Omri Barak, Melissa R Warden, Xiao-Jing Wang, Nathaniel D Daw, Earl K Miller,
and Stefano Fusi. The importance of mixed selectivity in complex cognitive tasks. Nature, 497
(7451):585–590, 2013. ISSN 0028-0836. doi: 10.1038/nature12160. URL http://dx.doi.
org/10.1038/nature12160.
Ravid Shwartz-Ziv and Naftali Tishby. Opening the Black Box of Deep Neural Networks via Information. arXiv, pp. 1–19, 2017. URL http://arxiv.org/abs/1703.00810.
Samuel L. Smith and Quoc V. Le. Understanding Generalization and Stochastic Gradient Descent.
pp. 1–11, 2017. URL http://arxiv.org/abs/1710.06451.
Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout : A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine
Learning Research (JMLR), 15:1929–1958, 2014. ISSN 15337928.
Ashia C. Wilson, Rebecca Roelofs, Mitchell Stern, Nathan Srebro, and Benjamin Recht. The
Marginal Value of Adaptive Gradient Methods in Machine Learning. pp. 1–14, 2017. URL
http://arxiv.org/abs/1705.08292.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey,
Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint
arXiv:1609.08144, 2016.
Matthew Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. Computer VisionECCV 2014, 8689:818–833, 2014. ISSN 978-3-319-10589-5. doi: 10.1007/
978-3-319-10590-1 53.
URL http://link.springer.com/chapter/10.1007/
978-3-319-10590-1{_}53.
Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding
deep learning requires rethinking generalization. 2017. URL http://arxiv.org/abs/
1611.03530.
Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Object Detectors
Emerge in Deep Scene CNNs. 2014. URL http://arxiv.org/abs/1412.6856.
Joel Zylberberg. Untuned but not irrelevant: The role of untuned neurons in sensory information
coding. September 2017. URL https://www.biorxiv.org/content/early/2017/
09/21/134379.
12
Published as a conference paper at ICLR 2018
A
A.1
A PPENDIX
C OMPARISON OF ABLATION METHODS
To remove the influence of a given direction, its value should be fixed or otherwise modified such
that it is no longer dependent on the input. However, the choice of such a fixed value can have a
substantial impact. For example, if its value were clamped to one which is highly unlikely given
its distribution of activations across the training set, network performance would likely suffer drastically. Here, we compare two methods for ablating directions: ablating to zero and ablating to
the empirical mean over the training set. Using convolutional networks trained on CIFAR-10, we
performed cumulative ablations, either ablating to zero or to the feature map’s mean (means were
calculated independently for each element of the feature map), and found that ablations to zero were
significantly less damaging than ablations to the feature map’s mean (Fig. A1). Interestingly, this
corresponds to the ablation strategies generally used in the model pruning literature (Li et al., 2017,
Anwar et al., 2015, Molchanov et al., 2017).
Figure A1: Ablation to zero vs. ablation to the empirical feature map mean.
A.2
T RAINING DETAILS
MNIST MLPs For class selectivity, generalization, early stopping, and dropout experiments, each
layer contained 128, 512, 2048 and 2048 units, respectively. All networks were trained for 640
epochs, with the exception of dropout networks which were trained for 5000 epochs.
CIFAR-10 ConvNets Convolutional networks were all trained on CIFAR-10 for 100 epochs. Layer
sizes were: 64, 64, 128, 128, 128, 256, 256, 256, 512, 512, 512, with strides of 1, 1, 2, 1, 1, 2, 1,
1, 2, 1, 1, respectively. All kernels were 3x3. For the hyperparameter sweep used in Section 3.2,
learning rate and batch size were evaluated using a grid search.
ImageNet ResNet 50-layer residual networks (He et al., 2015) were trained on ImageNet using distributed training with 32 workers and a batch size of 32 for 200,000 steps. Blocks were structured
as follows (stride, filter sizes, output channels): (1x1, 64, 64, 256) x 2, (2x2, 64, 64, 256), (1x1,
128, 128, 512) x 3, (2x2, 128, 128, 512), (1x1, 256, 256, 1024) x 5, (2x2, 256, 256, 1024), (1x1,
512, 512, 2048) x 3. For training with partially corrupted labels, we did not use any data augmen13
Published as a conference paper at ICLR 2018
a
b
Figure A2: Class selectivity increases with depth. Class selectivity distributions as a function of
depth for CIFAR-10 (a) and ImageNet (b).
a
b
Figure A3: Class selectivity is uncorrelated with L1 -norm. Relationship between class selectivity and the L1 -norm of the filter weights for CIFAR-10 (a) and ImageNet (b).
tation, as it would have dramatically increasing the effective training set size, and hence prevented
memorization.
A.3
D EPTH - DEPENDENCE OF CLASS SELECTIVITY
Here, we evaluate the distribution of class selectivity as a function of depth. In both networks trained
on CIFAR-10 (Fig. A2a) and ImageNet (Fig. A2b), selectivity increased as a function of depth. This
result is consistent with Bau et al. (2017), who show that concept-selectivity increases with depth.
It is also consistent with Alain & Bengio (2016), who show depth increases the linear decodability
of class information (though they evaluate linear decodability based on an entire layer rather than a
single unit).
A.4
R ELATIONSHIP BETWEEN CLASS SELECTIVITY AND THE FILTER WEIGHT L1 - NORM
Importantly, our results on the lack of relationship between class selectivity and importance do not
suggest that there are not directions which are more or less important to the network’s output, nor
do they suggest that these directions are not predictable; they merely suggest that class selectivity
is not a good predictor of importance. As a final test of this, we compared class selectivity to the
14
Published as a conference paper at ICLR 2018
L1 -norm of the filter weights, a metric which has been found to be a strongly correlated with the
impact of removing a filter in the model pruning literature (Li et al., 2017). Since the L1 -norm of
the filter weights is predictive of impact of a feature map’s removal, if class selectivity is also a good
predictor, the two metrics should be correlated. In the ImageNet network, we found that there was
no correlation between the L1 -norm of the filter weights and the class selectivity (Fig. A3a), while
in the CIFAR-10 network, we found there was actually a negative correlation (Fig. A3b).
A.5
R ELATIONSHIP BETWEEN MUTUAL INFORMATION AND IMPORTANCE
MNIST MLP
a
CIFAR-10 ConvNet
b
c
ImageNet ResNet
d
e
Figure A4: Mutual information is not a good predictor of unit importance. Impact of ablation
as a function of mutual information for MNIST MLP (a), CIFAR-10 convolutional network (b-c),
and ImageNet ResNet (d-e). c and e show regression lines for each layer separately.
To examine whether mutual information, which, in contrast to class selectivity, highlights units
with information about multiple classes, is a good predictor of importance, we performed the same
experiments as in Section 3.4 with mutual information in place of class selectivity. We found, that
while the results were a little less consistent (e.g., there appears to be some relationship in very
early and very late layers in CIFAR-10), mutual information was generally a poor predictor of unit
importance (Fig. A4).
15
| 2cs.AI
|
arXiv:1512.00180v1 [math.AT] 1 Dec 2015
Interactive Visualization of 2-D Persistence
Modules
Michael Lesnick∗
Matthew Wright†
Abstract
The goal of this work is to extend the standard persistent homology pipeline for
exploratory data analysis to the 2-D persistence setting, in a practical, computationally
efficient way. To this end, we introduce RIVET, a software tool for the visualization of
2-D persistence modules, and present mathematical foundations for this tool. RIVET
provides an interactive visualization of the barcodes of 1-D affine slices of a 2-D persistence module M . It also computes and visualizes the dimension of each vector space in
M and the bigraded Betti numbers of M . At the heart of our computational approach
is a novel data structure based on planar line arrangements, on which we can perform
fast queries to find the barcode of any slice of M . We present an efficient algorithm
for constructing this data structure and establish bounds on its complexity.
Contents
∗
†
1
Introduction
2
2
Algebra Preliminaries
17
3
Augmented Arrangements of 2-D Persistence Modules
24
4
Querying the Augmented Arrangement
30
5
Computing the Arrangement A(M )
34
6
Computing the Barcode Templates
41
7
Cost of Computing and Storing the Augmented Arrangement
54
8
Speeding up the Computation of the Augmented Arrangement
60
9
Preliminary Runtime Results
64
10 Conclusion
66
A Appendix
68
References
71
Notation Index
74
Columbia University, New York, NY, USA; [email protected]
St. Olaf College, Northfield, MN, USA; [email protected]
1
1
1.1
Introduction
Overview
Topological data analysis (TDA) is a relatively new branch of statistics whose goal is to
apply topology to develop tools for studying the coarse-scale, global, non-linear, geometric
features of data. Persistent homology, one of the central tools of TDA, provides invariants
of data, called barcodes, by associating to the data a filtered topological space F and
then applying standard topological and algebraic techniques to F. In the last 15 years,
persistent homology has been widely applied in the study of scientific data [9, 31], and has
been the subject of extensive theoretical work [5, 10, 17, 20, 24, 47].
For many data sets of interest, such as point cloud data with noise or non-uniformities
in density, a single filtered space is not a rich enough invariant to encode the structure
of interest in our data [11, 13, 19]. This motivates the consideration of multidimensional
persistent homology, which in its most basic form associates to the data a topological space
simultaneously equipped with two or more filtrations. Multi-D persistent homology yields
algebraic invariants of data far more complex than in the 1-D setting; new methodology is
thus required for working with these invariants in practice.
In the TDA community, it is widely appreciated that there is a need for practical data
analysis tools for handling multidimensional persistent homology. However, whereas the
community has been quick to develop fast algorithms and good publicly available software
for 1-D persistent homology, it has been comparatively slow to extend these to the multi-D
setting. Indeed, to date there is, to the best of our knowledge, no publicly available software
which extends the usual persistent homology pipeline for exploratory data analysis to the
multidimensional persistence.
This work seeks to address this gap in the case of 2-D persistence: building on ideas
presented in [13] and [14], we introduce a practical tool for working with 2-D persistent
homology in exploratory data analysis applications, and develop mathematical and algorithmic foundations for this tool. Our tool can be used in much the same way that 1-D
persistent homology is used in TDA, but offers the user significantly more information and
flexibility than standard 1-D persistent homology does.
Our tool, which we call RIVET (the Rank Invariant Visualization and Exploration
Tool), is interactive–it allows the user to dynamically navigate a two-parameter collection
of persistence barcodes derived from a 2-D persistence module. This is in contrast to
previous visualization tools for persistent homology, which have presented static displays.
Because the invariants considered in this paper are larger and more complex than the
standard 1-D persistence invariants, it is essential for the user have some nice way of
“browsing” the invariant on the computer screen. We expect that as TDA moves towards
the use of richer invariants in practical applications, interactive visualization paradigms
will play an increasingly prominent role in the TDA workflow.
It should be possible to extend our approach for 2-D persistence to 3-D, at a compu-
2
tational cost. However, there are a number of practical challenges in this, not the least
of which is designing and implementing a suitable graphical user interface. Since there is
already plenty to keep us busy in 2-D, we restrict attention to the 2-D case in this paper.
In the remainder of this section, we review multidimensional persistent homology, introduce the RIVET visualization paradigm, and provide an overview of RIVET’s mathematical and computational underpinnings. Given the length of this paper, we expect that
some readers will be content to limit their first reading to this introduction. We invite you
to begin this way.
Availability of the RIVET Software. We plan to make our RIVET software publicly
available at http://rivet.online within the next few months. In the meantime, a demo
of RIVET can be accessed through the website.
1.2
Multidimensional Filtrations and Persistence Modules
We start our introduction to RIVET by defining multidimensional filtrations and persistence modules, and reviewing the standard 1-D persistent homology pipeline for TDA.
Here and throughout, we freely use basic language from category theory. An accessible
introduction to such language in the context of persistence theory can be found in [8].
Notation. For categories, C and D, let DC denote the category whose objects are functors
C → D and whose objects are natural transformations.
For C a poset category, F : C → D a functor, and c ∈ obj(c), let Fc = F(c), and for
c ≤ d ∈ obj(c), let F(c, d) : Fc → Fd denote the image under F of the unique morphism
in homC (c, d). Let Simp denote the category of simplicial complexes and simplicial maps.
For a fixed field K, let Vect denote the category of K-vector spaces and linear maps.
Define a partial order on Rn by taking
(a1 , . . . , an ) ≤ (b1 , . . . bn )
if and only if ai ≤ bi for all i, and let Rn denote the corresponding poset category.
For S a set, we let S ∪ ∞ denote the set S ∪ {∞}.
Finite Multidimensional Filtrations. Define an n-D filtration to be a functor F :
Rn → Simp such that for all a ≤ b, F(a, b) : Fa → Fb is an inclusion. We will usually
refer to a 2-D filtration as a bifiltration. n-D filtrations are the basic topological objects of
study in multidimensional persistence.
In the computational setting, we of course work with filtrations that are specified by a
finite amount of data. Let us now introduce language and notation for such filtrations.
We say an n-D filtration stabilizes if there exists a0 ∈ Rn such that Fa = Fa0 whenever
a ≥ a0 . We write Fmax = Fa0 . We say a simplex s in Fmax appears a finite number of
times if there is a finite set A ∈ Rn such that
3
1. for each a ∈ A, s ∈ Fa , and
2. for each b ∈ Rn with s ∈ Fb , there is some a ∈ A with a ≤ b.
If s appears a finite number of times, then a minimal such A is unique; we denote it A(s),
and call it the set of grades of appearance of s.
We say an n-D filtration F is finite if
1. F stabilizes,
2. Fmax is finite, and
3. each simplex s ∈ Fmax appears a finite number of times.
For F finite, we define |F|, the size of F, by
X
|F| =
|A(s)|.
s∈Fmax
In the computational setting, we can represent a finite n-D filtration F in memory by
storing the simplicial complex Fmax , along with the set A(s) for each s ∈ Fmax .
Multidimensional Persistence Modules. Define an n-D persistence module to be a
functor M : Rn → Vect. We say M is pointwise finite dimensional (p.f.d.) if dim(Ma ) < ∞
n
for all a ∈ Rn . As we explain in Section 2.1, the category VectR of persistence modules
is isomorphic to a category of multi-graded modules over suitable rings. In view of this,
we may define presentations of n-D persistence modules; see Section 2.2 for details.
For i ≥ 0, let Hi : Simp → Vect denote the ith simplicial homology functor with
coefficients in K. For F a finite n-D filtration, there exists a finite presentation for Hi F,
which implies in particular that Hi F is p.f.d.; see Section 2.2 for details on presentations.
Barcodes of 1-D Persistence Modules. [27] shows that p.f.d. 1-D persistent homology
modules decompose in an essentially unique way into indecomposable summands, and that
the isomorphism classes of these summands are parameterized by (non-empty) intervals in
R.
Thus, we may associate to each p.f.d. 1-D persistence module M a multiset B(M ) of
intervals in R, which records the isomorphism classes of indecomposable summands of M .
We call B(M ) the barcode of M ; in general, we refer to any multiset of intervals as a
barcode.
The barcode of a finitely presented 1-D persistence module consists of a finite set of
intervals of the form [a, b), with a ∈ R, b ∈ R ∪ ∞. A barcode of this form can be
represented as a persistence diagram, i.e., a multiset of points in (a, b) ∈ R × (R ∪ ∞)
with a ≤ b. The right side of Fig. 1 depicts a barcode, together with its corresponding
persistence diagram.
4
Figure 1: A point cloud circle (left) and its 1st persistence barcode (right), obtained via the
Vietoris-Rips construction. A barcode can be visualized either by directly plotting each interval
(green bars, oriented vertically) or a via persistence diagram (green dots, right). The single long
interval in the barcode encodes the presence of a cycle in the point cloud.
1.3
Persistence Barcodes of Data
The standard 1-D persistent homology pipeline for data analysis associates a barcode Bi (P )
to a data set P , for each i ≥ 0. We regard each interval of each barcode Bi (P ) as a
topological feature of the data, and we interpret the length of the interval as a measure of
robustness of the feature. The pipeline for constructing Bi (P ) proceeds in three steps:
1. We associate to the data set P a finite 1-D filtration F(P ).
2. We apply Hi to obtain a finitely presented 1-D persistence module Hi F(P ).
3. We take Bi (P ) = B(Hi F(P )).
This pipeline for the construction of barcodes of data is quite flexible: in principle,
we can work with a data set P of any kind, and may consider any number of choices of
the filtration F(P ). Different choices of F(P ) topologically encode different aspects of the
structure of our data.
The barcodes Bi (P ) are readily computed in practice; see [47] or Section 6.2 for details.
Persistence Barcodes of Finite Metric Spaces. Here is one standard choice of P
and F(P ) in the TDA pipeline: Let P be a finite metric space, and let F(P ) = Rips(P ),
the Vietoris-Rips filtration of P , be defined as follows. For t ≥ 0, Rips(P )t is the maximal
simplicial complex with 0-skeleton P , and 1-simplices the pairs [p, q] with d(p, q) ≤ 2t; for
t < 0, Rips(P )t is the empty simplicial complex.
5
Informally, the long intervals in the barcodes Bi (P ) correspond to cycles in the data;
see Fig. 1 for an illustration.
Stability. The following well-known result shows that these barcode invariants of finite
metric spaces are robust to certain perturbations of the data:
Theorem 1.1 (Stability of Barcodes of Finite Metric Spaces [18]). For all finite metric
spaces P , Q and i ≥ 0,
dGH (P , Q) ≥ dB (Bi (P ), Bi (Q)),
where dGH denotes Gromov-Hausdorff distance and dB denotes the bottleneck distance on
barcodes; see [18] for definitions.
See also [21] for a generalization of this theorem to compact metric spaces.
1.4
Multidimensional Persistent Homology
In many cases of interest, a 1-D filtration is not sufficient to capture the structure of interest
in our data. In such cases, we are naturally led to associate to our data an n-D filtration,
for some n ≥ 2. As in the 1-D case, applying homology with field coefficients to an n-D
filtration yields an n-D persistence module. The question then arises of whether we can
associate to this n-D filtration an n-dimensional generalization of a barcode in a way that
is useful for data analysis.
In what follows, we motivate the study of multidimensional persistence by describing
one natural way that bifiltrations arise in the study of finite metric spaces; other ways that
bifiltrations arise in TDA applications are discussed, for example, in [13], [40], and [19]. We
then discuss the algebraic difficulties involved in defining a multidimensional generalization
of barcode.
Bifiltrations of Finite Metric Spaces. In spite of Theorem 1.1, which tells us that
the barcodes Bi (P ) = Bi (Hi Rips(P )) of a finite metric space P are well behaved in a
certain sense, these invariants have a couple of important limitations. First, they are
highly unstable to the addition and removal of outliers; see Fig. 2 for an illustration.
Second, and relatedly, when P exhibits non-uniformities in density, the barcodes Bi (P )
can be insensitive to interesting structure in the high density regions of P ; see Fig. 2.
To address these issues, [13] proposed that we associate a bifiltration to P , and study
the persistent homology of this bifiltration. We describe here both the proposal of [13],
which depends on a choice of bandwidth parameter, and a simple parameter-free variant
which, to the best of our knowledge, has not appeared elsewhere.
The construction of the bifiltration proposed in [13] depends on a choice of codensity
function γ : P → R, a function on P whose value is high at dense points and low at outliers
[13, 45]. For example, [13] proposes to take γ to be a K-nearest neighbors density function;
6
X
Y
B1 (X)
Z
B1 (Y )
B1 (Z)
Figure 2: Barcodes are unstable with respect to addition of outliers, and can be insensitive to
interesting structure in high density regions of the data. Thus, though the point clouds X and Y
share a densely sampled circle and differ only by the addition of a few low-density outliers, B1 (X)
and B1 (Y ) are quite different from one another: B1 (X) has a long interval not appearing in B1 (Y ).
In contrast, the point cloud Z contains no densely sampled circle, but the longest intervals of B1 (Y )
and B1 (Z) are of similar length.
in general, the choice of a density function depends on a choice of a bandwidth parameter
K.
Given γ, we may define a 2-dimensional filtration Rips(γ) by taking
Rips(γ)(a,b) = Rips(γ −1 (−∞, a])b .
If a ≥ a0 and b ≤ b0 , then Rips(γ)(a,b) ⊆ Rips(γ)(a0 ,b0 ) . Thus, the collection of all such
simplicial complexes, together with these inclusions, yields a functor
Rips(γ) : Rop × R → Simp.
Rop × R is naturally isomorphic to R2 —for example there is an isomorphism sending each
object (a, b) to (−a, b)—so upon the identification of these two categories, we may regard
Rips(γ) as a bifiltration. Note that the definition of Rips(γ) in fact makes sense for any
function γ : P → R. As discussed in [13], there are numerous possibilities for interesting
choices of γ in our study of data, aside from density estimates.
To obtain a parameter-free variant of Rips(γ), for (a, b) ∈ R2 we first define the graph
Ga,b to be the subgraph of the 1-skeleton of Rips(P )b consisting of vertices whose degree
is at least a. We then define B-Rips(P )(a,b) to be the maximal simplicial complex with
1-skeleton Ga,b . As above, upon the identification of Rop × R with R2 , the collection of
simplicial complexes {B-Rips(P )(a,b) }(a,b)∈R2 defines a bifiltration B-Rips(P ).
7
Hi B-Rips(P ) is a richer algebraic invariant than Hi Rips(P ), and in particular is sensitive to interesting structure in high density regions of P in a way that Hi Rips(P ) is
not.
Barcodes of Multi-D Persistence Modules? We now explain the algebraic difficulties
with defining the barcode of a n-D persistence module for n > 1, closely following a
discussion in [40].
As for the case n = 1, for n > 1, finitely presented n-D persistence modules also decompose in an essentially unique way into indecomposable summands; this follows easily
from a standard formulation of the Krull-Schmidt theorem [2]. However, it is a consequence of standard quiver theory results, as described for example in [30], that the set of
isomorphism classes of indecomposable multi-D persistence modules is, in contrast to the
1-D case, extremely complicated.1 In particular, for n ≥ 2 the dimension of a vector space
in a finitely presented indecomposable n-D persistence module can be arbitrarily large.
Thus, while in principle we could define the barcode of a multi-D persistence module to be
its multiset of isomorphism classes of indecomposables, as in the 1-D case, for n > 1 this
invariant will typically not be useful for data visualization and exploration in the way that
the 1-D barcode is.
In general, it seems that for the purposes of TDA, there is no entirely satisfactory way of
defining the barcode of an n-D persistence module for n > 1, even if we consider invariants
which are incomplete (i.e., invariants which can take the same value on two non-isomorphic
modules.)
Three Simple Invariants of a Multidimensional Persistence Module. Nevertheless, it is possible to define simple, useful, computable invariants of a multidimensional
persistence module. Our tool RIVET computes and visualizes three such invariants of a
2-D persistence module M :
1. The dimension function of M , i.e., the function which maps a ∈ R2 to dim Ma .
2. The fibered barcode of M , i.e., the collection of barcodes of 1-D affine slices of M .
3. The (multigraded) Betti numbers of M .
The dimension function of M is a simple, intuitive, and easily visualized invariant, but
is unstable and provides no information about persistent features in M . The next two
subsections introduce the fibered barcode and the multigraded Betti numbers.
1
For example, there is a fully faithful functor from the category of representations of the wild quiver
2
E9 to VectR , which maps indecomposables to indecomposables; see also [13] for a study of the possible
isomorphism types of a multidimensional persistence module.
8
1.5
The Rank Invariant and Fibered Barcodes
The Rank Invariant. For n ≥ 1, let Hn ⊆ Rn × Rn denote the set of pairs (s, t) with
s ≤ t, let N denote the non-negative integers, and let M be an n-D persistence module.
Following [13], we define
rank(M ) : Hn → N,
the rank invariant of M , by rank(M )(a, b) = rank M (a, b). Using the structure theorem
for 1-D persistence modules [27], it’s easy to check that for M a p.f.d. 1-D persistence
module M , rank(M ) and B(M ) determine each other [13].
For M an n-D persistence module, n ≥ 2, rank(M ) does not encode the isomorphism
class of M ; see example Example 2.2. Nevertheless, the rank invariant does capture interesting “first order information” about the structure of a persistence module.
[14] observed that if M is a p.f.d. n-D persistence module, rank(M ) carries the same
data as a (2n − 2)-parameter family of barcodes, each obtained by restricting M to an
affine line in Rn ; we call this parameterized family of barcodes the fibered barcode of M .
In particular, if M is a 2-D persistence module, the fibered barcode of M is a 2-parameter
family of barcodes. In what follows, we give the definition of a fibered barcode of a 2-D
persistence module.
The Space of Affine Lines in 2-D with Non-Negative Slope. Let L̄ denote the
collection of all (unparameterized) lines in R2 with non-negative (possibly infinite) slope,
and let L ⊂ L̄ denote the collection of all (unparameterized) lines in R2 with non-negative,
finite slope.
Note that L̄ is a submanifold (with boundary) of an affine Grassmannian of dimension 2.
With the induced topology, L̄ is homeomorphic to [0, 1] × R. In this sense, it is appropriate
to think of L̄ as a 2-parameter family of lines.
A standard point-line duality, to be described in Section 3.1, provides an identification
between L and the half-plane [0, ∞) × R; we will make extensive use of this point-line
duality later in the paper. The duality does not extend to L̄ (i.e., to vertical lines) in a
natural way.
Definition of the Fibered Barcode. For L ∈ L̄, let L denote the associated full
subcategory of R2 . The inclusion L ,→ R2 induces a functor iL : L → R2 . For M a p.f.d.
2-D persistence module, we define M L = M ◦ iL .
Define an interval I in L to be a non-empty subset of L such that whenever a < b <
c ∈ L and a, c ∈ I, we also have that b ∈ I. As L is isomorphic to R, the structure theorem
for p.f.d. 1-D persistence modules yields a definition of the barcode B(M L ) as a collection
of intervals in L.
We define B(M ), the fibered barcode of M , to be the map which sends each line L ∈ L̄
to the barcode B(M L ).
9
Proposition 1.2 ([14]). B(M ) and rank(M ) determine each other.
Proof. For each a ≤ b ∈ R2 , there exists a unique line L ∈ L̄ such that a, b ∈ L. Clearly
then, rank(M ) and the collection of 1-D rank invariants {rank(M L ) | L ∈ L̄} determine
each other. As noted above, for N a p.f.d. 1-D persistence module, B(N ) and rank(N )
determine each other. The result follows.
Stability of the Fibered Barcode. Adapting arguments introduced in [14] and [6], a
recent note by Claudia Landi [38] establishes that B(M ) is stable in two senses. The results
of [38] in fact hold for persistent homology modules of arbitrary dimension. However, in
this paper we are primarily interested in the 2-D case.
In 2-D, Landi’s first stability result says that for L a line which is neither horizontal
nor vertical, the map M → B(M L ) is Lipschitz continuous with respect to the interleaving
distance on 2-D persistence modules and the bottleneck distance on barcodes. The Lipschitz
constant c depends on the slope of L; c = 1 when the slope of L is 1, and c grows larger as
the slope of L deviates from 1, tending towards infinity as the slope of L approaches 0 or
∞. We refer to this stability result as the external stability of B(M ).
[38] also presents an internal stability result for B(M ) which tells us that when M
is finitely presented, the map L → B(M L ) is continuous, in a suitable sense, at lines L
which are neither horizontal nor vertical. In fact the result says something stronger, which
put loosely, is that the closer the slope of the line L is to 1, the more stable B(M L ) is to
perturbations of L.
In sum, the stability results of [38] tell us that for M a 2-D persistence module and L a
line which is neither horizontal or vertical, the barcode B(M L ) is robust to perturbations
both of M and L; the more diagonal L is, the more robust B(M L ) is.
1.6
Multigraded Betti Numbers
We next briefly introduce (multigraded) Betti numbers (called bigraded Betti numbers in
the case of 2-D persistence); see Section 2.4 for a precise definition and examples.
For M a finitely presented n-D persistence module and i ≥ 0, the ith (graded) Betti
number of M is a function ξi (M ) : Rn → N. It follows from the Hilbert Basis Theorem, a
classical theorem of commutative algebra, that ξi (M ) is identically 0 for i > n, so ξi (M )
is only of interest for i ≤ n.
We will be especially interested in ξ0 (M ) and ξ1 (M ) in this paper. For a ∈ Rn ,
ξ0 (M )(a) and ξ1 (M )(a) are the number of generators and relations, respectively, at index
a in a minimal presentation for M ; see Section 2.3. ξ2 (M ) has an analogous interpretation
in terms of a minimal resolution of M .
Neither of the invariants B(M ) nor {ξi (M ) | i = 0, 1, . . . , n} determines the other, but
the invariants are intimately connected. This connection in the case that n = 2 plays a
central role in the present work.
10
One of the main mathematical contributions of this project is a fast algorithm for
computing the multigraded Betti numbers of a 2-D persistence module M ; our algorithm
is described in the companion paper [41].
1.7
The RIVET Visualization Paradigm
Overview. We propose to use the fibered barcode in exploratory data analysis in much
the same way that 1-D barcodes are typically used. In particular, this requires us to have
a good way of visualizing the fibered barcode. Though discretizations of fibered barcodes
have been used in shape matching applications [15], to the best of our knowledge, there is
no prior work on visualization of fibered barcodes.
This work introduces a paradigm, called RIVET, for the interactive visualization of the
fibered barcode of a 2-D persistence module, and presents an efficient computational framework for implementing this paradigm. Our paradigm also provides for the visualization
of the dimension function and bigraded Betti numbers of the module. The visualizations
of the three invariants complement each other, and work in concert: Our visualizations of
the dimension function and bigraded Betti numbers provide a coarse, global view of the
structure of the persistence module, while our visualization of the fibered barcodes, which
focuses on a single barcode at a time, provides a sharper but more local view.
We now give a brief description of our RIVET visualization paradigm. Additional
details are in Appendix A.1.
Given a 2-D persistence module M , RIVET allows the user to interactively select a line
L ∈ R2 via a graphical interface; the software then displays the barcode B(M L ). As the
user moves the line L by clicking and dragging, the displayed barcode is updated in real
time.
The RIVET interface consists of two main windows, the Line Selection Window (left)
and the Persistence Diagram Window (right). Fig. 3 shows screenshots of RIVET for a
single choice of M and four different lines L.
The Line Selection Window. For a given finitely presented persistence module M ,
the Line Selection Window plots a rectangle in R2 containing the union of the supports of
functions ξi (M ), i ∈ {0, 1, 2}.
The greyscale shading at a point a in this rectangle represents dim Ma : a is unshaded
when dim Ma = 0, and larger dim Ma corresponds to darker shading. Scrolling the mouse
over a brings up a popup box which gives the precise value of dim Ma .
Points in the supports of ξ0 (M ), ξ1 (M ), and ξ2 (M ) are marked with green, red, and
yellow dots, respectively. The area of each dot is proportional to the corresponding function
value. The dots are translucent, so that, for example, overlaid red and green dots appear
brown on their intersection. This allows the user to read the values of the Betti numbers
at points which are in the support of more than one of the functions.
11
12
Figure 3: Screenshots of RIVET for a single choice of 2-D persistence module M and four different lines L. RIVET provides
visualizations of the dimension of each vector space in M (greyscale shading); the 0th , 1st , and 2nd bi-graded Betti numbers of
M (green, red, and yellow dots); and the barcodes of the 1-D slices M L , for each L (in purple).
The line selection window contains a blue line of non-negative slope, with endpoints
on the boundary of the displayed region of R2 . This line represents a choice of L ∈ L̄.
The intervals in the barcode B(M L ) are displayed in purple, offset from the line L in the
perpendicular direction.
The Persistence Diagram Window. In the Persistence Diagram Window (right), a
persistence diagram representation of B(M L ) is displayed. To represent B(M L ) via a
persistence diagram, we need to choose an order-preserving parameterization γL : R → L
of L, which we regard as a functor; we then display B(M L ◦ γL ). Our choice of the
parameterizations γL is described in Appendix A.1.
As with the multi-graded Betti numbers, the multiplicity of a point in the persistence
diagram is indicated by the area of the corresponding dot.
Interactivity. The user can click and drag the blue line in the left window with the
mouse, thereby changing the choice of L ∈ L̄. Clicking the blue line away from its endpoints
and dragging moves the line in the direction perpendicular to its slope, while keeping the
slope constant. The clicking and dragging an endpoint of the line moves that endpoint
while keeping the other fixed; this allows the user to change the slope of the line.
As the line moves, the both the interval representation of B(M L ) in the left window
and its persistence diagram representation in the right window are updated in real time.
1.8
Querying the Fibered Barcode
Our algorithm for fast computation of Betti numbers of 2-D persistence modules, described
in [41], performs an efficient computation of the dimension function of M as a subroutine.
Thus, in explaining the computational underpinnings of RIVET, we will focus on the
RIVET’s interactive visualization of the fibered barcode.
Because our visualization paradigm needs to update the plot of B(M L ) in real time as
we move the line L, it must be able to very quickly access B(M L ) for any choice of L ∈ L̄.
In this paper, we introduce an efficient data structure A• (M ), the augmented arrangement
of M , on which we can perform fast queries to determine B(M L ), for L ∈ L̄. We present a
theorem which guarantees that our query procedure correctly recovers B(M L ), and describe
an efficient algorithm for computing A• (M ).
Structure of the Augmented Arrangement. A• (M ) consists of a line arrangement
A(M ) in [0, ∞)×R (that is, a cell decomposition of [0, ∞)×R induced by a set of intersecting
lines), together with a collection T e of pairs (a, b) ∈ R2 × (R2 ∪ ∞) stored at each 2-cell e
of A(M ). We call T e the barcode template at e. As we explain in Section 3.3, T e is defined
in terms of the barcode of a discrete 1-D persistence module derived from M .
13
Queries of A• (M ). We now briefly describe how we query A• (M ) for the barcodes
B(M L ). Further details are given in Section 4.
As noted above, a standard point-line duality, described in Section 3.1, provides an
identification of L with [0, ∞) × R. For simplicity, let us restrict attention for now to the
generic case where L lies in a 2-cell e of A(M ); the general case is similar and is treated in
Section 4. To obtain B(M L ), for each pair (a, b) ∈ T e , we “push” the points of each pair
(a, b) ∈ T e onto the line L, by moving a and b upwards or rightwards in the plane along
horizontal and vertical lines; this gives a pair of points pushL (a), pushL (b) ∈ L; if b = ∞,
we take pushL (b) = ∞.
Our Theorem 4.1, the central mathematical result underlying the RIVET paradigm,
tells us that
B(M L ) = { [pushL (a), pushL (b)) | (a, b) ∈ T e };
see Fig. 4 for an illustration. Thus, to obtain the barcode of B(M L ) it suffices to identify
the cell e and then compute pushL (a) and pushL (b) for each (a, b) ∈ T e .
L
I2
a2
I1
b2
a1
b1
Figure 4: An illustration of how we recover the barcode B(M L ) from the barcode template T e
by “pushing” the points of each pair in the barcode template onto L. In this example, T e =
{(a1 , b1 ), (a2 , b2 )}, and B(M L ) consists of two disjoint intervals I1 , I2 .
1.9
Complexity Results for Computing, Storing, and Querying The Augmented Arrangement
Once A• (M ) has been computed, computing B(M L ) via a query to A• (M ) is far more
efficient than computing B(M L ) from scratch; for typical persistence modules arising from
data, the query of B(M L ) can be performed in real time, as desired. At the same time, the
cost of computing and storing A• (M ) is reasonable. The following two theorems provide
a theoretical basis for these claims.
14
For M a 2-D persistence module and i ∈ {0, 1, 2}, let
supp ξi (M ) = {a ∈ Rn | ξi (M )(a) > 0}.
Let κ = κx κy , for κx and κy the number of unique x and y coordinates, respectively, of
points in supp ξ0 (M ) ∪ supp ξ1 (M ); we call κ the coarseness of M .
Theorem 1.3 (computational cost of querying the augmented arrangement).
(i) For L ∈ L lying in a 2-cell of A(M ), we can query A• (M ) for B(M L ) in time
O(log κ + |B(M L )|), where |B(M L )| denotes the number of intervals in B(M L ).
0
(ii) For all other lines L ∈ L̄, we can query A• (M ) for B(M L ) in time O(log κ+|B(M L )|),
for L0 some arbitrarily small perturbation of L.
Theorem 1.4. For F a bifiltration of size m and M = Hi (F),
(i) A• (M ) is of size O(mκ2 ),
(ii) Our algorithm computes A• (M ) from F using O(m3 κ + (m + log κ)κ2 ) elementary
operations and O(m2 + mκ2 ) storage.
We prove Theorem 1.3 in Section 4.3 and Theorem 1.4 in Section 7.
To keep our exposition brief in this introduction, we have assumed in the statement
of Theorem 1.4 that M is a persistent homology module of a bifiltration. However, our
algorithm for computing augmented arrangements does handle purely algebraic input; see
Section 5.1. We give more general complexity bounds in the algebraic setting in Section 7.
Coarsening and the Interpretation of Theorem 1.4. For a fixed choice of L, B(M L )
is of size O(m), and the time required to compute B(M L ) via an application of the standard
persistence algorithm is O(m3 ). Thus Theorem 1.4 indicates that, as one would expect,
computation and storage of A• (M ) is more expensive than computation and storage of
B(M L ) for some fixed L.
For m as above, |κ| is O(m2 ) in the worst case. Thus, in the worst case, our bounds
on the size and time to compute A• (M ) grow like m5 , which on the surface may appear
problematic for practical applications. However, as we explain in Section 8.2, we can
always employ a simple coarsening procedure to approximate M by a module M 0 for which
κ is a small constant (say, κ ≤ 400). A• (M 0 ) encodes B(M 0 ) exactly, and so in view of
Landi’s external stability result [38], A• (M ) encodes B(M ) approximately. More details
on coarsening are given in Section 8.2.
15
1.10
Computation of the Augmented Arrangement
Our algorithm for computing A• (M ) decouples into three main parts:
1. Computing ξ0 (M ) and ξ1 (M ),
2. Constructing the line arrangement A(M ),
3. Computing the barcode template T e at each 2-cell e of A(M ).
We next say a few words about each of these.
Computing Bigraded Betti Numbers. As noted above, one of the main mathematical
contributions underlying RIVET is a fast algorithm for computing the bigraded Betti
numbers of a 2-D persistence module M . Not only does RIVET provide a visualization
of the Betti numbers, but it also makes essential use of the Betti numbers in constructing
A• (M ).
For M a persistent homology module of a bifiltration with n simplices, our algorithm
for computes the bigraded Betti numbers of M in O(m3 ) time; see [41] and Section 5.3.
In Section 9 of this paper, we present preliminary experimental results on the performance of our algorithm for computing bigraded Betti numbers. These indicate that the
cost of the algorithm is very reasonable in practice.
Computing the Line Arrangement. The second phase of computation constructs the
line arrangement A(M ) underlying A• (M ). Line arrangements have been the object of intense study by computational geometers for decades, and there is well-developed machinery
for constructing and working with line arrangements in practice [29]. Our algorithms for
constructing and querying A• (M ) leverage this machinery; see Section 4.3 and Section 5.
Computing the Barcode Templates. The third phase of our computation of A• (M )
computes the barcode templates T e stored at each 2-cell e ∈ A(M ). As noted above, each
T e is defined in terms of the barcode B(M e ) of a certain 1-D persistence module M e . To
compute each T e for each 2-cell e, we need to compute each B(M e ). This is the most
expensive part of the computation of A• (M ), both in theory and in practice. In Section 6,
we introduce our core algorithm for this, based on the vineyard algorithm for updating
persistent homology computations [25]. In Section 8.1, we present a modification of this
algorithm which is much faster in practice.
In fact, as explained in Section 8.3, our algorithm for computing barcode templates is
embarrassingly parallelizable.
16
Computational Experiments. Section 9 presents preliminary results on the performance of our algorithm for computing augmented arrangements. As explained there, our
present implementation of RIVET is not yet fully optimized, and our timing results should
be regarded as loose upper bounds on what can be achieved using the algorithms of this
paper.
Still, the results demonstrate that even with our current code, computation of an augmented arrangement of a bifiltration containing millions of simplices is feasible on a standard personal computer, provided we employ some modest coarsening. Thus, the current
implementation already performs well enough to be used in the analysis of modestly sized
real world data sets. With more implementation work, including the introduction of parallelization, we expect RIVET to scale well enough to be useful in many of the same settings
where 1-D persistence is currently used for exploratory data analysis.
1.11
Outline
We conclude this introduction with an outline of the remainder of the paper.
Section 2 reviews basic algebraic facts about persistence modules, their (minimal) presentations, and graded Betti numbers. We also discuss the connection between Rn -indexed
persistence modules and their Zn -indexed discretizations. Section 3 defines the augmented
arrangement A• (M ) of a 2-D persistence module M . Sections 4.1 and 4.2 give our main
result on querying A• (M ) for the barcodes B(M L ); this is Theorem 4.1. In Section 4.3,
we describe how A• (M ) is stored in memory, and apply Theorem 4.1 to give an algorithm
for querying A• (M ).
The remaining sections introduce our algorithm for computing A• (M ): First, Section 5
specifies how persistence modules are represented as input to our algorithm, and explains
our algorithm for computing A(M ). Section 6 explains our core algorithm for computing
the barcode templates T e ; this completes the specification of our algorithm for computing A• (M ), in its basic form. Section 7 analyzes the time and space complexity of our
algorithm for computing A• (M ). Section 8 describes several practical strategies to speed
up the computation of A• (M ). Section 9 presents our preliminary timing results for the
computation of A• (M ).
Appendix A.1 expands on the introduction to the RIVET interface given in Section 1.7,
providing additional details.
2
Algebra Preliminaries
In this section, we present the basic algebraic definitions and facts we will need to define
and study augmented arrangements of 2-D persistence modules.
17
2.1
A Module-Theoretic Description of Multi-D Persistence Modules
In Section 1.2 we defined a n-D persistence module to be an object of the functor category
n
n
VectR . Here we give a module-theoretic description of VectR .
The Ring Pn . Let the ring Pn be the analogue of the usual polynomial ring K[x1 , . . . , xn ]
in n variables, where exponents of the indeterminates in Pn are allowed to take on arbitrary
values√ in [0, ∞) rather than only values in N. For example, if K = R, then 1 + x2 + xπ1 +
2
2 3
5 x1 x2 is an element of P2 .
More formally, Pn can be defined as a monoid ring over the monoid ([0, ∞)n , +) [39].
For a = (a1 , . . . , an ) ∈ [0, ∞)n , we let xa denote the monomial xa11 xa22 · · · xann ∈ Pn . Let
I ⊂ Pn be the ideal generated by the set {xαi | α > 0, 1 ≤ i ≤ n}.
Rn -graded Pn -Modules. Since the field K is a subring of Pn , any Pn -module comes
naturally equipped with the structure of a K-vector space.
We define an Rn -graded Pn -module
L to be a Pn -module M with a direct sum decomposition as a K-vector space M ' a∈Rn Ma such that the action of Pn on M satisfies
xb (Ma ) ⊂ Ma+b for all a ∈ Rn , b ∈ [0, ∞)n .
The Rn -graded Pn -modules form a category whose morphisms are the module homomorphisms f : M → N such that f (Ma ) ⊂ Na for all a ∈ Rn . There is an obvious
n
isomorphism between this VectR and this category, so that we may identify the two categories. Henceforth, we will refer to Rn -graded Pn -modules as (n-D) persistence modules,
or n-modules for short.
Remark 2.1. As a rule, the familiar definitions and constructions for modules make sense
n
in the category VectR . For example, as the reader may check, we can define submodules,
quotients, direct sums, tensor products, projective/injective resolutions, and Tor functors
n
in VectR . As we next explain, we also can define free n-modules and presentations of
n-modules.
2.2
Free n-Modules and Presentations
n-graded Sets. Define an n-graded set to be a pair W = (W , grW ) for some set W and
function grW : W → Rn . Formally, we may regard W as the set of pairs
{(w, grW (w)) | w ∈ W },
and we’ll sometimes make use of this representation. We’ll often abuse notation and write
W to mean the set W . Also, when W is clear from context we’ll abbreviate grW as gr.
We say a subset Y of an n-module M is homogeneous if
[
Y⊆
Ma .
a∈Rn
Clearly, we may regard Y as an n-graded set.
18
Shifts of Modules. For M an n-module and v ∈ Rn , we define M (v) to be the n-module
such that for a ∈ Rn , M (v)a = Ma+v , and for a ≤ b ∈ Rn , M (v)(a, b) = M (a + v, b + v).
For example, when n = 2, M (1, 1) is obtained from M by shifting all vector spaces of
M down by one and to the left by one.
Free n-Modules. The usual notion of a free module
L extends to the setting of n-modules
as follows: For W an n-graded set, let Free[W] = w∈W Pn (− gr(w)). We identify W with
a set of generators in Free[W], in the obvious way. A free n-module F is an n-module such
that F ' Free[W] for some n-graded set W. Equivalently, we can define a free n-module
as an n-module which satisfies a certain universal property; see [13].
For Y a homogeneous subset of a free n-module F , let hYi denote the submodule of F
generated by Y.
Matrix Representations of Morphisms of Free Modules. Let W, W 0 be finite n0 }. For a
graded sets with ordered underlying sets W = {w1 , . . . , wl }, W 0 = {w10 , . . . , wm
morphism f : Free[W] → Free[W 0 ], we can represent f by a |W | × |W 0 | matrix [f ] with
coefficients in K: If gr(wi0 ) ≤ gr(wj ), we define [f ]ij to be the unique solution to
0
[f ]ij wi0 = xgr(wi )−gr(wj ) ρi ◦ f (wj ),
where ρi : Free[W] → Free[{wi }] is the projection. If gr(wi0 ) 6≤ gr(wj ), we define [f ]ij = 0.
Presentations of n-Modules. A presentation of an n-module M is a pair (W, Y), where
W is an n-graded set and Y ⊆ Free[W] is a homogeneous, with M ' Free[W]/hYi. We
denote the presentation (W, Y) as hW|Yi. If there exists a presentation hW|Yi for M with
W and Y finite, then we say M is finitely presented.
Note that the inclusion Y ,→ Free[W] induces a morphism Free[Y] → Free[W]; we
denote this morphism as ΦhW|Yi .
Example 2.2. Consider the 2-D persistence modules
M ' h(a, (1, 0)), (b, (0, 1)), (c, (1, 1)) | x2 a − x1 bi
N ' h(a, (1, 0)), (b, (0, 1)) | ∅i.
The induced linear maps M(1,0) ⊕ M(0,1) → M(1,1) and N(1,0) ⊕ N(0,1) → N(1,1) do not have
equal ranks. Hence M and N are not isomorphic. However, rank(M ) = rank(N ). This
shows that the rank invariant does not completely determine the isomorphism type of a
persistence module.
19
2.3
Minimal Presentations of n-Modules
Let M be an n-module. We say a presentation hW|Yi of M is minimal if hYi ⊆ I Free[W]
and ker ΦhW|Yi ⊆ I Free[Y].
The following proposition is a variant of [33, Lemma 19.4], and is proved in the same
way. It makes clear that minimal presentations are indeed minimal, in a reasonable sense.
Proposition 2.3. A finite presentation hW|Yi is minimal if and only if W descends to a
minimal set of generators for coker ΦhW|Yi and Y is a minimal set of generators for hYi.
Remark 2.4. It follows immediately from Proposition 2.3 that every finitely presented
n-D persistence module has a minimal presentation.
2.4
Graded Betti Numbers of Persistence Modules
For M an n-module, define dimf(M ) : Rn → N, the dimension function of M , by
dimf(M )(a) = dim(Ma ).
For i ≥ 0, define
ξi (M ) = dimf(Tori (M , Pn /IPn )).
The functions ξi (M ) are called the (graded) Betti numbers of M . Betti numbers of multigraded K[x1 , · · · , xn ]-modules are defined analogously; these are discussed in many places,
e.g., in [13]. In our study of augmented arrangements and fibered barcodes, we will only
need to consider ξ0 (M ) and ξ1 (M ).
We omit the straightforward proof the following result:
Proposition 2.5. If hW|Yi is a minimal presentation for M , then for all a ∈ Rn ,
ξ1 (M )(a) = | gr−1
Y (a)|.
ξ0 (M )(a) = | gr−1
W (a)|,
Example 2.6. The presentations of the modules M and N given in Example 2.2 are
minimal. Using this, it’s easy to see that
(
(
1 if a ∈ {(1, 0), (0, 1), (1, 1)}
1 if a ∈ {(1, 1)}
ξ0 (M )(a) =
ξ1 (M )(a) =
0 otherwise,
0 otherwise,
(
1 if a ∈ {(1, 0), (0, 1)}
ξ0 (N )(a) =
ξ1 (N ) = ξ2 (N ) = ξ2 (M ) = 0.
0 otherwise.
Example 2.7. For M ' h(a, (0, 0)) | x1 a, x2 ai,
(
(
1 if a ∈ {(0, 0)}
1 if a ∈ {(1, 0), (0, 1)}
ξ0 (M )(a) =
ξ1 (M )(a) =
0 otherwise,
0 otherwise,
(
1 if a ∈ {(1, 1)}
ξ2 (M )(a) =
0 otherwise.
20
Grades of Influence. For M a persistence module and a ∈ Rn let
IM (a) = {y ≤ a | ξ0 (y) > 0 or ξ1 (y) > 0}.
Lemma 2.8. For M finitely presented and a, b ∈ Rn with IM (a) = IM (b), M (a, b) is an
isomorphism.
Proof. Let hW|Yi be a minimal presentation for M . Let M 0 be the n-module with Ma0 =
Free[W]a /hYia , and M 0 (a, a0 ) the map induced by the inclusion Free[W](a, a0 ). Clearly,
M 0 isomorphic to M . Using Proposition 2.5, it’s easy to see that for a, b ∈ Rn with
IM (a) = IM (b), the map Free[W](a, b) is an isomorphism sending hYia isomorphically to
hYib . Hence M 0 (a, b) is an isomorphism. Since M and M 0 are isomorphic, M (a, b) is an
isomorphism as well.
2.5
Continuous Extensions of Discrete Persistence Modules
In the computational setting, the persistence modules we encounter are always finitely presented. It turns out that finitely presented persistence modules are, in a sense, essentially
discrete; we now explain this.
A discrete (or Zn -indexed) persistence module is a functor Zn → Vect, where Zn is
the poset category of Zn . Let An = K[x1 , . . . , xn ] denote the ordinary polynomial ring
in n variables. In analogy with the Rn -indexed case, we can regard a discrete persistence
module as a multi-graded An -module, in the obvious way.
All of the basic definitions and machinery we’ve described above for Rn -indexed persistence modules can be defined for discrete persistence modules in essentially the same way.
In particular, we may define the Betti numbers ξi (Q) : Zn → N of a discrete persistence
module Q.
Grid Functions. For n ≥ 1, we define an n-D grid to be a function G : Zn → Rn , given
by
G(z1 , . . . , zn ) = (G1 (z1 ), . . . , Gn (zn ))
for some non-decreasing functions Gi : Z → R with
lim Gi (z) = −∞
z→−∞
and
lim Gi (z) = ∞
z→∞
for each i.
We define flG (a) to be the maximum element of im(G) ordered before a in the partial
order on Rn . That is, for G a 1-D grid, we let flG : R → im(G) be given by
flG (t) = max{s ∈ im(G) | s ≤ t},
and for G an n-D grid function, we define flG : Rn → im(G) by
flG (a1 , . . . , an ) = (flG1 (a1 ), . . . , flGn (an )).
21
Continuous Extensions of Discrete Persistence Modules. For G an n-D grid, we
n
n
define a functor EG : VectZ → VectR as follows:
1. For Q a Zn -indexed persistence module and a, b ∈ Rn ,
EG (Q)a = Qy ,
EG (Q)(a, b) = Q(y, z),
where
y = max {w ∈ Zn | G(w) = flG (a)},
z = max {w ∈ Zn | G(w) = flG (b)}.
2. The action of EG on morphisms is the obvious one.
We say that an n-module M is a continuous extension of Q along G if M ∼
= EG (Q).
Proposition 2.9. Any finitely presented n-module M is a continuous extension of a finitely
generated discrete persistence module along some n-D grid.
Proof. Let G : Zn → Rn be any n-D grid such that supp ξ0 (M ) ∩ supp ξ1 (M ) ⊆ im G. We
regard G as a functor Zn → Rn in the obvious way. Using Lemma 2.8, it’s easy to check
that M is a continuous extension of M ◦ G along G. Further, M ◦ G is finitely generated; a
finite presentation for M induces one for M ◦ G of the same size.
Betti Numbers of Continuous Extensions.
Proposition 2.10. Suppose an n-module M is a continuous extension of Q along an
injective n-D grid G. Then for all i,
(
ξi (Q)(z) if a = G(z) for some z ∈ Zn ,
ξi (M )(a) =
0
otherwise.
RIVET exploits Proposition 2.10 to compute the Betti numbers of finitely presented
R2 -indexed persistence modules, by appealing to local formulae for the Betti numbers of
Z2 -indexed persistence modules; see Section 5.3.
Proof of Proposition 2.10. Let
α3
α2
α1
· · · −→ F 2 −→ F 1 −→ F 0
be a free resolution of Q. It’s easy to see that EG preserves exactness, so
EG (α3 )
EG (α2 )
EG (α1 )
· · · −−−−→ EG (F 2 ) −−−−→ EG (F 1 ) −−−−→ EG (F 0 )
22
is a free resolution for M . Write Gi = EG (F i ) and β i = EG (αi ).
By definition, ξi (M ) = dimf(Hi ), where Hi is the ith homology module of the following
chain complex:
β3 ⊗Id
β2 ⊗Id
β1 ⊗Id
· · · −−−−→ G2 ⊗ Pn /I −−−−→ G1 ⊗ Pn /I −−−−→ G0 ⊗ Pn /I.
n
n
We have two functors VectR → VectR acting respectively on objects by
N → N ⊗ Pn /I,
N → N/IN ,
with the action on morphisms defined in the obvious way; note that these are naturally
isomorphic. Thus the above chain complex is isomorphic to the chain complex
β̄ 3
β̄ 2
β̄ 1
· · · −→ G2 /IG2 −→ G1 /IG1 −→ G0 /IG0 ,
where β̄ i is the map induced by β i . Since Gi is a continuous extension along G, it is clear
that if a 6∈ im G, then (Gi /IGi )a = 0. Hence, if a 6∈ im G then ξi (M )(a) = 0, as claimed.
It remains to consider the case that a ∈ im G. Let J denote the maximal homogeneous
ideal of An . ξi (Q) = dimf(Ki ), for Ki the ith homology module of the chain complex
ᾱ3
ᾱ2
ᾱ1
· · · −→ F 2 /JF 2 −→ F 1 /JF 1 −→ F 0 /JF 0 ,
where ᾱi is the map induced by αi . If a = G(z), then by the way we have defined the functor
EG , it is clear that we have isomorphisms {ψi : Fzi → Gia }i≥0 sending JFzi isomorphically
to IGia , such that the following diagram commutes:
/ F2
···
α2z
z
ψ2
/ G2
a
···
βa2
/ F1
z
α1z
ψ1
/ G1
a
βa1
/ F0
z
ψ0
/ G0 .
a
Taking quotients, we obtain a commutative diagram:
···
/ (F 2 /JF 2 )z
···
ᾱ2z
/ (F 1 /JF 1 )z
'
/ (G2 /IG2 )a
β̄a2
ᾱ1z
'
/ (G1 /IG1 )a
β̄a1
It follows that Kzi ' Hai , so ξi (M )a = ξi (Q)z as desired.
23
/ (F 0 /JF 0 )z
'
/ (G0 /IG0 )a .
Barcodes of Discrete Persistence Modules. We discussed the barcodes of (Rindexed) 1-D persistence modules in Section 1.2. The structure theorem of [46] tells us
that the barcode B(Q) of a discrete 1-D persistence module Q is also well defined, provided
dim(Va ) < ∞ for all a less than some a0 ∈ Z. When Q is finitely generated, B(Q) is a
finite multiset of intervals [a, b) with a < b ∈ Z ∪ ∞.
Barcodes Under Continuous Extension. We omit the easy proof of the following:
Proposition 2.11. For Q a finitely generated discrete 1-D persistence module and M a
continuous extension of Q along G,
B(M ) = {[G(a), G(b)) | [a, b) ∈ B(Q), G(a) < G(b)},
where we define G(∞) = ∞.
Remark 2.12. As already noted in Section 1.5, for L the poset category corresponding to
a line L ∈ L̄, L is isomorphic to R. Hence, by adapting the definitions given above in the
R-indexed setting, we can define the grid function G : Z → L, a function flG : L → im G,
and the functor EG : VectZ → VectL . As in the R-indexed case, we say M : L → Vect
is a continuous extension of a Z-indexed persistence module Q if M ∼
= EG (Q). Clearly
then, Proposition 2.11 also holds for continuous extensions in the L-indexed setting. In
Section 4.2, we will use Proposition 2.11 in the L-indexed setting to prove our main result
on queries of augmented arrangements.
3
Augmented Arrangements of 2-D Persistence Modules
In this section we define the augmented arrangement A• (M ) associated to a finitely presented 2-D persistence module M .
First, in Section 3.1 we define the line arrangement A(M ) associated to M . Next, in
Section 3.2 we present a characterization of the 1-skeleton of A(M ). Finally, using this
characterization, in Section 3.3 we define the barcode template T e stored at a 2-cell e of
A(M ).
The augmented arrangement A• (M ) is defined to be the arrangement A(M ), together
with the additional data T e at each 2-cell e.
3.1
Definition of A(M )
Let S = supp ξ0 (M ) ∪ supp ξ1 (M ). To keep our exposition simple, we will assume that
each element of S has non-negative x-coordinate. Using the shift construction described
in Section 2.2, we can always translate the indices of M so that this assumption holds, so
there is no loss of generality in this assumption.
24
Point-Line Duality. Recall the definitions of L and L̄ from Section 1.5. As mentioned
there, a standard point-line duality gives a parameterization of L by the half-plane [0, ∞)×
R. We now explain this. Define dual transforms D` and Dp as follows:
D` : L → [0, ∞) × R
Dp : [0, ∞) × R → L
y = ax + b 7→ (a, −b)
(c, d) 7→ y = cx − d
This duality does not extend naturally to vertical lines, i.e. lines in L̄ − L.
Dp (w)
Dp (v)
w
u
D` (L)
L0
v
D` (L0 )
Dp (u)
L
Figure 5: Illustration of point-line duality
The following lemma, whose proof we omit, is illustrated in Fig. 5 by the point w and
line L.
Lemma 3.1. The transforms D` and Dp are inverses and preserve incidence, in the sense
that for w ∈ [0, ∞) × R and L ∈ L, w ∈ L if and only if D` (L) ∈ Dp (w).
Line Arrangements in [0, ∞) × R. A cell is a topological space homeomorphic to Rn
for some n ≥ 0. We define a cell complex on [0, ∞) × R to be a decomposition of [0, ∞) × R
into a finite number of cells, so that the topological boundary of each cell lies in the union
of cells of lower dimension. By standard point-set topology, each cell in a cell complex
on [0, ∞) × R has dimension at most 2. According to the definition, a cell complex on
[0, ∞) × R is not a CW-complex, as some cells will necessarily be unbounded.
By a line arrangement in [0, ∞) × R, we mean the cell complex on [0, ∞) × R induced
by a set W of lines in [0, ∞) × R. In this cell complex, the 1-skeleton consists of the union
of all lines in W , together with the line x = 0.
Definition of A(M ). For W = {(a1 , b1 ), (a2 , b2 ), . . . , (al , bl )} a finite subset of R2 , let
lub(S), the least upper bound of W be given by
lub(S) = (max ai , max bi ).
i
25
i
For example, lub((3, 5), (7, 4)) = (7, 5).
Say that a pair of distinct elements u, v ∈ S is weakly incomparable if one of the following
is true:
• u and v are incomparable with respect to the partial order on R2 ,
• u and v share either their first or second coordinate.
Call α ∈ [0, ∞) × R an anchor if α = lub(u, v) for u, v ∈ S weakly incomparable.
We define A(M ) to be the line arrangement in [0, ∞) × R induced by the set of lines
{Dp (α) | α is an anchor}.
In view of Lemma 3.1, then, the 1-skeleton A1 (M ) of A(M ) is given by
A1 (M ) = {D` (L) | L ∈ L contains an anchor} ∪ ({0} × R).
(3.1)
It is clear that A(M ) is completely determined by S.
0-cells of A(M ). Note that for α 6= β two anchors, Dp (α) and Dp (β) intersect in [0, ∞)
if and only if there exists some L ∈ L containing both α and β; such a line L exists if and
only if α and β are comparable and have distinct x-coordinate.
Size of A(M ). We now bound the number of cells in A(M ) of each dimension. As
in Section 1.9, let κ = κx κy , for κx and κy the number of unique x and y coordinates,
respectively, of points in S. Clearly, the number of anchors for S is bounded above by κ.
Hence the number of lines in A(M ) is also bounded above by κ.
Precise bounds on the number of vertices, edges, and faces in an arbitrary line arrangement are well known, and can be computed by simple counting arguments. These bounds
tell us that the number of vertices, edges, and faces in A(M ) is each not greater than κ2 .
3.2
Characterization of the 1-Skeleton of A(M )
We next give our alternate description of the 1-skeleton of A(M ). To do so, we first define
the set crit(M ) of critical lines in L◦ . Here L◦ denotes the topological interior of L, i.e.
the set of affine lines in R2 with positive, finite slope.
The Push Map. Note that for each L ∈ L̄, the partial order on R2 restricts to a total
order on L. This extends to a total order on L ∪ ∞ by taking v < ∞ for each v ∈ L. For
L ∈ L̄, define the push map pushL : R2 → L ∪ ∞ by taking
pushL (a) = min{v ∈ L | a ≤ v}.
Note that:
26
• ∞ ∈ im pushL if and only if L is horizontal or vertical.
• For a = (a1 , a2 ) ∈ R2 and pushL (u) = (v1 , v2 ) ∈ L, either a1 = v1 or a2 = v2 ; see
Fig. 6.
• For r < s ∈ R2 , pushL (r) ≤ pushL (s).
L
pushL (b)
a
pushL (a)
b
Figure 6: Illustration of the push map for lines of positive, finite slope
Continuity of Push Maps. For any a ∈ R2 , the maps {pushL }L∈L̄ induce a map
pusha : L̄ → R2 ∪ ∞, defined by pusha (L) = pushL (a).
Recall that we consider L̄ as a topological space, with the topology the restriction of
topology on the affine Grassmannian of 1-D lines in R2 .
Lemma 3.2. For each a ∈ R2 , pusha is continuous on L◦ .
Proof. Note that for any a = (a1 , a2 ) and L ∈ L◦ , pusha (L) is the unique intersection of L
with
{(a1 , y) | y ≥ a2 } ∪ {(x, a2 ) | x ≥ a1 }.
From this, the result follows readily.
Critical Lines. For L ∈ L◦ , pushL induces a totally ordered partition S L of S: elements
of the partition are restrictions of levelsets of pushL to S, and the total order on S L is the
pullback of the total order on L. This partition is illustrated in Fig. 7.
0
We call L ∈ L◦ regular if there is an open ball B ∈ L◦ containing L such that S L = S L
for all L0 ∈ B. We call L ∈ L◦ critical if it is not regular. Let crit(M ) denote the set of
critical lines in L◦ .
Theorem 3.3 (Characterization of the 1-Skeleton of A(M )). The 1-skeleton of A(M ) is
exactly
D` (crit(M )) ∪ ({0} × R) .
27
L
S5L
S2L
S1L
S3L S L
4
Figure 7: The totally ordered partition S L of S. The ith element of the partition is labeled as SiL .
u
L0
lub(u, v)
L00
lub(u, v)
u
v
v
Figure 8: An illustration of the geometric idea behind Theorem 3.3: For u, v, ∈ S and a line L0
lying just above lub(u, v), pushL0 (u) < pushL0 (v), whereas for a line L00 lying just below lub(u, v),
pushL00 (v) < pushL00 (u). Thus any line passing through lub(u, v) is critical.
Proof. In view of the description of the 1-skeleton of A(M ) given by Eq. (3.1), proving
the proposition amounts to showing that L ∈ L◦ is critical if and only if L contains some
anchor w.
Suppose L ∈ L◦ contains an anchor w, and let u, v ∈ S be weakly incomparable,
with w = lub(u, v) ∈ L. Then we must have that pushL (u) = pushL (v) = w. Further,
it’s easy to see that we can find an arbitrarily small perturbation of L so that either
pushL (u) < pushL (v) or pushL (v) < pushL (u). Thus L is critical; see Fig. 8 for an
illustration in the case that u and v are incomparable.
To prove the converse, assume that L ∈ L◦ does not contain any anchor, and consider
distinct u, v ∈ S with pushL (u) = pushL (v). Note that u, v must lie either on the same
horizontal line or the same vertical line; otherwise u and v would be incomparable and
we would have pushL (u) = pushL (v) = lub(u, v) ∈ L. Assume without loss of generality
that u < v and u, v lie on the same horizontal line H. Then we must also have that
pushL (u) = pushL (v) lies on H. Since v = lub(u, v), v is an anchor. However, L does not
contain any anchor, so we must have v < pushL (v).
Any sufficiently small perturbation L0 of L will also intersect H at a point p to the
right of v, so that we have pushL0 (u) = pushL0 (v) = p. Thus for all L0 in a neighborhood
of L, push(u) = push(v).
28
In fact, since S is finite, we can choose a single neighborhood N of L in L◦ such
that for any u, v ∈ S with pushL (u) = pushL (v), we have pushL0 (u) = pushL0 (v) for all
L0 ∈ N . By Lemma 3.2, choosing N to be smaller if necessary, may further assume that if
0
pushL (u) 6= pushL (v), then pushL0 (u) 6= pushL0 (v) for all L0 ∈ N . Thus, the partition S L
is independent of the choice of L0 ∈ N . Moreover, by Lemma 3.2 again, the total order on
0
S L is also independent of the choice of L0 ∈ N . Therefore, L is regular.
Corollary 3.4. If the duals of L, L0 ∈ L are contained in the same 2-cell in A(M ), then
0
SL = SL .
Proof. Each 2-cell is connected and open, so this follows from Theorem 3.3.
Remark 3.5. In fact, Corollary 3.4 can be strengthened to show that for any L, L0 ∈ L◦ ,
0
the duals of L and L0 lie in the same cell of A(M ) if and only if S L = S L . However, we
will not need this stronger result.
3.3
The Barcode Templates T e
Using Corollary 3.4, we now define the barcode templates T e stored at each 2-cell e of
A(M ). This will complete the definition of the augmented arrangement A• (M ).
By Corollary 3.4, we can associate to each 2-cell e in A(M ) a totally ordered partition
S e of S. Let Sie denote the ith element of the partition.
A 1-D Persistence Module at the 2-Cell e. We next use S e to define a discrete (i.e.,
Z-indexed) persistence module M e . If M ∼
6 0.
= 0, we take M e = 0. Assume then that M ∼
=
e
2
First, we define a map ζ : N \ {0} → R by
(
lub(∪zi=1 Sie ) if 1 ≤ z < |S e |,
ζ e (z) =
lub(S)
if z ≥ |S e |.
We define P e := im ζ e , and call this the set of template points at cell e. Note that the
restriction of ζ e to {1, . . . , |S e |} is an injection, so that |P e | = |S e |.
Let P denote the poset category of positive integers. ζ e (y) ≤ ζ e (z) whenever y < z,
so ζ e induces a functor P to R2 , which we also denote ζ e . Finally, the functor M ◦ ζ e :
P → Vect extends to a functor M e : Z → Vect (i.e., a Z-indexed persistence module) by
taking Mze = 0 whenever z ≤ 0.
The Definition of Barcode Templates. Clearly, M e is p.f.d., so it has a well-defined
barcode B(M e ), and it’s easy to see that B(M e ) consists of intervals [a, b), with a < b ∈
{1, . . . , |P e |, ∞}. Let us write ζ(∞) = ∞.
We define T e to be a collection of pairs of points in P e × (P e ∪ ∞), as follows:
T e = {(ζ e (a), ζ e (b)) | [a, b) ∈ B(M e )}.
29
This completes the definition of A• (M ).
Remark 3.6. A• (M ) is completely determined by the fibered barcode B(M ) and the set
S. Indeed, A(M ) is completely determined by S, and using Proposition 1.2, it’s easy to
see that T e is completely determined by B(M ) and S. The main result of the next section,
Theorem 4.1, shows that conversely, A• (M ) completely determines B(M ), and does so in
a simple way.
4
Querying the Augmented Arrangement
In the previous section, we defined the augmented arrangement A• (M ) of a finitely presented 2-D persistence module M . We now explain how A• (M ) encodes the fibered barcode
B(M ). That is, for a given L ∈ L̄, we explain how to recover B(M L ) from A• (M ). The
main result of this section, Theorem 4.1, is the basis of our algorithm for querying A• (M ).
We describe the computational details of our query algorithm in Section 4.3.
Recall that the procedure for querying A• (M ) for a barcode B(M L ) was discussed for
the case of generic lines L in Section 1.8. More generally, for any L ∈ L̄, querying A• (M )
for the barcode B(M L ) involves two steps. First, we choose a 2-cell e in A(M ); if L ∈ L,
then e is a 2-dimensional coface of the cell in containing D` (L). Second, we obtain the
intervals of B(M L ) from the pairs of T e by pushing the points in each pair (a, b) ∈ T e onto
the line L, via the map pushL of Section 3.2.
We now describe in more detail the first step of selecting the 2-D coface e.
4.1
Selecting a 2-D Coface e of L
For L ∈ L̄, we choose the 2-D coface e of L as follows:
• If L ∈ L◦ , then there exists a 2-cell of A(M ) whose closure contains D` (L); we take
e to be any such 2-cell.
• If L is horizontal, then the 2-D cofaces of the cell containing D` (L) are ordered
vertically in [0, ∞) × R; we take e to be the bottom coface. (Note that D` (L) has
only one 2-D coface unless L contains an anchor.)
• For L vertical, say L is the line x = a, let L0 be the line in the arrangement A(M )
of maximum slope, amongst those having slope less than or equal to a, if a unique
such line exists; if there are several such lines, take L0 to be the one with the largest
y-intercept. If such L0 exists, it contains a unique unbounded 1-cell in a. We take
e to be the 2-cell lying directly above this 1-cell. If such a line L0 does not exist,
then we take e to be the bottom unbounded 2-cell of A(M ); since we assume all
x-coordinates of S to be non-negative, this cell is uniquely defined.
The selection of cofaces for several lines is illustrated in Fig. 9.
30
e3
L2
L1
e1
D` (L2 ) e
2
D` (L1 )
L3
(a)
(b)
Figure 9: Three anchors are drawn as black dots in (a); the corresponding line arrangement A(M )
is shown in (b). For each line Li in (a), the dual point D` (Li ) and the corresponding 2-cell ei , chosen
as in Section 4.1, are shown in (b) in the same color.
4.2
The Query Theorem
Here is the main mathematical result underlying RIVET:
Theorem 4.1 (Querying the Augmented Arrangement). For any line L ∈ L̄ and e a 2-cell
chosen as in Section 4.1, the barcode obtained by restricting M to L is:
B(M L ) = { [pushL (a), pushL (b)) | (a, b) ∈ T e , pushL (a) < pushL (b) }.
Note that if L is such that D` (L) lies in a 2-cell of A(M ), then pushL (a) < pushL (b)
for all (a, b) ∈ T e , so the theorem statement simplifies for such L. In general, however, it
is possible to have pushL (a) = pushL (b).
Proof of Theorem 4.1. We prove the result for the case L ∈ L◦ ; the proof for L horizontal
or vertical is similar, but easier, and is left to the reader. The result holds trivially if
M∼
6 0.
= 0. Assume then that M ∼
=
Let e be a 2-D coface of the cell containing Dp (L). To keep notation simple, we will
write push = pushL and ζ = ζ e .
Keeping Remark 2.12 in mind, we define a 1-D grid G : Z → L. To do so, we first define
the restriction of G to {1, . . . , |P e |} by taking
G(z) = push ◦ ζ(z);
note that this is non-decreasing. We choose an arbitrary extension of this to a 1-D grid
G : Z → L.
By Proposition 2.11, to finish the proof it suffices to show that M L is a continuous
extension of M e along G; i.e., that there exists an isomorphism α : EG (M e ) → M L .
31
Given t ∈ L, let
z = max {w ∈ Z | G(w) = flG (t)}.
Note that we have
e
EG (M )t =
Mze
(
Mζ(z)
=
0
if z ≥ 1,
if z < 1.
Note also that
G(z) ≤ t < G(z + 1).
To define the maps αt : EG
(M e )
z < 1,
t
→
MtL ,
(4.1)
we will consider separately the three cases
1 ≤ z < |P e |,
and
z ≥ |P e |.
For a ∈ Rn , recall the definition of IM (a) from immediately above Lemma 2.8, and
note that G(1) = push ◦ ζ(1) is the minimal element a of L (with respect to the partial
order on R2 ) such that IM (a) 6= ∅. Hence by Eq. (4.1), if z < 1 then IM (t) = ∅, so
MtL = 0. Further, if z < 1, then EG (M e )t = 0. Thus for z < 1, we (necessarily) take the
isomorphism αt : EG (M e )t → MtL to be the zero map.
For 1 ≤ z < |P e |, Eq. (4.1) and the definition of G give that
push ◦ ζ(z) ≤ t < push ◦ ζ(z + 1).
This implies
ζ(z) ≤ t 6≥ ζ(z + 1),
so IM (ζ(z)) = IM (t). Thus M (ζ(z), t) is an isomorphism by Lemma 2.8. Since EG (M e )t =
Mζ(z) , we may regard the map M (ζ(z), t) as an isomorphism αt : EG (M e )t → Mt = MtL .
For z ≥ |P e |, we have that ζ(z) = lub(S), so IM (ζ(z)) = S. Further, the chain of
inequalities
ζ(z) = lub(S) ≤ push ◦ lub(S) = G(|P e |) ≤ G(z) ≤ t
gives that ζ(z) ≤ t, and further that S = IM (ζ(z)) ⊆ IM (t) ⊆ S, so in fact
IM (ζ(z)) = IM (t) = S.
Then by Lemma 2.8 again, M (ζ(z), t) is an isomorphism, which as above, can be interpreted
as an isomorphism αt : EG (M e )t → MtL .
We have now defined isomorphisms αt : EG (M e )t → MtL for all t ∈ L. Clearly, these
isomorphisms commute with internal maps in EG (M e ) and M L , so they define an isomorphism α : EG (M e ) → M L , as desired.
4.3
Computational Details of Queries
We next explain the computational details of storing and querying A• (M ). We also give
the complexity analysis of our query algorithm, proving Theorem 1.3.
32
DCEL Representation of A(M ). As noted in the introduction, we represent the line
arrangement A(M ) that underlies A• (M ) using the DCEL data structure [29], a standard
data structure for representing line arrangements in computational geometry. The DCEL
consists of a collection of vertices, edges, and 2-cells, together with a collection of pointers
specifying how the cells fit together to form a decomposition of [0, ∞) × R.
Representing the Barcode Templates. To represent the augmented arrangement
A• (M ), we store the barcode template T e at each 2-cell e in the DCEL representation
of A(M ). Recall that T e is a multiset; this means a pair (a, b) may appear in T e multiple
times. We thus store T e as a list of triples (a, b, k), where k ∈ N gives the multiplicity of
(a, b) in T e .
Our Query Algorithm. Given a line L ∈ L̄, the query of A• (M ) for B(M L ) proceeds in
two steps. The first step performs a search for the 2-cell e of A(M ) specified in Section 4.1.
Once the 2-cell e is selected, we obtain B(M L ) from T e by applying pushL to the endpoints
of each pair (a, b) ∈ T e .
Let us describe our algorithm to find the 2-cell e in detail. In the case that L ∈ L
(i.e., L is not vertical), it suffices to find the cell of A(M ) containing D` (L). In general,
the problem of finding the cell in a line arrangement containing a given query point is
known as the point location problem; this is a very well studied problem in computational
geometry. When we need to perform many point location queries on an arrangement, or
when we need to perform the queries in real time, it is standard practice to precompute a
data structure on the which point locations can be performed very efficiently [44]; this is
the approach we take. For v the number of vertices in the arrangement, there are a number
of different strategies which, in time O(v log v), compute a data structure of size O(v) on
which we can perform a point location query in time O(log v). A(M ) has O(κ2 ) vertices,
so computing such a data structure for A(M ) takes time O(κ2 log κ). The data structure
is of size O(κ2 ), and the point location query takes time O(log κ).
In the case that L is vertical, D` (L) is not defined, and we need to take a different
approach to find the 2-cell e. We precompute a separate (simpler) search data structure to
handle this case: Let Y denote the set of lines J in A(M ) such that there is no other line
in A(M ) with the same slope lying above J. We compute a 1-D array which contains, for
each J ∈ Y, a pointer to the rightmost (unbounded) 1-cell of A(M ) contained in J, sorted
according to slope. Given A(M ), computing this array takes O(nl ) time, where nl = O(κ)
is the number of anchor lines. Once the array has been computed, for any vertical line L,
we can find the appropriate 2-cell e via a binary search over the array. This takes log nl
time.
We are now ready to prove our result from the introduction on the cost of querying
•
A (M ).
Proof of Theorem 1.3. From the discussion above, it is clear that once we have precom33
puted the appropriate data structures, finding the cell e takes O(log κ) time. Each evaluation of pushL takes constant time, so computing B(M L ) from T e takes total time O(|T e |).
Thus, the total time to query A• (M ) for B(M L ) is O(|T e | + log κ). If D` (L) ∈ e, then
|T e | = |B(M L )|; this gives Theorem 1.3 (i). If, on the other hand, D` (L) 6∈ e, then we
0
may not have |T e | = |B(M L )|, but we do have |T e | = |B(M L )|, for L0 an arbitrarily small
perturbation of L with D` (L0 ) ∈ e. Theorem 1.3 (ii) follows.
Computing the Arrangement A(M )
5
We now turn to the specification of our algorithm for computing A(M ). First, in Sections 5.1 and 5.2, we specify the algebraic objects which serve as the input to our algorithm,
and explain how these objects arise from bifiltrations.
5.1
Free Implicit Representations of Persistence Modules: The Input to
our Algorithm
For k ≥ 1, define [k] to be the set of integers {1, . . . , k}, and let [0] denote the empty
set. Define a free implicit representation (FI-rep) of an n-D persistence module M to be a
quadruple Φ = (gr1 , gr2 , D1 , D2 ), such that:
• For i = 1, 2, gri : [mi ] → Rn is a function, for some mi ≥ 0.
• D1 and D2 are matrices with coefficients in K, of respective dimensions m0 × m1 and
m1 × m2 , for some m0 ≥ 0. Note that either D1 or D2 may be an empty matrix if
mi = 0.
• If m1 = 0, then M ∼
= 0.
• If m1 > 0, let gr0 : [m0 ] → Rn denote the constant function mapping to the greatest
lower bound of im gr1 . Then, defining ordered n-graded sets Wi := ([mi ], gri ) for
i = 0, 1, 2, we require that (in the notation of Section 2.2), D1 = [∂1 ] and D2 = [∂2 ]
are the matrix representations, respectively, of maps
∂1 : Free(W1 ) → Free(W0 ),
∂2 : Free(W2 ) → Free(W1 )
such that ∂1 ◦ ∂2 = 0 and M ∼
= ker ∂1 / im ∂2 .
We refer to (m0 , m1 , m2 ), defined above, as the dimensions of Φ, and write M ∼
= HΦ. Note
that a presentation of M is an FI-rep of M in the degenerate case that m0 = 0, so that D1
is an empty matrix.
Our algorithm for computing the augmented arrangement of a finitely presented 2-D
persistence module M takes as input an FI-rep of M .
34
Storing a Free Implicit Representation in Memory. We store the matrices D1
and D2 in a column-sparse data structure, as used in the standard persistent homology
algorithm [47]. Columns of Dj are stored in an array of size mj ; the ith column is stored
as a list in position i of the array. We also store grj (i) at position i of the same array.
5.2
Motivation: Free Implicit Representations From Finite Filtrations
We are interested in studying the ith persistent homology module Hi F of a finite bifiltration
F arising from data. Our choice to represent a 2-D persistence module via an FI-rep is
motivated by the fact that in practice, one typically has ready access to an FI-rep of Hi F.
In contrast, we generally do not have direct access to a presentation of Hi F at the outset,
and while there are known algorithms for computing one [36], they are computationally
expensive [12].
We’ll now describe in more detail how FI-reps of persistent homology modules arise in
practice.
The Chain Complexes of Multi-filtrations. Recall from Section 1.2 that an n-D
filtration is a functor F : Rn → Simp such that the map Fa → Fb is an inclusion
whenever a ≤ b ∈ Rn . F gives rise to a chain complex CF of n-D persistence modules,
given by
∂j+2
∂j+1
∂j
∂j−1
· · · −−−→ Cj+1 F −−−→ Cj F −→ Cj−1 F −−−→ · · · ,
where
• we define Cj F by taking the vector space (Cj F)a to be generated by the j-simplices
of Fa , and taking the map Cj F(a, b) to be induced by the inclusion Fa ,→ Fb ,
• the morphism ∂j is induced by the j th boundary maps of the simplicial complexes
Fa .
Note that Hj F ∼
= ker ∂j / im ∂j+1 .
1-Critical and Multi-critical Filtrations. To explain how FI-reps arise from finite
bifiltrations, it is helpful to first consider a special case: Following [12], we define a 1critical filtration F to be a finite n-D filtration where for each s ∈ Fmax , |A(s)| = 1; here,
as in Section 1.2, A(s) denotes the set of grades of appearance of s. If F is finite and not
1-critical, we say F is multi-critical.
Bifiltrations arising in TDA applications are often (but not always) 1-critical. For
example, for P a finite metric space and γ : P → R a function, the bifiltration Rips(γ)
of Section 1.3 is 1-critical, but the filtration B-Rips(P ) of that section generally is not
1-critical.
35
It’s easy to see that if F is 1-critical, then each Cj F is free; we have an obvious
isomorphism between Cj F and Free[Fj ], for Fj the n-graded set given by
Fj ≡ {(s, A(s)) | s a j-simplex in Fmax }.
Thus, choosing an order for each Fj , the boundary map ∂j : Cj F → Cj−1 F can be
represented with respect to Fj , Fj−1 by a matrix [∂j ] with coefficients in the field K, as
explained in Section 2.2; [∂j ] is exactly the usual matrix representation of the j th boundary
map of Fmax .
Free Implicit Representations of Hi F in the 1-Critical Case. The ordered ngraded sets Fj and matrices [∂j ] determine the chain complex CF, and hence each of the
homology modules Hj F up to isomorphism. In fact, for Oj : Fj → {1, 2, . . . , |Fj |} the
total order on Fj , we have that
−1
, [∂j ], [∂j+1 ], )
(grFj ◦ Oj−1 , grFj+1 ◦ Oj+1
is an FI-rep of Hj F, for any j ≥ 0.
Free Implicit Representations of Hi F in the Multi-Critical case. For F a multicritical filtration, the modules Cj F are not free. Nevertheless, as explained in [16], there
is an easy construction of an FI-rep
Φ = (gr1 , gr2 , D1 , D2 )
of Hj F, generalizing the construction given above in the 1-critical setting. Letting
X
lj =
|A(s)|,
s a j-simplex
in Fmax
the dimensions (m0 , m1 , m2 ) of Φ satisfy the following bounds when F is a bifiltration:
m0 ≤ lj−1 ,
m1 ≤ lj ,
m2 ≤ lj+1 + lj − 1;
(5.1)
see [16] for details.
Computation of the Free Implicit Representation of a Bifiltration. As explained
in Section 1.2, we can store a finite simplicial bifiltration in memory as a simplicial complex,
together with a list of grades of appearance for every simplex. Let F be a finite (one-critical
or multi-critical) bifiltration of size l. It’s not hard to show that given this input, for any
i ≥ 0 we can compute the FI-rep of Hi F described above in time O(l log l).
36
One Homology Index at a Time, or All Homology Indices at Once? The standard algorithms for computing persistence barcodes of a 1-D filtration F compute B(Hi F)
for all i up to a specified integer, in a single pass. When one is interested in the barcodes
at each homology index, this is more efficient than doing the computations one index at a
time, because the computations of B(Hi F) and B(Hi+1 F) share some computational work.
In contrast, the algorithm described in this paper, and implemented in the present
version of RIVET, computes A• (Hi F) of a finite bifiltration F, for a single choice of i ≥ 0.
This approach allows us to save computational effort when we are only interested in a
single homology module, and seems to be the more natural approach when working with
multi-critical filtrations.
That said, within the RIVET framework, for a 1-critical bifiltration F one can also
handle all persistence modules Hi F, for i up to a specified integer, in a single pass.2
5.3
Computation of Betti Numbers
Our first step in the computation of A(M ) is to compute S = supp ξ0 (M ) ∪ supp ξ1 (M ).
Since RIVET also visualizes the Betti numbers of M directly, we choose to compute this
by fully computing ξ0 and ξ1 . (In any case, we do not know of any algorithm for computing
S that is significantly more efficient than our algorithm for fully computing ξ0 and ξ1 .)
Computing Betti Numbers of Z2 -indexed Persistence Modules. We can define a
FI-rep Φ = (gr1 , gr2 , D1 , D2 ) of a Z2 -indexed persistence module Q just as we have for an
R2 -indexed persistence module above; in this case the functions gr1 and gr2 take values in
Z2 .
In a companion article [41], we show how to fully compute ξ0 (Q), ξ1 (Q), and ξ2 (Q),
given Φ; the algorithm runs in time runs in time O(m3 ), for (m0 , m1 , m2 ) the dimensions
of Φ and m = m0 + m1 + m2 .
One way to compute the bigraded Betti numbers of M , quite standard in computational
commutative algebra, is to compute a free resolution for M [37]. However, this gives us more
than we need for our particular application. Instead of following this route, our algorithm
computes the Betti numbers via carefully scheduled column reductions on matrices, taking
advantage of a well-known characterization of Betti numbers in terms of the homology of
Kozul complexes [34].
The natural way to do this is not to compute
A• (Hi F) for each 1 ≤ i ≤ l, but rather to compute
• Ll
the single augmented line arrangement A ( i=1 Hi F), labeling the intervals of the discrete barcode at
L
L
each 2-cell of A( li=1 Hi F) by homology degree, so that a query of A• ( li=1 Hi F) provides the homology
Ll
L
degree of each interval of B( i=1 (Hi F)L ). This “labeled” variant of A• ( li=1 Hi F) can be computed
using essentially the same algorithm as presented in this paper for computation of a single augmented
arrangement.
L
When F is not 1-critical, to compute A• ( li=1 Hi F), we need to first replace C(F) by a chain complex
of free 2-D persistence modules. As noted in [12], this can be done via a mapping telescope construction,
though this may significantly increase the size of C(F).
2
37
Computing Betti Numbers of R2 -indexed Persistence Modules. In fact, our algorithm for computing Betti numbers in the discrete setting can also be used to compute
the bigraded Betti numbers of a finitely presented R2 -indexed persistence module. Indeed,
as we will now explain, a FI-rep (gr1 , gr2 , D1 , D2 ) of an R2 -indexed persistence module M
induces a discrete FI-rep Φ = (gr01 , gr02 , D1 , D2 ) and an injective 2-D grid G : Z2 → R2 such
that M is a continuous extension of M along HΦ. (The matrices D1 , D2 are the same in
the two free implicit representations). Given this, we can compute the multigraded Betti
numbers of HΦ using the algorithm of [41] and deduce the multigraded Betti numbers of
M from those of HΦ, by Proposition 2.10.
The construction of Φ and the 2-D grid G are simple: Let ox (respectively, oy ) denote
the ordered set of unique x-coordinates (y-coordinates) of elements of im gr1 ∪ im gr2 . Let
nx = |ox | and ny = |oy |. Let Ωx : [nx ] → ox be the bijection sending i to the ith element of
ox ; define Ωy : [ny ] → oy analogously. We choose G : Z2 → R2 to be an arbitrary extension
of
Ωx × Ωy : [nx ] × [ny ] → ox × oy .
For j = 1, 2, we define gr0j to be the Z2 -valued function (Ωx × Ωy )−1 ◦ grj .
We leave to the reader the easy check that M is a continuous extension of HΦ along G.
5.4
Computation and Storage of Anchors and Template Points
Recall from Section 3.1 that an anchor is the least upper bound of a weakly incomparable
pair of points in S, and that the set of anchors determines the line arrangement A(M ).
In this section, we will let A denote the set of anchors. To compute the line arrangement
A(M ), we need to first compute a list anchors of all elements of A.
Moreover, our algorithm for computing the barcode templates, described in Section 6,
requires us to represent the set
[
P :=
Pe
e a 2-cell in A(M )
of all template points using a certain sparse matrix data structure. It’s easy to see that
P = A ∪ S.
We will see that because of this, it is convenient to compute the list anchors and the sparse
matrix representation of P at the same time.
In this section, we specify our data structure for P and describe our algorithm for
simultaneously computing both this data structure and the list anchors.
Sparse Matrix Data Structure for P. Note that S ⊆ im(Ωx × Ωy ). Given this, it’s
easy to see that also P ⊆ im(Ωx × Ωy ). Thus, to store P, it suffices to store (Ωx × Ωy )−1 (P)
and the maps Ωx and Ωy . We store Ωx and Ωy in two arrays of size nx and ny , and we
38
store (Ωx × Ωy )−1 (P) in a sparse matrix tPtsMat of size nx × ny . The triple (tPtsMat, Ωx ,
Ωy ) is our data structure for P.
Henceforth, to keep notation simple, we will assume that Ωx and Ωy are the identity
maps on [nx ] and [ny ] respectively, so that (Ωx × Ωy )−1 (P) = P.
Let us now describe tPtsMat in detail. An example tPtsMat is shown in Fig. 10. Each
element u ∈ P is represented in tPtsMat by a quintuple
(u, pl , pd , levSet1 (u), levSet2 (u)).
In this quintuple, pl and pd are pointers (possibly null); pl points to the element of P
immediately to the left of u, and pd points to the element of P immediately below u. The
objects levSet1 (u), levSet2 (u) are lists used in the computation of the barcode templates
T e . Initially, these lists are empty. We discuss them further in Section 6.
The data structure tPtsMat also contains an array rows of pointers, of length ny ; the
ith entry of rows points to the rightmost element of P with y-coordinate i.
1
2
3
4
5
6
7
8
9 10
7
6
5
4
3
2
1
rows
Figure 10: Example of tPtsMat for nx = 10 and ny = 7. Each element of P is represented
by a square. Shaded squares represent elements of S, and squares with solid borders represent
anchors. Each entry contains pointers to the next entries down and to the left; non-null pointers
are illustrated by arrows. The lists levSet1 (u) and levSet2 (u) stored at each u ∈ P are not shown.
Computation of tPtsMat and anchors. Our algorithm for computing Betti numbers,
given in [41], computes ξ0 (M )(u) and ξ1 (M )(u) at each bigrade u ∈ [nx ] × [ny ] by iterating
through [nx ] × [ny ] in lexicographical order. As we iterate through [nx ] × [ny ], it is easy to
also compute both tPtsMat and the list anchors. Let us explain this in detail.
Upon initialization, anchors is empty, tPtsMat contains no entries, and all pointers in
rows are null. We create a temporary array columns of pointers, of length nx , with each
pointer initially null.
39
At each u ∈ [nx ] × [ny ], once the Betti numbers at u have been computed, if u ∈
A, then we add u to the list anchors. If u ∈ S ∪ A = P, then we add the quintuple
(u, rows[u2 ], columns[u1 ], [ ], [ ]) to tPtsMat, and set rows[u2 ] and columns[u1 ] to both point
to u. These updates to columns and rows ensure that:
• For each i ∈ [ny ], rows[i] always points to the rightmost entry with y-coordinate i
added to tPtsMat thus far;
• For each j ∈ [nx ], columns[j] always points to the topmost entry with x-coordinate j
added to tPtsMat thus far.
It remains to explain how we determine whether u ∈ A. Note that u ∈ A if and only if
at least one of the following two conditions holds when we visit u:
1. Both rows[u2 ] and columns[u1 ] are not null.
2. u ∈ S and either rows[u2 ] or columns[u1 ] is not null.
Using this fact, we can check whether u ∈ A in constant time.
Beyond the O(m3 ) time required to compute S, this algorithm for computing tPtsMat
and anchors takes O(nx × ny ) = O(m2 ) time.
5.5
Building the Line Arrangement
Recall that the anchors of M correspond under point-line duality to the lines in the arrangement A(M ). Thus, once the list of anchors has been determined, we are ready to
build the DCEL representation of A(M ). For this, our implementation of RIVET uses the
well-known Bentley-Ottmann algorithm [29], which constructs the DCEL representation of
a line arrangement with n lines and k vertices in time O((n + k) log n). Since our arrangement contains O(κ) lines and O(κ2 ) vertices, the algorithm requires O(κ2 log κ) elementary
operations.
As explained in Section 3.1, the number of cells in A(M ) is O(κ2 ). The size of the DCEL
representation of any arrangement is of order the number of cells in the arrangement, so
the size of the DCEL representation of A(M ) is also O(κ2 ).
Remark 5.1. The κ2 log κ term in the bound of Theorem 1.4 (ii) arises from our use of the
Bentley-Ottmann algorithm. There are asymptotically faster algorithms for constructing
line arrangements that would give a slightly smaller term in the bound of Theorem 1.4 (ii)—
in fact, we can remove the log κ factor. However, the Bentley-Ottmann algorithm, which
is relatively simple and performs well in practice, is a standard choice.
Each 1-cell e in A(M ) lies on the line dual to some anchor α. In our DCEL representation of A(M ), we store a pointer at e to the entry of tPtsMat corresponding to
α.
40
Numerical Considerations. Line arrangement computations, as with many computations in computational geometry, are notoriously sensitive to numerical errors that arise
from floating-point arithmetic. Much effort has been invested in the development of smart
arithmetic models for computational geometry which allow us to avoid the errors inexact
arithmetic can produce, without giving up too much computational efficiency [7]. Because
exact arithmetic is generally far more computationally expensive than floating-point arithmetic, these models typically take a hybrid approach, relying on floating-point arithmetic
in cases where the resulting errors are certain to not cause problems, and switching over
to exact arithmetic for calculations where the errors could be problematic. Our implementation of RIVET relies on a simple such hybrid model specially tailored to the problem at
hand.
6
Computing the Barcode Templates
Once we have found all anchors and constructed the line arrangement A(M ), we are ready
to complete the computation of A• (M ) by computing the barcode templates T e for all
2-cells e of A(M ). This section describes our core algorithm for this. In Section 8.1, we
describe a refinement of the algorithm which performs significantly faster in practice.
The input to our algorithm consists of three parts:
1. A FI-rep Φ of M , represented in the way described in Section 5.1,
2. Our sparse matrix representation tPtsMat of the set P of all template points,
3. A DCEL representation of the line arrangement A(M ).
Recall that Φ is given to us as input to our algorithm for computing A• (M ); the computation of tPtsMat and A(M ) from Φ has already been described above.
Recall from Section 3.3 that
T e = {(ζ e (y), ζ e (z)) | [y, z) ∈ B(M e )}.
Thus, to compute T e at each 2-cell e, it suffices to compute the pair (B(M e ), ζ e ) at each
2-cell e. This is essentially what our algorithm does, though it turns out to be unnecessary
to explicitly store either B(M e ) or ζ e at any point in the computation.
Note that M ∼
= 0 if and only if S = ∅, if and only if each T e = ∅. Thus, we may assume
∼
that M =
6 0.
6.1
Trimming the Free Implicit Representation
Let
box(S) = {y ∈ R2 | y ≤ lub S}.
41
We say a FI-rep Φ0 = (gr01 , gr02 , D10 , D20 ) for M is trimmed if
im gr01 ∪ im gr02 ∈ box(S).
As a preliminary step in preparation for the computation of the barcode templates, if
Φ is not already trimmed we replace Φ with a smaller trimmed FI-rep. While it is possible
to work directly with an untrimmed FI-rep to compute the barcode templates, it is more
efficient to work with a trimmed one. In addition, assuming that our FI-rep is trimmed
allows for some simplifications in the description of our algorithm.
For j = 1, 2, let lj = | gr−1
j (box(S))|. There is a unique order-preserving map
ηj : [lj ] → gr−1
j (box(S)).
We define gr0j = grj ◦ ηj . We define D10 to be the submatrix of D1 whose columns correspond
to elements of im η1 , and we define D20 to be the submatrix of D2 whose rows and columns
correspond to elements of im η1 and im η2 , respectively. Let
Φ0 = (gr01 , gr02 , D10 , D20 ).
Proposition 6.1. H(Φ0 ) ∼
= M.
Proof. Associated to Φ and Φ0 we have respective chain complexes of free modules
∂
∂
∂0
∂0
2
1
F2 −→
F1 −→
F0
1
2
F00 ,
F10 −→
F20 −→
with ker ∂1 / im ∂2 = H(Φ) ∼
= M and ker ∂10 / im ∂20 = H(Φ0 ), as in the definition of a FI-rep;
further, since Φ0 is a trimming of Φ, we have obvious maps ψ j : Fj0 ,→ Fj for j = 0, 1, 2,
making the following diagram commute:
F2
∂2
ψ2
F20
F1
∂1
ψ1
∂20
F10
F0
ψ0
∂10
F00
The maps ψ i induce a map ψ : H(Φ0 ) → H(Φ) ∼
= M . It’s easy to see that ψa is an
isomorphism for a ∈ box(S). To finish the proof, it remains to check that ψb is also
an isomorphism for b 6∈ box(S). For b 6∈ box(S), there is a unique element a ∈ box(S)
minimizing the distance to b. Note that a < b. By commutativity, it suffices to see
that M (a, b) is an isomorphism and H(Φ0 )(a, b) is an isomorphism. Lemma 2.8 gives that
M (a, b) is an isomorphism, and since im gr01 ∪ im gr02 ∈ box(S), it’s easy to see directly that
H(Φ0 )(a, b) is an isomorphism.
42
Clearly, Φ0 is trimmed. Given tPtsMat and our column-sparse representation of Φ, we
can compute Φ0 from Φ in O(m + d) time, where d = O(m2 ) is the number of non-zero
entries of D2 .
Henceforth we will assume that Φ, the FI-rep of M given as input to our algorithm, is
trimmed.
6.2
RU -Decompositions and Computation of Persistence Barcodes
To prepare for a description of our algorithm for computing the barcode templates, we begin
with some preliminaries on the computation of persistence barcodes. There is a large and
growing body of work on this topic; see [42] for a recent overview with an emphasis on
publicly available software. We restrict attention here to what is needed to explain our
algorithm.
The standard algorithm for computing persistence barcodes was introduced in [47],
building on ideas in [32]; see also [31] for a succinct description of the algorithm, together
with implementation details.
The algorithm takes as input a FI-rep Φ = (gr1 , gr2 , D1 , D2 ) of a 1-D persistence module, with gr1 and gr2 non-decreasing, and returns B(HΦ). Of course, in applications Φ
typically comes from the chain complex of a 1-D filtration, with the simplices in each
dimension ordered according to their grade of appearance.
The algorithm is a variant of Gaussian elimination; it performs column additions to
construct certain factorizations of D1 and D2 , from which the barcode B(M ) can be read
off directly. Let us explain this in more detail, drawing on ideas introduced in [25]: Let
R be an m × n matrix, and for j the index of a non-zero column of R, let Λ(R, j) denote
the maximum row index of a non-zero entry in column j of R. We say R is reduced if
Λ(R, j) 6= Λ(R, j 0 ) whenever j 6= j 0 are the indices of non-zero columns in R.
The standard persistence algorithm yields a decomposition D = RU of any matrix D
with coefficients in the field K, where R is reduced, and U is an upper-triangular matrix.
For D an r × s matrix, the algorithm runs in time O(rs2 ).
We define an RU -decomposition of Φ simply to be a pair of RU -decompositions D1 =
R1 U1 and D2 = R2 U2 . We can read B(M ) off of R1 and R2 . To explain this, suppose Φ is
of dimensions (m0 , m1 , m2 ), and let Ri (∗, j) denote the j th column of Ri . Define
Pairs(Φ) := {(Λ(R2 , j), j) ∈ [m1 ] × [m2 ] | R2 (∗, j) 6= 0}
Ess(Φ) := {j ∈ [m1 ] | R1 (∗, j) = 0 and j 6= Λ(R2 , k) for any column k of R2 .}
While the RU -decomposition of a matrix is not unique, it is shown in [25] that Pairs(Φ)
and Ess(Φ) are independent of the choice of RU -decomposition of Φ.
Theorem 6.2 ([31, 47]).
B(M ) = [gr1 (j), gr2 (k)) | (j, k) ∈ Pairs(Φ) ∪ [gr1 (j), ∞) | j ∈ Ess(Φ) .
43
Vineyard Updates to Barcode Computations. Suppose that D is a d1 × d2 matrix,
and that D0 is obtained from D by transposing either two adjacent rows or two adjacent
columns of D. [25] introduces an algorithm, known as the vineyard algorithm, for updating
an RU -decomposition of D to obtain an RU -decomposition of D0 in time O(d1 + d2 ). This
algorithm is an essential subroutine in our algorithm for computing the barcode templates.
6.3
Permutations of Free Implicit Representations
As mentioned above, the standard persistence algorithm takes as input a FI-rep of a 1-D
persistence module Φ = (gr1 , gr2 , D1 , D2 ), with gr1 and gr2 non-decreasing. The reason
we need gr1 and gr2 to be non-decreasing is that the formula Theorem 6.2 for reading the
barcode off of the RU -decomposition holds only under this assumption on gr1 and gr2 .
Now suppose that we are given a 1-D FI-rep Φ = (gr1 , gr2 , D1 , D2 ) with either gr1
or gr2 not non-decreasing. How can we modify Φ to obtain a FI-rep Φ0 of H(Φ) with
non-decreasing grade functions, so that we can read the barcode of H(Φ) off of an RU decomposition of Φ0 ? We now answer this question.
It is easy to check that the following lemma holds.
Lemma 6.3. Suppose Φ = (gr1 , gr2 , D1 , D2 ) is a FI-rep of an n-D persistence module
M of dimensions (m0 , m1 , m2 ). Then for σ1 and σ2 any permutations on [m1 ] and [m2 ],
respectively, and P1 and P2 the corresponding permutation matrices, we have that
Φ0 := (gr1 ◦ σ1 , gr2 ◦ σ2 , D1 P1 , P1T D2 P2 )
is also a FI-rep of M .
Thus, in the case that M is 1-D, finding a FI-rep of M with non-decreasing grade
functions amounts to finding permutations σ1 and σ2 as above with gr1 ◦ σ1 and gr2 ◦ σ2
non-decreasing, and applying the corresponding permutations to the rows and columns of
P1 and P2 .
Finding the Permutations σj by Sorting. For j = 1, 2, we may use a sorting algorithm to find a permutation σj−1 which puts the list
grj (1), grj (2), . . . , grj (mj )
in non-decreasing order. The function gri ◦ σj is then non-decreasing. To take advantage
of the vineyard algorithm in our main algorithm, we will want to work with a sorting
algorithm which generates the permutation σj−1 as a product of transpositions of adjacent
elements in [mj ]. For this, we use the well known insertion-sort algorithm [26]. This yields
σj as a minimum length product of adjacent transpositions.
44
6.4
Induced Free Implicit Representations at Each 2-Cell
Using Lemma 6.3 above, we next show that for any 2-cell e in A(M ), Φ yields a FI-rep
Φe = (gre1 , gre2 , D1e , D2e )
of the discrete persistence module M e introduced in Section 3.3, with gre1 and gre2 nondecreasing. Thus, we can compute B(M e ) by computing an RU -decomposition of Φe .
Lift Maps.
a function
Recall the definition of the set of template points P e from Section 3.3. Define
lifte : box(S) → P e
by taking lifte (a) = u, for u the minimum element of P e such that a ≤ u (where as
0
elsewhere, < denotes the partial order order R2 ). In Fig. 11 we illustrate lifte and lifte for
a pair of adjacent 2-cells e and e0 .
u
L
u
L0
α
α
v
v
0
0
Figure 11: Illustration of lifte (left) and lifte (right) at two adjacent cells e and
S e e, containing
0
the duals of lines L and L , respectively. The black dots represent points of P = e P . Note that
0
0
0
u, α ∈ P e and v, α ∈ P e , but u 6∈ P e and v 6∈ P e . The maps lifte , lifte are illustrated by red
arrows, for a few sample points (purple dots). The shaded region in each figure is the subset of
0
box(S) on which lifte 6= lifte .
Let
orde : P e → {1, 2, . . . , |P e |}
denote the unique order-preserving bijection, i.e., the inverse of the restriction of ζ e to
{1, 2, . . . , |P e |}.
Free Implicit Representation of M e . Suppose that Φ = (gr1 , gr2 , D1 , D2 ), our FI-rep
of M , is of dimensions (m0 , m1 , m2 ). For j = 1, 2, let σje : [mj ] → [mj ] be any permutation
such that
grej := orde ◦ lifte ◦ grj ◦ σje : [mj ] → Z
45
is non-decreasing, and write σ e = (σ1e , σ2e ).
Let P1 and P2 denote the permutation matrices corresponding to σ1e and σ2e , respectively.
Let D1e = D1 P1 and D2e = P1T D2 P2 .
Proposition 6.4. Φe := (gre1 , gre2 , D1e , D2e ) is a FI-rep of M e .
Proof. It’s not hard to check that
(orde ◦ lifte ◦ gr1 , orde ◦ lifte ◦ gr2 , D1 , D2 )
is a FI-rep of M e . Given this, the result follows from Lemma 6.3.
Non-Uniqueness of Φe . In general, Φe is not uniquely defined, because it depends on
the pair of permutations σ e , which needn’t be unique. We will sometimes write Φe = Φ(σ e )
to emphasize the dependence on σ e . We say any σ e chosen as above is valid.
Reading off T e from an RU -Decomposition of Φe .
For j = 1, 2, let
fje = lifte ◦ grj ◦ σje = (orde )−1 ◦ grej ,
and write f e = (f1e , f2e ). We call f e the template map for Φe . Note that f e is independent
of the choice of a valid σ e .
From Theorem 6.2, Proposition 6.4, and the definition of the barcode template T e in
Section 3.3, we have the following relationship between Φe , f e , and T e :
T e = (f1e (j), f2e (k)) | (j, k) ∈ Pairs(Φe ) ∪ (f1e (j), ∞) | j ∈ Ess(Φe ) .
(6.1)
6.5
Our Algorithm
Eq. (6.1) tells us that for a 2-cell e in A(M ), to compute the barcode template T e it suffices
to compute the template map f e , along with the RU -decomposition of Φe = Φ(σ e ), for
some valid σ e . This is exactly what our algorithm does; we give a description of the
algorithm here, deferring some details to later sections.
We need to compute T e at every 2-cell e. A naı̈ve approach would be to do the
computation from scratch at each 2-cell e. However, we can do better, by leveraging the
work done at one cell to expedite the computation at a neighboring cell.
We proceed as follows: Let G denote the dual graph of A(M ); this is the undirected
graph with a vertex for each 2-cell e of A(M ), and an edge [e, e0 ] for each pair of adjacent
2-cells e, e0 ∈ A(M ). The dual graph is illustrated in Fig. 12. We compute a path Γ =
e0 , e1 , . . . ew in G which visits each 2-cell at least once. Our algorithm for computing Γ is
discussed below, in Section 6.6.
Let us adopt the convention of abbreviating an expression of the form (−)ei by (−)i .
For example, we write f ei as f i .
46
b
a
c
d
f
i
e
g
h
Figure 12: The line arrangement A(M ) (in grey), together with its dual graph G (in blue). The
path Γ through G might visit the vertices in the order a, b, c, d, e, f , e, g, h, i.
Once we have computed Γ, for each i = 1, . . . , w, we compute the template map f i
and an RU -decomposition of Φ(σ i ) = (gri1 , gri2 , D1i , D2i ), for some valid choice of σ i . We
proceed in order of increasing i.
For j = 1, 2, we store fji in memory by separately storing its factors lifti−1 ◦ grj and
σji−1 ; we discuss the data structures for this in Section 6.7. Thus, to compute fji , we
compute both lifti−1 ◦ grj and σji−1 .
The initial cell e0 is chosen in a way that allows for a simple combinatorial algorithm
to compute lift0 ◦ grj and σj0 ; see Section 6.8. Letting Pj denote the matrix representation
of σj0 , we have that
D10 = D1 P1 ,
D20 = P1T D2 P2 .
Thus, D10 and D20 can be obtained from D1 and D2 by performing row and column permutations. Given D10 and D20 , we compute an RU -decomposition of Φ0 via an application of
the standard persistence algorithm.
For 1 ≤ i ≤ w, we compute f i as an update of f i−1 , and we compute the RU decomposition of Φi as an update of the RU -decomposition of Φi−1 . Let us explain this in
more detail (with yet more detail to come in later sections). To update fji−1 , we update
both lifti−1 ◦ grj and σji−1 . First, we update lifti−1 ◦ grj to obtain lifti ◦ grj ; details of this
computation are given in Section 6.9. Second, we update σji−1 to obtain σji as follows.
Define
gr
˜ ij := ordi ◦ lifti ◦ grj ◦ σji−1 .
Note the distinction between gr
˜ ij and grij : the former is defined in terms of σji−1 , while the
i
latter is defined in terms of σj .
Applying the insertion-sort algorithm to gr
˜ ij , we compute a sequence of transpositions
of adjacent elements in [mj ] such that, for τj−1 : [mj ] → [mj ] the composition of these
47
transpositions, gr
˜ ij ◦ τj is non-decreasing. We take σji = σji−1 ◦ τj . Clearly, then,
grij = ordi ◦ lifti ◦ grj ◦ σji
is non-decreasing, so σ i is valid.
Note that for Pj the matrix representation of τj , we have
D1i = D1i−1 P1 ,
D2i = P1T D2i−1 P2 .
To compute an RU -decomposition of Φi (σ e ) from our RU -decomposition of Φi−1 (σ e ),
we exploit the decomposition of τj−1 as a sequence of transpositions provided by the
insertion-sort algorithm, and apply the RU -update algorithm of [25] repeatedly, performing an update of the RU -decomposition for each transposition in the sequence. We note
that neither τ , nor τ −1 , nor the decomposition of τj−1 into transpositions ever needs to
be stored explicitly in memory. Rather, we use each transposition to perform part of the
update as soon as it is computed; after this, there is no need to store the transposition.
It remains to explain how we compute Γ, lift0 ◦ grj , and σ 0 , and how we update
lifti−1 ◦ grj to obtain lifti ◦ grj . In what follows, we explain all of this, and also fill in
some details about the data structures used by our algorithm.
6.6
Computing the Path Γ
We first explain how we choose the path Γ = e0 , e1 , . . . , ew . As above, let G be the dual
graph of the line arrangement A(M ). To compute Γ, we first compute a weight for each edge
of G. The weight of [e, e0 ] is chosen to be an estimate of the amount of work our algorithm
must do to pass between cells e and e0 , in either direction. We defer the details of how
we define and compute these edge weights until Section 6.10. While the choice of edge
weights impacts our choice of Γ, and hence the speed of our algorithm for computing the
barcode templates, our asymptotic complexity bounds for our algorithm are independent
of the choice of edge weights.
We take e0 to be the topmost 2-cell in A(M ); the points of e0 correspond under pointline duality to lines that pass to the right of all points in S.
Call a path in G starting at e0 and visiting every vertex of G a valid path. We’d like to
choose Γ to be a valid path of minimum length, but we do not have an efficient algorithm
for computing such a path. (Indeed, we expect the problem is NP-hard). Instead, we
compute a path Γ whose length is approximately minimum.
Let Γ∗ be a valid path of minimum length. It is straightforward to compute a valid
path Γ such that length(Γ) ≤ 2 length(Γ∗ ): First we compute a minimum spanning tree
M for G via a standard algorithm such as Kruskal’s Algorithm [26]. Via depth-first search
of M starting at e0 , we can find a valid path Γ in M which traverses each edge of M at
most twice. Since length(M) ≤ length(Γ∗ ), we have that length(Γ) ≤ 2 length(Γ∗ ).
48
In fact, an algorithm with a better approximation ratio is known: [35] shows that a
variant of the Christofides 23 -approximation algorithm for the traveling salesman problem
on a metric graph [23] yields a valid path Γ with length(Γ) ≤ 32 length(Γ∗ ).
6.7
Data Structures
Before completing the specification of our algorithm for computing the barcode templates
T e , we need to describe the data structures used internally by the algorithm.
Persistent Homology and Vineyard Update Data Structures. First, we mention
that our algorithm uses each of the data structures specified in [25] for computing and
updating RU -decompositions. These consist consists of 4 sparse matrices—one to store
each of R1 , R2 , U1 and U2 —as well as several additional 1-D arrays which aid in performing
the persistence and vineyard algorithms, and in reading barcodes off of the matrices R1 ,
R2 . Since we use these data structures only in the way described in [25], we refer the reader
to that paper for details.
Array Data Structures. For j = 1, 2, we also maintain arrays gradesj , liftsj , sigj , and
sigInvj , each of length [mj ]:
• gradesj is a static array with gradesj [k] = grj (k).
• liftsj [k] is an array of pointers to entries of tPtsMat.
• After our computations at cell ei are complete,
liftsj [k] = lifti ◦ grj (k),
sigj [k] = σji (k),
sigInvj [k] = (σji )−1 (k).
Remark 6.5. Note that after all computations at cell ei are complete, we can use liftsj
and sigj together to perform constant time evaluations of the template map
fji = lifti ◦ grj ◦ σji .
Together with the RU -decomposition of Φi , this allows us to efficiently read off T e using
Eq. (6.1).
The Lists levSetj (u). We mentioned in Section 5.4 that for each u ∈ P, we store lists
levSet1 (u) and levSet2 (u) at the entry of tPtsMat corresponding to u. We now specify
what these lists store: After our computations at cell ei are complete, levSetj (u) stores
(lifti ◦ grj ◦ σji )−1 (u) for u ∈ P i , and levSetj (u) is empty for u 6∈ P i . As we will see
in Section 6.9, our algorithm uses the lists levSetj (u) to efficiently perform the required
updates as we pass from cell ei−1 to cell ei .
49
6.8
Computations at the Initial Cell e0
We next describe in more detail the computations performed at the initial cell e0 , building
on the explanation of Section 6.5.
To begin our computations at cell e0 , for each u ∈ P, we initialize
levSetj (u) = (lift0 ◦ grj )−1 (u).
Note that (lift0 ◦ grj )−1 (u) is nonempty only if u is the rightmost element of P on the
horizontal line passing through u. Thus, the elements u ∈ P such that levSetj (u) is nonempty have unique y-coordinates.
To efficiently initialize the lists levSetj (u), we use an O(m log m + mκ) time sweep
algorithm, described in Appendix A.2. When we add k ∈ [mj ] to the list levSetj (u), we
also set liftsj [k] = u.
Next, we concatenate the non-empty lists levSetj (u) into single list of length mj , in
increasing order of the y-coordinate of u, and set sigInvj equal to this list. For j = 1, 2,
given sigInvj , we construct sigj in the obvious way in time O(mj ). We define σj0 to be the
permutation whose array representation is sigj .
Letting Pj denote the matrix representation of σj0 , we use the arrays sigInv1 and sigInv2
to compute
D10 = D1 P1 , D20 = P1T D2 P2 .
Using the column sparse representation of D1 and D2 described in [25], which allows for
implicit representations of row permutations, this takes O(m) time.
As already explained in Section 6.5, we then apply the standard persistence algorithm
to compute the RU-decompositions D10 = R10 U10 , D20 = R20 U20 .
This completes the work done by the algorithm at cell e0 .
6.9
Computations at Cell ei , for i ≥ 1
In Section 6.5, we outlined our algorithm for updating the template map f i−1 and RU decomposition Φi−1 , as we pass from cell ei−1 to e1 . We now give a more detailed account
of this algorithm, filling in some details omitted earlier. As explained in Section 6.5, to
update
fji−1 = lifti−1 ◦ grj ◦ σ i−1 ,
our algorithm separately updates the factors lifti−1 ◦ grj and σji−1 .
While it is possible to first completely update lifti−1 ◦ grj and then update σji−1 via an
application of insertion sort, it is slightly more efficient to interleave the updates of the
two factors, so that when we update the value of lifti−1 ◦ grj (k) for some k ∈ [mj ], we
immediately perform the insertion-sort transpositions necessitated by that update, along
with the corresponding updates to the RU-decomposition. This is the approach we take.
50
We assume without loss of generality that ei−1 lies below ei . The shared boundary of
ei−1 and ei lies on the line dual to some anchor α = (α1 , α2 ). tPtsMat provides constant
time access to the element u = (u1 , u2 ) ∈ P immediately to the left of α and the element
v = (v1 , v2 ) ∈ P immediately below α, if such u and v exist. To keep our exposition simple,
we will assume that u and v do both exist; the cases that either u or v do not exist are
similar, but simpler. Note that
u, α ∈ P i−1 ,
v, α ∈ P i ,
u 6∈ P i ,
v 6∈ P i−1 ;
see Fig. 11.
Recall that for each i, lifti ◦ grj is represented in memory using the data structures liftsj
and levSetj (·), whereas σji is represented using sigj and sigInvj .
To perform the required updates as we pass from cell ei−1 to cell ei , we first iterate
through the list levSetj (α) in decreasing order. For each k ∈ levSetj (α), if the y-coordinate
of grj (k) is less than or equal to v2 , then lifti ◦ grj (k) = v; we remove k from levSetj (α),
add k to the beginning of the list levSetj (v), and set liftsj [k] = v.
If, on the other hand, the y-coordinate of grj (k) is greater than v2 , then lifti ◦ grj (k) = α,
and we do not perform any updates to liftsj , levSetj (α), or levSetj (v) for this value of k.
If in addition, liftsj [sigInvj [sigj [k] + 1]] = v, then we apply insertion sort to update sigj ,
sigInvj , and the RU -decomposition; specifically, we compute sortOneElement(sigj [k]), for
sortOneElement the algorithm defined below.
Algorithm 1 sortOneElement(w)
Input: w ∈ {1, . . . , mj − 1} such that liftsj [sigInvj [y]] ≤ liftsj [sigInvj [z]] whenever w < y < z
Output: Updated sigj and sigInvj such that liftsj [sigInvj [y]] ≤ liftsj [sigInvj [z]] whenever w ≤ y < z; correspondingly updated RU -decomposition
1: y ← w;
2: while y < mj and liftsj [sigInvj [y]] > liftsj [sigInvj [y + 1]] do
3:
swap sigInvj [y] and sigInvj [y + 1]
4:
sigj [sigInvj [y]] ← y
5:
sigj [sigInvj [y + 1]] ← y + 1
6:
Perform the corresponding updates to the RU -decomposition, as described in Section 6.5.
7:
y ←y+1
Once we have finished iterating through the list levSetj (α), we next iterate through the
list levSetj (u) in decreasing order. For each k ∈ levSetj (u), we perform updates exactly as
we did above for elements of levSetj (α), with one difference: If the second coordinate of
grj (k) is greater than v2 , then we must remove k from levSetj (u), add k to the beginning
of the list levSetj (α), and set liftsj [k] = α.
When we have finished iterating through the list levSetj (u), our updates at cell ei are
complete.
51
6.10
Choosing Edge Weights for G
We have seen in Section 6.6 that the path Γ depends on our choice of weights w(E) on
the edges E of G. We now explain how we choose and compute these weights. As we
will explain in Section 8, computing these weights is also the first step in two practical
improvements to our algorithm.
In practice, the cost of our algorithm for computing the barcode templates is dominated
by the cost of updating the RU -decompositions. On average, we expect the cost of updating
the RU -decomposition as we traverse edge E in G to be roughly proportional to the total
number of insertion-sort transpositions performed. Thus, if it were the case that the average
number t(E) of transpositions performed as we traverse E were independent of the choice
of path Γ, then it would be reasonable to take w(E) = t(E). In fact, t(E) does depend on
Γ. Nevertheless, we can give a simple, computable estimate of t(E) which is independent
of Γ. We choose w(E) to be this estimate. For E = [e1 , e2 ], our definition of w(E) in fact
depends only on the anchor line L containing the common boundary of e1 , e2 , so that we
may write w(E) = w(L).
To prepare for the definition of w(L), we introduce some terminology, which we will
also need in Section 7.3.
Switches and Separations. For e and e0 adjacent 2-cells of A(M ), we say r, s ∈ box(S)
switch at (e, e0 ) if either
lifte (r) < lifte (s)
or
0
0
lifte (r) < lifte (s)
and
and
0
0
lifte (s) < lifte (r),
lifte (s) < lifte (r).
Similarly, for r, s incomparable, we say r, s separate at (e, e0 ) if either
0
0
0
0
lifte (r) 6= lifte (s)
and
lifte (r) = lifte (s),
lifte (r) = lifte (s)
and
lifte (r) 6= lifte (s).
or
We omit the straightforward proof of the following:
Lemma 6.6. Suppose (e, e0 ) and (f , f 0 ) are pairs of adjacent 2-cells of A(M ), with the
shared boundary of each pair lying on the same anchor line L. Then
(i) r, s ∈ R2 switch at (e, e0 ) if and only if r, s switch at (f , f 0 ).
(ii) r, s separate at (e, e0 ) if and only if r, s separate at (f , f 0 ).
In view of Lemma 6.6, for L an anchor line, we say that a, b ∈ R2 switch at L if a, b
switch at e, e0 for any adjacent 2-cells e, e0 whose boundary lies on L. Analogously, we
speak of a, b ∈ R2 separating at L.
52
Lemma 6.7. If r, s switch at any anchor line L, then r, s are incomparable.
Proof. Suppose L = Dp (α) for some anchor α. If r, s switch at L, then there exist u, v as
in Fig. 11, and exchanging and r and s if necessary, we have that
r1 ≤ u1 < s1 ,
s2 ≤ v2 < r2 .
Remark 6.8. Every time we cross an anchor line L, our algorithm performs one insertionsort transposition for each pair k, l ∈ [mj ] such that grj (k), grj (l) switch at L. For each
pair k, l ∈ [mj ] such that grj (k), grj (l) separate at L, the algorithm may only perform
a corresponding insertion-sort transposition when crossing L in one direction—e.g., from
above to below—and may sometimes not perform such a transposition even when crossing
L in this direction.
It is reasonable to estimate, then, that for each pair k, l ∈ [mj ] such that grj (k), grj (l)
separate at L, the algorithm performs a corresponding transposition roughly 41 of the time.
Definition of w(L). For an anchor line L, a finite set Y , and a function f : Y → R2 ,
define swL (f ) (respectively sepL (f )) to be the number of unordered pairs a, b ∈ Y such
that f (a), f (b) switch (respectively, separate) along L. Motivated by Remark 6.8, we define
1
1
w(L) = swL (gr1 ) + swL (gr2 ) + sepL (gr1 ) + sepL (gr2 ).
4
4
Computing the weights w(L). The weights w(L) can be be computed using a simplified version of our main algorithm for computing all barcode templates: First, we choose a
path Q through the 2-cells of A(M ) starting at e0 and crossing every anchor line L once;
for example, we can choose Q to be a path through the rightmost cells of A(M ). We then
run a variant of the algorithm for computing the barcode templates described above, using
the path Q in place of P , and omitting all of the steps involving matrices and updates of
RU -decompositions. For e, e0 adjacent 2-cells in Q with shared boundary on the anchor
line L = Dp (α), we compute w(L) as we pass from cell e to cell e0 .
To explain how this works, let u, v be as in Section 6.9; for simplicity, assume that u
and v exist, as we did there. For any pair of elements r, s that switch or separate at L,
lifte (r), lifte (s) ∈ {u, α}, so to compute w(L) we only need to consider pairs whose elements
lie in the lists levSetj (u) and levSetj (α). Further, the lines x = u1 and y = v2 determine
a decomposition of the plane into four quadrants, and whether r, s switch or separate is
completely determined by which of these quadrants contain r and s; see Fig. 11. Using
these observations, we can easily extend the update procedure described in Section 6.9 to
compute the weight w(L) as we cross from e into e0 .
53
7
Cost of Computing and Storing the Augmented Arrangement
In this section, we prove Theorem 1.4, which bounds the cost of computing and storing
A• (M ). Recall that Theorem 1.4 is stated for persistence modules arising as the ith persistent homology of a bifiltration. Using language of FI-reps, we may state the result in a
more general algebraic form:
Proposition 7.1. Let M be 2-D persistence module of coarseness κ, and let Φ be a FI-rep
of M of dimensions (m0 , m1 , m2 ). Letting m = m0 + m1 + m2 , we have that
1. A• (M ) is of size O(mκ2 ),
2. Our algorithm computes A• (M ) using O(m3 κ + mκ2 log κ) elementary operations,
and requires O(m2 + mκ2 ) storage.
To see that Theorem 1.4 follows from Proposition 7.1, let F be a (1-critical or multicritical) bifiltration of size l, and recall from section Section 5.2 that, using the construction
of [16], in O(l log l) time we can compute a FI-rep of Hi (F) of dimensions (m0 , m1 , m2 ),
with m0 + m1 + m2 = O(l).
7.1
Size of the Augmented Arrangement
We prove Proposition 7.1 (i) first. As noted in Section 5.5, the DCEL representation of
A(M ) is of size O(κ2 ).
At each 2-cell of A(M ), we store the barcode template T e . By considering the RU decomposition, we see that if
Φ = (gr1 , gr2 , D1 , D2 )
is a FI-rep of a 1-D persistence module N of dimensions (m0 , m1 , m2 ), then |B(N )| ≤ m1 .
Hence, Proposition 6.4 implies that |T e | ≤ m1 for all 2-cells e. Therefore, our representation
of A• (M ) in memory is of total size O(m1 κ2 ) = O(mκ2 ), as claimed.
7.2
Cost of Computing the Augmented Arrangement
We now turn to the proof of Proposition 7.1 (ii). As we have seen, our algorithm for
computing A• (M ) involves several sub-computations. For each sub-computation, a row
in Table 1 lists the data computed by this sub-computation, a bound on the number
of elementary operation required, and the sections in this paper where the details were
discussed.
The bounds in the first four rows of Table 1 were explained earlier, so it remains to analyze the cost of our algorithm for computing the barcode templates T e . The computation
of T e itself involves a number of steps, whose individual time complexities we again list in
Table 2.
54
Table 1: Cost of augmented arrangement sub-computations
Data
# Elem. Operations
3
Details in
Set S = supp ξ0 (M ) ∪ supp ξ1 (M )
O(m )
Section 5.3
Template points P (stored in
tPtsMat) and list anchors
O(m2 )
Section 5.4
Arrangement A(M ) (constructed
via Bentley-Ottmann algorithm)
O(κ2 log κ)
Section 5.5
Data structures for point location
O(κ2 log κ)
Barcode templates T
e
3
Section 4.3
2
O(m κ + (m + log κ)κ )
Section 6
The bounds in all but the last four rows in Table 2 (above the double horizontal line)
were either explained earlier, or are clear from the discussion presented. In the remainder
of this section, we verify the last four bounds.
Cost of Updates of the levSetj (·) and liftsj at ei , i ≥ 1. In the notation of Section 6.9,
to update the lists levSetj (·) to their proper values at ei , our algorithm considers performing
an update for each element k in the lists levSetj (u) and levSetj (α). In the worst case, for
each cell ei there are O(m) such elements k to consider in total. For each k, the updates
of levSetj (·) and liftsj take constant time. Thus the total amount of work we need to do
at cell ei is O(m). The path P contains O(κ2 ) 2-cells, so the total work to perform the
updates over all 2-cells is O(mκ2 ), as claimed.
A Bound on the Total Number of Transpositions. To establish the next two
bounds, we take advantage of the following result, which we prove in Section 7.3:
Proposition 7.2. Our algorithm for computing all barcode templates performs a total of
O(m2 κ) insertion-sort transpositions on sigInv1 and sigInv2 .
Cost of Updates of sigj , sigInvj at cells ei , i ≥ 1. The cost of updating sig is proportional to the cost of updating sigInvj . Thus it is immediate from Proposition 7.2 that the
total cost of updating these arrays is O(m2 κ).
Cost of Updates of RU -decomposition at cells ei , i ≥ 1. There are O(κ2 ) cells to
consider; this gives us the O(κ2 ) term. For each transposition performed on sigj , we call
the vineyard algorithm, described in Section 6.2, at most twice. Each call to the vineyard
algorithm takes time O(m). By Proposition 7.2, then, the total cost of all vineyard updates
performed is O(m3 κ). This gives the desired bound.
55
Table 2: Cost of barcode template sub-computations
Data
# Elem. Operations
Details in
Trimming the FI-rep
2
O(m )
Section 6.1
Path Γ (found via the 2-approximation algorithm for the optimal path, using Kruskal’s
MST algorithm)
O(κ2 log κ)
Section 6.6
levSetj (u), liftsj , sigj , sigInvj at cell e0
O(m log m + mκ)
Section 6.8, Appendix A.2
0
3
RU-decomposition of Φ
O(m )
Section 6.2
Reading the barcode template T off of the
RU -decomposition of Φi , for all i ≥ 0
O(mκ2 )
Section 6.4, Section 6.7
levSetj (u) and liftsj at all cells ei , i ≥ 1
O(mκ2 )
Section 6.9
e
2
sigj , sigInvj at all ei , i ≥ 1
O(m κ)
Section 6.9
RU-decompositions at all ei , i ≥ 1
O(m3 κ + κ2 )
Section 6.9
Weights w(L) for all anchor lines L
O(m2 + mκ)
Section 6.10
Cost of Computing Weights w(L). As explained in Section 6.10, we compute the edge
weights w(L) using a stripped-down variant of our algorithm for computing the barcode
templates T e . Using Lemma 7.5 below, it can be checked that computing all edge weights
takes time O(m2 + mκ).
Storage Requirements. By Proposition 7.1 (i), A• (M ) itself is of size O(mκ2 ), so our
algorithm for computing A• (M ) requires at least this much storage. Our algorithm for
computing the Betti numbers requires O(m2 ) storage, as do the persistence algorithm and
the vineyard updates to RU -decompositions. The Bentley-Ottmann algorithm requires
O(κ2 ) storage, as does Kruskal’s algorithm. Constructing the search data structures used
for queries of A• (M ) also requires O(κ2 ) storage. From our descriptions of the data structures used in the remaining parts of our algorithm, it is clear that other steps of our
algorithm for computing A• (M ) do not require more than O(m2 + mκ2 ) storage. The
bound of Proposition 7.1 (ii) on the storage requirements of our algorithm follows.
56
7.3
Bounding Total Number of Transpositions Required to Compute all
Barcode Templates
To complete our proof of Proposition 7.1 (ii), it remains to prove Proposition 7.2. To this
end, for r ∈ box(S), let
P ≥r := {u ∈ P | u ≥ r},
lift(r) := {u ∈ P | u = lifte (r) for some 2-cell e}.
We leave to the reader the proofs of the following two lemmas.
Lemma 7.3. For e a 2-cell, L a line with D` (L) ∈ e, and u ∈ P ≥r , lifte (r) = u if and
only if the following two conditions hold:
1. pushL (u) is the minimum element of pushL (P ≥r ),
2. for all u0 ∈ P ≥r , u0 6= u, with pushL (u0 ) = pushL (u), we have u0 < u.
Lemma 7.4. For u ∈ P ≥r , we have that u ∈ lift(r) if and only if the following two
conditions hold:
1. there exists no w ∈ P ≥r with
w1 < u1
and
w2 < u2 .
2. there exist no pair v, w ∈ P ≥r with
v1 < u1 ,
v2 = u2 ,
w1 = u1 ,
and
w2 < w2 .
Fig. 13 illustrates the shape of the set lift(r), as described by Lemma 7.4.
r
Figure 13: The shape of a set lift(r), as described by Lemma 7.4.
The next lemma shows that the number of anchor lines at which a given pair of points
in box(S) can switch or separate is at most two; it is the key step in our proof of Proposition 7.2.
57
Lemma 7.5. For r, s ∈ box(S) incomparable,
(i) there is at most one anchor line L at which r and s switch, and if such L exists, then
there is no anchor line at which r and s separate.
(ii) there are at most two anchor lines at which r and s separate.
Proof. Assume without loss of generality that r1 < s1 . Then, since r and s are incomparable, r2 > s2 . Let
R = {(x, y) ∈ lift(r) | x < s1 },
S = {(x, y) ∈ lift(s) | y < r2 },
Q = {(x, y) = lifte (r) = lifte (s) for some e ∈ A(M )}.
The following observations, illustrated in Fig. 14, follow from Lemma 7.4:
• If R is nonempty, there is an element r0 ∈ R such that for all u ∈ R, u1 ≤ r10 and
u2 ≥ r20 .
• Symmetrically, if S is nonempty, there is an element s0 ∈ R such that for all u ∈ S,
u1 ≥ s01 and u2 ≤ s02 .
• If Q is nonempty, there is an element q x ∈ Q such that for all u ∈ Q, u1 ≥ q1x and
u2 ≤ q2x .
• Symmetrically, if Q is nonempty, there is an element q y ∈ Q such that for all u ∈ Q,
u1 ≤ q1y and u2 ≥ r2 .
Clearly, r0 , s0 , q x , and q y are unique, when they exist. Using Lemma 7.3, it is straightforward
to check that u ∈ Q if and only if u ∈ lift(r) ∩ lift(s) and one of the following is true:
1. u is incomparable to every element of R ∪ S.
2. R is non-empty, u2 = r20 , and there is no v ∈ lift(s) with v1 = u1 and v2 < u2 ,
3. S is non-empty, u1 = s01 , and there is no v ∈ lift(r) with v1 < u1 and v2 = u2 .
See Fig. 14 for an illustration of R, S, and Q.
To finish the proof of Lemma 7.5, we consider seven cases. For each, we explicitly
describe the lines where r and s either switch or separate. The verification of the claimed
behavior in each case, which uses Lemma 7.3 and the observations above, is left to the
reader. Fig. 14 illustrates case 7.
1. R and S empty. lifte (r) = lifte (s) for every 2-cell e ∈ A(M ). Therefore, r and s
never switch or separate.
58
2. R nonempty, S and Q empty. lifte (r) < lifte (s) for every 2-cell e ∈ A(M ). Again, no
switches or separations.
3. S nonempty, R and Q empty. Symmetric to the above, no switches or separations.
4. R and S nonempty, Q empty. lifte (r) < lifte (s) whenever e lies below L =
Dp (lub(r0 , s0 )), and lifte (s) < lifte (r) whenever e lies above L. Hence r and s switch
at L.
5. R and Q nonempty, S empty. lifte (r) < lifte (s) whenever e lies below L =
Dp (lub(r0 , qx )), and lifte (r) = lifte (s) whenever e lies above L. Hence r and s separate at L.
6. S and Q nonempty, R empty.
Dp (lub(qy , s0 )).
Symmetric to the above, r and s separate at
7. R, S, and Q all nonempty. lifte (r) < lifte (s) whenever e lies below L =
Dp (lub(r0 , qx )); lifte (r) = lifte (s) whenever e lies above L and below L0 =
Dp (lub(qy , s0 )); and lifte (s) < lifte (r) whenever e lies above L0 . Hence r and s
separate at L and L0 .
R
r0
qx
Q
qy
r
s0
S
s
Figure 14: An illustration of the case where R, S, and Q are each non-empty (case 7). Points
in lift(r) ∪ lift(s) are drawn as black dots. Observe that r and s separate at Dp (lub(r0 , qx )) and
Dp (lub(qy , s0 )), and at no other anchor line do r and s either switch or separate.
Proof of Proposition 7.2. Fix j ∈ {1, 2}, and let k 6= k 0 ∈ [mj ]. First, we note that if grj (k)
and grj (k 0 ) are comparable, then as we pass from cell ei−1 to cell ei , our algorithm for
computing barcode templates never performs an insertion-sort transposition of the values
59
k and k 0 in sigInvj . This is because our initialization procedure at cell e0 , described in
Section 6.8 and Appendix A.2, chooses sigInvj such that if grj (k) < grj (k 0 ), then sigj [k] <
sigj [k 0 ]. Since lifte (k) ≤ lifte (k 0 ) for all 2-cells e, there thus is never any need to swap k
and k 0 .
Therefore, as we pass from cell ei−1 to cell ei , our algorithm performs an insertion-sort
transposition of the values k and k 0 in sigInvj only if grj (k) and grj (k 0 ) either switch or
separate at (ei−1 , ei ).
Clearly, the number of pairs k, k 0 ∈ [mj ] such that grj (k) and grj (k 0 ) either switch
or separate at any anchor line is less than m2j /2, the total number of pairs. The path Γ,
constructed via the minimum spanning tree construction in Section 6.6, crosses each anchor
line at most 2κ times. By Lemma 7.5 then, for each pair k, k 0 ∈ [mj ] , the insertion-sort
component of our algorithm performs a total of at most 4κ transpositions of that pair in
sigInvj . Hence, the total number of transpositions performed by the algorithm altogether
is at most 2(m21 + m22 )κ < 2m2 κ.
8
Speeding up the Computation of the Augmented Arrangement
In this section, we describe several simple, practical strategies to speed up the runtime
of our computation of A• (M ). Used together, these strategies allow us to the compute
augmented arrangements of the persistent homology modules of much larger datasets than
would otherwise be possible.
8.1
Persistence Computation from Scratch When RU -Updates are too
Slow
Three Options for Computing a Barcode Template. While an update to an RU decomposition involving few transpositions is very fast in practice, an update to an RU decomposition requiring many transpositions can be quite slow; when many transpositions
are required, it is sometimes much faster to simply recompute the RU -decomposition from
scratch using the standard persistence algorithm. In our setting, the practical performance
of our algorithm can be greatly improved if for consecutive 2-cells ei−1 , ei ∈ Γ with the
edge weight w(ei−1 , ei ) greater than some suitably chosen threshold t, we simply compute
the RU -decomposition of Φi from scratch, directly from D1i and D2i .
Moreover, we can obtain significant additional speedups by avoiding the computation
of the full RU -decomposition of Φi altogether at some cells ei . To explain this, we first note
that to obtain T ei via Eq. (6.1), we do not need the full RU -decomposition Φi , but only
PE(Φi ) = (Pairs(Φi ), Ess(Φi )); in particular we do not need U1i and U2i . The algorithm for
computing barcode templates described in Section 6 maintains the full RU -decomposition
of each Φi because the vineyard algorithm requires this. But if we are willing to compute
60
PE(Φi+1 ) from scratch, then it is not necessary to compute the full RU -decomposition of
Φi ; it suffices to compute PE(Φi ). Further, if in this case we have that ei = ej for some
j < i, then we do not even need to compute PE(Φi ) at all at cell ei , since we have already
done so at an earlier step.
In recent years, several algorithms for computing barcodes have been introduced which
are much faster than the standard persistence algorithm [3, 4, 42]. For example, a few such
algorithms are implemented in the software library PHAT [4]. Given a FI-rep Φ as input,
these algorithms compute PE(Φ), but do not compute the full RU -decomposition of Φ.3
Let us restrict our attention a single such algorithm, say, the clear and compress algorithm implemented in PHAT [3].
To compute PE(Φi ), then, we have three options available to us:
(A) Use the clear and compress algorithm if ei 6= ej for all j < i; do nothing if ei = ej for
some j < i.
(B) Compute the full RU -decomposition of Φi from scratch using the standard persistence
algorithm.
(C) Use vineyard updates. This option is only available if we chose option B or C at cell
ei−1 , so that the full RU -decomposition of Φi−1 was computed.
Clearly, there is a tradeoff between options A and B: Option A is much faster, but
choosing option A at cell ei precludes the use of option C at cell ei+1 . How, then, do we
choose between these three options at each cell ei ? We formulate this problem as a discrete
optimization problem, which can solved efficiently by reduction to a min-cut problem.
Estimating Runtimes of the Different Options. Our formulation of the problem
requires us to first estimate the respective runtimes ci (A), ci (B), and ci (C) of options A,
B, and C at each cell ei in the path Γ = e0 , e1 , . . . , ew . We will describe a simple strategy
for this here, and then explain below how to modify our approach to correct for a drawback
of the strategy.
We take ci (A) = 0 if ei = ej for some j < i, and otherwise we take ci (A) to be some
constant c(A) independent of i. Similarly, we take ci (B) = c(B) to be independent of i.
To compute c(A), we compute PE(Φ0 ) using option A, and set c(A) to be the runtime of
this computation. Similarly, to compute c(B), we compute the the full RU -decomposition
of Φ0 from scratch and take c(B) to be the runtime.
We set c0 (C) arbitrarily, say c0 (C) = 0. To compute ci (A) for each i ≥ 1, we perform
several thousand random vineyard updates to the RU -decomposition of Φ0 . Using timing
data from these computations, we compute, for j = 1, 2, the average runtime cvine
of an
j
3
However, some fast algorithms for persistence computation can be readily adapted to compute R1 , R2 ,
and U2−1 . For example, as explained to us by Ulrich Bauer, this is true for the “twist” variant of the
standard persistence algorithm [22].
U1−1 ,
61
update to the RU -decomposition corresponding to a transposition of adjacent elements in
[mj ]. Letting L denote the anchor line containing the shared boundary of ei−1 and ei , and
recalling the notation of Section 6.10, we take
1
1
vine
ci := cvine
1 (swL (gr1 ) + sepL (gr1 )) + c2 (swL (gr2 ) + sepL (gr2 )).
4
4
Some motivation for this choice of ci is provided by Remark 6.8.
The Optimization Problem. To decide between options A, B and C at each cell ei ∈ Γ,
we solve the following optimization problem:
minimize:
w
X
ci (Xi )
i=0
subject to: Xi ∈ {A, B, C},
Xi = A =⇒ Xi+1 6= C,
X0 6= C.
Clearly, this problem is equivalent to the integer linear program (ILP):
minimize:
w
X
ci (A)xi + ci (B)yi + ci (C)zi
i=0
subject to: xi , yi , zi ∈ {0, 1}
xi + yi + zi = 1
xi + zi+1 ≤ 1,
z0 = 0.
Using the constraints xi + yi + zi = 1, we can eliminate the variables yi from this ILP to
obtain an equivalent ILP with a simpler set of constraints:
minimize:
w
X
(ci (A) − ci (B))xi + (ci (C) − ci (B))zi
i=0
subject to: xi , zi ∈ {0, 1}
xi + zi+1 ≤ 1,
z0 = 0.
The constraint matrix associated to the latter ILP is of a standard form, well known to be
totally modular [43]. While an ILP with totally unimodular constraint matrix can always
be solved directly via linear programming relaxation, it is often the case that such an ILP
62
can be cast as a network flow problem, in which case we can take advantage of very efficient
specialized algorithms.
In fact, as explained to us by John Carlsson, the simplified ILP above can be cast
as the problem of finding a minimum cut in a network: First, the ILP can be cast as a
maximum-weight independent set problem in a bipartite graph with non-negative vertex
weights. In any graph, the complement of an independent set is a vertex cover and vice
versa, so the latter problem is in turn equivalent to a minimum-cost vertex cover problem
in a bipartite graph. It is well known that such a problem can be solved by computing a
minimum cut in a flow network [1, Section 8.4.3.1].
Dynamic Updates to our Estimates of Runtime Cost. As mentioned above, there
is a drawback to the approach to barcode template computation we have proposed here:
c(A) and c(B) may not be very good estimates of the respective average costs of option A
and option B; after all, our estimates c(A) and c(B) are computed using very little data.
Here is one way to correct for this: We first solve the optimization problem above for
just the first few cells (say 5%) in the path Γ. Using the solution, we then compute the
barcode template for each of these cells. As we do this, we record the runtime of the
computation at each cell. We next update the value of c(A) to the average run time of all
vine in
computations performed using option A thus far. We also update c(B), cvine
1 , and c2
the analogous way. We then use these updated values as input to the optimization problem
for the next 5% of cells in Γ. We continue in this way until the estimates of c(A), c(B),
vine have stabilized. Finally, we solve the optimization problem for all of the
cvine
1 , and c2
remaining cells in Γ and use the solution to compute the remaining barcode templates.
8.2
Coarsening of Persistence Modules
We have seen that size of the augmented arrangement A• (M ) depends quadratically on the
coarseness κ of M , and that computing A• (M ) requires O(m3 κ+(m+log κ)κ2 ) elementary
operations.
Thus, to keep our computations small, we typically want to limit the size of κ by
coarsening our module. As we now explain, doing this quite simple. A similar coarsening
scheme is mentioned in [12].
Let G = (G1 , G2 , . . . , Gn ) : Z2 → R2 be any grid function, as defined in Section 2.5.
G extends to a functor Zn → Rn , which we also denote by G. For M a 2-D persistence
module, let M G be a continuous extension of M ◦ G. In view of Proposition 2.10, the
coarseness of the grid G controls the coarseness κ of the module M G .
Let dI denote the multidimensional interleaving distance, as defined in [40]. As explained there, dI is a particularly well-behaved metric on persistence modules. The following proposition, whose easy proof we omit, makes precise the intuitive idea that a small
amount of coarsening leads to a small change in our persistence module:
Proposition 8.1. If |Gj (z)−Gj (z +1)| ≤ δ for i = 1, 2 and all z ∈ Z, then dI (M , M G ) ≤ δ.
63
The “external stability theorem” of Landi [38], mentioned earlier in Section 1.5, shows
that if 2-D persistence modules M and N are close in the interleaving distance, then the
fibered barcodes B(M ) and B(N ) will be close, in a precise sense. This justifies the use of
coarsening in conjunction with our visualization paradigm.
Coarsening Free Implicit Representations. As explained in Section 5.2, in practice,
we typically have access to M via a FI-rep Φ = (gr1 , gr2 , D1 , D2 ) for M . Since our algorithm
for computing an augmented arrangement takes a FI-rep as input, to compute A• (M G ),
we want to first construct a FI-rep ΦG of M G from Φ.
Let clG : Rn → G be the function which takes each a ∈ Rn to the minimal z ∈ G with
a ≤ z.
Define ΦG = (clG ◦ gr1 , clG ◦ gr2 , D1 , D2 ). We leave the proof of the following to the
reader:
Proposition 8.2. H(ΦG ) ' M G .
Thus, to obtain a FI-rep of a coarsening of M , it suffices to simply coarsen the grade
functions in our FI-rep of M .
8.3
Parallelization
The problem of computing the barcode templates T e is embarrassingly parallelizable. Here
is one very simple parallelization scheme: Given l2 processors, where l is less than or equal
to the number of anchors, we can choose the l anchor lines L with the largest values of w(L).
These lines divide A(M ) into at most l2 polygonal cells {Ck } with disjoint interiors, and
the remaining anchor lines induce a line arrangement A(M )k on each Ck . On processor k,
we can run our main serial algorithm, described above, to compute the barcode templates
at each 2-cell of A(M )k .
For this, we need to make just one modification to the algorithm: When choosing our
path Γ through the 2-cells of A(M )k , we generally cannot choose our initial cell e to be
the cell e0 described in Section 6.6, since e0 may not be be contained in A(M )k . Instead,
we choose the initial cell e arbitrarily.
This means we cannot use the approach of Section 6.8 to initialize the data structures
sigj , sigInvj , liftsj , and levSetj (·) at cell e. One way to initialize these data structures is
to chose an arbitrary affine line L with D` (L) ∈ e, and consider the behavior of the map
pushL on S ∪ im(gr1 ) ∪ im(gr2 ), using exact arithmetic where necessary.
9
Preliminary Runtime Results
We now present runtimes for the computation of augmented arrangements arising from
synthetic data.
64
We emphasize that these computational results are preliminary, as our implementation
of RIVET does not yet take advantage of some key optimizations. First, the code that
produced these results employs a highly simplified variant of the scheme detailed in Section 8.1 for computing barcode templates; this variant chooses only between options (B)
and (C) at each edge crossing, which is less efficient than what is proposed in Section 8.1.
Secondly, this code stores the columns of our sparse matrices using linked lists; it is known
that a smarter choice of data structure for storing columns can lead to major speedups in
persistence computation [4]. Finally, as mentioned in Section 1.10, our current implementation runs only on a single processing core. We expect to see substantial speedups after
parallelizing the computation of barcode templates as proposed in Section 8.3.
Our computations were run on a single (slow) 800 MHz core in a 32-core server with 96
GB of RAM. However, for the computations reported here, only a fraction of the memory
was required. For example, RIVET used approximately 9.3 GB of RAM for our largest
computation.
Noisy Circle Data. Data sets we consider are point clouds sampled with noise from an
annulus, such as the center point cloud in Fig. 2. Specifically, 90% of the points in each
data set X are sampled randomly from a thick annulus in a plane, and 10% are sampled
randomly from a square containing the annulus. We define a codensity function γ : X → R
by taking γ(p) to be equal to the number of points of X within some fixed distance to p.
We then construct the Vietoris-Rips bifiltration F := Rips(γ), described in Section 1.4,
taking the metric on X to be the Euclidean distance, with the scale parameter for the
Vietoris-Rips complexes capped at a value slightly larger than the inner diameter of the
annulus.
Computing the Graded Betti Numbers. Table 3 displays the average runtimes for
computing the graded Betti numbers ξ0 , ξ1 , and ξ2 of Hi (F), for i ∈ {1, 2}. Each row
gives the averages from three point clouds of the specified size. For example, we generated
three point clouds of 100 points, and the average number of 1-simplices in the resulting
bifiltrations was 3,820, so computing H0 homology required working with a bifiltration of
average size 3,920 simplices.
Building the Augmented Arrangement. Table 4 displays the average runtimes (in
seconds) to build A• (H0 (F)). As before, each row gives the averages from three point
clouds of the specified size. The average runtimes for computing A• (H0 (F)) for each of
four different coarsenings are displayed in the table. Similarly, Table 5 displays the average
runtimes (in seconds) to build A• (H1 (F)).
65
Table 3: Average runtimes for computing the bigraded Betti numbers of the noisy circle data
Points
Simplices
Runtime (sec.)
H0
100
200
300
400
3,920
15,719
35,324
62,415
0.11
0.70
2.38
3.99
H1
100
200
300
400
91,876
755,211
2,560,718
6,052,584
4.01
59.4
264
790
Table 4: Runtimes for computing the augmented arrangement for 0th homology of the noisy circle
data
Runtimes (seconds)
Points
100
200
300
400
Simplices
Bins: 5 × 5
10 × 10
15 × 15
20 × 20
3,920
15,719
35,324
62,415
0.17
0.40
0.46
1.25
0.39
7.90
10.3
22.0
0.97
21.9
46.7
180
1.51
46.1
113
637
Table 5: Runtimes for computing the augmented arrangement for 1st homology of the noisy circle
data
Runtimes (seconds)
10
Points
Simplices
Bins: 5 × 5
10 × 10
15 × 15
20 × 20
100
200
300
400
91,876
755,211
2,560,718
6,052,584
0.60
13.1
39.9
228
1.33
211
374
6,893
2.35
1,012
3,364
18,346
3.61
3,368
7,510
59,026
Conclusion
In this paper, we have introduced RIVET, a practical tool for visualization of 2-D persistence modules. RIVET provides an interactive visualization of the barcodes of 1-D affine
slices of a 2-D persistence module, as well as visualizations of the dimension function and
bigraded Betti numbers of the module. We have presented a mathematical theory for our
visualization paradigm, centered around a novel data structure called an augmented ar-
66
rangement. We have also introduced and analyzed an algorithm for computing augmented
arrangements, and described several strategies for improving the runtime of this algorithm
in practice.
In addition, we have presented timing data from preliminary experiments on the computation of augmented arrangements. Though we have yet to incorporate several key
optimizations into our code, the results demonstrate that our current implementation already scales well enough to be used to study bifitrations with millions of simplices. With
more implementation work, we expect RIVET to scale well enough to be used in many of
the same settings where 1-D persistence is currently used for exploratory data analysis.
From here, there are several natural directions to pursue. Beyond continuing to improve
our implementation of RIVET, we would like to:
• Apply RIVET to the exploratory analysis of scientific data.
• Develop statistical foundations for our data analysis methodology.
• Adapt the RIVET paradigm in the setting of 0-D homology to develop a tool for
hierarchical clustering and interactive visualization of bidendrograms [11].
• Extend the RIVET methodology to other generalized persistence settings, such cosheaves of vector spaces over R2 [28], or cosheaves of 1-D persistence modules over
R.
We hope that RIVET will prove to be a useful addition to the existing arsenal of TDA
tools. Regardless of how it ultimately fares in that regard, however, we feel that the broader
program of developing practical computational tools for multidimensional persistence is
a promising direction for TDA, and we hope that this work can draw attention to the
possibilities for this. We believe that there is room for a diverse set of approaches.
Acknowledgements
This paper has benefited significantly from conversations with John Carlsson about pointline duality and discrete optimization. We also thank Ulrich Bauer, Magnus Botnan, and
Dmitriy Morozov, and Francesco Vaccarino for helpful discussions. The bulk of the work
presented in this paper was carried out while the authors were postdoctoral fellows at the
Institute for Mathematics and its Applications, with funds provided by the National Science
Foundation. Some of the work was completed while Mike was visiting Raul Rabadan’s lab
at Columbia University. Thanks to everyone at the IMA and Columbia for their support
and hospitality.
67
A
Appendix
A.1
Details of the RIVET Interface
Expanding on Section 1.7, we now provide some more detail about RIVET’s graphical
interface.
As discussed in Section 5, the module M is input to RIVET as a free implicit representation
Φ = (gr1 , gr2 , D1 , D2 ).
RIVET uses gr1 and gr2 to choose the bounds for the Line Selection Window and Persistence Diagram Window. To explain, let A and B denote the greatest lower bound and least
upper bound, respectively, of im gr1 ∪ im gr2 . In order to avoid discussion of uninteresting
edge cases, we will assume that A1 < B1 and A2 < B2 .
Choice of Bounds for the Line Selection Window. We take the lower left corner and
upper right corner of the Line Selection Window to be A and B, respectively.4 By default,
the Line Selection Window is drawn to scale. A toggle switch rescales (i.e., normalizes)
the window so that it is drawn as a square on the screen.
Parameterization of Lines for Plotting Persistence Diagrams. We next explain
how, given a line L ∈ L̄, RIVET represents B(M L ) as a persistence diagram. Let us first
assume that the Line Selection Window is unnormalized; we treat the case where it is
normalized at the end of this section. Further, by translating the indices of M if necessary,
we may assume without loss of generality that A = 0.
As noted in Section 1.7, to plot B(M L ) as a persistence diagram, we need to first choose
a parameterization γL : R → L of L. We choose γL to be the unique order-preserving
isometry such that:
1. If L has finite positive slope, then γL (0) is the unique point in the intersection of L
with the union of the non-negative portions of the coordinate axes.
2. If L is the line x = a, then γL (0) = (a, 0).
3. If L is the line y = a, then γL (0) = (0, a).
4
For a more intrinsic choice of bounds for the Line Selection Window, one could instead take the lower left
and upper right corners of the window to be the greatest lower bound and least upper bound, respectively, of
the set {a ∈ R2 | ξi (M )(a) > 0 for some i ∈ {0, 1, 2}}. However, we feel that for a typical TDA application,
the extrinsic bounds for the Line Selection Window we have proposed provide a more intuitive choice of
scale.
68
The Choice of Bounds for the Persistence Diagram Window. The bounds for
the Persistence Diagram Window are are chosen statically, depending on M but not on
the choice of L. RIVET chooses the viewable region of the persistence diagram to be
[0, |B|] × [0, |B|].
Representation of Points Outside of the Viewable Region of the Persistence
Diagram. It may be that for some choices of L, B(M L ) contains intervals [α, β), for
α ≥ |B| or β ≥ |B|, so that [α, β) falls outside of the viewable region of the persistence
diagram; indeed, the coordinates of some points in the persistence diagram can become
huge (but finite), as the slope of the L approaches 0 or ∞. Thus, our persistence diagrams
include some information on the top of the diagram, not found in typical persistence
diagram visualizations, to record the points in the persistence diagram which fall outside
of the viewable region:
• Above the main square region of the persistence diagram are two narrow horizontal
strips, separated by a dashed horizontal line. The upper strip is labeled “inf,” while
the lower is labeled “< inf”. In the higher strip, we plot a point with x-coordinate
α for each interval [α, ∞) with α ≤ |B|. In the lower strip, we plot a point with
x-coordinate α for each interval [α, β) ∈ B(M L ) with α ≤ |B| and |B| < β < ∞.
• Just to the right of each of the two horizontal strips is a number, separated from the
strip by a vertical dashed line. The upper number is the count of intervals [α, ∞) ∈
B(M L ) with |B| < α; the lower number is the count of intervals [α, β) ∈ B(M L ) with
|B| < α, β < ∞.
Persistence Diagrams Under Rescaling. If we choose to normalize the Line Selection Window, then RIVET also normalizes the persistence diagrams correspondingly. To
do this, it computes respective affine normalizations gr01 , gr02 of gr1 , gr2 , so that after normalization, A = (0, 0) and B = (1, 1). RIVET then chooses the parameterizations of lines
L ∈ L̄ and computes bounds on the Persistence Diagram Window exactly as described
above in the unnormalized case, but taking the input FI-rep to be (gr01 , gr02 , D1 , D2 ).
A.2
Computing the Lists levSetj (u) at Cell e0
As mentioned in Section 6.8, to efficiently compute the lists levSetj (u) at cell e0 , we use
a sweep algorithm. As we did in Section 5.4, to keep notation simple we assume that Ωx
and Ωy are the identity maps on [nx ] and [ny ] respectively, so that (Ωx × Ωy )−1 (P) = P.
We assume that to start, gr1 and gr2 are both increasing with respect to colexicographical
order; if this is not the case, then by Lemma 6.3, we can apply a sorting algorithm to
modify Φ so that the assumption does hold. This sorting can be done in O(m log m) time.
69
Our sweep algorithm maintains a linked list Frontier of pointers to elements of P. (These
elements are stored in tPtsMat.) Both the x- and y-coordinates of the entries of Frontier
are always strictly decreasing.
The algorithm iterates through the rows of the grid [nx ] × [ny ] from top to bottom.
The list Frontier is initially empty, and is updated at each row s, with the help of tPtsMat.
If row s is empty, then no update is necessary at row s. Otherwise, let u be the rightmost
entry of P in row s, and let r be the column containing u. If the last element of Frontier is
also in column r, then this element is removed from Frontier and replaced by u. Otherwise,
we append u to the end of Frontier.
The algorithm then inserts each k ∈ [mj ] with grj (k) = (r, s) for some r into the
appropriate list levSetj (−). (Since grj is assumed to be increasing with respect to colexicographical order, we have immediate access to all such k.) Specifically, k is added to the
list levSetj (u), for u the leftmost element of Frontier with r ≤ u1 . The lists levSetj (u) are
maintained in lexicographical order. It is easy to check that u = lift0 ◦ grj (k), as desired.
The frontier-sweep algorithm is stated in pseudocode in Algorithm 2.
Updating Frontier at each row of tPtsMat takes constant time, so the total cost of
updating Frontier is O(ny ). For each row s, we must iterate over Frontier once to identify
the lists levSetj (−) into which we insert elements of [mj ]. There are ny rows, and the length
of Frontier is O(κx ), so the total cost of these iterations over Frontier is O(ny κx ). Inserting
each k ∈ [mj ] into the appropriate list levSetj (−) takes constant time, so the total cost
of such insertions is O(m). Thus, the total number of operations for the frontier-sweep
algorithm (including the cost of the initial sorting to put gr1 and gr2 in the right form), is
O(m log m + ny κx ) = O(m log m + mκ).
Algorithm 2 Frontier-sweep algorithm for building the lists levSetj (u)
Input: grj (represented as a list in colexicographical order), tPtsMat
Output: tPtsMat, updated so that for each u ∈ P e0 , levSetj (u) = (lift0 ◦ grj )−1 (u), with each list sorted
in lexicographical order
1: initialize Frontier as an empty linked list
2: for s = ny to 1 do
3:
if tPtsMat has an entry in row s then
. update Frontier for row s
4:
let r be the column containing rightmost element of P in row s
5:
if last element of Frontier is in column r then
6:
remove the last entry from Frontier
7:
append the entry of P at row s, column r to the end of Frontier
8:
for each entry u in Frontier do
. add elements with grades in row s to the lists levSetj (−)
9:
if u is the last element of Frontier then
10:
add all k ∈ [mj ] such that grj (k) = (r, s) with r ≤ u1 to levSetj (u)
11:
else
12:
let v be the element after u in Frontier
13:
add all k ∈ [mj ] such that grj (k) = (r, s) with v1 < r ≤ u1 to levSetj (u)
70
References
[1] M. J. Atallah. Algorithms and theory of computation handbook. CRC press, 1998.
[2] M. F. Atiyah. On the krull-schmidt theorem with application to sheaves. Bulletin de la Société
Mathématique de France, 84:307–317, 1956.
[3] U. Bauer, M. Kerber, and J. Reininghaus. Clear and compress: Computing persistent homology
in chunks. In Topological Methods in Data Analysis and Visualization III, pages 103–117.
Springer, 2014.
[4] U. Bauer, M. Kerber, J. Reininghaus, and H. Wagner. Phat – persistent homology algorithms
toolbox. In Mathematical Software – ICMS 2014, volume 8592 of Lecture Notes in Computer
Science, pages 137–143. Springer Berlin Heidelberg, 2014.
[5] U. Bauer and M. Lesnick. Induced matchings of barcodes and the algebraic stability of persistence. In Proceedings of the 2014 Annual Symposium on Computational Geometry, page 355.
ACM, 2014.
[6] S. Biasotti, A. Cerri, P. Frosini, and D. Giorgi. A new algorithm for computing the
2-dimensional matching distance between size functions.
Pattern Recognition Letters,
32(14):1735–1746, 2011.
[7] H. Brönnimann, A. Fabri, G.-J. Giezeman, S. Hert, M. Hoffmann, L. Kettner, S. Pion, and
S. Schirra. 2D and 3D geometry kernel. In CGAL User and Reference Manual. CGAL Editorial Board, 4.5.2 edition, 2015. http://doc.cgal.org/4.5.2/Manual/packages.html#
PkgKernel23Summary.
[8] P. Bubenik, V. de Silva, and J. Scott. Metrics for generalized persistence modules. Foundations
of Computational Mathematics, pages 1–31, 2014.
[9] G. Carlsson. Topological pattern recognition for point cloud data. Acta Numerica, 23:289–368,
May 2014.
[10] G. Carlsson, V. de Silva, and D. Morozov. Zigzag persistent homology and real-valued functions. In Proceedings of the 25th annual symposium on Computational geometry, pages 247–256.
ACM, 2009.
[11] G. Carlsson and F. Mémoli. Multiparameter hierarchical clustering methods. Classification as
a Tool for Research, pages 63–70, 2010.
[12] G. Carlsson, G. Singh, and A. Zomorodian. Computing multidimensional persistence. Algorithms and Computation, pages 730–739, 2009.
[13] G. Carlsson and A. Zomorodian. The theory of multidimensional persistence. Discrete and
Computational Geometry, 42(1):71–93, 2009.
[14] A. Cerri, B. Di Fabio, M. Ferri, P. Frosini, and C. Landi. Betti numbers in multidimensional persistent homology are stable functions. Mathematical Methods in the Applied Sciences,
36(12):1543–1557, 2013.
71
[15] A. Cerri, B. Di Fabio, G. Jablonski, and F. Medri. Comparing shapes through multi-scale
approximations of the matching distance. Computer Vision, 2014.
[16] W. Chacholski, M. Scolamiero, and F. Vaccarino. Combinatorial resolutions of multigraded
modules and multipersistent homology. arXiv preprint arXiv:1206.1819, 2012.
[17] F. Chazal, D. Cohen-Steiner, M. Glisse, L. Guibas, and S. Oudot. Proximity of persistence
modules and their diagrams. In Proceedings of the 25th annual symposium on Computational
geometry, pages 237–246. ACM, 2009.
[18] F. Chazal, D. Cohen-Steiner, L. Guibas, F. Mémoli, and S. Oudot. Gromov-Hausdorff stable signatures for shapes using persistence. In Proceedings of the Symposium on Geometry
Processing, pages 1393–1403. Eurographics Association, 2009.
[19] F. Chazal, D. Cohen-Steiner, and Q. Mérigot. Geometric inference for probability measures.
Foundations of Computational Mathematics, pages 1–19, 2011.
[20] F. Chazal, V. de Silva, M. Glisse, and S. Oudot. The structure and stability of persistence
modules. arXiv preprint arXiv:1207.3674, 2012.
[21] F. Chazal, V. De Silva, and S. Oudot. Persistence stability for geometric complexes. Geometriae
Dedicata, 173(1):193–214, 2014.
[22] C. Chen and M. Kerber. Persistent homology computation with a twist. In Proceedings 27th
European Workshop on Computational Geometry, volume 11, 2011.
[23] N. Christofides. Worst-case analysis of a new heuristic for the travelling salesman problem.
Report 388, Graduate School of Industrial Administration, CMU, 1976.
[24] D. Cohen-Steiner, H. Edelsbrunner, and J. Harer. Stability of persistence diagrams. Discrete
and Computational Geometry, 37(1):103–120, 2007.
[25] D. Cohen-Steiner, H. Edelsbrunner, and D. Morozov. Vines and vineyards by updating persistence in linear time. In Proceedings of the twenty-second annual symposium on Computational
geometry, pages 119–126. ACM, 2006.
[26] T. Cormen, C. Leiserson, R. Rivest, and C. Stein. Introduction to Algorithms. MIT Press,
2009.
[27] W. Crawley-Boevey. Decomposition of pointwise finite-dimensional persistence modules. Journal of Algebra and Its Applications, 14(05):1550066, 2015.
[28] J. Curry. Sheaves, cosheaves and applications. Ph.D. Dissertation, University of Pennsylvania,
2014.
[29] M. de Berg, O. Cheong, M. van Kreveld, and M. Overmars. Computational Geometry: Algorithms and Applications. Springer, 2008.
[30] H. Derksen and J. Weyman. Quiver representations. Notices of the AMS, 52(2):200–206, 2005.
72
[31] H. Edelsbrunner and J. Harer. Computational topology: an introduction. American Mathematical Society, 2010.
[32] H. Edelsbrunner, D. Letscher, and A. Zomorodian. Topological persistence and simplification.
Discrete and Computational Geometry, 28(4):511–533, 2002.
[33] D. Eisenbud. Commutative Algebra: with a view toward algebraic geometry, volume 150.
Springer Science & Business Media, 1995.
[34] D. Eisenbud. The geometry of syzygies: a second course in algebraic geometry and commutative
algebra, volume 229. Springer Science & Business Media, 2005.
[35] J. Hoogeveen. Analysis of christofides’ heuristic: Some paths are more difficult than cycles.
Operations Research Letters, 10(5):291–295, 1991.
[36] M. Kreuzer and L. Robbiano. Computational commutative algebra 2, volume 1. Springer, 2005.
[37] R. La Scala and M. Stillman. Strategies for computing minimal free resolutions. Journal of
Symbolic Computation, 26(4):409–431, 1998.
[38] C. Landi. The rank invariant stability via interleavings. arXiv preprint arXiv:1412.3374, 2014.
[39] S. Lang. Algebra, revised third edition. Graduate Texts in Mathematics, 2002.
[40] M. Lesnick. The theory of the interleaving distance on multidimensional persistence modules.
Foundations of Computational Mathematics, 15(3):613–650, 2015.
[41] M. Lesnick and M. Wright. Computing multigraded betti numbers of 2-D persistent homology
modules in cubic time. In preparation, 2015.
[42] N. Otter, M. A. Porter, U. Tillmann, P. Grindrod, and H. A. Harrington. A roadmap for the
computation of persistent homology. arXiv preprint arXiv:1506.08903, 2015.
[43] A. Schrijver. Theory of linear and integer programming. John Wiley & Sons, 1998.
[44] C. D. Toth, J. O’Rourke, and J. E. Goodman. Handbook of discrete and computational geometry. CRC press, 2004.
[45] L. Wasserman. All of statistics: a concise course in statistical inference. Springer Verlag, 2004.
[46] C. Webb. Decomposition of graded modules. Proceedings of the American Mathematical
Society, 94(4):565–571, 1985.
[47] A. Zomorodian and G. Carlsson. Computing persistent homology. Discrete and Computational
Geometry, 33(2):249–274, 2005.
73
Notation Index
A• (M )
augmented arrangement of M , page 13
A(M )
line arrangement, page 13
B(M )
barcode of a 1-D persistence module, page 4; fibered barcode of a 2-D persistence module,
page 9
box(S)
quadrant in R2 with upper right corner lub S, page 42
D` , Dp
point-line duality transforms, page 25
e
cell in A(M ), page 13
EG
f
continuous extension functor, page 22
e
template map, page 46
FI-rep
free implicit representation, page 34
flG
floor function for im G, page 21
G
n-D grid function, page 21
G
dual graph of A(M ), page 46
Γ
path through all 2-cells of G, page 46
grW
grade function of the n-graded set W, page 18
Hi
ith homology functor, page 4
[k]
integers {1, . . . , k}, page 34
κ
coarseness of a 2-D persistence module, page 15
L̄
space of affine lines in R2 with non-negative slope, page 9
L
space of affine lines in R2 with non-negative, finite slope, page 9
L◦
space of affine lines in R2 with finite, positive slope, page 26
lifte
lift map at cell e, page 45
lub
least upper bound, page 26
M
multidimensional persistence module, page 4
mj
dimension of Φ, page 34
non-negative integers, page 9
N
e
ord
order map at cell e, page 45
P
set of all template points, page 38
P
e
set of template points at cell e, page 29
p.f.d.
pointwise finite dimensional, page 4
Φ
a free implicit representation, page 34
pushL
R
n
push map, page 26
poset category of Rn , page 3
rank(M ) rank invariant of M , page 9
74
S
union of the supports of the 0th and 1st bigraded Betti numbers of M , page 24
Se
totally ordered partition of S, page 29
σj
permutation of [mj ], page 44
e
T
barcode template at cell e, page 13
ξi (M )
ith (graded) Betti number of M , page 10
ζe
map from positive integers to template points at cell e, page 29
n
Z
Poset category of Zn , page 21
75
| 0math.AC
|
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
The Journal of MacroTrends in
Technology and Innovation
MACROJOURNALS
Automatic Knot Adjustment Using Dolphin
Echolocation Algorithm for B-Spline Curve
Approximation
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**
*Necmettin Erbakan University, School of Applied Sciences, Department of Management Information
Sciences
**Selcuk University, Faculty of Engineering, Department of Computer Engineering
Abstract
In this paper, a new approach to solve the cubic B-spline curve fitting problem is
presented based on a meta-heuristic algorithm called “dolphin echolocation”. The
method minimizes the proximity error value of the selected nodes that measured
using the least squares method and the Euclidean distance method of the new curve
generated by the reverse engineering. The results of the proposed method are
compared with the genetic algorithm. As a result, this new method seems to be
successful.
Keywords: B-spline Curve Approximation, Cubic B-spline, Data Parameterization on B-Spline, Dolphin
Echolocation Algorithm, Knot adjustment
1. INTRODUCTION
B-Spline curve fitting is a classical problem for computer aided geometric design [1]. For
example, de facto for the CAD / CAM and related graphic design industries, and in most
geometric modeling areas is that non-polynomial parametric curves should be transformed into
non-uniform rational B-splines. Similarly, in vector font modeling problems, fonts are often
fitted with a soft-pass B-Spline. In practical applications, the distance between the target curve
and the fitted B-Spline curve must be less than a predetermined tolerance, and the resulting
curve is called an error-bound approach. Euclidean distance method is used to measure the
value corresponding to the distance between two curves.
100
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
2. B-SPLINE CURVE FITTING PROBLEM
The problem of B-Spline curve fitting is expressing the target curve with minimum tolerance
through b-spline curves. The target curve can be two or three dimensional [2]. The scope of this
paper
The parameterization of the target data points
The convergence of the minimum error tolerance with the B-Spline curves using the
automatically placed minimum control point constitute.
A B-spline curve is expressed as equation (1).
𝑛
𝑃(𝑢) = ∑ 𝑝𝑖 𝑁𝑖 ,𝑝 (𝑢) #(1)
𝑖=0
where 𝑝𝑖 is the ith control point and 𝑁𝑖,𝑝 is the main function of B-Spline curves.
The main function 𝑁𝑖,𝑝 of B-Spline curve for given knot vector t with degree p is expressed as
equation (2).
1,
𝑡𝑖 ≤ 𝑢 ≤ 𝑡𝑖+1
𝑁𝑖,0 (𝑢) = {
0,
𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝑡𝑖+𝑝+1 − 𝑢
𝑢 − 𝑡𝑖
(𝑢) #(2)
𝑁𝑖 ,𝑝−1 (𝑢) +
𝑁
𝑡𝑖+𝑝 − 𝑡𝑖
𝑡𝑖+𝑝+1 − 𝑡𝑖+1 𝑖+1,𝑝−1
Further information on the B-Spline curves can be found [3].
𝑁𝑖,𝑝 (𝑢) =
A. Methods on Data Parameterization
Because of B-Spline curves are parametric curves, the target data points need to be
parameterized in the B-Spline curve fitting. However calculation of optimum data
parameterization is theoretically quite difficult, different ways of data parameterization are
used in applications. Three methods of uniform parameterization, chord-length
parameterization, and centripetal parameterization are emerging in researches based on
previous studies [4,5,6]. In this study centripetal parameterization method is used.
B. Euclidean Distance Minimization
The Euclidean distance is used to calculate the error between the target curve and the BSpline fitted curve. The Euclidean distance is calculated by an equation (3)
𝐿
2
𝐷 = √∑(𝐶(𝑖) − 𝐵(𝑖)) #(3)
𝑖=1
where 𝐶(𝑖) is the ith data in original dataset, 𝐵(𝑖) is the ith data in the fitted curve. The general
101
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
approach of this paper is to minimize this distance and express the B-Spline curve with
minimum control point at the same time. Thus the Euclidean distance and the number of
control points is treated together in the fitness function.
C. Dolphin Echolocation Algorithm
The dolphin echolocation algorithm presented by Kaveh and Ferhoudi is an optimization
algorithm that is inspired by the hunting principles of bottlenose dolphins through sonar waves
[7]. The dolphins explore the entire search area for a specific effect to hunt. As they approach
their prey, they try to focus on the target by limiting the number of waves they send by limiting
their search. This algorithm implements search by reducing the distance to the target. The
search space must be sorted before beginning to search. The alternatives of each variable to be
optimized must be sorted in ascending or descending order. If these alternatives have more
than one characteristic, they should be sorted according to the most important one. In the use
of this technique, for example, for the variable j, the vector Aj in length LAj forms the columns
of the Alternatives Matrix. In addition, a convergence curve is used to change the convergence
factor during the optimization process. The variation of this trend throughout the iterations is
calculated by equation (4).
𝐿𝑜𝑜𝑝𝑖𝑃𝑜𝑤𝑒𝑟 − 1
)
(1
)
𝑃𝑃(𝐿𝑜𝑜𝑝𝑖 = 𝑃𝑃1 + − 𝑃𝑃𝑖 ∗
#(4)
(𝐿𝑜𝑜𝑝𝑠𝑁𝑢𝑚𝑏𝑒𝑟)𝑃𝑜𝑤𝑒𝑟 − 1
where PP is the probability of being pre-defined, PP1 is randomly selected probability for the
first iteration, Loopi is the number of the current iteration, Power is the rank of the curve, and
LoopsNumber is the total number of iterations. Algorithm requires a location matrix LNL*NV in the
variable number NV at the location count NL.
The main steps of Dolphin Echolocation (DE) for discrete optimization are as follows:
1. Create NL locations randomly for dolphin.
2. PP of current iteration is calculated using the equation (4).
3. Fitness is calculated for every location.
4. Calculate cumulative fitness according to the following dolphin rules
(a)
for i = 1 to the number of locations
for j = 1 to the number of variables
find the position of L(i,j) in jth column of the Alternatives matrix and name it as A.
for k = -Re to Re
1
𝐴𝐹
=
∗ (𝑅𝑒 − |𝑘|) ∗ 𝐹𝑖𝑡𝑛𝑒𝑠𝑠(𝑖) + 𝐴𝐹 (A+k) #(5)
(𝐴+𝑘)𝑗
where AF(A+k)j is the cumulative fitness
for the jth variable
𝑅𝑒 of the selected alternative (𝐴+𝑘)𝑗
(the numbering of the alternatives is the same as the ordering of the alternative matrix); Re is
end
end
end
102
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
the diameter of the influence of the neighbor affected by the cumulative fitness of alternative
A. It is recommended that this diameter should not be more than 1/4 of the search space;
Fitness(i) is the fitness of the ith location. The fitness should be defined as the best answers will
get higher value. In other words, the goal of optimization should be to maximize the fitness.
AF must be calculated using a reflective property by adding alternatives near the edges (if A +
k is not valid, i.e. A + k <0 or A + k> LAj).
In this case, if the distance of the alternative to the edge is small, the same alternatives
appear in the mirror as if a mirror were placed on the edge.
(b) A small έ value is added to AF sequences AF=AF+έ in order to distribute probabilities
uniformly in search space. Here έ should be chosen according to the way of describing the
fitness. The best choice is lower than the lowest fitness value achieved.
(c) Find the best location for this loop and call it as "Best location". Find the alternatives
assigned to the best location variables and set their AF to zero.
In another saying:
for j = 1: Number of variables
for i = 1: Number of alternatives
if i = The best location(j)
𝐴𝐹𝑖𝑗 = 0
5. For the variable j (j = 1 to NV), calculate the probability by choosing alternative i (i = 1 to ALj)
endequation (6).
according to the
𝐴𝐹𝑖𝑗
end
𝑃𝑖𝑗 = 𝐿𝐴𝑗
#(6)
∑𝑖=1 𝐴𝐹𝑖𝑗
6. end
Assign PP probability to all alternatives of all selected variables for the best location and
distribute the remaining probability to other alternatives according to the form below:
for j = 1 to Number of variables
for i = 1 to Number of alternatives
if i = The best location(j)
𝑃𝑖𝑗 = 𝑃𝑃
7. Calculate the else
next step locations according to the assigned probability to each alternate.
Repeat steps 2-6 maximum
iteration number times.
𝑃 = (1 − 𝑃𝑃) ∗ 𝑃
𝑖𝑗
𝑖𝑗
end
end
end
103
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
3. AUTOMATIC KNOT ADJUSTMENT BY DOLPHIN ECHOLOCATION ALGORITHM
In the problem of B-Spline curve fitting, the fitted curve is tried to converge to the target
curve with minimum tolerance and with minimum control point. In that case, such nodes must
be selected for the given N points so that the error tolerance and the number of control points
of the nearest curve are minimum. Thus, an array of N bits is expressed 1 as selected nodes and
0 as non-selected. Thus, the alternatives for each variable are [0,1]. Each location for a dolphin
echolocation is called as solution. These solutions can be illustrated as Figure 1.
Figure 1. Sample solution illustration
For example, it is possible to express 10 points in this way with the control points to be
calculated for the 5 selected nodes. The aim of dolphin echolocation is maximizing the fitness
for equation (7) can be used as fitness function.
𝑓=
1
#(7)
𝑁𝑢𝑚𝑏𝑒𝑟𝑜𝑓𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑃𝑜𝑖𝑛𝑡 ∗ 𝐸𝑟𝑟𝑜𝑟𝑉𝑎𝑙𝑢𝑒
The B-Spline curve fitting process with the dolphin echolocation algorithm is as follows.
1. Create random solutions for the startup population.
2. Calculate the PP of current iteration.
3. Calculate the fitness value for all possible solutions.
4. Calculate the cumulative fitness of the variables in each possible solution.
5. Find the best solution according to maximum fitness.
6. Set the cumulative fitness of all solutions variables to 0 which variables equal to the
variables of the best solution.
7. Calculate the probabilities of alternatives for each variable in all solutions.
8. Set the probabilities of all alternatives equal to the variables of the best solution to
probability of the current iteration.
104
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
9. Find the possible solutions to be used in the next iteration by the probabilities of the
alternatives for each variable.
10. Repeat steps 2-9 for the number of iteration times.
4. EXPERIMENTAL RESULTS
A. Experimental Curve
The target is a curve of 200 points. The approximation results of the 3rd degree B-Spline
curves are as shown in Table 1.
GENETIC ALGORITHM
ITERATION
RMSE
10
25
50
100
250
500
1000
2000
4,36660
3,90026
2,18640
2,93898
2,49795
1,75746
1,97076
1,66949
Euclidean Number
Distance
of
Control
Point
670,6340
92
607,9586
96
308,7266
108
418,3969
93
350,5132
97
243,1649
101
278,8363
97
256,1887
98
DOLPHIN ECHOLOCATION ALGORITHM
Fitness
RMSE
61698,32
58364,02
33342,48
38910,91
33999,78
24559,65
27047,12
25106,49
3,92919
3,25922
1,62066
1,88715
1,81226
1,75954
1,41782
1,59712
Euclidean Number
Distance
of
Control
Point
572,6328
95
471,9693
93
233,4690
97
252,2174
98
244,5128
84
226,2702
93
201,0667
89
201,9198
87
Fitness
54400,11
43893,14
22646,50
24717,30
20539,08
21043,13
17894,94
17567,02
Table 1. Experimental results for different number of iteration.
Plotted experimental results is shown in Figure 2.
Figure 2. A) Original Curve, B) Genetic Algorithm, C) Dolphin Echolocation Algorithm
105
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
B. Epitrochoid Curve
The
target
is
a
curve
of
361
points.
The
curve equation is as follows.
𝑎
𝑥(𝑡) = (𝑎 + 𝑏) ∗ cos(𝑡) − ℎ ∗ cos (( + 1) ∗ 𝑡)
𝑏
𝑎
𝑦(𝑡) = (𝑎 + 𝑏) ∗ sin(𝑡) − ℎ ∗ sin (( + 1) ∗ 𝑡)
𝑏
For the parameters a = 5, b = 1 and h = 4, The approximation results of the 3rd degree B-Spline
curve for the curve calculated at -180 <= t <= 180 are as in Table 2.
GENETIC ALGORITHM
ITERATION
10
25
50
100
250
500
1000
2000
RMSE
Euclidea
n
Distance
5,2401
Number of
Control
Point
174
DOLPHIN ECHOLOCATION ALGORITHM
Fitnes
s
RMSE
Euclidea
n
Distance
3,3247
Number of
Control
Point
189
Fitness
192
663,78
185
680,02
195
583,69
185
553,83
181
565,97
179
312,68
166
328,61
0,015
911,7 0,0068
10
8
9
0,007
3,1472
196
616,8 0,0085
3,4572
95
4
1
0,009
3,6672
179
656,4 0,0090
3,6758
37
4
6
0,004
2,4077
193
464,6 0,0065
2,9933
60
9
2
0,007
3,2174
164
527,6 0,0073
2,9937
67
5
9
0,005
2,5488
170
433,3 0,0063
3,1269
71
0
8
0,002
1,7965
189
339,5 0,0025
1,7468
45
4
2
0,003
1,8436
198
365,0 0,0030
1,9796
14
3
8
Table 2. Experimental results for different number of iteration.
628,37
106
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
Plotted experimental results is shown in Figure 3.
Figure 3. A) Original Curve, B) Genetic Algorithm, C) Dolphin Echolocation Algorithm
C. Archimedean Spiral
The target is a curve of 100 points. The curve equation is as follows.
𝑟=𝑎∗ 𝑡
𝑥(𝑡) = 𝑟 ∗ cos(𝑡)
𝑦(𝑡) = 𝑟 ∗ sin(𝑡)
For the a = 2, The approximation results of the 3rd degree B-Spline curve for the curve
calculated at 0 <= t <= π are as shown in Table 3.
GENETIC ALGORITHM
ITERATION
10
25
50
100
250
500
1000
2000
RMSE
Euclidean
Distance
DOLPHIN ECHOLOCATION ALGORITHM
Number Fitness
RMSE
Euclidean Number
of
Distance of Control
Control
Point
Point
0,00468
1,3419
46
61,73
0,00279
1,2520
47
0,00356
1,2649
43
54,39
0,00468
1,3680
48
0,00150
1,1490
50
57,45
0,00417
1,3182
44
0,00178
1,1159
48
53,56
0,00340
1,2512
38
0,00209
1,1604
38
44,09
0,00535
1,3127
32
0,00205
1,1405
38
43,34
0,00313
1,2247
34
0,00209
1,1435
41
46,88
0,00257
1,2135
36
0,00318
1,2054
33
39,78
0,00277
1,1890
36
Table 3. Experimental results for different number of iteration.
Fitness
58,84
65,67
58,00
47,54
42,01
41,64
43,69
42,81
107
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
Plotted experimental results is shown in Figure 4.
Figure 4. A) Original Curve, B) Genetic Algorithm, C) Dolphin Echolocation Algorithm
D. Vivaldi Curve
The target curve is a curve of 241 points. The curve equation is as follows.
𝑥(𝑡) = 𝑎 ∗ (1 + cos(𝑡))
𝑦(𝑡) = 𝑎 ∗ sin(𝑡)
1
𝑧(𝑡) = 2 ∗ 𝑎 ∗ sin ( ∗ 𝑡)
2
For a = 0.5, The approximation results of the 3rd degree B-Spline curve for the curve
calculated at -360 <= t <= 360 are as shown in Table 4.
GENETIC ALGORITHM
ITERATION
10
25
50
100
250
500
1000
2000
RMSE
Euclidea
n
Distance
Number
of Control
Point
DOLPHIN ECHOLOCATION ALGORITHM
Fitness
RMSE
Euclidea
n
Distance
0,00036 1,0711
108
115,67 0,00026 1,0599
0,00021 1,0578
101
106,84 0,00029 1,0636
0,00024 1,0431
90
93,88
0,00054 1,0957
0,00015 1,0271
96
98,60
0,00037 1,0805
0,00013 1,0231
95
97,20
0,00062 1,1046
0,00015 1,0256
80
82,05
0,00074 1,1149
0,00018 1,0327
71
73,32
0,00036 1,0759
0,00028 1,0419
59
61,47
0,00066 1,1431
Table 4. Experimental results for different number of iteration.
Plotted experimental results is shown in Figure 5.
Number
of
Control
Point
92
96
77
79
65
53
54
40
Fitness
97,52
102,11
84,37
85,36
71,80
59,09
58,10
45,72
108
Hasan Ali AKYÜREK*, Erkan ÜLKER**, Barış KOÇER**, JMTI Vol 4 Issue 1 2016
Figure 5. A) Original Curve, B) Genetic Algorithm, C) Dolphin Echolocation Algorithm
5. CONCLUSION AND FEATURE WORK
This paper addresses the problem of curve fitting of noisy data points by using B-spline
curves. Given a set of noisy data points, the goal is to compute all parameters of the
approximating polynomial B-spline curve that best fits the set of data points in the least-squares
sense. This is a very difficult, overdetermined, continuous, multimodal, and multivariate
nonlinear optimization problem. Our proposed method solves it by applying the dolphin
echolocation algorithm. Our experimental results show that the presented method performs
very well by fitting the data points with a high degree of accuracy. A comparison with the most
popular previous approach genetic algorithm to this problem is also carried out. It shows that
our method outperforms previous approaches for the examples discussed in this paper. Future
work includes the extension of this method to other families of curves, such as NURBS and the
parametric B-spline curves. The extension of these results to the case of explicit surfaces is also
part of our future work.
REFERENCES
Park, H., & Lee, J. H. (2007). B-spline curve fitting based on adaptive curve refinement using dominant
points. Computer-Aided Design, 39(6), 439-451.
De Boor, C., De Boor, C., Mathématicien, E. U., De Boor, C., & De Boor, C. (1978). A practical guide to
splines (Vol. 27, p. 325). New York: Springer-Verlag.
Piegl, L., & Tiller, W. (2012). The NURBS book. Springer Science & Business Media.
Park, H. (2001). Choosing nodes and knots in closed B-spline curve interpolation to point data. ComputerAided Design, 33(13), 967-974.
Vassilev, T. I. (1996). Fair interpolation and approximation of B-splines by energy minimization and points
insertion. Computer-Aided Design, 28(9), 753-760.
Wang, X., Cheng, F. F., & Barsky, B. A. (1997). Energy and B-spline interproximation. Computer-Aided
Design, 29(7), 485-496.
Kaveh, A., & Farhoudi, N. (2013). A new optimization method: Dolphin echolocation. Advances in
Engineering Software, 59, 53-70.
109
| 9cs.NE
|
On Integer Programming and the Path-width of the Constraint
Matrix
Fedor V. Fomin1 , Fahad Panolan1 , M. S. Ramanujan2 , and Saket Saurabh1,3
arXiv:1607.05342v2 [cs.DS] 3 Nov 2016
1
Department of Informatics, University of Bergen, Norway.
{fomin|fahad.panolan}@ii.uib.no
2
Technische Universität Wien, Vienna, Austria. [email protected]
3
The Institute of Mathematical Sciences, HBNI, Chennai, India. [email protected]
Abstract
In the classic Integer Programming (IP) problem, the objective is to decide whether, for a
given m × n matrix A and an m-vector b = (b1 , . . . , bm ), there is a non-negative integer n-vector
x such that Ax = b. Solving (IP) is an important step in numerous algorithms and it is important
to obtain an understanding of the precise complexity of this problem as a function of natural
parameters of the input.
Two significant results in this line of research are the pseudo-polynomial time algorithms
for (IP) when the number of constraints is a constant [Papadimitriou, J. ACM 1981] and when
the branch-width of the column-matroid corresponding to the constraint matrix is a constant
[Cunningham and Geelen, IPCO 2007]. In this paper, we prove matching upper and lower bounds
for (IP) when the path-width of the corresponding column-matroid is a constant. These lower
bounds provide evidence that the algorithm of Cunningham and Geelen, are probably optimal.
We also obtain a separate lower bound providing evidence that the algorithm of Papadimitriou
is close to optimal.
1
Introduction
In the classic Integer Programming problem, the input is an m × n integer matrix A, and an m-vector
b = (b1 , . . . , bm ). The objective is to find a non-negative integer n-vector x (if one exists) such that
Ax = b. Solving this problem, denoted by (IP) is an important step in numerous algorithms and it
is important to obtain an understanding of the precise complexity of this problem as a function of
natural parameters of the input.
In 1981, Papadimitriou [26] showed that (IP) is solvable in pseudo-polynomial time on instances
for which the number of constraints m is a constant. His proof consists of two steps. The first step
is combinatorial, showing that if the entries of A and b are from {0, ±1, . . . , ±d}, and (IP) has a
solution, then there is also a solution which is in {0, 1, . . . , n(md)2m+1 }n . The second, algorithmic
step shows that if (IP) has a solution with the maximum entry at most B, then the problem is
solvable in time O((nB)m+1 ). In particular, when the matrix A happens to be non-negative, his
algorithm for IP runs in time O((nd)m+1 ) where d = max{b1 , . . . , bm }. A natural question therefore
is whether the algorithm of Papadimitriou can be improved significantly in general and in particular
for the case when A is non-negative. Our first theorem provides a conditional lower bound indicating
that any significant improvements are unlikely. To be precise, we prove the following theorem.
Theorem 1.1. Unless the Exponential Time Hypothesis (ETH) fails, (IP) with m × n matrix A
o( m )
cannot be solved in time n log m do(m) , where d = max{b1 , . . . , bm }, even when the constraint matrix
A is non-negative and each entry in any feasible solution is at most 2.
ETH is the conjecture that 3-SAT cannot be solved in time 2o(n) on n-variable formulas [21].
Due to Theorem 1.1, the simple dynamic programming algorithm [26] for (IP) when the maximum
entry in a solution, as well as in the constraint matrix, is bounded, is already close to optimal. In
fact, when the constraint matrix is non-negative, our lower bound asymptotically almost matches
the O((nd)m+1 ) running time of Papadimitriou’s algorithm [26]. Hence, we conclude that obtaining
a significant improvement over the algorithm of Papadimitriou for non-negative matrices is at least
as hard as obtaining a sub-exponential time (2o(n) ) algorithm for 3-SAT. In fact, observe that based
on the setting of the parameters m, d, n our lower bound rules out several interesting running times.
For instance, if m = Θ(n) and d = O(1), we immediately get a 2o(n) lower bound.
Continuing the quest for faster algorithms for (IP), Cunningham and Geelen [8] suggested a
new approach for solving (IP) which utilizes a branch decomposition of the matrix A. They were
motivated by the fact that the result of Papadimitriou can be interpreted as a result for matrices of
constant rank and branch-width is a parameter which is upper bounded by rank plus one. Robertson
and Seymour [28] introduced the notion of branch decompositions and the corresponding notion of
branch-width for graphs and more generally for matroids. Branch decompositions have immense
algorithmic significance because numerous NP-hard problems can be solved in polynomial time on
graphs or matroids of constant branchwidth [7, 16, 19, 18]. For a matrix A, the column-matroid
of A denotes the matroid whose elements are the columns of A and whose independent sets are
precisely the linearly independent sets of columns of A. We postpone the formal definitions of
branch decomposition and branch-width till the next section. For (IP) with a non-negative matrix
A, Cunningham and Geelen [8] showed that when the branch-width of the column-matroid of A is
constant, (IP) is solvable in pseudo-polynomial time.
Theorem 1.2 (Cunningham and Geelen [8]). (IP) with non-negative m × n matrix A given
together with a branch decomposition of its column matroid of width k, is solvable in time O((d +
1)2k mn + m2 n), where d = max{b1 , . . . , bm }.
1
Upper Bounds
O(nd)m+1 [26]
(non-negative matrix A)
O((d + 1)pw+1 mn + m2 n) (Theorem 1.3)
(non-negative matrix A)
Lower bounds
m )
o( log
m do(m)
no n
algorithm under ETH (Theorem 1.1)
(even for non-negative matrix A)
no f (pw)(d + 1)(1−)pw (mn)O(1) algorithm under SETH (Theorem 1.4)
(even for non-negative matrix A)
no f (d)(d + 1)(1−)pw (mn)O(1) algorithm under SETH (Theorem 1.5)
(even for non-negative matrix A)
Figure 1: A summary of our lower bound results in comparison to the known/new upper bound
results for non-negative A. Here, n and m are the number of variables and constraints respectively,
pw denotes the path-width of the column matroid of A and d denotes a bound on the largest entry
in b.
In fact, they also show that the assumption of non-negativity is unavoidable (without any further
assumptions such as a bounded domain for the variables) in this setting because (IP) is NP-hard
when the constraint matrix A is allowed to have negative values (in fact even when restricted to
{−1, 0, 1}) and the branchwidth of the column matroid of A is at most 3. A close inspection of the
instances they construct in their NP-hardness reduction shows that the column matroids of the
resulting constraint matrices are in fact direct sums of circuits, implying that even their path-width
is bounded by 3. The parameter path-width is closely related to the notion of trellis-width of a
linear code, which is a parameter commonly used in coding theory [20]. For a matrix A ∈ Rm×n ,
computing the path-width of the column matroid of A is equivalent to computing the trellis-width
of the linear code generated by A. Roughly speaking, the path-width of the column matroid of A is
at most k, if there is a permutation of the columns of A such that in the matrix A0 obtained from
A by applying this column-permutation, for every 1 ≤ i ≤ n − 1, the dimension of the subspace of
Rm obtained by taking the intersection of the subspace of Rm spanned by the first i columns and
the subspace of Rm spanned by the remaining columns, is at most k − 1.
The value of the parameter path-width is always at least the value of branch-width and at
most rank+1. As a result, any upper bounds on the complexity of (IP) in terms of path-width will
translate to upper bounds in terms of the larger parameter rank (number of constraints) and any
lower bounds on the complexity of (IP) in terms of path-width will translate to lower bounds in
terms of the smaller parameter branch-width. Motivated by this fact, we study the question of
designing an ‘optimal’ pseudo-polynomial time algorithm for (IP) when the column matroid of A
has constant path-width. We first obtain the following upper bound.
Theorem 1.3. (IP) with non-negative m×n matrix A given together with a path decomposition of its
column matroid of width k is solvable in time O((d + 1)k+1 mn + m2 n), where d = max{b1 , . . . , bm }.
As mentioned earlier, the NP-hardness of (IP) on constant branch-width instances also holds for
constant path-width instances and hence the assumption of non-negativity is unavoidable here as
well. Furthermore, while the proof of this theorem is not hard and is in fact almost identical to
the proof of Theorem 1.2, this upper bound becomes really interesting when placed in context and
compared to the tight lower bounds we provide in our next two theorems, which form the main
technical part of the paper. In these theorems, we provide tight conditional (subject to Strong ETH)
lower bounds for (IP) matching the running time of the algorithm of Theorem 1.3 (see Figure 1).
Strong ETH (SETH) is the conjecture that CNF-SAT cannot be solved in time (2 − )n mO(1) on
n-variable m-clause formulas for any constant . Both ETH and SETH were first introduced in
2
the work of Impagliazzo and Paturi [21], which built upon earlier work of Impagliazzo, Paturi and
Zane [22]. We obtain the following lower bounds for (IP). The first result shows that we cannot
relax the (d + 1)k factor in Theorem 1.3 even if we allow in the running time, an arbitrary function
depending on k. The second result shows a similar lower bound in terms of d instead of k. Put
together the results imply that no matter how much one is allowed to compromise on either the
path-width or the bound on d, it is unlikely that the algorithm of Theorem 1.3 can be improved.
Theorem 1.4. Unless SETH fails, (IP) with even a non-negative m × n constraint matrix A cannot
be solved in time f (k)(d+1)(1−)k (mn)O(1) for any function f and > 0, where d = max{b1 , . . . , bm }
and k is the path-width of the column matroid of A.
Theorem 1.5. Unless SETH fails, (IP) with even a non-negative m × n constraint matrix A cannot
be solved in time f (d)(d+1)(1−)k (mn)O(1) for any function f and > 0, where d = max{b1 , . . . , bm }
and k is the path-width of the column matroid of A.
Although the proofs of both lower bounds have a similar structure, we believe that there are
sufficiently many differences in the proofs to warrant stating them separately. Finally, since the
branch-width of a matroid never exceeds its path-width, our lower bounds hold when the parameter
of interest is chosen to be the branch-width of the column matroid of A as well. That is, under
SETH, there is no f (bw)(d + 1)(1−)bw (mn)O(1) or f (d)(d + 1)(1−)bw (mn)O(1) algorithm for (IP)
with non-negative constraint matrices (where bw denotes the branchwidth of the column matroid of
A), almost matching the upper bound of O((d + 1)2bw mn + m2 n) from Theorem 1.4.
Related work. Currently ETH is a commonly accepted conjecture and it serves as the basic
tool used for establishing asymptotically optimal lower bounds for various parameterized and
exact exponential algorithms. While there is no such consensus on SETH, the hypothesis has
already played a crucial role in the recent spectacular and rapid development in the analyses of
polynomial, parameterized and exact exponential algorithms. In particular, SETH was used to
establish conditional tight lower bounds for a number of fundamental computational problems,
including k-domination [27], the diameter of sparse graphs [29], dynamic connectivity problems [2],
the Frechet distance computation [5], string editing distance [3], dynamic programming on graphs of
bounded tree-width and clique-width [24, 12, 15, 9], Steiner tree and subset sum [10], finding the
longest common subsequence and the dynamic time warping distance [1, 6], and matching regular
expressions [4]. For further overview of applications of ETH and SETH, we refer to surveys [25, 31]
as well as [11, Chapter 14]. Our work extends this line of research by adding the fundamental (IP)
problem to the classes of “SETH-hard” and “ETH-hard” problems.
Organization of the paper. The remaining part of the paper is organized as follows. The
main technical part of the paper is devoted to proving Theorem 1.4 and Theorem 1.5. Therefore,
once we have set up the requisite preliminary definitions, we begin with Section 3 where we prove
Theorem 1.4. The first part of this section contains an overview of both reductions and could
be helpful for the reader in navigating the paper. We then prove Theorem 1.5 in Section 4 and
Theorem 1.3 in Section 5 (completing the results for constant path-width) and that of Theorem 1.1
in Section 6.
2
Preliminaries
We assume that the reader is familiar with basic definitions from linear algebra, matroid theory and
graph theory.
3
Notations. We use Z≥0 and R to denote the set of non negative integers and real numbers,
respectively. For any positive integer n, we use [n] and Zn to denotes the sets {1, . . . , n} and
{0, 1, . . . , n − 1}, respectively. For convenience, we say that [0] = ∅. For any two vectors b, b0 ∈ Rm
and i ∈ [m], we use b[i] to denote the ith coordinate of b and we write b0 ≤ b, if b0 [i] ≤ b[i] for all
i ∈ [m]. We often use 0 to denote the zero-vector whose length will be clear from the context.
For a matrix A ∈ Rm×n , I ⊆ [m] and J ⊆ [n], A[I, J] denote the submatrix of A obtained by the
restriction of A to the rows indexed by I and columns indexed by J. For an m × n matrix A and
P
n-vector v, we can write Av = ni=1 Ai v[i], where Ai is the ith column of A. Here we say that v[i]
is a multiplier of column Ai .
Branch-width of matroids. The notion of the branch-width of graphs, and implicitly of matroids,
was introduced by Robertson and Seymour in [28]. Let M = (U, F) be a matroid with universe set
U and family F of independent sets over U . We use rM to denote the rank function of M . That is,
for any S ⊆ U , rM (S) = maxS 0 ⊆S,S 0 ∈F |S 0 |. For X ⊆ U , the connectivity function of M is defined as
λM (X) = rM (X) + rM (U \ X) − rM (U ) + 1
For matrix A ∈ Rm×n , we use M (A) to denote the column-matroid of A. In this case the
connectivity function λM (A) has the following interpretation. For E = {1, . . . , n} and X ⊆ E, we
define
S(A, X) = span(A|X) ∩ span(A|E \ X),
where A|X is the set of columns of A restricted to X and span(A|X) is the subspace of Rm spanned
by the columns A|X. It is easy to see that the dimension of S(A, X) is equal to λM (A) (X) − 1.
A tree is cubic if its internal vertices all have degree 3. A branch decomposition of matroid M
with universe set U is a cubic tree T and mapping µ which maps elements of U to leaves of T . Let
e be an edge of T . Then the forest T − e consists of two connected components T1 and T2 . Thus
every edge e of T corresponds to the partitioning of U into two sets Xe and U \ Xe such that µ(Xe )
are the leaves of T1 and µ(U \ Xe ) are the leaves of T2 . The width of edge e is λM (Xe ) and the
width of branch decomposition (T, µ) is the maximum edge width, where maximum is taken over all
edges of T . Finally, the branch-width of M is the minimum width taken over all possible branch
decompositions of M .
The path-width of a matroid is defined as follows. Let us remind that a caterpillar is a tree
which is obtained from a path by attaching to some vertices of the paths some leaves. Then the
path-width of a matroid is the minimum width of a branch decomposition (T, µ), where T is a
cubic caterpillar. Let us note that every mapping of elements of a matroid to the leaves of a
cubic caterpillar, correspond to their ordering. Jeong, Kim, and Oum [23] gave a constructive
fixed-parameter tractable algorithm to construct a path decomposition of width at most k for a
column matroid of a given matrix.
ETH and SETH. For q ≥ 3, let δq be the infimum of the set of constants c for which there exists an
algorithm solving q-SAT with n variables and m clauses in time 2cn · mO(1) . The Exponential-Time
Hypothesis (ETH) and Strong Exponential-Time Hypothesis (SETH) are then formally defined as
follows. ETH conjectures that δ3 > 0 and SETH that limq→∞ δq = 1.
3
Proof of Theorem 1.4
In this section we prove that unless SETH fails, (IP) with non-negative matrix A cannot be solved
in time f (k)(d + 1)(1−)k (mn)O(1) for any function f and > 0, where d = max{b[1], . . . , b[m]} and
k is the path-width of the column matroid of A. In Subsection 3.1, we give an overview of our
reductions and in Subsection 3.2 we give a detailed proof of Theorem 1.4
4
3.1
Overview of our reductions
We prove Theorems 1.4 and 1.5 by giving reductions from CNF-SAT, where the parameters in
the reduced instances are required to obey certain strict conditions. For example, the reduction we
give to prove Theorem 1.4 must output an instance of (IP), where the path-width of the column
matroid M (A) of the constraint matrix A is a constant. Similarly, in the reduction used to prove
Theorem 1.5, we need to construct an instance of (IP), where the largest entry in the target
vector is upper bounded by a constant. These stringent requirements on the parameters make
the SETH-based reductions quite challenging. However, reductions under SETH can take super
polynomial time—they can even take 2(1−)n time for some > 0, where n is the number of variables
in the instance of CNF-SAT. This freedom to avail exponential time in SETH-based reductions is
used crucially in the proofs of Theorems 1.4 and 1.5.
Now we give an overview of the reduction used to prove Theorem 1.4. Let ψ be an instance of
CNF-SAT with n variables and m clauses. Given ψ and a fixed constant c ≥ 2, we construct an
instance A(ψ,c) x = b(ψ,c) , x ≥ 0 of (IP) satisfying certain properties. Since for every c ≥ 2, we have
a different A(ψ,c) and b(ψ,c) , this can be viewed as a family of instances of (IP). In particular our
main technical lemma is the following.
Lemma 3.1. Let ψ be an instance of CNF-SAT with n variables and m clauses. Let c ≥ 2 be a
n
fixed integer. Then, in time O(m2 2 c ), we can construct an instance A(ψ,c) x = b(ψ,c) , x ≥ 0, of (IP)
with the following properties.
(a.) ψ is satisfiable if and only if A(ψ,c) x = b(ψ,c) , x ≥ 0 is feasible.
n
(b.) The matrix A(ψ,c) is non-negative and has dimension O(m) × O(m2 c ).
(c.) The path-width of the column matroid of A(ψ,c) is at most c + 4.
n
(d.) The largest entry in b(ψ,c) is at most 2d c e − 1.
Once we have Lemma 3.1, the proof of Theorem 1.4 follows from the following observation: if
we have an algorithm A solving (IP) in time f (k)(d + 1)(1−)k (mn)a for some , a > 0, then we can
use this algorithm to refute SETH. In particular, given an instance ψ of CNF-SAT, we choose an
appropriate c depending only on and a, construct an instance A(ψ,c) x = b(ψ,c) , x ≥ 0, of (IP), and
run A on it. Our careful choice of c will imply a faster algorithm for CNF-SAT, refuting SETH.
More formally, we choose c to be an integer such that (1 − ) + 4(1−)
+ ac < 1. Then the total
c
running time to test whether ψ is satisfiable, is the time require to construct A(ψ,c) x = b(ψ,c) , x ≥ 0
plus the time required by A to solve the constructed instance of (IP). That is, the time required to
test whether ψ is satisfiable is
n
n
O(m2 2 c ) + f (c + 4)2 c (1−)(c+4) 2
a·n
c
mO(1) = 2
(1−)+
4(1−)
+ ac
c
n
0
mO(1) = 2 n mO(1) ,
where 0 < 1 is a constant depending on the choice of c. It is important to note that the utility of the
reduction described in Lemma 3.1 is extremely sensitive to the value of the numerical parameters
involved. In particular, even when the path-width blows up slightly, say up to δc, or when the
n
largest entry in b(ψ,c) blows up slightly, say up to 2δ c , for some δ > 1, then the calculation above will
not give us the desired refutation of SETH. Thus, the challenging part of the reduction described in
Lemma 3.1 is making it work under these strict restrictions on the relevant parameters.
As stated in Lemma 3.1, in our reduction, we need to obtain a constraint matrix with small
path-width. An important first step towards this is understanding what a matrix of small path-width
looks like. We first give an intuitive description of the structure of such matrices. Let A be a m × n
matrix of small path-width and let M (A) be the column matroid of A. For any i ∈ {1, . . . , n − 1},
5
1 1
1 1
1 1
..
.
1 1
B1
B2
..
.
Bm−1
Bm
(a) A matrix B for which path-width of its column
matroid is 1
(b) A pictorial representation of the matrix A(ψ,c) .
Figure 2: Comparison of A(ψ,c) with a low path-width matrix.
let A|{1, . . . i} denote the set of columns (or vectors) in A whose index is at most i (that is, the first
i columns) and let A|{i + 1, . . . n} denote the set of columns with index strictly greater than i. The
path-width of M (A) is at most
max dimhspan(A|{1, . . . , i}) ∩ span(A|{i + 1, . . . , n})i + 1.
i
Hence, in order to obtain a bound on the pathwidth, it is sufficient to bound dimhspan(A|{1, . . . , i})∩
span(A|{i + 1, . . . , n})i for every i ∈ [n]. Consider for example, the matrix B given in Figure 2a.
The path-width of M (B) is clearly at most 1. In our reduced instance, the constructed constraint
matrix A(ψ,c) will be an appropriate extension of B. That is A(ψ,c) will have the “same form” as B
but with each 1 replaced by a submatrix of order O(c) × n0 for some n0 . See Fig. 2b for a pictorial
representation of A(ψ,c) .
The construction used in Lemma 3.1 takes as input an instance ψ of CNF-SAT with n
variables and a fixed integer c ≥ 2, and outputs an instance A(ψ,c) x = b(ψ,c) , x ≥ 0, of (IP), that
satisfies all four properties of the lemma. Let X denote the set of variables in the input CNFformula ψ = C1 ∧ C2 ∧ . . . ∧ Cm . For the purposes of the present discussion we assume that c
divides n. We partition the variable set X into c blocks X0 , . . . , Xc−1 , each of size nc . Let Xi ,
i ∈ {0, . . . , c − 1}, denote the set of assignments of variables corresponding to Xi . Set ` = nc and
n
L = 2` . Clearly, the size of Xi is upper bounded by 2 c = 2` = L. We denote the assignments in Xi
by φ0 (Xi ), φ1 (Xi ), . . . , φL−1 (Xi ). To construct the matrix A(ψ,c) , we view “each of these assignments
as a different assignment for each clause”. In other words we have separate sets of column vectors in
the constraint matrix A(ψ,c) corresponding to different pairs (Cr , Xi ), where Cr is a clause and Xi
is a block in the partition of X. All the values set in these columns are based on the assignments
of Xi and the clause Cr . That is, based on the clause Cr and assignments in Xi . In total we have
2L columns corresponding to (Cr , Xi ). The set of columns corresponding to Cr , that is, the set of
columns corresponding to (Cr , Xi ), for all i, together forms a bigger block of columns, denoted by
Cr
r
AC
(ψ,c) , in A(ψ,c) . The columns of A(ψ,c) appears consecutively in A(ψ,c) . In other words, the clauses
r
of ψ partition the set of columns of A(ψ,c) into AC
(ψ,c) , r ∈ {1, . . . , m}, where columns in each of
Cr
r
the parts (A(ψ,c)
) occur consecutively. Thus, we can identify each column in the matrix AC
(ψ,c)
with a pair (Cr , φj (Xi )), i ∈ {0, . . . , c − 1} and j ∈ {0, . . . , L − 1}. For a pair (Cr , φj (Xi )), we refer
r
to φj (Xi ) as the assignment part of the pair. The non-zero values in AC
(ψ,c) are covered by 4c + 1
specific consecutive rows. These 4c + 1 rows are divided into 3 parts according to their roles in the
6
reduction:
• the first 2c rows comprise the predecessor matching part,
• the middle row is called the evaluation part, and
• the 2c rows after the evaluation part comprise the successor matching part.
The entries in the row corresponding to the evaluation part get values 0 or 1 depending on whether
the assignment part of the pair (Cr , φj (Xi )) satisfies Cr or not.
The matrix A(ψ,c) and the target vector b(ψ,c) are constructed in such a way that all the feasible
0
solutions to A(ψ,c) x = b(ψ,c) are from the set {0, 1}n , where n0 is the set of columns in A(ψ,c) . Hence,
setting a coordinate of x to 1 corresponds to choosing a particular column from A(ψ,c) . In our
reduction we use a selector gadget to enforce that any feasible solution will choose exactly one column
from the set of columns corresponding to a pair (Cr , Xi ). That is, it corresponds to choosing a
column identified by (Cr , φj (Xi )). Thus it results in choosing an assignment φj (Xi ) to the variables
in the set Xi . Note that this implies that we will choose an assignment in Xi for each clause Cr .
That way we might choose m assignments from Xi corresponding to m different clauses. However,
for the backward direction of the proof, it is important that we choose the same assignment from Xi
for each clause. This will ensure that we have selected an assignment to the variables in Xi . Towards
this we assign values in each of the columns in a way that all the assignments chosen by a feasible
solution for a particular block across different clauses are the same. Then choosing two columns, one
from the set of columns corresponding to (Cr , Xi ) and the other from the columns of (Cr0 , Xi ) in a
feasible solution would imply that both of these columns correspond to one particular assignment of
Xi . In this case, we say that these two columns are consistent. We enforce these consistencies in a
sequential manner. That is, for any block Xi , we make sure that the two columns chosen among
the columns corresponding to (Cr , Xi ) and (Cr+1 , Xi ) are consistent for any r ∈ {1, . . . , m − 1},
as opposed to checking the consistency for every pair (Cr , Xi ) and (Cr0 , Xi ) for r 6= r0 . Thus in
some sense these consistencies propagate. Such a propagation of consistencies is realized through
rows corresponding to the predecessor matching part and the successor matching part. For that
r
the rows corresponding to predecessor matching part of AC
(ψ,c) will be the same as the successor
C
r−1
r
matching part of A(ψ,c)
and the rows corresponding to the successor matching part of AC
(ψ,c) will
C
r+1
be the same as the predecessor matching part of A(ψ,c)
. Both the predecessor matching part as
well as the successor matching part contain designated rows for each block Xi of variables to handle
consistencies between (Cr , Xi ) and (Cr+1 , Xi ). Recall that Xi denotes the set of assignments of Xi
and |Xi | = 2` = L. Furthermore, assignments in Xi are denoted by φ0 (Xi ), . . . , φL−1 (Xi ). Thus, we
can identify the assignment φj (Xi ) by a non-negative integer j ≤ L − 1. These values are assigned
in a co-ordinated manner at designated places in the predecessor matching part as well as in the
successor matching part, enabling us to argue consistency. The largest entry in b(ψ,c) is upper
bounded by L − 1. Furthermore, the idea of making consistency in a sequential manner also allows
us to bound the path-width of column matroid of A(ψ,c) by c + 4.
The proof technique for Theorem 1.5 is similar to that for Theorem 1.4. This is achieved by
modifying the matrix A(ψ,c) constructed in the reduction described for Lemma 3.1. The largest
n
entry in A(ψ,c) is 2 c − 1. So each of these values can be represented by a binary string of length at
most ` = nc . We remove each row, say row indexed by γ, with entries greater than 1 and replace
it with nc rows, γ1 , . . . , γ` . Where, for any j, if the value A(ψ,c) [γ, j] = W then A(ψ,c) [γk , j] = ηk ,
where ηk is the k th bit in the `-sized binary representation of W . This modification reduces the
largest entry in A(ψ,c) to 1 and increases the path-width from constant to approximately n. Finally,
we set all the entries in b(ψ,c) to be 1. This concludes the overview of our reductions and we now
proceed to a detailed exposition.
7
3.2
Detailed Proof of Theorem 1.4
Towards the proof of Theorem 1.4, we first present the proof of our main technical lemma (Lemma 3.1),
which we restate here for the sake of completeness.
Lemma 3.1. Let ψ be an instance of CNF-SAT with n variables and m clauses. Let c ≥ 2 be a
n
fixed integer. Then, in time O(m2 2 c ), we can construct an instance A(ψ,c) x = b(ψ,c) , x ≥ 0, of (IP)
with the following properties.
(a.) ψ is satisfiable if and only if A(ψ,c) x = b(ψ,c) , x ≥ 0 is feasible.
n
(b.) The matrix A(ψ,c) is non-negative and has dimension O(m) × O(m2 c ).
(c.) The path-width of the column matroid of A(ψ,c) is at most c + 4.
n
(d.) The largest entry in b(ψ,c) is at most 2d c e − 1.
Let ψ = C1 ∧ C2 ∧ . . . ∧ Cm be an instance of CNF-SAT with variable set X = {x1 , x2 , . . . , xn }
and let c ≥ 2 be a fixed constant given in the statement of Lemma 3.1. We construct the instance
A(ψ,c) x = b(ψ,c) , x ≥ 0 of (IP) as follows.
Construction. Let C = {C1 , . . . , Cm }. Without loss of generality, we assume that n is divisible
by c, otherwise we add at most c dummy variables to X such that |X| is divisible by c. We divide
X into c blocks X0 , X1 , . . . , Xc−1 . That is Xi = {x i·n +1 , x i·n +2 , . . . , x (i+1)·n } for each i ∈ Zc . Let
c
c
c
` = nc and L = 2` . For each block Xi , there are exactly 2` assignments. We denote these assignments
by φ0 (Xi ), φ1 (Xi ), . . . , φL−1 (Xi ).
Now we create m matrices, one for each clause C ∈ C. These matrices will be submatrices of
the constraint matrix A(ψ,c) . For each clause Cr ∈ C, we create a (4c + 1) × c · 2`+1 matrix Br . For
each block Xi and all possible assignments to the variables of Xi , we allocate 2`+1 columns in Br .
For each assignment φj (Xi ) there are two columns in Br corresponding to it. Then the first 2`+1
columns of Br correspond to assignments of X0 , the second 2`+1 columns correspond to assignments
of X1 , etc.
Matrices Br for 1 < r < m. We first define Br for indices 1 < r < m. Matrices B1 and Bm have a
slightly different structure compared to the other matrices and so we define them separately. The
non-zero values of Br are defined as follows. Each assignment φj (Xi ) is identified by the number j.
Each φj (Xi ) defines 8 entries in Br : four in the column numbered 2`+1 i + 2j + 1 and four in the
column numbered 2`+1 i + 2j + 2. The rows of Br are partitioned into 3 parts. The part composed
of the first 2c rows is called the predecessor matching part, the part composed of the row indexed by
2c + 1 is called the evaluation part, and the part composed of the last 2c rows is called the successor
matching part, see Fig. 3a.
The predecessor matching part is defined by
Br [2i + 1, i · 2`+1 + 2j + 1] = Br [2i + 1, i · 2`+1 + 2j + 2] = j, i ∈ Zc .
(1)
For i ∈ Zc , the evaluation part is defined by
Br [2c + 1, i · 2`+1 + 2j + 1] = 0,
and
(
Br [2c + 1, i · 2
`+1
+ 2j + 2] =
8
1, if φj (Xi ) satisfies Cr ,
0, otherwise.
(2)
(3)
predecessor matching part
evaluation part
successor matching part
evaluation part
successor matching part
(a) Parts of Br for 1 < r ≤ m.
(b) Parts of B1 .
Figure 3: Parts of Bm .
The successor matching part for Br is defined for i ∈ Zc as
Br [2c + 2i + 2, i · 2`+1 + 2j + 1] = Br [2c + 2i + 2, i · 2`+1 + 2j + 2] = L − 1 − j,
(4)
Br [2c + 2i + 3, i · 2
(5)
`+1
+ 2j + 1] = Br [2c + 2i + 3, i · 2
`+1
+ 2j + 2] = 1.
All other entries in Br , which are not defined above, are set to zero. That is, for all i, i0 ∈ Zc
and a ∈ [2`+1 ] such that i 6= i0 ,
Br [2i + 1, i0 · 2`+1 + a] = Br [2i + 2, i0 · 2`+1 + a] = 0,
(6)
Br [2i + 2, i · 2
(7)
`+1
0
+ a] = 0, and
Br [2c + 2i + 2, i · 2
`+1
0
+ a] = Br [2c + 2i + 3, i · 2
`+1
+ a] = 0.
(8)
Before describing the construction of B1 and Bm , we provide a brief description of certain
desirable properties possessed by Br . We have designated set of columns per pair (Cr , Xi ), which
are indexed by [(i + 1)2`+1 ] \ [i · 2`+1 ] and (5) ensures that at most one of the columns from this set
is chosen by a feasible solution. This will be forced by putting 1 in the corresponding coordinate
of vector b(ψ,c) . In the construction of A(ψ,c) , we will only add zeros to the entries in the row
of A(ψ,c) corresponding to the (2c + 2i + 3)th row of Br , but outside the submatrix Br of A(ψ,c) .
This guarantees that exactly one of them is chosen by a feasible solution. The purpose of (1) is
to ensure consistency with the column selected from (Cr−1 , Xi ), and purpose of (4) is to ensures
consistency with the column selected from (Cr+1 , Xi ). We construct the matrix A(ψ,c) in such a way
that the row of Br indexed by 2i + 1 and the row of Br−1 indexed by 2c + 2i + 2 are equal in A(ψ,c) .
Suppose that this row is indexed by h in A(ψ,c) . Then (1) and (4) ensure that if we choose consistent
columns from the columns of (Cr−1 , Xi ) and (Cr , Xi ), then the sum of the values in coordinate h of
the selected columns will be equal to L − 1. So we will set b(ψ,c) [h] = L − 1 in the target vector
b(ψ,c) . For each assignment φj (Xi ), we have two designated columns in Br , they are indexed by
i · 2`+1 + 2j + 1 and i · 2`+1 + 2j + 2. The reason for creating two columns instead of just one per
φj (Xi ) is the following. The coordinate j 0 of the target vector b(ψ,c) corresponding to the row which
contains the row of Br indexed by 2c + 1 will be set to 1. For any satisfying assignment of φ of
ψ, more than one partial assignments of φ (i.e, assignments of φ restricted to different blocks of
X) may satisfy the clause Cr . So among the pairs of columns corresponding to these satisfying
partial assignments a feasible solution will choose the first column from the pair for all but one.
For a partial assignment (assignment of a block of X) which satisfies Cr , the feasible solution will
choose the second column corresponding to it. Equations (2) and (3) make sure that the entries
corresponding to the coordinate j 0 from the set of chosen columns by a feasible solution will add
up to 1; hence at least one selected column would correspond to an assignment (of a block of X)
satisfying clause Cr .
9
0
0
0
0
0
3
1
0
0
0
0
0
0
1
3
1
0
0
1
0
0
0
0
2
1
0
0
1
0
0
0
0
2
1
0
0
2
0
0
0
0
1
1
0
0
2
0
0
0
1
1
1
0
0
3
0
0
0
0
0
1
0
0
3
0
0
0
1
0
1
0
0
?
?
?
?
?
?
0
0
0
0
0
0
0
3
1
0
0
0
0
0
0
0
3
1
0
0
1
0
0
0
0
2
1
?
?
?
?
?
?
?
?
?
?
?
?
?
0
0
2
0
0
0
0
1
1
0
0
2
0
0
0
0
1
1
0
0
3
0
0
0
0
0
1
0
0
3
0
1
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
1
1
0
1
0
0
0
0
1
0
1
0
0
0
0
1
0
2
0
0
0
0
1
0
2
0
0
0
1
1
0
3
0
0
0
0
1
0
3
0
0
0
1
1
0
?
?
?
?
?
?
0
0
0
0
0
0
1
0
0
0
0
0
0
1
0
0
1
0
0
0
1
0
0
1
0
1
0
1
0
0
2
0
0
0
1
0
0
2
0
0
0
1
0
0
3
0
0
0
1
0
0
3
0
1
0
1
?
?
?
?
0
0
1
0
1
0
0
2
1
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
Figure 4: Let n = 4, c = 2, ` = 2 and Cr = x1 ∨ x2 ∨ x4 . The assignments are φ0 (X0 ) = {x1 =
x2 = 0}, φ1 (X0 ) = {x1 = 0, x2 = 1}, φ2 (X0 ) = {x1 = 1, x2 = 0}, φ3 (X0 ) = {x1 = x2 = 1}, φ0 (X1 ) =
{x3 = x4 = 0}, φ1 (X1 ) = {x3 = 0, x4 = 1}, φ2 (X1 ) = {x3 = 1, x4 = 0}, φ3 (X1 ) = {x3 = x4 = 1}.
The entries defined according to φ1 (X0 ) and φ3 (X1 ) are colored red and blue respectively. If
1 < r < m, then the matrix on the left represents Br and if r = 1, then Br can be obtained by
deleting the yellow colored portion from the left matrix. The matrix on the right represents Bm .
Sometimes, it is helpful to focus on the positions containing non-zero elements. This can be found
in the two matrices at the bottom of the figure.
Matrices B1 and Bm . The matrix B1 is created as above but with the exception that we remove
the predecessor matching part (see Fig. 3b). The matrix Bm is created as above with the exception
that we remove the rows numbered 2c + 2, 2c + 4, . . . , 4c. An illustration of Bm is given in Fig. 4.
Formally, the entries of B1 and Bm , defined by φj (Xi ) are given below.
For B1 we define its entries as
B1 [2i + 2, i · 2`+1 + 2j + 1] = B1 [2i + 2, i · 2`+1 + 2j + 2] = L − j − 1,
B1 [2i + 3, i · 2
B1 [1, i · 2
`+1
`+1
+ 2j + 1] = B1 [2i + 3, i · 2
`+1
+ 2j + 2] = 1,
+ 2j + 1] = 0, and
(
B1 [1, i · 2`+1 + 2j + 2] =
(9)
(10)
(11)
1 if φj (Xi ) satisfies C1 ,
0 otherwise.
(12)
For Bm ,
Bm [2i + 1, i · 2`+1 + 2j + 1] = Bm [2i + 1, i · 2`+1 + 2j + 2] = j,
Bm [2c + i + 2, i · 2`+1 + 2j + 1] = Bm [2c + i + 2, i · 2`+1 + 2j + 2] = 1,
Bm [2c + 1, i · 2
`+1
Bm [2c + 1, i · 2
`+1
(13)
+ 2j + 1] = 0, and
(
+ 2j + 2] =
1, if φj (Xi ) satisfies Cm ,
0, otherwise.
All other entries in B1 and Bm , which are not defined above, are set to zero. That is, for all
10
i, i0 ∈ Zc and a ∈ [2`+1 ] such that i 6= i0 ,
B1 [2i + 2, i0 · 2`+1 + a] = B1 [2i + 3, i0 · 2`+1 + a] = 0,
0
0
Bm [2i + 1, i · 2
+ a] = Bm [2i + 2, i · 2
`+1
Bm [2i + 2, i · 2
`+1
`+1
+ a] = 0,
+ a] = 0,
0
Bm [2c + i + 2, i · 2
`+1
(14)
(15)
(16)
+ a] = 0.
(17)
Matrix A(ψ,c) and vector b(ψ,c) . Now we explain how to construct the constraint matrix A(ψ,c) and
vector b(ψ,c) , which would serve as instance of (IP). In what follows,
we simplify the notation by using A instead of A(ψ,c) and b instead of b(ψ,c) .
The matrices B1 , . . . , Bm are disjoint submatrices of A and they cover all non zero entries of A.
Informally, the submatrices B1 , . . . , Bm form a chain such that the rows corresponding to the successor
matching part of Br will be the same as the rows in the predecessor matching part of Br+1 . A pictorial
representation of A can be found in Fig. 2b. Formally, A is a ((m − 1)(2c + 1) + (c + 1)) × (m · c · 2`+1 )
matrix. Let I1 = [2c + 1] and Im = [(m − 1)(2c + 1) + (c + 1)] \ [(m − 1)(2c + 1) − 2c]. For every
1 < r < m, let Ir = [r(2c+1)]\[(r−1)(2c+1)−2c], and for r ∈ [m], let Jr = [r·c·2`+1 ]\[(r−1)·c·2`+1 ].
Now for each r ∈ [m],we put matrix A[Ir , Jr ] := Br . All other entries of A not belonging to any of
the submatrices A[Ir , Jr ] are set to zero. This completes the construction of A.
Now we define the ((m − 1)(2c + 1) + (c + 1))-dimensional vector b. Let
P = {(r − 1)(2c + 1) + 2j + 2 | r ∈ [m − 1], j = Zc }.
In other words, P contains the indices of some rows: for each r ∈ [m − 1], alternating rows in
the successor matching part (and thus the alternating rows in the predecessor matching part) of
A[Ir , Jr ] belong to P (refer to Fig. 2b again). Then the entries of b are defined as
(
b[q] =
L − 1, if q ∈ P,
1,
otherwise.
This completes the construction of the matrix A and vector b which together make up the required
instance of (IP).
Correctness. Now we prove that ψ is satisfiable if and only if there is a non-negative integer vector
x such that Ax = b. We start with some notations. We partition the set of columns of A into m
parts J1 , . . . , Jm (we have already defined these sets) with one part per clause. For each r ∈ [m], Jr
is the set of columns associated with Cr . We further divide Jr into c equal parts, one per variable
set Xi . These parts are
Pr,i = {(r − 1)c · 2`+1 + i · 2`+1 + 1, . . . , (r − 1)c · 2`+1 + (i + 1) · 2`+1 }, i ∈ Zc .
In other words, Pr,i is the set of columns associated with the tuple (Cr , Xi ) and |Pr,i | = 2`+1 . The
set Pr,i is divided into 2` parts of size two each, one per tuple (Cr , φj (Xi )), where Cr ∈ C, j ∈ ZL and
i ∈ Zc . The two columns associated with tuple (Cr , φj (Xi )) are indexed by (r−1)2`+1 +i·2`+1 +2j +1
and (r − 1)2`+1 + i · 2`+1 + 2j + 2 in A. We also put n0 = m · c · 2`+1 to be the number of columns
in A.
0
Lemma 3.1. Formula ψ is satisfiable if and only if there exists x∗ ∈ Zn≥0 such that Ax∗ = b.
11
Proof. Suppose ψ is satisfiable and let φ be its satisfying assignment. There exists j0 , j1 , . . . , jc−1 ∈
ZL such that φ is the union of φj0 (X0 ), φj1 (X1 ), . . . , φjc−1 (Xc−1 ). Each clause C ∈ C, C is satisfied
by at least one of the assignments φj0 (X0 ), φj1 (X1 ), . . . , φjc−1 (Xc−1 ). For each C, we fix an arbitrary
i ∈ Zc such that the assignment φji (Xi ) satisfies clause C. Let α be a function which fixes these
assignments for each clause. That is, α : C → Zc such that the assignment φjα(C) (Xα(C) ) satisfies
0
the clause C for every C ∈ C. Now we define x∗ ∈ Zn≥0 and prove that Ax∗ = b. Let
Q1 = {(r − 1)2`+1 + i · 2`+1 + 2ji + 1 | r ∈ [m], i ∈ Zc , α(Cr ) 6= i},
Q2 = {(r − 1)2`+1 + i · 2`+1 + 2ji + 2 | r ∈ [m], i ∈ Zc , α(Cr ) = i}, and
Q = Q1 ∪ Q2 .
Then the vector x∗ is defined by setting
(
1, if q ∈ Q,
0, otherwise.
∗
x [q] =
The q-th entry in x∗ is the multiplier of column q and so we say that entry x∗ [q] corresponds to
column q. For each tuple (C, φji (Xi )), one entry of x∗ among the two entries corresponding to the
columns associated with (C, φji (Xi )) is set to 1. If α(C) = i, then the second column corresponding
to (C, φji (Xi )) is set to 1, otherwise the first column corresponding to (C, φji (Xi )) is set to 1. All
other entries are set to zero. Also note that for every r ∈ [m] and i ∈ Zc , we have that |Pr,i ∩ Q| = 1
and let {qr,i } = Pr,i ∩ Q. Here notice that among the 2`+1 columns of Pr,i , exactly one column,
which is indexed by qr,i , belongs to Q. The column qr,i corresponds to one of the two columns
corresponding to (Cr , φji (Xi )).
We need the following auxiliary claims.
Claim 3.2. For every r ∈ [m − 1] and i, i0 ∈ Zc such that i 6= i0 , we have
(a)
P
A[(r − 1)(2c + 1) + 2i + 2, z] · x[z] = L − 1 − ji ,
(b)
P
A[(r − 1)(2c + 1) + 2i + 3, z] · x[z] = 1,
(c)
P
A[(r − 1)(2c + 1) + 2i0 + h + 1, z] · x[z] = 0, for any h ∈ {1, 2} .
z∈Pr,i
z∈Pr,i
z∈Pr,i
Proof. First consider the case when r = 1. Let Pr,i ∩ Q = {i · 2`+1 + 2ji + g}, where g ∈ {1, 2}. Then
X
A[2i + 2, z] · x[z]
=
B1 [2i + 2, z]
X
z∈P1,i ∩Q
z∈P1,i
= B1 [2i + 2, i · 2`+1 + 2ji + g]
= L − ji − 1
(by (9))
and (a) follows.
To prove (b), we have
X
z∈P1,i
A[2i + 3, z] · x[z] =
X
B1 [2i + 3, z]
z∈P1,i ∩Q
= B1 [2i + 3, i · 2`+1 + 2ji + g]
= 1.
(By (10))
12
To show (c), observe that for h ∈ {1, 2},
X
A[2i0 + h + 1, z] · x[z] =
B1 [2i0 + h + 1, z]
X
z∈Pr,i ∩Q
z∈Pr,i
= B1 [2i0 + h + 1, i · 2`+1 + 2ji + g]
= 0.
(By (14))
Now consider the case when r > 1. Let Pr,i ∩ Q = {(r − 1)c · 2`+1 + i · 2`+1 + 2ji + g}, where
g ∈ {1, 2}. For this case we have
A[(r − 1)(2c + 1) + 2i + 2, z] · x[z] =
X
Br [2c + 2i + 2, z − ((r − 1)c · 2`+1 )]
X
z∈Pr,i ∩Q
z∈Pr,i
= Br [2c + 2i + 2, i · 2`+1 + 2ji + g]
= L − ji − 1.
(By (4))
A[(r − 1)(2c + 1) + 2i + 3, z] · x[z] = Br [2c + 2i + 3, i · 2`+1 + 2ji + g]
X
z∈Pr,i
= 1.
(By (5))
Finally, for h ∈ {1, 2},
X
A[(r − 1)(2c + 1) + 2i0 + h + 1, z] · x[z] =
Br [2c + 2i0 + h + 1, z − ((r − 1)c · 2`+1 )]
X
z∈Pr,i ∩Q
z∈Pr,i
= Br [2c + 2i0 + h + 1, i · 2`+1 + 2ji + g]
= 0.
(By (8))
Claim 3.3. For every r ∈ [m − 1] and i, i0 ∈ Zc such that i 6= i0 , we have
(a)
P
A[(r − 1)(2c + 1) + 2i + 2, z] · x[z] = ji ,
(b)
P
A[(r − 1)(2c + 1) + 2i + 3, z] · x[z] = 0,
(c)
P
A[(r − 1)(2c + 1) + 2i0 + h + 1, z] · x[z] = 0, where h ∈ {1, 2}.
z∈Pr+1,i
z∈Pr+1,i
z∈Pr+1,i
Proof. The proof of the claim is similar to the proof of Claim 3.2. Let Pr+1,i ∩ Q = {rc · 2`+1 + i ·
2`+1 + 2ji + g} where g ∈ {1, 2}. Then
A[(r − 1)(2c + 1) + 2i + 2, z] · x[z] = Br+1 [2i + 1, i · 2`+1 + 2ji + g]
X
z∈Pr+1,i
= ji
X
A[(r − 1)(2c + 1) + 2i + 3, z] · x[z] =
(By (1))
X
Br+1 [2i + 2, z − (rc · 2`+1 )]
z∈Pr+1,i ∩Q
z∈Pr+1,i
= Br+1 [2i + 2, i · 2`+1 + 2ji + g]
= 0
13
(By (7) or (16))
For any h ∈ {1, 2},
X
A[(r − 1)(2c + 1) + 2i0 + h + 1, z] · x[z] =
Br+1 [2i0 + h, z − (rc · 2`+1 )]
X
z∈Pr+1,i ∩Q
z∈Pr+1,i
= Br+1 [2i0 + h, i · 2`+1 + 2ji + g]
= 0
(By (6) or (15))
Now we show that Ax∗ = b. Recall that P = {(r − 1)(2c + 1) + 2j 0 + 2 | r ∈ [m − 1], j 0 ∈ Zc } .
Let m0 = (m − 1)(2c + 1) + c + 1, be the number of rows in A. To prove Ax∗ = b, we need to show
that,
(
n0
X
L − 1, if q ∈, P
A[q, j] · x[j] =
1,
otherwise.
j=1
We consider the following exhaustive cases.
Case 1: q ∈ P . Let q = (r − 1)(2c + 1) + 2i + 2 for some r ∈ [m − 1] and i ∈ Zc . Notice that
q ∈ Ir , q ∈ Ir+1 and q ∈
/ Ir0 for every r0 ∈ [m] \ {r, r + 1}. This implies that
for every j ∈
/ Jr ∪ Jr+1 , A[q, j] = 0.
(18)
Then
0
n
X
j=1
A[q, j] · x[j] =
X X
A[q, j] · x[j]
s∈[m] j∈Js
=
X
A[q, j] · x[j] +
j∈Jr
X
i0 ∈Zc
X
(By (18))
A[q, j] · x[j]
A[q, j] · x[j] +
j∈Pr,i0
=
X
j∈Jr+1
=
X
X
i0 ∈Z
j∈Pr+1,i0
c
A[q, j] · x[j] +
j∈Pr,i
X
A[q, j] · x[j]
X
A[q, j] · x[j]
(By Claims 3.2(c) and 3.3(c))
j∈Pr+1,i
= L − 1 − ji + ji
(By Claims 3.2(a) and 3.3(a))
= L − 1.
Case 2: q ∈ [m0 ] \ P . We partition [m0 ] \ P into R1 ] R2 ] R3 , and consider sub-cases based on
these parts. Let
R1 = (r − 1)(2c + 1) + 2j 0 + 3 | r ∈ [m − 1], j 0 ∈ Zc ,
R2 = {(r − 1)(2c + 1) + 1 | r ∈ [m]} ,
R3 = [(m − 1)(2c + 1) + 1 + c] \ [(m − 1)(2c + 1) + 1].
14
Case 2(a): q ∈ R1 . Let q = (r − 1)(2c + 1) + 2i + 3 for some r ∈ [m − 1] and i ∈ Zc . Notice
that q ∈ Ir , q ∈ Ir+1 and q ∈
/ Ir0 for any r0 ∈ [m] \ {r, r + 1}. This implies that,
for any j ∈
/ Jr ∪ Jr+1 , A[q, j] = 0.
(19)
Hence
0
n
X
A[q, j] · x[j] =
j=1
X X
A[q, j] · x[j]
s∈[m] j∈Js
=
A[q, j] · x[j] +
X
j∈Jr
X
j∈Jr+1
=
X
A[q, j] · x[j] +
X
i0 ∈Zc
j∈Pr,i0
=
(By (19))
A[q, j] · x[j]
X
i0 ∈Z
j∈Pr+1,i0
c
A[q, j] · x[j] +
X
X
j∈Pr,i
A[q, j] · x[j]
X
A[q, j] · x[j]
(By Claims 3.2(c) and 3.3(c))
j∈Pr+1,i
= 1+0
(By Claims 3.2(b) and 3.3(b))
= 1
Case 2(b): q ∈ R2 . Let q = (r − 1)(2c + 1) + 1 for some r ∈ [m]. By construction of A, we
have that for all j ∈
/ Jr , A[q, j] = 0. This implies that
0
n
X
A[q, j] · x[j] =
j=1
X
(20)
A[q, j] · x[j].
j∈Jr
We consider two cases based on r = 1 or r > 1. When r > 1,
A[{(r − 1)(2c + 1) + 1}, Jr ] = Br [{2c + 1}, Jr ].
Recall the function α. That is, if α(Cr ) = g, then φjg (Xg ) satisfies Cr . By Equation (20),
0
n
X
A[q, j] · x[j] =
j=1
`+1
c·2
X
Br [2c + 1, j] · x[j]
j=1
=
X
X
Br [2c + 1, i0 · 2`+1 + j] · x[i0 · 2`+1 + j]
i0 ∈Zc j∈[2`+1 ]
=
X
Br [2c + 1, α(Cr ) · 2`+1 + j] · x[α(Cr ) · 2`+1 + j]
j∈[2`+1 ]
+
X
X
Br [2c + 1, i0 · 2`+1 + j] · x[i0 · 2`+1 + j]
i0 ∈Zc \{α(Cr )} j∈2`+1
=
X
Br [2c + 1, α(Cr ) · 2`+1 + j] · x[α(Cr ) · 2`+1 + j] + 0
j∈[2`+1 ]
(By (2) and definition of Q)
= Br [2c + 1, α(Cr ) · 2
`+1
+ 2α(Cr ) + 2]
= 1.
(By (3) using the fact that φjα(Cr ) satisfies Cr )
15
When r = 1 we have that q = 1 and A[{1}, J1 ] = B1 [{1}, J1 ]. Hence, by Equation (20),
0
n
X
A[q, j] · x[j] =
j=1
`+1
c·2
X
B1 [1, j] · x[j]
j∈1
=
X
B1 [1, i0 · 2`+1 + j] · x[i0 · 2`+1 + j]
X
i0 ∈Zc j∈[2`+1 ]
=
X
B1 [1, α(C1 ) · 2`+1 + j] · x[α(C1 ) · 2`+1 + j]
j∈[2`+1 ]
+
X
X
B1 [1, i0 · 2`+1 + j] · x[i0 · 2`+1 + j]
i0 ∈Zc \{α(C1 )} j∈[2`+1 ]
=
B1 [1, α(C1 ) · 2`+1 + j] · x[α(C1 ) · 2`+1 + j] + 0
X
j∈[2`+1 ]
(By (11) and definition of Q)
= B1 [1, α(C1 ) · 2
`+1
+ 2α(C1 ) + 2]
= 1
(By (11) and definition of Q)
(By (12) using the fact that φjα(C1 ) satisfies C1 )
Case 2(c): q ∈ R3 . Let q = (m − 1)(2c + 1) + 1 + i where i ∈ [c]. By the definition of A. we
have that for every j ∈
/ {(m − 1) · c · 2`+1 + (i − 1)2`+1 + 1, . . . , (m − 1) · c · 2`+1 + i · 2`+1 }, A[q, j] = 0.
That is, for j ∈
/ Pm,i−1 , A[q, j] = 0. Let Pm,i−1 ∩ Q = {(m − 1)c · 2`+1 + (i − 1)2`+1 + 2ji−1 + g},
where g ∈ {1, 2}. Hence
0
n
X
A[q, j] · x[j] =
j=1
X
A[q, j] · x[j]
j∈Pm,i−1
=
X
Bm [2c + 1 + i, j − (m − 1)c2`+1 ]
j∈Pm,i−1 ∩Q
= Bm [2c + 1 + i, (i − 1)2`+1 + 2ji−1 + g]
= 1
(By (13))
Lemma 3.4. The path-width of the column matroid of A is at most c + 4
Proof. Recall that n0 = m · c · 2`+1 , be the number of columns in A and m0 be the number of rows in
A. To prove that the path-width of A is at most c + 4, it is sufficient to show that for all j ∈ [n0 − 1],
dimhspan(A|{1, . . . , j}) ∩ span(A|{j + 1, . . . , n0 })i ≤ c + 3.
(21)
The idea for proving Equation (21) is based on the following observation. For V 0 = A|{1, . . . , j}
and V 00 = A|{j + 1, . . . , n0 }, let
I = {q ∈ [m0 ] | there exists v 0 ∈ V 0 and v 00 ∈ V 00 such that v 0 [q] 6= v 00 [q] 6= 0}.
Then the dimension of span(V 0 ) ∩ span(V 00 ) is at most |I|. Thus to prove (21), for each j ∈ [n0 − 1],
we construct the corresponding set I and show that its cardinality is at most c + 3.
16
We proceed with the details. Let v1 , v2 , . . . , vn0 be the column vectors of A. Let j ∈ [n0 − 1]. Let
V1 = {v1 , . . . , vj } and V2 = {vj+1 , . . . , vn0 }. We need to show that dimhspan(V1 ) ∩ span(V2 )i ≤ c + 3.
Let
I 0 = {q ∈ [m0 ] | there exists v ∈ V1 and v 0 ∈ V2 such that v[q] 6= 0 6= v 0 [q]}.
We know that [n0 ] is partitioned into parts Pr0 ,i0 , r0 ∈ [m], i0 ∈ Zc .
Fix r ∈ [m] and i ∈ Zc such that j ∈ Pr,i .
Let j = (r − 1)c · 2`+1 + i · 2`+1 + 2j 0 + g, where j 0 ∈ ZL and g ∈ {1, 2}. Let q1 = max{0, (r −
1)(2c + 1) − 2c}, q2 = r(2c + 1), j1 = (r − 1) · c · 2`+1 , and j2 = r · c · 2`+1 Then [q2 ] \ [q1 ] = Ir and
[j2 ] \ [j1 ] = Jr (recall the definition of sets Ir and Jr from the construction of matrix A).
The way we constructed matrix A, for every q > q2 and for every vector v ∈ V1 , we have v[q] = 0.
Also, for every q ≤ q1 and for any v ∈ V2 , we have that v[q] = 0. This implies that I 0 ⊆ [q2 ]\[q1 ] = Ir .
Now we partition Ir into 5 parts: R1 , R10 , R, R2 , and R20 . These parts are defined as follows.
(
R1 =
(
R10
=
∅,
if r = 1,
{(r − 2)(2c + 1) + 2i0 + 2 | i0 ∈ Zc }, otherwise,
∅,
if r = 1,
{(r − 2)(2c + 1) + 2i0 + 3 | i0 ∈ Zc }, otherwise,
R = {(r − 1)(2c + 1) + 1},
(
∅,
if r = m,
{(r − 1)(2c + 1) + 2i0 + 2 | i0 ∈ Zc }, otherwise,
(
{(r − 1)(2c + 1) + i0 + 1 | i0 ∈ [c]}, if r = m,
{(r − 1)(2c + 1) + 2i0 + 3 | i0 ∈ Zc }, otherwise.
R2 =
(22)
and
R20
=
Claim 3.5. For each r ∈ [m], q ∈
/ Ir and j 00 ∈ Jr , vj 00 [q] = 0.
Proof. The non-zero entries in A are covered by the disjoint sub-matrices A[Ir , Jr ] = Br , r ∈ [m].
Hence the claim follows.
Claim 3.6. |I 0 ∩ R1 | ≤ c − (i − 1).
Proof. When r = 0, R1 = ∅ and the claim trivially follows. Let r > 1, and let q ∈ R1 be such that
q < (r−2)(2c+1)+2i+2. Then q = (r−2)(2c+1)+2i0 +2 for some i0 < i. Notice that q ∈
/ Ir0 for every
S
S
r0 > r. By Claim 3.5, for every v ∈ r0 >r Jr0 , v[q] = 0. Now consider the vector vj 00 ∈ V2 \ ( r0 >r Jr0 ).
Notice that j 00 > j and j 00 ∈ Jr . Let j 00 = j + a = (r − 1)c · 2`+1 + i · 2`+1 + 2j 0 + g + a for some
a ∈ [rc2`+1 − j]. From the construction of A, vj 00 [q] = Br [2i0 + 1, i · 2`+1 + 2j 0 + g + a] = 0, by (6).
Thus for every q ∈ R, q < (r − 2)(2c + 1) + 2i + 2 and v ∈ V2 , v[q] = 0. This implies that
|I 0 ∩ R1 | ≤ |{q ≥ (r − 2)(2c + 1) + 2i + 2} ∩ R1 | ≤ c − (i − 1).
Claim 3.7. |I 0 ∩ R10 | = 0.
17
Proof. When r = 1, R10 = ∅ and the claim holds. So, now assume that r > 1. Consider any q ∈ R10 .
Let i0 ∈ Zc be such that q = (r − 2)(2c + 1) + 2i0 + 3. Notice that q ∈
/ Ir0 for any r0 > r, and hence,
S
00
by Claim 3.5, for any v ∈ r0 >r Jr0 , v[q] = 0. Now consider any j ∈ Jr . Let j 00 = (r − 1)c · 2`+1 + a,
for some a ∈ [c · 2`+1 ]. From the construction of A, vj 00 [q] = Br [2i0 + 2, a] = 0, by (6) or (7). This
completes the proof of the claim.
Claim 3.8. |I 0 ∩ R2 | ≤ i.
Proof. When r = m, R2 = ∅ and the claim trivially holds. So, now let r < m and consider any
q ∈ R1 ∩ {q > (r − 1)(2c + 1) + 2i + 2}. Let i0 > i such that q = (r − 1)(2c + 1) + 2i0 + 2. Notice that
S
q∈
/ Ir0 for any r0 < r. Hence, by Claim 3.5, for any v ∈ r0 <r Jr0 , v[q] = 0. Now consider any vector
S
vj 00 ∈ V1 \ ( r0 <r Jr0 ). Notice that j 00 ≤ j and j 00 ∈ Jr . Let j 00 = (r − 1)c · 2`+1 + i00 · 2`+1 + a for some
a ∈ [2`+1 ] and i00 ≤ i < i0 . From the construction of A, vj 00 [q] = Br [2c + 2i0 + 2, i00 · 2`+1 + a] = 0, by
(8). Hence we have shown that for any q ∈ R, q > (r − 2)(2c + 1) + 2i + 2 and v ∈ V1 , v[q] = 0.
This implies that
|I 0 ∩ R2 | ≤ |{q ≤ (r − 1)(2c + 1) + 2i + 2} ∩ R1 | ≤ i
Claim 3.9. |I 0 ∩ R20 | ≤ 1.
Proof. Consider the case when r = m. Consider q ∈ R20 . We claim that if q ∈ I 0 ∩ R20 , then
q = (m − 1)(2c + 1) + i + 2. Suppose q ∈ I 0 ∩ R20 and q < (m − 1)(2c + 1) + i + 2. Let
q = (m − 1)(2c + 1) + i0 + 2, where 0 ≤ i0 < i. Then by the construction of A, for any j 00 > j,
vj 00 [q] = Bm [2c + i0 + 2, j 00 − (m − 1)c2`+1 ] = Bm [2c + i0 + 2, i1 2`+1 + a], where c − 1 ≥ i1 ≥ i and
a ∈ [2`+1 ]. Thus by (17), Vj 00 [q] = Bm [2c + i0 + 2, i1 2`+1 + a] = 0. This contradicts the assumption
that q ∈ I 0 ∩ R20 .
Suppose that q ∈ I 0 ∩ R20 and q > (m − 1)(2c + 1) + i + 2. Let q = (m − 1)(2c + 1) + i0 + 2,
where i < i0 ≤ c − 1. Then by the construction of A, for any j 00 ≤ j, vj 00 [q] = Bm [2c + i0 +
2, j 00 − (m − 1)c2`+1 ] = Bm [2c + i0 + 2, i1 2`+1 + a], where 0 ≤ i1 ≤ i, a ∈ [2`+1 ]. Thus by (17),
vj 00 [q] = Bm [2c + i0 + 2, i1 2`+1 + a] = 0. This contradicts the assumption that i ∈ I 0 ∩ R20 . Hence, in
this case, we have proved that |I 0 ∩ R20 | ≤ 1.
So, now assume that r ∈ [m − 1]. Consider any q ∈ R20 . Let i0 ∈ Zc such that q = (r − 1)(2c +
S
1) + 2i0 + 3. Notice that q ∈
/ Ir0 for any r0 < r, and hence, by Claim 3.5, for any v ∈ r0 <r Jr0 ,
S
v[q] = 0. Also notice that q ∈
/ Ir0 for any r0 > r + 1, and hence, by Claim 3.5, for any v ∈ r0 >r+1 Jr0 ,
v[q] = 0. So the only potential j 00 for which vj 00 [q] 6= 0, are from Jr ∪ Jr+1 . Let j 00 ∈ Jr+1 . Then by
the definition of A, vj 00 [q] = Br+1 [2i0 + 2, j 00 − rc2` ] = 0, by (6) or (7) or (15) or (16). Hence, we
conclude that the only possible j 00 for which vj 00 [q] 6= 0, are from Jr .
Now the proof is similar to case when r = m. We claim that if q ∈ I 0 ∩ R20 , then q =
(r − 1)(2c + 1) + 2i + 3. Suppose q ∈ I 0 ∩ R20 and q < (r − 1)(2c + 1) + 2i + 3. Let q =
(r − 1)(2c + 1) + 2i0 + 3, where 0 ≤ i0 < i. Then by the construction of A, for any j 00 > j,
vj 00 [q] = Br [2c + 2i0 + 3, j 00 − (r − 1)c2`+1 ] = Br [2c + 2i0 + 3, i1 2`+1 + a], where c − 1 ≥ i1 ≥ i and
a ∈ [2`+1 ]. Thus by (8), vj 00 [q] = Br [2c + 2i0 + 3, i1 2`+1 + a] = 0. This contradicts the assumption
that q ∈ I 0 ∩ R20 .
Suppose q ∈ I 0 ∩R20 and q > (r−1)(2c+1)+2i+3. Let q = (r−1)(2c+1)+2i0 +3, where i < i0 < c.
Then by the construction of A, for any j 00 ≤ j, vj 00 [q] = Br [2c + 2i0 + 3, j 00 − (r − 1)c2`+1 ] = Br [2c +
1 + i0 , i1 2`+1 + a], where 0 ≤ i1 ≤ i, a ∈ [2`+1 ]. Thus by (8), vj 00 [q] = Br [2c + 2i0 + 3, i1 2`+1 + a] = 0.
This contradicts the assumption that i ∈ I 0 ∩ R20 . Hence, in this case as well, |I 0 ∩ R20 | ≤ 1. This
completes the proof of the claim.
18
Therefore, we have
|I 0 | = |I 0 ∩ Ir |
(Because I 0 ⊆ Ir )
= |I 0 ∩ R1 | + |I 0 ∩ R10 | + |I 0 ∩ R| + |I 0 ∩ R2 | + |I 0 ∩ R20 |
(By (22) and Claims 3.6,3.7,3.8 and 3.9)
≤ c − (i − 1) + 0 + 1 + i + 1
= c+3
This completes the proof of the lemma.
Proof of Theorem 1.4. We prove the theorem by assuming a fast algorithm for (IP) and use it to give
a fast algorithm for CNF-SAT, refuting SETH. Let ψ be an instance of CNF-SAT with n1 variables
and m1 clauses. We choose a sufficiently large constant c such that (1 − ) + 4(1−)
+ ac < 1 holds.
c
We use the reduction mentioned in Lemma 3.1 and construct an instance A(ψ,c) x = b(ψ,c) , x ≥ 0, of
n1
(IP) which has a solution if and only if ψ is satisfiable. The reduction takes time O(m21 2 c ). Let
` = d nc1 e. The constraint matrix A(ψ,c) has dimension ((m1 − 1)(2c + 1) + 1 + c) × (m1 · c · 2`+1 ) and
the largest entry in vector b(ψ,c) does not exceed 2` − 1. The path-width of M (A(ψ,c) ) is at most
c + 4.
Assuming that any instance of (IP) with non-negative constraint matrix of path-width k is
solvable in time f (k)(d + 1)(1−)k (mn)a , where d is the maximum value in an entry of b and , a > 0
are constants, we have that A(ψ,c) x = b(ψ,c) , x ≥ 0, is solvable in time
O(1)
f (c + 4) · 2`·(1−)(c+4) · 2`·a · m1
=2
n1
(1−)(c+4)
c
·2
n1 ·a
c
O(1)
Here the constant f (c + 4) is subsumed by the term m1
whether ψ is satisfiable or not, is,
O(m21 2
n1
c
) + 2n1
where 0 = (1 − ) +
4
4(1−)
c
(1−)+
+
a
c
4(1−)
+ ac
c
O(1)
m1
= 2n1
O(1)
· m1
= 2n1
(1−)+
4(1−)
+ ac
c
O(1)
· m1
.
. Hence the total running time for testing
(1−)+
4(1−)
+ ac
c
O(1)
m1
0
O(1)
= 2 ·n1 m1
,
< 1. This completes the proof of Theorem 1.4.
Proof Sketch of Theorem 1.5
In this section we prove that (IP) with non-negative matrix A cannot be solved in time f (d)(d +
1)(1−)k (mn)O(1) for any function f and > 0, unless SETH fails, where d = max{b[1], . . . , b[m]}
and k is the path-width of the column matroid of A.
In Section 3, we gave a reduction from CNF-SAT to (IP). However in this reduction the values
n
in the constraint matrix A(ψ,c) and target vector b(ψ,c) can be as large as 2d c e − 1, where n is the
number of variables in the CNF-formula ψ and c is a constant. Let m be the number of clauses in ψ.
In this section we briefly explain how to get rid of these large values, at the cost of making large, but
still bounded path-width. From a CNF-formula ψ, we construct a matrix A = A(ψ,c) as described
in Section 3. The only rows in A which contain values strictly greater than 1 (values other than 0 or
1) are the rows indexed from the set P = {(r − 1)(2c + 1) + 1 + 2i + 1 | r ∈ [m − 2], i ∈ Zc }. That is
the values greater than 1 are in the alternate rows in yellow/green colored portion except the last c
rows in A in Figure 2b. Recall that ` = d nc e and the largest value in A is 2` − 1. Any number less
than or equal to 2` − 1 can be represented by a binary string of length ` = nc . Now we construct
a new matrix A0 from A by replacing each row of A whose index is in the set P with ` rows and
19
for any value A[i, j], i ∈ P we write its `-bit binary representation in the column corresponding
to j and the newly added ` rows of A0 . That is, for any γ ∈ P , we replace the row γ with ` rows,
γ1 , . . . , γ` . Where, for any j, if the value A[γ, j] = W then A0 [γk , j] = ηk , where ηk is the k th bit in
the `-sized binary representation of W .
Let m0 be the number of rows in A0 . Now the target vector b0 is defined as b0 [i] = 1 for all
i ∈ [m0 ]. This completes the construction of the reduced (IP) instance A0 x = b0 . The correctness
proof of this reduction is using arguments similar to those used for the correctness of Lemma 3.1.
Lemma 4.1. The path-width of the column matroid of A0 is at most (c + 1) nc + 3.
Proof. We sketch the proof, which is similar to the proof of Lemma 3.4. We define Ir0 and Jr0 for
any r ∈ [m] like Ir and Jr in Section 3. In fact, the rows in Ir0 are the rows obtained from Ir in the
process explained above to construct A0 from A. We need to show that dimhspan(A0 |{1, . . . , j}) ∩
span(A0 |{j + 1, . . . , n0 })i ≤ (c + 1) nc + 2 for all j ∈ [n0 − 1], where n0 is the number of columns in A0 .
The proof proceeds by bounding the number of indices I such that for any q ∈ I there exist vectors
v ∈ A0 |{1, . . . , j} and u ∈ A0 |{j + 1, . . . , n0 } with v[q] 6= 0 6= u[q]. By arguments similar to the ones
used in the proof of Lemma 3.4, we can show that for any j ∈ [n0 − 1], the corresponding set I 0
of indices is a subset of Ir0 for some r ∈ [m]. Recall the partition of Ir into R1 , R10 , R, R2 and R20
in Lemma 3.4. We partition Ir0 into parts S1 , S10 , S, S2 and S20 . Here S10 = R1 , S = R and S20 = R2 .
Notice that R1 , R2 ⊆ P , where P is the set of rows which covers all values strictly greater than 1.
The set S1 and S2 are obtained from R1 and R2 , respectively, by the process mentioned above to
construct A0 from A. That is, each row in Ri , i ∈ {1, 2} is replaced by ` rows in Si . This allows us
to bound the following terms for some i ∈ Zc :
|I 0 ∩ S1 | ≤ (c − (i − 1))` = (c − (i − 1))`,
|I 0 ∩ S10 | = 0,
|I 0 ∩ S2 | ≤ i · `,
|I 0 ∩ S20 | ≤ 1, and
|I 0 ∩ S| ≤ |S| = 1,
By using the fact that I 0 ⊆ Ir0 and the above system of inequalities, we can show that
n
dimhspan(A0 |{1, . . . , j}) ∩ span(A0 |{j + 1, . . . , n0 })i ≤ (c + 1)d e + 2.
c
This completes the proof of the lemma.
Now the proof of the theorem follows from Lemma 4.1 and the correctness of the reduction (it is
similar to the arguments in the proof of Theorem 1.4).
5
Proof of Theorem 1.3
In this section, we sketch how the proof of Cunningham and Geelen [8] of Theorem 1.2, can be
adapted to prove Theorem 1.3. Recall that a path decomposition of width k can be obtained in
f (k) · nO(1) time for some function f by making use of the algorithm by Jeong et al. [23]. However,
we do not know if such a path decomposition can be constructed in time O((d + 1)k+1 )nO(1) , so the
assumption that a path decomposition is given is essential.
Roughly speaking, the only difference in the proof is that when parameterized by the branchwidth, the most time-consuming operation is the “merge" operation, when we have to construct a
20
new set of partial solutions with at most (d + 1)k vectors from two already computed sets of sizes
(d + 1)k each. Thus to construct a new set of vectors, one has to go through all possible pairs of
vectors from both sets, which takes time roughly (d + 1)2k . For path-width parameterization, the
new partial solution set is constructed from two sets, but this time one set contains at most (d + 1)k
vectors while the second contains at most d + 1 vectors. This allows us to construct the new set in
time roughly (d + 1)k+1 .
Recall that for X ⊆ [n], we define S(A, X) = span(A|X) ∩ span(A|E \ X), where E = [n]. The
key lemma in the proof of Theorem 1.2 is the following.
Lemma 5.1 ([8]). Let A ∈ {0, 1, . . . , d}m×n and X ⊆ [n] such that λM (A) (X) = k. Then the
number of vectors in S(A, X) ∩ {0, . . . , d}m is at most (d + 1)k−1 .
To prove Theorem 1.3, without loss of generality, we assume that the columns of A are ordered
in such a way that for every j ∈ [n − 1],
dimhspan(A|{1, . . . , i}) ∩ span(A|{i + 1, . . . , n})i ≤ k − 1.
Let A0 = [A, b]. That is A0 is obtained by appending the column-vector b to the end of A. Then
for each i ∈ [n],
dimhspan(A0 |{1, . . . , i}) ∩ span(A0 |{i + 1, . . . , n + 1})i ≤ k.
(23)
Now we use dynamic programming to check whether the following conditions are satisfied. For
X ⊆ [n + 1], let B(X) be the set of all vectors b0 ∈ Zm
≥0 such that
(1) 0 ≤ b0 ≤ b,
|X|
(2) there exists z ∈ Z≥0 such that (A0 |X)z = b0 , and
(3) b0 ∈ S(A0 , X).
Then (IP) has a solution if and only if b ∈ B([n]). Initially the algorithm computes for all
i ∈ [n], B({i}) and by Lemma 5.1, we have that |B({i})| ≤ d + 1. In fact B({i}) ⊆ {a ·
v | v is the ith column vector of A0 and a ∈ [d + 1]}. Then for each j ∈ [2, . . . n] the algorithm
computes B([j]) in increasing order of j and outputs Yes if and only if b ∈ B([n]). That is, B([j]) is
computed from the already computed sets B([j − 1]) and B({j}). Notice that b0 ∈ B([j]) if and only
if
(a) there exist b1 ∈ B({1, . . . , j − 1}) and b2 ∈ B({j}) such that b0 = b1 + b2 ,
(b) b0 ≤ b and
(c) b0 ∈ S(A0 , [j]).
So the algorithm enumerates vectors b0 satisfying condition (a), and each such vector b0 is included in
B([j]), if b0 satisfy conditions (b) and (c). Since by (23) and Lemma 5.1, |B([j − 1])| ≤ (d + 1)k and
|B({j})| ≤ d + 1, the number of vectors satisfying condition (a) is (d + 1)k , and hence the exponential
factor of the required running time follows. This provides the bound on the claimed exponential
dependence in the running time of the algorithm. The bound on the polynomial component of the
running time follows from exactly the same arguments as in [8].
21
6
Proof of Theorem 1.1
o(
m
)
In this section we prove that unless ETH fails, (IP) cannot be solved in time n log m do(m) , where
d = max{b[1], . . . , b[m]}, even when the constraint matrix is non-negative and all entries in any
feasible solution is at most 2.
Our proof is by a reduction from a 3-CNF SAT to (IP). From a 3-CNF formula ψ on n variables
and m clauses we create an equivalent (IP) instance Aψ x = bψ , x ≥ 0, where Aψ is a non-negative
integer matrix of order (2m + n) × (2(m + n)) and the largest entry in bψ is 3. Our reduction work in
polynomial time. Let ψ be the input of 3-CNF SAT. Let X = {x1 , . . . , xn } be the set of variables
in ψ and C = {C1 , . . . , Cm } be the set of clauses in ψ. Now we create 2n + 2m number of 0 − 1
vectors of length 2m + n, two per variable and two per clause. For each xi ∈ X we make two vectors
vxi and vxi . They are defined as follows. For j ∈ [m], we set
(
1 if xi ∈ Cj ,
0 otherwise,
(
1 if xi ∈ Cj ,
0 otherwise.
vxi [j] =
and
vxi [j] =
For j = m + i, we put vxi [j] = vxi [j] = 1 and for all j ∈ {m + 1, . . . , 2m + n} \ {m + i}, we define
vxi [j] = vxi [j] = 0.
For every clause Cj ∈ C, we define two vectors vCj and vCj0 as follows. For i ∈ [m], we define
(
vCj [i] =
and
1 if i = j,
0 otherwise,
(24)
vCj0 [i] = 0.
For i = m + n + j, we set vCj [i] = vCj0 [i] = 1. For every m + n + j 6= i ∈ {m + 1, . . . , 2m + n}, we
put vCj [i] = vCj0 [i] = 0.
Matrix Aψ is constructed using these vectors as columns. The columns of Aψ are ordered as
0 .
vx1 , vx1 , vx2 , vx2 , . . . , vxn , vxn , vC1 , vC10 , vC2 , vC20 , . . . vCm , vCm
Vector bψ is defined as follows.
bψ [i] =
3 if i ∈ [m],
1 if i ∈ [m + n] \ [m],
2 if i ∈ [2m + n] \ [m + n].
Lemma 6.1. Formula ψ is satisfiable if and only if Aψ x = bψ , x ≥ 0 is feasible.
Proof. Suppose that the formula ψ is satisfiable and let φ be a satisfying assignment of ψ. We
define a (2n + 2m)-vector x∗ and prove that Aψ x∗ = bψ . For any i ∈ Zn ,
(
∗
x [2i + 1] =
1 if φ(xi+1 ) = 1,
0 otherwise,
22
and
(
∗
x [2i + 2] =
1 if φ(xi+1 ) = 1
0 otherwise.
This completes the definition of first 2n entries of x∗ . The other 2m entries (the last 2m entries) of
x∗ is defined as follows. For every i ∈ Zm , we define
0
if the number of literals set to 1 in Ci+1 by φ is 3,
1 if the number of literals set to 1 in Ci+1 by φ is 2,
x [2n + 2i + 1] =
2 otherwise,
∗
and
x∗ [2n + 2i + 2] =
2 if the number of literals set to 1 in Ci+1 by φ is 3,
1 if the number of literals set to 1 in Ci+1 by φ is 2,
0 otherwise.
(25)
(26)
We now proceed to prove that x∗ is indeed a feasible solution.
Claim 6.2. Aψ x∗ = bψ .
Proof. Towards this, we need to show that for every i ∈ [2m + n],
We consider the following exhaustive cases.
P2n+2m
j=1
Aψ [i, j] · x∗ [j] = bψ [i].
Case i ∈ [m]. The fact that each clause has 3 literals along with the definition of Aψ implies that
the number of entries set to 1 in {Aψ [i, j] | j ∈ [2n]} is 3. Also, the indices j for which Aψ [i, j] is
set to one, correspond to a literal in Ci . By the definition of x∗ and the fact that φ is a satisfying
P
P2n
∗
∗
assignment, we have that 2n
j=1 Aψ [i, j]x [j] ∈ [3]. Let r =
j=1 Aψ [i, j]x [j]. Hence
2n+2m
X
Aψ [i, j]x∗ [j] =
j=1
2n
X
Aψ [i, j]x∗ [j] +
j=1
= r+
= r+
2n+2m
X
Aψ [i, j]x∗ [j]
j=2n+1
2n+2m
X
vCi [j]x∗ [j]
j=2n+1
vCi [i]x∗ [2n
+ 2(i − 1) + 1]
= r + 1 · (3 − r)
(By (24) and construction of Aψ )
(By (24) and (25))
= 3 = bψ [i].
Case i ∈ [m + n] \ [m]. By the definition of Aψ and vectors vxj [i], vxj [i], j ∈ [2n], we have that
Aψ [i, j] = 1, j ∈ [2n] if and only if j ∈ {2(i − 1) + 1, 2(i − 1) + 2}. By the definition of x∗ , exactly
P
∗
one from {x∗ [2(i − 1) + 1], x∗ [2(i − 1) + 2]} is set to 1. This implies that 2n
j=1 Aψ [i, j]x [j] = 1.
By the construction of Aψ , we have that Aψ [i, j] = 0 for every j ∈ [2m + 2n] \ [2n]. Therefore,
P2n+2m
Aψ [i, j]x∗ [j] = 1 = bψ [i]
j=1
Case i ∈ [2m + n] \ [m + n]. Let i = m + n + i0 . From the construction Aψ , we have that Aψ [i, j]
is set to zero for all j ∈ [2n], and for any j ∈ [2m + 2n] \ [2n], Aψ [i, j] is set to 1 if and only if
j ∈ {2n + 2(i0 − 1) + 1, 2n + 2(i0 − 1) + 2}. This implies that
2n+2m
X
Aψ [i, j]x∗ [j] = x∗ [2n + 2(i0 − 1) + 1] + x∗ [2n + 2(i0 − 1) + 2]
j=1
= 2 = bψ [i]
(By (25) and (26))
This completes the proof of the claim.
23
For the converse direction of the statement of the lemma, suppose that there exists x∗ ∈ Z2n+2m
≥0
such that Aψ x∗ = bψ . Now we need to show that ψ is satisfiable. We first argue that for any
i ∈ [n], exactly one of {x∗ [2(i − 1) + 1], x∗ [2(i − 1) + 2]} is set to 1 with the other set to 0. This
follows from the fact that (i) Aψ [m + i, 2(i − 1) + 1] = Aψ [m + i, 2(i − 1) + 2] = 1, (ii) for all
j ∈ [2n + 2m] \ {2(i − 1) + 1, 2(i − 1) + 2}, Aψ [m + i, j] = 0 and (iii) bψ [m + i] = 1. Now we define
an assignment φ and prove that φ is a satisfying assignment for ψ. For i ∈ [n] we define
(
φ(xi ) =
1 if x∗ [2(i − 1) + 1] = 1,
0 if x∗ [2(i − 1) + 2] = 1.
We claim that φ satisfies all the clauses. Consider a clause Cj where j ∈ [m]. Since Aψ [m + n +
j, 2n + 2(j − 1) + 1] = 1, bψ [m + n + j] = 2, and x∗ is a feasible solution, we have that
Aψ [m + n + j, 2n + 2(j − 1) + 1] · x∗ [2n + 2(j − 1) + 1] ≤ bψ [m + n + j] = 2.
This implies that x∗ [2n + 2(j − 1) + 1] ≤ 2. Let Cj = x ∨ y ∨ z where x, y, z ∈ {xi , xi | i ∈ [n]}.
Notice that from the construction Aψ there are 3 distinct columns ix , iy , iz ∈ [2n] such that ith
w
column of Aψ is same as the vector vw , where w ∈ {x, y, z}. From the construction of Aψ , the only
non-zero entries in row numbered j are Aψ [j, ix ], Aψ [j, iy ], Aψ [j, iz ] and Aψ [j, 2n + 2(j − 1) + 1]. We
have proved that x∗ [2n + 2(j − 1) + 1] ≤ 2 and notice that bψ [j] = 3. This implies that at least
one among {x∗ [ix ], x∗ [iy ], x∗ [iz ]} is 1 and the corresponding entry in row j is 1. This implies that φ
satisfies Cj . This completes the proof of the lemma.
Now we show that for every feasible solution x, the largest entry in x is at most 2. Notice
that bψ [i] = 3 for all i ∈ [m] and bψ [i] < 3 for all i ∈ [2m + n] \ [m]. This implies that for any
feasible solution x, x[i] ≤ 2 for all i ∈ [2m + n] \ [m]. From the construction of Aψ we have that
for Aψ [i, j] 6= 0, i ∈ [m], there exists an i0 ∈ [2m + n] \ [m] such that Aψ [i0 , j] = 1. This along with
the fact that bψ [i0 ] < 3 implies that in every feasible solution x, x[i] ≤ 2 for all i ∈ [2m + n] \ [m].
Hence the largest entry in any feasible solution is at most 2. The following lemma completes the
proof of the theorem.
Lemma 6.3. If there is an algorithm for (IP) runs in time n
then ETH fails.
m
o( log
)
m do(m) ,
where d = max{b[1], . . . , b[m]},
Proof. By the sparsification lemma [22], we know that 3-CNF SAT on n0 variables and cn0 clauses,
0
where c is a constant, cannot be solved in time 2o(n ) time. Suppose there is an algorithm ALG for
m
o(
)
(IP) running in time n log m do(m) . Then for a 3-CNF formula ψ with n0 variables and m0 = cn
clauses we create an instance Aψ x = bψ , x ≥ 0 of (IP) as discussed in this section, in polynomial
time, where Aψ is a matrix of dimension (2cn0 + n0 ) × (2(n0 + cn0 )) and the largest entry in bψ is 3.
The rank of Aψ is at most (2cn0 + n). Then by Lemma 6.1, we can run ALG to test whether ψ is
satisfiable or not. This takes time
0
0
2cn0 +n0 )
o( log(2cn0 +n0 ) )
(2(cn + n ))
0
0
0
· 3o(2cn +n ) = 2o(n ) ,
hence refuting ETH.
7
Conclusion
We conclude with several open questions. First of all, while our SETH-based lower bounds for (IP)
with non-negative constraint matrix are tight for path-width parameterization, there is a “(d + 1)k
24
to (d + 1)2k gap” between lower and upper bounds for branch-width parameterization. Closing this
gap is the first natural question.
The proof of Cunningham-Geelen of Theorem 1.2 consists of two parts. The first part bounds the
number of potential partial solutions corresponding to any edge of the branch decomposition tree by
(d + 1)k . The second part is the dynamic programming over the branch decomposition using the fact
that the number of potential partial solutions is bounded. The bottleneck in Cunningham-Geelen’s
algorithm is the following subproblem. We are given two vector sets A and B of partial solutions,
each set of size at most (d + 1)k . We need to construct a new vector set C of partial solutions,
where the set C will have size at most (d + 1)k and each vector from C is the sum of a vector from
A and a vector from B. Thus to construct the new set of vectors, one has to go through all possible
pairs of vectors from both sets A and B, which takes time roughly (d + 1)2k .
A tempting approach towards improving the running time of this particular step could be the use
of fast subset convolution or matrix multiplication tricks, which work very well for “join” operations
in dynamic programming algorithms over tree and branch decompositions of graphs [14, 30, 13], see
also [11, Chapter 11]. Unfortunately, we have reason to suspect that these tricks may not help for
matrices: solving the above subproblem in time (d + 1)(1−)2k nO(1) for any > 0 would imply that
3-SUM is solvable in time n2− , which is believed to be unlikely. (The 3-SUM problem asks whether
a given set of n integers contains three elements that sum to zero.) Indeed, consider an equivalent
version of 3-SUM, named 3-SUM0 , which is defined as follows. Given 3 sets of integers A, B and
C each of cardinality n, and the objective is to check whether there exist a ∈ A, b ∈ B and c ∈ C
such that a + b = c. Then, 3-SUM is solvable in time n2− if and only if 3-SUM0 is as well (see
Theorem 3.1 in [17]). However, the problem 3-SUM0 is equivalent to the most time consuming step
in the algorithm of Theorem 1.2, where the integers in the input of 3-SUM0 can be thought of as
length-one vectors. While this observation does not rule out the existence of an algorithm solving
(IP) with constraint matrices of branch-width k in time (d + 1)(1−)2k nO(1) , it indicates that any
interesting improvement in the running time would require a completely different approach.
Our final open question is to obtain a refined lower bound for (IP) with bounded rank. Recall
that the constraint matrix of the algorithm of Papadimitriou [26] can contain negative values and
improving the running time of his algorithm or showing that its running time is tight up to SETH,
is still a very interesting question.
25
References
[1] A. Abboud, A. Backurs, and V. V. Williams, Tight hardness results for LCS and other
sequence similarity measures, in Proceedings of the 56th Annual Symposium on Foundations of
Computer Science (FOCS), IEEE Computer Society, 2015, pp. 59–78. 3
[2] A. Abboud and V. V. Williams, Popular conjectures imply strong lower bounds for dynamic
problems, in Proceedings of the 55th Annual Symposium on Foundations of Computer Science
(FOCS), IEEE Computer Society, 2014, pp. 434–443. 3
[3] A. Backurs and P. Indyk, Edit distance cannot be computed in strongly subquadratic time
(unless SETH is false), in Proceedings of the 47th Annual ACM Symposium on Theory of
Computing (STOC), ACM, 2015, pp. 51–58. 3
[4]
, Which regular expression patterns are hard to match?, in Proceedings of the 57th Annual
Symposium on Foundations of Computer Science (FOCS), IEEE Computer Society, 2016, p. to
appear. 3
[5] K. Bringmann, Why walking the dog takes time: Frechet distance has no strongly subquadratic
algorithms unless SETH fails, in Proceedings of the 55th Annual Symposium on Foundations
of Computer Science (FOCS), IEEE Computer Society, 2014, pp. 661–670. 3
[6] K. Bringmann and M. Künnemann, Quadratic conditional lower bounds for string problems
and dynamic time warping, in Proceedings of the 56th Annual Symposium on Foundations of
Computer Science (FOCS), IEEE Computer Society, 2015, pp. 79–97. 3
[7] W. Cook and P. D. Seymour, Tour merging via branch-decomposition, INFORMS Journal
on Computing, 15 (2003), pp. 233–248. 1
[8] W. H. Cunningham and J. Geelen, On integer programming and the branch-width of the
constraint matrix, in Proceedings of the 12th International Conference on Integer Programming
and Combinatorial Optimization (IPCO), vol. 4513 of Lecture Notes in Comput. Sci., Springer,
2007, pp. 158–166. 1, 20, 21
[9] R. Curticapean and D. Marx, Tight conditional lower bounds for counting perfect matchings
on graphs of bounded treewidth, cliquewidth, and genus, in Proceedings of the 26th Annual
ACM-SIAM Symposium on Discrete Algorithms (SODA), SIAM, 2016, pp. 1650–1669. 3
[10] M. Cygan, H. Dell, D. Lokshtanov, D. Marx, J. Nederlof, Y. Okamoto, R. Paturi,
S. Saurabh, and M. Wahlström, On problems as hard as CNF-SAT, in Proceedings of the
27th IEEE Conference on Computational Complexity (CCC), IEEE, 2012, pp. 74–84. 3
[11] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk,
M. Pilipczuk, and S. Saurabh, Parameterized Algorithms, Springer, 2015. 3, 25
[12] M. Cygan, S. Kratsch, and J. Nederlof, Fast hamiltonicity checking via bases of perfect
matchings, in Proceedings of the 45th Annual ACM Symposium on Theory of Computing
(STOC), ACM, 2013, pp. 301–310. 3
[13] M. Cygan, J. Nederlof, M. Pilipczuk, M. Pilipczuk, J. M. M. van Rooij, and J. O.
Wojtaszczyk, Solving connectivity problems parameterized by treewidth in single exponential
time, in Proceedings of the 52nd Annual Symposium on Foundations of Computer Science
(FOCS), IEEE, 2011, pp. 150–159. 25
26
[14] F. Dorn, Dynamic programming and fast matrix multiplication, in Proceedings of the 14th
Annual European Symposium on Algorithms (ESA), vol. 4168 of Lecture Notes in Comput.
Sci., Springer, Berlin, 2006, pp. 280–291. 25
[15] F. V. Fomin, P. A. Golovach, D. Lokshtanov, and S. Saurabh, Almost optimal
lower bounds for problems parameterized by clique-width, SIAM J. Computing, 43 (2014),
pp. 1541–1563. 3
[16] F. V. Fomin and D. M. Thilikos, Dominating sets in planar graphs: Branch-width and
exponential speed-up, SIAM J. Computing, 36 (2006), pp. 281–309. 1
[17] A. Gajentaan and M. H. Overmars, On a class of o(n2 ) problems in computational
geometry, Comput. Geom., 5 (1995), pp. 165–185. 25
[18] P. Hliněný, Branch-width, parse trees, and monadic second-order logic for matroids, J.
Combinatorial Theory Ser. B, 96 (2006), pp. 325–351. 1
[19] P. Hlinený, The tutte polynomial for matroids of bounded branch-width, Combinatorics,
Probability & Computing, 15 (2006), pp. 397–409. 1
[20] G. B. Horn and F. R. Kschischang, On the intractability of permuting a block code to
minimize trellis complexity, IEEE Trans. Information Theory, 42 (1996), pp. 2042–2048. 2
[21] R. Impagliazzo and R. Paturi, On the complexity of k-SAT, J. Computer and System
Sciences, 62 (2001), pp. 367–375. 1, 3
[22] R. Impagliazzo, R. Paturi, and F. Zane, Which problems have strongly exponential
complexity, J. Computer and System Sciences, 63 (2001), pp. 512–530. 3, 24
[23] J. Jeong, E. J. Kim, and S. Oum, Constructive algorithm for path-width of matroids, in
Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
SIAM, 2016, pp. 1695–1704. 4, 20
[24] D. Lokshtanov, D. Marx, and S. Saurabh, Known algorithms on graphs on bounded
treewidth are probably optimal, in Proceedings of the 21st Annual ACM-SIAM Symposium on
Discrete Algorithms (SODA), SIAM, 2011, pp. 777–789. 3
[25] D. Lokshtanov, D. Marx, and S. Saurabh, Lower bounds based on the exponential time
hypothesis, Bulletin of the EATCS, 105 (2011), pp. 41–72. 3
[26] C. H. Papadimitriou, On the complexity of integer programming, J. ACM, 28 (1981), pp. 765–
768. 1, 2, 25
[27] M. Pătraşcu and R. Williams, On the possibility of faster SAT algorithms, in Proceedings
of the 20th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), SIAM, 2010,
pp. 1065–1075. 3
[28] N. Robertson and P. D. Seymour, Graph minors. X. Obstructions to tree-decomposition,
J. Combinatorial Theory Ser. B, 52 (1991), pp. 153–190. 1, 4
[29] L. Roditty and V. V. Williams, Fast approximation algorithms for the diameter and radius
of sparse graphs, in Proceedings of the 45th Annual ACM Symposium on Theory of Computing
(STOC), ACM, 2013, pp. 515–524. 3
27
[30] J. M. M. van Rooij, H. L. Bodlaender, and P. Rossmanith, Dynamic programming
on tree decompositions using generalised fast subset convolution, in Proceedings of the 17th
Annual European Symposium on Algorithms (ESA), vol. 5757 of Lecture Notes in Comput.
Sci., Springer, 2009, pp. 566–577. 25
[31] V. V. Williams, Hardness of Easy Problems: Basing Hardness on Popular Conjectures
such as the Strong Exponential Time Hypothesis (Invited Talk), in Proceedings of the 10th
International Symposium on Parameterized and Exact Computation (IPEC), vol. 43 of Leibniz
International Proceedings in Informatics (LIPIcs), Dagstuhl, Germany, 2015, Schloss Dagstuhl–
Leibniz-Zentrum fuer Informatik, pp. 17–29. 3
28
| 8cs.DS
|
DIMENSION RIGIDITY OF LATTICES IN
SEMISIMPLE LIE GROUPS
arXiv:1607.03652v1 [math.AT] 13 Jul 2016
CYRIL LACOSTE
Abstract. We prove that if Γ is a lattice in the group of isometries of a symmetric space of non-compact type without euclidean
factors, then the virtual cohomological dimension of Γ equals its
proper geometric dimension.
1. Introduction
Let Γ be a discrete virtually torsion-free group. There exist several
notions of "dimension" for Γ. One of them is the virtual cohomological
dimension vcd(Γ), which is the cohomological dimension of any torsionfree finite index subgroup of Γ. Due to a result by Serre, it does not
depend on the choice of such a subgroup (see [9]). Another one is the
proper geometric dimension. A Γ-CW-complex X is said to be a model
for EΓ if the stabilizers of the action of Γ on X are finite and for every
finite subgroup H of Γ, the fixed point space X H is contractible. Note
that two models for EΓ are Γ-equivariantly homotopy equivalent to
each other. The proper geometric dimension gd(Γ) of Γ is the smallest
possible dimension of a model for EΓ.
These two notions are related. In fact, we always have the inequality
vcd(Γ) 6 gd(Γ)
but this inequality may be strict, see for instance the construction of
Leary and Nucinkis in [23], or other examples in [8], [24], [28], [11],
[12].
However there are also many examples of virtually torsion-free groups
Γ with vcd(Γ) = gd(Γ). For instance in [10] Degrijse and MartinezPerez prove that this is the case for a large class of groups containing
all finitely generated Coxeter groups. Other examples for equality can
be found in [1], [2], [25] and [33].
In this paper we will prove that equality holds for groups acting by
isometries, discretely and with finite covolume on symmetric spaces of
non-compact type without euclidean factors:
Theorem 1.1. Let S be a symmetric space of non-compact type without
euclidean factors. Then
gd(Γ) = vcd(Γ)
for every lattice Γ ⊂ Isom(S).
1
2
CYRIL LACOSTE
Recall that a symmetric space of non-compact type without euclidean factors is of the form G/K where G is a semisimple Lie group,
which can be assumed to be connected and centerfree, and K ⊂ G is a
maximal compact subgroup. Then Isom(S) = Aut(g) = Aut(G) where
g is the Lie algebra of G, and note that this group is semisimple, linear
and algebraic but may be not connected. In [1] the authors prove Theorem 1.1 for lattices in classical simple Lie groups G. We will heavily
rely on their results and techniques.
We discuss now some applications of Theorem 1.1. First note that
the symmetric space S is a model for EΓ. Theorem 1.1 yields then
that:
Corollary 1.1. If S is a symmetric space of non-compact type and
without euclidean factors, and if Γ ⊂ Isom(S) is a lattice, then S is
Γ-equivariantly homotopy equivalent to a proper cocompact Γ-CW complex of dimension vcd(Γ).
We stress again that in the setting of Theorem 1.1 we are considering
the full group of isometries of S. This has the consequence that we are
able to deduce that there is equality between the virtual cohomological
dimension and the proper geometric dimension not only for lattices
in Isom(S), but also for groups abstractly commensurable to them.
Here, two groups Γ1 and Γ2 are said abstractly commensurable if for
f of finite index in Γ , such that Γ
f
i = 1, 2, there exists a subgroup Γ
i
i
1
f
is isomorphic to Γ2 . Then we obtain from Theorem 1.1 that:
Corollary 1.2. If a group Γ is abstractly commensurable to a lattice
in the group of isometries of a symmetric space of non-compact type
without euclidean factors, then gd(Γ) = vcd(Γ).
Remark 1.1 Note that in general the equality between the proper
geometric dimension and the virtual cohomological dimension behaves
badly under commensuration. For instance, the fact that there exist
virtully torsion-free groups Γ with vcd(Γ) > 3 and such that vcd(Γ) <
gd(Γ) proves that if Γ0 is a torsion-free subgroup of Γ of finite index,
then vcd(Γ0 ) = cd(Γ0 ) = gd(Γ0 ) = gd(Γ0 ), whereas Γ is commensurable to Γ0 and vcd(Γ) < gd(Γ). In fact, we have concrete exemples of
groups for which Corollary 1.2 fails among familiar classes of groups.
For instance, in [10] the authors prove that if Γ is a finitely generated Coxeter group then vcd(Γ) = gd(Γ) and in [24] the authors construct finite extensions of certain right-angled Coxeter groups such that
vcd(Γ) < gd(Γ).
Returning to the applications of Theorem 1.1, we obtain from Corollary 1.2 that lattices in Isom(S) are dimension rigid in the sense of
[12]: we say that a virtually torsion-free group Γ is dimension rigid if
e = vcd(Γ)
e for every group Γ
e which contains Γ as a finite
one has gd(Γ)
index normal subgroup.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
3
Dimension rigidity has a strong impact on the behaviour of the
proper geometric dimension under group extensions, and we obtain
from Corollary 1.2 and [11, Cor. 2.3] that:
Corollary 1.3. If Γ is a lattice in the group of isometries of a symmetric space of non-compact type without euclidean factors and
1→Γ→G→Q→1
is a short exact sequence, then gd(G) 6 gd(Γ) + gd(Q).
We sketch now the strategy of the proof of Theorem 1.1. To begin
with, note that while symmetric spaces, both Riemannian and nonRiemannian, will play a key role in our considerations, most of the
time we will be working in the ambient Lie group. In fact it will be
convenient to reformulate Theorem 1.1 as follows:
Main Theorem Let g be a semisimple Lie algebra. Then
gd(Γ) = vcd(Γ)
for every lattice Γ ⊂ Aut(g).
The key ingredient in the proof of the Main Theorem, and hence of
Theorem 1.1, is a result of Lück and Meintrup [26], which basically
asserts that the proper geometric dimension gd(Γ) equals the Bredon
cohomological dimension cd(Γ) - see Theorem 2.3 for a precise statement. In the light of this theorem it suffices to prove that the two
cohomological notions of dimension vcd(Γ) and cd(Γ) coincide. In [1]
the authors noted that to prove the equality vcd(Γ) = cd(Γ) it suffices
to ensure that the fixed point sets S α of finite order elements α ∈ Γ
are of small dimension - see Section 2.8 for details. Still in [1] the
authors checked that this was the case for lattices contained in the
classical simple Lie groups. We will use a similar strategy to prove the
Main Theorem for lattices in groups of automorphisms of all simple Lie
algebras. Recall that any non-compact finite dimensional simple Lie
algebra over R is either isomorphic to one of the classical types or to
one of the exceptional ones. The classical Lie algebras are the complex
ones
sl(n, C), so(n, C), sp(2n, C)
and their real forms
sl(n, R), sl(n, H), so(p, q), su(p, q), sp(p, q), sp(2n, R), so∗ (2n).
Similarly, the exceptional Lie algebras are the five complex ones
gC2 , fC4 , eC6 , eC7 , eC8
and their twelve real forms
g2(2) , f4(4) , f4(−20) , e6(6) , e6(2) , e6(−14) , e6(−26) ,
e7(7) , e7(−5) , e7(−25) , e8(8) , e8(−24) .
4
CYRIL LACOSTE
Here the number in brackets is the difference between the dimension
of the adjoint group and twice the dimension of a maximal compact
subgroup (which equals 0 for a complex Lie group).
We illustrate now the basic steps of the proof of the Main Theorem
in the example of g = sl(n, C). Suppose that Γ ⊂ Aut(g) is a lattice,
and consider the symmetric space S = PSL(n, C)/PSUn . To prove that
gd(Γ) = vcd(Γ), it will suffice to establish that
(1.1)
dim S α < dim S − rkR (PSL(n, C)) = (n2 − 1) − (n − 1)
for every α ∈ Γ of finite order and non central (see Lemma 2.2). First
note that α is the composition of an inner automorphism and an outer
automorphism. Since every non-trivial element in Out(g) has order 2,
it follows that α2 is an inner automorphism. If it is non trivial then
we use the results of Section 6.1 in [1]. In general, if α = Ad(A) is a
non-trivial inner automorphism of sl(n, C), we get from [1] that (1.1)
holds.
We are reduced to the case where α2 is trivial, meaning that α is
of order 2. Then the automorphism α ∈ Aut(g) is induced by an
automorphism of the adjoint group Gad = PSL(n, C) which is still
denoted α and is also an involution. The fixed point set S α is the
Riemannian symmetric space associated to Gαad , where Gαad is the set
of fixed points of α. Now, notice that the quotient Gad /Gαad is a (nonRiemannian) symmetric space. The symmetric spaces associated to
simple groups have been classified by Berger in [4]. In the case of
Gad = PSL(n, C), we obtain from this classification that the Lie algebra
of Gαad is either compact or isomorphic to so(n, C), s(gl(k, C) ⊕ gl(n −
k, C)), sp(n, C), sl(n, R), su(p, n − p) or sl( n2 , H), where sp(n, C) and
sl( n2 , H) only appear if n is even. Armed with this information, we
check (1.1) for every involution α, which leads to the Main Theorem
for g = sl(n, C). The argument we just sketched will be applied in
Section 3 to all complex simple Lie algebras and in Section 4 to the
real ones. Since the arguments are similar, and since the complex case
is somewhat easier, we advise the reader to skip Section 4 in a first
reading.
Having dealt with the simple Lie algebras, we treat in section 5 the
semisimple case. The method for the simple algebras will not work at
first sight, but the proof will eventually by simpler. The idea is to restrict to irreducible lattices, i.e. those who cannot be decomposed into
a product. Then we will show that the rational rank of an irreducible
lattice is lower than the real rank of any factor of the adjoint group,
meaning that we get a much improved bound than in (1.1). This fact
will lead rapidly to the Main Theorem.
Finally note that in the proof of the Main Theorem, we do not construct a concrete model for EΓ of dimension vcd(Γ), we just prove its
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
5
existence. It is however worth mentioning that in a few cases such
models are known. For instance if Γ = SL(n, Z), the symmetric space
S = SL(n, R)/SOn admits a Γ-equivariant deformation retract of dimension vcd(Γ) called the "well-rounded retract" (see [3], [31] and [32]).
It will be interesting to do the same for groups such as Sp(2n, Z).
Acknowledgements. The author thanks Dave Witte Morris for his
help, Dieter Degrijse for interesting discussions and Juan Souto for his
useful advice and instructive discussions.
2. Preliminaries
In this section we recall some basic facts and definitions about algebraic groups, Lie groups and Lie algebras, symmetric spaces, lattices
and arithmetic groups, virtual cohomological dimension and Bredon
cohomology.
2.1. Algebraic groups and Lie groups. An algebraic group is a
subgroup G of SL(N, C) determined by a collection of polynomials. It
is defined over a subfield k of C if those polynomials can be chosen to
have coefficients in k. The Galois criterion (see [6, Prop. 14.2 p.30])
says that G is defined over k if and only if G is stable under the Galois
group Gal(C/k). If G is an algebraic group and R ⊂ C is a ring we note
GR the set of elements of G with entries in R. If G is an algebraic group
defined over R, it is well-known that the groups GC and GR are Lie
groups with finitely many connected components. In fact, G is Zariski
connected if and only if GC is a connected Lie group, whereas GR may
not be connected in this case. A non-abelian algebraic group (or Lie
group) is said simple if every connected normal subgroup is trivial, and
semisimple if every connected normal abelian subgroup is trivial. Note
that if G is a semisimple algebraic group defined over k = R or C
then Gk is a semisimple Lie group. Any connected semisimple complex
linear Lie group is algebraic and any connected semisimple real linear
Lie group is the identity component of the group of real points of an
algebraic group defined over R. Recall that two Lie groups G1 and
G2 are isogenous if they are locally isomorphic, meaning that there
exist finite normal subgroups N1 ⊂ G01 and N2 ⊂ G02 of the identity
components of G1 and G2 such that G01 /N1 is isomorphic to G02 /N2 .
A semisimple linear Lie group is isogenous to a product of simple Lie
groups.
The center Z(G) of a semisimple algebraic group G is finite (it is
also the case for semisimple linear Lie groups but not for semisimple
Lie groups in general) and the quotient G/Z(G) is again a semisimple
algebraic group (see [6, Thm. 6.8 p.98]). Moreover, if G is defined over
k then so is G/Z(G).
6
CYRIL LACOSTE
A connected algebraic group T ⊂ SL(N, C) is a torus if it is diagonalizable, meaning there exists A ∈ SL(N, C) such that for every B ∈ T,
ABA−1 is diagonal. A torus is in particular abelian and isomorphic,
as an algebraic group, to a product C∗ × · · · × C∗ . If T is defined over
k, it is said to be k-split if the conjugating element A can be chosen
in SL(N, k). A torus in an algebraic group G is a subgroup that is a
torus. It is said to be maximal if it is not strictly contained in any
other torus. An important fact is that any two maximal tori in G are
conjugate in G, and that if G is defined over k, then any two maximal
k-split tori are conjugate by an element in Gk . The k-rank of G (or
of Gk ), denoted by rkk G (or rkk Gk ), is the dimension of any maximal
k-split torus in G, and the rank of G is just the C-rank.
We refer to [6], [22] and [30] for basic facts about algebraic groups
and Lie groups.
2.2. Lie algebras and their automorphisms. Recall that the Lie
algebra g of a Lie group G is the set of left-invariant vector fields. A
subalgebra of g is a subspace closed under Lie bracket. An ideal is a
subalgebra I such that [g, I] ⊂ I. The Lie algebra g is simple if it is
not abelian and has no non-trivial ideals, and semisimple if it has no
non-zero abelian ideals. A Lie group is simple (resp. semisimple) if
and only if its Lie algebra is simple (resp. semisimple). A semisimple
Lie algebra is isomorphic to a finite direct sum of simple ones.
By Lie’s third theorem, if g is a finite dimensional real Lie algebra
(which will be always the case here), there exists a connected Lie group,
unique up to covering, whose Lie algebra is g. This means that their
exists a unique simply connected Lie group G associated to g, and every
other connected Lie group whose Lie algebra is g is a quotient of G by a
subgroup contained in the center. In particular, Gad = G/Z(G) is the
unique connected centerfree Lie group associated to g. The group Gad
is called the adjoint group of g. The adjoint group is a linear algebraic
group, whereas its universal cover may be not linear (see for instance
the universal cover of PSL(2, R)). It follows that the classification of
simple Lie algebras is in correspondance with that of simple Lie groups.
A Lie algebra is said compact if the adjoint group is.
An automorphism of a Lie algebra g is a bijective linear endomorphism which preserves the Lie bracket. The group of automorphisms
of g is denoted Aut(g), it is linear and algebraic but not connected in
general. If G is a Lie group associated to g, then the differential of a
Lie group automorphism of G is an automorphism of g. Conversely,
if G is either simply connected or connected and centerfree, any automorphism of g comes from an automorphism of G. In this case, we
will often identify these two automorphisms and denote them by the
same letter. An inner automorphism is the derivative of the conjugation in G by an element A ∈ G - we denote it Ad(A). The group
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
7
Inn(g) of inner automorphisms is a normal subgroup of Aut(g). It is
also the identity component of Aut(g) and is isomorphic to the adjoint
group Gad . If g is semisimple, the subgroup Inn(g) is of finite index
in Aut(g) and the quotient Aut(g)/Inn(g) is the (finite) group of outer
automorphisms Out(g). Moreover if g is simple, Out(g) can be seen as
a subgroup of Aut(g) and Aut(g) is the semidirect product of Out(g)
and Inn(g), that is Aut(g) = Inn(g) o Out(g) (see [16]).
Note that even if g is complex, we let Aut(g) be the group of real
automorphisms. If g is complex and simple then Aut(g) contains the
complex automorphism group AutC (g) as a subgroup of index 2, the
quotient being generated by complex conjugation (see [13, Prop. 4.1]).
Recall that if g is a complex Lie algebra, a real form of g is a real Lie
algebra whose complexification is g. Any real form is the group of fixed
points by a conjugation of g, meaning an involutive real automorphism
which is antilinear over C.
We refer to [30], [22], [17] and [13] for other facts about Lie algebras
and their automorphisms.
2.3. Simple Lie groups, simple Lie algebras and their outer
automorphisms. As mentioned in the previous section, the classification of simple Lie groups (up to isogeny) and of simple Lie algebras
are in correspondance. Both are due to by Cartan. We will now see
that of simple Lie groups. Every linear simple Lie group is isogenous
to either a classical group or to one of the finitely many exceptional
groups. We denote the transpose of a matrix A by At and its conjugate
transpose by A∗ and we consider the particular matrices
Jn =
Idn
−Idn
!
, Qp,q =
!
−Idp
Idq
.
The non-compact classical simple Lie groups are the groups in the
following list
SL(n, C) = {A ∈ GL(n, C)| det A = 1}
n>2
SO(n, C) = {A ∈ SL(n, C)|At A = Id}
n > 3, n 6= 4
Sp(2n, C) = {A ∈ SL(2n, C)|At Jn A = Jn }
n>1
SL(n, R) = {A ∈ GL(n, R)| det A = 1}
n>2
SL(n, H) = {A ∈ GL(n, H)| det A = 1}
n>2
SO(p, q) = {A ∈ SL(p + q, R)|At Qp,q A = Qp,q } 1 6 p 6 q, p + q > 3
SU(p, q) = {A ∈ SL(p + q, C)|A∗ Qp,q A = Qp,q } 1 6 p 6 q, p + q > 3
Sp(p, q) = {A ∈ GL(p + q, H)|A∗ Qp,q A = Qp,q } 1 6 p 6 q, p + q > 3
Sp(2n, R) = {A ∈ SL(2n, R)|At Jn A = Jn }
n>1
∗
t
SO (2n) = {A ∈ SU(n, n)|A Qn,n Jn A = Qn,n Jn }
n>2
Similarly we give the list of the compact ones
SOn = {A ∈ SL(n, R)|At A = Id}
SUn = {A ∈ SL(n, C)|A∗ A = Id}
On = {A ∈ GL(n, R)|At A = Id}
Un = {A ∈ GL(n, C)|A∗ A = Id}
8
CYRIL LACOSTE
Spn = {A ∈ GL(n, H)|A∗ A = Id}
The compact exceptional Lie groups are
G2 , F4 , E6 , E7 , E8
and the non-compact ones are the complex ones (which are the complexifications of the previous compact groups)
GC2 , F4C , E6C , E7C , E8C
and their real forms
G2(2) , F4(4) , F4(−20) , F4(4) , E6(6) , E6(2) , E6(−14) , E6(−26) ,
E7(7) , E7(−5) , E7(−25) , E8(8) , E8(−24) .
We refer to [34] for definitions and complete descriptions of the simply connected versions of the exceptional Lie groups. Note that in this
paper we will always consider the centerless versions with the same
notations.
As usual, the simple Lie algebra associated to a simple Lie group will
be denoted by gothic caracters, for instance sl(n, R) is the Lie algebra
of SL(n, R). Note that the adjoint group of sl(n, R) is PSL(n, R). The
classification of simple Lie algebras runs in parallel to that of simple
Lie groups. The following table summarizes the structure of the outer
automorphisms groups of simple Lie algebras (see [17] section 3.2). We
denote by Sn the symmetric group and D2n the dihedral group.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
g
sl(n, C), n > 3
so(8, C)
so(2n, C), n > 5
eC6
all others complex Lie algebras
sl(2, R)
sl(n, R), n > 3 odd
sl(n, R), n > 4 even
su(p, q), p 6= q
su(p, p), p > 2
sl(n, H)
so(p, q), p + q odd
so(p, q), p and q odd, p 6= q
so(p, q) p and q even, p 6= q
so(p, p) p > 5 odd
so(p, p) p > 6 even
so(4, 4)
sp(2n, R)
sp(p, p)
so∗ (2n)
e6(j) , j = 6, 2, −14, −26
e7(j) , j = 7, −25
all others real Lie algebras
Table 1. Outer automorphisms groups of
9
Out(g)
Z2 × Z2
S3 × Z2
Z2 × Z2
Z2 × Z2
Z2
Z2
Z2
Z2 × Z2
Z2
Z2 × Z2
Z2
Z2
Z2
Z2 × Z2
Z2 × Z2
D4
S4
Z2
Z2
Z2
Z2
Z2
1
simple Lie algebras
Note that we have the isomorphisms: so(3, C) ∼
= sl(2, C) ∼
= sp(2, C),
∼
∼
so(5, C) = sp(4, C), so(6, C) = sl(4, C) and the corresponding ones
between their real forms, and that so(4, C) is not simple but isomorphic
to sl(2, C) ⊕ sl(2, C).
2.4. Symmetric spaces. Let G be a Lie group. A symmetric space
is a space of the form G/Gρ where ρ is an involutive automorphism
of G and Gρ its fixed points set. It is said irreducible if it can not
be decomposed as a product. From an algebraic point of view, the
irreducibility of G/H implies that the Lie algebra h of H is a maximal
subalgebra of the Lie algebra g of G. Equivalently, the irreducibility of
G/H implies that the identity component of H is a maximal connected
Lie subgroup of the identity component of G.
Another point of view on symmetric spaces is based on Lie algebras.
If G/Gρ is a symmetric space and g is the Lie algebra of G, the involutive automorphism ρ induces an involutive automorphism of g whose
fixed point set is the Lie algebra h of H = Gρ . We can thus always
associate to a symmetric space G/H a linear space g/h, called a local
symmetric space. The Lie subalgebra h is called the isotropy algebra of
10
CYRIL LACOSTE
g/h (more generally we say that h is an isotropy algebra if it is the fixed
point set of an involutive automorphism) . Conversely, if g is a Lie algebra, G a simply connected or connected and centerless Lie group whose
Lie algebra is g, and h ⊂ g an isotropy algebra, the local symmetric
space g/h lifts to a symmetric space G/H, because Aut(g) = Aut(G).
So the classification of symmetric spaces are in correspondance with
those of local symmetric spaces, and has been done by Berger in [4].
Note that if g is simple and complex, and if ρ ∈ Aut(g) is an involution, then ρ is either C-linear and in this case h = gρ is also complex,
or ρ is anti-linear (that means it is a conjugation) and h = gρ is a real
form of g. Note also that is g is real and ρ is an involution, then ρ can
be extended to a C-linear involution ρC of the complexification gC of g,
C
and the isotropy algebra (gC )ρ is the complexification of h = gρ , that
C
is (gC )ρ = hC .
We give now the list of the non-compact isotropy algebras of the
local symmetric spaces associated to sl(n, C) and its real forms.
gC = sl(n, C)
g = sl(n, R)
n
,H
2
gl n2 , C
Table 2. Non-compact isotropy algebras of sl(n, C) and
its real forms
Table 2 is organized as follows: in the first line we give the complex
isotropy algebras hC of sl(n, C) (fixed by a complex involution). Each
column consists of real forms of the complex algebra in the first entry. The local symmetric spaces associated to sl(n, C) are then those
of the form sl(n, C)/hC , for instance sl(n, C)/so(n, C), or of the form
sl(n, C)/g, for instance sl(n, C)/sl(n, R). The ones associated to a real
form g are of the form g/h, for instance sl(n, R)/so(k, l) with k + l = n.
The following tables summarize the classification for other simple
Lie algebras. They are organized in a similar way.
gC = so(n, C)
hC = so(k, C) ⊕ so(l, C)
g = so(p, q)
h = so(kp , kq ) ⊕ so(lp , lq ))
(p =
q) h = so(p, C)
∗
h = so 2 k2 ⊕ so∗ 2 2l
g = so∗ 2
h = so(p, q)
h = s(u(kp , kq ) ⊕ u(lp , lq )) h = sp p2 , 2q
∗
(p = q) h = so
(2p)
h = gl(p, C)
h = sp(2p,
R)
n
k
l
∗
h = so 2 2
h = s(gl( 2 , H) ⊕ gl( 2 , H)) h = sp k2 , 2l
g = su(p, q)
g = sl
hC = s(gl(k, C) ⊕ gl(l, C)) hC = sp(n, C)
h = s(gl(k, R)
h = sp(n, R)
⊕ gl(l,
R))
n
h = gl 2 , C
hC = so(n, C)
h = so(k, l)
n
2
h = so
n
,C
2
n
,C
2
u p2 , 2q
hC = gl
h=
h = gl(p,
R)
k l
h = u 2, 2
h = gl
n
,H
4
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
11
Table 3. Non-compact isotropy algebras of so(n, C)
and its real forms
gC = sp(2n, C) hC = sp(2k, C) ⊕ sp(2l, C)
g = sp(p, q)
h = sp(kp , kq ) ⊕ sp(lp , lq )
(p = q) h = sp(p, C)
g = sp(2n, R)
h = sp(2k, R) ⊕ sp(2l, R)
h = sp(n, C)
Table 4. Non-compact isotropy algebras
and its real forms
hC = gl(n, C)
h = u(p, q)
h = gl(p, H)
h = u(k, l)
h = gl(n, R)
of sp(2n, C)
gC = gC2 hC = sl(2, C) ⊕ sl(2, C)
g = g2(2) h = sl(2, R) ⊕ sl(2, R)
Table 5. Non-compact isotropy algebras of gC2 and its
real form
hC = sp(6, C) ⊕ sp(2, C) hC = so(9, C)
h = sp(6, R) ⊕ sp(2, R)
h = so(4, 5)
h = sp(1, 2) ⊕ sp(1)
g = f4(−20)
h = sp(1, 2) ⊕ sp(1)
h = so(1, 8)
Table 6. Non-compact isotropy algebras of fC4 and its
real forms
gC = fC4
g = f4(4)
hC = sp(8, C) hC = sl(6, C) ⊕ sl(2, C) hC = so(10, C) ⊕ so(2, C) hC = fC4
h = sp(2, 2)
h = sl(6, R) ⊕ sl(2, R)
h = so(5, 5) ⊕ so(1, 1)
h = f4(4)
h = sp(8, R)
h = sl(3, H) ⊕ su(2)
g = e6(2)
h = sp(1, 3)
h = su(2, 4) ⊕ su(2)
h = so(4, 6) ⊕ so(2)
h = f4(4)
∗
h = sp(8, R) h = su(3, 3) ⊕ sl(2, R)
h = so (10) ⊕ so(2)
g = e6(−14)
h = sp(2, 2)
h = su(2, 4) ⊕ su(2)
h = so(2, 8) ⊕ so(2)
h = f4(−20)
∗
h = su(1, 5) ⊕ sl(2, R)
h = so (10) ⊕ so(2)
g = e6(−26) h = sp(1, 3)
h = sl(3, H) ⊕ sp(1)
h = so(1, 9) ⊕ so(1, 1)
h = f4(−20)
C
Table 7. Non-compact isotropy algebras of e6 and its
real forms
gC = eC6
g = e6(6)
hC = sl(8, C) hC = so(12, C) ⊕ sl(2, C) hC = eC6 ⊕ so(2, C)
h = su(4, 4)
h = so(6, 6) ⊕ sl(2, R)
h = e6(6) ⊕ so(1, 1)
h = sl(8, R)
h = so∗ (12) ⊕ sp(1)
h = e6(2) ⊕ so(2)
h = sl(4, H)
g = e7(−5)
h = su(4, 4)
h = so(4, 8) ⊕ su(2)
h = e6(2) × so(2)
∗
h = su(2, 6)
h = so (12) ⊕ sl(2, R)
h = e6(−14) ⊕ so(2)
g = e7(−25) h = sl(4, H) h = so(2, 10) ⊕ sl(2, R)
h = e6(−14) ⊕ so(2)
∗
h = su(2, 6)
h = so (12) ⊕ sp(1)
h = e6(−26) ⊕ so(1, 1)
C
Table 8. Non-compact isotropy algebras of e7 and its
real forms
gC = eC7
g = e7(7)
12
CYRIL LACOSTE
hC = eC7 ⊕ sl(2, C) hC = so(16, C)
h = e7(7) ⊕ sl(2, R)
h = so(8, 8)
h = e7(−5) ⊕ su(2)
h = so∗ (16)
g = e8(−24) h = e7(−25) ⊕ sl(2, R) h = so(4, 12)
h = e7(−5) ⊕ su(2)
h = so∗ (16)
Table 9. Non-compact isotropy algebras of eC8 and its
real forms
gC = eC8
g = e8(8)
Note that not all the symmetric spaces given in these tables are
irreducible. For instance sl(n, C)/s(gl(k, C) ⊕ gl(l, C)) is not. The
results of [4] are more precise and we refer to them for the list of the
irreducible symmetric spaces and the non-irreducibles ones.
We refer to [18] and [4] for facts about symmetric spaces and local
symmetric spaces.
2.5. Riemannian symmetric spaces. We stress that the symmetric
spaces G/H associated to the isotropy subalgebras h of g in Tables
2 to 9 are non-Riemannian. We discuss now a few features about
Riemannian symmetric spaces, which are of the form G/Gρ with Gρ
compact. The symmetric spaces which are Riemannian spaces of nonpositive curvature are called symmetric spaces of non-compact type.
They are all of the form S = G/K where G = Isom(S)0 and K is a
maximal compact subgroup. If it has no euclidean factors, then G is
semisimple, linear and centerless.
Recall that if G is a Lie group, all maximal compact subgroups are
conjugated. If G is semisimple, or more generally reductive, and if K is
a maximal compact subgroup, then the symmetric space G/K is called
the Riemannian symmetric space associated to G. It follows that we
can identify the smooth manifold
S = G/K
with the set of all maximal compact subgroups of G. Remark that
isogenous Lie groups have isometric associated Riemannian symmetric
spaces. In particular, if G is a semisimple linear Lie group, the associated Riemannian symmetric space is the same as that associated to
its identity component G0 or to G/Z(G). We can thus assume that G
is connected and centerless. In this case, as the image of a maximal
compact subgroup by an automorphism of G is again a maximal compact subgroup, we have an action of Aut(G) = Aut(g) by isometries
on S = G/K. Finally we have that the group of isometries of a symmetric space S = G/K of non-compact type without euclidean factors
is Aut(g) where g is the Lie algebra of G.
An important part of our work will be to compute dimensions of
fixed point sets
S α = {x ∈ S | α(x) = x}
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
13
where α ∈ Isom(S) = Aut(g). Assuming that G is connected and
centerless, the fixed point set S α is the Riemannian symmetric space
associated to Gα (recall that we denote by the same letter the automorphism of g and that of G). If A ∈ G we will denote by S A the fixed
point set of the inner automorphism Ad(A). In the case where A is of
finite order, it can be conjugated in the maximal compact subgroup K.
Then the fixed point set of G by Ad(A) is the centralizer of A in G,
that is GAd(A) = CG (A) = {B ∈ G | AB = BA}. A maximal compact
subgroup of CG (A) is CK (A), the centralizer of A in K. So we can
identify S A with CG (A)/CK (A), and we can write:
dim S A = dim CG (A) − dim CK (A).
We refer to [18] for other facts about Riemannian symmetric spaces.
2.6. Lattices and arithmetic groups. A discrete subgroup Γ of a
Lie group G is said to be a lattice if the quotient Γ\G has finite Haar
measure. It is said uniform (or cocompact) if this quotient is compact
and non-uniform otherwise. The Borel density theorem (see [29, Cor.
4.5.6]) says that if G is the group of real points of a connected semisimple algebraic group defined over R, and if a lattice Γ ⊂ G projects
densely into the maximal compact factor of G, then Γ is Zariski-dense
in G. For instance, if G is a connected semisimple algebraic group
defined over Q, then the group GZ is a lattice in GR and thus Zariskidense. The group GZ is the paradigm of an arithmetic group, which
will be defined now.
Let G be a semisimple Lie group with identity component G0 and
Γ ⊂ G a lattice. The lattice Γ is said to be arithmetic if there are a connected algebraic group G defined over Q, compact normal subgroups
K ⊂ G0 , K 0 ⊂ G0R and a Lie group isomorphism
ϕ : G0 /K → G0R /K 0 ,
such that ϕ(Γ) is commensurable to GZ , where Γ and GZ are the images
of Γ ∩ G0 and GZ ∩ G0R in G0 /K and G0R /K (recall that two subgroups
H and H 0 of G are commensurable if their intersection is of finite index
in both subgroups).
We say that the lattice Γ ⊂ G is irreducible if ΓN is dense in G for
every non-compact, closed, normal subgroup N of G0 . The Margulis
arithmeticity theorem (see [27, Ch. IX] and [29, Thm. 5.2.1]) tells us
that in a way, most irreducible lattices are arithmetic.
Theorem 2.1 (Margulis arithmeticity theorem). Let G be the group of
real points of a semisimple algebraic group defined over R and Γ ⊂ G an
irreducible lattice. If G is not isogenous to SO(1, n)×K or SU(1, n)×K
for any compact group K, then Γ is arithmetic.
14
CYRIL LACOSTE
Observe that SO(1, n) × K and SU(1, n) × K have real rank 1, so the
arithmeticity theorem applies to every irreducible lattice in a group of
real rank at least 2.
The definition of arithmeticity can be simplified in some cases. If
G is connected, centerfree and has no compact factors, the compact
subgroup K in the definition must be trivial. Moreover, if Γ is nonuniform and irreducible, then the compact subgroup K 0 is not needed
either (see [29, Cor. 5.3.2]). Under the same assumptions, we can
also assume that the algebraic group G is centerfree, and in this case
the commensurator of GZ in G is GQ and ϕ(Γ) ⊂ GQ . Under the
same hypotheses on G, if Γ is non irreducible, it is almost a product
of irreducible lattices. In fact (see [29, Prop. 4.3.3]), there is a direct
decomposition G = G1 × · · · × Gr such that Γ is commensurable to
Γi × · · · × Γr where Γi = Γ ∩ Gi is an irreducible lattice in Gi .
The rational rank of the arithmetic group Γ, denoted by rkQ Γ, is
by definition the Q-rank of the algebraic group G in the definition of
arithmeticity, and we have
rkQ Γ 6 rkR G.
Note that rkQ Γ = 0 if and only if Γ is cocompact (see [5, Thm. 8.4]).
We refer to [5] and [29] for other facts about lattices and arithmetic
groups.
2.7. Virtual cohomological dimension and proper geometric
dimension. Recall that the virtual cohomological dimension of a virtually torsion-free discrete subgroup Γ is the cohomological dimension
of any torsion-free subgroup Γ0 of finite index of Γ, that is
vcd(Γ) = cd(Γ0 ) = max{n | H n (Γ0 , A) 6= 0 for a certain ZΓ0 -module A}.
If X is a cocompact model for EΓ, we can compute the virtual
cohomological dimension of Γ as
(2.1)
vcd(Γ) = max{n ∈ N | Hnc (X) 6= 0}
where Hnc (X) denotes the compactly supported cohomology of X (see
[9, Cor. VIII.7.6]).
The proper geometric dimension gd(Γ) is the smallest possible dimension of a model for EΓ.
If G is the group of real points of a semisimple algebraic group, K ⊂
G a maximal compact subgroup, S = G/K the associated Riemannian
symmetric space and Γ ⊂ G a uniform lattice of G, S is a model for
EΓ and has dimension vcd(Γ), so we have vcd(Γ) = gd(Γ), that is why
we will be mostly interested in non-uniform lattices.
We will also rule out the case when the adjoint group Gad of g has
real rank 1, in fact we have the following (see [1, Cor. 2.8])
Proposition 2.1. Let G be an algebraic group defined over R and
Γ ⊂ GR a lattice. If rkR G = 1, then vcd(Γ) = gd(Γ).
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
15
For the case of higher real rank, recall that by Margulis arithmeticity
theorem, Γ is arithmetic as long as it is irreducible.
If Γ is non-uniform, Γ \ S is not compact. However Borel and Serre
constructed in [7] a Γ-invariant bordification of S called the Borel-Serre
bordification X which is a cocompact model for EΓ (see [19, Th. 3.2]).
Using their bordification, Borel and Serre proved in [7] the following theorem which links the virtual cohomological dimension and the
rational rank of such an arithmetic lattice.
Theorem 2.2 (Borel-Serre). Let G be a semisimple Lie group, K ⊂ G
a maximal compact subgroup and Γ ⊂ G an arithmetic lattice. Then:
vcd(Γ) = dim(G/K) − rkQ Γ.
In particular:
vcd(Γ) > dim(G/K) − rkR G.
Before moving on, note that we will often in this article consider
groups up to isogeny, and the philosophy behind it is that normal finite
subgroups do not change the dimensions, indeed we have:
Lemma 2.1. Let Γ be an infinite discrete group, and N ⊂ Γ a finite
normal subgroup. Then:
gd(Γ) = gd (Γ/N ) ,
vcd(Γ) = vcd (Γ/N ) .
Proof. For the first equality: if X is a model for EΓ it follows easily
that X N is a model for E(Γ/N ) of dimension lower than those of X,
so:
gd(Γ) > gd (Γ/N ) .
Reciprocally, a model for E(Γ/N ) is also a model for EΓ and we have
the other inequality.
For the second equality, it suffices to recall that vcd(Γ) = cd(Γ0 )
where Γ0 ⊂ Γ is a torsion-free subgroup of finite index, and in this case
Γ0 N/N is a torsion-free subgroup of finite index of Γ/N isomorphic to
Γ0 .
We refer to [9] and [7] for facts about the (virtual) cohomological
dimension and geometric dimension.
2.8. Bredon cohomology. The Bredon cohomological dimension cd(Γ)
is an algebraic counterpart to the proper geometric dimension gd(Γ).
We recall how cd(Γ) is defined and a few of its properties.
Let Γ be a discrete group and F be the family of subgroups of Γ. The
orbit category OF Γ is the category whose objects are left coset spaces
Γ/H with H ∈ F and where the morphisms are all Γ-equivariant maps
between them. An OF Γ-module is a contravariant functor
M : OF Γ → Z-mod
16
CYRIL LACOSTE
to the category of Z-modules. The category of OF Γ-modules, denoted
by Mod-OF Γ, has as objects all the OF Γ-modules and all the natural transformations between them as morphisms. One can show that
Mod-OF Γ is an abelian category and that we can construct projective resolutions on it. The Bredon cohomology of Γ with coefficients
in M ∈ Mod-OF Γ, denoted by H∗F (Γ, M ), is by definition the cohomology associated to the cochain of complexes HomOF Γ (C∗ , M ) where
C∗ → Z is a projective resolution of the functor Z which maps all objects to Z and all morphisms to the identity map. If X is a model for
EΓ, the augmented cellular chain complexes C∗ (X H ) → Z of the fixed
points sets X H for H ∈ F form such a projective (even free) resolution
C∗ (X − ) → Z. Thus we have
HnF (Γ, M ) = Hn (HomOF Γ (C∗ (X − ), M )).
The Bredon cohomological dimension of Γ for proper actions, denoted
by cd(Γ) is defined as
cd(Γ) = sup{n ∈ N | ∃M ∈ Mod-OF Γ : HnF (Γ, M ) 6= 0}.
As we said above, this invariant can be viewed as an algebraic counterpart to gd(Γ). Indeed Lück and Meintrup proved in [26] the following
theorem:
Theorem 2.3 (Lück-Meintrup). If Γ is a discrete group with cd(Γ) >
3, then gd(Γ) = cd(Γ).
We explain now the strategy to prove that vcd(Γ) = cd(Γ), beginning
with some material and definitions. Recall that if G is the group of
real points of a semisimple algebraic group and Γ ⊂ G a lattice, then
the Borel-bordification X is a model for EΓ. Note also that if H is
a finite subgroup of Γ, dim(X H ) = dim(S H ). If we denote F0 the
family of finite subgroups of Γ containing properly the kernel of Γ,
Xsing the subspace of the Borel-bordification X consisting of points
whose stabilizer is stricly larger than the kernel of Γ, and
S = {X H | H ∈ F0 and @H 0 ∈ F0 with X H ( X H },
then we have
Xsing =
[
XH.
X H ∈S
Also every fixed point set X H ∈ S is of the form X α where α ∈ Γ is of
finite order and non-central.
In general, computing cd(Γ) is not an easy task. However, if Γ admits
a cocompact model X for EΓ, then there is a version of the formula
(2.1) for the Bredon cohomological dimension. In fact, from [10, Th.
1.1] we get that
K
cd(Γ) = max{n ∈ N | ∃K ∈ F0 s.t. Hnc (X K , Xsing
) 6= 0}
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
17
K
where X K is the fixed point set of X under K and Xsing
is the subcomK
plex of X consisting of those cells that are fixed by a finite subgroup
of Γ that strictly contains K.
Using the above caracterisations of vcd(Γ) and cd(Γ), one can show
(see [1, Prop. 3.3])
Proposition 2.2. Let G be the group of real points of a semisimple
algebraic group G of real rank at least two, Γ ⊂ G a non-uniform
lattice of G, K ⊂ G a maximal compact subgroup and S = G/K the
associated Riemannian symmetric space. If
(1) dim(X α ) 6 vcd(Γ) for every X α ∈ S, and
(Xsing ) is surjective
(X) → Hvcd(Γ)
(2) the homomorphism Hvcd(Γ)
c
c
then vcd(Γ) = cd(Γ).
Note that in [1] the authors assume that G is connected but this
hypothesis is not needed as the Borel-Serre bordification is still a model
for EΓ if G is not connected (see [19, Th. 3.2]). As dim(X α ) = dim(S α )
we have immediately the following lemma as a corollary of the previous
proposition (see [1, Cor. 3.4])
Lemma 2.2. With the same notations as above, if dim S α < vcd(Γ)
for all α ∈ Γ of finite order and non central, then cd(Γ) = vcd(Γ).
This lemma will be the key argument to prove the Main Theorem.
However, as it is the case in [1], in some cases we will need the following
result (see [1, Cor 3.7])
Lemma 2.3. With the same notations as above, suppose that
(1) dim(S α ) 6 vcd(Γ) for every non-central finite order element
α ∈ Γ,
(2) dim(S α ∩ S β ) 6 vcd(Γ) − 2 for any distinct S α , S β ∈ S, and
(3) for any finite set of non-central finite order elements α1 , . . . , αr
with S αi 6= S αj for i 6= j, dim(S αi ) = vcd(Γ), and such that
CΓ (αi ) is a cocompact lattice in CG (αi ), there exists a rational
flat F in S that intersects S αi in exactly one point and is disjoint
from S αi for i ∈ {2, . . . , n}.
Then vcd(Γ) = cd(Γ).
We refer to [26] and [10] for other facts about Bredon cohomology.
3. Complex simple Lie algebras
In this section we prove the Main Theorem for all complex simple
Lie algebras:
Proposition 3.1. Let g be a complex simple Lie algebra, G = Aut(g)
its group of automorphisms and S the associated Riemannian symmetric space. We assume that rkR G > 2. Then
dim S α < dim S − rkR G
18
CYRIL LACOSTE
for every α ∈ G of finite order and non central. In particular
gd(Γ) = vcd(Γ)
for every lattice Γ ⊂ G.
Recall that the adjoint group Gad is the identity component of G =
Aut(g). It agrees with the group of inner automorphisms, that is Gad =
Inn(g). Note that Gad is centerfree has the same dimension and real
rank as G, and their associated Riemannian symmetric spaces agree.
The quotient Aut(g)/Inn(g) is the group of outer automorphism Out(g)
and can be realized as a subgroup of Aut(g). The group Aut(g) is then
the semi-direct product of Inn(g) and Out(g) (see [16]). Recall also
that if A ∈ Gad , S A is the fixed point set of the inner automorphism
Ad(A).
For further use, note that if ρ ∈ Aut(g) is an involution, then it is
induced by an involution on Gad that will be still denoted ρ. The group
of its fixed points Gρad has Lie algebra gρ and the fixed point set S ρ is
the associated Riemannian symmetric space. In particular, dim S ρ = 0
if Gρad is compact.
The proof of Proposition 3.1 relies on the following lemmas:
Lemma 3.1. Let g be a semisimple Lie algebra such that every element
of Out(g) has order at most 2. Let G be the group of automorphisms
of g and let S be the associated Riemannian symmetric space. If
(3.1)
dim S A < dim S − rkR G and dim S ρ < dim S − rkR G
for all A ∈ Gad non trivial of finite order, and for all involutions ρ ∈ G,
then we also have
dim S α < dim S − rkR G
for every α ∈ G of finite order and non central.
Proof. Every element α ∈ Aut(g) is of the form Ad(A) ◦ ρ where A ∈
Gad and ρ ∈ Out(g) ⊂ Aut(g). We know that ρ is of order at most 2
by hypothesis. Then α2 = Ad(Aρ(A)) is an inner automorphism and
2
we have the inclusion S α ⊂ S α = S Aρ(A) . So, if Aρ(A) is not central
in Gad , then we have
dim S α 6 dim S Aρ(A) < dim S − rkR G.
Note now that if Aρ(A) is central, then it is actually the identity because Gad is centerfree. This means that α2 = Id. In other words, α is
an involution, and we again have
dim S α < dim S − rkR G
by assumption. We have proved the claim.
To check the first part of (3.1) we will use the following:
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
19
Lemma 3.2. Let G be the group of complex points of a semisimple
connected algebraic group and K ⊂ G a maximal compact subgroup.
Suppose that there exists a group H isogenous to a subgroup H 0 of
K such that K/H 0 is an irreducible symmetric space, rkK = rkH,
dim H < dim K − rkR G, and satisfying
dim CH (A) < dim H − rkR G
for all A ∈ H of finite order and non central. Then we have
dim S A < dim S − rkR G
for every A ∈ G of finite order and non central.
Proof. As all maximal compact subgroups are conjugated, we can conjugate such an A ∈ G into K. Since K is connected, A is then contained
in a maximal torus. Since all maximal tori are conjugated, we can conjugate A into any one of them. Since the subgroup H 0 has the same
rank as K, a maximal torus in H 0 is also maximal in K. We can then
assume, up to replacing A by a conjugate element, that A ∈ H 0 .
Taking now into account that G is the group of complex points of
a reductive algebraic group, CG (A) is the complexification of its maximal compact subgroup CK (A) and then its dimension is twice that of
CK (A). As a result we get
dim S A = dim CG (A) − dim CK (A) = dim CK (A)
because S A ' CG (A)/CK (A) as seen in section 2.5.
Similarly we have
dim S = dim(G/K) = dim K.
In particular, the claim follows once we show that
dim CK (A) < dim K − rkR G.
Now, as H and H 0 are isogenous, assume for simplicity that H =
H /F with F a finite normal subgroup of H 0 . We denote by A the class
of A in H 0 /F . As F is finite, we have
0
dim CH (A) = dim CH 0 (A)
and in particular, A is central in H 0 if and only if A is central in H.
Suppose for a moment that A is non central in H 0 , then we write
dim CK (A) 6 dim CH 0 (A)+dim K−dim H = dim CH (A)+dim K−dim H
and by assumption we have
dim CH (A) < dim H − rkR G
and finally
dim CK (A) < dim K − rkR G.
It remains to treat the case that A is central in H 0 but not in K, that
is H 0 ⊂ CK (A) ( K. Since the symmetric space K/H 0 is irreducible,
20
CYRIL LACOSTE
it follows that the identity component of H 0 is a maximal connected
Lie group of K, so dim CK (A) = dim H 0 = dim H, and we have
dim H < dim K − rkR G
by assumption.
In the course of the proof of Proposition 3.1, the subgroups H will
all be classical groups of the forms SO(n) or SU(n) and we will need
the following bounds for the dimension of centralizers in those groups
(see [1] section 5)
1. Let A ∈ SO(n) (n > 3) of finite order and non central, then
(n − 1)(n − 2)
.
2
2. Let A ∈ SU(n) (n > 2) of finite order and non central, then
(3.2)
(3.3)
dim CSO(n) (A) 6
CSU(n) (A) 6 (n − 1)2 .
For more simplicity we will sometimes consider H as a subgroup of
K and denote K/H the symmetric space K/H 0 . For the convenience
of the reader, we summarize in the following table the informations we
need to prove Proposition 3.1 for exceptional Lie algebras.
Gad K
H
dim K dim H rkH = rkK rkR Gad
GC2 G2
SO(4)
14
6
2
2
C
F4 F4
SO(9)
52
36
4
4
E6C E6 U(1) × SO(10)
78
46
6
6
C
E7 E7
SU(8)
133
63
7
7
C
E8 E8
SO(16)
248
120
8
8
Table 10. Exceptional complex simple centerless Lie
groups Gad , maximal compact subgroups K, classical
subgroups H, dimensions and ranks.
We are now ready to launch the proof of Proposition 3.1.
Proof of Proposition 3.1. The second claim follows from Lemma 2.2
because we have
vcd(Γ) > dim S − rkR G
for every lattice Γ ⊂ G by Theorem 2.2. So it suffices to prove the first
claim.
Recall that every complex simple Lie algebra is isomorphic to either
one of the classical algebras sl(n, C), so(n, C) and sp(2n, C) (with conditions on n to ensure simplicity), or to one of the 5 exceptional ones:
gC2 , fC4 , eC6 , eC7 and eC8 .
To prove Proposition 3.1 we will consider all these cases individually.
Classical complex simple Lie algebras:
Let g be a classical complex simple Lie algebra and Γ ⊂ G = Aut(g)
a lattice. From a brief inspection of the table in section 2.3 we obtain
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
21
that, unless g = so(8, C), every outer automorphism of g has order 2.
We will assume that g 6= so(8, C) for a while, treating this case later.
To begin with note that we get from parts 6.1, 6.2 and 6.3 of [1] that
dim S A < dim S − rkR G
for all A ∈ Gad non trivial and of finite order. In other words, the first
part of condition (3.1) in Lemma 3.1 holds. To check the second part
we make use of the classification of local symmetric spaces in [4]. For
instance if g = sl(n, C) with n > 3 because of our assumption on the
rank, and if ρ ∈ G is an involution, then the Lie algebra gρ is isomorphic
to either a Lie algebra whose adjoint group is compact or to one of the
follows: so(n, C), s(gl(k, C) ⊕ gl(n − k, C)), sp(n, C), sl(n, R), su(p, n −
p) and sl( n2 , H), where sp(n, C) and sl( n2 , H) only appear if n is even.
The associated Riemannian symmetric space S ρ is obtained by taking
the quotient of the adjoint group by a maximal compact subgroup, for
example in the case of so(n, C) it is PSO(n, C)/PSOn . The Lie algebra
gρ for which dim S ρ is maximal is s(gl(1, C) ⊕ gl(n − 1, C)) for n 6= 4
(where dim S ρ = (n−1)2 ), and sp(4, C) for n = 4 (where dim S ρ = 10).
In all these cases we have
dim S ρ < (n2 − 1) − (n − 1) = dim S − rkR G.
We get then by Lemma 3.1 that the first claim of Proposition 3.1 holds.
The cases of sp(2n, C) and so(n, C) for n 6= 8 are similar, we leave the
details to the reader.
Now we treat the case of the Lie algebra so(8, C). Its group of
complex outer automorphisms is isomorphic to the symmetric group
S3 and contains an order 3 element τ called triality. (See section 1.14
in [34] for an interpretation of triality in terms of octonions.) The
group Out(so(8, C)) of real outer automorphisms is then isomorphic to
S3 ×Z/2Z where the second factor corresponds to complex conjugation.
Consequently the only order 3 outer automorphisms are τ and τ −1 and
those of order 6 are their compositions with complex conjugation. If
ρ ∈ Aut(so(8, C)) is of order 6 and α = Ad(A) ◦ ρ, then α3 is the
composition of an inner automorphism and ρ3 . As ρ3 is of order 2 and
3
we have the inclusions S α ⊂ S α , we can consider α3 instead of α and
we just have to treat the cases when ρ is of order 2 or 3.
If ρ is of order 2 we apply the same method than for other classical
simple Lie algebras, using the classification of local symmetric spaces.
It remains to treat the case when ρ = τ the triality automorphism (or
its inverse). In this case α = Ad(A) ◦ τ is a complex automorphism.
Proceeding like in the proof of Lemma 3.1, α3 is an inner automorphism and the result follows if it is non trivial. If α3 = 1 then α
belongs to the set Aut3C (so(8, C)) of complex automorphisms of order
3. A result of Gray and Wolf (see [15, Thm. 5.5]) says that if ∼i is
the equivalence relation of conjugation by an inner automorphism in
22
CYRIL LACOSTE
Aut3C (so(8, C)), then Aut3C (so(8, C))/ ∼i contains, besides the classes
of inner automorphisms, four other classes: those of τ and τ −1 and two
others represented by order 3 automorphisms τ 0 and τ 0−1 . The Lie algebra of the fixed point set of triality is the exceptional Lie algebra gC2 ,
and those of the fixed point set of τ 0 is isomorphic to the Lie algebra
sl(3, C). In both cases we have
dim S α < dim S − rkR G
and Proposition 3.1 holds for g = so(8, C).
Lie algebra gC2 :
Here we consider the simple exceptional Lie algebra g = gC2 . Its outer
automorphism group is of order 2 and its adjoint group is Gad = GC2 ,
a connected algebraic group of real rank 2 and complex dimension 14.
The compact group G2 is a maximal compact subgroup of GC2 .
The group G2 contains a subgroup H isomorphic to SO(4), fixed by
an involution of G2 which extends to a conjugation of GC2 , giving the
split real form G2(2) (see section 1.10 of [34] for an explicit description). Then G2 /SO(4) is an irreducible symmetric space, rk(G2 ) =
rk(SO(4)) = 2 and
dim H + rkR Gad = 6 + 2 = 8 < 14 = dim K.
Moreover if A ∈ H is of finite order and non central in H ' SO(4) we
have by inequality (3.2)
dim CH (A) + rkR Gad 6 3 + 2 = 5 < 6 = dim H.
So by Lemma 3.2, the first part of (3.1) in Lemma 3.1 holds.
To check the second part we have to list the local symmetric spaces
associated to an involution ρ ∈ Aut(gC2 ). By the classification of Berger
in [4], the only non compact cases are when gρ is isomorphic to sl(2, C)⊕
sl(2, C) or g2(2) . The associated Riemannian symmetric spaces are
S ρ = (PSL(2, C) × PSL(2, C))/(PSU2 × PSU2 ) and G2(2) /SO(4) and
we have in both cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 3.1 holds for G = Aut(gC2 ).
Lie algebra fC4 :
We proceed like previously for the simple Lie algebra g = fC4 , with
|Out(fC4 )| = 2, Gad = F4C of maximal compact subgroup K = F4 ,
rkR Gad = 4 and dim K = 52. We know that there exists a subgroup
H ⊂ K isogenous to SO(9), with rkH = rkK = 4, and such that K/H
is an irreducible symmetric space. In addition to that
dim H + rkR Gad = 36 + 4 = 40 < 52 = dim K
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
23
and if A ∈ H is of finite order and non central in H, we have
dim CH (A) + rkR Gad 6 28 + 4 = 32 < 36 = dim H
by inequality (3.2).
So by Lemma 3.2, the first part of (3.1) in Lemma 3.1 holds.
Then by the classification of local symmetric spaces, the ones we
have to study is those when gρ is isomorphic to sp(6, C) ⊕ sp(2, C),
so(9, C), f4(4) or f4(−20) , in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 3.1 holds for G = Aut(fC4 ).
Lie algebra eC6 :
For the algebra g = eC6 , the outer automorphism group is a product
of two groups of order 2, Gad = E6C of maximal compact subgroup
K = E6 , rkR Gad = 6 and dim K = 78. We know there exists H ⊂
K isogenous to U(1) × SO(10), with rkH = rkK = 6, K/H is an
irreducible symmetric space and we have the following
dim H + rkR Gad = 46 + 6 = 52 < 78 = dim K
and if A ∈ H is of finite order and non central in H, by inequality (3.2)
we have
dim CH (A) + rkR Gad 6 1 + 36 + 6 = 43 < 46 = dim H.
So by Lemma 3.2, the first part of (3.1) in Lemma 3.1 holds.
Then by the classification of local symmetric spaces, the ones we have
to study is those when gρ is isomorphic to sp(8, C), sl(6, C) ⊕ sl(2, C),
so(10, C) ⊕ so(2, C), fC4 , e6(6) , e6(2) , e6(−14) or e6(−26) , in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 3.1 holds for G = Aut(eC6 ).
Lie algebra eC7 :
Consider now the simple Lie algebra g = eC7 , of order 2 outer automorphism group, and of adjoint group Gad = E7C , whose compact
maximal subgroup is K = E7 , rkR Gad = 7 and dim K = 133. We know
there exists H ⊂ K isogenous to SU(8) with rkH = rkK = 7 and K/H
is an irreducible symmetric space. We have the inequality
dim H + rkR Gad = 63 + 7 = 70 < 133 = dim K
and if A ∈ H is of finite order and non central in H, by inequality (3.3)
we have
dim CH (A) + rkR Gad 6 49 + 7 < 63 = dim H.
So by Lemma 3.2, the first part of (3.1) in Lemma 3.1 holds.
24
CYRIL LACOSTE
Then by the classification of local symmetric spaces, the ones we have
to study is those when gρ is isomorphic to sl(8, C), so(12, C) ⊕ sp(2, C),
eC6 ⊕ so(2, C), e7(7) , e7(5) or e7(−25) , in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 3.1 holds for G = Aut(eC7 ).
Lie algebra eC8 :
The last exceptional Lie algebra is g = eC8 , again its outer automorphism group is of order 2, its adjoint group is Gad = E8C , of maximal
compact subgroup K = E8 , rkR Gad = 8 and dim K = 248. We know
there exists H ⊂ K isogenous to SO(16), with rkH = rkK = 8 and
K/H is an irreducible symmetric space. We have also the inequality
dim H + rkR Gad = 120 + 8 = 128 < 248 = dim K
and if A ∈ H is of finite order and non central in H, by inequality (3.2)
we have
dim CH (A) + rkR Gad 6 105 + 8 < 120 = dim H.
So by Lemma 3.2, the first part of (3.1) in Lemma 3.1 holds.
Then by the classification of local symmetric spaces, the ones we have
to study is those when gρ is isomorphic to so(16, C), eC7 ⊕ sp(2, C), e8(8)
or e8(−24) , in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 3.1 holds for G = Aut(eC8 ) and it
concluded its proof.
4. Real simple Lie algebras
We will in this section extend the previous proposition to the real
simple Lie algebras. They are the real forms of the complex ones studied in the previous section. The ideas of the proof are similar to those of
the complex case, although we face some additional difficulties. Maybe
the reader can skip this section in a first reading.
Proposition 4.1. Let g be a real simple Lie algebra, G = Aut(g) its
group of automorphisms and S the associated Riemannian symmetric
space. Then
gd(Γ) = vcd(Γ)
for every lattice Γ ⊂ G. Moreover
dim S α 6 dim S − rkR G
for every α ∈ G of finite order and non central.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
25
We will again use Lemma 3.1, but in the case of exceptional real
simple Lie algebras, we cannot use Lemma 3.2 to establish inequalities
of the form
dim S A < dim S − rkR G
for A in the adjoint group Gad of g. The difficulty is that the dimension
of Gad is not anymore twice that of a maximal compact subgroup. To
some extent, we will bypass this problem using the following lemma:
Lemma 4.1. Let G be a connected Lie group which is the group of
real points of a semisimple algebraic group defined over R, and K ⊂ G
a maximal compact subgroup. Suppose there exists a subgroup G ⊂ G
such that G/G is an irreducible symmetric space and whose compact
maximal subgroup K ⊂ K has the same rank as K. Let S = G/K and
S = G/K be the associated Riemannian symmetric spaces. If we have
dim S < dim S − rkR G,
and
A
dim S < dim S − rkR G
for every A ∈ K of finite order non central, then we also have
dim S A < dim S − rkR G
for every A ∈ G of finite order and non central.
Proof. As in the proof of Lemma 3.2 we can conjugate such an A into
a maximal torus in K.
If A is central in G: as G/G is irreducible, we have that the identity
component of G is a maximal connected Lie subgroup of G. It follows
thus from G ⊂ CG (A) ( G that the Riemannian symmetric spaces of
CG (A) and G are the same, that is S A = S. Then the result follows
from the assumption
dim S < dim S − rkR G.
Suppose now that A is not central in G, then we have
dim S − dim S A > dim S − dim S
A
A
as S = S ∩ S A . Then the result follows because
A
dim S − dim S > rkR G
by assumption.
In all cases if interest, the group G will be a classical group. We will
A
use the following inequalities to majorate dim S (see [1] sections 6.5,
6.6 and 6.8):
1. Let A ∈ SU(p, q) (p 6 q, p + q > 3) of finite order non central,
and S = SU(p, q)/S(U(p) × U(q)) the associated symmetric space, then
(4.1)
dim S A 6 2p(q − 1).
26
CYRIL LACOSTE
2. Let A ∈ Sp(p, q) (p 6 q, p + q > 3) of finite order non central, and
S = Sp(p, q)/(Sp(p) × Sp(q)) the associated symmetric space, then
(4.2)
dim S A 6 4p(q − 1).
3. Let A ∈ SO∗ (2n) (n > 2) of finite order non central, et S =
SO∗ (2n)/U(n) the associated symmetric space, then
(4.3)
dim S A 6 n2 − n − 2(n − 1).
The tables below list exceptional real simple Lie groups, the subgroups G we will use and the informations we need to know for the
proof of Proposition 4.1. Note that for more simplicity, the compact
maximal subgroups K are given up to isogeny.
Gad
K
G
K
E6(6)
Sp(4)
Sp(2, 2)
Sp(2) × Sp(2)
E6(2)
SU(6) × SU(2)
SO∗ (10) × SO(2)
U(5) × SO(2)
∗
E6(−14) SO(10) × SO(2) SO (10) × SO(2)
U(5) × SO(2)
E6(−26)
F4
Sp(1, 3)
Sp(1) × Sp(3)
E7(7)
SU(8)
E6(2) × SO(2)
SU(6) × SU(2) × SO(2)
E7(−5) SO(12) × SU(2)
SU(4, 4)
S(U(4) × U(4))
E7(−25)
E6 × SO(2)
SU(2, 6)
S(U(2) × U(6))
∗
E8(8)
SO(16)
SO (16)
U(8)
E8(−24)
E7 × SU(2)
SO∗ (16)
U(8)
G2(2)
SO(4)
SL(2, R) × SL(2, R)
SO(2) × SO(2)
F4(4)
Sp(3) × Sp(1)
SO(4, 5)
S(O(4) × O(5))
Table 11. Real exceptional simple centerless Lie groups
Gad , certain classical subgroups G ⊂ Gad and the respective maximal compact subgroups
Gad
dim S dim S rkK = rkK rkR Gad
E6(6)
42
16
4
6
E6(2)
40
20
6
4
E6(−14)
32
20
6
2
E6(−26)
26
12
4
2
E7(7)
70
40
7
7
E7(−5)
64
32
7
4
E7(−25)
54
24
7
3
E8(8)
128
56
8
8
E8(−24) 112
56
8
4
G2(2)
8
4
2
2
F4(4)
28
20
4
4
Table 12. With the same notations as in Table 11, dimensions of the Riemannian symmetric spaces associated
to Gad and G, together with the ranks of K, K and Gad
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
27
We are now ready to prove Proposition 4.1.
Proof of Proposition 4.1. Recall that the first claim holds when the adjoint group has real rank 1 by Proposition 2.1, that is when g is isomorphic to sl(2, R) ∼
= sp(2, R), sl(2, H), so∗ (4), so∗ (6), so(1, n), su(1, n),
sp(1, n) and f4(−20) . The second claim is also true because if α ∈ Aut(g)
is of finite order and non central, then S α is a strict submanifold of S
so we have
dim S α 6 dim S − 1.
We suppose from now on that rkR G > 2. By inspection of the
table of outer automorphisms in section 2.3, we see that every outer
automorphism of g has order 2, except if g = so(p, p) with p > 4 even.
As in the proof of Proposition 3.1 We will again do a case-by-base
analysis.
Classical real simple Lie algebras other that sl(n, R) and
so(p, q):
We start dealing with the classical Lie algebras su(p, q), sl(n, H),
sp(2n, R) and so∗ (2n). Note that we rule out su(2, 2) ∼
= so(2, 4) and
∼
sp(4, R) = so(2, 3).
We use again Lemma 3.1. We want then to establish
dim S A < dim S − rkR G and dim S ρ < dim S − rkR G
for every A in the adjoint group Gad of finite order and non central
and for every involution ρ ∈ G = Aut(g). The first condition holds by
the computations in sections 6.4 to 6.7 of [1]. Using the classification
of local symmetric spaces, we can check the second condition as we
did in the complex case. For instance, if g = sp(2n, R) with n >
3, then gρ is either compact or isomorphic to one of the following:
sp(2k, R) ⊕ sp(2(n − k), R), u(k, n − k), gl(n, R) or sp(n, C), the last
case only appearing if n is even. The Lie algebra gρ for which dim S ρ
is maximal is sp(2(n − 1), R) ⊕ sp(2, R), for which we have
dim S ρ = n2 − n + 2 < n2 = dim S − rkR G.
Hence by Lemma 3.1 and Lemma 2.2, Proposition 4.1 holds for G =
Aut(sp(2n, R)). The cases of su(p, q), sl(n, H) and so∗ (2n) are similar.
Lie algebras sl(n, R) and so(p, q) The remaining classical cases are
sl(n, R) and so(p, q). If p > 6 is even, then Out(so(p, p)) is isomorphic
to D4 so every outer automorphism ρ has order 2 or 4. The only case
where we have order 3 outer automorphisms is so(4, 4), as Out(so(4, 4))
is isomorphic to S4 .
As already noted in [1], where the argument for lattices in SL(n, R)
and SO(p, q) was more involved than in the other cases, we face the
28
CYRIL LACOSTE
problem that there exists α ∈ Aut(g) such that
dim S α = dim S − rkR G.
Our next goal is to characterize the automorphisms α for which this
happens.
Lemma 4.2. If g = sl(n, R) with n 6= 4 or so(p, q) with p + q = n and
(p, q) 6= (3, 3), and α ∈ Aut(g), then
dim S α 6 dim S − rkR G
with equality if and only if n is odd and S α = S A with A conjugated to
!
Qn−1,1
−In−1 0
=
0
1
in PSL(n, R) or PSO(p, q), or n is even and α is conjugated to the
outer automorphism corresponding to the conjugation by Qn−1,1 .
By abusing notations we will still denote S A the fixed point set by
the automorphism corresponding to the conjugation by A with A conjugated to Qn−1,1 even if it is not an inner automorphism.
Proof. We begin with the case g 6= so(4, 4) and we use the same strategy that for the proof of Lemma 3.1. An automorphism α of g is the
composition of an inner automorphism and an outer automorphism ρ
of order 2 or 4. If ρ has order 4 then α2 is the composition of an inner
automorphism and ρ2 which is of order 2, and we have the inclusion
2
S α ⊂ S α so we can replace α by α2 and it will suffice to consider the
outer automorphisms of order 2. Then if ρ has order 2, α2 is an inner
automorphism, that is α2 = Ad(A) with A ∈ Gad . If A is not trivial,
by the computations in sections 7 and 8 in [1] we have
2
dim S α 6 dim S A 6 dim S − rkR G
with equality in the last inequality if and only if n is odd and A conjugated by Qn−1,1 . The first inequality is an equality if and only if
2
S α = S A . We have proved the claim if A is not trivial. If A is trivial,
then α is an involution, so we use the classification of local symmetric
spaces. For instance if g = sl(n, R) with n 6= 4, the non-compact associated
isotropy
algebra h = gα are so(k, n−k), s(gl(k, R)⊕gl(n−k, R)),
gl n2 , C or sp(n, R), the last two cases only appearing if n is even. In
all theses cases we have
dim S α 6 dim S − rkR G
with equality if and only if h = s(gl(1, R) ⊕ gl(n − 1, R)), which corresponds to an automorphism α conjugated to the inner automorphism
Ad(Qn−1,1 ) if n is odd or to the outer automorphism of conjugation by
Qn−1,1 if n if even.
It remains to consider the case g = so(4, 4). In this case the group
of outer automorphism is isomorphic to the symmetric group S4 so we
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
29
can have elements of order 2, 3 or 4. If ρ is an outer automorphism of
order 2 or 4, and α = Ad(A) ◦ τ , we apply the same method using the
classification of local symmetric spaces and we see that
dim S α 6 dim S − rkR G
with equality if and only if α is an outer automorphism corresponding
to the conjugation by a matrix conjugated to Q7,1 . If ρ is of order 3 and
α = Ad(A) ◦ τ then α3 is inner and we have just to treat the case where
it is trivial, that is α is of order 3. Then its complexification αC is an
order 3 complex automorphism of gC = so(8, C), a case already treated
C
in the previous section. We know that the fixed point set (gC )α is
C
isomorphic to gC2 , sl(3, C) or is compact. As gα is a real form of (gC )α ,
it is isomorphic to g2(2) , sl(3, R), su(2, 1), s(u(dp , dq ) ⊕ o(sp , sq )) with
2dp + sp = 4 and 2dq + sq = 4 or is compact. In all these cases we have
dim S α < dim S − rkR G
so we have proved the claim.
Let us assume for a while that g 6= sl(3, R) and g 6= sl(4, R) ∼
=
so(3, 3). We will conclude that vcd(Γ) = gd(Γ) using Lemma 2.3. The
first condition in the said lemma holds by Lemma 4.2. To check the
second condition, take S α and S β maximal and distinct. We want to
establish
dim(S α ∩ S β ) 6 vcd(Γ) − 2.
First remark that by maximality S α and S β are not contained in each
other. If one of them is not of the form S A with A conjugated to
Qn−1,1 , let us say S α , then dim S α 6 vcd(Γ) − 1 and S α ∩ S β is a
strict submanifold of S α so the result holds. If we have S α = S A and
S β = S B with A and B conjugated to Qn−1,1 , then we refer to the
computations in the proofs of Lemma 7.2 and Lemma 8.4 in [1]. The
proof of the third point is the same as for Lemma 7.5 and Lemma 8.8
in [1]. Note that in [1] the authors consider only inner automorphisms,
so the case n odd, but their argument also works without modifications
of any kind for n even.
It must be enlighted why the argument we just gave fails for sl(3, R)
and sl(4, R) ∼
= so(3, 3). For sl(3, R), the second condition of Lemma
2.3 does not hold anymore. In the case that sl(4, R) ∼
= so(3, 3), the
conclusion of Lemma 4.2 does not apply, because we have that
dim S α = dim S − rkR G
when α is either the conjugation by Q5,1 in PSO(3, 3) or the conjugation by Q3,1 in PSL(4, R). These two are not conjugated, as the
conjugation by Q5,1 in PSO(3, 3) corresponds in PSL(4, R) to an outer
automorphism whose fixed point set is isomorphic to PSp(4, R).
However the proof of Lemma 7.6 in [1] concerning lattices in SL(3, R)
can be adapted to Aut(sl(3, R)) and Aut(sl(4, R)). In fact it can be
30
CYRIL LACOSTE
adapted to Aut(sl(n, R)) for all n > 3, because a lattice in PSL(n, R) of
Q-rank n−1 can be conjugated to a lattice commensurable to PSL(n, Z)
(see the classification of arithmetic groups of classical groups in Section
18.5 in [29]).
As a result Proposition 4.1 holds for all real classical simple Lie
algebras.
Lie algebra e6(6)
Here we consider the simple exceptional Lie algebra g = e6(6) . Its
outer automorphism group is of order 2 and its adjoint group is Gad =
E6(6) , which is the group of real points of an algebraic group of real rank
6. This group contains a maximal compact subgroup K isogenous to
Sp(4). We will use Lemma 4.1 to check the first condition of Lemma
3.1.
The group Gad = E6(6) contains a subgroup G isogenous to Sp(2, 2)
whose maximal compact subgroup is K = Sp(2) × Sp(2). We see in [4]
that E6(6) /Sp(2, 2) is an irreducible symmetric space, furthermore we
have rk(K) = rk(K) = 4 and
dim S = 16 < 42 − 6 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.2)
A
dim S 6 8 < 16 − 6 = dim S − rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e6(6) ). By the classification of
Berger in [4], the only non compact cases are when gρ is isomorphic to
sp(2, 2), sp(8, R), sl(6, R) ⊕ sl(2, R), sl(3, H) ⊕ su(2), so(5, 5) ⊕ so(1, 1)
or f4(4) . We have in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e6(6) ).
Lie algebra e6(2)
Here we consider the simple exceptional Lie algebra g = e6(2) . Its
outer automorphism group is of order 2 and its adjoint group is Gad =
E6(2) , which is the group of real points of an algebraic group of real rank
4. This group contains a maximal compact subgroup K isogenous to
SU(6) × SU(2). We will use Lemma 4.1 to check the first condition of
Lemma 3.1.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
31
The group Gad = E6(2) contains a subgroup G isogenous to SO∗ (10)×
SO(2) whose maximal compact subgroup is K = U(5) × SO(2). We see
in [4] that E6(2) /(SO∗ (10) × SU(2)) is an irreducible symmetric space,
furthermore we have rk(K) = rk(K) = 6 and
dim S = 20 < 40 − 4 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.3)
A
dim S 6 12 < 20 − 4 = dim S − rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e6(2) ). By the classification of
Berger in [4], the only non compact cases are when gρ is isomorphic
to sp(1, 3), sp(8, R), su(2, 4) ⊕ su(2), su(3, 3) ⊕ sl(2, R), so(4, 6) ⊕ so(2),
so∗ (10) ⊕ so(2) or f4(4) . We have in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e6(2) ).
Lie algebra e6(−14)
Here we consider the simple exceptional Lie algebra g = e6(−14) .
Its outer automorphism group is of order 2 and its adjoint group is
Gad = E6(−14) , which is the group of real points of an algebraic group
of real rank 2. This group contains a maximal compact subgroup K
isogenous to SO(10) × SO(2). We will use Lemma 4.1 to check the first
condition of Lemma 3.1.
The group Gad = E6(−14) contains a subgroup G isogenous to SO∗ (10)×
SO(2) whose maximal compact subgroup is K = U(5) × SO(2). We
see in [4] that E6(−14) /(SO∗ (10) × SU(2)) is an irreducible symmetric
space, furthermore we have rk(K) = rk(K) = 6 and
dim S = 20 < 32 − 2 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.3)
A
dim S 6 12 < 20 − 2 = dim S − rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e6(−14) ). By the classification of
32
CYRIL LACOSTE
Berger in [4], the only non compact cases are when gρ is isomorphic to
sp(2, 2), su(2, 4) ⊕ su(2), su(1, 5) ⊕ sl(2, R), so(2, 8) ⊕ so(2), so∗ (10) ⊕
so(2) or f4(−20) . We have in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e6(−14) ).
Lie algebra e6(−26)
Here we consider the simple exceptional Lie algebra g = e6(−26) .
Its outer automorphism group is of order 2 and its adjoint group is
Gad = E6(−26) , which is the group of real points of an algebraic group
of real rank 2. This group contains a maximal compact subgroup K
isogenous to F4 . We will use Lemma 4.1 to check the first condition of
Lemma 3.1.
The group Gad = E6(−26) contains a subgroup G isogenous to Sp(1, 3)
whose maximal compact subgroup is K = Sp(1) × Sp(3). We see in
[4] that E6(−26) /Sp(1, 3) is an irreducible symmetric space, furthermore
we have rk(K) = rk(K) = 4 and
dim S = 12 < 26 − 2 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.2)
A
dim S 6 8 < 12 − 2 = dim S − rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e6(−26) ). By the classification of
Berger in [4], the only non compact cases are when gρ is isomorphic to
sp(1, 3), sl(3, H) ⊕ sp(1), so(1, 9) ⊕ so(1, 1), or f4(−20) . We have in all
cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e6(−26) ).
Lie algebra e7(7)
Here we consider the simple exceptional Lie algebra g = e7(7) . Its
outer automorphism group is of order 2 and its adjoint group is Gad =
E7(7) , which is the group of real points of an algebraic group of real rank
7. This group contains a maximal compact subgroup K isogenous to
SU(8). We will use Lemma 4.1 to check the first condition of Lemma
3.1.
The group Gad = E7(7) contains a subgroup G isogenous to E6(2) ×
SO(2) whose maximal compact subgroup is K = SU(6)×SU(2)×SO(2).
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
33
We see in [4] that E7(7) /(E6(2) × SO(2)) is an irreducible symmetric
space, furthermore we have rk(K) = rk(K) = 7 and
dim S = 40 < 70 − 7 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
the results about e6(2)
A
dim S − dim S > 20 − 12 = 8 > rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e7(7) ). By the classification of
Berger in [4], the only non compact cases are when gρ is isomorphic
to su(4, 4), sl(8, R), sl(4, H), so(6, 6) ⊕ sl(2, R), so∗ (12) ⊕ sp(1), e6(6) ⊕
so(1, 1) or e6(2) ⊕ so(2). We have in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e7(7) ).
Lie algebra e7(−5)
Here we consider the simple exceptional Lie algebra g = e7(−5) . Its
outer automorphism group is trivial so G = Aut(g) is equal to the
adjoint group Gad = E7(7) , which is the group of real points of an
algebraic group of real rank 4. Thus we only have to check the first
condition of Lemma 3.1 and we will again use Lemma 4.1.
The group Gad = E7(−5) contains a maximal compact subgroup K
isogenous to SO(12) × SU(2). It also contains a subgroup G isogenous
to SU(4, 4) whose maximal compact subgroup is K = S(U(4) × U(4)).
We see in [4] that E7(−5) /SU(4, 4) is an irreducible symmetric space,
furthermore we have rk(K) = rk(K) = 7 and
dim S = 32 < 64 − 4 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.1)
A
dim S 6 24 < 32 − 4 = dim S − rkR Gad .
So by Lemma 4.1 and Lemma 3.1, Proposition 4.1 holds for G =
Aut(e7(−5) ).
Lie algebra e7(−25)
Here we consider the simple exceptional Lie algebra g = e7(−25) .
Its outer automorphism group is of order 2 and its adjoint group is
34
CYRIL LACOSTE
Gad = E7(−25) , which is the group of real points of an algebraic group
of real rank 3. This group contains a maximal compact subgroup K
isogenous to E6 × SO(2). We will use Lemma 4.1 to check the first
condition of Lemma 3.1.
The group Gad = E7(−25) contains a subgroup G isogenous to SU(2, 6)
whose maximal compact subgroup is K = S(U(2) × U(6)). We see in
[4] that E7(−25) /SU(2, 6) is an irreducible symmetric space, furthermore
we have rk(K) = rk(K) = 7 and
dim S = 24 < 54 − 3 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.1)
A
dim S 6 20 < 24 − 3 = dim S − rkR Gad .
Lemma 4.1 applies and shows that the first condition of Lemma 3.1
holds.
To check the second condition we list the local symmetric spaces
associated to an involution ρ ∈ Aut(e7(−25) ). By the classification of
Berger in [4], the only non compact cases are when gρ is isomorphic to
su(2, 6), sl(4, H), so(2, 10) ⊕ sl(2, R), so∗ (12) ⊕ sp(1), e6(−14) ⊕ so(2) or
e6(−26) ⊕ so(1, 1). We have in all cases
dim S ρ < dim S − rkR G.
So by Lemma 3.1, Proposition 4.1 holds for G = Aut(e7(−25) ).
Lie algebra e8(8)
Here we consider the simple exceptional Lie algebra g = e8(8) . Its
outer automorphism group is trivial so G = Aut(g) is equal to the
adjoint group Gad = E8(8) , which is the group of real points of an
algebraic group of real rank 8. Thus we only have to check the first
condition of Lemma 3.1 and we will again use Lemma 4.1.
The group Gad = E8(8) contains a maximal compact subgroup K
isogenous to SO(16). It also contains a subgroup G isogenous to
SO∗ (16) whose maximal compact subgroup is K = U(8). We see in
[4] that E8(8) /SO(16) is an irreducible symmetric space, furthermore
we have rk(K) = rk(K) = 8 and
dim S = 56 < 128 − 8 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.3)
A
dim S 6 42 < 56 − 8 = dim S − rkR Gad .
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
35
So by Lemma 4.1 and Lemma 3.1, Proposition 4.1 holds for G =
Aut(e8(8) ).
Lie algebra e8(−24)
Here we consider the simple exceptional Lie algebra g = e8(−24) . Its
outer automorphism group is trivial so G = Aut(g) is equal to the
adjoint group Gad = E8(−24) , which is the group of real points of an
algebraic group of real rank 4. Thus we only have to check the first
condition of Lemma 3.1 and we will again use Lemma 4.1.
The group Gad = E8(−24) contains a maximal compact subgroup K
isogenous to E7 × SU(2). It also contains a subgroup G isogenous to
SO∗ (16) whose maximal compact subgroup is K = U(8). We see in [4]
that E8(−24) /SO(16) is an irreducible symmetric space, furthermore we
have rk(K) = rk(K) = 8 and
dim S = 56 < 112 − 4 = dim S − rkR Gad
where S and S are the Riemannian symmetric spaces associated to
respectively Gad and G.
Moreover, if A ∈ K is of finite order and non central, we get from
inequality (4.3)
A
dim S 6 42 < 56 − 4 = dim S − rkR Gad .
So by Lemma 4.1 and Lemma 3.1, Proposition 4.1 holds for G =
Aut(e8(−24) ).
Lie algebra g2(2)
Here we consider the simple exceptional Lie algebra g = g2(2) . Its
outer automorphism group is trivial, so the group G = Aut(g) equals
the adjoint group Gad = G2(2) , which is the group of real points of
an algebraic group of real rank 2. Thus we only have to check the
conditions of Lemma 3.2.
The group G2(2) contains a maximal compact subgroup K isomorphic
to SO(4). It also contains a subgroup G isogenous to SL(2, R)×SL(2, R)
whose maximal compact subgroup is K = SO(2) × SO(2). We see in
[4] that G2(2) /(SL(2, R) × SL(2, R)) is an irreducible symmetric space,
furthermore we have rk(K) = rk(K) = 2 and
dim S = 4 < 8 − 2 = dim S − rkR G
where S and S are the Riemannian symmetric spaces associated to
respectively G = Gad and G.
Moreover, if A ∈ K is of finite order and non central, we have
A
dim S 6 2 = 4 − 2 = dim S − rkR G.
36
CYRIL LACOSTE
The equality case in the last inequality happens when A is conjugated
to a matrix of the form:
!
±I2 0
0 Rθ
with
!
cos(θ) − sin(θ)
Rθ =
.
sin(θ) cos(θ)
Assume that A is of this form and that the first block is I2 , we will
prove directly that
dim S A < dim S − rkR G.
First of all, CK (A) = CSO(4) (A) = SO(2) × SO(2) is of dimension 2.
To study CG (A), we have to know to which element of G = G2(2)
this matrix corresponds to. Recall that G2(2) is the group of automorphisms of the (non associative) algebra O0 of split octonions, which is
of dimension 8 over R, and equiped with a quadratic form of signature
(4,4) (see section 1.13 of in [34]). We can decompose O0 into the direct
sum H ⊕ He04 where H = Vect{1, e1 , e2 , e3 } is the quaternion algebra.
So G is a subgroup of the special orthogonal group SO(3, 4) which
preserves the standard form of signature (4,4) over R8 and fixes 1.
The maximal compact subgroup K corresponds to the stabilizer of
H, meaning the elements α ∈ G such that α(H) = H. Automatically
we have α(He04 ) = He04 as this is the orthogonal of H. K is isomorphic
to SO(4) via the isomorphism who sends α to its restriction to He04 .
Consequently, the matrix A we consider corresponds to the matrix of
the restriction of an element α ∈ K to He04 . This element α is entirely
determined by the matrix A, indeed for example we have:
α(e1 e04 ) = α(e1 )α(e04 ) = α(e1 )e04 = e1 e04
so we deduce α(e1 ) = e1 . Similarly we find:
α(e2 ) = cos(θ)e2 + sin(θ)e3 ,
α(e3 ) = − sin(θ)e2 + cos(θ)e3 .
Knowing α(1) = 1, we have completely described α. The matrix of
SO(3, 4) which corresponds to is:
1 0 0 0
0 R
0 0
θ
Ae =
.
0 0 I2 0
0 0 0 Rθ
e so:
Then we remark that CG (A) ⊂ CSO(3,4) (A)
e = dim S(O(1, 2) × U(1, 1)) = 6.
dim CG (A) 6 dim CSO(3,4) (A)
Finally:
dim S A 6 6 − 2 = 4 < 6 = dim S − rkR G.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
37
Thus we have that
dim S A < dim S − rkR G
for every A ∈ G of finite order and non central. Then by Lemma 2.2,
Proposition 4.1 holds for g = g2(2) .
Lie algebra f4(4)
Here we consider the simple exceptional Lie algebra g = f4(4) . Its
outer automorphism group is trivial, so the group G = Aut(g) equals
the adjoint group Gad = F4(4) , which is the group of real points of
an algebraic group of real rank 4. Thus we only have to check the
conditions of Lemma 3.2.
The group F4(4) contains a maximal compact subgroup K isogenous
to Sp(3) × Sp(1). It also contains a subgroup G isogenous to SO(4, 5)
whose maximal compact subgroup is K = S(O(4) × O(5)). We see in
[4] that F4(4) /SO(4, 5) is an irreducible symmetric space, furthermore
we have rk(K) = rk(K) = 4 and
dim S = 20 < 28 − 4 = dim S − rkR G
where S and S are the Riemannian symmetric spaces associated to
respectively G = Gad and G.
Moreover, if A ∈ K is of finite order and non central, we have by the
computations in section 8 of [1]
A
dim S 6 16 = 20 − 4 = dim S − rkR G.
The equality case in the last inequality happens when A is conjugated
to the matrix:
!
−I8 0
∈ SO(4) × SO(5).
0 1
Assuming that A is of this form, the conjugation by A is an involutive
automorphism of G = Gad = F4(4) so the quotient G/GA is a symmetric
space and we know by the classification in [4] that GA is isogenous to
either Sp(6, R) × Sp(2, R), Sp(2, 1) × Sp(1) or SO(4, 5). In all these
cases the inequality
dim S A < dim S − rkR G
holds (in fact GA is isogenous to SO(4, 5)).
Thus we have that
dim S A < dim S − rkR G
for every A ∈ G of finite order and non central. Then by Lemma 2.2,
Proposition 4.1 holds for g = f4(4) , and it concludes the proof.
38
CYRIL LACOSTE
5. Semisimple Lie algebras
We prove in this section the Main Theorem:
Main Theorem Let g be a semisimple Lie algebra and G = Aut(g).
Then
gd(Γ) = vcd(Γ)
for every lattice Γ ⊂ G.
Recall that if g is semisimple, it is isomorphic to a sum of simple Lie
algebras g1 ⊕ · · · ⊕ gr . The adjoint group Gad of g is then isomorphic
to a product of simple Lie groups, that is Gad = G1 × G2 × · · · × Gn
where the Gi are the adjoint groups of the gi . We can also assume that
Gad has no compact factors, indeed the symmetric spaces S and S α do
not change if we replace Gad by its quotient by the compact factors.
An automorphism of g is the composition of a permutation σ of the
isomorphic factors of g and a diagonal automorphism ρ of the form
ρ = ρ1 ⊕ · · · ⊕ ρr with ρi ∈ Aut(gi ).
We explain now why does the strategy used in the previous sections
not work. The point is that the inequality
dim S α 6 dim S − rkR G
for α ∈ Aut(g) needed to apply Lemma 2.3 does not hold even
simplest cases.
−1 0
In fact, if Gad = SL(3, R) × SL(3, R) and A = I3 , 0 −1
0
0
A
we have dim S = 5 + 3 = 8 > 6 = dim S − rgR G.
We bypass this this problem by improving the lower bound
in the
0
0
,
1
vcd(Γ) > dim S − rkR G
used above. Remember that by Theorem 2.2
vcd(Γ) = dim S − rkQ Γ
as long as Γ is arithmetic, so we want to majorate rkQ Γ. To do that
we will restrict our study to irreducible lattices. Recall that in this
context, a lattice Γ in G is irreducible if ΓH is dense in G for every
non-compact, closed, normal subgroup H of Gad .
We prove the following result, which is probably known to experts:
Proposition 5.1. Let g = g1 ⊕ · · · ⊕ gn be a semisimple Lie algebra
and Gi the adjoint group of gi for i = 1 . . . n. Then
rkQ Γ 6 min rkR Gi
i=1...n
for every irreducible arithmetic lattice Γ ⊂ G = Aut(g).
Proposition 5.1 will follow from the following theorem proved in [20]
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
39
Theorem 5.1. Let G = G1 × · · · × Gn (n > 2) be a product of noncompact connected simple Lie groups. The following statements are
equivalent:
(1) G contains an irreducible lattice.
(2) G is isomorphic as a Lie group to (GR )0 , where G is an Qsimple algebraic group.
(3) G is isotypic, that is the complexifications of the Lie algebras of
the Gi are isomorphic.
In addition to that, in this case G contains both cocompact and non
cocompact irreducible lattices.
Recall that an algebraic group G defined over Q is said to be Qsimple if it does not contain non-trivial connected normal subgroups
defined over Q.
Then we can prove Proposition 5.1.
Proof of Proposition 5.1. Remark that rkQ Γ = rkQ (Γ ∩Gad ). Moreover
if Γ is an irreducible lattice of G then Γ ∩ Gad is an irreducible lattice
of Gad , so we can assume that Γ ⊂ Gad . Remember that Gad = G1 ×
· · · × Gn and that we assumed that none of the Gi is compact. If n = 1
the result is trivial so assume that n > 2. Then rkR Gad > 2 so Γ is
arithmetic by Theorem 2.1 and there exists a Lie group isomorphism
ϕ : Gad → (GR )0 , where G is Q-simple by Theorem 5.1. Then we have
rkQ Γ = rkQ G. The algebraic group G is isomorphic to a product G1 ×
· · ·×Gn where the Gi are R-groups with (Gi )R isomorphic to G
for i =
Qi
1 . . . n (we can define Gi as the centralizer in G of the product k6=i Gk ).
We note πi the canonical projection of G on Gi . Let T ⊂ G be maximal
Q-split torus. Our goal is to prove that the restriction πi |T is of finite
kernel. On the one hand, Ker(πi |GQ ) ⊂ GQ is a normal subgroup of
GQ and G is defined over Q, so the Zariski closure of Ker(πi |GQ ) is
defined over Q by the Galois rationality criterion. However it is a non
trivial normal subgroup of G which is Q-simple so Ker(πi |GQ ) is finite
(it may be not connected). So Ker(πi |TQ ) is finite too. But Ker(πi |T ) is
a subgroup of the Q-split torus T, so its identity component is a Q-split
torus, and we have just seen its group of rational points is finite, so
Ker(πi |T ) is finite too.
Then the image of T by πi is a torus of Gi of the same dimension
than T (see [6, Cor. 8.4 p.114]). It may not be Q-split because the
projection is not defined over Q, but it is R-split as the projection is
defined over R, so:
rgQ G = dim T 6 rgR Gi = rgR Gi .
We can now conclude the proof of our Main Theorem.
40
CYRIL LACOSTE
Proof of the Main Theorem. If g is simple then the result follows from
Propositions 2.1, 3.1 and 4.1. Then we assume that g = g1 ⊕ · · · ⊕ gn
with n > 2. We can also assume that the adjoint group Gad of g is of
the form Gad = G1 × · · · × Gn where the Gi are simple, non-compact
and Gi is the adjoint group of gi .
We begin with the case where Γ is irreducible. Then we have
rkQ Γ 6 min rkR Gi .
i=1...r
As rkR G > 2, Γ is arithmetic by Theorem 2.1. Remember that Γ ∩ Gad
is also an irreducible arithmetic lattice of Gad . We can then assume
that Gad = (GR )0 , where G = G1 × · · · × Gn is a semisimple Q-group,
which is Q-simple by Theorem 5.1. As Gad has trivial center, we can
assume that G is centerfree. In this case we have Γ ∩ Gad ⊂ GQ .
We want to use Lemma 2.2. Let α ∈ Γ of finite order non central.
Then α is of the form σ ◦ ρ where σ is a permutation of the isomorphic
factors of g and ρ = ρ1 ⊕ . . . ρr with ρi ∈ Aut(gi ). Assume for a while
that σ is trivial. We identify α ∈ Aut(g) (resp. ρi ∈ Aut(gi )) with
the corresponding automorphism of Gad (resp. Gi ). The key point is
to remark that for all i between 1 and n, the automorphism ρi is not
trivial. In fact if A ∈ Γ ∩ Gad ⊂ GQ , we can identify it with the inner
automorphism Ad(A) and we have
α ◦ Ad(A) ◦ α−1 = Ad(α(A)) ∈ Γ ∩ Gad ⊂ GQ ,
so α(A) lies also in GQ . Recall that we have seen in the proof of
Proposition 5.1 that the projections πi |GQ : GQ → Gi are injective. So
if ρi is trivial, we have πi (A) = πi (α(A)) for each A ∈ Γ ∩ Gad ⊂ GQ ,
which leads to α(A) = A. Then α is trivial on Γ ∩ Gad which is Zariskidense in Gad , so α is trivial.
Finally, α = ρ = ρ1 ⊕ · · · ⊕ ρn where each ρi is a non trivial automorphism of Gi . By Proposition 5.1 we also have
rkQ Γ 6 min rkR Gi .
i=1...n
Then if we note S the symmetric space associated to G and Si those
associated to Gi and we have by Propositions 3.1 and 4.1
dim S α =
n
X
i=1
dim Siρi 6
n
X
(dim Si − rkR Gi )
i=1
6 dim S −
n
X
rkR Gi
i=1
< dim S − rkQ Γ.
as we assumed n > 2.
By Theorem 2.2, dim S A < vcd(Γ) and Lemma 2.2 gives us the
result.
If σ is not trivial, the fixed point set will be even smaller. Indeed,
assume for simplicity that g = g1 ⊕ g2 where g1 and g2 are isomorphic,
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
41
and α ∈ Aut(g) is of the form
α(x1 , x2 ) = (ρ1 (x2 ), ρ2 (x1 ))
with ρ1 , ρ2 ∈ Aut(g1 ) = Aut(g2 ). Then the fixed point set S α is S1ρ1 ρ2
where S1 is the symmetric space associated to g1 . In fact the elements
fixed by α are of the form (x1 , ρ2 (x1 )) where x1 is a fixed point of ρ1 ρ2 .
So we have
dim S α 6 dim S1 < dim S − rkR G1 6 dim S − rkQ Γ = vcd(Γ)
as dim S = 2 dim S1 and dim S1 > rkR G1 . The same argument works
for a higher number of summands by decomposing the permutation σ
into disjoint cycles.
Finally if Γ is reducible, there exists a decomposition of G = H1 ×H2
such that the projections π1 (Γ) and π2 (Γ) are lattices in H1 and H2 , and
then Γ ⊂ π1 (Γ) × π2 (Γ) (see the proof of [29, Prop. 4.3.3]). It follows
by induction that Γ is contained in a product of irreducible lattices of
factors of G. We will treat the case where G = H1 × H2 , Γ ⊂ Γ1 × Γ2
and Γ1 and Γ2 are irreducible lattices of H1 and H2 . As Γ is of finite
index in G, it is also of finite index in Γ1 × Γ2 so vcd(Γ) = vcd(Γ1 × Γ2 ).
If we note S, S1 , S2 the symmetric spaces associated to G, H1 , H2 , by
Theorem 2.2 we have:
vcd(Γ) = vcd(Γ1 × Γ2 ) = dim S1 + dim S2 − rkQ Γ1 − rkQ Γ2
= vcd(Γ1 ) + vcd(Γ2 ).
Finally, we have:
gd(Γ) 6 gd(Γ1 × Γ2 ) 6 gd(Γ1 ) + gd(Γ2 )
because if X1 and X2 are models for EΓ1 and EΓ2 , X1 × X2 is a model
for E(Γ1 ×Γ2 ). As Γ1 and Γ2 are irreducible, we have gd(Γ1 ) = vcd(Γ1 )
and gd(Γ2 ) = vcd(Γ2 ) so:
gd(Γ) 6 vcd(Γ).
The other inequality is always true, so it concludes the proof of the
Main Theorem.
We will end with the proof of Corollaries 1.1 and 1.2.
Proof of Corollary 1.1. The case of real rank 1 is treated in Proposition
2.6 in [1]. For higher real rank, we know by the Main Theorem that
there exists a model for EΓ of dimension vcd(Γ). We also know that
the Borel-Serre bordification is a cocompact model for EΓ. Then using
the same construction as in the proof of Corollary 1.1 in [1], one has
a cocompact model for EΓ of dimension vcd(Γ). As all models of EΓ
are Γ-equivariantly homotopy equivalent and the symmetric space S is
also a model for EΓ, we conclude that S is Γ-equivariantly homotopy
equivalent to a cocompact model for EΓ of dimension vcd(Γ).
42
CYRIL LACOSTE
Proof of Corollary 1.2. We have to prove that if Γ1 ⊂ Aut(g) and Γ2
e of finite index, then gd(Γ ) = vcd(Γ ). To
have a common subgroup Γ
2
2
that end, we will prove that Γ2 is essentially also a lattice in Aut(g).
e is a lattice in Aut(g), so we can assume that Γ
e =Γ ⊂
First note that Γ
1
Γ2 . We can also assume that Γ1 is a normal finite index subgroup of Γ2 .
Then Γ2 acts by conjugation on Γ1 . By Mostow rigidity Theorem (see
for example [29, Thm. 15.1.2]), automorphisms of Γ1 can be extended
to automorphisms of Gad , so we have a morphism Γ2 → Aut(Gad ). The
kernel N of this morphism does not intersect Γ1 (since Γ1 is centerfree,
as it is a lattice and thus it is Zariski-dense in Gad ) and Γ1 is of finite
index in Γ2 , so N is finite. Then Γ2 /N is isomorphic to a lattice in
Aut(Gad ). The result follows now from the Main Theorem and Lemma
2.1.
Note that Mostow rigidity theorem does not apply to the group
PSL(2, R), whose associated symmetric space is the hyperbolic plane.
In this case the lattice Γ1 is either a virtually free group or a virtually surface group. In the first case the group Γ2 is also virtually
free, so there exists a model for EΓ2 which is a tree (see [21]), and
gd(Γ2 ) = vcd(Γ2 ) = 1. In the second case, Γ2 acts as a convergence
group on S1 = ∂∞ Γ1 , so it is also a Fuchsian group (see [14]), that is
Γ2 is isomorphic to a cocompact lattice of PSL(2, R). Finally we have
gd(Γ2 ) = vcd(Γ2 ) = 2.
References
[1] J. Aramayona, D. Degrijse, C. Martinez-Perez, and J. Souto. Geometric dimension of lattices in classical simple Lie groups. arXiv:1504.02607, 2015.
[2] J. Aramayona and C. Martinez-Perez. The proper geometric dimension of the
mapping class group. Algebraic and Geometric Toplogy 14, 217-227, 2014.
[3] A. Ash. Small-dimensional classifying spaces for arithmetic subgroups of general linear groups. Duke Math. J. 51, no. 2, 459-468, 1984.
[4] M. Berger. Les espaces symétriques non compacts. Annales scientifiques de
l’ENS, 1957.
[5] A. Borel. Introduction aux groupes arithmétiques. Hermann, 1969.
[6] A. Borel. Linear Algebraic Groups. Springer, 1969.
[7] A. Borel and J-P. Serre. Corners and arithmetic groups. Comment.Math.Helv.48, 436-491, 1973.
[8] N. Brady, I. Leary, and B. Nucinkis. On algebraic and geometric dimensions
for group with torsion. London. Math. Soc.(2), 64, 489–500, 2001.
[9] K.S. Brown. Cohomology of groups. Graduate texts in Mathematics 77,
Springer, 1982.
[10] D. Degrijse and C. Martinez-Perez. Dimension invariants for groups admitting a cocompact model for proper actions. Journal für Reine und Angewandte
Mathematik: Crelle’s Journal, 2014.
[11] D. Degrijse and N. Petrosyan. Geometric dimension of groups for the family
of virtually cyclic subgroups. J. Topol. 7, 697–726, 2014.
[12] D. Degrijse and J. Souto. Dimension invariants of outer automorphism groups.
arXiv:1602.04354, 2016.
DIMENSION RIGIDITY OF LATTICES IN SEMISIMPLE LIE GROUPS
43
[13] D.Z. Djokovic. On real form of complex semisimple Lie algebras. Aequationes
Math. 58, 73-84, 1999.
[14] D. Gabai. Convergence groups are Fuchsian groups. Annals of Mathematics,
136, 447-510, 1992.
[15] A. Gray and J.A. Wolf. Homogeneous spaces defined by Lie groups automorphisms. I. J. Diff. Geom. 2, 77-114, 1968.
[16] H. Gündogan. The Component Group of the Automorphism Group of a Simple Lie Algebra and the Splitting of the Corresponding Short Exact Sequence.
Journal of Lie Theory 20, 709-737, 2010.
[17] H. Gündogan. Classification and structure theory of Lie algebras of smooth
section. Logos Verlag Berlin GmbH, 2011.
[18] S. Helgason. Differential Geometry, Lie Groups, and Symmetric spaces. American Mathematical Society, 2001.
[19] L. Ji. Integral Novikov conjectures and arithmetic groups containing torsion
elements. Communications in analysis and geometry, Volume 15, Number 3,
509–533, 2007.
[20] F.E.A. Johnson. On the existence of irreducible discrete subgroups in isotypic
Lie groups of classical type. Proc.London Math.Soc., 1988.
[21] A. Karass, A. Pietrowski, and D. Solitar. Finite and infinite cyclic extensions
of free group. J. Austral. Math. Soc. 16, 458–466, 1973.
[22] A.W. Knapp. Lie groups beyond an introduction. Progress in Mathematics,
Birkäuser, 2002.
[23] I. Leary and B. Nucinkis. Some groups of type VF. Invent. Math. 151, 135–165,
2001.
[24] I. Leary and N. Petrosyan. On dimensions of groups with cocompact classifying
spaces for proper actions. arXiv:1504.02704, 2015.
[25] W. Lück. Survey on classifying spaces for families of subgroups. Infinite Groups
: Geometric, Combinatorial and Dynamical Aspects, Springer, 269-322, 2005.
[26] W. Lück and D. Meintrup. On the universal space for group actions with compact isotropy. Proc. of the conference "Geometry and Topology" in Aarhus,
293-305, 1998.
[27] G.A. Margulis. Discrete subgroups of semisimple Lie groups. Springer-Verlag,
1989.
[28] C. Martinez-Perez. Euler classes and Bredon cohomology for groups with restricted families of finite torsion. Math. Z. 2013, 761–780, 2013.
[29] D. Witte Morris. Introduction to arithmetic groups. arXiv:math/0106063, 1990.
[30] A.L. Onishchik and E.B. Vinberg. Lie groups and Algebraic groups. SpringerVerlag, 1980.
[31] A. Pettet and J.Souto. The spine which was no spine. L’Enseignement Mathematique 54, 2008.
[32] A. Pettet and J. Souto. Minimality of the well-rounded retract. Geometry and
Topology 12, 2008.
[33] K. Vogtmann. Automorphisms of free groups and outer space. Geometriae Dedicata, Vol.94 (1), 1-31, 2002.
[34] I. Yokota. Exceptional Lie Groups. arXiv:0902.0431, 2009.
IRMAR, Université de Rennes 1
E-mail address: [email protected]
| 4math.GR
|
arXiv:1710.09358v2 [math.AC] 28 Oct 2017
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
GIUSEPPE FAVACCHIO
Abstract. We introduce the fractal expansions, sequences of integers associated to a number. These
can be used to characterize the O-sequences. We generalize them by introducing numerical functions
called fractal functions. We classify the Hilbert functions of bigraded algebras by using fractal functions.
1. Introduction
In Commutative Algebra (and other fields of pure mathematics) it often happens that easy numerical
conditions describe some deeper algebraic results. A significant example are the O-sequences.
Let S := k[x1 , . . . , xn ] be the standard graded polynomial ring and let I ⊆ S be a homogeneous ideal.
The quotient ring S/I is called a standard graded k-algebra. The Hilbert function of S/I is defined as
HS/I : N → N such that
HS/I (t) := dimk (S/I)t = dimk St − dimk It .
A famous theorem, due to Macaulay (cf. [9]) and pointed out by Stanley (cf. [11]), characterizes the
numerical functions that are Hilbert functions of a standard graded k-algebra, i.e. the functions H such
that H = HS/I for some homogeneous ideal I ⊆ S. To introduce this fundamental result we need some
preparatory material.
Let h, i > 0 be integers, we can uniquely write h as
mj
mi−1
mi
+ ···+
+
(1.1)
h=
j
i−1
i
where mi > mi−1 > · · · > mj ≥ j ≥ 1. This expression is called the i-binomial expansion of the integer
h.
If h > 0 has i-binomial expansion as in 1.1, then we set
mj + 1
mi + 1
mi
+ ··· +
hhii =
.
+
i
j+1
i+1
We use the convention that 0hii = 0.
For example, since 7 = 43 + 32 , the 3-binomial expansion of 7 is and 7h3i =
5
4
+
4
3
= 9.
Definition 1.1. A sequence of non-negative integers (h0 , h1 , h2 · · · ) is called an O-sequence if
i) h0 = 1;
hii
ii) hi+1 ≤ hi for all i > 0.
hii
An O-sequence is said to have maximal growth from degree i to degree i + 1 if hi+1 = hi
2000 Mathematics Subject Classification. 13F20, 13A15, 13D40.
Key words and phrases. Hilbert function, multigraded albegra, numerical function.
Version:October 31, 2017.
1
2
GIUSEPPE FAVACCHIO
We are now ready to enunciate the Macaulay’s theorem. It characterizes the Hilbert function of
standard graded k-algebras bounding the growth from any degree to the next. The proof of this theorem
and more details about O-sequences are also discussed in [2] Chapter 4. We represent HS/I as a sequence
of integers (h0 , h1 , h2 , · · · ) where ht := HS/I (t).
Theorem 1.2 (Macaulay, [9]). Let H := (h0 , h1 , h2 , · · · ) be a sequence of integers, then the following are
equivalent:
1) H is the Hilbert function of a standard graded k-algebra;
2) H is an O-sequence.
It is therefore interesting to find an extension of the above theorem to the multi-graded case. Multigraded Hilbert functions arise in many contexts and properties related to the Hilbert function of multigraded algebras are currently studied. See for instance [6] and [10] for several examples. The generalization of Macaulay’s theorem to multi-graded rings is an open problem. A first answer was given by the
author in [5] where the Hilbert functions of a bigraded algebra in k[x1 , x2 , y1 , y2 ] are classified.
The goal of this work is to generalize the Macaulay’s theorem to any bigraded algebras.
In order to reach our purpose we first, in Section 2, introduce Φ(n) a list of finite sequences called
fractal expansion of n. Then we define a coherent truncation of these vectors and we show that these
objects are strictly related to the O-sequences. Indeed, in Section 3 we show that they also characterize
the Hilbert function of standard graded k-algebras. Furthermore, we show that these sequences can be
used to compute the Betti numbers of a lex ideal. In Section 4 we extend some of these results and we
classify the Hilbert function of bigraded algebras, Theorem 4.13.
The computer program CoCoA [3] was indispensable for all the computations.
2. Expansion of a Fractal sequence
In this section we describe a new approach to classify the Hilbert functions of standard graded algebras.
We introduce a sequence of tuplas, called coherent fractal growth, and we study its properties. The main
result of this section is Theorem 2.14. We prove that these sequences have the same behavior of the
O-sequences.
Roughly speaking a numerical sequence σ is called fractal if once we delete the first occurrence of
each number it remains identical to the original. Such property thus implies we can repeat this process
indefinitely, and σ contains infinitely many copy of itself. It has something like a fractal behavior. See
[8] for a formal definition and further properties.
For instance, one can show that the sequence σ = (1, 1, 2, 1, 2, 3, 1, 2, 3, 4, 1, 2, 3, 4, 5, . . .) is fractal.
Indeed, after removing the first occurrence of each number we get a sequence that it is the same as the
starting one
(✁
1, 1, ✁
2, 1, 2, ✁3, 1, 2, 3, ✁4, 1, 2, 3, 4, ✁5, . . .) = σ.
We introduce some notation. Given a positive integer a ∈ N we denote by [a] := (1, 2, . . . , a) ∈ Na the
tupla of length a consisting af all the positive integers less then or equal to a written in increasing order.
Given a (finite or infinite) sequence σ of positive integers we construct a new sequence, named the
expansion of σ, denoted by [σ]. If σ := (s1 , s2 , s3 , . . .) we set [σ] := [s1 ] [s2 ], [s3 ] · · · , where the symbol
" " denotes the associative operation of concatenation of two vectors. E.g. (3, 5, 4) (2, 3) = (3, 5, 4, 2, 3).
This construction can be recursively applied. We denote by [σ]d := [[σ]d−1 ] where we set [σ]0 := σ.
For a positive integer a, we also denote by [a]d := [[a]d−1 ], where [a]0 := (a).
For instance we have [3] = (1, 2, 3), [3]2 = (1, 1, 2, 1, 2, 3), [3]3 = (1, 1, 1, 2, 1, 1, 2, 1, 2, 3).
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
3
Lemma 2.1. Let σ := (s1 , s2 , s3 , . . .) be a sequence of positive integers. Then [σ]d = [s1 ]d [s2 ]d [s3 ]d
··· .
Proof. If d = 0 the statement is true. Assume d > 0. By definition [σ]d = [[σ]d−1 ] then by the inductive
hypothesis we have [σ]d = [[s1 ]d−1 ] [[s2 ]d−1 ] [[s3 ]d−1 ] · · · = [s1 ]d [s2 ]d [s3 ]d · · · .
Corollary 2.2. Let a ∈ N be a positive integer. Then [a]d = [1]d−1 [2]d−1
···
[a]d−1 .
Proof. The statement follows by Lemma 2.1 since [a]d = [(1, 2, . . . , a)]d−1 .
Remark 2.3. The sequence [N] := [1] [2]
···
[n]
· · · is a fractal sequence.
Given a sequence σ, the symbolsPσi and σ(i) both denote the i-th entry of σ. If σ if finite then |σ|
denotes the number of entries and
σ their sum. We use the convention that
P these values are "∞" for
infinite sequences of positive integers. For instance, [3]34 = 2, [3]3 = 10 and [3]3 = 15.
Throughout this paper we use the convention that, for a finite sequence σ, the notation σa or σ(a)
implies a ≤ |σ|.
Remark 2.4.
P Note that, for a finite sequence of positive integers σ, the definition of [σ] easily implies
the equality
σ = |[σ]| .
Given a positive integer n, we define the fractal expansion of n as the set
Φ(n) := {[n]0 , [n]1 , . . . , [n]d , . . .}.
Each element in Φ(n) is a finite sequence of positive integers. In the following lemma we compute the
number of their entries.
P
Lemma 2.5. Let n be a positive integer. Then [n]d = d+n−1
and [n]d = d+n
d
d+1 .
n−1
n
Proof. By definition we have [n]0 = n−1
= 1 and [n]1 = n−1
= n. By Lemma 2.1 we have
[n]d = [1]d−1 [2]d−1 · · · [n]d−1 , therefore
n
n
X
X
d+j−2
d+n−1
d
d−1
[n] =
[j]
=
.
=
n−1
n−1
j=1
j=1
Moreover, by Remark 2.4 we have
P
[n]d = [n]d+1 =
d+n
d+1
.
Next lemma introduces a way to decompose a number as sum of binomial
coefficients that is slight
different to the Macaulay decomposition. We use the convention that ab = 0 whenever a < b.
Lemma 2.6. Let d be a positive integer. Any a ∈ N can be written uniquely in the form
k1
k2
kd−1
kd
+
+ ··· +
+
(2.1)
a=
1
2
d−1
d
where kd > kd−1 > · · · > k2 ≥ k1 ≥ 1
Proof. In order to prove the existence, we choose kd maximal such that kdd < a. By the inductive
Pd−1
d−1
hypothesis on d, 0 < a− kdd = i=1 kii , where kd−1 > · · · > k2 ≥ k1 ≥ 1. Since k1 ≥ 1 a− kdd > kd−1
,
moreover since kdd+1 ≥ a it follows that
kd−1
kd
kd + 1
kd
kd
.
>
≥a−
=
−
d−1
d
d
d−1
d
4
GIUSEPPE FAVACCHIO
Hence kd > kd−1
uniqueness
follows
. The
by induction on d. If d = 1 it is trivial. Now assume that d > 1
d−1
and let a = kdd + kd−1
+ · · · + k22 + k11 be a decomposition of a. Then kd is the maximal integer such
d−1
that kdd < a. Otherwise, if kdd+1 < a we get a > kdd+1 ≥ kdd + kd−1
+ · · · + k22 + k11 .
Remark 2.7. The decomposition in Lemma 2.6 is different from the Macaulay decomposition since it is
always required that k1 ≥ 1. Moreover, for any j ≥ 2 we only have that kj ≥ j − 1, thus some binomial
1
1
coefficient could be zero. For instance, we have 1 = d−1
+ d−2
d
d−1 + · · · + 2 + 1 where the first d − 1
binomial coefficients in the sum are equal to zero.
Definition 2.8. We refer to equation 2.1 as the d-fractal decomposition of a. We denote by
[a](d) := (kd − d + 2, kd−1 − d + 3, . . . , k4 − 2, k3 − 1, k2 , k1 ) ∈ Nd .
We call these numbers the d-fractal coefficients of a.
(d)
[a](d) is a not increasing sequence of positive integers. Indeed, [a]d > 0 and by construction k1 ≤ k2 .
Moreover, for j > 1, since kj > kj−1 , we have kj − j + 2 ≥ kj−1 − j + 1.
Next result explains the name "d-fractal decomposition". We show that k1 is the a-th entry in [n]d .
We need the convention that [0]d is the empty sequence and [0]d = 0.
(d)
Theorem 2.9. [n]da = [a]d .
(1)
Proof. If d = 1 then [n]1a = a and a = a1 thus [a]1 = a = [n]1a . We now assume d > 1.
d−1
Let a = kdd + kd−1
+ · · · + k22 + k11 be the d-fractal decomposition of a. By Lemma 2.5 we have
d−1
[kd − d + 2]d ≥ a > [kd − d + 1]d and [n]da = [kd − d + 2]d−1kd .Thus, since kd−1
+ · · · + k22 + k11 is
a−( d )
the d − 1 fractal decomposition of a − kdd , by the inductive hypothesis we have k1 = [n]d−1 kd .
(a−( d ))
Given α, β ∈ Nd then α <lex β iff for some i ≤ d we have αj = βj for j < i and αi < βi . The following
lemma is crucial for our intent. We prove that the d-fractal coefficients have a good behavior with respect
the lex order.
Lemma 2.10. [a](d) <lex [b](d) iff a < b.
Proof. If d = 1 the assertion is trivial. Let d > 1 and let a, b be two integers with fractal decomposition
a1
a2
ad−1
ad
, ad > · · · > a2 ≥ a1 ≥ 1
+
+ ···+
+
a=
1
2
d−1
d
and
b=
b1
b2
bd−1
bd
, bd > · · · > b2 ≥ b1 ≥ 1.
+
+ ···+
+
1
2
d−1
d
(d)
(d)
(d)
(d)
If [a](d) <lex [b](d) then there is an index j such that [a]i = [b]i for any i < j and [a]j < [b]j . Hence,
ai
bi
aj
bj
i = i for any i < j and j < j . If j = d then easily b > a, otherwise
ai + 1
ad−1
ad
bj
bd−1
bd
+ ···+
≥ a.
+
≥
+ ···+
+
b>
d−1
j
d
j
d−1
d
Vice versa let b > a we claim that bd ≥ ad . Indeed, if bd < ad we get a > add ≥ bdd+1 ≥ b
contradicting b > a. So if bd > ad we are done otherwise the statement follows by induction.
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
5
Given two sequences τ and σ we say that τ is a truncation of σ if |τ | ≤ |σ| and τ (j) = σ(j) for any
j ≤ |τ |.
Next definition introduces the main tool of the paper. The coherent fractal growths are suitable
truncations the elements in the fractal expansion of n.
Definition 2.11. We say that T := {τ0 , τ1 , τ2 , . . .} is a coherent fractal growth if τ0 := (n) and τj is a
truncation of [τj−1 ] for each j ≥ 1.
For instance {(3), (1, 2, 3), (1, 1, 2, 1, 2), (1, 1, 1, 2, 1, 1), (1, 1, 1, 1)} is a coherent fractal growth. Indeed
one can check that each elements is truncation of the expansion of the previous one. On the other hand,
for instance, {(3), (1, 2, 3), (1, 1, 2), (1, 1, 1, 2, 1)} is not a coherent fractal growth. Indeed, (1, 1, 1, 2, 1) is
not a truncation of [(1, 1, 2)] = (1, 1, 1, 2).
Remark 2.12. Note that, in a coherent fractal growth, τd consists of the first |τd | elements in [n]d .
Moreover, the length of the elements in T := {τ0 , τ1 , τ2 , . . .}, a coherent fractal growth, is bounded for
any d. Indeed, by Remark 2.4 we have
X
(2.2)
|τd | ≤
τd−1
for each d ≥ 1.
In the next part of this section we prove that the bound in Remark 2.12 is equivalent to the binomial
expansion for a O-sequence. In order to relate the
coherent
fractal growth with O-sequences we need the
a−1
following lemma. It uses the equality ab = a−1
+
b
b−1 .
Lemma 2.13. Let [a](d) = (cd , cd−1 , . . . , c1 ) be the d-fractal coefficients of a. Then the (d + 1)-fractal
coefficients of ahdi are [ahdi ](d+1) = (cd , cd−1 , . . . , c2 , c1 , c1 ).
Proof. The d-fractal decomposition of a is, by definition
cd + d − 2
c1
cd−1 + d − 3
c3 + 1
c2
.
a=
+
+
+ ···+
+
1
2
d
d−1
3
If c1 < c2 , we get the Macaulay decomposition of a by removing the binomials ji equal to 0. Since
j
j+1
i = 0 implies i+1 = 0 we have
cd + d − 1
cd + d − 1
c3 + 2
c2 + 1
c1 + 1
ahdi =
+
+ ···+
+
+
.
d+1
d
4
3
2
Since c12+1 = c21 + c11 we are done. Now we consider the case c1 = c2 . Then we can write the following
decomposition of a
cd + d − 2
cd−1 + d − 1
c3 + 1
c2 + 1
a=
+
+ ···+
+
.
d
d−1
3
2
Thus if c3 > c2 this representation is the Macaulay decomposition of a once we remove the binomials ji
equal to 0.
cd + d − 1
cd−1 + d
c3 + 2
c2 + 2
ahdi =
+
+ ···+
+
=
d+1
d
4
3
cd + d − 1
cd−1 + d
c3 + 2
c2 + 1
c1 + 1
=
+
+ ···+
+
+
=
d+1
d
4
3
2
cd + d − 1
cd−1 + d
c3 + 2
c1
c2 + 1
c1
=
.
+
+ ···+
+
+
+
1
d+1
d
2
4
3
The proof follows in a finite number of steps by iterating this argument.
6
GIUSEPPE FAVACCHIO
The following theorem is the main result of this section. We show that the length of the elements in
a coherent fractal growth is an O-sequence.
Theorem 2.14. Let T = {τ0 , τ1 , τ2 . . .} be a list of truncations of Φ(n) = {(n), [n], [n]2 , . . .}. Then the
following are equivalent
i) T = {τ0 , τ1 , τ2 . . .} is a coherent fractal growth;
ii) (|τ0 |, |τ1 |, |τ2 |, . . .) is an O-sequence.
Proof. In order to prove (i) ⇒ (ii), we need to show |τd+1 | ≤ |τd |hdi , for each d ≥ 0. Set a = |τd | and
take the d-fractal decomposition of a
k1
k2
k3
kd−1
kd
.
+
+
+ ···+
+
a=
1
2
3
d−1
d
If a = n+d+1
= [n]d the statement follows by Lemma 2.5. Assume now a <
d
d
truncation of [n] = [1]d−1 [2]d−1 · · · [n]d−1 and, by Lemma 2.5,
kd + 1
kd
d
<a≤
= [kd − d + 2]d
[kd − d + 1] =
d
d
n+d+1
d
. Since τd is a
we get (denoted by [0] = ())
τd = [kd − d + 1]d τd′
where τd′ is a truncation of [kd − d + 2]d−1 . Therefore, reiterating this argument, we have
τd = [kd − d + 1]d [kd−1 − d + 2]d−1
···
[k3 − 1]3 [k2 − 1]2 [k1 ].
By equation 2.2 in Remark 2.12, we have
d
X
X
ki + 1
kd + 1
kd−1 + 1
k1
k2 + 1
k1
.
=
+
+
+ ···+
+
|τd+1 | ≤
τd =
1
d+1
d+1
2
d
3
i=1
The last sum, by Lemma 2.13, is equal to ahdi .
Vice versa, to prove (ii) ⇒ (i), we have to prove that, for each d ≥ 0, the sequence τd+1
P is a truncation
of [τd ] i.e. |τd+1 | ≤ |[τd ]|. It follows by using the same argument as above, |τd |hdi =
τdP
. Indeed, by
hypothesis, we have the bound |τd+1 | ≤ |τd |hdi and, by Remark 2.12, we know that |[τd ]| = τd .
Let’s check, for instance, that H := (1, 3, 3, 4) is an O-sequence. We write a sequence of truncations
of Φ(3) of length 1, 3, 3, 4 respectively. We get
T := {(3), (1, 2, 3), (1, 1, 2), (1, 1, 1, 2)}.
It is a coherent fractal growth. Indeed, by definition each sequence is a truncation of the previous one,
e.g. [(1, 2, 3)] = (1, 1, 2) (1, 2, 3) and [(1, 1, 2)] = (1, 1, 1, 2). On the other hand we can also check that
H := (1, 3, 5, 8) is not an O-sequence. Indeed, in a coherent fractal growth such
{(3), (1, 2, 3), (1, 1, 2, 1, 2), τ3}
τ3 need to be a truncation of [(1, 1, 2, 1, 2)] = (1, 1, 1, 2, 1, 1, 2) that has length 7. Therefore |τ3 | ≤ 7 that
is the maximal growth allowed.
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
7
3. Fractal expansion and Homological Invariants
In Section 2 we introduced a novel approach to describe the O-sequences. In this section we show
the "algebraic" meaning of a coherent fractal growth. We directly relate this sequences with lex segment
ideals and its homological invariants. In particular, the Eliahou-Kervaire formula is naturally applied to
our case. Therefore, the fractal expansion of n is used in Proposition 3.4 to compute the Betti numbers
of a lex algebra.
Let a and d be positive integers. Let [a](d) := (cd , cd−1 , . . . , c2 , c1 ) be the d-fractal coefficient of a, see
(d)
Lemma 2.6 and Definition 2.8. We associate to a and d a monomial Xa of degree d in the variables
X := {x1 , . . . xn } in such a way
Xa(d) := xc1 xc2 · · · xcd−1 xcd .
Vice versa, a monomial T = xc1 xc2 · · · xcd−1 xcd of degree d in the variables in X identifies a d-upla
(cd , cd−1 , . . . , c1 ) such that ci ≥ ci+1 for any i.
(d)
(d)
Remark 3.1. An immediate consequence of Lemma 2.10 is that Xa >lex Xb iff a > b, with respect
(d)
the lex order induced by x1 < x2 < · · · < xn . Therefore, with respect the same order, Xa is the a-th
greatest monomial of degree d in the set of variables X.
Let S := k[X] = k[x1 , . . . , xn ] be the standard graded polynomial ring.
We set
(d)
G(X)≤t := {Xa(d) | a ≤ t}
and
(d)
G(X)>t := {Xa(d) | a > t}.
(d)
(d)
Sd is spanned by the monomials in G(X)≤t ∪ G(X)>t .
(d)
By Remark 3.1 G(X)>t is a lex set of monomials of degree d with respect the degree lexicographic
order x1 < x2 < · · · < xn .
(d)
Given T := {τ0 , τ1 , τ2 , . . .} a coherent fractal growth, we set I(T )d := hG(X)>|τd | ik the k-vector space
(d)
spanned by the monomial in G(X)>|τd | . Then, by Theorem, 2.14 and Theorem 1.2, the following result
holds.
Proposition 3.2. I(T ) := ⊕d Id (T ) ⊆ R is a lex segment ideal and HR/I (d) = |τd |.
Given a minimal free resolution of an lex segment ideal
M
M
S(−j)β1,1+j → S → S/I → 0
S(−j)βp,p+j → · · · →
0→
j
j
the Betti numbers can be computed by the Eliahou-Kervaire formula, see [4]. See also [7] equation 7.7 of
Section 7.
Theorem 3.3 (Eliahou-Kervaire formula). Let I be a lex segment ideal. For u ∈ G(I), a monomial
minimal generator of I, let m(u) denotes the largest index j such that xj divides u. Let mkj be the
number of monomials u ∈ G(I) with m(u) = k. Then
n
X m(u) − 1 X
k−1
mkj .
=
βi,i+j (S/I) =
i−1
i
i=1
u∈G(I)j
This result can be written in terms of coherent fractal growth.
8
GIUSEPPE FAVACCHIO
Proposition 3.4. Given T := {τ0 , τ1 , τ2 , . . .} a coherent fractal growth. Then
X
|[τi−1 ]|
n
(j)
X
[a]j − 1
k−1
βi,i+j (S/I(T )) =
=
wkj
i
i−1
i=1
a=|τj |+1
where wkj is the number of occurrence of "k" in τi′ with [τi−1 ] = τi τi′ .
Proof. It is an immediate consequence of Theorem 3.3 and Theorem 2.9.
4. The Hilbert function of a bigraded algebra and the Fractal Functions
Let k be an infinite field, and let R := k[x1 , · · · xn , y1 , · · · , ym ] be the polynomial ring in n + m
indeterminates with the grading defined by deg xi = (1, 0) and deg yj = (0, 1). Then R = ⊕(i,j)∈N2 R(i,j)
where R(i,j) denotes the set of all homogeneous elements in R of degree (i, j). Moreover, R(i,j) is generated,
jm
as a k-vector space, by the monomials xi11 · · · xinn y1j1 · · · ym
such that i1 + · · ·+ in = i and j1 + · · ·+ jm = j.
An ideal I ⊆ R is called a bigraded ideal if it is generated by homogeneous elements with respect to
this grading. A bigraded algebra R/I is the quotient of R with a bigraded ideal I. The Hilbert function
of a bigraded algebra R/I is defined such that HR/I : N2 → N and HR/I (i, j) := dimk (R/I)(i,j) =
dimk R(i,j) − dimk I(i,j) where I(i,j) = I ∩ R(i,j) is the set of the bihomogeneous elements of degree (i, j)
in I.
From now on, we will work with the degree lexicographical order on R induced by xn > · · · > x1 >
ym > · · · > y1 . With this ordering, we recall the definition of bilex ideal, introduced and studied in [1].
We refer to [1] for all preliminaries and for further results on bilex ideals.
Definition 4.1 ([1], Definition 4.4). A set of monomials L ⊆ R(i,j) is called bilex if for every monomial
uv ∈ L, where u ∈ R(i,0) and v ∈ R(0,j) , the following conditions are satisfied:
• if u′ ∈ R(i,0) and u′ > u, then u′ v ∈ L;
• if v ′ ∈ R(0,j) and v ′ > v, then uv ′ ∈ L.
A monomial ideal I ⊆ R is called a bilex ideal if I(i,j) is generated as k-vector space by a bilex set of
monomials, for every i, j ≥ 0.
Bilex ideals play a crucial role in the study of the Hilbert function of bigraded algebras.
Theorem 4.2 ([1],Theorem 4.14). Let J ⊆ R be a bigraded ideal. Then there exists a bilex ideal I such
that HR/I = HR/J .
In [5] was solved the question of characterize the Hilbert functions of bigraded algebras of k[x1 , x2 , y1 , y2 ]
by introducing the Ferrers functions. In this section we generalize these functions by introducing the fractal functions, see Definition 4.6. We prove, Theorem 4.13, that these classify the Hilbert functions of
bigraded algebras.
We need some preparatory material. We denote by Ma×b (U ) the set of all the matrices with size
(a, b), i.e. a rows and b columns, and entries in a set U ⊆ k. Given a matrix M = (mij ) ∈ Ma×b (U ) we
denote by
X
XX
M :=
mi,j
i≤a j≤b
named the weight of M.
Next definition introduces the objects we need in this section.
Definition 4.3. A Ferrers matrix of size (a, b) is a matrix M = (mij ) ∈ M({0, 1})a×b such that
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
9
if mij = 1 then mi′ j ′ = 1 for any (i′ , j ′ ) ≤ (i, j).
We set by F a×b the family of all the Ferrers matrices of size (a, b).
In the next definition we introduce expansions of a matrix.
Definition 4.4. Let M ∈ M(U )a×b be a matrix of size (a, b) and let v := (v1 , . . . , va ) ∈ Na and P
w :=
b
hv,•i
(w1 , . . . , wb ) ∈ N be vectors of non negative integers. We denote by M
an element in M (U ) v×b
constructed by
repeating vi times the i-th row of M , for i = 1, . . . , a.
We denote by M
h•,wi
an element in M (U )a×
P
v2
constructed by
repeating wj times the j-th column of M , for j = 1, . . . , b.
Remark 4.5. The expansions of a Ferrers matrix are
1 1 1
1 1 0
M =
1 1 0
1 0 0
Set v := (2, 1, 0, 3) and w := (3, 1, 3). Then
1 1 1
1 1 1
1 1 0
hv1 ,•i
∈ F (6,3)
M
=
1 0 0
1 0 0
1 0 0
M h•,v2 i
also Ferres matrices. Take, for instance,
∈ F (4,3) .
1
1
=
1
1
1
1
1
1
1
1
1
1
1
1
1
0
1
0
0
0
1
0
0
0
1
0
∈ F (4,7) .
0
0
Given M, N ∈ F a×b we define a new matrix M ∗ N := (mij ∗ nij ) = (min(mij , nij )) ∈ F a×b . We say
that M ≤ N iff mij ≤ nij for any i, j.
We are ready to introduce the fractal functions.
Definition 4.6. Let H : N2 → N be a numerical function. We say that H is a fractal function if
j−1+m
i−1+n
P
H(0, 0) = 1 and, for any (i, j) ∈ N2 , there exists a matrix of Mij ∈ F ( n−1 )×( m−1 ) with Mij = H(i, j)
and such that all the matrices satisfy the condition
h[n]i−1 ,•i
if i > 0
Mij ≤ Mi−1j
h•,[m]j−1 i
Mij ≤ Mij−1
if j > 0
i−1+m
Remark 4.7. Let H : N2 → N be the numerical function H(i, j) = i−1+n
n−1
m−1 . For any i, j ∈ N,
i−1+m
i−1+n
×
)
)
(
(
m−1
satisfying the condition in Definition 4.6 that is the
there is only one element in M ∈ F n−1
ij
matrix with all "1" entries. Therefore H is a fractal function.
Remark 4.8. If n = m = 2, the definition of fractal functions agrees with Definition 3.3. in [5]. Indeed
it is enough to write each partition αij = (a1 , a2 , . . .) as a matrix Mij = (mhk ) ∈ F (i+1)×(j+1) where
mhk = 1 iff k ≤ ak otherwise mhk = 0. In this case the expansions are given by the elements in
Φ(2) := {(2), (1, 2), (1, 1, 2), (1, 1, 1, 2), . . .}.
In the following we denote by X := {x1 , · · · xn } and Y := {y1 , · · · , ym } the set of the variables of
degree (1, 0) and (0, 1) respectively.
Next lemma is useful for our purpose. It is an immediate consequence of Lemma 2.13.
10
GIUSEPPE FAVACCHIO
(d)
Lemma 4.9. x1 · Xa
(d+1)
= Xahdi .
To shorten the notation we set αi := n+i−1
and βj := m+j−1
n−1
m−1 . In order to relate fractal functions
and Hilbert functions of bigraded algebras we need to introduce a correspondence between Ferrers matrices
and monomials.
Let M = (mab ) ∈ F αi ×βj . We denote by λ(M )(ij) the set of the monomials
(j)
λ(M ) := {Xa(i) · Yb
Proposition 4.10. Let M = (mab ) ∈ F
αi ×βj
| mab = 0}.
. Then λ(M ) is a bilex set of monomials of bidegree (i, j).
(i)
(j)
(i)
(i)
Proof. We use Lemma 2.10 and Remark 3.1. Let Xa · Yb be an element in λ(M ) and Xu > Xa .
(i)
(j)
Since (u, b) > (a, b) and mab = 0 we get mub = 0 i.e. Xu · Yb ∈ λ(M ). In a similar way it follows that
(i)
(j)
Xa · Yv ∈ λ(M ) for v > b.
Let L ⊆ R(i,j) be a bilex set of monomials of bidegree (i, j). We denote by µ(L) ∈ M({0, 1})αi ×βj the
(j)
(i)
matrix (mab ) such that mab = 1 iff Xa · Yb
∈
/ L otherwise mab = 0.
Proposition 4.11. Let L ⊆ R(i,j) be a bilex set of monomials of bidegree (i, j). Then µ(L) ∈ F αi ×βj .
Proof. If follows by using Lemma 2.10 and Remark 3.1. Indeed, say mab = 0 for an entry of µ(L). This
(i)
(j)
(i)
(j)
implies Xa · Yb ∈ L. Thus, for u > a, we have Xu · Yb ∈ L i.e. mub = 0. Analogously, we see that
mav = 0 for v > b.
Proposition 4.10 and Proposition 4.11 together imply the following result.
Corollary 4.12. There is a one to one correspondence between the bilex sets of monomials of degree
(i, j) and the elements in F αi ×βj .
We are ready to prove the main result of this paper.
Theorem 4.13. Let H : N × N → N be a numerical function. Then the following are equivalent
1) H is a fractal function;
2) H = HR/I for some bilex ideal I ⊆ R = k[x1 , . . . , xn , y1 , . . . , ym ].
Proof. (1) → (2) Let H be a fractal function. For each (i, j) ∈ N2 , let I(i,j) be the k-vector space
spanned by the elements in λ(Mij ). Then we claim that I := ⊕(i,j)∈N2 I(i,j) is an ideal of R. To prove the
(i)
(j)
(i)
(j)
claim it is enough to show that if Xa · Yb ∈ I(i,j) then xu · Xa · Yb
(i)
(j)
yv · Xa · Yb ∈ I(i,j+1) for any yv ∈ Y. We have, see Lemma 4.9,
(j)
xu · Xa(i) · Yb
(j)
≥ x1 · Xa(i) · Yb
(i+1)
∈ I(i+1,j) for any xu ∈ X and
(j)
= Xahii · Yb .
(i+1)
Then, by Definition 4.6 and Theorem 2.14, the entry (ahii , b) of the matrix Mi+1,j is 0 and then Xahii ·
(j)
(i)
(j)
Yb ∈ I(i+1,j) and furthermore xu · Xa · Yb
I(i,j+1) .
(i)
(j)
∈ I(i+1,j) . In a similar way it follows that yv · Xa · Yb
∈
(2) → (1) Let I ⊆ R be a bilex ideal such that HR/I = H. Set Mij := µ(Iij ), we claim that the
Mij s satisfy the condition in Definition 4.6. By Theorem 2.14 it is enough to show that if Mij (a, b) = 0
(the entry (a, b) in Mij is 0), then also Mi+1,j (ahii , b) = 0 (the entry (ahii , b) in Mi+1j is 0). Set
(h)
J := (Xu | Mhj (u, b) = 0), then the claim is an immediate consequence of the fact that J is a lex ideal
of k[X].
FRACTAL SEQUENCES AND HILBERT FUNCTIONS
11
The following question is motivated by the argument of Section 3.
Question 4.14. Can the bigraded Betti numbers of a bilex ideal I = ⊕I(i,j) be computed from the
matrices µ(I(i,j) )?
References
[1] Aramova A, Crona K, De Negri E. Bigeneric initial ideals, diagonal subalgebras and bigraded Hilbert functions. Journal
of Pure and Applied Algebra. 2000 Jul 3;150(3):215-35.
[2] Bruns W, Herzog HJ. Cohen-macaulay rings. Cambridge University Press; 1998 Jun 18.
[3] CoCoATeam,
CoCoA: a system for doing Computations in Commutative Algebra. Available at
http://cocoa.dima.unige.it
[4] Eliahou S., Kervaire M., Minimal resolutions of some monomial ideals. J. Algebra 129 (1990) 1–25.
[5] Favacchio G. The Hilbert Function of Bigraded Algebras in k[P1 × P1 ]. Journal of Commutative Algebra. In Press 2017.
[6] Guardo E, Van Tuyl A. Arithmetically Cohen-Macaulay sets of points in P1 × P1 . SpringerBriefs in Mathematics,
Springer 2015.
[7] Herzog J, Hibi T. Monomial ideals. Springer. 2011.
[8] Kimberling C. Fractal sequences and interspersions. Ars Combinatoria. 1997; 45:157-68.
[9] Macaulay FS. Some properties of enumeration in the theory of modular systems. Proceedings of the London Mathematical Society. 1927 Jan 1;2(1):531-55.
[10] Peeva I, Stillman M. Open problems on syzygies and Hilbert functions. Journal of Commutative Algebra. 2009;1(1):15995.
[11] Stanley RP. Hilbert functions of graded algebras. Advances in Mathematics. 1978 Apr 1;28(1):57-83.
Dipartimento di Matematica e Informatica, Viale A. Doria, 6 - 95100 - Catania, Italy
E-mail address: [email protected]
URL: www.dmi.unict.it/~gfavacchio
| 0math.AC
|
Noname manuscript No.
(will be inserted by the editor)
A. B. Duncan · N. Nüsken · G. A. Pavliotis
arXiv:1705.00170v1 [math.PR] 29 Apr 2017
Using Perturbed Underdamped Langevin Dynamics
to Efficiently Sample from Probability Distributions
Received: date / Accepted: date
Abstract In this paper we introduce and analyse Langevin samplers that consist of perturbations
of the standard underdamped Langevin dynamics. The perturbed dynamics is such that its invariant
measure is the same as that of the unperturbed dynamics. We show that appropriate choices of the
perturbations can lead to samplers that have improved properties, at least in terms of reducing the
asymptotic variance. We present a detailed analysis of the new Langevin sampler for Gaussian target
distributions. Our theoretical results are supported by numerical experiments with non-Gaussian target
measures.
1 Introduction and Motivation
Sampling from probability measures in high-dimensional spaces is a problem that appears frequently
in applications, e.g. in computational statistical mechanics and in Bayesian statistics. In particular, we
are faced with the problem of computing expectations with respect to a probability measure π on Rd ,
i.e. we wish to evaluate integrals of the form:
ˆ
π(f ) :=
f (x)π(dx).
(1)
Rd
As is typical in many applications, particularly in molecular dynamics and Bayesian inference, the
density (for convenience denoted by the same symbol π) is known only up to a normalization constant;
furthermore, the dimension of the underlying space is quite often large enough to render deterministic
quadrature schemes computationally infeasible.
A standard approach to approximating such integrals is Markov Chain Monte Carlo (MCMC)
techniques [GCS+ 14, Liu08, RC13], where a Markov process (Xt )t≥0 is constructed which is ergodic
with respect to the probability measure π. Then, defining the long-time average
πT (f ) :=
1
T
ˆ
T
f (Xs )ds
(2)
0
A. B. Duncan
School of Mathematical and Physical Sciences, University of Sussex, Falmer, Brighton, BN1 9RH United
Kingdom E-mail: [email protected]
N. Nüsken
Imperial College London, Department of Mathematics, South Kensington Campus, London SW7 2AZ,England
E-mail: [email protected]
G. A. Pavliotis
Imperial College London, Department of Mathematics, South Kensington Campus, London SW7 2AZ,England
E-mail: [email protected]
2
for f ∈ L1 (π), the ergodic theorem guarantees almost sure convergence of the long-time average πT (f )
to π(f ).
There are infinitely many Markov, and, for the purposes of this paper diffusion, processes that
can be constructed in such a way that they are ergodic with respect to the target distribution. A
natural question is then how to choose the ergodic diffusion process (Xt )t≥0 . Naturally the choice
should be dictated by the requirement that the computational cost of (approximately) calculating (1)
is minimized. A standard example is given by the overdamped Langevin dynamics defined to be the
unique (strong) solution (Xt )t≥0 of the following stochastic differential equation (SDE):
√
(3)
dXt = −∇V (Xt )dt + 2dWt ,
where V = − log π is the potential associated with the smooth positive density π. Under appropriate
assumptions on V , i.e. on the measure π(dx), the process (Xt )t≥0 is ergodic and in fact reversible with
respect to the target distribution.
Another well-known example is the underdamped Langevin dynamics given by (Xt )t≥0 = (qt , pt )t≥0
defined on the extended space (phase space) Rd × Rd by the following pair of coupled SDEs:
dqt = M −1 pt dt,
dpt = −∇V (qt )dt − Γ M −1 pt dt +
√
2Γ dWt ,
(4)
where mass and friction tensors M and Γ , respectively, are assumed to be symmetric positive definite
matrices. It is well-known [Pav14, LS16] that (qt , pt )t≥0 is ergodic with respect to the measure π
b :=
π ⊗ N (0, M ), having density with respect to the Lebesgue measure on R2d given by
1
1
−1
π
b(q, p) = exp −V (q) − p · M p ,
(5)
b
2
Z
b is a normalization constant. Note that π
where Z
b has marginal π with respect to p and thus for
´t
functions f ∈ L1 (π), we have that 1t 0 f (qt ) dt → π(f ) almost surely. Notice also that the dynamics
restricted to the q-variables is no longer Markovian. The p-variables can thus be interpreted as giving
some instantaneous memory to the system, facilitating efficient exploration of the state space. Higher
order Markovian models, based on a finite dimensional (Markovian) approximation of the generalized
Langevin equation can also be used [CBP09].
As there is a lot of freedom in choosing the dynamics in (2), see the discussion in Section 2, it is
desirable to choose the diffusion process (Xt )t≥0 in such a way that πT (f ) can provide a good estimation of π(f ). The performance of the estimator (2) can be quantified in various manners. The
ultimate goal, of course, is to choose the dynamics as well as the numerical discretization in such a
way that the computational cost of the longtime-average estimator is minimized, for a given tolerance.
The minimization of the computational cost consists of three steps: bias correction, variance reduction
and choice of an appropriate discretization scheme. For the latter step see Section 5 and [DLP16, Sec.
6].
Under appropriate conditions on the potential V it can be shown that both (3) and (4) converge
to equilibrium exponentially fast, e.g. in relative entropy. One performance objective would then be
to choose the process (Xt )t≥0 so that this rate of convergence is maximised. Conditions on the potential V which guarantee exponential convergence to equilibrium, both in L2 (π) and in relative entropy
can be found in [MV00, BGL13]. A powerful technique for proving exponentially fast convergence to
equilibrium that will be used in this paper is C. Villani’s theory of hypocoercivity [Vil09]. In the case
when the target measure π is Gaussian, both the overdamped (3) and the underdamped (4) dynamics
become generalized Ornstein-Uhlenbeck processes. For such processes the entire spectrum of the generator – or, equivalently, the Fokker-Planck operator – can be computed analytically and, in particular,
an explicit formula for the L2 -spectral gap can be obtained [MPP02, OPPS12, OPPS15]. A detailed
analysis of the convergence to equilibrium in relative entropy for stochastic differential equations with
linear drift, i.e. generalized Ornstein-Uhlenbeck processes, has been carried out in [AE14].
In addition to speeding up convergence to equilibrium, i.e. reducing the bias of the estimator (2),
3
one is naturally also interested in reducing the asymptotic variance. Under appropriate conditions
on the target measure π and the observable f , the estimator πT (f ) satisfies a central limit theorem
(CLT) [KLO12], that is,
1
d
√ (πT (f ) − π(f )) −−−−→ N (0, 2 σf2 ),
T
→∞
T
where σf2 < ∞ is the asymptotic variance of the estimator πT (f ). The asymptotic variance characterises
how quickly fluctuations of πT (f ) around π(f ) contract to 0. Consequently, another natural objective is
to choose the process (Xt )t≥0 such that σf2 is as small as possible. It is well known that the asymptotic
variance can be expressed in terms of the solution to an appropriate Poisson equation for the generator
of the dynamics [KLO12]
ˆ
2
− Lφ = f − π(f ), σf =
φ(−Lφ) π(dx).
(6)
Rd
Techniques from the theory of partial differential equations can then be used in order to study the
problem of minimizing the asymptotic variance. This is the approach that was taken in [DLP16], see
also [HNW15], and it will also be used in this paper.
Other measures of performance have also been considered. For example, in [RBS15b, RBS15a],
performance
of the estimator is quantified in terms of the rate functional of the ensemble measure
´
1 t
δ
(dx).
See also [JO10] for a study of the nonasymptotic behaviour of MCMC techniques,
t 0 X(t)
including the case of overdamped Langevin dynamics.
Similar analyses have been carried out for various modifications of (3). Of particular interest to us
are the Riemannian manifold MCMC [GC11] (see the discussion in Section 2) and the nonreversible
Langevin samplers [HHMS93, HHMS05]. As a particular example of the general framework that was introduced in [GC11], we mention the preconditioned overdamped Langevin dynamics that was presented
in [AMO16]
√
dXt = −P ∇V (Xt ) dt + 2P dWt .
(7)
In this paper, the long-time behaviour of as well as the asymptotic variance of the corresponding
estimator πT (f ) are studied and applied to equilibrium sampling in molecular dynamics. A variant of
the standard underdamped Langevin dynamics that can be thought of as a form of preconditioning
and that has been used by practitioners is the mass-tensor molecular dynamics [Ben75].
The nonreversible overdamped Langevin dynamics
√
dXt = − (∇V (Xt ) − γ(Xt )) dt + 2 dWt ,
(8)
where the vector field γ satisfies ∇ · (πγ) = 0 is ergodic (but not reversible) with respect to the target
measure π for all choices of the divergence-free vector field γ. The asymptotic behaviour of this process
was considered for Gaussian diffusions in [HHMS93], where the rate of convergence of the covariance
to equilibrium was quantified in terms of the choice of γ. This work was extended to the case of
non-Gaussian target densities, and consequently for nonlinear SDEs of the form (8) in [HHMS05].
The problem of constructing the optimal nonreversible perturbation, in terms of the L2 (π) spectral
gap for Gaussian target densities was studied in [LNP13] see also [WHC14]. Optimal nonreversible
perturbations with respect to miniziming the asymptotic variance were studied in [DLP16, HNW15].
In all these works it was shown that, in theory (i.e. without taking into account the computational cost
of the discretization of the dynamics (8)), the nonreversible Langevin sampler (8) always outperforms
the reversible one (3), both in terms of converging faster to the target distribution as well as in
terms of having a lower asymptotic variance. It should be emphasized that the two optimality criteria,
maximizing the spectral gap and minimizing the asymptotic variance, lead to different choices for the
nonreversible drift γ(x).
The goal of this paper is to extend the analysis presented in [DLP16, LNP13] by introducing the
following modification of the standard underdamped Langevin dynamics:
dqt = M −1 pt dt − µJ1 ∇V (qt )dt,
dpt = −∇V (qt )dt − νJ2 M −1 pt dt − Γ M −1 pt dt +
√
2Γ dWt ,
(9)
4
where M, Γ ∈ Rd×d are constant strictly positive definite matrices, µ and ν are scalar constants and
J1 , J2 ∈ Rd×d are constant skew-symmetric matrices. As demonstrated in Section 3, the process defined
by (9) will be ergodic with respect to the Gibbs measure π
b defined in (5).
Our objective is to investigate the use of these dynamics for computing ergodic averages of the form
(2). To this end, we study the long time behaviour of (9) and, using hypocoercivity techniques, prove
that the process converges exponentially fast to equilibrium. This perturbed underdamped Langevin
process introduces a number of parameters in addition to the mass and friction tensors which must be
tuned to ensure that the process is an efficient sampler. For Gaussian target densities, we derive estimates for the spectral gap and the asymptotic variance, valid in certain parameter regimes. Moreover,
for certain classes of observables, we are able to identify the choices of parameters which lead to the
optimal performance in terms of asymptotic variance. While these results are valid for Gaussian target
densities, we advocate these particular parameter choices also for more complex target densities. To
demonstrate their efficacy, we perform a number of numerical experiments on more complex, multimodal distributions. In particular, we use the Langevin sampler (9) in order to study the problem of
diffusion bridge sampling.
The rest of the paper is organized as follows. In Section 2 we present some background material
on Langevin dynamics, we construct general classes of Langevin samplers and we introduce criteria for
assessing the performance of the samplers. In Section 3 we study qualitative properties of the perturbed
underdamped Langevin dynamics (9) including exponentially fast convergence to equilibrium and the
overdamped limit. In Section 4 we study in detail the performance of the Langevin sampler (9) for the
case of Gaussian target distributions. In Section 5 we introduce a numerical scheme for simulating the
perturbed dynamics (9) and we present numerical experiments on the implementation of the proposed
samplers for the problem of diffusion bridge sampling. Section 6 is reserved for conclusions and suggestions for further work. Finally, the appendices contain the proofs of the main results presented in
this paper and of several technical results.
2 Construction of General Langevin Samplers
2.1 Background and Preliminaries
In this section we consider estimators of the form (2) where (Xt )t≥0 is a diffusion process given by the
solution of the following Itô SDE:
√
(10)
dXt = a(Xt ) dt + 2b(Xt ) dWt ,
with drift coefficient a : Rd → Rd and diffusion coefficient b : Rd → Rd×m both having smooth
components, and where (Wt )t≥0 is a standard Rm –valued Brownian motion. Associated with (10) is
the infinitesimal generator L formally given by
L f = a · ∇f + Σ : D2 f,
f ∈ Cc2 (Rd )
(11)
where Σ = bb> , D2 f denotes the Hessian of the function f and : denotes the Frobenius inner product.
In general, Σ is nonnegative definite, and could possibly be degenerate. In particular, the infinitesimal
generator (11) need not be uniformly elliptic. To ensure that the corresponding semigroup exhibits
sufficient smoothing behaviour, we shall require that the process (10) is hypoelliptic in the sense of
Hörmander. If this condition holds, then irreducibility of the process (Xt )t≥0 will be an immediate
consequence of the existence of a strictly positive invariant distribution π(x)dx, see [Kli87].
Suppose that (Xt )t≥0 is nonexplosive. It follows from the hypoellipticity assumption that the process
(Xt )t≥0 possesses a smooth transition density p(t, x, y) which is defined for all t ≥ 0 and x, y ∈ Rd ,
[Bas98, Theorem VII.5.6]. The associated strongly continuous Markov semigroup (Pt )t≥0 is defined by
ˆ
Pt f (x) =
p(t, x, y)f (y) dy, t ≥ 0.
(12)
Rd
5
Suppose that (Pt )t≥0 is invariant with respect to the target distribution π(x) dx, i.e.
ˆ
ˆ
Pt f (x)π(x) dx =
f (x)π(x) dx, t ≥ 0,
Rd
Rd
for all bounded continuous functions f . Then (Pt )t≥0 can be extended to a positivity preserving
contraction semigroup on L2 (π) which is strongly continuous. Moreover, the infinitesimal generator
corresponding to (Pt )t≥0 is given by an extension of (L, Cc2 (Rd )), also denoted by L.
Due to hypoellipticity, the probability measure π on Rd has a smooth and positive density with respect
to the Lebesgue measure, and (slightly abusing the notation) we will denote this density also by π.
Let L2 (π) be the Hilbert space of π-square integrable functions equipped with inner product h·, ·iL2 (π)
and norm k·kL2 (π) . We will also make use of the Sobolev space
H 1 (π) = {f ∈ L2 (π) :
k∇f k2L2 (π) < ∞}
(13)
of L2 (π)-functions with weak derivatives in L2 (π), equipped with norm
kf k2H 1 (π) = kf k2L2 (π) + k∇f k2L2 (π).
2.2 A General Characterisation of Ergodic Diffusions
A natural question is what conditions on the coefficients a and b of (10) are required to ensure that
(Xt )t≥0 is invariant with respect to the distribution π(x) dx. The following result provides a necessary
and sufficient condition for a diffusion process to be invariant with respect to a given target distribution.
Theorem 1 Consider a diffusion process (Xt )t≥0 on Rd defined by the unique, non-explosive solution
to the Itô SDE (10) with drift a ∈ C 1 (Rd ; Rd ) and diffusion coefficient b ∈ C 1 (Rd ; Rd×m ). Then
(Xt )t≥0 is invariant with respect to π if and only if
a = Σ∇ log π + ∇ · Σ + γ,
(14)
where Σ = bb> and γ : RD → RD is a continuously differentiable vector field satisfying
∇ · (πγ) = 0.
(15)
If additionally γ ∈ L1 (π), then there exists a skew-symmetric matrix function C : Rd → Rd×d such
that
1
γ = ∇ · (πC) .
π
In this case the infinitesimal generator can be written as an L2 (π)-extension of
Lf =
1
∇ · ((Σ + C)π∇f ) ,
π
f ∈ Cc2 (Rd ).
The proof of this result can be found in [Pav14, Ch. 4]; similar versions of this characterisation can
be found in [Vil09] and [HHMS05]. See also [MCF15].
Remark 1 If (14) holds and L is hypoelliptic it follows immediately that (Xt )t≥0 is ergodic with unique
invariant distribution π(x) dx.
More generally, we can consider Itô diffusions in an extended phase space:
√
dZt = b(Zt ) dt + 2σ(Zt ) dWt ,
(16)
where (Wt )t≥0 is a standard Brownian motion in RN , N ≥ d. This is a Markov process with generator
L = b(z) · ∇z + Σ(z) : Dz2 ,
(17)
6
where Σ(z) = σσ T (z). We will consider dynamics (Zt )t≥0 that is ergodic with respect to πz (z) dz
such that
ˆ
πz (x, y) dy = π(x).
(18)
Rm
where z = (x, y), x ∈ Rd , y ∈ Rm , d + m = N .
There are various well-known choices of dynamics which are invariant (and indeed ergodic) with
respect to the target distribution π(x)dx.
1. Choosing b = I and γ = 0 we immediately recover the overdamped Langevin dynamics (3).
2. Choosing b = I, and γ 6= 0 such that (15) holds gives rise to the nonreversible overdamped
equation defined by (8). As it satisfies the conditions of Theorem 1, it is ergodic with respect to π.
In particular choosing γ(x) = J∇V (x) for a constant skew-symmetric matrix J we obtain
dXt = −(I + J)∇V (Xt ) dt +
√
2 dWt ,
(19)
which has been studied in previous works.
3. Given a target density π > 0 on Rd , if we consider the augmented target density π
b on R2d given in
(5), then choosing
M −1 p
γ((q, p)) =
(20)
−∇V (q)
and
b=
√0
Γ
∈ R2d×d ,
(21)
where M and Γ are positive definite symmetric matrices, the conditions of Theorem 1 are satisfied
for the target density π
b. The resulting dynamics (qt , pt )t≥0 is determined by the underdamped
Langevin equation (4). It is straightforward to verify that the generator is hypoelliptic, [LRS10,
Sec 2.2.3.1], and thus (qt , pt )t≥0 is ergodic.
4. More generally, consider the augmented target density π
b on R2d as above, and choose
M −1 p − µJ1 ∇V (q)
γ((q, p)) =
(22)
−∇V (q) − νJ2 M −1 p
and
b=
√0
Γ
∈ R2d×d ,
(23)
where µ and ν are scalar constants and J1 , J2 ∈ Rd×d are constant skew-symmetric matrices. With
this choice we recover the perturbed Langevin dynamics (9). It is straightforward to check that
(22) satisfies the invariance condition (15), and thus Theorem 1 guarantees that (9) is invariant
with respect to π
b.
5. In a similar fashion, one can introduce an augmented target density on R(m+2)d , with
|p|
b
π
b(q, p, u1 , . . . , um ) ∝ e− 2
where p, q, ui ∈ Rd , for i = 1, . . . , m. Clearly
now define γ : R(m+2)d → R(m+2)d by
´
Rd ×Rmd
2
−
2
u2
1 +...+um
2
−V (q)
,
b
π
b(q, p, u1 , . . . , um ) dp du1 . . . dum = π(q). We
pP
m
−∇q V (q) + j=1 λj uj
−λ1 p
γ(q, p, u1 , . . . , um ) =
..
.
−λm p
7
and b : R(m+2)d → R(m+2)d×(m+2)d by
00
0
0
0 0 √ 0
0
0 0 α1 Id×d √ 0
b(q, p, u1 , . . . , um ) =
α2 Id×d
0
0 0
. .
.
..
.. ..
..
.
00
0
0
...
...
...
...
..
.
...
√
0
0
0
0
..
.
,
αm Id×d
where λi ∈ R and αi > 0, for i = 1, . . . , m. The resulting process (10) is given by
dqt = pt dt
dpt = −∇q V (qt ) dt +
d
X
λj uj (t) dt
j=1
du1t = −λ1 pt dt − α1 u1t dt +
..
.
√
m
dum
t = −λm pt dt − αm ut dt +
2α1 dWt1
√
(24)
2αm dWtm ,
where (Wt1 )t≥0 , . . . (Wtm )t≥0 are independent Rd –valued Brownian motions. This process is ergodic
b
with unique invariant distribution π
b, and under appropriate conditions on V , converges exponentially fast to equilibrium in relative entropy [OP11]. Equation (24) is a Markovian representation
of a generalised Langevin equation of the form
dqt = pt dt
ˆ
t
dpt = −∇q V (qt ) dt −
F (t − s)ps ds + N (t),
0
where N (t) is a mean-zero stationary Gaussian process with autocorrelation function F (t), i.e.
E [N (t) ⊗ N (s)] = F (t − s)Id×d ,
and
F (t) =
m
X
λ2i e−αi |t| .
i=1
6. Let π
e(z) ∝ exp(−Φ(z)) be a positive density on RN where N > d such that
ˆ
π(x) =
RN −d
π
e(x, z) dz,
where (x, y) ∈ Rd × RN −d . Then choosing b = ID×D and γ = 0 we obtain the dynamics
dXt = −∇x Φ(Xt , Yt ) dt +
dYt = −∇y Φ(Xt , Yt ) dt +
√
√
then (Xt , Yt )t≥0 is immediately ergodic with respect to π
e.
2 dWt1
2 dWt2 ,
8
2.3 Comparison Criteria
For a fixed observable f , a natural measure of accuracy of the estimator πT (f ) = t−1
the mean square error (MSE) defined by
´t
0
f (Xs ) ds is
2
M SE(f, T ) := Ex |πT (f ) − π(f )| ,
(25)
where Ex denotes the expectation conditioned on the process (Xt )t≥0 starting at x. It is instructive to
introduce the decomposition M SE(f, T ) = µ2 (f, T ) + σ 2 (f, T ), where
µ(f, T ) = |Ex [πT (f )] − π(f )|
2
σ 2 (f, T ) = Ex |πT (f ) − π(f )| = Var[πT (f )].
and
(26)
Here µ(f, T ) measures the bias of the estimator πT (f ) and σ 2 (f, T ) measures the variance of fluctuations of πT (f ) around the mean.
The speed of convergence to equilibrium of the process (Xt )t≥0 will control both the bias term µ(f, T )
and the variance σ 2 (f, T ). To make this claim more precise, suppose that the semigroup (Pt )t≥0 associated with (Xt )t≥0 decays exponentially fast in L2 (π), i.e. there exist constants λ > 0 and C ≥ 1
such that
kPt g − π(g)kL2 (π) ≤ Ce−λt kg − π(g)kL2 (π) , g ∈ L2 (π).
(27)
Remark 2 If (27) holds with C = 1, this estimate is equivalent to − L having a spectral gap in L2 (π).
Allowing for a constant C > 1 is essential for our purposes though in order to treat nonreversible and
degenerate diffusion processes by the theory of hypocoercivity as outlined in [Vil09].
The following lemma characterises the decay of the bias µ(f, T ) as T → ∞ in terms of λ and C. The
proof can be found in Appendix A.
Lemma 1 Let (Xt )t≥0 be the unique, non-explosive solution of (10), such that X0 ∼ π0 π and
dπ0
dπ0
2
dπ ∈ L (π), where dπ denotes the Radon-Nikodym derivative of π0 with respect to π. Suppose that
the process is ergodic with respect to π such that the Markov semigroup (Pt )t≥0 satisfies (27). Then
for f ∈ L∞ (π),
1
C
dπ0 2
−λT
µ(f, T ) ≤
1−e
kf kL∞ Varπ
.
λT
dπ
The study of the behaviour of the variance σ 2 (f, T ) involves deriving a central limit theorem for
´t
the additive functional 0 f (Xt ) − π(f ) dt. As discussed in [CCG12], we reduce this problem to proving
well-posedness of the Poisson equation
− L χ = f − π(f ),
π(χ) = 0.
(28)
The only complications in this approach arise from the fact that the generator L need not be symmetric
in L2 (π) nor uniformly elliptic. The following result summarises conditions for the well-posedness of
the Poisson equation and it also provides with us with a formula for the asymptotic variance. The
proof can be found in Appendix A.
Lemma 2 Let (Xt )t≥0 be the unique, non-explosive solution of (10) with smooth drift and diffusion
coefficients, such that the corresponding infinitesimal generator is hypoelliptic. Syppose that (Xt )t≥0 is
ergodic with respect to π and moreover, (Pt )t≥0 decays exponentially fast in L2 (π) as in (27). Then
for all f ∈ L2 (π), there exists a unique mean zero solution χ to the Poisson equation (28). If X0 ∼ π,
then for all f ∈ C ∞ (Rd ) ∩ L2 (π)
√
d
T (πT (f ) − π(f )) −−−−→ N (0, 2σf2 ),
(29)
T →∞
where
σf2
is the asymptotic variance defined by
σf2 = hχ, (− L)χiL2 (π) = h∇χ, Σ∇χiL2 (π) .
Moreover, if X0 ∼ π0 where π0 π and
dπ0
dπ
(30)
∈ L2 (π) then (29) holds for all f ∈ C ∞ (Rd ) ∩ L∞ (π).
9
Clearly, observables that only differ by a constant have the same asymptotic variance. In the sequel,
we will hence restrict our attention to observables f ∈ L2 (π) satisfying π(f ) = 0, simplifying expressions (28) and (29). The corresponding subspace of L2 (π) will be denoted by
L20 (π) := {f ∈ L2 (π) : π(f ) = 0}.
(31)
If the exponential decay estimate (27) is satisfied, then Lemma 2 shows that − L is invertible on L20 (π),
so we can express the asymptoptic variance as
σf2 = hf, (− L)−1 f iL2 (π) ,
f ∈ L20 (π).
Let us also remark that from the proof of Lemma 2 it follows that the inverse of L is given by
ˆ ∞
Pt dt.
L−1 =
(32)
(33)
0
We note that the constants C and λ appearing in the exponential decay estimate (27) also control the
speed of convergence of σ 2 (f, T ) to zero. Indeed, it is straightforward to show that if (27) is satisfied,
then the solution χ of (28) satisfies
σf2 = hχ, f − π(f )iL2 (π) ≤
C
2
kf kL2 (π) .
λ
(34)
Lemmas 1 and 2 would suggest that choosing the coefficients Σ and γ to optimize the constants
C and λ in (34) would be an effective means of improving the performance of the estimator πT (f ),
especially since the improvement in performance would be uniform over an entire class of observables.
When this is possible, this is indeed the case. However, as has been observed in [LNP13, HHMS93,
HHMS05], maximising the speed of convergence to equilibrium is a delicate task. As the leading order
term in M SE(f, T ), it is typically sufficient to focus specifically on the asymptotic variance σf2 and
study how the parameters of the SDE (10) can be chosen to minimise σf2 . This study was undertaken
in [DLP16] for processes of the form (8).
3 Perturbation of Underdamped Langevin Dynamics
The primary objective of this work is to compare the performances of the perturbed underdamped
Langevin dynamics (9) and the unperturbed dynamics (4) according to the criteria outlined in Section
2.3 and to find suitable choices for the matrices J1 , J2 , M and Γ that improve the performance of the
sampler. We begin our investigations of (9) by establishing ergodicity and exponentially fast return to
equilibrium, and by studying the overdamped limit of (9). As the latter turns out to be nonreversible
and therefore in principle superior to the usual overdamped limit (3),e.g. [HHMS05], this calculation
provides us with further motivation to study the proposed dynamics.
For the bulk of this work, we focus on the particular case when the target measure is Gaussian,
i.e. when the potential is given by V (q) = 12 q T Sq with a symmetric and positive definite precision matrix S (i.e. the covariance matrix is given by S −1 ). In this case, we advocate the following conditions
for the choice of parameters:
M = S,
Γ = γS,
SJ1 S = J2 ,
µ = ν.
(35a)
(35b)
(35c)
(35d)
Under the above choices (35), we show that the large perturbation limit limµ→∞ σf2 exists and is finite
and we provide an explicit expression for it (see Theorem 5). From this expression, we derive an algorithm for finding optimal choices for J1 in the case of quadratic observables (see Algorithm 2).
10
√
If the friction coefficient is not too small (γ > 2), and under certain mild nondegeneracy conditions, we prove that adding a small perturbation will always decrease the asymptotic variance for
observables of the form f (q) = q · Kq + l · q + C:
d 2
σ
dµ f
=0
and
µ=0
d2 2
σ
dµ2 f
< 0,
µ=0
see Theorem 3. In fact, we conjecture that this statement is true for arbitrary observables f ∈ L2 (π),
but we have not been able to prove this. The dynamics (9) (used in conjunction with the conditions
(35a)-(35c)) proves to be especially effective when the observable is antisymmetric (i.e. when it is invariant under the substitution q 7→ −q) or when it has a significant antisymmetric part. In particular,
in Proposition 5 we show that under certain conditions on the spectrum of J1 , for any antisymmetric
observable f ∈ L2 (π) it holds that limµ→∞ σf2 = 0.
Numerical experiments and analysis show that departing significantly from 35c in fact possibly decreases the performance of the sampler. This is in stark contrast to (8), where it is not possible to
increase the asymptotic variance by any perturbation. For that reason, until now it seems practical to
use (9) as a sampler only when a reasonable estimate of the global covariance of the target distribution
is available. In the case of Bayesian inverse problems and diffusion bridge sampling, the target measure
π is given with respect to a Gaussian prior. We demonstrate the effectiveness of our approach in these
applications, taking the prior Gaussian covariance as S in (35a)-(35c).
Remark 3 In [LNP13, Rem. 3] another modification of (4) was suggested (albeit with the simplifications
Γ = γ · I and M = I):
dqt = (1 − J)M −1 pt dt,
dpt = −(1 + J)∇V (qt )dt − Γ M −1 pt dt +
√
2Γ dWt ,
(36)
J again denoting an antisymmetric matrix. However, under the change of variables p 7→ (1 + J)p̃ the
above equations transform into
dqt = M̃ −1 pt dt,
dp˜t = −∇V (qt )dt − Γ̃ M̃ −1 p̃t dt +
p
2Γ̃ dW̃t ,
where M̃ = (1 + J)−1 M (1 − J)−1 and Γ̃ = (1 + J)−1 Γ (1 − J)−1 . Since any observable f depends only
on q (the p-variables are merely auxiliary), the estimator πT (f ) as well as its associated convergence
characteristics (i.e. asymptotic variance and speed of convergence to equilibrium) are invariant under
this transformation. Therefore, (36) reduces to the underdamped Langevin dynamics (4) and does not
represent an independent approach to sampling. Suitable choices of M and Γ will be discussed in
Section 4.5.
3.1 Properties of Perturbed Underdamped Langevin Dynamics
In this section we study some of the properties of the perturbed underdamped dynamics (9). First,
note that its generator is given by
L = M −1 p · ∇q − ∇q V · ∇p −Γ M −1 p · ∇p + Γ : Dp2 −µJ1 ∇V · ∇q − νJM −1 p · ∇p ,
|
{z
}|
{z
}
{z
}|
Lham
|
Ltherm
{z
L0
(37)
Lpert
}
decomposed into the perturbation Lpert and the unperturbed operator L0 , which can be further split
into the Hamiltonian part Lham and the thermostat (Ornstein-Uhlenbeck) part Ltherm , see [Pav14,
LRS10, LS16].
Lemma 3 The infinitesimal generator L (37) is hypoelliptic.
11
t
u
Proof See Appendix B.
An immediate corollary of this result and of Theorem 1 is that the perturbed underdamped Langevin
process (9) is ergodic with unique invariant distribution π
b given by (5).
As explained in Section 2.3, the exponential decay estimate (27) is crucial for our approach, as in
particular it guarantees the well-posedness of the Poisson equation (28). From now on, we will therefore make the following assumption on the potential V, required to prove exponential decay in L2 (π):
Assumption 1 Assume that the Hessian of V is bounded and that the target measure π(dq) =
1 −V
dq satisfies a Poincare inequality, i.e. there exists a constant ρ > 0 such that
Ze
ˆ
ˆ
|∇φ|2 dπ,
(38)
φ2 dπ ≤ ρ
Rd
Rd
holds for all φ ∈ L20 (π) ∩ H 1 (π).
Sufficient conditions on the potential so that Poincaré’s inequality holds, e.g. the Bakry-Emery criterion, are presented in [BGL13].
Theorem 2 Under Assumption 1 there exist constants C ≥ 1 and λ > 0 such that the semigroup
(Pt )t≥0 generated by L satisfies exponential decay in L2 (π) as in (27).
Proof See Appendix B.
Remark 4 The proof uses the machinery of hypocoercivity developed in [Vil09]. However, it seems
likely that using the framework of [DMS15], the assumption on the boundedness of the Hessian of V
can be substantially weakened.
3.2 The Overdamped Limit
In this section we develop a connection between the perturbed underdamped Langevin dynamics
(9) and the nonreversible overdamped Langevin dynamics (8). The analysis is very similar to the one
presented in [LRS10, Section 2.2.2] and we will be brief. For convenience in this section we will perform
the analysis on the d-dimensional torus Td ∼
= (R/Z)d , i.e. we will assume q ∈ Td . Consider the following
scaling of (9):
1 −1
M pt , dt − µJ1 ∇q V (qt )dt,
1
1
1
1√
dpt = − ∇q V (qt )dt − 2 νJ2 M −1 pt dt − 2 Γ M −1 pt dt +
2Γ dWt ,
dqt =
(39a)
(39b)
valid for the small mass/small momentum regime
M → 2 M,
pt → pt .
Equivalently, those modifications can be obtained from subsituting Γ → −1 Γ and t 7→ −1 t, and so in
the limit as → 0 the dynamics (39) describes the limit of large friction with rescaled time. It turns
out that as → 0, the dynamics (39) converges to the limiting SDE
√
dqt = −(νJ2 + Γ )−1 ∇q V (qt )dt − µJ1 ∇q V (qt )dt + (νJ2 + Γ )−1 2Γ dWt .
(40)
The following proposition makes this statement precise.
Proposition 1 Denote by (qt , pt ) the solution to (39) with (deterministic) initial conditions (q0 , p0 ) =
(qinit , pinit ) and by qt0 the solution to (40) with initial condition q00 = qinit . For any T > 0, (qt )0≤t≤T
converges to (qt0 )0≤t≤T in L2 (Ω, C([0, T ]), Td ) as → 0, i.e.
lim E sup |qt − qt0 |2 = 0.
→0
0≤t≤T
12
Remark 5 By a refined analysis, it is possible to get information on the rate of convergence; see,
e.g. [PS03, PS05].
The limiting SDE (40) is nonreversible due to the term −µJ1 ∇q V (qt )dt and also because the matrix
(νJ2 + Γ )−1 is in general neither symmetric nor antisymmetric. This result, together with the fact that
nonreversible perturbations of overdamped Langevin dynamics of the form (8) are by now well-known
to have improved performance properties, motivates further investigation of the dynamics (9).
Remark 6 The limit we described in this section respects the invariant distribution, in the sense that
the limiting dynamics (40) is ergodic with respect to the measure π(dq) = Z1 e−V dq. To see this, we
have to check that (we are using the notation ∇ instead of ∇q )
L† (e−V ) = −∇ · (νJ2 + Γ )−1 ∇e−V + ∇ · (µJ1 ∇e−V ) + ∇ · (νJ2 + Γ )−1 Γ (−νJ2 + Γ )−1 ∇e−V = 0,
where L† refers to the L2 (Rd )-adjoint of the generator of (40), i.e. to the associated Fokker-Planck
operator. Indeed, the term ∇ · (µe−V J1 ∇V ) vanishes because of the antisymmetry of J1 . Therefore, it
remains to show that
∇ · (νJ2 + Γ )−1 Γ (−νJ2 + Γ )−1 − (νJ2 + Γ )−1 ∇e−V = 0,
i.e. that the matrix (νJ2 + Γ )−1 Γ (−νJ2 + Γ )−1 − (νJ2 + Γ )−1 is antisymmetric. Clearly, the first term
is symmetric and furthermore it turns out to be equal to the symmetric part of the second term:
1
1
(νJ2 + Γ )−1 + (−νJ2 + Γ )−1 = = (νJ2 + Γ )−1 (−νJ2 + Γ )(−νJ2 + Γ )−1
2
2
+(νJ2 + Γ )−1 (νJ2 + Γ )(−νJ2 + Γ )−1
= (νJ2 + Γ )−1 Γ (−νJ2 + Γ )−1 ,
so π is indeed invariant under the limiting dynamics (40).
4 Sampling from a Gaussian Distribution
In this section we study in detail the performance of the Langevin sampler (9) for Gaussian target
densities, first considering the case of unit covariance. In particular, we study the optimal choice for the
parameters in the sampler, the exponential decay rate and the asymptotic variance. We then extend
our results to Gaussian target densities with arbitrary covariance matrices.
4.1 Unit covariance - small perturbations
In our study of the dynamics given by (9) we first consider the simple case when V (q) = 21 |q|2 , i.e. the
task of sampling from a Gaussian measure with unit covariance. We will assume M = I, Γ = γI and
J1 = J2 =: J (so that the q− and p−dynamics are perturbed in the same way, albeit posssibly with
different strengths µ and ν). Using these simplifications, (9) reduces to the linear system
dqt = pt dt − µJqt dt,
dpt = −qt dt − νJpt dt − γpt dt +
p
2γdWt .
The above dynamics are of Ornstein-Uhlenbeck type, i.e. we can write
p
dXt = −BXt dt + 2QdW̄t
(41)
(42)
with X = (q, p)T ,
µJ
−I
B=
I γI + νJ
0 0
Q=
0 γI
,
(43)
(44)
13
and (W̄t )t≥0 denoting a standard Wiener process on R2d . The generator of (42) is then given by
L = −Bx · ∇ + ∇T Q∇.
(45)
We will consider quadratic observables of the form
f (q) = q · Kq + l · q + C,
d
2
with K ∈ Rd×d
sym , l ∈ R and C ∈ R, however it is worth recalling that the asymptotic variance σf
does not depend on C. We also stress that f is assumed to be independent of p as those extra degrees
of freedom are merely auxiliary. Our aim will be to study the associated asymptotic variance σf2 , see
equation (30), in particular its dependence on the parameters µ and ν. This dependence is encoded in
the function
R2 → R
Θ:
(µ, ν) 7→ σf2 ,
assuming a fixed observable f and perturbation matrix J. In this section we will focus on small
perturbations, i.e. on the behaviour of the function Θ in the neighbourhood of the origin. Our main
theoretical tool will be the Poisson equation (28), see the proofs in Appendix C. Anticipating the
forthcoming analysis, let us already state our main result, showing that in the neighbourhood of the
origin, the function Θ has favourable properties along the diagonal µ = ν (note that the perturbation
strengths in the first and second line of (46) coincide):
Theorem 3 Consider the dynamics
dqt = pt dt − µJqt dt,
dpt = −qt dt − µJpt dt − γpt dt +
p
2γdWt ,
(46)
√
with γ > 2 and an observable of the form f (q) = q · Kq + l · q + C. If at least one of the conditions
[J, K] 6= 0 and l ∈
/ ker J is satisfied, then the asymptotic variance√of the unperturbed sampler is at a
local maximum independently of K and J (and γ, as long as γ > 2), i.e.
∂µ σf2
µ=0
∂µ2 σf2
µ=0
=0
and
< 0.
4.1.1 Purely quadratic observables
Let us start with the case l = 0, i.e. f (q) = q · Kq + C. The following holds:
Proposition 2 The function Θ satisfies
∇Θ|(µ,ν)=(0,0) = 0
(47)
and
Hess Θ|(µ,ν)=(0,0) =
Proof See Appendix C.
−(γ +
1
γ3
+ γ 3 ) Tr(JKJK) − Tr(J 2 K 2 )
( γ13 + γ1 − γ) Tr(J 2 K 2 )
2
1
1
− γ Tr(JKJK)
+(− γ 3 + γ + γ) Tr(JKJK)
.
1
1
2 2
( γ13 − γ1 ) Tr(J 2 K 2 )
( γ 3 + γ − γ) Tr(J K )
1
1
1
1
+(− γ 3 + γ + γ) Tr(JKJK)
−( γ 3 + γ ) Tr(JKJK)
(48)
t
u
14
The above proposition shows that the unperturbed dynamics represents a critical point of Θ, independently of the choice of K, J and γ. In general though, Hess Θ|(µ,ν)=(0,0) can have both positive
and negative eigenvalues. In particular, this implies that an unfortunate choice of the perturbations
will actually increase the asymptotic variance of the dynamics (in contrast to the situation of perturbed overdamped Langevin dynamics, where any nonreversible perturbation leads to an improvement in asymptotic variance as detailed in [HNW15] and [DLP16]). Furthermore, the nondiagonality
of Hess Θ|(µ,ν)=(0,0) hints at the fact that the interplay of the perturbations J1 and J2 (or rather their
relative strengths µ and ν) is crucial for the performance of the sampler and, consequently, the effect
of these perturbations cannot be satisfactorily studied independently.
Example 1 Assuming J 2 = −I and [J, K] = 0 it follows that
∂µ2 Θ
µ=0
= ∂ν2 Θ
µ=0
=
1
Tr(K 2 ) > 0,
γ
for all nonzero K. Therefore in this case, a small perturbation of J1 only or J2 only will increase the
asymptotic variance, uniformly over all choices of K and γ.
However, it turns out that it is possible to construct an improved sampler by combining both perturbations in a suitable way. Indeed, the function Θ can be seen to have good properties along µ = ν. We
set µ(s) = s, ν(s) := s and compute
d2
Θ
ds2
= (1, 1) · Hess Θ|(µ,ν)=(0,0) (1, 1)
s=0
2
1
= −(γ + 3 + γ 3 ) Tr(JKJK) − Tr(J 2 K 2 ) − Tr(JKJK)
γ
γ
1
1
1
1
+ 2 · ( 3 + − γ) Tr(J 2 K 2 ) + (− 3 + + γ) Tr(JKJK)
γ
γ
γ
γ
1
1
1
1
+ ( 3 − ) Tr(J 2 K 2 ) − ( 3 + ) Tr(JKJK)
γ
γ
γ
γ
4
1
= γ − 3 − γ3 −
· (Tr(JKJK) − Tr(J 2 K 2 )) ≤ 0.
γ
γ
The last inequality follows from
γ−
4
1
− γ3 − < 0
γ3
γ
and
Tr(JKJK) − Tr(J 2 K 2 ) ≥ 0
(both inequalities are proven in the Appendix, Lemma 10), where the last inequality is strict if [J, K] 6=
0. Consequently, choosing both perturbations to be of the same magnitude (µ = ν) and assuring that
J and K do not commute always leads to a smaller asymptotic variance, independently of the choice
of K, J and γ. We state this result in the following corrolary:
Corollary 1 Consider the dynamics
dqt = pt dt − µJqt dt,
dpt = −qt dt − µJpt dt − γpt dt +
p
2γdWt ,
(49)
and a quadratic observable f (q) = q · Kq + C. If [J, K] 6= 0, then the asymptotic variance of the
unperturbed sampler is at a local maximum independently of K, J and γ, i.e.
∂µ σf2
µ=0
∂µ2 σf2
µ=0
=0
and
< 0.
15
Remark 7 As we will see in Section 4.3, more precisely Example 3, if [J, K] = 0, the asymptotic
variance is constant as a function of µ, i.e. the perturbation has no effect.
Example 2 Let us set µ(s) := s and ν(s) := −s (this corresponds to a small perturbation with
J∇V (qt )dt in q and −Jpt dt in p). In this case we get
d2 Θ
ds2
s=0
1 γ 4 + 3γ 2 + 5
Tr(J 2 K 2 )
=− ·
Tr(JKJK) − Tr(J 2 K 2 ) −4
,
2
γ
γ
|
{z
}|
{z
}
≤0
≥0
which changes its sign depending on J and K as the first term is negative and the second is positive.
Whether the perturbation improves the performance of the sampler in terms of asymptotic variance
therefore depends on the specifics of the observable and the perturbation in this case.
4.1.2 Linear observables
Here we consider the case K = 0, i.e. f (q) = l · q + C, where again l ∈ Rd and C ∈ R. We have the
following result:
Proposition 3 The function Θ satisfies
∇Θ|(µ,ν)=(0,0) = 0
and
Hess Θ|(µ,ν)=(0,0) =
−2γ 3 |Jl|2 2γ|Jl|2
2γ|Jl|2
0
.
t
u
Proof See Appendix C.
Let us assume that l ∈
/ ker J. Then ∂µ2 Θ|µ,ν=0 < 0, and hence Theorem 3 shows that a small perturbation by µJ∇V (qt )dt alone always results in an improvement of the asymptotic variance. However,
if we combine both perturbations µJ∇V (qt )dt and νJpt dt, then the effect depends on the sign of
−2γ 3 |Jl|2 2γ|Jl|2
µ
µν
= −(2µ2 γ 3 − 4µνγ)|Jl|2 .
ν
2γ|Jl|2
0
This will be negative if µ and ν have different signs, and also if they have the same sign and γ is big
enough. Following Section 4.1.1, we require µ = ν. We then end up with the requirement
2µ2 γ 3 − 4µνγ > 0,
√
which is satisfied if γ > 2
Summarizing the results of this section, for observables of the form f (q)√= q · Kq + l · q + C,
choosing equal perturbations (µ = ν) with a sufficiently strong damping (γ > 2) always leads to an
improvement in asymptotic variance under the conditions [J, K] 6= 0 and l ∈
/ ker J. This is finally the
content of Theorem 3.
Let us illustrate the results of this section by plotting the asymptotic variance as a function of the
perturbation strength µ (see Figure 1), making the choices d = 2, l = (1, 1)T ,
20
0 1
K=
and J =
.
(50)
01
−1 0
The asymptotic variance has been computed according to (114), using (113a) and (113b) from Appendix C. The graphs confirm the results summarized in Corollary 3 concerning the asymptotic variance
in the neighbourhood of the unperturbed dynamics (µ = 0). Additionally, they give an impression of
the global behaviour, i.e. for larger values of µ.
Figures 1a, 1b and 1c show the asymptotic variance associated with the quadratic observable
f (q) = q · Kq. In accordance with Corollary 1, the asymptotic variance is at a local maximum at zero
perturbation in the case µ = ν (see Figure 1a). For increasing perturbation strength, the graph shows
that it decays monotonically and reaches a limit for µ → ∞ (this limiting behaviour will be explored
16
Quadratic observable, µ=0.9ν, γ=2.5
29
28.5
28.5
asymptotic variance σ2f
asymptotic variance σ2f
Quadratic observable, µ=ν, γ=2.5
29
28
27.5
27
28
27.5
27
26.5
26.5
26
−5
−4
−3
−2
−1
0
1
2
3
4
26
−5
5
−4
−3
−2
(a) Equal perturbations: µ = ν
−1
0
1
2
3
4
5
perturbation strength µ
perturbation strength µ
(b) Approximately equal perturbations: µ = 0.9ν
Linear observable, µ=ν, γ=2.5
10
Quadratic observable, µ=−ν, γ=2.5
400
9
350
asymptotic variance σ2f
8
asymptotic variance σ2f
300
250
200
150
7
6
5
4
3
2
100
1
50
0
−5
0
−5
−4
−3
−2
−1
0
1
2
3
4
5
perturbation strength µ
−4
−3
−2
−1
0
1
2
3
4
5
perturbation strength µ
(d) Equal perturbations: µ = ν (sufficiently large friction γ)
(c) Opposing perturbations: µ = −ν
Linear observable, µ=ν, γ=1
6
asymptotic variance σ2f
5
4
3
2
1
0
−5
−4
−3
−2
−1
0
1
2
3
4
5
perturbation strength µ
(e) Equal perturbations: µ = ν (small friction γ)
Fig. 1: Asymptotic variance for linear and quadratic observables, depending on relative perturbation
and friction strengths
17
analytically in Section 4.3). If the condition µ = ν is only approximately satisfied (Figure 1b), our
numerical examples still exhibits decaying asymptotic variance in the neighbourhood of the critical
point. In this case, however, the asymptotic variance diverges for growing values of the perturbation
µ. If the perturbations are opposed (µ = −ν) as in Example 2, it is possible for certain observables
that the unperturbed dynamics represents a global minimum. Such a case is observed in Figure 1c.
In Figures
1d and 1e the observable f (q) = l · q is considered. If the damping is sufficiently strong
√
(γ > 2), the unperturbed dynamics is at a local maximum of the asymptotic variance (Figure 1d).
Furthermore, the asymptotic variance approaches zero as µ → ∞ (for a theoretical explanation see
again Section 4.3). The graph in Figure 1e shows that the assumption of γ not being too small cannot
be dropped from Corollary 3. Even in this case though the example shows decay of the asymptotic
variance for large values of µ.
4.2 Exponential decay rate
Let us denote by λ∗ the optimal exponential decay rate in (27), i.e.
λ∗ = sup{λ > 0 | There exists C ≥ 1 such that (27) holds}.
(51)
Note that λ∗ is well-defined and positive by Theorem 2. We also define the spectral bound of the
generator L by
s(L) = inf(Re σ(− L) \ {0}).
(52)
In [MPP02] it is proven that the Ornstein-Uhlenbeck semigroup (Pt )t≥0 considered in this section
is differentiable (see Proposition 2.1). In this case (see Corollary 3.12 of [EN00]), it is known that
the exponential decay rate and the spectral bound coincide, i.e. λ∗ = s(L), whereas in general only
λ∗ ≤ s(L) holds. In this section we will therefore analyse the spectral properties of the generator (45).
In particular, this leads to some intuition of why choosing equal perturbations (µ = ν) is crucial for
the performance of the sampler.
In [MPP02] (see also [OPPS12]), it was proven that the spectrum of L as in (45) in L2 (π
b) is given
by
)
( r
X
nj λj : nj ∈ N, λj ∈ σ(B) .
(53)
σ(L) = −
j=1
Note that σ(L) only depends on the drift matrix B. In the case where µ = ν, the spectrum of B can
be computed explicitly.
Lemma 4 Assume µ = ν. Then the spectrum of B is given by
r
r
γ 2
γ
γ 2
γ
− 1 + |λ ∈ σ(J)} ∪ {µλ −
− 1 + |λ ∈ σ(J) .
σ(B) = µλ +
2
2
2
2
Proof We will compute σ B − γ2 I and then use the identity
n
γ
γ o
σ(B) = λ + |λ ∈ σ B − I
.
2
2
We have
γ
γ
γ
det B − I − λI = det µJ − I − λI µJ + I − λI + I
2
2
2
γ 2
2
= det (µJ − λI) −
I +I
2
!
!!
r
r
γ 2
γ 2
= det
µJ − λI +
− 1I · µJ − λI −
− 1I
2
2
!
!
r
r
γ 2
γ 2
= det µJ − λI +
− 1I · det µJ − λI −
− 1I ,
2
2
(54)
(55)
18
4
8=4
3
2
8=3
1
0
8=2
-1
-2
8=1
-3
-4
8=0
0
(a) σ(− L) in the case µ = ν. The arrows indicate
the movement of the spectrum as the perturbation
strength µ increases.
0.5
1
1.5
2
(b) σ(B) in the case J1 = 0, i.e. the dynamics is
only perturbed by −νJ2 pdt. The arrows indicate
the movement of the eigenvalues as ν increases. .
Fig. 2: Effects of the perturbation on the spectra of − L and B.
where I is understood to denote the identity matrix of appropriate dimension. The above quantity is
zero if and only if
r
γ 2
λ−
− 1 ∈ σ(µJ)
2
or
r
γ 2
λ+
− 1 ∈ σ(µJ).
2
t
u
Together with (55), the claim follows.
Using formula (53), in Figure 2a we show a sketch of the spectrum σ(−L) for the case of equal
perturbations (µ = ν) with the convenient choices n = 1 and γ = 2. Of course, the eigenvalue at
0 is associated to the invariant measure since σ(−L) = σ(−L† ) and L† π
b = 0, where L† denotes
the Fokker-Planck operator, i.e. the L2 (R2d )-adjoint of L. The arrows indicate the movement of the
eigenvalues as the perturbation µ increases in accordance with Lemma 4. Clearly, the spectral bound
of L is not affected by the perturbation. Note that the eigenvalues on the real axis stay invariant
under the perturbation. The subspace of L20 (π
b) associated to those will turn out to be crucial for the
characterisation of the limiting asymptotic variance as µ → ∞.
To illustrate the suboptimal properties of the perturbed dynamics when the perturbations are not
equal, we plot the spectrum of the drift matrix σ(B) in the case when the dynamics is only perturbed
by the term νJ2 pdt (i.e. µ = 0) for n = 2, γ = 2 and
J2 =
0 −1
1 0
,
(56)
(see Figure 2b). Note that the full spectrum σ(−L) can be inferred from (53). For ν = 0 we have that
the spectrum σ(B) only consists of the (degenerate) eigenvalue 1. For increasing ν, the figure shows
that the degenerate eigenvalue splits up into four eigenvalues, two of which get closer to the imaginary
axis as ν increases, leading to a smaller spectral bound and therefore to a decrease in the speed of
convergence to equilibrium. Figures (2a) and (2b) give an intuitive explanation of why the fine-tuning
of the perturbation strengths is crucial.
19
4.3 Unit covariance - large perturbations
In the previous subsection we observed that for the particular perturbation J1 = J2 and µ = ν, i.e.
dqt = pt dt − µJqt dt
dpt = −qt dt − µJpt dt − γpt dt +
p
2γ dWt ,
(57)
the perturbed Langevin dynamics demonstrated an improvement in performance for µ in a neighbourhood of 0, when the observable is linear or quadratic. Recall that this dynamics is ergodic with respect
to a standard Gaussian measure π
b on R2d with marginal π with respect to the q–variable. In the following we shall consider only observables that do not depend on p. Moreover, we assume without loss
of generality that π(f ) = 0. For such an observable we will write f ∈ L20 (π) and assume the canonical
embedding L20 (π) ⊂ L2 (π
b). The infinitesimal generator of (57) is given by
L = p · ∇q − q · ∇p + γ(−p · ∇p + ∆p ) +µ (−Jq · ∇q − Jp · ∇p ) =: L0 + µA,
{z
}
|
{z
}
|
L0
(58)
A
where we have introduced the notation Lpert = µA. In the sequel, the adjoint of an operator B in
L2 (π
b) will be denoted by B ∗ . In the rest of this section we will make repeated use of the Hermite
polynomials
gα (x) = (−1)|α| e
|x|2
2
∇α e −
|x|2
2
,
α ∈ N2d ,
(59)
2d
invoking the notation x = (q, p) ∈ R . For m ∈ N0 define the spaces
Hm = span{gα : |α| = m},
with induced scalar product
hf, gim := hf, giL2 (πb) ,
f, g ∈ Hm .
The space (Hm , h·, ·im ) is then a real Hilbert space with (finite) dimension
m + 2d − 1
dim Hm =
.
m
The following result (Theorem 4) holds for operators of the form
L = −Bx · ∇ + ∇T Q∇,
(60)
where the quadratic drift and diffusion matrices B and Q are such that L is the generator of an
ergodic stochastic process (see [AE14, Definition 2.1] for precise conditions on B and Q that ensure
ergodicity). The generator of the SDE (57) is given by (60) with B and Q as in equations (43) and
(44), respectively. The following result provides an orthogonal decomposition of L2 (π
b) into invariant
subspaces of the operator L.
Theorem 4 [AE14, Section 5]. The following holds:
(a) The space L2 (π
b) has a decomposition into mutually orthogonal subspaces:
M
L2 (π
b) =
Hm .
m∈N0
(b) For all m ∈ N0 , Hm is invariant under L as well as under the semigroup (e−tL )t≥0 .
(c) The spectrum of L has the following decomposition:
[
σ(L) =
σ(L|Hm ),
m∈N0
where
σ(L|Hm ) =
( 2d
X
j=1
)
αj λj : |α| = m, λj ∈ σ(B) .
(61)
20
Remark 8 Note that by the ergodicity of the dynamics, ker L consists of constant functions and so
ker L = H0 . Therefore, L20 (π
b) has the decomposition
L20 (π
b) = L2 (π
b)/ ker L =
M
Hm .
m≥1
Our first main result of this section is an expression for the asymptotic variance in terms of the
unperturbed operator L0 and the perturbation A:
Proposition 4 Let f ∈ L20 (π) (so in particular f = f (q)). Then the associated asymptotic variance
is given by
σf2 = hf, −L0 (L20 + µ2 A∗ A)−1 f iL2 (πb) .
(62)
Remark 9 The proof of the preceding Proposition will show that L20 + µ2 A∗ A is invertible on L20 (π
b)
and that (L20 + µ2 A∗ A)−1 f ∈ D(L0 ) for all f ∈ L20 (π
b).
To prove Proposition 4 we will make use of the generator with reversed perturbation
L− = L0 − µA
and the momentum flip operator
P : L20 (π
b) → L20 (π
b)
φ(q, p) 7→ φ(q, −p).
Clearly, P 2 = I and P ∗ = P . Further properties of L0 , A and the auxiliary operators L− and P are
gathered in the following lemma:
Lemma 5 For all φ, ψ ∈ C ∞ (R2d ) ∩ L2 (π
b) the following holds:
(a) The generator L0 is symmetric in L2 (π
b) with respect to P :
hφ, P L0 P ψiL2 (πb) = hL0 φ, ψiL2 (πb) .
(b) The perturbation A is skewadjoint in L2 (π
b):
A∗ = −A.
(c) The operators L0 and A commute:
[L0 , A]φ = 0.
(d) The perturbation A satisfies
P AP φ = Aφ.
(e) L and L− commute,
[L, L− ]φ = 0,
and the following relation holds:
hφ, P LP ψiL2 (πb) = hL− φ, ψiL2 (πb) .
(63)
(f ) The operators L, L0 , L− , A and P leave the Hermite spaces Hm invariant.
Remark 10 The claim (c) in the above lemma is crucial for our approach, which itself rests heavily on
the fact that the q− and p−perturbations match (J1 = J2 ).
21
Proof (of Lemma 5) To prove (a), consider the following decomposition of L0 as in (37):
L0 = p · ∇q − q · ∇p + γ (−p · ∇p + ∆p ) .
{z
} |
{z
}
|
Lham
Ltherm
By partial integration it is straightforward to see that
hφ, Lham ψiL2 (πb) = −hLham φ, ψiL2 (πb)
and
hφ, Ltherm ψiL2 (πb) = hLtherm φ, ψiL2 (πb) ,
for all φ, ψ ∈ C ∞ (R2d ) ∩ L2 (π
b), i.e. Lham and Ltherm are antisymmetric and symmetric in L2 (π
b)
respectively. Furthermore, we immediately see that P Lham P φ = −Lham φ and P Ltherm P φ = Ltherm φ,
so that
hφ, P L0 P ψiL2 (πb) = hφ, −Lham ψ + Ltherm ψiL2 (πb) = hL0 φ, ψiL2 (πb) .
We note that this result holds in the more general setting of Section 3 for the infinitesimal generator
(37). The claim (b) follows by noting that the flow vector field b(q, p) = (−Jq, −Jp) associated to
A is divergence-free with respect to π
b, i.e. ∇ · (π
bb) = 0. Therefore, A is the generator of a strongly
continuous unitary semigroup on L2 (π
b) and hence skewadjoint by Stone’s Theorem. To prove (c) we
use the decomposition L0 = Lham + Ltherm to obtain
[L0 , A]φ = [Lham , A]φ + [Ltherm , A]φ,
φ ∈ C ∞ (R2d ) ∩ L2 (π
b).
(64)
The first term of (64) gives
[p · ∇q − q · ∇p , −Jq · ∇q − Jp · ∇p ]φ
= [p · ∇q , −Jq · ∇q ] + [p · ∇q , −Jp · ∇p ] + [−q · ∇p , −Jq · ∇q ]
+ [−q · ∇p , −Jp · ∇p ] φ
= Jp · ∇q φ − Jp · ∇q φ + Jq · ∇p φ − Jq · ∇p φ = 0.
The second term of (64) gives
[−p · ∇p + ∆p , A]φ = [−p · ∇p , −Jp · ∇p ]φ + [∆p , −Jp · ∇p ]φ,
(65)
since Jq · ∇q commutes with p · ∇p + ∆p . Both terms in (65) are clearly zero due the antisymmetry of
J and the symmetry of the Hessian Dp2 φ.
The claim (d) follows from a short calculation similar to the proof of (a). To prove (e), note that
the fact that L and L− commute follows from (c), as
[L, L− ]φ = [L0 + µA, L0 − µA]φ = −2µ[L0 , A]φ = 0,
φ ∈ C ∞ ∩ L2 (π
b),
while the property hφ, P L0 P ψiL2 (πb) = hL− φ, ψiL2 (πb) follows from properties (a), (b) and (d). Indeed,
hφ, P LP ψiL2 (πb) = hφ, P (L0 + µA)P ψiL2 (πb) = hφ, (P L0 P + µA) ψiL2 (πb)
= h(L0 − µA)φ, ψiL2 (πb) = hL− φ, ψiL2 (πb) ,
as required. To prove (f) first notice that L, L0 and L− are of the form (60) and therefore leave the
spaces Hm invariant by Theorem 4. It follows immediately that also A leaves those spaces invariant.
The fact that P leaves the spaces Hm invariant follows directly by inspection of (59).
t
u
Now we proceed with the proof of Proposition 4:
22
Proof (of Proposition 4) Since the potential V is quadratic, Assumption 1 clearly holds and thus
Lemma 2 ensures that L and L− are invertible on L20 (π
b) with
ˆ ∞
L−1 =
e−tL dt,
(67)
0
and analogously for L−1
− . In particular, the asymptotic variance can be written as
σf2 = hf, (−L)−1 f iL2 (πb) .
Due to the respresentation (67) and Theorem 4, the inverses of L and L− leave the Hermite spaces
Hm invariant. We will prove the claim from Proposition 4 under the assumption that P f = f which
includes the case f = f (q). For the following calculations we will assume f ∈ Hm for fixed m ≥ 1.
Combining statement (f) with (a) and (e) of Lemma 5 (and noting that Hm ⊂ C ∞ (R2d ) ∩ L2 (π
b)) we
see that
P LP = L∗−
(68)
and
P L0 P = L∗0
(69)
when restricted to Hm . Therefore, the following calculations are justified:
1
hf, (−L)−1 f iL2 (πb) + hf, (−L∗ )−1 f iL2 (πb)
2
1
= hf, (−L)−1 f iL2 (πb) + hP f, (−L∗ )−1 P f iL2 (πb)
2
1
= hf, (−L)−1 f iL2 (πb) + hf, (−L− )−1 f iL2 (πb)
2
1
= hf, (−L)−1 + (−L− )−1 f iL2 (πb) ,
2
hf, (−L)−1 f iL2 (πb) =
where in the third line we have used the assumption P f = f and in the fourth line the properties
P 2 = I, P ∗ = P and equation (68). Since L and L− commute on Hm according to Lemma 5(e),(f) we
can write
(−L)−1 + (−L− )−1 = L− (−LL− )−1 + L(−LL− )−1 = −2L0 (LL− )−1
for the restrictions on Hm , using L + L− = 2L0 . We also have
LL− = (L0 + µA)(L0 − µA) = L20 + µ2 A∗ A,
since L0 and A commute. We thus arrive at the formula
σf2 = hf, −L0 (L20 + µ2 A∗ A)−1 f iL2 (πb) ,
f ∈ Hm .
(70)
Now since (L20 + µ2 A∗ A)−1 f = (LL− )−1 f ∈ D(L0 ) for all f ∈ L2 (π
b), it follows that the operator
−L0 (L20 + µ2 A∗ A)−1 is bounded. We L
can therefore extend formula (70) to the whole of L2 (π
b) by
continuity, using the fact that L20 (π
b) = m≥1 Hm .
t
u
Applying Proposition 4 we can analyse the behaviour of σf2 in the limit of large perturbation strength
µ → ∞. To this end, we introduce the orthogonal decomposition
L20 (π) = ker(Jq · ∇q ) ⊕ ker(Jq · ∇q )⊥ ,
(71)
where Jq · ∇q is understood as an unbounded operator acting on L20 (π), obtained as the smallest
closed extension of Jq · ∇q acting on Cc∞ (Rd ). In particular, ker(Jq · ∇q ) is a closed linear subspace of
L20 (π). Let Π denote the L20 (π)-orthogonal projection onto ker(Jq · ∇q ). We will write σf2 (µ) to stress
the dependence of the asymptotic variance on the perturbation strength. The following result shows
that for large perturbations, the limiting asymptotic variance is always smaller than the asymptotic
variance in the unperturbed case. Furthermore, the limit is given as the asymptotic variance of the
projected observable Πf for the unperturbed dynamics.
23
Theorem 5 Let f ∈ L20 (π), then
2
lim σf2 (µ) = σΠf
(0) ≤ σf2 (0).
µ→∞
Remark 11 Note that the fact that the limit exists and is finite is nontrivial. In particular, as Figures
1b and 1c demonstrate, it is often the case that limµ→∞ σf2 (µ) = ∞ if the condition µ = ν is not
satisfied.
Remark 12 The projection Π onto ker(Jq · ∇q ) can be understood in terms of Figure 2a. Indeed,
the eigenvalues on the real axis (highlighted by diamonds) are not affected by the perturbations. Let
us denote by Π̃ the projection onto the span of the eigenspaces of those eigenvalues. As µ → ∞,
the limiting asymptotic variance is given as the asymptotic variance associated to the unperturbed
dynamics of the projection Π̃f . If we denote by Π0 the projection of L2 (π
b) onto L20 (π), then we have
that ΠΠ0 = Π0 Π̃.
Proof (of Theorem 5) Note that L0 and A∗ A leave the Hermite spaces Hm invariant and their restrictions to those spaces commute (see Lemma 5, (b), (c) and (f)). Furthermore, as the Hermite
∗
spaces Hm are finite-dimensional, those operators have discrete
Lspectrum. As A A is nonnegative self2
adjoint, there exists an orthogonal decomposition
L0 (π) =
iW
L
Li into eigenspaces of the operator
−L0 (L20 + µ2 A∗ A)−1 , the decomposition
Wi being finer then
Hm in the sense that every Wi is
a subspace of some Hm . Moreover,
−L0 (L20 + µ2 A∗ A)−1 |Wi = −L0 (L20 + µ2 λi )−1 |Wi ,
where λi ≥ 0 is the eigenvalue of A∗ A associated to the subspace Wi . Consequently, formula (62) can
be written as
X
σf2 =
hfi , −L0 (L20 + µ2 λi )−1 fi iL2 (πb) ,
(72)
i
where f = i fi and fi ∈ Wi . Let us assume now without loss of generality that W0 = ker A∗ A, so in
particular λ0 = 0. Then clearly
P
lim σf2 = 2hf0 , −L0 (L20 )−1 f0 iL2 (πb) = 2hf0 , (−L0 )−1 f0 iL2 (πb) = σf20 (0).
µ→∞
2
Now note that W0 = ker A∗ A = ker A due to ker A∗ = (im A)⊥ . It remains to show that σΠf
(0) ≤
σf2 (0). To see this, we write
σf2 (0) = 2hf, (−L0 )−1 f iL2 (πb) = 2hΠf + (1 − Π)f, (−L0 )−1 Πf + (1 − Π)f iL2 (πb)
2
2
= σΠf
(0) + σ(1−Π)f
(0) + R,
where
R = 2hΠf, (−L0 )−1 (1 − Π)f iL2 (πb) + 2h(1 − Π)f, (−L0 )−1 Πf iL2 (πb) .
Note that since we only consider observables that do not depend on p, Πf ∈ ker(Jq·∇
L
Lq ) and (1−Π)f ∈
−1
W
.
Since
L
commutes
with
A,
it
follows
that
(−L
)
leaves
both
W
and
i
0
0
0
i≥1
i≥1 Wi invariant.
Therefore, as the latter spaces are orthogonal to each other, it follows that R = 0, from which the
result follows.
t
u
From Theorem 5 it follows that in the limit as µ → ∞, the asymptotic variance σf2 (µ) is not decreased
by the perturbation if f ∈ ker(Jq · ∇q ). In fact, this result also holds true non-asymptotically, i.e.
observables in ker(Jq · ∇q ) are not affected at all by the perturbation:
Lemma 6 Let f ∈ ker(Jq · ∇q ). Then
σf2 (µ) = σf2 (0)
for all µ ∈ R.
Proof From f ∈ ker(Jq · ∇q ) it follows immediately that f ∈ ker A∗ A. Then the claim follows from
the expression (72).
t
u
24
d
Example 3 Recall the case of observables of the form f (q) = q · Kq + l · q + C with K ∈ Rd×d
sym , l ∈ R
and C ∈ R from Section 4.1. If [J, K] = 0 and l ∈ ker J, then f ∈ ker(Jq · ∇q ) as
Jq · ∇q (q · Kq + l · q + C) = 2Jq · Kq + Jq · l = q · (KJ − JK)q − q · Jl = 0.
From the preceding lemma it follows that σf2 (µ) = σf2 (0) for all µ ∈ R, showing that the assumption
in Theorem 3 does not exclude nontrivial cases.
The following result shows that the dynamics (57) is particularly effective for antisymmetric observables
(at least in the limit of large perturbations):
Proposition 5 Let f ∈ L20 (π) satisfy f (−q) = −f (q) and assume that ker J = {0}. Furthermore,
assume that the eigenvalues of J are rationally independent, i.e.
σ(J) = {±iλ1 , ±iλ2 , . . . , ±iλd }
with λi ∈ R>0 and
P
i
(73)
ki λi 6= 0 for all (k1 , . . . , kd ) ∈ Zd \ (0, . . . , 0). Then limµ→∞ σf2 (µ) = 0.
Proof (of Proposition 5) The claim would immediately follow from f ∈ ker(Jq · ∇)⊥ according to
Theorem 5, but that does not seem to be so easy to prove directly. Instead, we again make use of the
Hermite polynomials.
Recall from the proof of Proposition 4 that L is invertible on L20 (π
b) and its inverse leaves the
Hermite spaces Hm invariant. Consequently, the asymptotic variance of an observable f ∈ L20 (π
b) can
be written as
σf2 = hf, (−L)−1 f iL2 (πb)
∞
X
=
hΠm f, (−L|Hm )−1 Πm f iL2 (πb) ,
(74a)
(74b)
m=1
where Πm : L20 (π
b) → Hm denotes the orthogonal projection onto Hm . From (59) it is clear that ga
is symmetric for |α| even and antisymmetric for |α| odd. Therefore, from f being antisymmetric it
follows that
M
f∈
Hm .
m≥1,m odd
In view of (54), (61) and (73) the spectrum of L|Hm can be written as
( 2d
)
X
σ(L|Hm ) = µ
αj βj + Cα,γ : |α| = m, βj ∈ σ(J)
j=1
(
=
iµ
d
X
)
(αj − αj+d )λj + Cα,γ : |α| = m
(75a)
j=1
with appropriate real constants Cα,γ ∈ R that depend on α and γ, but not on µ. For |α| =
odd, we have that
d
X
(αj − αj+d )λj 6= 0.
P2d
j=1
αj = m
(76)
j=1
Indeed, assume to the contrary that the above expression is zero. Then it follows that αj = αj+d for
all j = 1, . . . , d by rational independence of λ1 , . . . , λd . From (75a) and (76) it is clear that
µ→∞
sup {r > 0 : B(0, r) ∩ σ(L|Hm ) = ∅} −−−−→ ∞,
where B(0, r) denotes the ball of radius r centered at the origin in C. Consequently, the spectral radius
of (−L|Hm )−1 and hence (−L|Hm )−1 itself converge to zero as µ → ∞. The result then follows from
(74b).
t
u
25
Remark 13 The idea of the preceding proof can be explained using Figure 2a and Remark 12. Since
the real eigenvalues correspond to Hermite polynomials of even order, antisymmetric observables are
orthogonal to the associated subspaces. The rational independence condition on the eigenvalues of J
prevents cancellations that would lead to further eigenvalues on the real axis.
The following corollary gives a version of the converse of Proposition 5 and provides further intuition
into the mechanics of the variance reduction achieved by the perturbation.
Corollary 2 Let f ∈ L20 (π) and assume that limµ→∞ σf2 (µ) = 0. Then
ˆ
f dq = 0
B(0,r)
for all r ∈ (0, ∞), where B(0, r) denotes the ball centered at 0 with radius r.
2
Proof According to Theorem 5, limµ→∞ σf2 (µ) = 0 implies σΠf
(0) = 0. We can write
2
σΠf
(0) = hΠf, (−L0 )−1 Πf iL2 (πb)
1
= hΠf, (−L0 )−1 + (−L∗0 )−1 Πf iL2 (πb)
2
and recall from the proof of Proposition 4 that (−L0 )−1 and (−L∗0 )−1 leave the Hermite spaces Hm
invariant. Therefore
ker (−L0 )−1 + (−L∗0 )−1 = 0
(78)
2
in L20 (π
b), and in particular σΠf
(0) = 0 implies Πf = 0, which in turn shows that f ∈ ker(Jq · ∇)⊥ .
⊥
Using ker(Jq · ∇) = im(Jq · ∇), it follows that there exists a sequence (φn )n ∈ Cc∞ (Rd ) such that
Jq · ∇φn → f in L2 (π). Taking a subsequence if necessary, we can assume that the convergence is
pointwise π-almost everywhere and that the sequence is pointwise bounded by a function in L1 (π).
Since J is antisymmetric, we have that Jq · ∇φn = ∇ · (φn Jq). Now Gauss’s theorem yields
ˆ
ˆ
ˆ
f dq =
∇ · (φJq)dq =
φJq · dn,
B(0,r)
B(0,r)
∂B(0,r)
where n denotes the outward normal to the sphere ∂B(0, r). This quantity is zero due to the orthogonality of Jq and n, and so the result follows from Lebesgue’s dominated convergence theorem.
t
u
4.4 Optimal Choices of J for Quadratic Observables
d×d
Assume f ∈ L20 (π) is given by f (q) = q · Kq + l · q − Tr K, with K ∈ Rsym
and l ∈ Rd (note that
the constant term is chosen such that π(f ) = 0). Our objective is to choose J in such a way that
limµ→∞ σf2 (µ) becomes as small as possible. To stress the dependence on the choice of J, we introduce
⊥
the notation σf2 (µ, J). Also, we denote the orthogonal projection onto (ker J)⊥ by Πker
J.
⊥
Lemma 7 (Zero variance limit for linear observables). Assume K = 0 and Πker
J l = 0. Then
lim σf2 (µ, J) = 0.
µ→∞
Proof According to Proposition 5, we have to show that Πf = 0, where Π is the L2 (π)-orthogonal
projection onto ker(Jq · ∇). Let us thus prove that
f ∈ ker(Jq · ∇)⊥ = im(Jq · ∇)∗ = im(Jq · ∇),
⊥
where the second identity uses the fact that (Jq·∇)∗ = −Jq·∇. Indeed, since Πker
J = 0, by Fredholm’s
alternative there exists u ∈ Rd such that Ju = l. Now define φ ∈ L20 (π) by φ(q) = −u · q, leading to
f = Jq · ∇φ,
so the result follows.
t
u
26
Lemma 8 (Zero variance limit for purely quadratic observables.) Let l = 0 and consider the decomposition K = K0 + K1 into the traceless part K0 = K − TrdK · I and the trace-part K1 = TrdK · I. For
the corresponding decomposition of the observable
f (q) = f0 (q) + f1 (q) = q · K0 q + q · K1 q − Tr K
the following holds:
(a) There exists an antisymmetric matrix J such that limµ→∞ σf20 (µ, J) = 0, and there is an algorithmic
way (see Algorithm 1) to compute an appropriate J in terms of K.
(b) The trace-part is not effected by the perturbation, i.e. σf21 (µ, J) = σf21 (0) for all µ ∈ R.
Proof To prove the first claim, according to Theorem 5 it is sufficient to show that f0 ∈ ker(Jq · ∇)⊥ =
im(Jq · ∇). Let us consider the function φ(q) = q · Aq, with A ∈ Rd×d
sym . It holds that
Jq · ∇φ = q · (J T Aq) = q · [A, J]q.
The task of finding an antisymmetric matrix J such that
lim σf20 (µ, J) = 0
µ→∞
(79)
can therefore be accomplished by constructing an antisymmetric matrix J such that there exists a
symmetric matrix A with the property K0 = [A, J]. Given any traceless matrix K0 there exists an
orthogonal matrix U ∈ O(Rd ) such that U K0 U T has zero entries on the diagonal, and that U can be
obtained in an algorithmic manner (see for example [Kaz88] or [HJ13, Chapter 2, Section 2, Problem
3]; for the reader’s convenience we have summarised the algorithm in Appendix D.) Assume thus that
such a matrix U ∈ O(Rd ) has been found and choose real numbers a1 , . . . , ad ∈ R such that ai 6= aj if
i 6= j. We now set
Ā = diag(a1 , . . . , an ),
(80)
and
(
J¯ij =
(U K0 U T )ij
ai −aj
0
if i 6= j,
if i = j.
(81)
¯ =
Observe that since U K0 U T is symmetric, J¯ is antisymmetric. A short calculation shows that [Ā, J]
¯ to obtain [A, J] = K0 . Therefore, the J
U K0 U T . We can thus define A = U T ĀU and J = U T JU
constructed in this way indeed satisfies (79). For the second claim, note that f1 ∈ ker(Jq · ∇), since
Tr K
Tr K
Jq · ∇ q ·
q =2
q · Jq = 0
(82)
d
d
because of the antisymmetry of J. The result then follows from Lemma 6.
t
u
We would like to stress that the perturbation J constructed in the previous lemma is far from unique
due to the freedom of choice of U and a1 , . . . , ad ∈ R in its proof. However, it is asymptotically optimal:
Corollary 3 In the setting of Lemma 8 the following holds:
min
lim σf2 (µ, J) = σf21 (0).
J T =−J
µ→∞
Proof The claim follows immediately since f1 ∈ ker(Jq · ∇) for arbitrary antisymmetric J as shown in
(82), and therefore the contribution of the trace part f1 to the asymptotic variance cannot be reduced
by any choice of J according to Lemma 6.
As the proof of Lemma 8 is constructive, we obtain the following algorithm for determining optimal
perturbations for quadratic observables:
Algorithm 1 Given K ∈ Rd×d
sym , determine an optimal antisymmetric perturbation J as follows:
27
1. Set K0 = K − TrdK · I.
2. Find U ∈ O(Rd ) such that U K0 U T has zero entries on the diagonal (see Appendix D).
3. Choose ai ∈ R, i = 1, . . . d such that ai 6= aj for i 6= j and set
(U K0 U T )ij
J¯ij =
ai − aj
for i 6= j and J¯ii = 0 otherwise.
¯ .
4. Set J = U T JU
Remark 14 In [DLP16], the authors consider the task of finding optimal perturbations J for the nonreversible overdamped Langevin dynamics given in (19). In the Gaussian case this optimization problem
turns out be equivalent to the one considered in this section. Indeed, equation (39) of [DLP16] can be
rephrased as
f ∈ ker(Jq · ∇)⊥ .
Therefore, Algorithm 1 and its generalization Algorithm 2 (described in Section 4.5) can be used
without modifications to find optimal perturbations of overdamped Langevin dynamics.
4.5 Gaussians with Arbitrary Covariance and Preconditioning
In this section we extend the results of the preceding sections to the case when the target measure π
d×d
symmetric and
is given by a Gaussian with arbitrary covariance, i.e. V (q) = 21 q · Sq with S ∈ Rsym
positive definite. The dynamics (9) then takes the form
dqt = M −1 pt dt − µJ1 Sqt dt,
dpt = −Sqt dt − νJ2 M −1 pt dt − Γ M −1 pt dt +
√
2Γ dWt .
(83)
The key observation is now that the choices M = S and Γ = γS together with the transformation
qe = S 1/2 q and pe = S −1/2 p lead to the dynamics
dqet = pet dt − µS 1/2 J1 S 1/2 qet dt,
dpet = −qet dt − µS −1/2 J2 S −1/2 pet dt − γ pet dt +
p
2γdWt ,
(84)
which is of the form (41) if J1 and J2 obey the condition SJ1 S = J2 (note that both S 1/2 J1 S 1/2 and
S −1/2 J2 S −1/2 are of course antisymmetric). Clearly the dynamics (84) is ergodic with respect to a
Gaussian measure with unit covariance, in the following denoted by π
e. The connection between the
asymptotic variances associated to (83) and (84) is as follows:
For an observable f ∈ L20 (π) we can write
√
T
1
T
ˆ
T
√
f (qs )ds − π(f )
=
T
0
1
T
ˆ
0
T
e
e
f (qes )ds − π
e (f ) ,
where fe(q) = f (S −1/2 q). Therefore, the asymptotic variances satisfy
σf2 = σ
ef2e,
(85)
where σ
ef2e denotes the asymptotic variance of the process (qet )t≥0 . Because of this, the results from the
previous sections generalise to (83), subject to the condition that the choices M = S, Γ = γS and
SJ1 S = J2 are made. We formulate our results in this general setting as corollaries:
28
Corollary 4 Consider the dynamics
dqt = M −1 pt dt − µJ1 ∇V (qt )dt,
√
(86)
dpt = −∇V (qt )dt − µJ2 M −1 pt dt − Γ M −1 pt dt + 2Γ dWt ,
√
with V (q) = 21 q · Sq. Assume that M = S, Γ = γS with γ > 2 and SJ1 S = J2 . Let f ∈ L2 (π) be an
observable of the form
f (q) = q · Kq + l · q + C
(87)
d
with K ∈ Rd×d
/ ker J is
sym , l ∈ R and C ∈ R. If at least one of the conditions KJ1 S 6= SJ1 K and l ∈
satisfied, then the asymptotic variance is at a local maximum for the unperturbed sampler, i.e.
∂µ σf2
µ=0
=0
and
∂µ2 σf2
µ=0
< 0.
Proof Note that
e +e
fe(q) = f (S −1/2 q) = q · S −1/2 KS −1/2 q + S −1/2 l · q + C = q · Kq
l·q+C
e = S −1/2 KS −1/2 and
is again of the form (87) (where in the last equality, K
been defined). From (84), (85) and Theorem 3 the claim follows if at least one
e S 1/2 J1 S 1/2 ] 6= 0 and e
[K,
l ∈
/ ker(S 1/2 J1 S 1/2 ) is satisfied. The first of those can
equivalent to
S −1/2 (KJS − SJK)S −1/2 6= 0,
e
l = S −1/2 l have
of the conditions
easily seen to be
which is equivalent to KJ1 S 6= SJ1 K since S is nondegenerate. The second condition is equivalent to
S 1/2 J1 l 6= 0,
which is equivalent to J1 l 6= 0, again by nondegeneracy of S.
t
u
Corollary 5 Assume the setting from the previous corollary and denote by Π the orthogonal projection
onto ker(J1 Sq · ∇). For f ∈ L2 (π) it holds that
2
lim σf2 (µ) = σΠf
(0) ≤ σf2 (0).
µ→∞
Proof Theorem 5 implies
2
lim σ
ef2e(µ) = σ
eΠ
ef2e(0)
e fe(0) ≤ σ
µ→∞
e denotes
for the transformed system (84). Here fe(q) = f (S −1/2 q) is the transformed observable and Π
2
1/2
1/2
L (π)-orthogonal projection onto ker(S J1 S q · ∇). According to (85), it is sufficient to show
e fe. This however follows directly from the fact that the linear transformation
that (Πf ) ◦ S −1/2 = Π
1/2
φ 7→ φ ◦ S
maps ker(S 1/2 J1 S 1/2 q · ∇) bijectively onto ker(J1 Sq · ∇).
t
u
Let us also reformulate Algorithm 1 for the case of a Gaussian with arbitrary covariance.
1
Algorithm 2 Given K, S ∈ Rd×d
sym with f (q) = q · Kq and V (q) = 2 q · Sq (assuming S is nondegenerate), determine optimal perturbations J1 and J2 as follows:
e = S −1/2 KS −1/2 and K
e0 = K
e − Tr Ke · I.
1. Set K
d
e 0 U T has zero entries on the diagonal (see Appendix D).
2. Find U ∈ O(Rd ) such that U K
3. Choose ai ∈ R, i = 1, . . . , d such that ai 6= aj for i 6= j and set
e 0 U T )ij
(U K
J¯ij =
.
ai − aj
¯ .
4. Set Je = U T JU
−1/2 e −1/2
5. Put J1 = S
JS
and J2 = S 1/2 JS 1/2 .
Finally, we obtain the following optimality result from Lemma 7 and Corollary 3.
29
⊥
Corollary 6 Let f (q) = q · Kq + l · q − Tr K and assume that Πker
J l = 0. Then
lim σf2 (µ, J1 , J2 ) = σf21 (0),
min
J1T =−J1 , J2 =SJ1 S
where f1 (q) = q ·K1 q, K1 =
2.
Tr(S −1 K)
S.
d
µ→∞
Optimal choices for J1 and J2 can be obtained using Algorithm
Remark 15 Since in Section 4.1 we analysed the case where J1 and J2 are proportional, we are not able
to drop the restriction J2 = SJ1 S from the above optimality result. Analysis of completely arbitrary
perturbations will be the subject of future work.
Remark 16 The choices M = S and Γ = γS have been introduced to make the perturbations considered in this article lead to samplers that perform well in terms of reducing the asymptotic variance.
However, adjusting the mass and friction matrices according to the target covariance in this way (i.e.
M = S and Γ = γS) is a popular way of preconditioning the dynamics, see for instance [GC11]
and, in particular mass-tensor molecular dynamics [Ben75]. Here we will present an argument why
such a preconditioning is indeed beneficial in terms of the convergence rate of the dynamics. Let us
first assume that S is diagonal, i.e. S = diag(s(1) , . . . , s(d) ) and that M = diag(m(d) , . . . , m(d) ) and
Γ = diag(γ (d) , . . . , γ (d) ) are chosen diagonally as well. Then (83) decouples into one-dimensional SDEs
of the following form:
(i)
dqt =
1 (i)
p dt,
m(i) t
(i)
(i)
dpt = −s(i) qt dt −
p
γ (i) (i)
p
dt
+
2γ (i) dWt ,
t
m(i)
i = 1, . . . , d.
(88)
Let us write those Ornstein-Uhlenbeck processes as
(i)
dXt
(i)
= −B (i) Xt dt +
with
B
(i)
=
0 − m1(i)
s(i)
p
(i)
2Q(i) dWt
!
γ (i)
m(i)
(i)
and Q
=
0 0
0 γ (i)
(89)
.
As in Section 4.2, the rate of the exponential decay of (89) is equal to min Re σ(B (i) ). A short calculation
shows that the eigenvalues of B (i) are given by
s
2
(i)
γ
γ (i)
s(i)
(i)
λ1,2 =
±
− (i) .
(i)
(i)
2m
2m
m
Therefore, the rate of exponential decay is maximal when
(i) 2
s(i)
γ
− (i) = 0,
(i)
2m
m
(90)
in which case it is given by
r
s(i)
.
m(i)
Naturally, it is reasonable to choose m(i) in such a way that the exponential rate (λ(i) )∗ is the same
for all i, leading to the restriction M = cS with c > 0. Choosing c small will result in fast convergence
to equilibrium, but also make the dynamics (88) quite stiff, requiring a very small timestep ∆t in
a discretisation scheme. The choice of c will therefore need to strike a balance between those two
competing effects. The constraint (90) then implies Γ = 2cS. By a coordinate transformation, the
preceding argument also applies if S, M and Γ are diagonal in the same basis, and of course M and
Γ can always be chosen that way. Numerical experiments show that it is possible to increase the rate
of convergence to equilibrium even further by choosing M and Γ nondiagonally with respect to S
(although only by a small margin). A clearer understanding of this is a topic of further investigation.
(i) ∗
(λ ) =
30
5 Numerical Experiments: Diffusion Bridge Sampling
5.1 Numerical Scheme
In this section we introduce a splitting scheme for simulating the perturbed underdamped Langevin
dynamics given by equation (9). In the unpertubed case, i.e. when J1 = J2 = 0, the right-hand side
can be decomposed into parts A, B and C according to
−1
0√
qt
0
M pt
,
d
=
dt +
dt +
pt
−∇V (qt )
0
−Γ M −1 + 2Γ dWt
|
{z
} |
{z
} |
{z
}
A
B
O
i.e. O refers to the Ornstein-Uhlenbeck part of the dynamics, whereas A and B stand for the momentum and position updates, respectively.
One particular splitting scheme which has proven to be efficient is the BAOAB scheme, (see [LM15] and
references therein). The string of letters refers to the order in which the different parts are integrated,
namely
1
pn+1/2 = pn − ∆t∇V (qn ),
2
1
qn+1/2 = qn + ∆t · M −1 pn+1/2 ,
2
p
p̂ = exp(−∆tΓ M −1 )pn+1/2 + I − e−2Γ ∆t N (0, I),
1
qn+1 = q n+1/2 + ∆t · M −1 p̂,
2
1
pn+1 = p̂ − ∆t · ∇V (qn+1 ).
2
(91a)
(91b)
(91c)
(91d)
(91e)
We note that many different discretisation schemes such as ABOBA, OABAO, etc. are viable, but
that analytical and numerical evidence has shown that the BAOAB-ordering has particularly good
properties to compute long-time ergodic averages with respect to q-dependent observables. Motivated
by this, we introduce the following perturbed scheme, introducing additional Runge-Kutta integration
steps between the A, B and O parts:
1
pn+1/2 = pn − ∆t∇V (qn ),
2
1
qn+1/2 = qn + ∆t · M −1 pn+1/2 ,
2
1
0
qn+1/2
= RK4 ( ∆t, qn+1/2 ),
2
p
p̂ = exp(−∆t(Γ M −1 + νJ2 M −1 ))pn+1/2 + I − e−2Γ ∆t N (0, 1),
1
0
00
qn+1/2
= RK4 ( ∆t, qn+1/2
),
2
1
00
qn+1 = qn+1/2
+ ∆t · M −1 p̂,
2
1
pn+1 = p̂ − ∆t · ∇V (qn+1 ),
2
(92a)
(92b)
(92c)
(92d)
(92e)
(92f)
(92g)
where RK4 (∆t, q0 ) refers to fourth order Runge-Kutta integration of the ODE
q̇ = −J1 ∇V (q),
q(0) = q0
(93)
up until time ∆t. We remark that the J2 -perturbation is linear and can therefore be included in the Opart without much computational overhead. Clearly, other discretisation schemes are possible as well,
for instance one could use a symplectic integrator for the ODE (93), noting that it is of Hamiltonian
31
type. However, since V as the Hamiltonian for (93) is not separable in general, such a symplectic
integrator would have to be implcit. Moreover, (92c) and (92e) could be merged since (92e) commutes
with (92d). In this paper, we content ourselves with the above scheme for our numerical experiments.
Remark 17 The aformentioned schemes lead to an error in the approximation for π(f ), since the
invariant measure π is not preserved exactly by the numerical scheme. In practice, the BAOABscheme can therefore be accompanied by an accept-reject Metropolis step as in [MWL16], leading to
an unbiased estimate of π(f ), albeit with an inflated variance. In this case, after every rejection the
momentum variable has to be flipped (p 7→ −p) in order to keep the correct invariant measure. We note
here that our perturbed scheme can be ’Metropolized’ in a similar way by ’flipping the matrices J1
and J2 after every rejection (J1 7→ −J1 and J2 7→ −J2 ) and using an appropriate (volume-preserving
and time-reversible) integrator for the dynamics given by (93). Implementations of this idea are the
subject of ongoing work.
5.2 Diffusion Bridge Sampling
To numerically test our analytical results, we will apply the dynamics (9) to sample a measure on path
space associated to a diffusion bridge. Specifically, consider the SDE
p
dXs = −∇U (Xs )ds + 2β −1 dWs ,
with Xs ∈ Rn , β > 0 and the potential U : Rn → R obeying adequate growth and smoothness conditions (see [HSV07], Section 5 for precise statements). The law of the solution to this SDE conditioned
on the events X(0) = x− and X(s+ ) = x+ is a probability measure π on L2 ([0, s+ ], Rn ) which poses
a challenging and important sampling problem, especially if U is multimodal. This setting has been
used as a test case for sampling probability measures in high dimensions (see for example [BPSSS11]
and [OPPS16]). For a more detailed introduction (including applications) see [BS09] and for a rigorous
theoretical treatment the papers [HSVW05, HSV07, HSV09, BS09] .
In the case U ≡ 0, it can be shown that the law of the conditioned process is given by a Gaussian
measure π0 with mean zero and precision operator S = − β2 ∆ on the Sobolev space H 1 ([0, s+ ], Rd )
equipped with appropriate boundary conditions. The general case can then be understood as a perturbation thereof: The measure π is absolutely continuous with respect to π0 with Radon-Nikodym
derivative
dπ
∝ exp − Ψ ,
(94)
dπ0
where
ˆ
β s+
G(x(s), β)ds
Ψ (x) =
2 0
and
1
1
|∇U (x)|2 − ∆U (x).
2
β
We will make the choice x− = x+ = 0, which is possible without loss of generality as explained in
[BRSV08, Remark 3.1], leading to Dirichlet boundary conditions on [0, s+ ] for the precision operator
S. Furthermore, we choose s+ = 1 and discretise the ensuing s-interval [0, 1] according to
G(x, β) =
[0, 1] = [0, s1 ) ∪ [s1 , s2 ) ∪ . . . ∪ [sn−1 , sn ) ∪ [sn , 1]
1
in an equidistant way with stespize sj+1 − sj ≡ δ = d+1
. Functions on this grid are determined by the
values x(s1 ) = x1 , . . . , x(sn ) = xn , recalling that x(0) = x(1) = 0 by the Dirichlet boundary conditions.
We discretise the functional Ψ as
d
Ψ̃ (x1 , . . . , xn ) =
β X
δ
G(xi , β)
2
i=1
d
β X
1
= δ
(U 0 (xi )2 − U 00 (xi ) ,
2
β
i=1
32
such that its gradient is given by
(∇Ψ̃ )i =
1
β
δ 2U 0 (xi )U 00 (xi ) − U 000 (xi ) ,
2
β
i = 1, . . . , d.
The discretised version A of the Dirichlet-Laplacian ∆ on [0, 1] is given by
−2 1
1 −2
...
A = δ −2
.
1
1 −2
Following (94), the discretised target measure π
b has the form
π
b=
1 −V
e dx,
Z
with
βδ
x · Ax, x ∈ Rd .
4
In the following we will consider the case n = 1 with potential U : R → R given by U (x) = 21 (x2 − 1)2
and set β = 1. To test our algorithm we adjust the parameters M , Γ , J1 and J2 according to the
recommended choice in the Gaussian case,
V (x) = Ψ̃ (x) −
M = S,
Γ = γS,
SJ1 S = J2 ,
µ = ν,
(95)
where we take S = β2 δ · A as the precision operator of the Gaussian target. We will consider the
linear observable f1 (x) = l · x with l = (1, . . . , 1) and the quadratic observable f2 (x) = |x|2 . In a
first experiment we adjust the perturbation J1 (and via (95) also J2 ) to the observable f2 according
to Algorithm 2. The dynamics (9) is integrated using the splitting scheme introduced in Section 5.1
with a stepsize of ∆t = 10−4 over the time interval [0, T ] with T = 102 . Furthermore, we choose
initial conditions q0 = (1, . . . , 1), p0 = (0, . . . , 0) and introduce a burn-in time T0 = 1, i.e. we take the
estimator to be
ˆ T
1
f (qt )dt.
π̂(f ) ≈
T − T0 T0
We compute the variance of the above estimator from N = 500 realisations and compare the results
for different choices of the friction coefficient γ and of the perturbation strength µ.
The numerical experiments show that the perturbed dynamics generally outperform the unperturbed dynamics independently of the choice of µ and γ, both for linear and quadratic observables.
One notable exception is the behaviour of the linear observable for small friction γ = 10−3 (see Figure
3a), where the asymptotic variance initially increases for small perturbation strengths µ. However,
this does not contradict our analytical results, since the small
√ perturbation results from Section 4.1
generally require γ √to be sufficiently big (for example γ ≥ 2 in Theorem 3). We remark here that
the condition γ ≥ 2, while necessary for the theoretical results from Section 4.1, is not a very advisable choice in practice (at least in this experiment), since Figures 3b and 4b clearly indicate that
the optimal friction is around γ ≈ 10−1 . Interestingly, the problem of choosing a suitable value for the
friction coefficient coefficient γ becomes mitigated by the introduction of the perturbation: While the
performance of the unperturbed sampler depends quite sensitively on γ, the asymptotic variance of
the perturbed dynamics is a lot more stable with respect to variations of γ.
In the regime of growing values of µ, the experiments confirm the results from Section 4.3, i.e. the
asymptotic variance approaches a limit that is smaller than the asymptotic variance of the unperturbed
dynamics.
As a final remark we report our finding that the performance of the sampler for the linear observable is qualitatively independent of the coice of J1 (as long as J2 is adjusted according to (95)). This
result is in alignment with Propostion 5 which predicts good properties of the sampler for antisymmetric observables. In contrast to this, a judicious choice of J1 is critical for quadratic observables. In
particular, applying Algorithm 2 significantly improves the performance of the perturbed sampler in
comparison to choosing J1 arbitrarily.
33
=10-3
40
30
=10-1
=10
35
25
standard deviation
standard deviation
Linear Observable
35
Linear Observable
45
30
25
20
15
=0
=40
20
15
10
10
5
5
0
10 -3
0
0
5
10
15
20
25
30
35
40
10 -2
10 -1
10 0
10 1
10 2
friction
perturbation strength
(a)
(b)
.
Fig. 3: Standard deviation of π̂(f ) for a linear observable as a function of friction γ and perturbation
strength µ
Quadratic Observable
12
=10
=10-1
=10
10
10
8
standard deviation
standard deviation
Quadratic Observable
12
-3
6
4
=0
=40
8
6
4
2
2
0
0
5
10
15
20
25
perturbation strength
(a)
30
35
40
0
10 -3
10 -2
10 -1
10 0
10 1
10 2
friction
(b)
.
Fig. 4: Standard deviation of π̂(f ) for a quadratic observable as a function of friction γ and perturbation
strength µ
6 Outlook and Future Work
A new family of Langevin samplers was introduced in this paper. These new SDE samplers consist
of perturbations of the underdamped Langevin dynamics (that is known to be ergodic with respect
to the canonical measure), where auxiliary drift terms in the equations for both the position and the
momentum are added, in a way that the perturbed family of dynamics is ergodic with respect to
the same (canonical) distribution. These new Langevin samplers were studied in detail for Gaussian
target distributions where it was shown, using tools from spectral theory for differential operators,
that an appropriate choice of the perturbations in the equations for the position and momentum
can improve the performance of the Langvin sampler, at least in terms of reducing the asymptotic
variance. The performance of the perturbed Langevin sampler to non-Gaussian target densities was
tested numerically on the problem of diffusion bridge sampling.
The work presented in this paper can be improved and extended in several directions. First, a
rigorous analysis of the new family of Langevin samplers for non-Gaussian target densities is needed.
34
The analytical tools developed in [DLP16] can be used as a starting point. Furthermore, the study of
the actual computational cost and its minimization by an appropriate choice of the numerical scheme
and of the perturbations in position and momentum would be of interest to practitioners. In addition,
the analysis of our proposed samplers can be facilitated by using tools from symplectic and differential
geometry. Finally, combining the new Langevin samplers with existing variance reduction techniques
such as zero variance MCMC, preconditioning/Riemannian manifold MCMC can lead to sampling
schemes that can be of interest to practitioners, in particular in molecular dynamics simulations. All
these topics are currently under investigation.
Acknowledgments
AD was supported by the EPSRC under grant No. EP/J009636/1. NN is supported by EPSRC
through a Roth Departmental Scholarship. GP is partially supported by the EPSRC under grants
No. EP/J009636/1, EP/L024926/1, EP/L020564/1 and EP/L025159/1. Part of the work reported in
this paper was done while NN and GP were visiting the Institut Henri Poincaré during the Trimester
Program ”Stochastic Dynamics Out of Equilibrium”. The hospitality of the Institute and of the organizers of the program is greatly acknowledged.
A Estimates for the Bias and Variance
Proof (of Lemma 1) Suppose that (Pt )t≥0 satisfies (27). Let π0 be an initial distribution of (Xt )t≥0 such that
0
∈ L2 (π). Slightly abusing notation, we denote by π0 Pt the law of Xt given X0 ∼ π. Then
π0 π and h = dπ
dπ
kπ0 Pt − πkT V = kPt∗ h − 1kL1 (π) ≤ kPt∗ kL2 (π)→L2 (π) kh − 1kL2 (π) ≤ Ce−λt kh − 1kL2 (π) ,
where Pt∗ denotes the L2 (π)-adjoint of Pt . Since f is assumed to be bounded, we immediately obtain
dπ0 1/2
|E[f (Xt )|X0 ∼ π0 ] − π(f )| ≤ C kf kL∞ e−λt Varπ
,
dπ
and so, for X0 ∼ π0 ,
|πT (f ) − π(f )| ≤
C
1 − e−λt kf kL∞
λT
Varπ
dπ0
dπ
1/2
,
u
t
as required.
Proof (of Lemma 2) Given f ∈ L2 (π), for fixed T > 0,
ˆ
T
(π(f ) − Pt f (x)) dt.
χT (x) :=
(96)
0
Then we have that χT ∈ D(L) and LχT = f − PT f , moreover
ˆ
kχT − χT 0 kL2 (π) =
T0
Pt (f ) − π(f ) dt
T
ˆ
≤ C kf kL2 (π)
L2 (π)
T0
so that {χT }T ≥0 is a Cauchy sequence in L2 (π) converging to χ =
(LχT , χT ) → (f − π(f ), χ),
e−λt dt,
T
´∞
0
(π(f ) − Pt f ) dt. Since L is closed and
T → ∞,
in L2 (π), it follows that χ ∈ D(L) and Lχ = f − π(f ). Moreover,
ˆ ∞
kχkL2 (π) ≤
kPt (f ) − π(f )kL2 (π) dt ≤ Kλ kf − π(f )kL2 (π) ,
0
35
´∞
where Kλ = C 0 e−λt dt. Since we assume that f is smooth, the coefficients are smooth and L is hypoelliptic,
then Lχ = f − π(f ) implies that χ ∈ C ∞ (Rd ), and thus we can apply Itô’s formula to χ(Xt ) to obtain:
ˆ
ˆ
1 T
1 T
1
[f (Xt ) − π(f )] dt = [χ(X0 ) − χ(XT )] +
∇χ(Xt )σ(Xt ) dWt .
T 0
T
T 0
One can check that the conditions of [EK86, Theorem 7.1.4] hold. In particular, the following central limit
theorem follows
ˆ T
1
d
√
∇χ(Xt )σ(Xt ) dWt −
→ N (0, 2σf2 ),
as T → ∞.
T 0
By Theorem 1, the generator L has the form
L = π −1 ∇ · (πΣ∇·) + γ · ∇,
where ∇ · (πγ) = 0. It follows that
σf2 = hΣ∇χ, ∇χiL2 (π) = − hL χ, χiL2 (π) = hχ, f iL2 (π) < ∞.
(97)
First suppose that X0 ∼ π. Then (χ(Xt ))t≥0 is a stationary process, and so
1
√ (χ(X0 ) − χ(XT )) → 0,
T
a.s as T → ∞.
From which (29) follows. More generally, suppose that X0 ∼ π0 , where π0 (x) = h(x)π(x) for h ∈ L2 (π). If
f ∈ L∞ (π), then by (1),
ˆ ∞
|χ(x)| ≤
|π(f ) − Pt f (x)| dt
ˆ0∞
≤
kf kL∞ kπ − π0 Pt kT V dt
0
C
≤ kf kL∞
λ
so that χ ∈ L∞ (π). Therefore
√1
T
dπ0 1/2
Varπ
,
dπ
p
(χ(X0 ) − χ(XT )) −
→ 0 as T → ∞, and so (29) holds in this case, similarly.
B Proofs of Section 3
Proof of Lemma 3 We first note that L in (37) can be written in the “sum of squares” form:
L = A0 +
where
d
1X 2
Ak ,
2 k=1
A0 = M −1 p · ∇q − ∇q V · ∇p − µJ1 ∇q V · ∇q − νJ2 M −1 p · ∇p − Γ M −1 p · ∇p
and
Ak = ek · Γ 1/2 ∇p ,
k = 1, . . . , d.
Here {ek }k=1,...,d denotes the standard Euclidean basis and Γ 1/2 is the unique positive definite square root of
the matrix Γ . The relevant commutators turn out to be
[A0 , Ak ] = ek · Γ 1/2 M −1 (Γ ∇p − ∇q − νJ2 ∇p ),
k = 1, . . . , k.
Because Γ has full rank on Rd , it follows that
span{Ak : k = 1, . . . d} = span{∂pk : k = 1, . . . , d}.
Since
and span({Γ
ek · Γ 1/2 M −1 (Γ ∇p − νJ2 ∇p ) ∈ span{Aj : j = 1, . . . d},
1/2
M
−1
k = 1, . . . , d,
∇q : k = 1, . . . , d}) = span{∂qk : k = 1, . . . , d}, it follows that
span({Ak : k = 0, 1, . . . , d} ∪ {[A0 , Ak ] : k = 1, . . . , d}) = R,
so the assumptions of Hörmander’s theorem hold.
u
t
36
B.1 The overdamped limit
The following is a technical lemma required for the proof of Proposition 1:
Lemma 9 Assume the conditions from Proposition 1. Then for every T > 0 there exists C > 0 such that
!
sup |pt |2
E
≤ C.
0≤t≤T
Proof Using variation of constants, we can write the second line of (39) as
pt = e
− t2 (νJ2 +Γ )M −1
p0 −
1
ˆ
t
−
e
(t−s)
(νJ2 +Γ )M −1
2
∇q V (qs )ds +
0
1√
2Γ
ˆ
t
−
e
(t−s)
(νJ2 +Γ )M −1
2
dWs .
0
We then compute
− t2 (νJ2 +Γ )M −1
E sup |pt |2 = sup
0≤t≤T
e
0≤t≤T
√
ˆ
t
2
p0
+
1
E sup
2 0≤t≤T
ˆ
t
−
e
(t−s)
(νJ2 +Γ )M −1
2
2
∇q V (qs )ds
0
2
1
dWs
2Γ
E sup
e
2 0≤t≤T
0
ˆ t
(t−s)
1
− t (νJ +Γ )M −1
(νJ2 +Γ )M −1
−
− E sup e 2 2
p0 ·
e 2
∇q V (qs )ds
0≤t≤T
0
√ ˆ t − (t−s) (νJ +Γ )M −1
−1
t
1
1
(νJ +Γ )M
−
2
p0 ·
dWs
+ E sup e 2 2
2Γ
e 2
0≤t≤T
0
ˆ t
√ ˆ t − (t−s) (νJ +Γ )M −1
(t−s)
−1
1
(νJ
+Γ
)M
−
2
2
− 2 E sup
e 2
∇q V (qs )ds · 2Γ
dWs .
e 2
0≤t≤T
0
0
(t−s)
(νJ2 +Γ )M −1
−
2
+
(98)
Clearly, the first term on the right hand side of (98) is bounded. For the second term, observe that
1
E sup
2 0≤t≤T
ˆ
t
−
e
(t−s)
(νJ2 +Γ )M −1
2
2
∇q V (qs )ds
≤
0
1
sup
2 0≤t≤T
ˆ
t
−
e
(t−s)
(νJ2 +Γ )M −1
2
2
ds
since V ∈ C 1 (Td ) and therefore ∇q V is bounded. By the basic matrix exponential estimate ke−t(νJ2 +Γ )M
Ce−ωt for suitable C and ω, we see that (99) can further be bounded by
1
C sup
2 0≤t≤T
ˆ
t
−2ω
e
0
(t−s)
2
ds =
(99)
0
−1
k≤
C
−2ω T2
,
1−e
2ω
so this term is bounded as well. The third term is bounded by the Burkholder–Davis–Gundy inequality and a
similar argument to the one used for the second term applies. The cross terms can be bounded by the previous
ones, using the Cauchy-Schwarz inequality and the elementary fact that sup(ab) ≤ sup a · sup b for a, b > 0, so
the result follows.
u
t
Proof (of Proposition 1) Equations (39) can be written in integral form as
1
(νJ2 + Γ )qt = (νJ2 + Γ )q0 +
and
ˆ
−
0
t
∇V (qs )ds −
1
ˆ
t
ˆ
t
ˆ
t
(νJ2 + Γ )M −1 ps ds − µ
0
(νJ2 + Γ )M −1 ps ds +
(νJ2 + Γ )J1 ∇q V (qs )ds
0
√
2Γ W (t) = (pt − p0 ),
(100)
0
where the first line has been multiplied by the matrix νJ2 + Γ . Combining both equations yields
ˆ
qt = q0 −
0
t
ˆ
(νJ2 + Γ )∇q V (qs )ds − (νJ2 + Γ )−1 (pt − p0 ) − µ
t
√
J1 ∇q V (qs )ds + (νJ2 + Γ )−1 2Γ Wt .
0
Now applying Lemma 9 gives the desired result, since the above equation differs from the integral version of
(40) only by the term (νJ2 + Γ )−1 (pt − p0 ) which vanishes in the limit as → 0.
u
t
37
B.2 Hypocoercivity
The objective of this section is to prove that the perturbed dynamics (9) converges to equilibrium exponentially
fast, i.e. that the associated semigroup (Pt )t≥0 satisfies the estimate (27). We we will be using the theory of
hypocoercivity outlined in [Vil09] (see also the exposition in [Pav14, Section 6.2]). We provide a brief review
of the theory of hypocoercivity.
Let (H, h·, ·i) be a real separable Hilbert space and consider two unbounded operators A and B with domains
D(A) and D(B) respectively, B antisymmetric. Let S ⊂ H be a dense vectorspace such that S ⊂ D(A) ∩ D(B),
i.e. the operations of A and B are authorised on S. The theory of hypocoercivity is concerned with equations
of the form
∂t h + Lh = 0,
(101)
and the associated semigroup (Pt )t≥0 generated by L = A∗ A−B. Let us also introduce the notation K = ker L.
b ), A = σ∇p and B = M −1 p · ∇q − ∇q V · ∇p − µJ1 ∇q V · ∇q − νJ2 M −1 p · ∇p , it
With the choices H = L2 (π
turns out that L is the (flat) L2 (R2d )-adjoint of the generator L given in (37) and therefore equation (101)
is the Fokker-Planck equation associated to the dynamics (9). In many situations of practical interest, the
operator A∗ A is coercive only in certain directions of the state space, and therefore exponential return to
equilibrium does not follow in general. In our case for instance, the noise acts only in the p-variables and
therefore relaxation in the q-variables cannot be concluded a priori. However, intuitively speaking, the noise
gets transported through the equations by the Hamiltonian part of the dynamics. This is what the theory of
hypocoercivity makes precise. Under some conditions on the interactions between A and B (encoded in their
iterated commutators), exponential return to equilibrium can be proved. To state the main abstract theorem,
we need the following definitions:
Definition 1 (Coercivity) Let T be an unbounded operator on H with domain D(T ) and kernel K. Assume
that there exists another Hilbert space (H̃, h·, ·iH̃ ), continuously and densely embedded in K ⊥ . The operator
is said to be λ-coercive if
hT h, hiH̃ ≥ λkhk2H̃
for all h ∈ K ⊥ ∩ D(T ).
Definition 2 An operator T on H is said to be relatively bounded with respect to the operators T1 , . . . , Tn if
the intersection of the domains ∩D(Tj ) is contained in D(T ) and there exists a constant α > 0 such that
kT hk ≤ α(kT1 hk + . . . + kTn hk)
holds for all h ∈ D(T ).
We can now proceed to the main result of the theory.
Theorem 6 [Vil09, Theorem 24] Assume there exists N ∈ N and possibly unbounded operators
C0 , C1 , . . . , CN +1 , R1 , . . . , RN +1 , Z1 , . . . , ZN +1 ,
such that C0 = A,
[Cj , B] = Zj+1 Cj+1 + Rj+1
(0 ≤ j ≤ N ),
CN +1 = 0,
(102)
and for all k = 0, 1, . . . , N
(a) [A, Ck ] is relatively bounded with respect to {Cj }0≤j≤k and {Cj A}0≤j≤k−1 ,
(b) [Ck , A∗ ] is relatively bounded with respect to I and {Cj }0≤j≤k ,
(c) Rk is relatively bounded with respect to {Cj }0≤j≤k−1 and {Cj A}0≤j≤k−1 and
(d) there are positive constants λi , Λi such that λj I ≤ Zj ≤ Λj I.
P
∗
Furthermore, assume that N
j=0 Cj Cj is κ-coercive for some κ > 0. Then, there exists C ≥ 0 and λ > 0 such
that
kPt kH1 /K→H1 /K ≤ Ce−λt ,
(103)
where H1 ⊂ H is the subspace associated to the norm
v
u
N
u
X
khkH1 = tkhk2 +
kCk hk2
(104)
k=0
and K = ker(A∗ A − B).
Remark 18 Property (103) is called hypocoercivity of L on H1 := (K ⊥ , k · kH1 ).
If the conditions of the above theorem hold, we also get a regularization result for the semigroup e−tL (see
[Vil09, Theorem A.12]):
38
Theorem 7 Assume the setting and notation of Theorem 6. Then there exists a constant C > 0 such that for
all k = 0, 1, . . . , N and t ∈ (0, 1] the following holds:
kCk Pt hk ≤ C
khk
1
tk+ 2
,
h ∈ H.
Proof (of Theorem 2) . We pove the claim by verifying the conditions of Theorem 6. Recall that C0 = A = σ∇p
and
B = M −1 p · ∇q − ∇q V · ∇p − µJ1 ∇q V · ∇q − νJ2 M −1 p · ∇p .
A quick calculation shows that
A∗ = σM −1 p − σ∇p ,
so that indeed
A∗ A = Γ M −1 p · ∇p − ∇T Γ ∇ = Ltherm
and
A∗ A − B = −L∗ .
We make the choice N = 1 and calculate the commutator
[A, B] = σM −1 (∇q + νJ2 ∇p ).
Let us now set C1 = σM −1 ∇q , Z1 = 1 and R1 = νσM −1 J2 ∇p , such that (102) holds for j = 0. Note that
[A, A] = 01 , [A, C1 ] = 0 and [A∗ , C1 ] = 0. Furthermore, we have that
[A, A∗ ] = σM −1 σ.
We now compute
[C1 , B] = −σM −1 ∇2 V ∇p + µσM −1 ∇2 V J1 ∇q
and choose R2 = [C1 , B], Z2 = 1 and recall that C2 = 0 by assumption (of Theorem 6). With those choices,
assumptions (a)-(d) of Theorem 6 are fulfilled. Indeed, assumption (a) holds trivially since all relevant commutators are zero. Assumption (b) follows from the fact that [A, A∗ ] = σM −1 σ is clearly bounded relative to I. To
verify assumption (c), let us start with the case k = 1. It is necessary to show that R1 = νσM −1 J2 ∇p is bounded
relatively to A = σ∇p and A2 . This is obvious since the p-derivatives appearing in R1 can be controlled by the pderivatives appearing in A. For k = 2, a similar argument shows that R2 = −σM −1 ∇2 V ∇p +µσM −1 ∇2 V J1 ∇q
is bounded relatively to A = σ∇p and C1 = σM −1 ∇q because of the assumption that ∇2 V is bounded. Note
that it is crucial for the preceding arguments to assume that the matrices σ and M have full rank. Assumption
(d) is trivially satisfied, since Z1 and Z2 are equal to the identity. It remains to show that
T :=
N
X
Cj∗ Cj
j=0
is κ-coercive for some κ > 0. It is straightforward to see that the kernel of T consists of constant functions and
therefore
b) : π
b (φ) = 0}.
(ker T )⊥ = {φ ∈ L2 (R2d , π
Hence, κ-coercivity of T amounts to the functional inequality
ˆ
|σM
R2d
−1
2
2
ˆ
b≥κ
∇q φ| + |σ∇p φ| dπ
R2d
2
2
ˆ
b−
φ dπ
R2d
b
φdπ
,
b ).
φ ∈ H 1 (π
b ), the above is equivalent
Since the transformation φ 7→ ψ, ψ(q, p) = φ(σ −1 M q, σ −1 p) is bijective on H 1 (R2d , π
to
ˆ
ˆ
2
ˆ
b≥κ
b−
b
b ),
|∇q ψ|2 + |∇p ψ|2 dπ
ψ 2 dπ
ψdπ
, ψ ∈ H 1 (π
R2d
R2d
R2d
b . Since π
b = π ⊗ N (0, M ), coercivity of T boils down to a Poincaré inequality
i.e. a Poincaré inequality for π
for π as in Assumption 1. This concludes the proof of the hypocoercive decay estimate (103). Clearly, the
b ), and therefore it follows that there exist
abstract H1 -norm from (104) is equivalent to the Sobolev norm H 1 (π
constants C ≥ 0 and λ ≥ 0 such that
kPt f kH 1 (πb ) ≤ Ce−λt kf kH 1 (πb ) ,
1
This is not true automatically, since [A, A] stands for the array ([Aj , Ak ])jk .
(105)
39
b ) \ K, where K = ker T consists of constant functions. Let us now lift this estimate to L2 (π
b ).
for all f ∈ H 1 (π
There exist a constant C̃ ≥ 0 such that
2
X
khkH 1 (πb ) ≤ C̃
kCk hkL2 (πb ) ,
b ).
f ∈ H 1 (π
(106)
k=0
Therefore, Theorem 7 implies
kP1 f kH 1 (πb ) ≤ C̃kf kL2 (πb ) ,
b ),
f ∈ L2 (π
(107)
2
b ) \ K. It holds that
for t = 1 and a possibly different constant C̃. Let us now assume that t ≥ 1 and f ∈ L (π
kPt f kL2 (πb ) ≤ kPt f kH 1 (πb ) = kPt−1 P1 f kH 1 (πb ) ≤ Ce−λ(t−1) kP1 f kH 1 (πb ) ,
(108)
where the last inequality follows from (105). Now applying (107) and gathering constants results in
kPt f kL2 (πb ) ≤ Ce−λt kf kL2 (πb ) ,
b ) \ K.
f ∈ L2 (π
(109)
Note that although we assumed t ≥ 1, the above estimate also holds for t ≥ 0 (although possibly with a
different constant C) since kPt kL2 (πb )→L2 (πb ) is bounded on [0, 1].
u
t
C Asymptotic Variance of Linear and Quadratic Observables in the Gaussian Case
We begin by deriving a formula for the asymptotic variance of observables of the form
f (q) = q · Kq + l · q − Tr K,
d
b (f ) = 0. The following calculations
with K ∈ Rd×d
sym and l ∈ R . Note that the constant term is chosen such that π
are very much along the lines of [DLP16, Section 4]. Since the Hessian of V is bounded and the target measure
π is Gaussian, Assumption 1 is satisfied and exponential decay of the semigroup (Pt )t≥0 as in (27) follows by
Theorem 2. According to Lemma 2, the asymptotic variance is then given by
σf2 = hχ, f iL2 (πb ) ,
(110)
where χ is the solution to the Poisson equation
− Lχ = f,
Recall that
b (χ) = 0.
π
(111)
L = −Bx · ∇ + ∇T Q∇ = −x · A∇ + ∇T Q∇
is the generator as in (45), where for later convenience we have defined A = B T , i.e.
−µJ
I
2d×2d
A=
.
−I γI − νJ ∈ R
(112)
In the sequel we will solve (111) analytically. First, we introduce the notation
K 0
K̄ = 0 0 ∈ R2d×2d
and
¯
l=
l
2d
0 ∈R ,
such that by slight abuse of notation f is given by
f (x) = x · K̄x + ¯
l · x − Tr K̄.
By uniqueness (up to a constant) of the solution to the Poisson equation (111) and linearity of L, g has to be
a quadratic polynomial, so we can write
g(x) = x · Cx + D · x − Tr C,
2d×2d
where C ∈ Rsym
and D ∈ R2d (notice that C can be chosen to be symmetrical since x · Cx does not depend
on the antisymmetric part of C). Plugging this ansatz into (111) yields
−Lg(x) = x · A 2Cx + D − γ Trp C = x · K̄x + ¯
l · x − Tr K̄,
40
where
2n
X
Trp C =
Cii
i=n+1
denotes the trace of the momentum component of C. Comparing different powers of x, this leads to the
conditions
AC + CAT
AD
= K̄,
=¯
l,
(113a)
(113b)
γ Trp C = Tr K̄.
(113c)
Note that (113c) will be satisfied eventually by existence and uniqueness of the solution to (111). Then, by the
calculations in [DLP16], the asymptotic variance is given by
σf2 = 2 Tr(C K̄) + D · ¯
l.
(114)
Proof (of Proposition 2) . According to (114) and (113a), the asymptotic variance satisfies
σf2 = 2 Tr(C K̄),
where the matrix C solves
AC + CAT = K̄
(115)
and A is given as in (112). We will use the notation
C1 (µ, ν) C2 (µ.ν)
C(µ, ν) =
T
C2 (µ.ν) C3 (µ, ν)
and the abbreviations C(0) := C(0, 0), C µ (0) := ∂µ C|µ,ν=0 and C ν (0) := ∂ν C|µ,ν=0 . Let us first determine
C(0), i.e. the solution to the equation
0 I
−I γI
C(0) + C(0)
0 I
−I γI
T
=
K 0
0 0
.
This leads to the following system of equations,
C2 (0) + C2 (0)T = K,
−C1 (0) + γC2 (0) + C3 (0) = 0,
(116a)
(116b)
−C1 (0) + γC2 (0)T + C3 (0) = 0,
(116c)
−C2 (0) − C2 (0)T + 2γC3 (0) = 0.
(116d)
(116e)
Note that equations (116b) and (116c) are equivalent by taking the transpose. Plugging (116a) into (116e)
yields
1
K.
(117)
C3 (0) =
2γ
Adding (116b) and (116c), together with (116a) and (117) leads to
C1 (0) =
1
γ
K + K.
2γ
2
Solving (116b) we obtain,
C2 (0) =
1
K,
2
so that
C(0) =
1
K + γ2 K 12 K
2γ
1
1
K
K
2
2γ
!
.
(118)
Taking the µ-derivative of (115) and setting µ = ν = 0 yields
Aµ (0)C(0) + A(0)C µ (0) + C µ (0)A(0)T + C(0)Aµ (0)T = 0.
(119)
41
Notice that
Aµ (0)C(0) + C(0)Aµ (0)T
J 0
−J 0
C(0)
+
C(0)
=
00
0 0
1
γ
1
+ 2 [K, J] − 2 JK
2γ
=
.
1
KJ
0
2
With computations similar to those in the derivation of (118) (or by simple substitution), equation (119) can
be solved by
!
2
1
JK − γ4 [K, J]
− γ4 + 4γ12 + 41 [K, J] 2γ
µ
C (0) =
.
(120)
1
− 2γ
KJ − γ4 [K, J] − 4γ12 + 14 [K, J]
We employ a similar strategy to determine C ν (0): Taking the ν-derivative in equation (115), setting µ = ν = 0
and inserting C(0) and A(0) as in (118) and (112) leads to the equation
0
− 12 KJ
0 I
0 I
ν
ν
=
C
(0)
+
C
(0)
,
1
1
−I γI
−I γI
JK − 2γ [K, J]
2
which can be solved by
+ 14 [K, J]
1
1
KJ − 4 [K, J]
2
−
ν
C (0) =
1
γ
1
4γ 2
1
γ
− 12 KJ + 41 [K, J]
− 4γ12 [K, J]
!
.
(121)
Note that Tr(C K̄) = Tr(C1 K), and so
∂µ Θ|µ,ν=0 = 2 Tr(C1µ (0)K) =
=−
γ2
1
1
+ 2 +
· Tr([K, J]K) = 0,
4
4γ
4
since clearly Tr([K, J], K) = Tr(KJK) − Tr(JK 2 ) = 0. In the same way it follows that
∂ν Θ|µ,ν=0 = 0,
proving (47).
Taking the second µ-derivative of (115) and setting µ = ν = 0 yields
2Aµ (0)C µ (0) + A(0)C µµ (0) + C µµ (0)A(0)T + 2C µ (0)Aµ (0)T = 0,
employing the notation C µµ (0) = ∂µ2 C|µ,ν=0 and noticing that ∂µ2 A = 0. Using (120) we calculate
!
γ2
1
+ 4γ12 + 41 [J, [K, J]] − 2γ
J 2 K + γ4 J[K, J]
µ
µ
µ
µ
T
4
.
A (0)C (0) + C (0)A (0) =
1
KJ 2 − γ4 [K, J]J
0
− 2γ
As before, we make the ansatz
C µµ (0) =
C1µµ (0) C2µµ (0)
(C2µµ (0))T C3µµ (0)
,
leading to the equations
1
γ2
1
+ 2 + [J, [K, J]]
4
4γ
4
γ
1 2
µµ
µµ
µ
−C1 (0) + γC2 (0) + C3 (0) = J K − J[K, J]
γ
2
γ
1
µµ
2
µµ
T
µ
−C1 (0) + γC2 (0) + C3 (0) = KJ + [K, J]J
γ
2
C2µµ (0) + C2µµ (0)T = −
−C2µµ (0) − C2µµ (0)T + 2γC3µµ (0) = 0.
(122a)
(122b)
(122c)
(122d)
Again, (122b) and (122c) are equivalent by taking the transpose. Plugging (122a) into (122d) and combing
with (122b) or (122c) gives
C1µµ (0) =
γ
1
γ3
1
+ 3 +
(2JKJ − J 2 K − KJ 2 ) − JKJ.
4
4γ
4
γ
42
Now
2
1
+ γ 3 ) Tr(JKJK) − Tr(J 2 K 2 ) − Tr(JKJK)
γ3
γ
gives the first part of (48). We proceed in the same way to determine C1νν (0). Analogously, we get
!
1
(KJ 2 − 21 [K, J]J)
0
ν
ν
ν
ν
T
γ
.
A (0)C (0) + C (0)A (0) = 1
(JKJ − 12 J[K, J]) 2γ12 ([K, J]J − J[K, J])
γ
∂µ2 Θ|µ,ν=0 = 2 Tr(C1µµ (0)K) = −(γ +
Solving the resulting linear matrix system (similar to (122a)-(122d)) results in
1
1
1
1
C1νν (0) =
(KJ 2 + J 2 K) −
JKJ,
−
+
4γ 3
4γ
2γ 3
2γ
leading to
1
1
1
1
Tr(J 2 K 2 ) −
Tr(JKJK).
−
+
γ3
2γ
2γ 3
2γ
2
To compute the cross term C1µν (0) we take the mixed derivative ∂µν
of (115) and set µ = ν = 0 to arrive at
∂ν2 Θ|µ,ν=0 = 2 Tr(C1νν (0)K) =
Aµ (0)C ν (0) + Aν (0)C µ (0) + A(0)C µν (0) + C µν (0)A(0)T + C µ (0)Aν (0)T + C ν (0)Aµ (0)T = 0.
Using (120) and (121) we see that
Aµ (0)C ν (0) + Aν (0)C µ (0) + C µ (0)Aν (0)T + C ν (0)Aµ (0)T
1
− 14 [J, [K, J]]
4γ 2
=
1
1
JKJ + 2γ
KJ 2 + γ4 J[K, J] −
2γ
1
JKJ
γ
1
[K, J]J
4γ
1
J[K, J] − γ4 [K, J]J
4γ
1
+ 14 [J, [K, J]]
4γ 2
−
!
.
The ensuing linear matrix system yields the solution
C1µν (0) = −
γ
1
1
1
+ −
[J, [K, J]] + JKJ,
4γ 3
4
4γ
γ
leading to
2
∂µν
Θ|µ,ν=0 = 2 Tr(C1µν (0)K) =
1
1
1
1
+ − γ Tr(J 2 K 2 ) + − 3 + + γ Tr(JKJK).
γ3
γ
γ
γ
(123)
u
t
This completes the proof.
Proof (Proof of Proposition 3) By (113b) and (114) the function Θ satisfies
l.
Θ(µ, ν) = ¯
l · A−1 ¯
Recall the following formula for blockwise inversion of matrices using the Schur complement:
−1
U V
(U − V X −1 W )−1 . . . ,
=
W X
...
...
(124)
provided that X and U − V X −1 W are invertible. Using this, we obtain
Θ(µ, ν) = l · − µJ + (γ − νJ)−1 l.
Taking derivatives, setting µ = ν = 0 and using the fact that J T = −J leads to the desired result.
u
t
Lemma 10 The following holds:
(a) γ − γ43 − γ 3 − γ1 < 0 for γ ∈ (0, ∞).
(b) Let J = −J T and K = K T . Then Tr(JKJK) − Tr(J 2 K 2 ) ≥ 0. Furthermore, equality holds if and only if
[J, K] = 0.
Proof To show (a) we note that γ − γ43 −γ 3 − γ1 < γ − γ43 −γ 3 = γ(1− γ44 −γ 2 ). The function f (γ) := 1− γ44 −γ 2
has a unique global maximum on (0, ∞) at γmin = 81/6 with f (γmin ) = −2, so the result follows.
For (b) we note that [J, K]T = [J, K], and that [J, K]2 is symmetric and nonnegative definite. We can write
X 2
Tr([J, K]2 ) =
λi ,
i
with λi denoting the (real) eigenvalues of [J, K]. From this it follows that Tr([J, K]2 ) ≥ 0 with equality if and
only if [J, K] = 0. Now expand
Tr([J, K]2 ) = 2 Tr(JKJK) − Tr(J 2 K 2 ),
which implies the advertised claim.
u
t
43
D Orthogonal Transformation of Tracefree Symmetric Matrices into a Matrix with
Zeros on the Diagonal
d
Given a symmetric matrix K ∈ Rd×d
sym with Tr K = 0, we seek to find an orthogonal matrix U ∈ O(R ) such
that U KU T has zeros on the diagonal. This is a crucial step in Algorithms 1 and 2 and has been addressed in
various places in the literature (see for instance [Kaz88] or [Bha97], Chapter 2, Section 2). For the convenience
of the reader, in the following we summarize an algorithm very similar to the one in [Kaz88].
Since K is symmetric, there exists an orthogonal matrix U0 ∈ O(Rd ) such that U0 KU0T = diag(λ1 , . . . , λd ).
Now the algorithm proceeds iteratively, orthogonally transforming this matrix into one with the first diagonal entry vanishing, then the first two diagonal entries vanishing, etc, until after d steps we are left with a
matrix with zeros on the diagonal. Starting with λ1 , assume that λ1 6= 0 (otherwise proceed with λ2 ). Since
P
Tr(K) = Tr(U0 KU0T ) =
λi = 0, there exists λj , j ∈ {2, . . . , d} such that λ1 λj < 0 (i.e. λ1 and λj have
opposing signs). We now apply a rotation in the 1j-plane to transform the first diagonal entry into zero. More
specifically, let
j
cos α 0 . . . 0 − sin α
0
. . . 0
..
0
1
0
.
.
..
.
.
.
1
0
0
j ∈ O(Rd )
U1 = sin α 0
0
cos α
0
..
0
0
1
.
..
..
. 0
.
0
0 ...
...
0
1
with α = arctan
q
− λλ1j . We then have (U1 U0 KU0T U1T )11 = 0. Now the same procedure can be applied to the
second diagonal entry λ2 , leading to the matrix U2 U1 U0 KU0T U1T U2T with
(U2 U1 U0 KU0T U1T U2T )11 = (U2 U1 U0 KU0T U1T U2T )22 = 0
Iterating this process, we obtain that Ud . . . U1 U0 KU0T U1T . . . UdT has zeros on the diagonal, so Ud . . . U1 U0 ∈
O(Rd ) is the required orthogonal transformation.
References
AE14.
A. Arnold and J. Erb. Sharp entropy decay for hypocoercive and non-symmetric Fokker-Planck
equations with linear drift. arXiv:1409.5425v2, 2014.
AMO16. H. Alrachid, L. Mones, and C. Ortner. Some remarks on preconditioning molecular dynamics. arXiv
preprint arXiv:1612.05435, 2016.
Bas98.
R. F. Bass. Diffusions and elliptic operators. Springer Science & Business Media, 1998.
Ben75.
C. H. Bennett. Mass tensor molecular dynamics. Journal of Computational Physics, 19(3):267 –
279, 1975.
BGL13.
D. Bakry, I. Gentil, and M. Ledoux. Analysis and geometry of Markov diffusion operators, volume
348. Springer Science & Business Media, 2013.
Bha97.
R. Bhatia. Matrix analysis, volume 169 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 1997.
BPSSS11. A. Beskos, F. J. Pinski, J. M. Sanz-Serna, and A. M. Stuart. Hybrid Monte Carlo on Hilbert spaces.
Stochastic Process. Appl., 121(10):2201–2230, 2011.
BRSV08. A. Beskos, G. Roberts, A. Stuart, and J. Voss. MCMC methods for diffusion bridges. Stoch. Dyn.,
8(3):319–350, 2008.
BS09.
A. Beskos and A. Stuart. MCMC methods for sampling function space. In ICIAM 07—6th International Congress on Industrial and Applied Mathematics, pages 337–364. Eur. Math. Soc., Zürich,
2009.
CBP09.
M. Ceriotti, G. Bussi, and M. Parrinello. Langevin equation with colored noise for constanttemperature molecular dynamics simulations. Physical Review Letters, 102(2):020601, 2009.
CCG12. P. Cattiaux, D. Chafaı, and A. Guillin. Central limit theorems for additive functionals of ergodic
markov diffusions processes. ALEA, 9(2):337–382, 2012.
DLP16.
A. B. Duncan, T. Lelievre, and G. A. Pavliotis. Variance reduction using nonreversible Langevin
samplers. Journal of Statistical Physics, 163(3):457–491, 2016.
DMS15. J. Dolbeault, C. Mouhot, and C. Schmeiser. Hypocoercivity for linear kinetic equations conserving
mass. Trans. Amer. Math. Soc., 367(6):3807–3828, 2015.
44
EK86.
EN00.
GC11.
GCS+ 14.
HHMS93.
HHMS05.
HJ13.
HNW15.
HSV07.
HSV09.
HSVW05.
JO10.
Kaz88.
Kli87.
KLO12.
Liu08.
LM15.
LNP13.
LRS10.
LS16.
MCF15.
MPP02.
MV00.
MWL16.
OP11.
OPPS12.
OPPS15.
OPPS16.
Pav14.
PS03.
PS05.
RBS15a.
S. N. Ethier and T. G. Kurtz. Markov processes. Wiley Series in Probability and Mathematical
Statistics: Probability and Mathematical Statistics. John Wiley & Sons, Inc., New York, 1986.
Characterization and convergence.
K.-J. Engel and R. Nagel. One-parameter semigroups for linear evolution equations, volume 194
of Graduate Texts in Mathematics. Springer-Verlag, New York, 2000. With contributions by S.
Brendle, M. Campiti, T. Hahn, G. Metafune, G. Nickel, D. Pallara, C. Perazzoli, A. Rhandi, S.
Romanelli and R. Schnaubelt.
M. Girolami and B. Calderhead. Riemann manifold Langevin and Hamiltonian Monte Carlo methods. J. R. Stat. Soc. Ser. B Stat. Methodol., 73(2):123–214, 2011. With discussion and a reply by
the authors.
A. Gelman, J. B. Carlin, H. S. Stern, D. B. Dunson, A. Vehtari, and D. B. Rubin. Bayesian data
analysis. Texts in Statistical Science Series. CRC Press, Boca Raton, FL, third edition, 2014.
C.-R. Hwang, S. Hwang-Ma, and S. J. Sheu. Accelerating Gaussian diffusions. Ann. Appl. Probab.,
3(3):897–913, 1993.
C.-R. Hwang, S.-Y. Hwang-Ma, and S.-J. Sheu. Accelerating diffusions. Ann. Appl. Probab.,
15(2):1433–1444, 2005.
Roger A. Horn and Charles R. Johnson. Matrix analysis. Cambridge University Press, Cambridge,
second edition, 2013.
C.-R. Hwang, R. Normand, and S.-J. Wu. Variance reduction for diffusions. Stochastic Process.
Appl., 125(9):3522–3540, 2015.
M. Hairer, A. M. Stuart, and J. Voss. Analysis of SPDEs arising in path sampling. II. The nonlinear
case. Ann. Appl. Probab., 17(5-6):1657–1706, 2007.
M. Hairer, A. Stuart, and J. Voss. Sampling conditioned diffusions. In Trends in stochastic analysis, volume 353 of London Math. Soc. Lecture Note Ser., pages 159–185. Cambridge Univ. Press,
Cambridge, 2009.
M. Hairer, A. M. Stuart, J. Voss, and P. Wiberg. Analysis of SPDEs arising in path sampling. I.
The Gaussian case. Commun. Math. Sci., 3(4):587–603, 2005.
A. Joulin and Y. Ollivier. Curvature, concentration and error estimates for Markov chain Monte
Carlo. Ann. Probab., 38(6):2418–2442, 2010.
J. Y. Kazakia. Orthogonal transformation of a trace free symmetric matrix into one with zero
diagonal elements. Internat. J. Engrg. Sci., 26(8):903–906, 1988.
W. Kliemann. Recurrence and invariant measures for degenerate diffusions. The Annals of Probability, pages 690–707, 1987.
T. Komorowski, C. Landim, and S. Olla. Fluctuations in Markov processes, volume 345 of
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer, Heidelberg, 2012. Time symmetry and martingale approximation.
J. S. Liu. Monte Carlo strategies in scientific computing. Springer Science & Business Media, 2008.
B. Leimkuhler and C. Matthews. Molecular dynamics, volume 39 of Interdisciplinary Applied Mathematics. Springer, Cham, 2015. With deterministic and stochastic numerical methods.
T. Lelièvre, F. Nier, and G. A. Pavliotis. Optimal non-reversible linear drift for the convergence to
equilibrium of a diffusion. J. Stat. Phys., 152(2):237–274, 2013.
T. Lelièvre, M. Rousset, and G. Stoltz. Free energy computations. Imperial College Press, London,
2010. A mathematical perspective.
T. Lelièvre and G. Stoltz. Partial differential equations and stochastic methods in molecular dynamics. Acta Numer., 25:681–880, 2016.
Y.-A. Ma, T. Chen, and E. Fox. A complete recipe for stochastic gradient MCMC. In Advances in
Neural Information Processing Systems, pages 2899–2907, 2015.
G. Metafune, D. Pallara, and E. Priola. Spectrum of Ornstein-Uhlenbeck operators in Lp spaces
with respect to invariant measures. J. Funct. Anal., 196(1):40–60, 2002.
P. A. Markowich and C. Villani. On the trend to equilibrium for the Fokker-Planck equation: an
interplay between physics and functional analysis. Mat. Contemp, 19:1–29, 2000.
C. Matthews, J. Weare, and B. Leimkuhler. Ensemble preconditioning for Markov Chain Monte
Carlo simulation. arXiv:1607.03954, 2016.
M. Ottobre and G. A. Pavliotis. Asymptotic analysis for the generalized Langevin equation. Nonlinearity, 24(5):1629, 2011.
M. Ottobre, G. A. Pavliotis, and K. Pravda-Starov. Exponential return to equilibrium for hypoelliptic quadratic systems. J. Funct. Anal., 262(9):4000–4039, 2012.
M. Ottobre, G. A. Pavliotis, and K. Pravda-Starov. Some remarks on degenerate hypoelliptic
Ornstein–Uhlenbeck operators. J. Math. Anal. Appl., 429(2):676–712, 2015.
M. Ottobre, N. S. Pillai, F. J. Pinski, and Andrew M. Stuart. A function space HMC algorithm
with second order Langevin diffusion limit. Bernoulli, 22(1):60–106, 2016.
G. A. Pavliotis. Stochastic processes and applications: Diffusion Processes, the Fokker-Planck and
Langevin Equations, volume 60. Springer, 2014.
G. A. Pavliotis and A. M. Stuart. White noise limits for inertial particles in a random field.
Multiscale Model. Simul., 1(4):527–533 (electronic), 2003.
G. A. Pavliotis and A. M. Stuart. Analysis of white noise limits for stochastic systems with two
fast relaxation times. Multiscale Model. Simul., 4(1):1–35 (electronic), 2005.
L. Rey-Bellet and K. Spiliopoulos. Irreversible Langevin samplers and variance reduction: a large
deviations approach. Nonlinearity, 28(7):2081–2103, 2015.
45
RBS15b.
RC13.
Vil09.
WHC14.
L. Rey-Bellet and K. Spiliopoulos. Variance reduction for irreversible Langevin samplers and diffusion on graphs. Electron. Commun. Probab., 20:no. 15, 16, 2015.
C. Robert and G. Casella. Monte Carlo statistical methods. Springer Science & Business Media,
2013.
C. Villani. Hypocoercivity. Number 949-951. American Mathematical Soc., 2009.
S.-J. Wu, C.-R. Hwang, and M. T. Chu. Attaining the optimal Gaussian diffusion acceleration. J.
Stat. Phys., 155(3):571–590, 2014.
| 10math.ST
|
DIVERGENCES MEASURES
Amadou Diadie Ba(1) , Gane Samb Lo(1,2) .
(1)
arXiv:1604.04242v2 [stat.ME] 17 Aug 2016
(2)
LERSTAD, Université Gaston Berger, Sénégal.
LSA, Université Pierre et Marie Curie, France.
1. Introduction
In this paper, we deal with divergence measures estimation using both wavelet and classical probability
density functions. Let P be a class of two probability measures on Rd , a divergence measure on P is an
application,
(1.1)
D:
P2
−→ R
(Q, L) 7−→ D(Q, L)
such that D(Q, Q) = 0 for any Q ∈ P.
A divergence measure then is not necessarily symmetrical and it does neither have to be a metric. To better
explain our concern, let us intoduce some of the most celebrated divergence measures. Most of them are based
on probability density functions. So let us suppose that all Q ∈ P have p.d.f. fQ with respect to a σ-finite
measure ν on (Rd , B(Rd )) that is usually the Lebesgues measure. We have the L22 -divergence measure
Z
(fQ (x) − fL (x))2 dν(x);
DL2 (Q, L) =
Rd
the family of Renyi divergence measures indexed by α 6= 1, more known under the name of Renyi-α,
Z
1
(1.2)
DR,α (Q, L) =
log
fQα (x)fL1−α (x)dν(x) ;
α−1
Rd
the family of Tsallis divergence measures indexed by α 6= 1, also known under the name of Tsallis-α
Z
1
1−α
α
(1.3)
DT,α (Q, L) =
fQ (x)fL (x) − 1 dν(x);
α−1
Rd
and finally the Kulback-Leibler divergence measure
Z
fQ (x) log(fQ (x)/fL (x)) dν(x).
(1.4)
DKL (Q, L) =
Rd
The latter, the Kullback-Leibler measure, may be interpreted as a limit case of both the Renyi’s family and
the Tsallis’ one by letting α → 1. As well, for α near 1, the Tsallis family may be seen as derived from a fisrt
order expansion DR,α (Q, L) based on the first order expansion of the logarithm function in the neigborhood
of the unity.
Although we are focusing on the aforementioned divergence measures, we have to attract the attention of
the reader that there exist quite a few number of them. Let us cite for example the ones denamed as : AliSilvey distance or f -divergence [[1]], Cauchy-Schwarz, Jeffrey’s divergence (see [2]), Chernoff, Jensen-Shannon
divergence etc. According to [3], there is more than a dozen of different divergence measures that one can
find in the literature.
1
2
DIVERGENCES MEASURES CONSISTENCY BANDS
Before coming back to our divergence measures of interest, we want to highlight some important applications of them. Indeed, divergence has proven to be useful in applications. Let us cite a few of them
(a) It may be as a similarity measure in image registration or multimedia classification (see [4]). It is
also applicable as a loss function in evaluating and optimizing the performance of density estimation
methods (see [6]).
(b) The estimation of divergence between the samples drawn from unknown distributions gauges the
distance between those distributions. Divergence estimates can then be used in clustering and in particular for deciding whether the samples come from the same distribution by comparing the estimate
to a threshold.
(c) Divergence estimates can also be used to determine sample sizes required to achieve given performance
levels in hypothesis testing.
(d) Divergence gauges how differently two random variables are distributed and it provides a useful
measure of discrepancy between distributions. In the frame of information theory , the key role of
divergence is well known.
(e) There has been a growing interest in applying divergence to various fields of science and engineering
for the purpose of estimation, classification, etc [[7], [9]].
(f) Divergence also plays a central role in the frame of large deviations results including the asymptotic
rate of decrease of error probability in binary hypothesis testing problems.
The reader may find more applications descriptions in the following papers: [10],[11], [12], [14], [15], [16],
[17],[18]
We may see two kinds of problems we encounter when dealing with these objects. First, the divergence
measures may not be finite on the whole support of the distributions.
These two remarks apply to too many divergence measures. Both these problems are avoided with some
boundedness assumption as in Singh et al. [25] and in Krishnamurthy et al. [23]. In the case where all Q ∈ P
have p.d.f. fQ with respect to a σ-finite measure ν on (Rd , B(Rd )), these authors suppose that there exist two
finite numbers 0 < κ1 < κ2 < +∞ such that
(1.5)
κ1 ≤ f Q , f L ≤ κ2 .
so that the quantities Iα (PX , PY , α), for example, are finite in the expressions of Renyi-α and Tsallis-α mea-
sures and that the Kullback-Leibler is also finite. We will follow these authors by adopting the Assumption
4 throughout this paper.
1.1. Divergence measures as goodness-of-fit tests. The divergence measures may be applied to two
statistical problems among others.
First, it may be used as a goodness-of-fit problem like that : let X1 , X2 , .... a sample from X with an
unkown probability distribution PX and we want to test the hypothesis that PX is equal to a known and
fixed probability P0 .
For example Jager et al. (in [5]) proposed P0 to be the uniform probability distribution on [0, 1].
DIVERGENCES MEASURES CONSISTENCY BANDS
3
Theoritically, if we want to test the null hypothesis H0 : F = F0 versus H1 : F 6= F0 we have to use any of
general phi-divergence test statistic φ(fn (x), f0 (x)). Then our test statistic is of the form
Z
D(fn , f0 ) = φ(fn (x), f0 (x))dx
Then we can answer this question by estimating a divergence measure D(PX , P0 ) by the plug-in estimator
(n)
D(PX , P0 ) based on the sequences of empirical probabilities
n
(n)
PX =
1X
δX .
n i=1 i
(n)
From there establishing an asymptotic theory of ∆n = D(PX , P0 ) − D(PX , P0 ) is necessary to conclude.
1.2. Divergence measures as a comparison tool problem. As a comparison tool for two distributions,
we may have two samples and wonder whether they come from the same probability measure. Here, we also
may two different cases. In the first, we have two independent samples X1 , X2 , .... and Y1 , Y2 , .... respectively
(n)
(n)
from a random variable X and Y. Here the empirical divergence D(PX , PY ) is the natural estimator of
D(PX , PY ) on which depends the statistical test of PX = PY .
But the data may aslo be paired (X, Y ), (X1 , Y2 ), (X2 , Y2 ), ..., that is Xi and Yi are measurements of the
same case i = 1, 2, ... In that case, testing the equality of the margins PX = PY should be based on the
empirical probabilities from the couple (X, Y ), that is
n
(n)
P(X,Y ) =
1X
δ(Xi ,Yi ) .
n i=1
1.3. Related work : Krisnamurthy et al. ([23]), Singh and Poczos ([25] ) studied mainly the independent
case of the two distributions comparaison. They both used divergence measures based on probability density
functions and concentrated of Renyi-α, Tsallis-α and Kullback-Leibler. (???). Reyni :
- Singh and Poczos [25] proposed divergence estimators that achieve the parametric convergence of rate
n−N (s) where 0 < N (s) ≤ 1/2 depends on the smoothness s of the densities f and g both in a Holder class
of smothness s
They showed that
and
i
h
(n)
(n)
E DT,α (PX , PY ) − DT,α (PX , PY ) = O (ǫn )
h
i
(n)
(n)
E DR,α (PX , PY ) − DR,α (PX , PY ) = O (ǫn )
4s
where ǫn = Ω(n−γ ) and γ = min{ 4s+d
, 1/2}.
Singh and Poczos [25] , and Krishnamurthy et al [23] each proposed divergence estimators that achieve the
parametric convergence rate O T1 under weaker conditions than those given in [39].
4
DIVERGENCES MEASURES CONSISTENCY BANDS
Krishnamurthy et al [23] proposed three estimators for the divergence measures DL2 (PX , PY ), DR,α (PX , PY ),
and for DT,α (PX , PY ): the plugging (pl), linear (lin), and the quadratic (qd) one. They showed that
s
(n)
(n)
pl
E DT,α
(PX , PY ) − DT,α (PX , PY ) = O n− 2s+d
and
(n)
(n)
lin
E DT,α
(PX , PY ) − DT,α (PX , PY )
(n)
(n)
qd
E DT,α
(PX , PY ) − DT,α (PX , PY )
with the quadratic estimator
(n)
(n)
qd
E DR,α
(PX , PY ) − DR,α (PX , PY )
(n)
(n)
qd
E DT,α
(PX , PY ) − DT,α (PX , PY )
≤
≤
1
2s
c n− 2 + n− 2s+d
1
3s
c n− 2 + n− 2s+d
1
3s
≤ c n− 2 + n− 2s+d
1
3s
≤ c n− 2 + n− 2s+d
Poczos and Jeff ([26]) considered two samples not necessarily with the same size and used the k−Nearest
Neighbour (kNN) based density estimators. They showed that, if |α − 1| < k, then Reyni estimator est
asymptotically unbiaised that is
(m)
(n)
lim E DR,α (PX , PY ) = DR,α (PX , PY )
n,m→∞
and it is consistent for L2 norm that is
2
(n)
(m)
lim E DR,α (PX , PY ) − DR,α (PX , PY ) = 0.
n,m→∞
All this is under conditions on the densities fQ and fL .
In Liu et al. [27] and worked with densities in Holder classes, whereas our work applies for densities in the
Bessov class
In any case, the asymptotic distributions of the estimators in [[30], [25],[23]] are currently unknown.
===========
But, in our view this case should rely on the available data so that using the same sample size may lead
to a reduction (?). To apply their method, one should take the minimum of the two sizes and then loose
(n)
(m)
information. We suggest to come back to a general case and then study the asymptotics of D(PX , PY )
based on samples X1 , X2 , .., Xn . and Y1 , Y2 , ..., Ym .
As for the fitting approach, we may cite Hamza et al. ([28]) who used modern techniques of Mason and
co-authors ([31] ,[32] ,[33]) on consistency bounds for p.d.f’s kernel estimators. But, these authors, hamza and
al., in the current version of their work, did not address the existence problem of the divergence measures.
We will seize the opportunity of these papers to correct this.
Also, for the fitting case, and when using Renyi-α, Tsallis-α and Kulback-Leibler measures, we do not have
(n)
symmetry. So we have to deal with the estimation of both D(PX , P0 ) by D(PX , P0 ) and that of D(P0 , PX )
(n)
by D(P0 , PX ) and decided which of these cases is better.
DIVERGENCES MEASURES CONSISTENCY BANDS
5
As to the paired case, we are not aware of works on this. Yet, this approach is very important and should
be addressed.
This paper will be devoted to a general study the estimation the Renyi-α, Tsallis-α, Kulback-Leibler, and
L2 measures in the three level : fitting, independent comparison and paired comparison.
We will use empirical estimations of the density functions both by the Parzen estimator and the wavelet
ones. The main novelty here resides in the wavelet approach. When using the Parzen statistics, our main
tool will be modern techniques of Mason and co-authors ([31] ,[32] ,[33]) on consistency bounds for p.d.f’s
kernel. For the wavelet approch, we will mainly back on the Giné and Nickl paper ([43]).
Since the tools we are using do not have the level of developpement, our results for the Parzen scheme will
use k-dimensional distributions, while those pertaining to the wavelet frame are set for univariate distributions.
But, we will have to give a precise account of wavelet theory and its applications to statistical estimation,
using Hardle et al.([37]), .....
The paper will be organized as follows. In section 3, we will describe how to use the density estimations
both for Parzen and wavelets as well as the statements of the main hypothesis. As for wavelets, a broader
account will be given in Appendix .. In Section 3, we deal with the fitting questions. Section 4 is devoted to
independent distribution comparison. Finally in Section 5, we deal with margins distribution comparison. In
all Sections 3, 4 and 5, we will establish strong efficiency and central limit theorems.
Under standards assumptions on the densities fQ (x), fL (x), on the scale function ϕ and on the wavelet
kernel K (formalized in the sequel) we establish the following properties.
a) We define the linear wavelet density estimators and establish the consistency of these density estimators
b) We establish the asymptotic consistency showing that . . .(Theorem ??)
b) When .... we prove that the estimator is asymptotically normal (Theorem ??).
c) We derive . . .
d) We also prove
e) Lastly, we prove
Organization of the paper (plan)
2. RESULTS
We are going to establish general results both for consistency and asymptotic normality. Next results for
particular divergences measures will follow as corollaries.
6
DIVERGENCES MEASURES CONSISTENCY BANDS
2.1. General conditions. Let J(f, g) be a functional of two densities functions f and g satisfying Assumption 4 below of the form
J(f, g) =
Z
φ(f (x), g(x))dx
D
where φ(s, t) is a function of (s, t) ∈ R2 of class C 2 . We adopt the following notations with respect to the partial
derivatives :
(1)
φ1 (s, t) =
∂φ
∂φ
(1)
(s, t), φ2 (s, t) =
(s, t)
∂s
∂t
and
∂2φ
∂ 2φ
∂2φ
(2)
(2)
(2)
(s, t), φ2 (s, t) = 2 (s, t), φ1,2 (s, t) = φ2,1 (s, t) =
(s, t).
2
∂s
∂t
∂s∂t
We require the following general conditions
(2)
φ1 (s, t) =
C1-φ : The following integrals are finite
Z n
o
(1)
(1)
|φ1 (f (x), g(x))| + |φ2 (f (x), g(x))| dx < +∞.
(1)
(2)
(1)
(2)
C2-φ : For any measurable sequences of functions δn (x), δn (x), ρn (x), and ρn (x) of x ∈ D, uniformly
converging to zero, that is
max
i=1,2, j=1,2
then
(2.1)
Z
φ1
(2.2)
Z
φ2
(2)
(2)
and
(2.3)
Z
(2)
φ1,2
sup
n
δn(i) (x) + ρ(j)
n (x)
o
< +∞,
Z
(2)
f (x) + δn(1) (x), g(x) dx → φ1 (f (x), g(x))dx,
Z
(2)
f (x), g(x) + δn(2) (x) dx → φ2 (f (x), g(x))dx
f (x) +
ρ(1)
n (x), g(x)
+
ρ(2)
n (x)
dx →
Z
(2)
φ1,2 (f (x), g(x))dx.
Remark 1. These results may result from the Dominated Convergence Theorem or the monotone Convergence
Theorem or from other limit theorems. We may either express conditions under which these results hold true
on the general function φ. But we choose here to state the final results and next, to check them for particular
cases, on which reside our real interests.
Our general results concern the estimations of J(f, g) in a one sample (see Theorem 2) and two samples
problems (see Theorem 3). In both case, we use the linear wavelet estimators of f and g, denoted fn and gn ,
and defined in [37]. From there we mainly use results for Giné and Nickl [43].
Under their conditions, we define
(2.4)
an = kfn − f k∞ , bn = kgn − gk∞ , cn = an ∨ bn ,
where khk∞ stands for supx∈D |h(x)|.
DIVERGENCES MEASURES CONSISTENCY BANDS
7
2.2. Wavelet setting. The wavelet setting involves two functions ϕ and ψ in L2 (R) such that
n
o
2j/2 ϕ(2j (.) − k), 2j/2 ψ(2j (.) − k), (j, k) ∈ Z2
be a orthonormal basis of L2 (R) . The associated kernel function of the wavelets ϕ and ψ is defined by
Kj (x, y) = 2j K(2j x, 2j y), j ∈ N
where K(x, y) =
P
k∈Z
ϕ(x − k)(y − k), x, y ∈ R.
For a mesurable function we define Kj (h)(x) =
Assuming the following :
R
Kj (x, y)h(y)dy.
Assumption 1. (S ). ϕ and ψ are bounded and have compact support and either (i) the father wavelet ϕ
R
has weak derivatives up to order S in Lp (R) or (ii) ψ has S vanishing moments, i.e xm ψ(x)dx = 0 for all
m = 0, . . . , S − 1.
Assumption 2. ϕ is of bounded p-variation for some 1 ≤ p < ∞ and vanishes on (B1 , B2 ]c for some
−∞ < B1 < B2 < ∞.
Assumption 3. The resolution level j := jn is such that 2jn ≈ n1/4 .
With this assumption, one has jn ր ∞ and
(2.5)
r
jn 2jn
+ 2−t/jn
n
jn
log log n
=
s
1 log n
+ n−t/4 → 0 as n → ∞, ∀t > 0
4 log 2 n3/4
→ ∞ as n → ∞,
sup (j2n − jn ) ≤ τ .
and
n≥n0
These conditions allow the use of results of Giné [43].
Definition 1. Given two independent samples with size n X1 , . . . Xn ∼ f and Y1 , . . . Yn ∼ g respectively from
a random variable X and Y and absolute continuous law PX and PY on R, straighforward wavelets estimators
of f and g are defined by
n
fn (x) = Pn,X (Kjn (x, .)) =
(2.6)
and
1X
Kj (x, Xi )
n i=1 n
n
1X
Kj (x, Yi )
gn (x) = Pn,Y (Kjn (x, .)) =
n i=1 n
(2.7)
t
In the sequel we suppose the densities f and g belong to the Besov space B∞,∞
(R) (see [37]),
h ∈ L∞ (R) : khks,∞,∞ := sup |αk (h)| + sup sup 2l(t+1/2) βlk (h) < ∞
k∈Z
where αk (h) =
the function h.
R
R
h(x) ϕ(x − k)dx and βlk (h) =
l≥0 k∈Z
R
R
2l/2 h(x) ψ(2l x − k)dx are the wavelet coefficients of
t
The spaces B∞,∞
(R) are the Holder-Zygmund spaces, which contain the classical Holder-Lipschitz spaces.
Given these definitions, we now describe how we will use the wavelet approach.
8
DIVERGENCES MEASURES CONSISTENCY BANDS
t
It is remarquable from Theorem 3 (in [43]) that, if the densities f and g belong to B∞,∞
(R), ϕ satisfies
Assumption2, and ϕ, ψ satisfy Assumption1 (T ) then an , bn and cn are all of them
s
!
1 log n
−t/4
O
+n
4 log 2 n3/4
almost surely and converge all to zero at this rate (with 0 < t < T ).
In order to establish the asymptotic normality of the divergences estimators, we need to recall some facts
about kernels wavelets.
t
For h ∈ B∞,∞
(R), the Theorem 1 below provides the asymptotic normality of
√
n
Z
(fn (x) − f (x))h(x)dx
necessary for setting the asymptotic normality of divergence measure, provided the finitness of PX (Kjn (h)(X))2 .
t
Theorem 1. Under Assumption 1 and 2 and if h ∈ B∞,∞
(R), then we have
√
n
Z
(fn (x) − f (x))h(x)dx
N (0, σh2 ) as n → ∞,
where
σh2 = PX (Kjn (h)(X))2 − (PX (Kjn (h)(X))2
(2.8)
The symbol
denotes the convergence in law.
R
PX (h) = h(x)f (x)dx denotes the expection of the measurable function h.
The proof of this theorem is postpooned to Subsection 4.
2.3. Main results. In the sequel J(f, g) is a functional of two densities functions f and g satisfying Assumption 4 and defined by
J(f, g) =
Z
φ(f (x), g(x))dx
D
where φ(s, t) is a function of (s, t) ∈ R2 of class C 2 .
Define the functions h1 and h2 by
(1)
(1)
h1 (x) = φ1 (f (x), g(x)) and h2 (x) = φ2 (f (x), g(x))
and the constants A1 and A2 by
A1 =
Z
|h1 (x)| dx and A2 =
Suppose that A1 and A2 are both finites.
Z
|h2 (x)| dx
DIVERGENCES MEASURES CONSISTENCY BANDS
9
2.3.1. One side estimation. Suppose that either we have a sample X1 , . . . , Xn with unknown p.d.f f and
a known p.d.f g and we want to study the limit behavior of J(fn , g) or we have a sample Y1 , . . . , Yn with
unknown p.d.f g and a known p.d.f. f and we want to study the limit behavior of J(f, gn ).
fn or gn are as in (2.6) or in (2.7).
Theorem 2. Under Assumption 1 and 2, we have :
• Consistency :
(2.9)
lim sup
n→∞
|J(fn , g) − J(f, g)|
≤ A1 , a.s
an
and
(2.10)
lim sup
n→∞
|J(f, gn ) − J(f, g)|
≤ A2 , a.s
bn
where an and bn are as in (2.4).
• Asymptotic normality :
(2.11)
√
n(J(fn , g) − J(f, g))
N 0, σ12
as n → ∞
√
n(J(f, gn ) − J(f, g))
N 0, σ22
as n → ∞
and
(2.12)
where
σ12 = PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X)))2
(2.13)
and
(2.14)
2
2
σ22 = PY (Kjn (h2 )(Y )) − (PY (Kjn (h2 )(Y ))) .
2.3.2. Two sides estimation. Suppose that we have two samples X1 , . . . , Xn and Y1 , . . . , Yn with respectively
unknown p.d.f f and g and we want to study the limit behavior of J(fn , gn ).
Theorem 3. Under Assumption 1 and 2, we have
(2.15)
lim sup
n→∞
J(fn , gn ) − J(f, g)
≤ A1 + A2
cn
and
(2.16)
√
n(J(fn , gn ) − J(f, g))
where σ32 = σ12 + σ22 .
N 0, σ32
as n → ∞
σ12 and σ22 are as in 2.13 and 2.14.
The proofs are given in Section 4.
Right now, we are going to apply these results to particular divergence measures estimations. We will have
to check the conditions (2.1),(2.2), and (2.3).
10
DIVERGENCES MEASURES CONSISTENCY BANDS
2.4. Particular cases. Results for Tsallis-α, Renyi-α, Kullback-Leibler and L2 divergences measures will
follow as corollaries since they are particular cases of J(f, g). To ensure the general conditions C1 − φ and
C2 − φ we begin by giving the main assumption on the densities f and g.
Assumption 4. There exists a compact K ⊂ R containing the supports of the densities f and g and such
that
∃(κ1 , κ2 ) ∈ R2 , such that , 0 < κ1 ≤ f (x), g(x) ≤ κ2 < ∞.
Throughout this subsection, we will use the Assumption 4. The integrales are on K and the constantes
are integrables. We use the Dominate convergence theorem based on this remark. Meaning that with
Assumption 4, then the conditions (2.1),(2.2), and (2.3) are satisfied.
In the following the divergence measures, the functions h1 and h2 should be updated in each cases, in the
t
same way that A1 , A2 , σ12 , and σ22 since they depend on the bessov functions f and g in B∞,∞
(R) and on the
randoms variables X and Y .
2.4.1. Case 1 : Hellinger integral of order α. We start by the Hellinger integral of order α defined by
Z
I(f, g) =
f α (x)g 1−α (x)dx
K
α 1−α
Here φ(s, t) = s t
and one has
(1)
(1)
φ1 (s, t) = αsα−1 t1−α , φ2 (s, t) = (1 − α)sα t−α
and
(2)
(2)
(2)
(2)
φ1 (s, t) = α(α − 1)sα−2 t1−α , φ2 (s, t) = −α(1 − α)sα t−α−1 , φ1,2 (s, t) = φ2,1 (s, t) = α(1 − α)sα−1 t−α .
Now let h1 (x) = αf α−1 (x)g 1−α (x) and h2 (x) = (1 − α)f α (x)g −α (x)
R
R
A1 = h1 (x)dx < ∞ and A2 = |h2 (x)| dx < ∞.
Corollary 1. (One sample estimation). We have
• Consistency
lim sup
n−→∞
lim sup
n−→∞
• Asymptotic normality
|I(fn , g) − I(f, g)|
≤ A1 , a.s
an
|I(f, gn ) − I(f, g)|
≤ A2 , a.s
bn
√
n(I(fn , g) − I(f, g))
√
2
N (0, σI,1
) as n → ∞
2
N (0, σI,2
) as n → ∞
n(I(f, gn ) − I(f, g))
where
2
σI,1
= PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X)))2
and
2
2
2
σI,2
= PY (Kjn (h2 )(Y )) − (PY (Kjn (h2 )(Y ))) .
DIVERGENCES MEASURES CONSISTENCY BANDS
11
whith h1 (x) = αf α−1 (x)g 1−α (x) and h2 (x) = (1 − α)f α (x)g −α (x).
Corollary 2. (Two side estimation) We have
• Consistency
lim sup
n−→∞
• Asymptotic normality
|I(fn , gn ) − I(f, g)|
≤ A1 + A2 , a.s
cn
√
n(I(fn , gn ) − I(f, g))
N (0, σ32 ) as n → ∞
where σ32 = σ12 + σ22 .
In the following, handling I the Hellinger integral of order α, conditions (2.1),(2.2), and (2.3) are satisfied
from Assumption 4
2.4.2. Case 2 : Tsallis Divergence measure.
DT,α (f, g) =
1
(I(f, g) − 1)
α−1
Corollary 3. (One side estimation) We have
• Consistency
lim sup
n−→∞
lim sup
n−→∞
|DT,α (fn , g) − DT,α (f, g)|
1
≤
A1 , a.s
an
|α − 1|
1
|DT,α (f, gn ) − DT,α (f, g)|
≤
A2 , a.s
bn
|α − 1|
• Asymptotic normality
√
n(DT,α (fn , g) − DT,α (f, g))
√
n(DT,α (f, gn ) − DT,α (f, g))
where
2
σT,1
=
and
2
N 0, σT,1
2
N 0, σT,2
as n → ∞
as n → ∞
1
PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X)))2
2
(α − 1)
1
2
2
.
(h
)(Y
)))
(h
)(Y
))
−
(P
(K
P
(K
2
2
Y
j
Y
j
n
n
(α − 1)2
whith h1 (x) = αf α−1 (x)g 1−α (x) and h2 = (1 − α)f α (x)g −α (x).
2
σT,2
=
Corollary 4. (Two sides estimation)
Under conditions of theorem , we have
• Consistency
lim sup
n−→∞
• Asymptotic normality
2
2
where σT2 = σT,1
+ σT,2
.
|DT,α (fn , gn ) − DT,α (f, g)|
≤ A1 + A2 , a.s
cn
√
n(DT,α (fn , gn ) − DT,α (f, g))
N 0, σT2
12
DIVERGENCES MEASURES CONSISTENCY BANDS
2.4.3. Case 3 : Reyni Divergence measure.
1
log I(f, g)
α−1
DR,α (f, g) =
Corollary 5. (One side estimation) We have
• Consistency
(2.17)
DR,α (fn , g) − DR,α (f, g) = Oa.s (an )
(2.18)
DR,α (f, gn ) − DR,α (f, g) = Oa.s (bn )
where an and bn are as in (2.4).
• Asymptotic normality
√
(2.19)
n(DR,α (fn , g) − DR,α (f, g))
2
N 0, σR,1
√
n(DR,α (f, gn ) − DR,α (f, g))
(2.20)
2
where σR,1
=
2
σI,1
2
(α−1) I(f,g)
2
and σR,2
=
2
σI,2
2
2
N 0, σR,2
(α−1) I(f,g)
.
as n → ∞
as n → ∞
.
Corollary 6. (Two sides estimation) We have
• Consistency
DR,α (fn , gn ) − DR,α (f, g) = Oa.s (cn )
(2.21)
where cn is as in (2.4)
• Asymptotic normality
√
n(DR,α (fn , gn ) − DR,α (f, g))
(2.22)
2
N 0, σR
2
2
2
where σR
= σR,1
+ σR,2
.
as n → ∞
The proofs of Corollaries 5 and 6 are postponed to Section4.
2.4.4. Case 4 : Kulback-Leib Divergence measure.
DKL (f, g) =
In this case φ(s, t) = s log st and one has
Z
K
f (x) log
f (x)
dx
g(x)
s (1)
s
(1)
φ1 (s, t) = 1 + log , φ2 (s, t) = −
t
t
and
1
s
1 (2)
(2)
(2)
, φ2 (s, t) = 2 , φ1,2 (s, t) = φ2,1 (s, t) = − .
s
t
t
R
R
(x)
(x)
(x)
Thus h1 (x) = 1 + log fg(x)
dx < ∞, and A2 = K
, h2 (x) = fg(x)
, A1 = K 1 + log fg(x)
(2)
φ1 (s, t) =
f (x)
g(x)
dx < ∞.
With the Assumption 4, the conditions (2.1),(2.2), and (2.3) are satisfied for any measurables sequences
(1)
(2)
(1)
(2)
of functions δn (x), δn (x), ρn (x), and ρn (x) of x ∈ D, uniformly converging to zero.
DIVERGENCES MEASURES CONSISTENCY BANDS
13
Corollary 7. (One side estimation) We have
• Consistency
lim sup
n−→∞
lim sup
n−→∞
|DKL (fn , g) − DKL (f, g)|
≤ A1 , a.s
an
|DKL (f, gn ) − DKL (f, g)|
≤ A2 , a.s
bn
• Asymptotic normality
√
n(DKL (fn , g) − DKL (f, g))
√
n(DKL (f, gn ) − DKL (f, g))
2
N 0, σK,1
2
N 0, σK,2
where
2
σK,1
=
and
2
σK,2
=
as n → ∞
as n → ∞
1
PX (Kjn (h1 )(X))2 − PX (Kjn (h1 )(X)))2
(α − 1)2
1
2
2
PY (Kjn (h2 )(Y )) − (PY (Kjn (h2 )(Y ))) .
2
(α − 1)
(x)
and h2 (x) =
with h1 (x) = 1 + log fg(x)
f (x)
g(x)
Corollary 8. (Two sides estimation) We have
• Consistency
lim sup
n−→∞
|DKL (fn , gn ) − DKL (f, g)|
≤ A1 + A2 , a.s
cn
• Asymptotic normality
√
n(DKL (fn , gn ) − DKL (f, g))
2
N 0, σK
2
2
2
where σK
= σK,1
+ σK,2
as n → ∞
2.4.5. Case 5 : L2 Divergence measure.
DL2 (f, g) =
Z
K
2
(f (x) − g(x)) dx
2
Here φ(s, t) = |f (x) − g(x)| , but we proceed by a different route.
One has
DL2 (fn , g) − DL2 (f, g) =
=
=
Z
2
(fn (x) − g(x))2 − (f (x) − g(x)) dx
ZK
(fn (x) − f (x)) (fn (x) + f (x) − 2g(x)) dx
K
Z
Z
(fn (x) − f (x))2 dx
2 (fn (x) − f (x)) (f (x) − g(x)) dx +
K
K
and also
DL2 (f, gn ) − DL2 (f, g) = −2
Z
K
(gn (x) − g(x)) (f (x) − g(x)) dx +
Z
K
(gn (x) − g(x))2 dx
14
DIVERGENCES MEASURES CONSISTENCY BANDS
Let h1 (x) = 2(f (x) − g(x)) and h2 = −h1 . Then we deduce
Z
√
√
n (DL2 (fn , g) − DL2 (f, g)) =
n (fn (x) − f (x))h1 (x)dx + oP (1)
Z
√
√
n (DL2 (f, gn ) − DL2 (f, g)) =
n (gn (x) − g(x))h2 (x)dx + oP (1)
Let A1 = A2 = 2
R
K
|f (x) − g(x)| dx. Then we give
Theorem 4. (One side estimation )
• Consistency
lim sup
n−→∞
lim sup
n−→∞
DL2 (fn , g) − DL2 (f, g)
≤ A1 , a.s
an
DL2 (f, gn ) − DL2 (f, g)
≤ A2 , a.s
bn
• Asymptotic normality
√
n(DL2 (fn , g) − DL2 (f, g))
√
n(DL2 (f, gn ) − DL2 (f, g))
2
N 0, σL
2 ,1
2
N 0, σL
2 ,2
where
as n → ∞
as n → ∞
2
σL
= PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X))2
2 ,1
and
2
σL
= PY (Kjn (h2 )(Y ))2 − (PY (Kjn (h2 )(Y ))2
2 ,2
with h1 (x) = 2(f (x) − g(x)) and h2 (x) = 2(g(x) − f (x)).
Theorem 5. (Two sides estimation)
• Consistency
lim sup
n→∞
• Normality
where
2
σL
2
DL2 (fn , gn ) − DL2 (f, g))
≤ A1 + A2 a.s
cn
√
n(DL2 (fn , gn ) − DL2 (f, g))
=
2
σL
2 ,1
+
2
N 0, σL
2
2
σL
2 ,2
a.s n → ∞
3. Applications
3.1. Statistics tests. The divergence measures may be applied to two statistical problems among others.
First, it may be used as a goodness-of-fit problem like that : let X1 , X2 , .... a sample from X with an unkown
probability density function f and we want to test the hypothesis that f is equal to a known and fixed
probability density function g . We want to test
H0 : f = g
versus H1 : f 6= g,
t
both unctions f and g in Besov space B∞,∞
(R).
For a fixed x ∈ D, we can test the (pointwise) null hypothesis
(3.1)
H0 : f (x) = g(x)
versus H1 : f (x) 6= g(x)
DIVERGENCES MEASURES CONSISTENCY BANDS
15
using particular divergences measure like α−Tsallis, α−Renyi, KB, or L2 divergences.
Then our proposed test statistics are of the form
Z
Dφ (fn , g) = φ(fn (x), g(x))dx
As particular cases we consider
φ1 (s, t)
φ2 (s, t)
φ3 (s, t)
= sα t1−α
s
= s log
t
= |s − t|2
3.1.1. Limit distribution under null hypothesi H0 . In testing the null hypothesis, we propose tests statistics
using Tsallis, Renyi, Kulback and L2 divergence measures. Suppose that the null hypothesis H0 holds so that
g is a known p.d.f.
Then it follows, from the previous work that
√
n(DT,α (fn , g) − DT,α (f, g))
(3.2)
2
N 0, σT,1
where
as n → ∞
1
PX (Kjn (hT )(X))2 − (PX (Kjn (hT )(X)))2
2
(α − 1)
R
α−1
1−α
with hT (x) = αf
(x)g
(x) and Kjn (hT )(X) = Kjn (x, t)h1 (t)f (t)dt
2
σT,1
=
Renya divergence measure
2
where σR,1
=
2
σI,1
.
(α−1)2 I(f,g)
√
n(DR,α (fn , g) − DR,α (f, g))
2
N 0, σR,1
as n → ∞
2
σI,1
= PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X)))2
whith hR (x) = αf α−1 (x)g 1−α (x)
Kulback-Leib :
√
n(DKL (fn , g) − DKL (f, g))
2
N 0, σK,1
where
2
σK,1
=
(x)
where hK (x) = 1 + log fg(x)
as n → ∞
1
PX (Kjn (hK )(X))2 − PX (Kjn (hK )(X)))2
2
(α − 1)
3.2. Confidence bands. We want to obtain
4. PROOFS
The rest of this section, proceeds as follows. In Subsection4.1 we establish the proof of the Theorem 1.
Subsection4.2 is devoted to the proof of the Theorem 2. In Subsection 4.3 we present the proof of the theorem
3 . The Subsection 4.4 is devoted to proofs of the corollaries 5 and 6
16
DIVERGENCES MEASURES CONSISTENCY BANDS
4.1. Proof of the Theorem 1.
t
Proof. Suppose assumptions 1 and 2 are satisfied and h ∈ B∞,∞
(R) .
√ R
We start by showing first that n (fn (x) − f (x))h(x)dx is a sum of an empirical process based on the
sample Xi and applied on the function Kjn (h) and a random variable.
R
We have, by definition Kjn (h)(.) = Kjn (x, (.))h(x)dx. Write
!
Z
Z
n
1X
Kj (x, Xi )h(x) − f (x)h(x) dx
(fn (x) − f (x))h(x)dx =
n i=1 n
Z
n Z
1X
=
Kjn (x, Xi )h(x)dx − f (x)h(x)dx
n i=1
Z
n
1X
=
Kjn (h)(Xi ) − f (x)h(x)dx
n i=1
=
Therefore
where
√
n
Z
Pn,X (Kjn (h)) − PX h(X) = (Pn,X − PX )(Kjn (h)) + PX ((Kjn (h))(X) − h(X)).
(fn (x) − f (x))h(x)dx =
√
√
nR1,n = n(PX ((Kjn (h))(X) − h(X)).
√
√
n(Pn,X − PX )(Kjn (h)) + nR1,n
One has
2
PX (Kjn (h)(X))
=
=
Z
(Kjn (h)(x))2 f (x)dx
Z Z
= 2
2jn
Kjn (x, t)h(t)dt
Z Z
jn
2
jn
f (x)dx
K(2 x, 2 t)h(t)dt
2
f (x)dx.
P
Boundedness and support compactness of ϕ and ψ give K(2jn x, 2jn t) = k ϕ(2jn x − k)ψ(2jn t − k) ≤ C1 .
2
R
Now
K(2jn x, 2jn t)h(t)dt ≤ C12 C2 since ϕ vanishes on (B1 , B2 ]c and h is bounded. Finally
PX (Kjn (h)(X))2 ≤ κ2 22jn C12 C2 with C1 , C2 > 0.
Now the usual C.L.T gives
(4.1)
√
n(Pn,X − PX )(Kjn (h))
where σh2 = PX (Kjn (h)(X))2 − (PX (Kjn (h)(X))2 .
Then the theorem will be proved if we show that
√
N (0, σh2 ) as n → ∞
nR1,n = oP (1) and it is in this step that we use the
s
fact that h ∈ B∞,∞
(R).
From Theorem 9.3 in [37] one has
|PX (Kjn (h) − h)| ≤
Z
|Kjn (h)(x) − h(x)| f (x)dx
≤ C3 kKjn (h) − hk∞ kf k∞
≤ κ2 C3 2−jn t .
DIVERGENCES MEASURES CONSISTENCY BANDS
Therefore
17
√
√
nR1,n (h) ≤ κ2 C3 n2−jn t = κ2 C3 n(1−2t)/8 = oP (1),
for any 1/2 < t < T .
Note the moment condition in theorem quoted above is equivallent to Assumption 1 (S ) (see [37] page
85). This justify its use in our context.
Finally we conclude by
√
Z
n
where σh2 is defined above.
N (0, σh2 ) as n → ∞
(fn (x) − f (x))h(x)dx
4.2. Proof of the Theorem 2.
Proof. In the following development we are going to use systematically the Mean Value Theorem (M.V.T) in
a bivariate dimensional and with real functions θi (i = 1, 2, . . . , 6) depending on x ∈ K but always satisfying
|θi (x)| ≤ 1.
For ease of notation, we introduce the two following notations used in the sequel
∆n f (x) = fn (x) − f (x)
and ∆n g(x) = gn (x) − g(x)
such that
an = k∆n f k∞
and bn = k∆n gk∞ .
Let cn = max(an , bn ). Recall an , bn and cn are all oP (1).
We start by the one side asymptotic estimation.
One has
φ(fn (x), g(x)) = φ(f (x) + ∆n f (x), g(x)).
By an application of the M.V.T to the function u1 (x) 7→ φ(u1 (x), g(x)) one has that there exists θ1 (x) ∈ (0, 1)
such that
(4.2)
φ(fn (x), g(x))
(1)
=
φ(f (x), g(x)) + ∆n f (x)φ1 (f (x) + θ1 (x)∆n f (x), g(x))
where
(1)
(1)
(2)
∆n f (x)φ1 (f (x) + θ1 (x)∆n f (x), g(x)) = ∆n f (x) φ1 (f (x), g(x)) + θ1 (x)(∆n f (x))2 φ1 (f (x) + θ2 (x)∆n f (x), g(x)) ,
(1)
by an application of the M.V.T to the function u2 (x) 7→ φ1 (u2 (x), g(x)) and with θ2 (x) ∈ (0, 1).
We can write (4.2) as
φ(fn (x), g(x))
(2)
(1)
= φ(f (x), g(x)) + ∆n f (x)φ1 (f (x), g(x)) + θ1 (x)(∆n f (x))2 φ1 (f (x) + θ2 (x)∆n f (x), g(x))
Now we has
(4.3)
J(fn , g) − J(f, g) =
hence
Z
(1)
∆n f (x) φ1 (f (x), g(x))dx +
|J(fn , g) − J(f, g)| ≤ an
Z
(1)
Z
(2)
θ1 (x)(∆n f (x))2 φ1 (f (x) + θ2 (x)∆n f (x), g(x)) dx,
φ1 (f (x), g(x)) dx + a2n
Z
(2)
φ1 (f (x) + θ2 (x)∆n f (x), g(x)) dx.
18
DIVERGENCES MEASURES CONSISTENCY BANDS
Therefore
|J(fn , g) − J(f, g)|
≤ A1 + an
lim sup
an
n→∞
where A1 =
R
Z
(2)
φ1 (f (x) + θ2 (x)∆n f (x), g(x)) dx,
(1)
φ1 (f (x), g(x)) dx.
This with 2.1 yield and prove (2.9).
Now let prove 2.10
By swapping the roles of f and g one obtains
(4.4)
J(f, gn ) − J(f, g) =
Z
(1)
∆n g(x) φ2 (f (x), g(x))dx +
then
|J(f, gn ) − J(f, g)| ≤ bn
one obtains
Z
(1)
R
Z
(2)
θ3 (x)(∆n g(x))2 φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx
φ2 (f (x), g(x)) dx + b2n
|J(f, gn ) − J(f, g)|
≤ A2 + bn
bn
where A2 =
Z
Z
(2)
φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx
(2)
φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx
(1)
φ2 (f (x), g(x)) dx. This and (2.2) give and prove (2.10).
We focus now on the asymptotic normality for one sample estimation.
Going back to (4.3), we have
Z
Z
√
√
(1)
(2)
2
n(J(fn , g) − J(f, g)) =
φ1 (f (x), g(x)) αn (x)dx + θ1 (x) n (∆n f (x)) φ1 (f (x) + θ2 (x)∆n f (x), g(x)) dx.
Z
√
√
=
n (fn (x) − f (x))h1 (x)dx + nR2,n
(1)
where h1 (x) = φ1 (f (x), g(x)).
√ R
Now by theorem 1, n (fn (x) − f (x))h1 (x)dx
N (0, σ12 ) as n → ∞ where
σ12 = PX (Kjn (h1 )(X))2 − (PX (Kjn (h1 )(X)))2
(4.5)
and provided that h1 ∈ B∞,∞ (R).
√
Thus, (2.11) will be proved if we show that nR2,n = 0P (1). One has
Z
√
√ 2
(2)
(4.6)
nR2,n ≤ nan φ1 (f (x) + θ2 (x)∆n f (x), g(x))dx
Let show that
√
na2n = oP (1).
By Chebyshev’s inequality, one has for any ǫ > 0
√
√ 2
n1/4
ǫ
P nan > ǫ = P an > 1/4 ≤ √ E a2n .
ǫ
n
DIVERGENCES MEASURES CONSISTENCY BANDS
19
From theorem 3 in Gine [43], one has
1/2
Ea2n
=
=
!
O
r
jn 2jn
+ 2−tjn
n
O
s
1 log n
+ n−t/4
4 log 2 n3/4
!
where we use the fact that 2jn ≈ n1/4 . Thus
Finally
2
√
P na2n > ǫ
=O
√ 2
nan = oP (1) since
s
s
1 log n
+ n(1−2t)/8
4 log 2 n1/2
!
1 log n
+ n(1−2t)/8 → 0 as n → +∞
4 log 2 n1/2
for any 1/2 < t < T.
Finally from (4.6) and using (2.1), one has
This yields and ends the proof of (2.11)
√
nR2,n →P 0 as n → +∞.
Going back to (4.4), one has
Z
Z
√
√
(1)
(2)
n(J(f, gn ) − J(f, g)) =
φ2 (f (x), g(x)) βn (x)dx + n θ4 (x)(∆n g(x))2 φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx.
Z
√
√
=
n (gn (x) − g(x))h2 (x)dx + nR3,n
(1)
where h2 (x) = φ2 (f (x), g(x)).
Then by Theorem 1, one has
√ R
n (gn (x) − g(x))h2 (x)dx
N (0, σ22 ) where
2
2
σ22 = PY (Kjn (h2 )(Y )) − (PY (Kjn (h2 )(Y )))
(4.7)
since P2Y (h2 ) < ∞ and provided that h2 ∈ B∞,∞ (R)).
Similarly,
√
while
nR3,n ≤
√ 2
nbn
Z
(2)
φ2 (f (x), g(x) + θ4 (x)∆n g(x))dx,
√ 2
√
nbn = oP (1) as previously. So this and (2.2) give nR3,n = oP (1).
Finally this shows that (2.12) holds and completes the proof of the Theorem 2.
20
DIVERGENCES MEASURES CONSISTENCY BANDS
4.3. Proof of Theorem 3.
Proof. We proceed by the same techniques that led to the prove of (2.9).
We begin by breaking φ(fn (x), gn (x)) − φ(f (x), g(x)) into two terms we have already handled :
= φ(fn (x), gn (x)) − φ(f (x), gn (x)) + φ(f (x), gn (x)) − φ(f (x), g(x))
{z
} |
{z
}
|
=
I1
+
I2
φ(fn (x), gn (x)) − φ(f (x), g(x))
By an application of the M.V.T to the function fn (x) 7→ φ(fn (x), gn (x)), one has that there exists θ5 (x) ∈
(0, 1) such that
I1
=
φ(f (x) + ∆fn (x), gn (x)) − φ(f (x), gn (x))
=
∆fn (x)φ1 (f (x) + θ5 (x)∆n f (x), g(x))
=
∆fn (x)φ1 (f (x), g(x)) + θ5 (x)(∆n f (x))2 φ1 (f (x) + θ6 (x)∆n f (x), g(x))
(1)
(1)
(2)
By a second application of the M.V.T to the function
(1)
f (x) + θ5 (x)∆n f (x) 7→ φ1 (f (x) + θ5 (x)∆n f (x), g(x))
with θ5 (x) ∈ (0, 1).
From (4.4), we get
(1)
(2)
I2 = ∆n g(x) φ2 (f (x), g(x)) + θ3 (x)(∆n g(x))2 φ2 (f (x), g(x) + θ4 (x)∆n g(x))
Therefore
(4.8)
J(fn , gn ) − J(f, g) =
and
Z
Z
(1)
(1)
∆n f (x)φ1 (f (x), g(x))dx + ∆n g(x)φ2 (f (x), g(x))dx
Z
(2)
+ θ5 (x)(∆n f (x))2 φ1 (f (x) + θ6 (x)∆n f (x), g(x))dx
Z
(2)
+
θ3 (x)(∆n g(x))2 φ2 (f (x), g(x) + θ4 (x)∆n g(x))dx
|J(fn , gn ) − J(f, g)| ≤
thus
|J(fn , gn ) − J(f, g)|
cn
Z
(2)
φ1 (f (x) + θ6 (x)∆n f (x), g(x)) dx
cn A1 + cn A2 + c2n
Z
(2)
φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx
+c2n
Z
(2)
≤ A1 + A2 + cn
φ1 (f (x) + θ6 (x)∆n f (x), g(x)) dx
Z
(2)
+ cn
φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx
(2.1) and (2.2) give
lim sup
x→∞
This proves the desired result.
It remains to prove (2.16)
J(fn , gn ) − J(f, g)
≤ A1 + A2 a.s
cn
DIVERGENCES MEASURES CONSISTENCY BANDS
21
Going back to (4.8) one has
Z
Z
√
√
√
√
(1)
(1)
n(J(fn , gn )−J(f, g)) = n (fn (x)−f (x))φ1 (f (x), g(x))dx+ n (gn (x)−g(x))φ2 (f (x), g(x))dx+ nR4,n
where
√
nR4,n
√
n
Then by Theorem 1, one has
√
n
Z
Z
(2)
θ5 (x)(∆n f (x))2 φ1 (f (x) + θ6 (x)∆n f (x), g(x))dx
Z
√
(2)
θ3 (x)(∆n g(x))2 φ2 (f (x), g(x) + θ4 (x)∆n g(x))dx.
+ n
=
(fn (x) − f (x))h1 (x)dx +
√
n
Z
(gn (x) − g(x))h2 (x)dx
N (0, σ12 + σ22 )
t
since P2X (h1 ) < ∞ and P2Y (h2 ) < ∞. Provided that h2 , h2 ∈ B∞,∞
(R).
Now one has
√
nR4,n
As previously, one has
≤
√ 2
ncn
Z
(2)
φ1 (f (x) + θ6 (x)∆n f (x), g(x)) dx
Z
√ 2
(2)
φ2 (f (x), g(x) + θ4 (x)∆n g(x)) dx.
+ ncn
√ 2
√
ncn = oP (1) and from conditions (2.1) and (2.2) one has | nR4,n | = oP (1).
Finally this shows that (2.16) holds and completes the proof of the Theorem 3.
4.4. Proofs of Corollaries 5 and 6.
Proof. of Corollary 5.
One has
1
(log I(fn , g) − log I(f, g))
α−1
but I(fn , g) − I(f, g) = Oa.s (an ) = oP (1). Then, by using a Taylor expansion of log(1 + y) it follows that
DR,α (fn , g) − DT,α (f, g)) =
almost surely,
I(fn , g) − I(f, g)
log 1 +
I(f, g)
log I(fn , g) − log I(f, g) =
I(fn , g) − I(f, g)
+ Oa.s (a2n ) = Oa.s (an ).
I(f, g)
=
That is
DR,α (fn , g) − DR,α (f, g) = Oa.s (an ).
This proves the desired result.
The proof of (2.18) is similar to the previous proof.
To prove (2.19), recall
√
√
n(I(fn , g) − I(f, g)) = n
then
I(fn , g)
=1+
I(f, g)
R
Z
(fn (x) − f (x))h1 (x)dx + oP (1) = OP (1)
(fn (x) − f (x))h1 (x)dx
+ oP (1)
I(f, g)
22
DIVERGENCES MEASURES CONSISTENCY BANDS
and by Taylor expansion of log(1 + y) it follows that almost surely,
R
(fn (x) − f (x))h1 (x)dx
log I(fn , g) − log I(f, g) = log 1 +
I(f, g)
R
(fn (x) − f (x))h1 (x)dx
1
+ OP
=
I(f, g)
n
therefore
√
n (DR,α (fn , g) − DR,α (f, g)) =
2
where σR,1
=
2
σI,1
.
(α−1)2 I(f,g)
1
α−1
√ R
n (fn (x) − f (x))h1 (x)dx
+ oP (1)
I(f, g)
2
N 0, σR,1
as n → ∞
(2.20) is proved similarly.
Finally this ends the proof of the Corollary 5.
Proof of the Corollary 6
Proof. We start by the consistency. From the previous work one gets
I(fn , gn ) − I(f, g)
+ Oa.s (c2n )
I(f, g)
= Oa.s (cn )
log I(fn , gn ) − log I(f, g) =
hence (DR,α (fn , gn ) − DR,α (f, g)) = Oa.s (cn ). That proves (2.21)
Let find the asymptotic normality. One gets
Z
Z
√
√
√
n (I(fn , gn ) − I(f, g)) =
n (fn (x) − f (x))h1 (x)dx + n (gn (x) − g(x))h2 (x)dx + oP (1) = Nn
=
OP (1)
where h1 (x) = αf α−1 (x)g 1−α (x) and h2 (x) = (1 − α)f α (x)g −α (x).
Hence we obtain
log I(fn , gn ) − log I(f, g) = √
Nn
+ OP
nI(f, g)
Therefore
√
n (DR,α (fn , gn ) − DR,α (f, g))
=
2
2
2
where σR
= σR,1
+ σR,2
.
1
n
1
Nn
+ oP (1)
α − 1 I(f, g)
2
as n → ∞.
N 0, σR
References
[1] Topsoe F. (2000), Some inequalities for information divergence and related measures of discrimination, IEEE Trans. Inf.
Theory, vol. 46, pp. 1602- 1609.
[2] Evren, A. Some applications of Kullback-Leibler and Jeffreys’divergences in multinomial populations.
[3] Cichocki, A.; Amari, S (2010). Families of Alpha-Beta-and Gamma-Divergences: Flexible and Robust Measures of Similarities. Entropy , 12, 1532-1568.
[4] P. J. Moreno, P. P. Ho, and N. Vasconcelos (2004), ”A Kullback-Leibler divergence based kernel for SVM classification in
multimedia applications”, HP Laboratories, Cambridge, MA, Tech. Rep. HPL-2004-4, 2004.
[5] Jager L. and Jon A. Wellner ”Goodness of fit tests via Phi-divergences”. The Annals of Statistics 2007, vol.35, N◦ 5,2018-2053
DIVERGENCES MEASURES CONSISTENCY BANDS
23
[6] P. Hall (1987), ”On Kullback-Leibler loss and density estimation,” Ann. Statist., vol. 15, no. 4, pp. 1491-1519, Dec. 1987.
[7] P. K. Bhattacharya (1967), Efficient estimation of a shift parameter from grouped data, The Annals of Mathematical
Statistics, vol. 38, no. 6,pp. 1770-1787, Dec., 1967.
[8] A. Berlinet, L. Devroye, and L. Gyorfi, ”Asymptotic normality of L1 error in density estimation,” Statistics, vol. 26, pp.
329-343, 1995.
[9] C. Liu, and H-Y Shum (2003), ”Kullback-Leibler boosting,” Proc. of the 2003 IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, vol. 1, pp. 587-594, June, 2003.
[10] Kullback, S.; Leibler, R. (1951) On information and sufficiency. Ann. Math. Stat. , 22, 79âe“86.
[11] Fukunaga, K.; Hayes, R. (1989) The reduced Parzen classifier. IEEE Trans. Pattern Anal. Mach. Intell. , 11, 423âe“425.
[12] Cardoso, J. Infomax and maximum likelihood for blind source separation. IEEE Signal Process. Lett. 1997, 4, 112âe“114.
[13] Cardoso, J. Blind signal separation: statistical principles. Proc. IEEE 1998, 86, 2009-2025.
[14] Ojala, T.; Pietik ainen, M.; Harwood, D. A comparative study of texture measures with classification based on featured
distributions. Pattern Recogn. 1996, 29, 51-59.
[15] Hastie, T.; Tibshirani, R. Classification by pairwise coupling. Ann. Stat. 1998, 26, 451-471.
[16] Buccigrossi, R.; Simoncelli, E. Image compression via joint statistical characterization in the wavelet domain. IEEE Trans.
Image Process. 1999, 8, 1688âe“1701.
[17] Moreno, P.; Ho, P.; Vasconcelos, N. A Kullback-Leibler divergence based kernel for SVM classification in multimedia
applications. Adv. Neural Inform. Process. Syst. 2004, 16, 1385-1392.
[18] MacKay, D. Information Theory, Inference, and Learning Algorithms; Cambridge University Press: Cambridge, UK, 2003.
[19] Cover, T. and Thomas, J.(1991). Elements of Information Theory. : Wiley, 1991.
[20] G. A. Darbellay and I. Vajda (1999), ”Estimation of the information by an adaptive partitioning of the observation space,”
IEEE Trans. Inf. Theory, vol. 45, no. 4, pp. 1315-1321, May 1999.
[21] Common, P.(1994). Independent component analysis, a new concept?, Signal Process., vol. 36, pp. 387-314, 1994.
[22] Tsitsiklis, J.N.(1993). Decentralized detection. In Advances in Statistical Signal Processing. New York: JAI, 1993, pp.
297-344.
[23] Akshay Krishnamurthy, Kirthevasan Kandasamy Barnaba Poczos, Larry Wasserman. Nonparametric Estimation of Rényi
Divergence and Friends.Journal of Selcuk. University Natural and Applied Science.
[24] Liu, H., Lafferty,J., and Wasserman, L. Exponential concentration inequality for
[25] Shashank Singh and Barnabas Poczos Generalized Exponential Concentration Inequality for Rényi Divergence Estimation
Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213 USA.
[26] Poczos B.,Jeff, S. On the estimation of α−Divergences
[27] Liu, H., Lafferty, J., and Wasserman, L. Exponential concentration inequality for mutual information estimation. In Neural
Information Processing Systems (NIPS), 2012.
[28] Hamza, D., Ngom, P., Deme, E. and Mendy, P.(2007). Kernel-type estimators of divergence measures and its strong uniform
consistency.
[29] XuanLong N.,Martin J.W., and Michael I.J.(2010). Estimating Divergence Functionals and the Likelihood Ratio by Convex
Risk Minimization. IEEE Transastions on information theory, vol.56, no.11, pp.5847 − 561, 2010;
[30] X. Nguyen, M. J. Wainwright, and M. I. Jordan, ”Estimating divergence functionals and the likelihood ratio by convex risk
minimization, IEEE Trans. Inform. Theory, vol. 56, no. 11, pp. 5847-5861, 2010.
[31] Einmahl, U. and Mason, D. M. (2000). An empirical process approach to the uniform consistency of kernel-type function
estimators. J. Theoret. Probab., 13 (1), 1-37.
[32] Einmahl, U. and Mason, D. M. (2005). Uniform in bandwidth consistency of kernel-type function estimators. Ann. Statist.,
33(3), 1380-1403.
[33] Giné, E. and Mason, D. M. (2008) Uniform in bandwidth estimation of integral functionals of the density function. Scand.
J. Statist., 35(4), 739-761
[34] H. Cai, S. R. Kulkarni, and S. Verdu (2002), ”Universal estimation of entropye and divergence via block sorting,” in Proc.
IEEE Int. Symp. Information Theory, Lausanne, Switzerland, Jun-Jul. 2002, p. 433.
24
DIVERGENCES MEASURES CONSISTENCY BANDS
[35] H. Cai, S. R. Kulkarni, and S. Verdu, ”Universal divergence estimation for finite-alphabet sources,” IEEE Trans. Inf. Theory,
submitted for publication
[36] J. Ziv and N. Merhav(1993), ”A measure of relative entropy between individual sequences with application to universal
classification,” IEEE Trans. Inf.Theory, vol. 39, no. 4, pp. 1270-1279, Jul. 1993.
[37] Wolfgang HardIe, Gerard Kerkyacharian,Dominique Picard, and Alexander Tsybakov. Wavelets, Approximation,
and Statistical Applications
[38] A. Hero, B. Ma, and O. Michel (1999), ”Estimation of RÃ c nyi information divergence via pruned minimal spanning trees,”
in IEEE Workshop on Higher Order Statistics, Caesaria, Israel, Jun. 1999.
[39] K. R. Moon and A. O. Hero III, ”Ensemble estimation of multivariate f -divergence, in IEEE International Symposium on
Information Theory, pp. 356-360, 2014.
[40] A. Berlinet, L. Gyorfi, and I. D nes, ”Asymptotic normality of relative entropy in multivariate density estimation,” Publications de l’Institut de Statistique de l’Université c de Paris, vol. 41, pp. 3-27, 1997.
[41] P. J. Bickel and M. Rosenblatt, ”On some global measures of the deviations of density function estimates,” The Annals of
Statistics, pp. 1071-1095, 1973.
[42] Kevin R.M.,AlfredO.H.III. ”Multivariate f −Divergence Estimation with Confidence”
[43] Giné E., Nickl R. (2009) ”Uniform Limit theorem for wavelet density estimators”. The Annals of Probability 2009, Vol. 37,
No. 4, 1605-1646 DOI: 10.1214 08-AOP447
| 10math.ST
|
arXiv:1801.01258v1 [cs.CV] 4 Jan 2018
Deep Learning Reconstruction for 9-View Dual
Energy CT Baggage Scanner
Yoseob Han
Jingu Kang
Jong Chul Ye
KAIST, Daejeon, Korea
Email: [email protected]
GEMSS Medical Co. Seongnam, Korea
Email: [email protected]
KAIST, Daejeon, Korea
Email: [email protected]
Abstract—For homeland and transportation security applications, 2D X-ray explosive detection system (EDS) have been
widely used, but they have limitations in recognizing 3D shape
of the hidden objects. Among various types of 3D computed
tomography (CT) systems to address this issue, this paper is
interested in a stationary CT using fixed X-ray sources and
detectors. However, due to the limited number of projection
views, analytic reconstruction algorithms produce severe streaking artifacts. Inspired by recent success of deep learning approach
for sparse view CT reconstruction, here we propose a novel image
and sinogram domain deep learning architecture for 3D reconstruction from very sparse view measurement. The algorithm
has been tested with the real data from a prototype 9-view dual
energy stationary CT EDS carry-on baggage scanner developed
by GEMSS Medical Systems, Korea, which confirms the superior
reconstruction performance over the existing approaches.
Index terms— Explosive detection system (EDS), sparseview X-ray CT, convolutional neural network (CNN)
I. I NTRODUCTION
In homeland and aviation security applications, there has
been increasing demand for X-ray CT EDS system for carryon baggage screening. A CT-EDS can produce an accurate
3D object structure for segmentation and threat detection,
which is often not possible when a 2D-EDS system captures
projection views in only one or two angular directions. There
are currently two types of CT EDS systems: gantry-based CT
and stationary CT. While gantry-based CT EDS is largely the
same as medical CT, baggage screening should be carried out
continuously, so it is often difficult to continuously screen
carry-on bags because of the possible mechanical overloading
of the gantry system. On the other hand, a stationary CT EDS
system uses fixed X-ray sources and detectors, making the
system suitable for routine carry-on baggage inspection.
For example, Fig. 1 shows source and detector geometry of
the prototype stationary CT-EDS system developed by GEMSS
Medical Systems, Korea. As shown in Fig. 1(a), nine pairs
of X-ray source and dual energy detector in the opposite
direction are distributed at the same angular interval. For
seamless screening without stopping convey belt, each pair
of source and detectors are arranged along the z-direction as
shown in Fig. 1(b) so that different projection view data can
be collected while the carry-on baggages moves continuously
on the conveyor belt. Then, 9-view fan beam projection data
is obtained for each z-slice by rebinning the measurement
data. This type of stationary CT system is suitable for EDS
Fig. 1. X-ray source positions in our prototype 9 view dual energy CT EDS:
(a) x − y direction and (b) θ − z direction, respectively.
applications because it does not require a rotating gantry, but
with only 9 projection views it is difficult to use a conventional
filtered backprojection (FBP) algorithm due to severe streaking
artifacts. Therefore, advanced reconstruction algorithms with
fast reconstruction time are required.
For sparse-view CT EDS, model-based iterative reconstruction (MBIR) with the total variation (TV) penalty have been
extensively investigated [1], [2]. Inspired by the recent success
of deep learning approach for sparse view and limited angle
CT [3], [4], [5], [6] that outperform the classical MBIR
approach, this paper aims at developing a deep learning
approach for real-world sparse view CT EDS. However, neural
network training using the retrospective angular subsampling
as in the existing works [3], [4], [5], [6] is not possible for
our prototype system, since there are no ground-truth data for
the real world sparse view CT EDS. We therefore propose a
novel deep learning approach composed of image domain and
sinogram domain learning that compensate for the imperfect
label data.
II. T HEORY
A. Problem Formulation
Recall that the forward model for sparse view CT EDS
system can be represented by
gΘ = PΘ Rf
(1)
where R denotes the 3D projection operator from an x − y − z
volume image to a s − θ − z domain sinogram data with s, θ
and z denoting the detector, projection angle, and the direction
Fig. 2. Sinogram interpolation flow for the proposed method. The final reconstruction is obtained by applying the FBP for the interpolated sinogram data.
of the conveyor belt travel, respectively. See Fig. 2 for the
coordinate systems. In (1), PΘ denotes the view sampling
operator for the measured angle set Θ, and gΘ refers to the
measured sinogram data. For each projection view data, we
use the notation gθ and Pθ , where θ denotes the specific view.
The main technical issue of the sparse view CT reconstruction is the non-uniqueness of the solution for (1). More
specifically, there exists a null spacce NΘ such that
PΘ Rh = 0,
∀h ∈ NΘ ,
which leads to infinite number of feasible solutions. To avoid
the non-uniqueness of the solution, constrained form of the
penalized MBIR can be formulated as :
min3 kLf k1 ,
f ∈R
subject to gΘ = PΘ R ,
(2)
where L refers to a linear operator and k · k1 denotes the l1
norm. For the case of the TV penalty, L corresponds to the
derivative. Then, the uniqueness of (2) is guaranteed that if
the NL ∩ NΘ = {0}, where NL denotes the null space of the
operator L.
Instead of designing a linear operator L such that the
common null space of NΘ and NL to be zero, we can design
a frame W, its dual W̃, and shrinkage operator Sλ such that
W̃ > W = I and
>
∗
W̃ Sλ W(f + g) = f
∗
∀g ∈ NΘ
∀h ∈ NΘ
.
More specifically, our sparse view reconstruction algorithm
finds the unknown f ∈ R3 that satisfy both data fidelity and
the so-called frame constraints [8]:
gΘ = PΘ Rf,
(3)
In other word, (3) directly removes the null space component.
Eq. (3) is the constraint we use for training our neural network.
QI (f ) = f ∗
,
(4)
where QI is the image domain CNN that satisfies (3) and f ∗
denotes the ground-truth images that are available for training
data. Now, by defining M as a right-inverse of PΘ R, i.e.
(PΘ R)MgΘ = gΘ , ∀gΘ , we have
MgΘ = f ∗ + h
for some h ∈ NΘ , since the right inverse is not unique due to
the existence of the null space. Thus, we can show that MgΘ
is the feasible solution for (4), since we have
QI MgΘ = QI (f ∗ + h) = f ∗
,
(5)
for the training data, and
PΘ RMgΘ = PΘ R(f ∗ + h) = gΘ
.
(6)
Therefore, the neural network training problem to satisfy (4)
can be equivalently represented by
min
QI
for the ground-truth image f ∗ . This frame-based regularization is also an active field of research for image denoising,
inpainting, etc [7]. One of the most important contributions
of the deep convolutional framelet theory [8] is that W
and W̃ > correspond to the encoder and decoder structure
of a convolutional neural network (CNN), respectively, and
the shrinkage operator Sλ emerges by controlling the number of filter channels and nonlinearities. More specifically,
a convolutional neural network can be designed such that
Q = W̃ > Sλ W and
Q(f ∗ + h) = f ∗ ,
B. Derivation of Image and Projection Domain CNNs
N
X
(i)
kf ∗(i) − QI MgΘ k2
(7)
i=1
(i)
where {(f ∗(i) , gΘ )}N
i=1 denotes the training data set composed of ground-truth image an its sparse view projection.
Since a representative right inverse for the sparse view projection is the inverse Radon transform after zero padding to
(i)
the missing view, MgΘ in (7) can be implemented using the
standard FBP algorithm. In fact, this is the main theoretical
ground for the success of image domain CNN when the
ground-truth data is available [3], [4], [5], [6]. Moreover, the
fan-beam rebinning makes the problem separable for each z
slices, so we use the 2D FBP for each slice as shown in Fig. 2.
However, the main technical difficulties in our 9-view CT
EDS system is that we do not have ground-truth image
{f ∗(i) }N
i=1 . One could use physical phantoms and atomic
number to form a set of ground-truth images, but those data
set may be different from the real carry-on bags, so we need a
new method to account for the lack of ground-truth for neural
network training. Thus, to overcome the lack of the groundtruth data, the approximate label images are generated using
an MBIR with TV penalty. Then, using MBIR reconstruction
I
as label data {f ∗(i) }N
i=1 , an 2D image domain network Q is
trained to learn the mapping between the artifact-corrupted 2D
image and MBIR reconstruction in x − y domain.
One downside of this approach is that the network training
by (7) is no more optimal, since the label data is not the
ground-truth image. Thus, the generated sinogram data from
the denoised 3D volume may be biased. Thus, we impose
additional frame constraint to the sinogram data in addition to
(4):
gθ∗ = QS (gθ ) ,
(8)
for the measured angle θ, where QS is the s − z sinogram
domain CNN and gθ∗ denotes the ground-truth sinogram data
measured at θ. Then, Eq. (8) leads to the following network
training:
min
QS
N
XX
∗(i)
kgθ
(i)
− QS Pθ RQI MgΘ k2
(9)
θ∈Θ i=1
More specifically, as shown in Fig. 2, 3D sinogram data is
generated in the s − θ − z domain by applying the forward
projection operator along 720-projection views after stacking
the image domain network output over multiple slices to form
3D reconstruction volume in the x − y − z domain. Next, a 2D
sinogram domain network QS is trained so that it can learn the
mapping between the synthetic s−z sinogram data and the real
projection data in the s − z domain. Since the real projection
data is available only in 9 views, this sinogram network
training is performed using synthetic and real projection data
in the measured projection views. The optimization problems
(7) and (9) can be solved sequentially or simultaneously, and
in this paper we adopt the sequential optimization approach
for simplicity.
After the neural networks QI and QS are trained, the
inference can be done simply by obtaining x − y − z volume
images from the 9 view projection data by slice-by-slice FBP
algorithm, which are then fed into QI to obtain the denoised
3D volume data. Then, by applying projection operator, we
generate 720 projection view data in s − θ − z domain,
which are fed into the QS to obtain denoised sinogram data
for each θ angle. Then, the final reconstruction is obtained
by applying FBP algorithms. One could use post-processing
using additional TV-based denosing. This algorithmic flow is
illustrated in Fig. 2.
III. M ETHODS
A. Real CT EDS data Acquisition
We collected CT EDS data using the prototype stationary
9 view dual energy CT-EDS system developed by GEMSS
Medical Systems, Korea as shown in Fig. 1. the distance from
source to detector (DSD) and the distance from source to
Fig. 3. CNN architecture for our image and singoram domain networks.
object (DSO) are 1202.6mm and 648.2mm, respectively. The
number of detector is 384 with a pitch of 1.5mm. The region
of interest (ROI) is 256 × 256 and the pixel size is 2mm2 .
The detectors collect low and high energy X-ray at 80KVp
and 120KVp, respectively.
We collect 47 sets of projection data from the prototype
CT EDS baggage scanner. Among the 47 sets, 32 dataset are
simple-objects and the other set are realistic carry-on bags.
The 47 set of 28 simple- and 13 baggage-objects was used
during the training phase, and the validation was performed
by two simple- and one baggage-object. The other set was
used for test.
B. Network Architecture and Training
Fig. 3 illustrates modified the U-Net structure [9] for the
image domain and the sinogram domain networks. To account
for the multi-energy image and sinogram data, the input for
the network is two channel multi-energy image and sinogram
data. The proposed network consists of convolution layer,
batch normalization, rectified linear unit (ReLU) [10], and
contracting path connection with concatenation [9]. A detail
parameters are illustrated as shown in Fig. 3.
The proposed networks were trained by stochastic gradient
descent (SGD). The regularization parameter was λ = 10−4 .
The learning rate has been set from 10−3 to 10−5 , which
has been reduced step by step in each epoch. The number
of epoch was 200. The batch size was 12 and the patch
size for image and projection data are 256 × 256 × 2 and
768 × 384 × 2, respectively. The network was implemented
using MatConvNet toolbox (ver.24) [11] in the MATLAB
2015a environment (MathWorks, Natick). Central processing
unit (CPU) and graphic processing unit (GPU) specification
are i7-7700 CPU (3.60 GHz) and GTX 1080 Ti GPU, respectively.
IV. E XPERIMENTAL R ESULTS
To evaluate the performance of the proposed method, we
perform image reconstruction from real 9-view CT EDS prototype system. Fig. 4 illustrates image reconstruction results
of bag using various methods such as FBP, MBIR with
TV penalty, image domain CNN [3], [5], and the proposed
method. The FBP reconstruction results suffered from severe
streaking artifacts, so it was difficult to see the threats in the
tomographic reconstruction and 3D rendering. The MBIR and
image domain CNN were slight better in their reconstruction
quality, but the detailed 3D structures were not fully recovered
and several objects were not detected as indicated by the red
arrow in Fig. 4. Moreover, the 3D rendering results in Fig.
Fig. 5. A s − z domain sinogram data from (a) measurement, (b) 9-views
FBP (c) MBIR, (d) image CNN, and (e) the proposed method. The number
written in the images is the NMSE value. Yellow and red arrows indicate
grenade and knife, respectively.
high quality images. Using real data from our prototype 9-view
CT EDS system, we demonstrated that the proposed method
outperforms the existing algorithms, delivering high quality
three reconstruction for threat detection.
ACKNOWLEDGMENT
Fig. 4. Reconstruction results by various methods from 9-views CT-EDS..
4 correctly identify the shape of grenade and knife as well
as the frame of the bag, which was not possible using other
methods.
Because we do not have the ground-truth in the image
domain, we perform quantitative evaluation using normalized
mean squares error (NMSE) in the sinogram domain. More
specifically, after obtaining the final reconstruction, we perform the forward projection to generate the sinogram data in
the measured projection view and calculated the normalized
mean square errors. Table I showed that the proposed method
provides the most accurate sinogram data compared to the
other methods. Moreover, the s − z projection data in Fig. 5
showed that the projection data from the proposed method is
much closer to the ground-truth measurement data.
TABLE I
NMSE VALUE COMPARISON OF VARIOUS METHODS .
Energy level
FBP
MBIR-TV
Image CNN
Ours
80 kvP
1.6647e+1
5.8247e-1
3.3207e-1
0.6845e-1
120 kvP
1.0536e+1
6.0440e-1
3.2249e-1
0.5450e-1
V. C ONCLUSION
In this paper, we proposed a novel deep learning reconstruction algorithm for a prototype 9-view dual energy CT
EDS for carry-on baggage scanner. Even though the number
of projection view was not sufficient for high equality 3D
reconstruction, our method learns the relationships between
the 2D tomographic slices in x − y domain as well as the 2D
projections in s − z domain such that the artifact-corrupted
image and sinogram data can be successively refined to obtain
This work is supported by Korea Agency for Infrastructure
Technology Advancement, Grant number 17ATRP-C07116405-000000.
R EFERENCES
[1] Sagar Mandava, David Coccarelli, Joel A Greenberg, Michael E Gehm,
Amit Ashok, and Ali Bilgin, “Image reconstruction for view-limited
x-ray ct in baggage scanning,” in Anomaly Detection and Imaging with
X-Rays (ADIX) II. International Society for Optics and Photonics, 2017,
vol. 10187, p. 101870F.
[2] Sherman J Kisner, Eri Haneda, Charles A Bouman, Sondre Skatter,
Mikhail Kourinny, and Simon Bedford, “Limited view angle iterative ct
reconstruction,” in Computational Imaging X, 2012, vol. 8296.
[3] Yoseop Han, Jaejoon Yoo, and Jong Chul Ye, “Deep residual learning for
compressed sensing ct reconstruction via persistent homology analysis,”
arXiv preprint arXiv:1611.06391, 2016.
[4] Yoseob Han and Jong Chul Ye, “Framing U-net via deep convolutional framelets: Application to sparse-view CT,” arXiv preprint
arXiv:1708.08333, 2017.
[5] Kyong Hwan Jin, Michael T McCann, Emmanuel Froustey, and Michael
Unser, “Deep convolutional neural network for inverse problems in
imaging,” IEEE Transactions on Image Processing, vol. 26, no. 9, pp.
4509–4522, 2017.
[6] Jawook Gu and Jong Chul Ye, “Multi-scale wavelet domain residual learning for limited-angle CT reconstruction,” arXiv preprint
arXiv:1703.01382, 2017.
[7] Jian-Feng Cai, Raymond H Chan, and Zuowei Shen, “A framelet-based
image inpainting algorithm,” Applied and Computational Harmonic
Analysis, vol. 24, no. 2, pp. 131–149, 2008.
[8] Jong Chul Ye, Yo Seob Han, and Eunjoo Cha, “Deep convolutional
framelets: A general deep learning framework for inverse problems,”
arXiv preprint arXiv:1707.00372, 2017.
[9] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International
Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241.
[10] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton, “Imagenet
classification with deep convolutional neural networks,” in Advances in
neural information processing systems, 2012, pp. 1097–1105.
[11] Andrea Vedaldi and Karel Lenc, “Matconvnet: Convolutional neural
networks for matlab,” in Proceedings of the 23rd ACM international
conference on Multimedia. ACM, 2015, pp. 689–692.
| 2cs.AI
|
arXiv:1612.05124v2 [math.ST] 7 Jun 2017
Adaptive nonparametric drift estimation for diffusion
processes using Faber-Schauder expansions
Frank van der Meulen*
Moritz Schauer†
Jan van Waaij‡
June 8, 2017
Abstract
We consider the problem of nonparametric estimation of the drift of a continuously observed one-dimensional diffusion with periodic drift. Motivated by computational considerations, van der Meulen et al. (2014) defined a prior on the drift as a randomly truncated
and randomly scaled Faber-Schauder series expansion with Gaussian coefficients. We study
the behaviour of the posterior obtained from this prior from a frequentist asymptotic point
of view. If the true data generating drift is smooth, it is proved that the posterior is adaptive
with posterior contraction rates for the L 2 -norm that are optimal up to a log factor. Contraction rates in L p -norms with p ∈ (2, ∞] are derived as well.
1 Introduction
Assume continuous time observations X T = {X t , : t ∈ [0, T ]} from a diffusion process X defined
as (weak) solution to the stochastic differential equation (sde)
dX t = b 0 (X t ) dt + dWt ,
X 0 = x0 .
(1)
Here W is a Brownian Motion and the drift b 0 is assumed to be a real-valued measurable function
on the real line that is 1-periodic and square integrable on [0, 1]. The assumed periodicity implies
that we can alternatively view the process X as a diffusion on the circle. This model has been
used for dynamic modelling of angles, see for instance Pokern (2007) and Hindriks (2011).
We are interested in nonparametric adaptive estimation of the drift. This problem has recently
been studied by multiple authors. Spokoiny (2000) proposed a locally linear smoother with a
data-driven bandwidth choice that is rate adaptive with respect to |b 00 (x)| for all x and optimal
* TU Delft, Mekelweg 4, 2628 CD Delft, The Netherlands, E-mail address: [email protected].
† Leiden University, Niels Bohrweg 1, 2333 CA Leiden, The Netherlands, E-mail address: m.r.schauer@math.
leidenuniv.nl.
‡ Korteweg-de Vries Institute for Mathematics, Science Park 107, 1098 XG Amsterdam, The Netherlands, E-mail
address: [email protected].
1
up to a log factors. Interestingly, the result is non-asymptotic and does not require ergodicity.
Dalalyan and Kutoyants (2002) and Dalalyan (2005) consider ergodic diffusions and construct
estimators that are asymptotically minimax and adaptive under Sobolev smoothness of the drift.
Their results were extended to the multidimensional case by Strauch (2015).
In this paper we focus on Bayesian nonparametric estimation, a paradigm that has become increasingly popular over the past two decades. An overview of some advances of Bayesian nonparametric estimation for diffusion processes is given in van Zanten (2013).
The Bayesian approach requires the specification of a prior. Ideally, the prior on the drift is chosen such that drawing from the posterior is computationally efficient while at the same time
ensuring that the resulting inference has good theoretical properties. which is quantified by a
contraction rate. This is a rate for which we can shrink balls around the true parameter value,
while maintaining most of the posterior mass. More formally, if d is a semimetric on the space
of drift functions, a contraction rate εT is a sequence of positive numbers εT ↓ 0 for which the
posterior mass of the balls {b : d (b, b 0 ) ≤ εT } converges in probability to 1 as T → ∞, under the
law of X with drift b 0 . For a general discussion on contraction rates, see for instance Ghosal et al.
(2000) and Ghosal and van der Vaart (2007).
For diffusions, the problem of deriving optimal posterior convergence rates has been studied
R1
recently under the additional assumption that the drift integrates to zero, 0 b 0 (x)d x = 0. In
Papaspiliopoulos et al. (2012) a mean zero Gaussian process prior is proposed together with an
algorithm to sample from the posterior. The precision operator (inverse covariance operator)
¡
¢
of the proposed Gaussian process is given by η (−∆)α+1/2 + κI , where ∆ is the one-dimensional
Laplacian, I is the identity operator, η, κ > 0 and α + 1/2 ∈ {2, 3, . . .}. A first consistency result was
shown in Pokern et al. (2013).
In van Waaij and van Zanten (2016) it was shown that this rate result can be improved upon for
a slightly more general class of priors on the drift. More specifically, in this paper the authors
consider a prior which is defined as
b=L
∞
X
k −1/2−α ϕk Zk ,
(2)
k=1
p
p
where ϕ2k (x) = 2 cos(2πkx), ϕ2k−1 (x) = 2 sin(2πkx) are the standard Fourier series basis
functions, {Zk } is a sequence of independent standard normally distributed random variables
and α is positive. It is shown that when L and α are fixed and b 0 is assumed to be α-Sobolev
smooth, then the optimal posterior rate of contraction, T −α/(1+2α) , is obtained. Note that this
result is nonadaptive, as the regularity of the prior must match the regularity of b 0 . For obtaining
optimal posterior contraction rates for the full range of possible regularities of the drift, two options are investigated: endowing either L or α with a hyperprior. Only the second option results
in the desired adaptivity over all possible regularities.
While the prior in (2) (with additional prior on α) has good asymptotic properties, from a computational point of view the infinite series expansion is inconvenient. Clearly, in any implementation this expansion needs to be truncated. Random truncation of a series expansion is a well
known method for defining priors in Bayesian nonparametrics, see for instance Shen and Ghosal
(2015). Exactly this idea was exploited in van der Meulen et al. (2014), where the prior is defined
as the law of the random function
2
Figure 1: Elements ψ1 and ψ j ,k , 0 ≤ j ≤ 2 of the Faber-Schauder basis
b
R,S
= S Z 1 ψ1 + S
R X
2j
X
Zjkψjk,
(3)
j =0 k=1
where the functions ψ j k constitute the Faber-Schauder basis (see fig. 1). These functions feature
prominently in the Lévy-Ciesielski construction of Brownian motion (see for instance (Bhattacharya and Waymire, 2007, paragraph 10.1)). The prior coefficients Z j k are equipped with a
Gaussian distribution, and the truncation level R and the scaling factor S are equipped with independent priors. Truncation in absence of scaling increases the apparent smoothness of the
prior (as illustrated for deterministic truncation by example 4.5 in van der Vaart and van Zanten
(2008)), whereas scaling by a number ≥ 1 decreases the apparent smoothness. (Scaling with a
number ≤ 1 only increases the apparent smoothness to a limited extent, see for example Knapik
et al. (2011).)
The simplest type of prior is obtained by taking the coefficients Z j k independent. We do however
also consider the prior that is obtained by first expanding a periodic Ornstein-Uhlenbeck process
into the Faber-Schauder basis, followed by random scaling and truncation. We will explain that
specific stationarity properties of this prior make it a natural choice.
Draws from the posterior can be computed using a reversible jump Markov Chain Monte Carlo
(MCMC) algorithm (cf. van der Meulen et al. (2014)). For both types of priors, fast computation is
facilitated by leveraging inherent sparsity properties stemming from the compact support of the
functions ψ j k . In the discussion of van der Meulen et al. (2014) it was argued that inclusion of
both the scaling and random truncation in the prior is beneficial. However, this claim was only
supported by simulations results.
In this paper we support this claim theoretically by proving adaptive contraction rates of the posterior distribution in case the prior (3) is used. We start from a general result in van der Meulen
et al. (2006) on Brownian semimartingale models, which we adapt to our setting. Here we take
into account that as the drift is assumed to be one-periodic, information accumulates in a different way compared to (general) ergodic diffusions. Subsequently we verify that the resulting
prior mass, remaining mass and entropy conditions appearing in this adapted result are satis-
3
fied for the prior defined in equation (3). An application of our results shows that if the true drift
β
function is B ∞,∞ -Besov smooth, β ∈ (0, 2), then by appropriate choice of the variances of Z j k , as
well as the priors on R and S, the posterior for the drift b contracts at the rate (T / log T )−β/(1+2β)
around the true drift in the L 2 -norm. Up to the log factor this rate is minimax-optimal (See for instance (Kutoyants, 2004, Theorem 4.48)). Moreover, it is adaptive: the prior does not depend on
β. In case the true drift has Besov-smoothness greater than or equal to 2, our method guarantees
contraction rates equal to essentially T −2/5 (corresponding to β = 2). A further application of our
results shows that for L p -norms we obtain contraction rate T −(β−1/2+1/p)/(1+2β) , up to log-factors.
The paper is organised as follows. In the next section we give a precise definition of the prior.
In section 3 a general contraction result for the class of diffusion processes considered here is
derived. Our main result on posterior contraction for L p -norms with p ≥ 2 is presented in section 4. Many results of this paper concern general properties of the prior and their application
is not confined to drift estimation of diffusion processes. To illustrate this, we show in section 5
how these results can easily be adapted to nonparametric regression and nonparametric density
estimation. Proofs are gathered in section 6. The appendix contains a couple of technical results.
2 Prior construction
2.1 Model and posterior
Let
¾
½
¯Z 1
¯
b(x)2 dx < ∞ and b is 1-periodic
L 2 (T) = b : R → R ¯
0
be the space of square integrable 1-periodic functions.
Lemma 1. If b 0 ∈ L 2 (T), then the SDE eq. (1) has a unique weak solution.
The proof is in section 6.1.
For b ∈ L 2 (T), let P b = P b,T denote the law of the process X T generated by eq. (1) when b 0 is
replaced by b. If P 0 denotes the law of X T when the drift is zero, then P b is absolutely continuous
with respect to P 0 with Radon-Nikodym density
µZ T
¶
Z
¡ T¢
1 T 2
p b X = exp
b(X t ) dX t −
b (X t ) dt .
(4)
2 0
0
Given a prior Π on L 2 (T) and path X T from (1), the posterior is given by
R
p b (X T ) Π( db)
T
Π(b ∈ A | X ) = RA
,
p b (X T ) Π( db)
(5)
where A is Borel set of L 2 (T). These assertions are verified as part of the proof of theorem 3.
2.2 Motivating the choice of prior
We are interested in randomly truncated, scaled series priors that simultaneously enable a fast
algorithm for obtaining draws from the posterior and enjoy good contraction rates.
4
To explain what we mean by the first item, consider first a prior that is a finite series prior. Let
{ψ1 , . . . , ψr } denote basis functions and Z = (Z1 , . . . , Zr ) a mean zero Gaussian random vector with
P
precision matrix Γ. Assume that the prior for b is given by b = ri=1 Zi ψi . By conjugacy, it follows
that Z | X T ∼ N(W−1 µ, W−1 ), where W = G + Γ,
µi =
T
Z
0
ψi (X t ) dX t
T
Z
and G
i ,i 0
=
0
ψi (X t )ψi 0 (X t ) dt
(6)
for i , i 0 ∈ {1, . . . , r }, cf. (van der Meulen et al., 2014, Lemma 1). The matrix G is referred to as the
Grammian. From these expressions it follows that it is computationally advantageous to exploit
compactly supported basis functions. Whenever ψi and ψi 0 have nonoverlapping supports, we
have G i ,i 0 = 0. Depending on the choice of such basis functions, the Grammian G will have a
specific sparsity structure (a set of index pairs (i , i 0 ) such that G i ,i 0 = 0, independently of X T .)
This sparsity structure is inherited by W as long as the sparsity structure of the prior precision
matrix matches that of G.
In the next section we make a specific choice for the basis functions and the prior precision
matrix Γ.
2.3 Definition of the prior
Define the “hat” function Λ by Λ(x) = (2x)1[0, 1 ) (x) + 2(1 − x)1[ 1 ,1] (x). The Faber-Schauder basis
2
2
functions are given by
ψ j ,k (x) = Λ(2 j x − k + 1),
j ≥ 0,
k = 1, . . . , 2 j
Let
¡
¢
ψ1 = ψ0,1 (x − 21 ) + ψ0,1 (x + 12 ) I[0,1] (x).
In figure 1 we have plotted ψ1 together with ψ j ,k where j ∈ {0, 1, 2}.
We define our prior as in (3) with Gaussian coefficients Z1 and Z j k , where the truncation level R
and the scaling factor S are equipped with (hyper)priors. We extend b periodically if we want to
consider b as function on the real line. If we identify the double index ( j , k) in (3) with the single
P R+1
index i = 2 j + k, then we can write b R,S = S 2i =1 ψi Zi . Let
`(i ) =
(
0
j
if i ∈ {1, 2}
.
©
ª
if i ∈ 2 j + 1, . . . , 2 j +1 and j ≥ 1
We say that ψi belongs to level j ≥ 0 if `(i ) = j . Thus both ψ1 and ψ0,1 belong to level 0, which is
convenient for notational purposes. For levels j ≥ 1 the basis functions are per level orthogonal
with essentially disjoint support. Define for r ∈ {0, 1, . . .}
©
ª
Ir = i : `(i ) ≤ r } = {1, 2, . . . , 2r +1 .
Let A = (Cov(Zi , Zi 0 ))i ,i 0 ∈N and define its finite-dimensional restriction by A r = (A i i 0 )i ,i 0 ∈Ir . If we
denote Z r = {Zi , i ∈ Ir }, and assume that Z r is multivariate normally distributed with mean zero
5
and covariance matrix A r , then the prior has the following hierarchy
b | R, S, Z R = S
X
Z i ψi
(7)
i ∈IR
Z R | R ∼ N(0, AR )
(8)
(R, S) ∼ Π(·).
(9)
Here, we use Π to denote the joint distribution of (R, S).
We will consider two choices of priors for the sequence Z1 , Z2 , . . . Our first choice consists of taking independent Gaussian random variables. If the coefficients Zi are independent with standard deviation 2−`(i )/2 , the random draws from this prior are scaled piecewise linear interpolations on a dyadic grid of a Brownian bridge on [0, 1] plus the random function Z1 ψ1 . The choice
¡
¢
of ψ1 is motivated by the fact that in this case Var b(t )|S = s, R = ∞ = s 2 is independent of t .
We construct this second type of prior as follows. For γ, σ2 > 0, define V ≡ (Vt , t ∈ [0, 1]) to be
the cyclically stationary and centred Ornstein-Uhlenbeck process. This is a periodic Gaussian
process with covariance kernel
Cov(V (s),V (t )) =
σ2 e −γh + e −γ(1−h)
, h = t − s ≥ 0.
2γ
1 − e −γ
(10)
This process is cyclically stationary, that is, the covariance only depends on |t − s| and 1 − |t − s|.
It is the unique Gaussian and Markovian prior with continuous periodic paths with this property. This makes the cyclically stationary Ornstein-Uhlenbeck prior an appealing choice which
respects the symmetries of the problem.
Each realisation of V is continuous and can be extended to a periodic function on R. Then V can
be represented as an infinite series expansion in the Faber-Schauder basis:
Vt =
X
Zi ψi (t ) = Z1 ψ1 (t ) +
2j
∞ X
X
Z j ,k ψ j ,k (t )
(11)
j =0 k=1
i ≥1
Finally by scaling by S and truncating at R we obtain from V the second choice of prior on the
drift function b. Visualisations of the covariance kernels Cov(b(s), b(t )) for first prior (Brownian
bridge type) and for the second prior (periodic Ornstein-Uhlenbeck process prior with parameter γ = 1.48) are shown in fig. 2 (for S = 1 and R = ∞).
2.4 Sparsity structure induced by choice of Zi
Conditional on R and S, the posterior of Z R is Gaussian with precision matrix G R + ΓR (here G R
is the Grammian corresponding to using all basis functions up to and including level R).
If the coefficients are independent it is trivial to see that the precision matrix Γ does not destroy
the sparsity structure of G, as defined in (6). This is convenient for numerical computations. The
next lemma details the situation for periodic Ornstein-Uhlenbeck processes.
Lemma 2. Let V be defined as in equation (10)
6
Figure 2: Heat maps of (s, t ) 7→ Cov(b(s), b(t )), in case S = 1 and R = ∞. Left: Brownian bridge
plus the random function Z1 ψ1 . Right: periodic Ornstein-Uhlenbeck process with parameter
γ = 1.48 and σ2 chosen such that Var(b(s)) = 1.
1. The sparsity structure of the precision matrix of the infinite stochastic vector Z (appearing
in the series representation (11)) equals the sparsity structure of G, as defined in (6).
2. The entries of the covariance matrix of the random Gaussian coefficients Zi and Zi 0 , A i ,i 0 =
2
E Zi Zi 0 , satisfy the following bounds: A 11 = A 22 = σ
2γ coth(γ/2) and for γ ≤ 1.5 and i ≥ 3,
0.95 · 2−`(i ) σ2 /4 ≤ A i i ≤ 2−`(i ) σ2 /4
and A 12 = A 21 =
σ2
2γ
sinh−1 (γ/2) and for i 6= i 0
|A i i 0 | ≤
(
0
0.20σ2 2−1.5(`(i )∨`(i ))
0.37σ2 2−1.5(`(i )+`(i
0
))
i ∧ i 0 ≤ 2 < i ∨ i 0,
otherwise.
The proof is given in section 6.2. By the first part of the lemma, also this prior does not destroy
the sparsity structure of the G. The second part asserts that while the off-diagonal entries of A r
are not zero, they are of smaller order than the diagonal entries, quantifying that the covariance
matrix of the coefficients in the Schauder expansion is close to a diagonal matrix.
3 Posterior contraction for diffusion processes
The main result in van der Meulen et al. (2006) gives sufficient conditions for deriving posterior
contraction rates in Brownian semimartingale models. The following theorem is an adaptation
and refinement of Theorem 2.1 and Lemma 2.2 of van der Meulen et al. (2006) for diffusions defined on the circle. We assume observations X T , where T → ∞. Let ΠT be a prior on L 2 (T) (which
henceforth may depend on T ) and choose measurable subsets (sieves) BT ⊂ L 2 (T). Define the
balls
B T (b 0 , ε) = {b ∈ BT : kb 0 − bk2 < ε}.
7
The ε-covering number of a set A for a semimetric ρ, denoted by N (ε, A, ρ), is defined as the
minimal number of ρ-balls of radius ε needed to cover the set A. The logarithm of the covering
number is referred to as the entropy.
The following theorem characterises the rate of posterior contraction for diffusions on the circle
in terms of properties of the prior.
Theorem 3. Suppose {εT } is a sequence of positive numbers such that T ε2T is bounded away from
zero. Assume that there is a constant ξ > 0 such that for every K > 0 there is a measurable set
BT ⊆ L 2 (T) and for every a > 0 there is a constant C > 0 such that for T big enough
¡
¢
log N aεT , B T (b 0 , εT ), k · k2 ≤ C T ε2T ,
(12)
¡
¢
2
ΠT B T (b 0 , εT ) ≥ e −ξT εT ,
(13)
and
¡
¢
2
ΠT L 2 (T) \ BT ≤ e −K T εT .
(14)
Then for every M T → ∞
¡
¢
P b0 ΠT b ∈ L 2 (T) : kb − b 0 k2 ≥ M T εT | X T → 0
and for K big enough,
¡
¢
ΠT L 2 (T) \ BT | X T → 0.
(15)
Equations (12), (13) and (14) are referred to as the entropy condition, small ball condition and
remaining mass condition of theorem 3 respectively. The proof of this theorem is in section 6.3.
4 Theorems on posterior contraction rates
The main result of this section, theorem 9 characterises the frequentist rate of contraction of the
posterior probability around a fixed parameter b 0 of unknown smoothness using the truncated
series prior from section 2.
We make the following assumption on the true drift function.
Assumption 4. The true drift b 0 can be expanded in the Faber-Schauder basis, b 0 = z 1 ψ1 +
P∞ P2 j
P
j =0 k=1 z j k ψ j k = i ≥1 z i ψi and there exists a β ∈ (0, ∞) such that
Jb0 Kβ := sup 2β`(i ) |z i | < ∞.
(16)
i ≥1
Note that we use a slightly different symbol for the norm, as we denote the L 2 -norm by k · k2 .
β
Remark 5. If β ∈ (0, 2), then assumption 4 on b 0 is equivalent to assuming b 0 to be B ∞,∞ -Besov
smooth. It follows from the definition of the basis functions that
´ 1 ³
´
³
´ 1 ³
z j k = b 0 (2k − 1)2−( j +2) − b 0 2−( j +2) (2k − 2) − b 0 2−( j −2) 2k .
2
2
8
Therefore it follows from equations (4.72) (with r = 2) and (4.73) (with p = ∞) in combination
with equation (4.79) (with q = ∞) in Giné and Nickl (2016), section 4.3, that kb 0 k∞ + Jb 0 Kβ is
β
equivalent to the B ∞,∞ -norm of b 0 for β ∈ (0, 2).
β
If β ∈ (0, 1), then β–Hölder smoothness and B ∞,∞ –smoothness coincide (cf. Proposition 4.3.23
in Giné and Nickl (2016)).
For the prior defined in eqs. (7) to (9) we make the following assumptions.
Assumption 6. The covariance matrix A satisfies one of the following conditions:
(A) For fixed α > 0, A i i = 2−2α`(i ) and A i i 0 = 0 for i 6= i 0 .
(B) There exists 0 < c 1 < c 2 and 0 < c 3 with 3c 3 < c 1 independent from r , such that for all
i , i 0 ∈ Ir
c 1 2−`(i ) ≤ A i i ≤ c 2 2−`(i ) ,
|A i i 0 | ≤ c 3 2−1.5(`(i )+`(i
0
))
if i 6= i 0 .
In particular the second assumption if fulfilled by the prior defined by eq. (10) if 0 < γ ≤ 3/2 and
any σ2 > 0.
Assumption 7. The prior on the truncation level satisfies for some positive constants c 1 , c 2 ,
P(R > r ) ≤ exp(−c 1 2r r ),
(17)
P(R = r ) ≥ exp(−c 2 2r r ).
For the prior on the scaling we assume existence of constants 0 < p 1 < p 2 , q > 0 and C > 1 with
p 1 > q|α − β| such that
¡
¢
P(S ∈ [x p 1 , x p 2 ]) ≥ exp − x q
for all x ≥ C .
(18)
The prior on R can be defined as R = b2 logY c, where Y is Poisson distributed. Equation (18) is
satisfied for a whole range of distributions, including the popular family of inverse gamma distributions. Since the inverse gamma prior on S 2 decays polynomially (lemma 17), condition (A2)
of Shen and Ghosal (2015) is not satisfied and hence their posterior contraction results cannot
be applied to our prior. We obtain the following result for our prior.
Theorem 8. Assume b 0 satisfies assumption 4. Suppose the prior satisfies assumptions 6 and 7.
Let {εn }∞
n=1 be a sequence of positive numbers that converges to zero. There is a constant C 1 > 0
such that for any C 2 > 0 there is a measurable set Bn ⊆ L 2 (T) such that for every a > 0 there is a
positive constant C 3 such that for n sufficiently large
¡
¢
−1/β
log P kb R,S − b 0 k∞ < εn ≥ −C 1 εn | log εn |
¡
¢
−1/β
log P b R,S ∉ Bn ≤ −C 2 εn | log εn |
−1/β
log N (aε, {b ∈ Bn : kb − b 0 k2 ≤ εn }, k · k∞ ) ≤ C 3 εn
9
| log εn |.
(19)
(20)
(21)
The following theorem is obtained by applying these bounds to theorem 3 after taking εn =
(T / log T )−β/(1+2β) .
Theorem 9. Assume b 0 satisfies assumption 4. Suppose the prior satisfies assumptions 6 and 7.
Then for all M T → ∞
Ã
P b0 Πn b : kb − b 0 k2 ≥ M T
µ
T
log T
¶−
β
1+2β
¯
!
¯
¯ T
¯X →0
¯
as T → ∞.
β
This means that when the true parameter is from B ∞,∞ [0, 1], β < 2 a rate is obtained that is op2−δ
[0, 1], for
timal possibly up to a log factor. When β ≥ 2 then b 0 is in particular in the space B ∞,∞
−2/5
every small positive δ, and therefore converges with rate essentially T
.
When a different function Λ is used, defined on a compact interval of R, and the basis elements
P
are defined by ψ j k = m∈Z Λ(2 j (x−m)+k−1); forcing them to be 1-periodic. Then theorem 9 and
derived results for applications still holds provided kψ j k k∞ = 1 and ψ j ,k · ψ j ,l ≡ 0 when |k − l | ≥
d for a fixed d ∈ N and the smoothness assumptions on b 0 are changed accordingly. A finite
number of basis elements can be added or redefined as long as they are 1-periodic.
It is easy to see that our results imply posterior convergences rates in weaker L p -norms, 1 ≤ p < 2,
with the same rate. When p ∈ (2, ∞] the L p -norm is stronger than the L 2 -norm. We apply ideas
of Knapik and Salomond (2014) to obtain rates for stronger L p -norms.
Theorem 10. Assume the true drift b 0 satisfies assumption 4. Suppose the prior satisfies assumptions 6 and 7. Let p ∈ (2, ∞]. Then for all M T → ∞
³
´
β−1/2+1/p
2β−2β/p ¯
¯
− 1+2β
P b0 Πn b : kb − b 0 kp ≥ M T T
(log T ) 1+2β ¯ X T → 0
as T → ∞.
These rates are similar to the rates obtained for the density estimation in Giné and Nickl (2011).
However our proof is less involved. Note that we have only consistency for β > 1/2 − 1/p.
5 Applications to nonparametric regression and density estimation
Our general results also apply to other models. The following results are obtained for b 0 satisfying
assumption 4 and the prior satisfying assumptions 6 and 7.
5.1 Nonparametric regression model
As a direct application of the properties of the prior shown in the previous section, we obtain the
following result for a nonparametric regression problem. Assume
X in = b 0 (i /n) + η i ,
10
0 ≤ i ≤ n,
(22)
with independent Gaussian observation errors η i ∼ N(0, σ2 ). When we apply Ghosal and van der
Vaart (2007), example 7.7 to theorem 8 we obtain, for every M n → ∞,
Ã
Π b : kb − b 0 k2 ≥ M n
µ
n
log n
¶−
β
1+2β
¯
!
¯
¯ n Pb0
¯ X −→ 0
¯
as n → ∞ and (in a similar way as in theorem 10) for every p ∈ (2, ∞],
¯
¶
µ
2β−2β/p ¯
β−1/2+1/p
b
− 1+2β
n P 0
¯
1+2β
(log n)
Π b : kb − b 0 k2 ≥ M n n
¯ X −→ 0
as n → ∞.
5.2 Density estimation
Let us consider n independent observations X n := (X 1 , . . . , X n ) with X i ∼ p 0 where p 0 is an unknown density on [0, 1] relative to the Lebesgue measure. Let P denote the space of densities
on [0, 1] relative to the Lebesgue measure. The natural distance for densities is the Hellinger
distance h defined by
Z 1 ³p
´2
p
2
p(x) − q(x) dx.
h(p, q) =
0
b
Define the prior on P by p = keeb k , where b is endowed with the prior of theorem 9 or its non1
periodic version. Assume that log p 0 is β-smooth in the sense of assumption 4. Applying Ghosal
et al. (2000), theorem 2.1 and van der Vaart and van Zanten (2008), lemma 3.1 to theorem 8, we
obtain for a big enough constant M > 0
Ã
n
Π p ∈ P : h(p, p 0 ) ≥ M
log n
µ
¶−
β
1+2β
¯
!
¯
¯ n P0
¯ X −→ 0,
¯
as n → ∞.
6 Proofs
6.1 Proof of lemma 1
Since conditions (ND) and (LI) of (Karatzas and Shreve, 1991, theorem 5.15) hold, the SDE eq. (1)
has a unique weak solution up to an explosion time.
Assume without loss of generality that X 0 = 0. Define τ0 = 0 and for i ≥ 1 the random times
τi = inf{t ≥ τi −1 : |X t − X τi −1 | = 1}.
By periodicity of drift and the Markov property the random variables Ui = τi − τi −1 are independent and identically distributed.
11
Note that
inf{t : X t = ±n} ≥
n
X
Ui
i =1
P
and hence non-explosion follows from limn→∞ ni=1 Ui = ∞ almost surely. The latter holds true
since U1 > 0 with positive probability, which is clear from the continuity of diffusion paths.
6.2 Proof of lemma 2
Proof of the first part. For the proof we introduce some notation: for any ( j , k), ( j 0 , k 0 ) we write
( j , k) ≺ ( j 0 , k 0 ) if supp ψ j 0 ,k 0 ⊂ supp ψ j ,k . The set of indices become a lattice with partial order ≺,
and by ( j , k) ∨ ( j 0 , k 0 ) we denote the supremum. Identify i with ( j , k) and similarly i 0 with ( j 0 , k 0 ).
For i > 1, denote by t i the time points in [0, 1] corresponding to the maxima of ψi . Without loss
of generality assume t i < t i 0 . We have G i ,i 0 = 0 if and only if the interiors of the supports of ψi and
ψi 0 are disjoint. In that case
max supp ψ j ,k ≤ t ( j ,k)∨( j 0 ,k 0 ) ≤ min supp ψ j 0 ,k 0 .
(23)
The values of Zi can be found by the midpoint displacement technique. The coefficients are
given by Z1 = V0 , Z2 = V 1 and for j ≥ 1
2
Z j ,k = V2− j (k−1/2) −
¢
1¡
V2− j (k−1) + V2− j k .
2
As V is a Gaussian process, the vector Z is mean-zero Gaussian, say with (infinite) precision
matrix Γ. Now Γi ,i 0 = 0 if there exists a set L ⊂ N such that L ∩ {i , i 0 } = ∅ for which conditional
on {Zi ? , i ? ∈ L }, Zi are Zi 0 are independent.
Define ( j ? , k ? ) = ( j , k) ∨ ( j 0 , k 0 ) and
L = {i ? ∈ N : i ? = 2 j + k, with j ≤ j ? }.
The set {Zi ? , i ? ∈ L } determine the process V at all times k2− j
?
−1
, k = 0...,2j
?
?
+1
Now Zi and Zi 0 are conditionally independent given {Vt , t = k2− j −1 , k = 0 . . . , 2 j
the Markov property of the nonperiodic Ornstein-Uhlenbeck process.
The result follows since σ({Zi ? , i ? ∈ L }) = σ({Vt , t = k2− j
Lemma 11. Let K (s, t ) = EVs Vt =
?
−1
, k = 0...,2j
σ
1
−γ|t −s|
+ e −γ(1−|t −s|) ).
2γ 1−e −γ (e
2
?
+1
.
?
+1
} by (23) and
}).
If x ∉ (s, t )
1
1
t +s
K (s, x) − K ( s+t
, x) + K (t , x) = 2 sinh2 (γ t −s
2
4 )K ( 2 , x)
2
2
Proof. Without loss of generality assume that t ≤ x ≤ 1. With m = (t + s)/2 and δ = (t − s)/2
(e −γ|s−x| + e −γ(1−|s−x|) ) − 2(e −γ|m−x| + e −γ(1−|m−x|) ) + (e −γ|t −x| + e −γ(1−|t −x|) )
= e −γ|t −x| e −2γδ − 2e −γ|t −x| e −γδ + e −γ|t −x| + e −γ(1−|s−x|) − 2e −γ(1−|s−x|) e −γδ + e −γ(1−|s−x|) e −2γδ
= (1 − e −γδ )2 (e −γ|t −x| + e −γ(1−|s−x|) ) = (1 − e −γδ )2 e γδ (e −γ|m−x| + e −γ(1−|m−x|) )
The result follows from (1 − e −γδ )2 e γδ = 4 sinh2 (γδ/2) and scaling both sides with
12
1 σ2
1
2 2γ 1−e −γ .
Proof of the second part. Denote by [a, b], [c, d ] the support of ψi and ψi 0 respectively and let
m = (b + a)/2 and n = (d + c)/2 but for i = 1, let m = 0. Z1 = V (0), Z2 = V1/2 and Var(Z1 ) =
2
σ2
−1
Var(Z2 ) = σ
2γ coth(γ/2), and Cov(Z1 , Z2 ) = 2γ sinh (γ/2). Note that the 2 × 2 covariance matrix of
Z1 and Z2 has eigenvalues
σ2
2γ
tanh(γ/4) and
σ2
2γ
coth(γ/4) and is strictly positive definite.
By midpoint displacement, 2Zi = 2Vm − Va − Vb , i > 2 and K (s, t ) = EVs Vt =
σ2
1
−γ|t −s|
+
2γ 1−e −γ (e
e −γ(1−|t −s|) ).
Assume without loss of generality b−a ≥ d −c. Define δ to be the halfwidth of the smaller interval,
0
so that δ := (d − c)/2 = 2− j −1 . Then
(b − a)/2 = 2− j −1 = hδ,
with h = 2 j
0
−j
.
Consider three cases:
1. The entries on diagonal, i = i 0 ;
2. The interiors of the supports of ψi and ψi 0 are non-overlapping;
3. The support of ψi 0 is contained in the support of ψi .
Case 1. By elementary computations for i > 2,
4
2γ
(1 − e −γ )A i i = 6(1 + e −γ ) + 2(e −γ2δ + e −γ(1−2δ) ) − 8(e −γδ + e −γ(1−δ) ) =
σ2
= 2(1 − e −γδ )(3 − e −γδ ) + 2e −γ (1 − e γδ )(3 − e γδ ).
As δ ≤
1
4
and under the assumption γ ≤ 3/2 the last display can be bounded by
0.9715 · 4γδ(1 − e −γ ) ≤ 4
2γ
(1 − e −γ )A i i ≤ 4γδ(1 − e −γ ).
σ2
Hence 0.9715 · 2− j σ2 /4 ≤ A i i ≤ 2− j σ2 /4.
Case 2. Necessarily i , i 0 > 2. By twofold application of lemma 11
A i j = (K (c, b) − 2K (n, b) + K (d , b))/4
− 2(K (c, m) − 2K (n, m) + K (d , m))/4
+ (K (c, a) − 2K (n, a) + K (d , a))/4
(24)
2
= 2 sinh (γ d −c
4 )(K (n, b) − 2K (n, m) + K (n, a))/2
2 b−a
= 4 sinh (γ 4 ) sinh2 (γ d −c
4 )K (n, m).
Using the convexity of sinh we obtain the bound
2 sinh2 (x/2) ≤ 0.55x 2
(25)
for 0 ≤ x ≤ 1. Note that f (x) = e −γx + e −γ(1−x) is convex on [0, 1], from which we derive f (x) ≤
1 + e −γ . Using this bound, and the fact that for γ ≤ 3/2,
γ2 K (n, m) ≤
2
σ2
2 γ coth(γ/2) ≤ σ (1 + γ/2),
13
(26)
which can be easily seen from a plot, that
|A i i 0 | ≤0.552 γ4 · 2−2 j −2 · 2−2 j
2 2
0
−2
≤0.0095σ γ (1 + γ/2)2
|K (n, m)|
−1.5( j + j 0 )
.
Case 3.
For i 0 > 2, i = 1 with m = 0 or i = 2 with m = 21 , using eq. (26), we obtain
1
1
|A i i 0 | = |K (m, n) − K (m, c) − K (m, d )|
2
2
≤2 sinh2 (γ d −c
)K
(m,
n)
4
≤0.55γ2 2−2 j
0
−2
(27)
K (m, n)
≤0.098σ2 (1 + γ/2)2−1.5 j .
When i , i 0 > 2 then, using the calculation eq. (24) and lemma 11 noting that a, b and m are not in
(c, d ), we obtain
A i i 0 = 2 sinh2 (γ d −c
4 )(K (n, b) − 2K (n, m) + K (n, a))/2.
Write x = γ|a − m| = γ|b − m| = γhδ and α =
|m−n|
|b−m|
∈ (0, 1). A simple computation then shows
e −γ|b−n| − 2e −γ|m−n| + e −γ|a−n|
=e −(1+α)x − 2e −αx + e −(1−α)x .
The derivative of f (α) := e −(1+α)x − 2e −αx + e −(1−α)x is nonnegative, for α, x > 0 hence f (α) is
increasing and so f (0) ≤ f (α) ≤ f (1). Note that f (0) = 2e −x − 2 ≥ −2x, for x > 0 and f (1) = e −2x −
2e −x + 1 =: g (x). Maximising g 0 (x) over x > 0 gives g 0 (x) ≤ 1/2 and g (0) = 0 and therefore f (1) =
g (x) ≤ x/2.
It follows that
−2γhδ ≤ e −γ|b−n| − 2e −γ|m−n| + e −γ|a−n| ≤ γhδ/2.
For the other terms we derive the following bounds. Write
e −γ(1−|b−n|) − 2e −γ(1−|m−n|) + e −γ(1−|a−n|)
=e −γ+(1+α)x − 2e −γ+αx + e −γ+(1−α)x =: h(α).
Now h(α) is decreasing for x ≤ log 2 and convex and positive for x ≥ log 2. In both case we can
bound h(α) by its value at the endpoints α = 0 and α = 1. Using that 2x ≤ γ we obtain 0 ≤ h(0) =
¡
¢
e −γ (2e x − 2) ≤ 2x and 0 ≤ h(1) = e −γ e 2x − 2e x + 1 ≤ 2x. So 0 ≤ h(α) ≤ 2γhδ.
Using the bound eq. (25) and x/(1 − exp(−x)) ≤ (1 + x) we obtain
0
|A i i 0 | ≤ 0.061σ2 γ(1 + γ)2−1.5( j + j ) .
14
6.3 Proof of theorem 3
A general result for deriving contraction rates for Brownian semi-martingale models was proved
in van der Meulen et al. (2006). Theorem 3 follows upon verifying the assumptions of this result
for the diffusion on the circle. These assumptions are easily seen to boil down to:
1. For every T > 0 and b 1 , b 2 ∈ L 2 (T) the measures P b1 ,T and P b2 ,T are equivalent.
2. The posterior as defined in equation eq. (5) is well defined.
3. Define the (random) Hellinger semimetric h T on L 2 (T) by
Z T³
´2
2
h T (b 1 , b 2 ) :=
b 1 − b 2 (X t ) dt , b 1 , b 2 ∈ L 2 (T).
(28)
0
There are constants 0 < c < C for which
´
³ p
p
lim P þ0 ,T c T kb 1 − b 2 k2 ≤ h T (b 1 , b 2 ) ≤ C T kb 1 − b 2 k2 , ∀ , b 1 , b 2 ∈ L 2 (T) = 1.
T →∞
We start by verifying the third condition. Recall that the local time of the process X T is defined
as the random process L T (x) which satisfies
Z T
Z
f (X t ) dt = f (x)L T (x) dx.
0
R
For every measurable function f for which the above integrals are defined. Since we are working
with 1-periodic functions, we define the periodic local time by
X
L̊ T (x) =
L T (x + k).
k∈Z
Note that t 7→ X t is continuous with probability one. Hence the support of t 7→ X t is compact
with probability one. Since x 7→ L T (x) is only positive on the support of t 7→ X t , it follows that the
sum in the definition of L̊ T (x) has only finitely many nonzero terms and is therefore well defined.
For a one-periodic function f we have
Z T
Z 1
f (X t ) dt =
f (x)L̊ T (x) dx,
0
0
provided the involved integrals exists. It follows from (Schauer and van Zanten, 2017, Theorem
5.3) that L̊ T (x)/T converges to a positive deterministic function only depending only on b 0 and
which is bounded away from zero and infinity. Since the Hellinger distance can be written as
s
Z 1
p
L̊ T (x)
h T (b 1 , b 2 ) = T
(b 1 (x) − b 2 (x))2
dt
T
0
p
it follows that the third assumption is satisfied with d T (b 1 , b 2 ) = T kb 1 − b 2 k2 .
Conditions 1 and 2 now follow by arguing precisely as in lemmas A.2 and 3.1 of van Waaij and
van Zanten (2016) respectively (the key observation being that the convergence result of L̊ T (x)/T
R1
also holds when 0 b(x) dx is nonzero, which is assumed in that paper).
p
The stated result follows from Theorem 2.1 in van der Meulen et al. (2006) (taking µT = T εT in
their paper).
15
6.4 Proof of theorem 8 with assumption 6 (A)
The proof proceeds by verifying the conditions of theorem 3. By assumption 4 the true drift can
P
P2 j
be represented as b 0 = z 1 ψ1 + ∞
j =0 k=1 z j k ψ j k . For r ≥ 0, define its truncated version by
b 0r
= z 1 ψ1 +
r X
2j
X
z jkψjk.
j =0 k=1
6.4.1 Small ball probability
For ε > 0 choose an integer r ε with
1/β
Cβε
−1/β
rε
≤ 2 ≤ 2C β ε
−1/β
where
Cβ =
Jb0 Kβ
(2β − 1)1/β
.
(29)
For notational convenience we will write r instead of r ε in the remainder of the proof. By
lemma 16 we have kb 0r − b 0 k∞ ≤ ε. Therefore
kb r,s − b 0 k2 ≤ kb r,s − b 0r k2 + kb 0r − b 0 k2 ≤ kb r,s − b 0r k∞ + ε
which implies
¢
¡
¢
¡
P kb r,s − b 0 k2 < 2ε ≥ P kb r,s − b 0r k∞ < ε .
Let f S denotes the probability density of S. For any x > 0, we have
Z ∞
¡
¢ X
¡
¢
P kb R,S − b 0 k2 < 2ε =
P kb r,s − b 0 k2 < 2ε f S (s) ds
P(R = r )
0
r ≥1
≥ P(R = r )
where
Lε = ε
inf
s∈[L ε ,Uε ]
p
− qβ1
and
¢
¡
P kb r,s − b 0r k∞ < ε
Uε = ε
Z
Uε
Lε
f S (s) ds,
(30)
p
− qβ2
and p 1 , p 2 and q are taken from assumption 7. For ε sufficiently small, we have by the second
part of assumption 7
Z Uε
¡
−1 ¢
f S (s) ds ≥ exp − ε β
Lε
By choice of r and the first part of assumption 7, there exists a positive constant C such that
³
´
³
´
−1
P(R = r ) ≥ exp − c 2 2r r ≥ exp −C ε β | log ε| ,
for ε sufficiently small.
For lower bounding the middle term in equation (30), we write
b r,s − b 0r = (s Z1 − z 1 )ψ1 +
r X
2j
X
j =0 k=1
16
(s Z j k − z j k )ψ j k
which implies
kb r,s − b 0r k∞ ≤ |s Z1 − z 1 | +
r
X
max |s Z j k − z j k | ≤ (r + 2) max |s Zi − z i |.
j =0 1≤k≤2
j
i ∈Ir
This gives the bound
¡
¢ Y ³
P kb r,s − b 0r k∞ < ε ≥
P |s Zi − z i | <
i ∈Ir
ε ´
.
r +2
By choice of the Zi , we have for all i ∈ {1, 2, . . .}, 2α`(i ) Zi is standard normally distributed and
hence
Ã
!
¯
¯ 2α`(i ) ε
³
ε ´
¯ α`(i )
¯
α`(i )
= log P ¯2
log P |s Zi − z i | <
Zi − 2
z i /s ¯ <
r +2
(r + 2)s
Ã
!
2α`(i ) 2
zi 1
¡ ¢
2α`(i ) ε
22α`(i ) ε2 2
≥ log
−
−
+ 2 log π2 ,
2
2
2
(r + 2)s
(r + 2) s
s
where the inequality follows from lemma 18. The third term can be further bounded as we have
22α`(i ) z i2 = 22(α−β)`(i ) 22β`(i ) z i2 ≤ 22(α−β)`(i ) Jb 0 K2β .
Hence
Ã
!
2(α−β)`(i )
Jb0 K2β
¡ ¢
2α`(i ) ε
22α`(i ) ε2 2
ε ´
−
+ 12 log π2 .
≥ log
−
log P |s Zi − z i | <
2
2
2
r +2
(r + 2)s
(r + 2) s
s
³
(31)
For s ∈ [L ε ,Uε ] and i ∈ Ir we will now derive bounds on the first three terms on the right of
eq. (31). For ε sufficiently small we have r ≤ r + 2 ≤ 2r and then inequality (29) implies
logC β ≤ r + 2 ≤ 2 log(4C β ) +
2
| log ε|.
β
Bounding the first term on the RHS of (31). For ε sufficiently small, we have
³
´¶
µ
¶
µ
¶
µ
p
(r + 2)s
(r + 2)Uε
− 1+ qβ2
log α`(i ) ≤ log
= log (r + 2)ε
ε
2
ε
½
¾ µ
¶
2
p2
≤ log 2 log(4C β ) + | log ε| + 1 +
| log ε|
β
qβ
≤ C̃ p 2 ,q,β | log ε|,
where C̃ p 2 ,q,β is a positive constant.
Bounding the second term on the RHS of (31). For ε sufficiently small, we have
(2C β )2α 2 (−α+β+p 1 /q )
22α`(i ) ε2
22αr ε2
≤
≤
εβ
≤ 1.
(r + 2)2 s 2 (logC β )2 L 2ε (logC β )2
The final inequality is immediate in case α = β, else if suffices to verify that the exponent is nonnegative under the assumption p 1 > q|α − β|.
17
Bounding the third term on the RHS of (31). For ε sufficiently small, in case β ≥ α we have
22(α−β)`(i ) Jb 0 K2β
s2
≤ Jb 0 K2β L −2
ε ≤ 1.
In case β < α we have
22(α−β)`(i ) Jb 0 K2β
s2
≤
22(α−β)r Jb 0 K2β
L 2ε
2
≤ (2C β )2(α−β) ε β (
p 1 /q−α+β)
≤1
as the exponent of ε is positive under the assumption p 1 > q|α − β|.
Hence for ε small enough, we have
³
log P |s Zi − z i | <
ε ´
≥ −C̃ p 2 ,q,β | log ε| − 3.
r +2
As −2r +1 ≥ −4C β ε−1/β we get
log
inf
p
s∈[x
1 ,x p 2 ]
¡
¢
¡
¢
P kb r,s − b 0r k∞ < ε ≥ −4C β ε−1/β C̃ p 2 ,q,β | log ε| + 3
& −ε−1/β | log ε|.
¡
¢
We conclude that the right hand side of eq. (30) is bounded below by exp −C 1 ε−1/β | log ε| , for
some positive constant C 1 and sufficiently small ε.
6.4.2 Entropy and remaining mass conditions
For r ∈ {0, 1, . . .} denote by C r the linear space spanned by ψ1 and ψ j k , 0 ≤ j ≤ r, k ∈ 1, . . . , 2 j , and
define
©
ª
C r,t := b ∈ C r , Jb Kα ≤ t .
Proposition 12. For any ε > 0
log N (ε, C r,t , k · k∞ ) ≤ 2r +1 log(3A α t ε−1 ),
where A α =
P∞
k=0
2−kα .
Proof. We follow (van der Meulen et al., 2006, §3.2.2). Choose ε0 , . . . , εr > 0 such that
Define
(£
¤2 j
−2−α j t , 2−α j t
if j ∈ {1, . . . , r }
Uj =
.
2
[−t , t ]
if j = 0
Pr
j =0 ε j
j
≤ ε.
For each j ∈ {1, . . . , r }, let E j be a minimal ε j -net with respect to the max-distance on R2 and let
E 0 be a minimal ε0 -net with respect to the max-distance on R2 . Hence, if x ∈ U j , then there exists
a e j ∈ E j such that maxk |x k − e k | ≤ ε j .
18
Take b ∈ C r,t arbitrary: b = z 1 ψ1 +
Pr
j =0
P2 j
k=1
z j k ψ j k . Let b̃ = e 1 ψ1 +
Pr
j =0
P2 j
e ψ ,
k=1 j k j k
where
j
(e 1 , e 0,1 ) ∈ E 0 and (e j 1 , . . . , e j 2 j ) ∈ E j (for j = 1, . . . , 2 ). We have
kb − b̃k∞ ≤ |z 1 − e 1 |kψ1 k∞ +
≤ |z 1 − e 1 | +
This can be bounded by
Pr
j =0 ε j
r
X
r
X
max |z j k − e j k |kψ j k k∞
j =0 1≤k≤2
j
max 2 j α |2− j α z j k − 2− j α e j k |.
j
j =0 1≤k≤2
by an appropriate choice of the coefficients in b̃. In that case we
obtain that kb − b̃k∞ ≤ ε. This implies
µ
¶
3 · 2−α j t
2 log
log |E j | ≤
.
log N (ε, C r,t , k · k∞ ) ≤
εj
j =0
j =0
r
X
r
X
j
The asserted bound now follows upon choosing ε j = ε2− j α /A α .
Proposition 13. There exists a constant a positive constant K such that
¡
¢
log N (aε, {b ∈ C r : kb − b 0 k2 ≤ ε}, k · k2 ) ≤ 2r +1 log 6A α K 2αr .
Proof. There exists a positive K such that
{b ∈ C r : kb − b 0 k2 ≤ aε} ⊂ {b ∈ C r : kbk2 ≤ K }.
By lemma 21, this set is included in the set
n
o
p
b ∈ C r : kbk∞ ≤ 32(r +1)/2 K .
By lemma 20, for any b = z 1 ψ1 +
Pr
j =0
P2 j
k=1
(32)
z j k ψ j k in this set we have
o
n
p
max |z 1 |, |z j k |, j = 0, . . . , r, k = 1 . . . , 2 j ≤ 2kbk∞ 32(r +1)/2 K .
©
ª
Hence, the set eq. (32) is included in the set b ∈ C r : Jb Kα ≤ a(r, ε) = C r,a(r,ε) , where a(r, ε) =
p
21+αr 32(r +1)/2 K .
Hence,
¡
¢
N (aε, {b ∈ C r : kb − b 0 k2 ≤ ε}, k · k2 ) ≤ N ε, C r,a(r,ε) , k · k2 .
Using lemma 21 again the latter can be bounded by
³ p
´
N ε 32(r +1)/2 , C r,a(r,ε) , k · k∞
The result follows upon applying proposition 12.
19
We can now finish the proof for the entropy and remaining mass conditions. Choose r n to be the
− β1
smallest integer so that 2r n ≥ Lεn , where L is a constant, and set Bn = C r n . The entropy bound
then follows directly from proposition 13.
For the remaining mass condition, using assumption 7, we obtain
¡
¢
¡
¢
¡
¢
− β1
P b R,S ∉ Bn = P(R > r n ) ≤ exp − c 1 2r n r n ≤ exp −C 3 εn | log εn | ,
and note that the constant C 3 can be made arbitrarily big by choosing L big enough.
6.5 Proof of theorem 8 under assumption 6 (B)
We start with a lemma.
Lemma 14. Assume there exists 0 < c 1 < c 2 and 0 < c 3 with c 3 < c 1 independent from r , such that
for all i , i 0 , 2 ≤ `(i ), `(i 0 ) ≤ r ,
c 1 2−`(i ) ≤ A i i ≤ c 2 2−`(i ) ,
|A i i 0 | ≤ c 3 2
−1.5(`(i )+`(i 0 ))
(33)
0
if i 6= i .
(34)
Let Ae = (A i i 0 )2≤`(i ),`(i 0 )≤r (so the right-lower submatrix of A r ). Then for all x ∈ R|Ir |−2
e ≤ x 0 Ax
e ≤ 2c 2 x 0 Λx.
e
(c 1 − c 3 )x 0 Λx
e = (Λ
e i i 0 )2≤`(i ),`(i 0 )≤r is the diagonal matrix with Λ
e i i = 2−`(i ) , .
where Λ
Proof. In the following the summation are over i , i 0 , 2 ≤ `(i ), `(i 0 ) ≤ r .
P 2
P
i x i A i i + i 6= j x i A i j x j . By the first inequality
c 1 x 0 Λ(r ) x = c 1
X
i
x i2 2−`(i ) <
X
i
x i2 A i i < c 2
X
i
Trivially, x 0 A r x =
x i2 2−`(i ) = c 2 x 0 Λ(r ) x.
On the other hand
¯
¯
Ã
!2
¯X
¯
X
X
0
¯
¯
x A 0 x 0 ¯ ≤ c3
|x i |2−1.5`(i ) |x i 0 |2−1.5`(i ) ≤ c 3
|x i |2−1.5`(i ) .
¯
¯i 6=i 0 i i i i ¯
i
i 6=i 0
At the first inequality we used the second part of of (33). The second inequality follows upon
including the diagonal. By Cauchy-Schwarz, this can be further bounded by
Ã
!Ã
!
X 2 −`(i ) X −2`(i )
c3
xi 2
2
≤ c 3 x 0 Λx,
i
i
P
P
P
−2`(i )
j −2 j
where the final inequality follows from i 2−2`(i ) ≤ ∞
= ∞
= 1. The result
i =3 2
j =1 2 2
follows by combining the derived inequalities.
20
We continue with the proof of theorem 8. Write A as block matrix
¸
·
A1 B 0
,
A=
B A2
with A 1 a 2 × 2-matrix, and B , A 2 defined accordingly. By lemma 2
·
¸
σ2
coth(γ/2) sinh−1 (γ/2)
A1 =
γ
.
sinh−1 (γ/2) coth(γ/2)
2
Define the 2 × 2-matrix
Λ1 = c σ2 γ tanh(γ/4)I,
2
c ∈ (0, 1).
where I is the 2 × 2-identity matrix. It is easy to see that A 1 − Λ1 is positive definite.
When A 2 − Λ2 − B (A 1 − Λ1 )−1 B 0 is positive definite, then it follows from the Cholesky decomposition that A − Λ is positive definite, where Λ = diag(Λ1 , Λ2 ) positive definite.
Note
where
!
Ã
0
(B A −1
1 B )i ,i 0
=
X
k,k 0
0 0
B i k (A 1 )−1
kk 0 B i k
Ã
!
X
k,k 0
(A 1 )−1
kk 0
=
≤
X
k,k 0
(A 1 )−1
kk 0
(B i ,1 ∨ B i ,2 )(B i 0 ,1 ∨ B i 0 ,2 )
2
2
2
≤ 2
.
−1
2
σ γ sinh (γ/2) + coth(γ/2) σ (1 + γ)
Therefore
0
2
−1.5(`(i )+`(i
|(B A −1
1 B )i i 0 | ≤ 0.020σ (1 + γ/4)2
0
))
0 0
Now consider à = A 2 − Λ2 − B (A 1 − Λ1 )−1 B 0 . By lemma 2 and the bound on |(B A −1
1 B )i i | and
choosing c > 0 in the definition of Λ1 small enough, under the assumption that γ ≤ 1.5,
0.945 · 2−`(i ) σ2 /4 < Ã i i < 1.03 · 2−`(i ) σ2 /4.
and for i 6= i 0 | Ã i i 0 | ≤ 0.9415 σ4 2−1.5(`(i )+`(i )) . Therefore by lemma 14 Ã − Λ2 is positive definite
with diagonal matrix Λ2 with diagonal entries 2−`(i ) .
2
0
It follows that x 0 Λx ³ x 0 Ax. This implies that the small ball probabilities and the mass outside
a sieve behave similar under assumption 6(B) as when the Zi are independent normally distributed with zero mean and variance ξ2i = Λi i . As this case corresponds to assumption 6(A) with
α = 21 for which posterior contraction has already been established, the stated contraction rate
under assumption 6(B) follows from Anderson’s lemma (lemma 19).
6.6 Proof of theorem 10: convergence in stronger norms
The linear embedding operator T : L p (T) → L 2 (T), x 7→ x is a well-defined injective continuous
operator for all p ∈ (2, ∞]. Its inverse is easily seen to be a densely defined, closed unbounded
linear operator. Following Knapik and Salomond (2014) we define the modulus of continuity m
as
©
ª
m(Bn , ε) := sup k f − f 0 kp : f ∈ Bn , k f − f 0 k2 ≤ ε .
Theorem 2.1 of Knapik and Salomond (2014) adapted to our case is
21
Theorem 15 (Knapik and Salomond (2014)). Let εn ↓ 0, Tn ↑ ∞ and Π be a prior on L p (T) such
that
¡
¢
E0 Π Bnc | X Tn → 0,
for measurable sets Bn ⊂ L p (T). Assume that for any positive sequence M n
¡
¢
E0 Π b ∈ Bn : kb − b 0 k2 ≥ M n εn | X Tn → 0,
then
¡
¢
E0 Π b ∈ L p (T) : kb − b 0 kp ≥ m(Bn , M n εn ) | X Tn → 0.
c
|
Note that the sieves C r,t which we define in section 6.4.2 have by eq. (15) the property Π(C r,t
T
r (1/2−1/p)
X ) → 0. By lemmas 21 and 23, the modulus of continuity satisfies m(C r,u , εn ) . 2
εn ,
for all p ∈ (2, ∞], (assume 1/∞ = 0), and the result follows.
A Lemmas used in the proofs
Lemma 16. Suppose z has Faber-Schauder expansion
2j
∞ X
X
z = z 1 ψ1 +
z jkψjk.
j =0 k=1
If Jz Kβ < ∞ (with the norm defined in (16)), then for r ≥ 1
°
°
X
°
°
z i ψi °
°z −
∞
i ∈Ir
≤ Jz Kβ
2−r β
2β − 1
.
(35)
Proof. This follows from
°
°
X
°
°
z i ψi °
°z −
i ∈Ir
∞
°
2j
∞ °X
X
°
°
z jkψjk°
°
≤
=
j =r +1
∞
X
j =r +1
∞
k=1
2− j β max 2 j β |z j k | ≤ Jz Kβ
1≤k≤2 j
∞
X
2− j β .
j =r +1
Lemma 17. If X ∼ IG(A, B ) then for any M > 0,
P(X ≥ M ) ≤
BA
M −A .
Γ(A)
Proof. This follows from
P(X ≥ M ) ≤
BA
Γ(A)
Z
∞
M
x −α−1 dx = −
22
B A £ −α ¤∞
BA
x
=
M −A .
x=M
Γ(A)
Γ(A)
Lemma 18. Let X ∼ N(0, 1), θ ∈ R and ε > 0.Then
q
2
p ´
e −θ ³
log ε−ε2 −θ 2 +log π2
P(|X − θ| ≤ ε) ≥ p P |X | ≤ 2ε ≥ e
.
2
Proof. Note that
θ+ε
Z
θ−ε
and
1
e − 2 (x+θ)
e
thus e
− 12 (x+θ)2
2
≥ e −θ e
Z
− 12 (
p 2
2x)
θ+ε
θ−ε
−θ 2
e
− 12 x 2
Now the elementary bound
e
1
2
e − 2 x dx =
Z
ε
1
2
e − 2 (x+θ) dx
−ε
2
= eθ
p
− 12 ( 2x)2
2
− 12 (x+θ)2 +x 2
1
2
= e 2 (x−θ) ≥ 1,
, hence
dx ≥ e
Ry
−y e
1
P(|X − θ| ≤ ε) = p
2π
−θ 2
− 12 x 2
Z
Z
e
p
− 21 ( 2x)2
−ε
1
2
e −θ
dx = p
2
Z
p
2ε
p e
− 2ε
− 21 u 2
du.
2
≥ 2ye − 2 y gives
2
θ+ε
θ−ε
ε
e
− 12 x 2
1 e −θ
dx ≥ p
p
2π 2
2
p
2ε
Z
p e
− 2ε
2
− 12 u 2
p
2
1 e −θ p
e −θ
= p P(|X | ≤ 2ε) ≥ p
p 2 2εe −ε =
2π 2
2
du
r
2 log ε−θ2 −ε2
e
π
Lemma 19 (Anderson’s lemma). Define a partial order on the space of n×n-matrices (n ∈ N ∪{∞})
by setting A ≤ B, when B − A is positive definite. If X ∼ N(0, ΣX ) and Y ∼ N(0, ΣY ) independently
with Σ X ≤ ΣY , then for all symmetric convex sets C
P(Y ∈ C ) ≤ P(X ∈ C ).
Proof. See Anderson (1955).
Lemma 20. Let
f = z 1 ψ1 +
r X
2j
X
z j ,k ψ j ,k .
j =1 k=1
Then
sup |z i | ≤ 2k f k∞ .
i :`(i )≤r
Proof. Note that |z 1 | = | f (0)| ≤ 2k f k∞ , and |z 0,1 | = | f (1/2)| ≤ 2k f k∞ and inductively, for j ≥ 1,
¡
¢
¡
¢
¡
¢
z j k = f (2k − 1)2−( j +2) − 12 f 2−( j +1) (k − 1) − 12 f 2−( j −1) k , hence |z j k | ≤ 2k f k∞ .
Lemma 21. Let C r as in section 6.4.2. Then
sup
06= f ∈C r
k f k∞ p
≤ 3 · 2(r +1)/2 .
k f k2
23
Proof. Let f ∈ C r be nonzero. Note that for any constant c > 0,
kc f k∞ k f k∞
=
.
kc f k2
k f k2
Hence, we may and do assume that k f k∞ = 1. Furthermore, since the L 2 and L ∞ norm of f and
| f | are the same, we also assume that f is nonnegative.
Let x 0 be a global maximum of f . Clearly f (x 0 ) = 1. Since f is a linear interpolation between
the points {k2− j −1 : k = 0, 1, . . . , 2r +1 }, we may also assume that x 0 is of the form x 0 = k2− j −1 . We
consider two cases
(i) 0 ≤ k < 2r +1 ,
(ii) k = 2r +1 .
¡
¢
In case (i) we have that f (x) ≥ 1 − 2r +1 (x − k2−r −1 ) I[k2−r −1 ,(k+1)2−r −1 ] (x), for all x ∈ [k2−r −1 , (k +
1)2−r −1 ]. In case (ii) f (x) ≥ 2r +1 (x −1+2−r −1 )I[1−2−r −1 ,1] (x), for all x ∈ [1−2−r −1 , 1]. Hence, in both
cases,
Z 2−r −1
1
1
2
2r +2
k f k2 ≥ 2
x 2 dx = 22r +2 2−3r −3 = 2−r −1 .
3
3
0
Thus
k f k∞
≤
k f k2
1
p1 2−(r +1)/2
3
=
p
3 · 2(r +1)/2 ,
uniformly over all nonzero f ∈ C r,s .
Lemma 22. Let a 1 , a 2 , x 1 , x 2 be positive numbers. Then
a1 + a2 a1 W a2
≤
.
x1 + x2 x1 x2
Proof. Suppose that the lemma is not true, so there are positive a 1 , a 2 , x 1 , x 2 such that,
·
¸ ·
¸
a1 W a2 a1 + a2
a1 a1 + a2 W a2 a1 + a2
−
=
−
−
x1 x2 x1 + x2
x1 x1 + x2
x2 x1 + x2
" a1
# " a2
#
x 1 (x 1 + x 2 ) − (a 1 + a 2 ) W x 2 (x 1 + x 2 ) − (a 1 + a 2 )
=
< 0.
x1 + x2
x1 + x2
Hence, both terms on the right-hand-side are negative. In particular, this means for the first term
that x 2 /x 1 < a 2 /a 1 . For the second term this gives x 1 /x 2 < a 1 /a 2 . These two inequalities cannot
hold simultaneously and we have reached a contradiction.
Lemma 23. Let C r and C r,s as in section 6.4.2. Then for p ∈ [2, ∞),
sup
06= f ∈C r
k f kp
k f k2
≤
31/2
(p + 1)1/p
24
2(r +1)(1/2−1/p) .
Proof. Let f ∈ C r . Just as in proof of lemma 21 we may assume that f is nonnegative and k f k2 =
1. Hence
Ã
!1/p Ã
p !1/p
k f kp
k f kp
p
=
sup
=
sup
k f kp
.
sup
2
06= f ∈C r ,k f k2 =1 k f k2
06= f ∈C r ,k f k2 =1
06= f ∈C r k f k2
Note that
p
k f kp =
+1
2r X
−1 Z (k+1)2−r −1
k=0
Hence, by repeatedly applying lemma 22
P2r +1 R (k+1)2−r −1
f (x)p dx
k=0 k2−r −1
≤
P2r +1 −1 R (k+1)2−r −1
f (x)2 dx
k=0
k2−r −1
f (x)p dx.
k2−r −1
R (k+1)2−r −1
_
k ∈ {0, . . . , 2r +1 − 1}
∃x ∈ (k2−r −1 , (k + 1)2−r −1 ) : f (x) 6= 0
k2−r −1
R (k+1)2−r −1
k2−r −1
f (x)p dx
f (x)2 dx
.
Note that f is a linear interpolation between the points k2−r −1 , k ∈ {0, 1, . . . , 2r +1 }.
Now study affine functions g : [0, 2−r −1 ] → R which are positive. A maximum of g is attained in
either 0 or 2−r −1 . Without lose of generality it is attained in 0. Using scaling in a later stadium of
the proof, we assume for the moment that g (0) = 1. Hence a := g (2−r −1 ) ∈ [0, 1]. Note that
g (x) = 1 − (1 − a)2r +1 x.
When a = 1, kg kp = kg k2 = 1. Now consider a < 1,
Z 2−r −1
Z 2−r −1
£
¤p
p
g (x) dx =
1 − (1 − a)2r +1 x dx.
0
0
−r −1
−r −1
Let y = −x + 21−a then x = −y + 21−a and dx = − dy. Hence
2−r −1
Z
0
Z
p
g (x) dx =
2−r −1
1−a
−r −1
−2−r −1 + 2 1−a
= 2−r −1
(1 − a)p 2r p+p y p dy
·
¸
1
1 − a p+1
.
p +1
1−a
Note that for a constant c > 0 and a function h,
p
kchkp
kchk22
p
=
c p khkp
c 2 khk22
Let
c 2 = 3 · 2r +1
p
= c p−2
khkp
khk22
.
1−a
.
1 − a3
Hence c g has L 2 -norm one and
p
p
kcg kp = c p kg kp
µ
¶p
·
¸
2
1 − a p+1
r +1 1 − a
−r −1 1
= 3·2
2
1 − a3
p +1
1−a
=
3p/2 (r +1)(p/2−1)
2
(1 − a)p/2−1 (1 − a 3 )−p/2 (1 − a p+1 ).
p +1
25
The maximum is attained for a = 0, then
p
kc g kp =
Hence
kcg kp =
3p/2 (r +1)(p/2−1)
2
p +1
31/2
(p + 1)1/p
2(r +1)(1/2−1/p)
and the result follows, using that k f I(k2−r −1 ,(k+1)2−r −1 ) k22 ≤ k f k22 and that for 0 < c 0 < c,
p
p
kc 0 g kp
p
(c 0 /c)p kc g kp kc g kp
=
≤
.
kc 0 g k22 (c 0 /c)2 kc g k22 kc g k22
B Acknowledgement
This work was partly supported by the Netherlands Organisation for Scientific Research (NWO)
under the research programme “Foundations of nonparametric Bayes procedures”, 639.033.110
and by the ERC Advanced Grant “Bayesian Statistics in Infinite Dimensions”, 320637.
References
Anderson, T. W. (1955). The integral of a symmetric unimodal function over a symmetric convex
set and some probability inequalities. Proc. Amer. Math. Soc., 6:170–176.
Bhattacharya, R. and Waymire, E. (2007). A Basic Course in Probability Theory. Universitext.
Springer New York.
Dalalyan, A. (2005). Sharp adaptive estimation of the drift function for ergodic diffusions. Ann.
Statist., 33(6):2507–2528.
Dalalyan, A. S. and Kutoyants, Y. A. (2002). Asymptotically efficient trend coefficient estimation
for ergodic diffusion. Math. Methods Statist., 11(4):402–427 (2003).
Ghosal, S., Ghosh, J. K., and van der Vaart, A. W. (2000). Convergence rates of posterior distributions. Ann. Statist., 28(2):500–531.
Ghosal, S. and van der Vaart, A. W. (2007). Convergence rates of posterior distributions for noniid
observations. Ann. Statist., 35(1):192–223.
Giné, E. and Nickl, R. (2011). Rates of contraction for posterior distributions in L r -metrics, 1 ≤
r ≤ ∞. Ann. Statist., 39(6):2883–2911.
Giné, E. and Nickl, R. (2016). Mathematical foundations of infinite-dimensional statistical models. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University
Press.
26
Hindriks, R. (2011). Empirical dynamics of neuronal rhythms. PhD thesis, Vrije Universiteit Amsterdam.
Karatzas, I. and Shreve, S. E. (1991). Brownian motion and stochastic calculus, volume 113 of
Graduate Texts in Mathematics. Springer-Verlag, New York, second edition.
Knapik, B. and Salomond, J.-B. (2014). A general approach to posterior contraction in nonparametric inverse problems. Bernoulli.
Knapik, B. T., van der Vaart, A. W., and van Zanten, J. H. (2011). Bayesian inverse problems with
Gaussian priors. Ann. Statist., 39(5):2626–2657.
Kutoyants, Y. A. (2004). Statistical inference for ergodic diffusion processes. Springer, New York.
Papaspiliopoulos, O., Pokern, Y., Roberts, G. O., and Stuart, A. M. (2012). Nonparametric estimation of diffusions: a differential equations approach. Biometrika, 99(3):511.
Pokern, Y. (2007). Fitting Stochastic Differential Equations to Molecular Dynamics Data. PhD
thesis, University of Warwick.
Pokern, Y., Stuart, A. M., and van Zanten, J. H. (2013). Posterior consistency via precision operators for Bayesian nonparametric drift estimation in SDEs. Stochastic Processes and their
Applications, 123(2):603 – 628.
Schauer, M. and van Zanten, J. H. (2017). Uniform central limit theorems for additive functionals
of diffusions on the circle. In preparation.
Shen, W. and Ghosal, S. (2015). Adaptive Bayesian procedures using random series priors. Scandinavian Journal of Statistics, 42(4):1194–1213.
Spokoiny, V. G. (2000). Adaptive drift estimation for nonparametric diffusion model. Ann. Statist.,
28(3):815–836.
Strauch, C. (2015). Sharp adaptive drift estimation for ergodic diffusions: the multivariate case.
Stochastic Process. Appl., 125(7):2562–2602.
van der Meulen, F. H., Schauer, M., and van Zanten, J. H. (2014). Reversible jump MCMC for
nonparametric drift estimation for diffusion processes. Comput. Statist. Data Anal., 71:615–
632.
van der Meulen, F. H., van der Vaart, A. W., and van Zanten, J. H. (2006). Convergence rates of
posterior distributions for Brownian semimartingale models. Bernoulli, 12(5):863–888.
van der Vaart, A. W. and van Zanten, J. H. (2008). Rates of contraction of posterior distributions
based on Gaussian process priors. Ann. Statist., 36(3):1435–1463.
van Waaij, J. and van Zanten, H. (2016). Gaussian process methods for one-dimensional diffusions: Optimal rates and adaptation. Electron. J. Statist., 10(1):628–645.
van Zanten, J. H. (2013). Nonparametric Bayesian methods for one-dimensional diffusion models. Mathematical biosciences, 243(2):215–222.
27
| 10math.ST
|
Work in progress by Shie Mannor, Vianney Perchet and Gilles Stoltz (2016)
Approachability in Unknown Games:
Online Learning Meets Multi-Objective Optimization
Shie Mannor
[email protected]
Israel Institute of Technology (Technion), Faculty of Electrical Engineering, 32000 Haifa, Israel
Vianney Perchet
arXiv:1402.2043v2 [stat.ML] 17 Jun 2016
[email protected]
ENSAE ParisTech, 3 avenue Pierre Larousse, 92240 Malakoff, France
Gilles Stoltz
[email protected]
GREGHEC: HEC Paris — CNRS, 1 rue de la Libération, 78351 Jouy-en-Josas, France
Abstract
In the standard setting of approachability there are two players and a target set. The
players play repeatedly a known vector-valued game where the first player wants to have
the average vector-valued payoff converge to the target set which the other player tries to
exclude it from this set. We revisit this setting in the spirit of online learning and do not
assume that the first player knows the game structure: she receives an arbitrary vectorvalued reward vector at every round. She wishes to approach the smallest (“best”) possible
set given the observed average payoffs in hindsight. This extension of the standard setting
has implications even when the original target set is not approachable and when it is not
obvious which expansion of it should be approached instead. We show that it is impossible,
in general, to approach the best target set in hindsight and propose achievable though
ambitious alternative goals. We further propose a concrete strategy to approach these goals.
Our method does not require projection onto a target set and amounts to switching between
scalar regret minimization algorithms that are performed in episodes. Applications to global
cost minimization and to approachability under sample path constraints are considered.
Keywords: Approachability, online learning, multi-objective optimization
1. Introduction
The approachability theory of Blackwell (1956) is arguably the most general approach available so far for online multi-objective optimization and it has received significant attention
recently in the learning community (see, e.g., Abernethy et al., 2011, and the references
therein). In the standard setting of approachability there are two players, a vector-valued
payoff function, and a target set. The players play a repeated vector-valued game where
the first player wants the average vector-valued payoff (representing the states in which the
different objectives are) to converge to the target set (representing the admissible values
for the said states), which the opponent tries to exclude. The target set is prescribed a
priori before the game starts and the aim of the decision-maker is that the average reward
be asymptotically inside the target set.
A theory of approachability in unknown games (i.e., for arbitrary vector-valued
bandit problems). The analysis in approachability has been limited to date to cases where
some underlying structure of the problem is known, namely the vector payoff function r
c (2016) by Mannor, Perchet, Stoltz.
Mannor, Perchet, Stoltz
(and some signalling structure if the obtained payoffs are not observed). We consider the
case of “unknown games” where only vector-valued rewards are observed and there is no a
priori assumption on what can and cannot be obtained. In particular, we do not assume
that there is some underlying game structure we can exploit. In our model at each round,
for every action of the decision maker there is a vector-valued reward that is only assumed
to be arbitrary. The minimization of regret could be extended to this setting (see, e.g.,
Cesa-Bianchi and Lugosi, 2006, Sections 7.5 and 7.10). And we know that the minimization
of regret is a special case of approachability. Hence our motivation question: can a theory
of approachability be developed for unknown games?
One might wonder if it is possible to treat an unknown game as a known game with a
very large class of actions and then use approachability. While such lifting is possible in
principle, it would lead to unreasonable time and memory complexity as the dimensionality
of the problem will explode.
In such unknown games, the decision maker does not try to approach a pre-specified
target set, but rather tries to approach the best (smallest) target set given the observed
vector-valued rewards. Defining a goal in terms of the actual rewards is standard in online
learning, but has not been pursued (with a few exceptions listed below) in the multiobjective optimization community.
A theory of smallest approachable set in insight. Even in known games it may
happen that no pre-specified target set is given, e.g., when the natural target set is not
approachable. Typical relaxations are then to consider uniform expansions of this natural
target set or its convex hull. Can we do better? To answer this question, another property
of regret minimization is our source of inspiration. The definition of a no-regret strategy
(see, e.g., Cesa-Bianchi and Lugosi, 2006) is that its performance is asymptotically as good
as the best constant strategy, i.e., the strategy that selects at each stage the same mixed
action. Another way to formulate this claim is that a no-regret strategy performs (almost)
as well as the best mixed action in hindsight. In the approachability scenario, this question
can be translated into the existence of a strategy that approaches the smallest approachable
set for a mixed action in hindsight. If the answer is negative (and, unfortunately, it is) the
next question is to define a weaker aim that would still be more ambitious than the typical
relaxations considered.
Short literature review. Our approach generalizes several existing works. Our proposed
strategy can be used for standard approachability in all the cases where the desired target
set is not approachable and where one wonders what the aim should be. We illustrate this
on the problems of global costs introduced by Even-Dar et al. (2009) and of approachability
with sample path constraints as described in the special case of regret minimization by
Mannor et al. (2009).
The algorithm we present does not require projection which is the Achilles’ heel of many
approachability-based schemes (it does so similarly to Bernstein and Shimkin, 2015). Our
approach is also strictly more general and more ambitious than one recently considered
by Azar et al. (2014). An extensive comparison to the results by Bernstein and Shimkin
(2015) and Azar et al. (2014) is offered in Section 5.2.
2
Approachability in Unknown Games
Outline. This article consists of four parts of about equal lengths. We first define the
problem of approachability in unknown games and link it to the standard setting of approachability (Section 2).
We then discuss what are the reasonable target sets to consider (Sections 3 and 4).
Section 3 shows by means of two examples that the best-in-hindsight expansion cannot be
achieved while its convexification can be attained but is not ambitious enough. Section 4 introduces a general class of achievable and ambitious enough targets: a sort of convexification
of some individual-response-based target set.
The third part of the paper (Section 5) exhibits concrete and computationally efficient
algorithms to achieve the goals discussed in the first part of the paper. The general strategy
of Section 5 amounts to playing a (standard) regret minimization in blocks and modifying
the direction as needed; its performance and merits are then studied in detail with respect
to the literature mentioned above. It bears some resemblance with the approach developed
by Abernethy et al. (2011).
Last but not least, the fourth part of the paper revisits two important problems, for
which dedicated methods were created and dedicated articles were written: regret minimization with global cost functions, and online learning with sample path constraints (Section 6).
We show that our general strategy has stronger performance guarantees in these problems
than the ad hoc strategies that had been constructed by the literature.
2. Setup (“unknown games”), notation, and aim
The setting is the one of (classical) approachability, that is, vector payoffs are considered.
The difference lies in the aim. In (classical) approachability theory, the average rT of the
obtained vector payoffs should converge asymptotically to some target set C, which can be
known to be approachable based on the existence and knowledge of the payoff function r.
In our setting, we do not know whether C is approachable because there is no underlying
payoff function. We then ask for convergence to some α–expansion of C, where α should be
as small as possible.
Setting: unknown game with vectors of vector payoffs. The following game is
repeatedly played between two players, who will be called respectively the decision-maker
(or first player) and the opponent (or second player). Vector payoffs in Rd , where d > 1,
will be considered. The first player has finitely many actions whose set we denote by
A = {1, . . . , A}. We assume A > 2 throughout the paper, to avoid trivialities. The
opponent chooses at each round t ∈ {1, 2, . . .} a vector mt = (mt,a )a∈A of vector payoffs
mt,a ∈ Rd . We impose the restriction that these vectors mt lie in a convex and bounded set
K of (Rd )A . The first player picks at each round t an action at ∈ A, possibly at random
according to some mixed action xt = (xt,a )a∈A ; we denote by ∆(A) the set of all such mixed
actions. She then receives mt,at as a vector payoff. We can also assume that mt,at is the
only feedback she gets on mt and that she does not see the other components of mt than
the one she chose. (This is called bandit monitoring but can and will be relaxed to a full
monitoring as we explain below.)
Remark 1 We will not assume that the first player knows K (or any bound on the maximal
norm of its elements); put differently, the scaling of the problem is unknown.
3
Mannor, Perchet, Stoltz
The terminology of “unknown game” was introduced in the machine learning literature,
see Cesa-Bianchi and Lugosi (2006, Sections 7.5 and 7.11) for a survey. A game is unknown
(to the decision-maker) when she not only does not observe the vector payoffs she would have
received has she chosen a different pure action (bandit monitoring) but also when she does
not even know the underlying structure of the game, if any such structure exists. Section 2.2
will make the latter point clear by explaining how the classical setting of approachability
introduced by Blackwell (1956) is a particular case of the setting described above: some
payoff function r exists therein and the decision-maker knows r. The strategy proposed
by Blackwell (1956) crucially relies on the knowledge of r. In our setting, r is unknown and
even worse, might not even exist. Section 2.3 (and Section A) will recall how a particular
case of approachability known as minimization of the regret could be dealt with for unknown
games.
Formulation of the approachability aim. The decision-maker is interested in controlling her average payoff
T
1X
reT =
mt,at .
T
t=1
She wants it to approach an as small as possible neighborhood of a given target set C, which
we assume to be closed. This concept of neighborhood could be formulated in terms of a
general filtration (see Remark 2 below); for the sake of concreteness we resort rather to
expansions of a base set C in some `p –norm, which we denote by k · k, for 0 < p 6 +∞.
Formally, we denote by Cα the closed α–expansion in `p –norm of C:
Cα = c ∈ Rd : ∃ c0 ∈ C s.t. kc − c0 kp 6 α = c ∈ Rd : dp (c, C) 6 α .
Here and in the sequel, dp ( · , S) denotes the distance in `p –norm to a set S.
As is traditional in the literature of approachability and regret minimization, we consider
the smallest set that would have been approachable in hindsight, that is, had the averages
of the vectors of vector payoffs be known in advance:
mT =
T
1X
mt ,
T
whose components equal
t=1
∀ a ∈ A,
mT,a =
T
1X
mt,a .
T
t=1
This notion of “smallest set” is somewhat tricky and the first part of this article will be
devoted to discuss it. The model we will consider is the following one. We fix a target
function ϕ : K → [0, +∞); it takes mT as argument. (Section 4 will indicate reasonable
such choices of ϕ.) It associates with it the ϕ(mT )–expansion of C. Our aim is then to
ensure the almost-sure convergence
dp reT , Cϕ(mT ) −→ 0
as T → ∞ .
As in the definition of classic approachability, uniformity will be required with respect to
the strategies of the opponent: the decision-maker should construct strategies such that for
all ε > 0, there exists a time Tε such that for all strategies of the opponent, with probability
at least 1 − ε,
sup dp reT , Cϕ(mT ) 6 ε .
T >Tε
4
Approachability in Unknown Games
Remark 2 More general filtrations α ∈ [0, +∞) 7→ Cα could have been considered than
expansions in some norm. By “filtration” we mean that Cα ⊆ Cα0 for all a 6 α0 . For
instance, if 0 ∈ C, one could have considered shrinkages and blow-ups, that is, C0 = {0} and
Cα = α C for α > 0. Or, given some compact set B with non-empty interior, Cα = C + αB
for α > 0. But for the sake of clarity and simplicity, we restrict the exposition to the more
concrete case of expansions of a base set C in some `p –norm.
Summary: the two sources of unknowness. As will become clearer in the concrete
examples presented in Section 6, not only the structure of the game is unknown and might
even not exist (first source of unknownness) but also the target is unknown. This second
source arises also in known games, in the following cases: when some natural target (e.g.,
some best-in-hindsight target) is proven to be unachievable or when some feasible target
is not ambitious enough (e.g., the least approachable uniform expansion of C as will be
discussed in Section 5.2.5). What to aim for, then? Convex relaxations are often considered
more manageable and ambitious enough targets; but we will show that they can be improved
upon in general.
See the paragraph “Discussion” on page 26 for more details on these two sources of
unknownness in the concrete example of global costs.
2.1. Two classical relaxations: mixed actions and full monitoring
We present two extremely classical relaxations of the general setting described above. They
come at no cost but simplify the exposition of our general theory.
The decision-maker can play mixed actions. First, because of martingale convergence results, for instance, the Hoeffding-Azuma inequality, controlling reT is equivalent to
controlling the averages rT of the conditionally expected payoffs rt , where
rt = xt
mt =
X
xt,a mt,a
and
rT =
T
T
1X
1X
rt =
xt
T
T
t=1
a∈A
mt .
t=1
Indeed, the boundedness of K and a component-by-component application of the said inequality ensure that there exists a constant C such that for all ε > 0, for all T > 1, for all
strategies of the opponent, with probability at least 1 − ε,
r
w
w
wreT − rT w 6 C ln(d/ε) .
p
T
Given ε > 0, we use these inequalities each with ε replaced by ε/T 2 : a union bound entails
that choosing Tε > 2 sufficiently large so that
r
ln(dT 2 /ε)
sup C
6 ε,
T
T >Tε
we then have, for all strategies of the opponent, with probability at least 1 − ε,
w
w
sup wreT − rT wp 6 ε .
T >Tε
(1)
Therefore, we may focus on rT instead of reT in the sequel and consider equivalently the
aim (3) discussed below.
5
Mannor, Perchet, Stoltz
The decision-maker can enjoy a full monitoring. Second, the bandit-monitoring
assumption can be relaxed to a full monitoring, at least under some regularity assumptions,
e.g., uniform continuity of the target function ϕ.
Indeed, we assumed that the decision-maker only gets to observe mt,at after choosing the
component at ∈ A of mt . However, standard estimation techniques presented by Auer et al.
(2002) and Mertens et al. (1994, Sections V.5.a and VI.6) provide accurate and unbiased
estimators m
b t of the whole vectors mt , at least in the case when the latter only depends on
what happened in the past and on the opponent’s strategy but not on the decision-maker’s
present1 choice of an action at . The components of these estimators m
b t equal, for a ∈ A,
m
b t,a =
mt,a
I
,
xt,a {a=at }
(2)
with the constraint on mixed actions that xt,a > γt for all t > 1. The decision-maker should
then base her decisions and apply her strategy on m
b t , and eventually choose as a mixed
action the convex combination of the mixed action she would have freely chosen based on
the m
b t , with weight 1 − γt , and of the uniform distribution, with weight γt .
Indeed, by the Hoeffding-Azuma inequality, the averages of the vector payoffs and of
the vectors of vector payoffs based respectively on the mt,a and on the m
b t,a , as well as the
corresponding average payoffs obtained by the decision-maker, differ by something of the
order of
v
u T
T
u
X
X
p
1 t
1
+
γt ln(1/ε)
2
T
γ
t=1 t
t=1
for each T with probability at least 1 − ε, and uniformly over the opponent’s strategies.
These differences vanish as T → ∞, e.g., at a T −2/3 rate when the γt are of the order of
t−1/3 . A treatment similar to the one performed to obtain (1) can also be applied to obtain
statements with uniformities both with respect to time T > Tε and to the strategies of
opponent.
Because our aim involves the average payoffs mT via the target function ϕ as in Cϕ(mT ) ,
we require the uniform continuity of ϕ for technical reasons, i.e., to carry over the negligible differences between the average payoffs and their estimation in the approachability
aim. (This assumption of uniform continuity can easily be dropped based on the result of
Theorem 11; details are omitted.)
Conclusion: approachability aim. The decision-maker, enjoying a full monitoring,
should construct a strategy such that almost surely and uniformly over the opponent’s
strategies,
dp rT , Cϕ(mT ) −→ 0
as T → ∞ ;
(3)
that is, for all ε > 0, there exists Tε > 0 such that for all strategies of the opponent, with
probability at least 1 − ε,
sup dp rT , Cϕ(mT ) 6 ε .
T >Tε
1. However, such a dependency can still be dealt with in some cases, see, e.g., the case of regret minimization
in Section 2.3 (and Section A): when the dependency on the decision-maker’s present action at comes
only through an additive term equal to the obtained payoff, which is known.
6
Approachability in Unknown Games
We note that we will often be able to provide stronger, uniform and deterministic controls,
of the form: there exists a function ζ : {1, 2, 3, . . .} → [0, +∞) such that ζ(T ) → 0 and for
all strategies of the opponent,
dp rT , Cϕ(mT ) 6 ζ(T ) .
To conclude this section, we point out again that the two relaxations considered come at
no cost in the generality of setting: they are only intended to simplify and clarify the
exposition. Full details of this standard reduction from the case of bandit monitoring to full
monitoring are omitted because they are classical, though lengthy and technical, to expose.
2.2. Link with approachability in known finite games
We link here our general setting above with the classical setting considered by Blackwell
(1956). Therein the decision-maker and the opponent have finite sets of actions A and B,
and choose at each round t respective pure actions at ∈ A and bt ∈ B, possibly at random
according to some mixed actions xt = (xt,a )a∈A ∈ ∆(A) and yt = (yt,b )b∈B ∈ ∆(B). A
payoff function r : A × B → Rd is given and is multilinearly extended to ∆(A) × ∆(B)
according to
XX
∀ (x, y) ∈ ∆(A) × ∆(B),
r(x, y) =
xa yb r(a, b) .
a∈A b∈B
From the decision-maker viewpoint, the game takes place as if the opponent was choosing
at each round the vector of vector payoffs
mt = r( · , bt ) = r(a, bt ) a∈A .
A target set C is to be approached, that is, the convergence
reT =
T
1X
r(at , bt ) −→ C
T
a.s.
t=1
should hold uniformly over the opponent’s strategies. (Of course, as recalled above, we can
equivalently require the uniform convergence of rT to C.)
A necessary and sufficient condition for this when C is closed and convex is that for all
y ∈ ∆(B), there exists some x ∈ ∆(A) such that r(x, y) ∈ C. Of course, this condition,
called the dual condition for approachability, is not always met. However, in view of the
dual condition, the least approachable α–expansion in `p –norm of such a non-empty, closed,
and convex set C is given by
αunif = max min dp r(x, y), C .
(4)
y∈∆(B) x∈∆(A)
Approaching Cαunif corresponds to considering the constant target function ϕ ≡ αunif in (3).
Better (uniformly smaller) choices of target functions exist, as will be discussed in Section 5.2.5. This will be put in correspondence therein with what is called “opportunistic
approachability.”
7
Mannor, Perchet, Stoltz
The knowledge of r is crucial (a first strategy). The general strategies used to
approach C (or Cαunif when C is not approachable and p = 2) rely crucially on the knowledge
of r.
Indeed, the original strategy of Blackwell (1956) proceeds as follows: at round t + 1, it
first computes the projection e
ct of ret onto C. Then it picks at+1 at random according to a
mixed action xt+1 such that
ret − e
ct , r(xt+1 , y) − e
ct 6 0 .
∀ y ∈ ∆(B),
(5)
When C is approachable, such a mixed action always exists; one can take, for instance,
xt+1 ∈ arg min max ret − e
ct , r(x, y) .
x∈∆(A) y∈∆(B)
In general, the strategy thus heavily depends on the knowledge of r.
When C is not approachable and p = 2, the set Cαunif is the target and the choice right
above is still suitable to approach Cαunif in `2 –norm. Indeed, the projection det of ret onto
Cαunif is such that ret − det is proportional to ret − e
ct , thus
arg min max ret − e
ct , r(x, y) = arg min max ret − det , r(x, y) .
x∈∆(A) y∈∆(B)
x∈∆(A) y∈∆(B)
The knowledge of r is crucial (a second strategy). There are other strategies to
perform approachability in known finite games, though the one described above may be the
most popular one. For instance, Bernstein and Shimkin (2015) propose a strategy based on
the√dual condition for approachability, which still performs approachability at the optimal
1/ T rate. We discuss it in greater details and generalize it to the case of unknown games
in Section 5.2. For now, we describe it shortly only to show how heavily it relies on the game
r being known. Assume that C is approachable. At round t = 1, choose an arbitrary mixed
action x1 to draw a1 and choose an arbitrary mixed action ye1 ∈ B. For rounds t + 1, assume
that mixed actions ye1 , . . . , yet ∈ ∆(B) have been chosen by the decision-maker in addition to
the pure actions b1 , . . . , bt actually played
by the opponent, and that corresponding mixed
actions x
e1 , . . . , x
et such that r x
es , yes ∈ C have been chosen as well. Denoting
δt =
t
X
τ =1
r(aτ , bτ ) −
the strategy selects
xt+1 ∈ arg min max hδt , r(x, y)i
x∈∆(A) y∈∆(B)
and
t
X
τ =1
r x
eτ , yeτ ,
yet+1 ∈ arg max min hδt , r(x, y)i ,
y∈∆(B) x∈∆(A)
et+1 exists since C is
as well as x
et+1 ∈ ∆(A) such that r x
et+1 , yet+1 ∈ C, where such an x
approachable. Thus, it is crucial that the strategy knows r; however, that C be approachable
is not essential: in case it is not approachable and Cαunif is to be approached instead, it
suffices to pick
x
es ∈ arg min dp r x, yes , C ,
x∈∆(A)
so that r x
es , yes ∈ Cαunif . Any `p –norm is suitable for this argument.
8
Approachability in Unknown Games
2.3. Link with regret minimization in unknown games
The problem of regret minimization can be encompassed as an instance of approachability.
For the sake of completeness, we recall in Appendix A why the knowledge of the payoff
structure is not crucial for this very specific problem. This, of course, is not the case at all
for general approachability problems.
3. Two toy examples to develop some intuition
The examples presented below will serve as guides to determine suitable target functions
ϕ : K → [0, +∞), that is, target functions for which the convergence (3) can be guaranteed
and that are ambitious (small) enough, in a sense that will be made formal in the next
section.
Example 1: minimize several costs at a time. The following example is a toy modeling of a case when the first player has to perform several tasks simultaneously and incurs
a loss (or a cost) for each of them; we assume that her overall loss is the worst (the largest)
of the losses thus suffered.
For simplicity, and because it will be enough for our purpose, we will assume that the
decision-maker only has two actions, that is, A = {1, 2}, while the opponent is restricted to
only pick convex combinations of the following vectors of vector payoffs:
and
†
m =
m†a a∈{1,2}
]
m]a a∈{1,2}
m =
∈R
2×2
2×2
∈R
with
m†1
with
m]1
=
3
4
=
4
3
and
m†2
and
m]2
=
0
5
,
=
5
0
.
The opponent’s actions can thus be indexed by ν ∈ [0, 1], where the latter corresponds to
the vector of vectors
m(ν) = νm† + (1 − ν)m] ∈ R2×2 .
The base target set C is the negative orthant C = (−∞, 0]2 and its α–expansions in
the supremum norm (p = +∞) are Cα = (−∞, α]2 . A graphical representation of these
expansions and of the vectors m† and m] is provided in Figure 1.
Example 2: control absolute values. In this example, the decision-maker still has only
two actions, A = {1, 2}, and gets scalar rewards, i.e., d = 1. The aim is to minimize the
absolute value of the average payoff, i.e., to control the latter from above and from below
(for instance, because these payoffs measure deviations in either direction from a desired
situation).
Formally, the opponent chooses vectors mt ∈ R2 , which we assume to actually lie in
K = [−1, 1]2 . The product
is then simply the standard inner product over R2 . We
consider C = {0} as a base target set to be approached. Its expansions (in any `p –norm)
are C = [−α, α], for α > 0.
9
5 − 5ν
?
ϕ (ν) = min αx (ν) = min α0 (ν), α1 (ν) =
x∈[0,1]
5ν
3 + ν
Mannor, Perchet, Stoltz
if ν ∈ [1/4, 1/2],
if ν ∈ [1/2, 3/4],
if ν ∈ [3/4, 1].
This last set of equalities can be seen to hold true by contemplating Figure 2.
m†
5
αx
4
ϕ?
m]
C
C2.5
2.5
m]
m†
C4
Figure 1: Graphical
of m† and m]ofand
different
.
Figurerepresentation
2: Graphical representation
m† of
and
m] , andexpansions
of differentCαexpansions
Cα (left); graph
?
of the functions ϕ (bold solid line), of α0 and α1 (dotted lines), and of some αx , fo
x ∈ [0, 1] (thin solid line).
3.1. The smallest set in hindsight cannot be achieved in general
We denote by ϕ? : K → [0, +∞) the function that associates with a vector of vector payoffs
Proof
[of Lemma
4] Assume
contradiction
that ϕ?combination
is achievableofinitsthe above example an
m ∈ K the index of the
smallest
`p –expansion
of Cbycontaining
a convex
consider any strategy of the decision maker to do so, which we denote by σ. Imagine in a first tim
components:
that Nature chooses, at every stage t > 1,othe vectors mt = m† , which amounts to playing νt =
n
?
ϕ (m) = min Given
α > 0that
: the
∃ x average
∈ ∆(A)ofs.t.
m ∈ CναT =
= 1 min
dp (x
m, C)
(6) ϕ? (1) = 4, the ai
thexνt equals
and that
its image
by, ϕ? equals
x∈∆(A)
is then to converge to C4 . But this can only be guaranteed if the average of the chosen xt converge
the infimum being achieved
byis,continuity.
This
defines
function
x? : large, integer Tε such that
to 1. That
given ε > 0,
there
exists asome,
possibly
∀ m ∈ K,
∃ x? (m) ∈ ∆(A) :
x? (m) m ∈3 Cϕ? (m) .
rTε −
6 ε.
(21
4
?
∞
Lemma 3 In Examples 1 and 2, the convergence (3) cannot be achieved for ϕ against all
strategies of the opponent.
Now, consider a second scenario. During the first Tε stages, Nature chooses the vectors mt = m
By construction, as σ is fixed, (21) is ensured. Now, for the next Tε stages, for Tε + 1 6 t 6 2T
The proofs (located in Appendix B.1) reveal thatthe difficulty ]in (3) is that it should
assume that Nature chooses the vectors mt = m , and denote by γ ∈ [0, 1] the average of th
hold along a whole path, while the value of ϕ? mT can change
more rapidly than the
x ∈ [0, 1] selected by σ. The average of the played vectors is m2Tε corresponds to ν = 1/2, whos
average payoff vectors trT do. ?
image by ϕ equals ϕ? (1/2) = 5/2. Therefore the target set is C2.5 . However, by definition of γ
They will formalize the following proof scheme. To accommodate a first situation,
which
we have
lasts a large number T 0 of stages, the decision-maker should play in a given way; but then,
is she
2Tε from where the decision-maker
X
the opponent changes drastically his 1strategy1 and
1
1
4
5
]
rTε +
x0 t. The
m situation
= rTε +is repeated.
γ
+ (1 − γ)
ε =
cannot catch up and is far fromr2T
the
target
at
stage
2T
3
0
2
2
2
2
t=Tε +1
1
1
5−γ
3.2. A concave relaxation is not ambitious enough
= rTε +
,
3γ
2
2
A classical relaxation in the game-theory literature for unachievable targets (see, e.g., how
Mannor et al., 2009 proceed) is to consider concavifications. Can the convergence (3) hold
23
with ϕ = cav[ϕ? ], the concavification of ϕ? ? The latter is defined as the least concave
function K → [0, +∞] above ϕ? . The next section will show that it is indeed always the
case but we illustrate on our examples why such a goal is not ambitious enough. (The proof
of the lemma below can be found in Appendix B.2.)
Lemma 4 In Examples 1 and 2, the decision-maker has a mixed action (x, 1 − x) that
she can play at each round to ensure the convergence (3) for a target function ϕx that is
uniformly smaller than ϕ? , and even strictly smaller at some points.
10
Approachability in Unknown Games
4. A general class of ambitious enough target functions
The previous section showed on examples that the best-in-hindsight target function ϕ? was
too ambitious a goal while its concavification cav[ϕ? ] seemed not ambitious enough. In this
section, based on the intuition given by the formula for concavification, we provide a whole
class of achievable target functions, relying on a parameter: a response function Ψ.
In the definition below, by uniformity over strategies of the opponent player, we mean
the uniform convergence stated right after (3). We denote by Gϕ the graph of the set-valued
mapping m ∈ K 7→ Cϕ(m) :
n
o
Gϕ = (m, r) ∈ K × Rd s.t. r ∈ Cϕ(m) ⊂ (Rd )A × Rd .
Definition 5 A continuous target function ϕ : K → [0, +∞) is achievable if the decisionmaker has a strategy ensuring that, uniformly over all strategies of the opponent player,
dp rT , Cϕ(mT ) −→ 0
as T → ∞ .
(7)
More generally, a (possibly non-continuous) target function ϕ : K → [0, +∞) is achievable
if Gϕ is approachable for the game with payoff function (x, m) ∈ ∆(A) × K 7→ (m, x m),
that is, if uniformly over all strategies of the opponent player,
(mT , rT ) −→ Gϕ
as T → ∞ .
(8)
We always have that (7) entails (8), with or without continuity of ϕ. The condition (8)
is however less restrictive in general and it is useful in the case of non-continuous target
functions (e.g., to avoid lack of convergence due to errors at early stages). But for continuous
target functions ϕ, the two definitions (7) and (8) are equivalent. We prove these two facts
in Section C.1 in the appendix.
The defining equalities (6) for ϕ? show that this function is continuous (it is even a
Lipschitz function with constant 1 in the `p –norm). We already showed in Section 3.1 that
the target function ϕ? is not achievable in general.
To be able to compare target functions, we consider the following definition and notation.
Definition 6 A target function ϕ : K → [0, +∞) is strictly smaller than another target
function ϕ0 if ϕ 6 ϕ0 and there exists m ∈ K with ϕ(m) < ϕ0 (m). We denote this fact by
ϕ ≺ ϕ0 .
For instance, in Lemma 4, we had ϕx ≺ ϕ? .
4.1. The target function cav[ϕ? ] is always achievable
We show below that the target function cav[ϕ? ] is always achievable... But of course,
Section 3.2 already showed that cav[ϕ? ] is not ambitious enough: in Examples 1 and 2,
there exist easy-to-construct achievable target functions ϕ with ϕ ≺ cav[ϕ? ]. We however
provide here a general study of the achievability of cav[ϕ? ] as it sheds light on how to achieve
more ambitious target functions.
So, we now only ask for convergence of (mT , rT ) to the convex hull of Gϕ? , not to
Gϕ? itself. Indeed, this convex hull is exactly the graph Gcav[ϕ? ] , where cav[ϕ? ] is the
11
Mannor, Perchet, Stoltz
concavification of ϕ? , defined as the least concave function K → [0, +∞] above ϕ? . Its
variational expression reads
X
X
cav[ϕ? ](m) = sup
λi ϕ? (mi ) : N > 1 and
λi mi = m ,
(9)
i6N
i6N
for all m ∈ K, where the supremum is over all finite convex decompositions of m as elements
of K (i.e., the mi belong to K and the λi factors are nonnegative and sum up to 1). By a
theorem of Fenchel and Bunt (see Hiriart-Urruty and Lemaréchal, 2001, Theorem 1.3.7) we
could actually further impose that 1 6 N 6 dA + 1. In general, cav[ϕ? ] is not continuous;
it is however so when, e.g., K is a polytope.
Lemma 7 The target function cav[ϕ? ] is always achievable.
Proof sketch; when K is known When the decision-maker knows K (and only in this
case), she can compute cav[ϕ? ] and its graph Gcav[ϕ? ] . As indicated after Definition 5,
it suffices to show that the convex set Gcav[ϕ? ] is approachable for the game with payoffs
(x, m) ∈ ∆(A) × K 7→ (m, x m); the decision-maker should then play any strategy approaching Gcav[ϕ? ] . Note that ϕ? is continuous, that Gϕ? is thus a closed set, and that
Gcav[ϕ? ] is a closed convex set containing Gϕ? . Now, the characterization of approachability
by Blackwell (1956) for closed convex sets (recalled already in Section 2.2) states that for
all m ∈ K, there should exist x ∈ ∆(A) such
that (m, x m) ∈ Gcav[ϕ? ] . But by the
definition (6), we even have m, x? (m) m ∈ Gϕ? , which concludes the proof.
We only proved Lemma 7 under the assumption that the decision-maker knows K, a
restriction which we are however not ready to consider as indicated in Remark 1. Indeed,
she needs to know K to compute Gϕ? and the needed projections onto this set to implement
Blackwell’s approachability strategy. Some other approachability strategies may not require
this knowledge, e.g., a generalized version of the one of Bernstein and Shimkin (2015) based
on the dual condition for approachability (see Section 2.2 for their original version, see
Section 5.2.4 for our generalization).
But anyway, we chose not to go into these details now because at least in the case
when C is convex, Lemma 7 will anyway follow from Lemmas 8 and 10 (or Theorem 11)
below, which are proved independently and wherein no knowledge of K is assumed2 . Even
better, they prove the strongest notion of convergence (7) of Definition 5, irrespectively of
the continuity or lack of continuity of cav[ϕ? ].
4.2. An example of a more ambitious target function
By (6) we can rewrite (9) as
X
cav[ϕ? ](m) = sup
λi dp x? (mi )
i6N
mi , C : N > 1 and
X
i6N
λi mi = m
.
2. Indeed, the functions x? and ϕ? at hand therein are independent of K as they are defined for each m ∈ K
as the solutions of some optimization program that only depends on this specific m and on C, but not
on K.
12
Approachability in Unknown Games
Now, whenever C is convex, the function dp ( · , C is convex as well over Rd ; see, e.g., Boyd
?
and Vandenberghe (2004, Example 3.16). Therefore, denoting by ϕx the function defined
as
!
X
X
?
ϕx (m) = sup dp
λi x? (mi ) mi , C : N > 1 and
λi mi = m
(10)
i6N
i6N
?
for all m ∈ K, we have ϕx 6 cav[ϕ? ]. The two examples considered in Section 3.2 actually
show that this inequality can be strict at some points. We summarize these facts in the
?
lemma below, whose proof can be found in Appendix B.3. That ϕx is achievable is a special
?
case of Lemma 10 stated in the next subsection, where a class generalizing the form of ϕx
will be discussed.
?
Lemma 8 The inequality ϕx 6 cav[ϕ? ] always holds when C is convex. For Examples 1
?
and 2, we even have ϕx ≺ cav[ϕ? ].
4.3. A general class of achievable target functions
The class is formulated by generalizing the definition (10): we call response function any
function Ψ : K → ∆(A) and we replace in (10) the specific response function x? by any
response function Ψ.
Definition 9 The target function ϕΨ based on the response function Ψ is defined, for all
m ∈ K, as
X
X
ϕΨ (m) = sup dp
λi Ψ(mi ) mi , C : N > 1 and
λi mi = m .
(11)
i6N
i6N
Lemma 10 For all response functions Ψ, the target functions ϕΨ are achievable.
The lemma actually follows from Theorem 11 below, which provides an explicit and efficient strategy to achieve any ϕΨ , in the stronger sense (7) irrespectively of the continuity
or lack of continuity of ϕΨ . For now, we provide a sketch of proof (under an additional
assumption of Lipschitzness for Ψ) based on calibration, because it further explains the
intuition behind (11). It also advocates why the ϕΨ functions are reasonable targets: resorting to some auxiliary calibrated strategy outputting accurate predictions m̂t (in the
sense of calibration) of the vectors mt almost amounts to knowing in advance the mt . And
with such a knowledge, what can we get?
Proof sketch; when Ψ is a Lipschitz function We will show below that there exists
a constant L0 ensuring the following: given any δ > 0, there exists randomized strategy of
the decision-maker such that for all ε > 0, there exists a time Tε such that for all strategies
of the opponent, with probability at least 1 − ε,
a.s.
(12)
sup dp rT , CϕΨ (mT ) 6 L0 (δ + ε)
T >Tε
13
Mannor, Perchet, Stoltz
In terms of approachability theory (see, e.g., Perchet, 2014 for a survey), this means that
GϕΨ is in particular an L0 δ–approachable set for all δ > 0, thus a 0–approachable set. But 0–
approachability and approachability are two equivalents notions (a not-so-trivial fact when
the sets at hand are not closed convex sets). That is, GϕΨ is approachable, or put differently,
ϕΨ is achievable.
Indeed, fixing δ > 0, there exists a randomized strategy picking predictions m
b t among
[j]
finitely many elements m ∈ K, where j ∈ {1, . . . , Nδ } so that the so-called calibration
score is controlled: for all ε > 0, there exists a time Tε such that for all strategies of the
opponent, with probability at least 1 − ε,
w
Nδ w
T
X
w1 X
I
sup
w
m
b t =m[j]
wT
T >Tε
j=1
t=1
w
w
w
m
b t − mt w
w
6δ+ε
a.s.;
(13)
p
see3 Foster and Vohra (1998). Now, the main strategy, based on such an auxiliary calibrated
strategy, is to play Ψ(m
b t ) at each round. The average payoff of the decision-maker is thus
rT =
T
1X
Ψ m
bt
T
mt .
t=1
We decompose it depending on the predictions m
b t made: for each j ∈ {1, . . . , Nδ }, the av[j]
erage number of times m was predicted and the average vectors of vector payoffs obtained
on the corresponding rounds equal
bj,T
λ
T
1X
=
I
m
b t =m[j]
T
and
t=1
m
b
j,T
=
PT
t=1 mt I m
b t =m[j]
PT
t=1 I m
b t =m[j]
j,T
bj,T > 0, otherwise, we take an arbitrary value for m
whenever λ
b . In particular,
mT =
Nδ
X
j=1
j,T
bj,T m
λ
b
rT =
and
Nδ
X
j=1
bj,T Ψ m[j]
λ
j,T
m
b
.
Using this convex decomposition of mT in terms of elements of K, the very definition of ϕΨ
leads to
Nδ
j,T
X
j,T
bj,T Ψ m
λ
b
m
b
∈ CϕΨ (mT ) ,
j=1
3. Actually, the latter reference only considers the case of calibrated predictions of elements in some simplex,
but it is clear from the method used in Mannor and Stoltz (2010) — a reduction to a problem of
approachability — that this can be performed for all subsets of compact sets, such as K here, with the
desired uniformity over the opponent’s strategies; see also Mannor et al. (2014, Appendix B). The result
holds for any `p –norm by equivalence of norms on vector spaces of finite dimension, even if the original
references considered the `1 or `2 –norms only.
14
Approachability in Unknown Games
hence
dp rT , CϕΨ (mT )
w
w
w
w
Nδ
j,T
X
w
j,T w
b
w
6 wrT −
λj,T Ψ m
b
m
b w
w
w
w
j=1
p
w
wX
Nδ
j,T
w
bj,T Ψ m[j] − Ψ m
b
=w
λ
w
w j=1
w
w
j,T w
m
b w
w .
w
p
We denote by Bp,max a bound on the maximal `p –norm of an element in the bounded set
K. A triangular equality shows that
w
w
wX
w
Nδ
j,T
j,T w j,T w
X
X
w Nδ
j,T w
w
w
[j]
bj,T Ψ m[j] − Ψ m
bj,T
w
w 6
b
m
b
m
b
ba w
λ
λ
Ψ
m
−
Ψ
wm
w
w
a
a
p
w j=1
w
j=1
a∈A
p
6 Bp,max
Nδ
X
j=1
bj,T
λ
X
Ψ m[j]
a∈A
j,T
m
b
−
Ψ
a
a
where ψ(m)a refers to the probability mass put on a ∈ A by ψ(m). As indicated above,
we assume for this sketch of proof that Ψ is a Lipschitz function, with Lipschitz constant L
with respect to the `1 –norm over ∆(A) and the `p –norm over K. We get
Bp,max
Nδ
X
j=1
bj,T
λ
X
a∈A
Ψ m
[j]
j,T
m
b
−
Ψ
a
a
6 Bp,max L
Nδ
X
j=1
w
w
j,T w
w [j]
b
b w
λj,T wm − m
p
w
Nδ w
T
X
w1 X
= Bp,max L
I
w
m
b t =m[j]
wT
j=1
t=1
w
w
w
m
b t − mt w .
w
p
Substituting (13), we proved (12) for L0 = Bp,max L, which concludes the proof.
4.4. Some thoughts on the optimality of target functions
The previous subsections showed that target functions of the form ϕΨ were achievable,
unlike the best-in-hindsight target function ϕ? , and that they were more ambitious than
the concavification cav[ϕ? ]. The question of their optimality can be raised — a question to
which we will not be able to answer in general. Our thoughts are gathered in Appendix D.
5. A strategy by regret minimization in blocks
In this section we exhibit a strategy to achieve the stronger notion of convergence (7) with
the target functions ϕΨ advocated in Section 4.3, irrespectively of the continuity or lack of
continuity of ϕΨ . The algorithm is efficient, as long as calls to Ψ are (a full discussion of
the complexity issues will be provided for each application studied in Section 6).
15
Mannor, Perchet, Stoltz
5.1. Description and analysis of the strategy
As in Abernethy et al. (2011), the considered strategy — see Figure 2 — relies on some
auxiliary regret-minimizing strategy R, namely, a strategy with the following property.
Assumption 1 The strategy R sequentially outputs mixed actions ut ∈ ∆(A) such that
for all ranges B > 0 (not necessarily known in advance), for all T > 1 (not necessarily
known in advance), for all sequences of vectors m0t ∈ RA of one-dimensional payoffs lying
in the bounded interval [−B, B], possibly chosen online by some opponent player, where
t = 1, . . . , T ,
T
T
X
X
√
max
u m0t 6 4B T ln A +
ut m0t .
u∈∆(A)
t=1
t=1
Note in particular that the auxiliary strategy R automatically adapts to the range B of
the payoffs and to the number of rounds T , and has a sublinear worst-case guarantee. (The
adaptation to B will be needed because K is unknown.) Such auxiliary strategies indeed
exist, for instance, the polynomially weighted average forecaster of Cesa-Bianchi
and Lugosi
√
(2003). Other ones with a possibly larger constant factor in front of the B T ln A term also
exist, for instance, exponentially weighted average strategies with learning rates carefully
tuned over time, as described by Cesa-Bianchi et al. (2007) or de Rooij et al. (2014).
For the sake of elegance (but maybe at the cost of not providing all the intuitions that
led us to this result), we only provide in Figure 2 the time-adaptive version of our strategy,
which does not need to know the time horizon T in advance. The used blocks are of
increasing lengths 1, 2, 3, . . .. Simpler versions with
√ fixed block length L would require a
tuning of L in terms of T (pick L of the order of T ) to optimize the theoretical bound.
Theorem 11 For all response functions Ψ, the strategy of Figure 2 is such that for all
T > 1, for all strategies of the opponent, there exists cT ∈ CϕΨ (mT ) ensuring
w
w
wrT − cT w 6 10 T −1/4 ln A + 3B2,max T −1/2 ,
(14)
2
where B2,max = max kmk2 is the maximal Euclidean norm of elements in K.
m∈K
In particular, denoting by κp a constant such that k · kp 6 κp k · k2 , for all T > 1 and
all strategies of the opponent,
dp rT , CϕΨ (mT ) 6 κp 10 T −1/4 ln A + 3B2,max T −1/2 .
(15)
Remark 12 With the notation of Figure 2, denoting in addition by nT the largest integer
such that nT (nT + 1)/2 6 T , by
mpart. =
1
T − nT (nT + 1)/2
T
X
mt
t=nT (nT +1)/2+1
the partial average of the vectors of vector payoffs mt obtained during the last and (nT +1)–th
block when nT (nT + 1)/2 < T (and an arbitrary element of K otherwise), we can take
!
nT
1 X
nT (nT + 1)
(k)
(k)
part.
part.
cT =
kΨ m
m + T−
Ψ m
m
.
(16)
T
2
k=1
16
Approachability in Unknown Games
Parameters: a regret-minimizing strategy R (with initial action u1 ) and a response function
Ψ : K → ∆(A)
Initialization: play x1 = u1 and observe m1 ∈ (Rd )A ; this is block n = 1
For blocks n = 2, 3 . . .,
1. compute the total discrepancy δn at the beginninga of block n (that is, till the end of
block n − 1),
n(n−1)/2
δn =
X
t=1
where
m(k)
xt
mt −
n−1
X
k=1
k
1X
=
mk(k−1)/2+t
k t=1
k Ψ m(k)
m(k) ∈ Rd ,
is the average vector of vector payoffs obtained in block k ∈ {1, . . . , n − 1};
2. run a fresh instance Rn of R for n rounds as follows: set un,1 = u1 ; then, for t = 1, . . . , n,
(a) play xn(n−1)/2+t = un,t and observe mn(n−1)/2+t ∈ (Rd )A ;
(b) feed Rn with the vector payoff m0n,t ∈ RA with components given, for a ∈ A, by
m0n,t,a = −hδn , mn(n−1)/2+t,a i ∈ R,
where h · , · i denotes the inner product in Rd ;
(c) obtain from Rn a mixed action un,t+1 .
a. Block n starts at round 1 +
n(n + 1)
n(n − 1)
, is of length n, thus lasts till round
.
2
2
Figure 2: The proposed strategy, which plays in blocks of increasing lengths 1, 2, 3, . . .
Important comments on the result. The strategy itself does not rely on the knowledge
of K, as promised in Remark 1; only its performance bound does, via the B2,max term. Also,
the convexity of C is not required. The convergence rates are independent of the ambient
dimension d.
Concerning the norms, even if the strategy and its bound (14) are based on the Euclidean
norm, the set CϕΨ (mT ) is defined in terms of the `p –norm as in (11). The constant κp exists
by equivalence of the norms on a finite-dimensional space.
Finally, we note that we obtained the uniformity requirement
stated after (3) in the
deterministic form with a function ζ where ζ(T ) = O T −1/4 .
Proof The convergence (15) follows from the bound (14) via the equivalence between `p –
and `2 –norms. That the stated cT in (16) belongs to CϕΨ (mT ) , where the latter set is defined
in terms of the `p –norm as in (11), is by construction of ϕΨ as a supremum. It thus suffices
to prove (14) with the cT defined in (16), which we do by induction.
The induction is on the index n > 1 of the blocks, and the quantities to control are the
squared Euclidean norms of the discrepancies at the end of these blocks, kδn+1 k22 . (We recall
17
Mannor, Perchet, Stoltz
that δn+1 denotes the discrepancy at the end of block n.) We have that δ2 is a difference
2
between two elements of K, thus that kδ2 k22 6 4B2,max
.
We use a self-confident approach: we consider a function β : {1, 2, . . .} → [0, +∞) to be
defined by the analysis and assume that we have proved that our strategy is such that for
some n > 1 and for all sequences of vectors of vector payoffs mt ∈ K, possibly chosen by
some opponent (i.e., for all strategies of the opponent),
w
w2
wn(n+1)/2
w
n
X
X
w
w
2
(k)
(k) w
w
kδn+1 k2 = w
xt mt −
kΨ m
m w 6 β(n) .
w t=1
w
k=1
2
2
4B2,max
.
For instance, we define β(1) =
We then study what we can guarantee for n + 2. We have
w
w2
w
w
(n+1)(n+2)/2
X
w
w
2
(n+1)
(n+1)
w
kδn+2 k2 = w
xt mt − (n + 1) Ψ m
m
wδn+1 +
w
w
w
t=n(n+1)/2+1
2
+
*
(n+1)(n+2)/2
X
m(n+1)
= kδn+1 k22 + 2 δn+1 ,
xt mt − (n + 1) Ψ m(n+1)
t=n(n+1)/2+1
w
w (n+1)(n+2)/2
X
w
+w
xt
w
wt=n(n+1)/2+1
mt − (n + 1) Ψ m(n+1)
w2
w
w
(n+1) w
m
w .
w
(17)
2
2
1)2 B2,max
,
We upper bound the two squared norms by β(n) and 4(n +
respectively. Using
(n+1)
(n+1)
, the inner product can be rewritten, with the
the short-hand notation u
=Ψ m
notation of Figure 2, as
*
+
(n+1)(n+2)/2
X
δn+1 ,
xt mt − (n + 1) Ψ m(n+1)
m(n+1)
t=n(n+1)/2+1
=−
n+1
X
t=1
un+1,t
m0n+1,t +
n+1
X
u(n+1)
m0n+1,t (18)
t=1
Now, the Cauchy–Schwarz inequality indicates that for all a and t,
p
m0n+1,t,a 6 kδn+1 k2 kmn(n+1)/2+t,a k2 6 B2,max β(n) ,
where we used again the induction hypothesis.
Assumption
1 therefore indicates that the
p
p
quantity (18) can be bounded by 4B2,max β(n) (n + 1) ln A.
Putting everything together, we have proved that the induction holds provided that
β(n + 1) is defined, for instance, as
p
p
2
β(n + 1) = β(n) + 8B2,max β(n) (n + 1) ln A + 4B2,max
(n + 1)2 .
√
2
By the lemma in Appendix C.2 (taking γ1 = 4B2,max ln A and γ2 = 4B2,max
), we thus get
first
√
2
∀ n > 1, β(n) 6 32 B2,max
(ln A) n3 ,
then
∀ n > 1, kδn+1 k2 6 B2,max 32 n3 ln A .
18
Approachability in Unknown Games
It only remains to relate the quantity at hand in (14) and (16) to the δn+1 . By separating
time till the end of the nT –block and starting from the beginning of block nT + 1 (should
the latter start strictly before T ), we get
T
1X
rT − cT =
xt
T
t=1
=
1
mt −
T
1
1
δnT +1 +
T
T
nT
X
(k)
kΨ m
k=1
T
X
t=nT (nT +1)/2+1
(k)
m
xt − Ψ mpart.
nT (nT + 1)
+ T−
2
Ψ m
part.
part.
m
mt .
The second sum contains at most nT elements, as the (nT + 1)-th regime is incomplete. A
triangular inequality thus shows that
q
1
nT
1
nT
krT − cT k2 6 kδnT +1 k2 + 2B2,max
6 B2,max 32 n3T ln A + 2B2,max
T
T
T
T
−1/4
−1/2
6 10 T
ln A + 3B2,max T
,
√
where we used the inequality nT (nTp
+ 1)/2 6 T , its implication nT 6 2T , as well as (for
√
the sake of readability) the bounds 32 × 23/2 < 10 and 2 2 < 3.
5.2. Discussion
In this section we gather comments, remarks, and pointers to the literature. We discuss
in particular the links and improvements over the concurrent (and independent) works
by Bernstein and Shimkin (2015) and Azar et al. (2014).
5.2.1. Do we have to play in blocks? Is the obtained T −1/4 rate optimal?
Our strategy proceeds in blocks, unlike the ones exhibited for the case of known games, as
the original strategy by Blackwell (1956) or the more recent one by Bernstein and Shimkin
(2015), see Section 2.2. The calibration-based strategy considered in the proof of Lemma 10
also performed some grouping, according to the finitely many possible values of the predicted
vectors of vector payoffs. This is because the target set to approach is unknown: the
decision-maker approaches a sequence of expansions of this set, where the sizes ϕΨ (mT ) of
the expansions vary depending on the sequence of realized averages mT of vectors of vector
payoffs. When an approachable target set C is given, the strategies, e.g., by Blackwell (1956)
or Bernstein and Shimkin (2015), do not need to perform any grouping.
Actually, it is easy to prove that the following quantity, which involves no grouping in
rounds, cannot be minimized in general:
krT −
where
c0T
c0T kp
T
1X
=
xt
T
T
1X
=
Ψ(mt )
T
t=1
mt .
t=1
19
T
1X
mt −
Ψ(mt )
T
t=1
mt
(19)
p
(20)
!
Mannor, Perchet, Stoltz
Indeed, consider a toy case where the mt = (gt,a )a∈A have scalar components gt,a ∈ R,
the negative orthant C = (−∞, 0] is to be approached, whose expansions
are given by
Cα = (−∞, α], for α > 0. Considering the response function Ψ (ga )a∈A ∈ arg maxa∈A ga ,
we see that (19) boils down to controlling
T
T
1 XX
1X
max
ga0 ,t ,
xt,a gt,a −
T
T
a0t ∈A t
t=1 a∈A
t=1
which is impossible4 . This is in contrast with the regret (38), which can be minimized.
The most severe issue here is not really the absolute value taken, but the fact that we
are comparing the decision-maker’s payoff to the sum of the instantaneous maxima of the
payoffs ga,t , instead of being interesting in the maximum of their sums as in (38).
So, the answer to the first question would be: yes, we have to play in blocks. Given
that, is the obtained T −1/4 rate optimal? We can answer this question in the positive
by considering the same toy case as above. With this example, the bound (14) given the
definition (16) of cT rewrites
k(k+1)/2
nT
T
T
X
X
1 XX
1 X
xt,a gt,a −
max
gt,a0k + 0 max
gt,a0k ,
T
T
a0k ∈A
an +1 ∈A
T
t=1 a∈A
k=1
t=k(k−1)/2+1
t=nT (nT −1)/2+1
which corresponds to the control (from above and from below) of what is called a (per-round)
“tracking regret” for nT shifts. This notion was introduced by Helmbold and M.Warmuth
(1998); see also Cesa-Bianchi and Lugosi (2006, Chapter 5) for a review of the results known
for tracking regret. In particular, the examples used therein to show the optimality of the
bounds (which are of the form of the one considered in Footnote 4) can be adapted in our
context, so thatpthe lower bound on tracking regret with nT shifts applies in our case: it is
of the order of nT /T , thus of T −1/4 .
In a nutshell, what we proved in these paragraphs is that if we are to ensure the convergence (3) by controlling a quantity of the form (14) and (16), then we have to proceed
in blocks and convergence cannot hold at a faster rate than T −1/4 . However, the associated
strategy is computationally efficient. Also, neither the convexity of C, nor the continuity of
ϕΨ or of Ψ are required, yet the stronger convergence (7) is achieved, not only (8).
5.2.2. Trading efficiency for a better rate; an interpretation of the
different rates
Theorem 11 shows that some set is approachable here, namely, the set GϕΨ defined in (8): it
is thus a B–set in the terminology of Spinat (2002), see also Hou (1971) as well as a remark
by Blackwell (1956). Therefore, there exists some (abstract√and possibly computationally
extremely inefficient) strategy which approaches it at a 1/ T –rate. Indeed, the proof of
existence of such a strategy does not rely on any constructive argument.
4. This can be seen, e.g., by taking A = {1, 2} and binary payoffs gt,a ∈ {0, 1}. The expectation of the
per-round regret is larger than a positive constant when the gt,a are realizations of independent random
variables Gt,a identically distributed according to a symmetric Bernoulli distribution. In particular, the
regret is larger than this constant for some sequence of binary payoffs gt,a ∈ {0, 1}.
20
Approachability in Unknown Games
Based on all remarks above, we may provide
an intuitive interpretation of the T −1/4
√
rate obtained in Theorem 11, versus the 1/ T rate achieved either in our context by the
abstract strategy mentioned right above, or associated with Blackwell’s original strategy or
variations of it as the one by Bernstein and Shimkin (2015) in the classical case of known
games and sets C being known to be approachable. The interpretation is in terms of the
number of significant (costly) computational units Ncomp (projections, solutions √
of convex or
linear programs, etc.) to be performed. The strategies with the faster rate 1/ T perform
at √
least one or two of these units at each round, while our strategy does it only of the order
of T times during T rounds—they are encompassed into the calls to Ψ and take
p place at
times t = k(k − 1)/2 + 1 for k > 1. In all these cases, the rate is proportional to Ncomp /T .
5.2.3. On the related framework of Azar et al. (2014)
The setting considered therein is exactly the one described in Section 2; our works are
concurrent and independent. Crucial differences lie however in the aims pursued and in the
nature of the results obtained.
The quality of a strategy is evaluated by Azar et al. (2014) based on some quasi-concave
and Lipschitz function f : Rd → R. With the notation of Theorem 11 (the straightforward
extension to an unknown horizon T of) their aim is to guarantee that
lim inf
T →∞
(
f
T
1X
xt
T
t=1
mt
!
−
min
m(k)
max f x
k∈{1,...,nT −1} x∈∆(A)
)
> 0,
(21)
√
where we recall that nT is of order T . Azar et al. (2014) mention that this convergence
can take place at an optimal T −1/4 rate.
Satisfying (21) and recovering this optimal rate is actually a direct consequence of our
Theorem 11 and of the assumptions on f . Indeed, (14) and (16) together with the Lipschitz
assumption on f entail that
lim inf
T →∞
(
f
T
1X
xt
T
mt
t=1
!
nT −1
√
1 X
− f O(1/ T ) +
k Ψ(m(k) )
T
m(k)
k=1
!)
> 0.
(22)
The quasi-concavity of f implies that the image by f of a convex combination is larger than
the minimum of the images by f of the convex combinations. Thus, (22) yields in particular
lim inf
T →∞
(
f
T
1X
xt
T
t=1
mt
!
−
min
k=1,...,nT −1
f Ψ(m(k) )
m(k)
)
> 0.
The convergence rate is the same as for (22), thus is of order at least T −1/4 . Defining the
response function Ψ by Ψ(m) ∈ arg max f (x m), we get (21).
x∈∆(A)
However, we need to underline that the aim (21) is extremely weak: assume, for instance,
that during some block Nature chooses m(k) with identical components such that
∀ x ∈ ∆(A),
f x
21
m(k) = min f .
Mannor, Perchet, Stoltz
Then (21) is satisfied irrespectively of the algorithm. On the contrary, the more demanding
aim (22) that we consider is not necessarily satisfied and an appropriate algorithm—as our
one—must be used.
In addition, the strategy designed by Azar et al. (2014) still requires some knowledge—
the set K of vectors of vector payoffs needs to be known (which is a severe restriction)—and
uses projections onto convex sets. The rate they obtain for their weaker aim is O(T −1/4 ),
as we get for our improved aim.
5.2.4. Links with the strategy of Bernstein and Shimkin (2015)
In this final paragraph of our discussion of Theorem 11 we review the strategy of Bernstein
and Shimkin (2015) and extend it, as much as it can be extended, to a setting as close as
possible to our setting of unknown games: see Figure 3. The extension however requires
that the set K of possible vectors of vector payoff is known to the decision-maker — an
assumption that we would not be ready to make.
Parameters: the set K, a response function Ψ : K → ∆(A)
Initialization: play an arbitrary x1 ∈ ∆(A), pick an arbitrary m
e1 ∈ K
For rounds t = 2, 3, . . .,
1. Update the discrepancy
∆t−1 =
t−1
X
xτ
τ =1
2. Play a mixed action
xt ∈ arg min max h∆t−1 , x
x∈∆(A) m∈K
m
e t ∈ arg max min h∆t−1 , x
3. Compute
mτ −
m∈K
x∈∆(A)
t−1
X
τ =1
Ψ m
es
mi ;
m
es ;
mi .
Figure 3: A generalization of the strategy of Bernstein and Shimkin (2015).
Theorem 13 For all response functions Ψ, the strategy of Figure 3 is such that for all
T > 1, for all sequences m1 , . . . , mT ∈ (Rd )A of vectors of vector payoffs, possibly chosen
by an opponent player,
w
w
T
T
w
w1 X
2B2,max
1X
w
w
xt mt −
Ψ m
et
m
e tw 6 √
.
(23)
w
w
wT
T
T
t=1
t=1
2
The obtained bound is deterministic and uniform over all strategies of the opponent, just
as the bound of Theorem 11 was. Of course, the control (23) is a much weaker statement
than trying to force the convergence of the quantity (19) towards 0: to which set can we
guarantee that
T
X
Ψ m
et
m
et
t=1
belongs? It seems difficult to relate this quantity to the set CϕΨ (mT ) and get the convergence
dp rT , CϕΨ (mT ) → 0 except in some special cases. The applications of Section 6 will further
underline this limitation.
22
Approachability in Unknown Games
One of these special cases is when the set C is approachable, i.e., that the null target
function ϕ ≡ 0 is achievable. This assumption of approachability translates in our more
general case into the existence of a response function ΨC such that ΨC (m) m ∈ C for
all m ∈ K. As advocated by Bernstein and Shimkin (2015), in such settings it is often
computationally feasible to access to ΨC (m) and less costly than performing projections
onto C.
In a nutshell, the strategy of Bernstein and Shimkin (2015) can be extended to the setting
of “almost unknown” games (the set K needs to be known), but the obtained convergence
guarantees are meaningful only under an assumption of approachability of the target set C.
One of the two sources of unknownness of our setting is then (almost) dealt with: the fact
that the underlying structure of the game is unknown, but not the fact that the target is
unknown as well.
Proof of Theorem 13 The construction of the strategy at hand and the proof of its
performance bound also follow some self-confident approach, as for Theorem 11; however,
no blocks are needed. We proceed as in (17) by developing the square Euclidian norm of
∆t+1 to relate it to the one of ∆t , where t > 1:
k∆t+1 k22 = k∆t k22 + 2 ∆t+1 , xt+1 mt+1 − Ψ m
e t+1
m
e t+1
|
{z
}
w
+ wxt+1
|
60
mt+1 − Ψ m
e t+1
{z
2
64B2,max
w2
m
e t+1 w2 .
}
We show below that the inner product is non-positive, which after an immediate recurrence
2
shows that k∆t+1 k22 6 4B2,max
(t + 1)2 and concludes the proof.
Indeed, by von Neumann’s minmax theorem, using the definitions of xt+1 and m
e t+1 ,
maxhδt+1 , xt+1
mi min maxh∆t+1 , x
m∈K
x∈∆(A) m∈K
mi = max min h∆t+1 , x
m∈K x∈∆(A)
= min
x∈∆(A)
δt+1 , x
In particular, for all m0 ∈ K and x0 ∆(A),
hδt+1 , xt+1
m0 i 6 maxhδt+1 , xt+1
m∈K
mi = min
x∈∆(A)
Choosing m0 = mt+1 and x0 = Ψ m
e t+1 entails
∆t+1 , xt+1
δt+1 , x
mt+1 − Ψ m
e t+1
as used above to complete the induction.
mi
m
e t+1 .
m
e t+1 6 δt+1 , x0
m
e t+1 .
m
e t+1 6 0
5.2.5. Link with classical approachability, opportunistic approachability
We recall that in the setting of known finite games described in Section 2.2, vectors of vector
payoffs m actually correspond to the r( · , b). This defines the closed convex set K as the set
23
Mannor, Perchet, Stoltz
of the r( · , y) for all mixed actions y ∈ ∆(B) of the opponent. Both strategies considered
therein relied on a response function Ψ = x? defined as
∀ y ∈ ∆(B),
Ψ r( · , y) = x? r( · , y) ∈ arg min dp r(x, y), C .
x∈∆(A)
Accessing to a value of this response function amounts to solving the convex program
min
X
a∈A
2
xa r(a, y) − c
s.t. x ∈ ∆(A), c ∈ C ,
which can be done efficiently. (It even reduces to a quadratic problem when C is a polytope.)
Our algorithm based on this response function approaches the set Cαunif , where the
quantity αunif is defined in (4); it is not required to compute the said quantity αunif . The
same guarantee with the same remark apply to the two strategies presented in Section 2.2:
Blackwell’s strategy for the case p = 2 only, and the strategy by Bernstein and Shimkin
(2015) for all 0 < p 6 +∞. These three algorithms ensure in particular that the average
payoffs rT are asymptotically inside of or on the border of the set Cαunif .
Now, that αunif is null or positive indicates whether a convex set C is approachable or not.
But the problem of determining the approachability of a set is actually an extremely difficult
problem as even the determination of the approachability of the singleton set C = {0} in
known games is NP–hard to perform; see Mannor and Tsitsiklis (2009). To see that there
is no contradiction between being able to approach Cαunif and not being able to say that
αunif > 0 or not, note that none of the algorithms discussed above does, neither in advance
nor in retrospect, issue any statement on the value of αunif . They happen to perform
approachability to Cαunif for the specific sequence of actions chosen by the opponent but
do not determine a minimal approachable set which would be suited for all sequences of
actions. In particular, they do not provide a certificate of whether a given convex set C is
approachable or not.
Opportunistic approachability. In general, in known games, one has that the target
?
?
function considered above, ϕx , satisfies ϕx ≺ αunif . That is, easy-to-control sequences of
vectors r( · , bt ) can lead to an average payoff rT being much closer to C than the uniform
distance αunif : we get some pathwise refinement of classical approachability. This should be
put in correspondence with the recent, but different, notion of opportunistic approachability
(see Bernstein et al., 2013). However, quantifying exactly what we gain here with the
pathwise refinement would require much additional work (maybe a complete paper as the
one mentioned above) and this is why we do not explore further this issue.
6. Applications
In this section we work out two applications: learning while being evaluated with global
cost functions, and approachability under sample path constraints.
6.1. Global cost functions
This problem was introduced by Even-Dar et al. (2009) and slightly generalized by Bernstein
and Shimkin (2015). We first extend it to our setting of unknown games and describe what
24
Approachability in Unknown Games
Theorem 11 guarantees in our case, and then compare our approach and results to the ones
of the two mentioned references. We keep the original terminology of global costs (thus to
be minimized) and do not switch to global gains (to be maximized), but such a substitution
would be straightforward.
Description of the problem in the case of unknown games. We denote by Kproj ⊂
Rd the closed convex and bounded set formed by the ma when m ∈ K and a ∈ A. A global
cost function is a mapping C : Kproj → R measuring the quality of any vector in Kproj . For
instance, the choice of a mixed action x ∈ ∆(A) given a vector of vector payoffs m ∈ K
is evaluated by C(x m); or the performance of the average payoff rT is equal to C(rT ).
Some regret is to be controlled to ensure that the latter quantity is small as well. Even-Dar
et al. (2009) and Bernstein and Shimkin (2015) defined this regret as
C(rT ) −
where
C ? (m) =
∀m ∈ K,
inf
x∈∆(A)
inf
C(x
x∈∆(A)
C(x
mT ) = C(rT ) − C ? (mT )
(24)
m) .
Assuming that C is continuous, the infimum in the defining equation of C ? is achieved and
we can thus construct a response function Ψ? : K → ∆(A) such that
∀m ∈ K,
C Ψ? (m) m = min C(x m) = C ? (m) .
(25)
x∈∆(A)
Actually, the proof techniques developed in the latter references (see the discussion below)
only ensure a vanishing regret for the convexification vex[C] of C and the concavification
cav[C ? ] of C ? , i.e., they can only issue statements of the form
lim sup vex[C](rT ) − cav[C ? ](mT ) 6 0 ;
(26)
T →∞
they additionally get convergence rates when vex[C] is a Lipschitz function.
We recall that vex[C] 6 C and that cav[C ? ] > C ? , so that the statements of the form
above are much weaker than the original aim (24), at least when C is not convex or C ?
is not concave. A natural case when the latter assumptions are however satisfied is when
C = Cp is the `p –norm, for p > 1 (including the supremum norm p = +∞):
1/p
d
X
Cp (u1 , . . . , ud ) =
upj
and
C+∞ (u1 , . . . , ud ) = max |uj | .
j=1,...,d
j=1
Our main contribution: a better notion of regret. We will directly bound C(rT ),
whether C is convex or not, and will similarly relax the assumption of concavity of C ?
needed in all mentioned references to tackle the desired regret (24).
To that end, we propose a notion of regret that is better in all cases (whether C and
C ? are respectively convex and concave, or not). More precisely, we compare C(rT ) to a
quantity φΨ (mT ) based on any response function Ψ and which generalizes the definition (11):
for all m ∈ K,
X
X
φΨ (m) = sup C
λi Ψ(mi ) mi : N > 1 and
λi mi = m .
i6N
i6N
25
Mannor, Perchet, Stoltz
The extended notion of regret is then defined as C(rT ) − φΨ (mT ).
We now explain why this new definition is always more ambitious than what could be
guaranteed so far by the literature, namely (26). Indeed, when C is convex and by definition
of Ψ? , we have in particular
X
X
?
φΨ (m) 6 sup
λi C Ψ? (mi ) mi : N > 1 and
λi mi = m = cav[C ? ](m) .
i6N
i6N
The inequality
stated above can be strict. For instance, as indicated in Section 4.2, when
C = dp ( · , C where C is convex, the global cost function C is indeed convex. We then have
?
φ Ψ = ϕx
?
cav[C ? ] = cav[ϕ? ] ,
and
?
and thus we possibly have φΨ ≺ cav[C ? ], as stated in Lemma 8. The function C = dp ( · , C
is also a Lipschitz function, which illustrates the interest of the second part of the following
corollary. We recall that B2,max denotes the maximal Euclidean norm of elements in K.
Corollary 14 For all response functions Ψ, when C is continuous and convex, the strategy
of Figure 2 ensures that, uniformly over all strategies of the opponent,
n
o
lim sup C(rT ) − φΨ (mT ) 6 0 .
(27)
T →∞
When C is in addition a Lipschitz function, with constant L for the `2 –norm on Kproj , we
more precisely have
C(rT ) − φΨ (mT ) 6 10L T −1/4 ln A + 3B2,max L T −1/2 .
Proof We apply Theorem 11 and use its notation. The function C is continuous thus
uniformly continuous on the compact set Kproj . Thus,
w
w
wrT − cT w −→ 0
entails
C(rT ) − C(cT ) −→ 0 ,
2
both convergences toward 0 being uniform over all strategies of the opponent. Now, by
definition of cT as a convex combination of elements of the form Ψ(mi ) mi , we have
C(cT ) 6 φΨ (mT ), which concludes the first part of the corollary.
The second part is proved in the same manner, simply by taking into account the
bound (14) and the fact that C is a Lipschitz function.
Discussion. As indicated in general in Section 2 we offered two extensions to the setting
of global costs: first, we explained how to deal with unknown games and second, indicated
what to aim for, given that the natural target is not necessarily approachable and that
sharper targets as the ones traditionally considered can be reached. The second contribution
is perhaps the most important one.
Indeed, the natural target (24) corresponds to ensuring the following convergence to a
set:
where
H = (r, m) : C(r) 6 C ? (m) .
(28)
(rT , mT ) −→ H
26
Approachability in Unknown Games
This target set H is not necessarily a closed, convex, and approachable set but its convex
hull co[H] is so, as proved by Even-Dar et al. (2009) and Bernstein and Shimkin (2015).
This convex hull is exactly equal to
co[H] = (r, m) : vex[C](r) 6 cav[C ? ](m) .
We replace the convergence of (rT , mT ) to the above convex hull co[H] by a convergence
to, e.g., the smaller set
n
?
(r, m) : C(r) 6 φΨ (m) .
Such a convergence is ensured by (27) and the continuity of C, and this set is smaller than
co[H] as follows from the discussion before Corollary 14.
Even-Dar et al. (2009) use directly Blackwell’s approachability strategy to approach
co[H], which requires the computation of projections onto co[H], a possibly computationally
delicate task. We thus only focus on how Bernstein and Shimkin (2015) proceed and will
explain why the obtained guarantee of convergence to co[H] cannot be easily improved with
their strategy. We apply Theorem 13 to a lifted space of payoffs K ⊂ Rd × (Rd )A . Namely,
with each m ∈ K, we associate m ∈ K defined as
ma
∈ Rd × (Rd )A .
(29)
∀ a ∈ A,
ma =
m
That is, the component a ∈ A of m contains the corresponding component ma of m as well
as the vector m itself. In particular,
T
1X
xt
T
mt =
t=1
rT
mT
.
We pick the response function Ψ? : K → ∆(A) corresponding to the base response function
Ψ? defined in (25): Ψ? (m) = Ψ? (m). Then, the convergence (23) reads
w
w
w
w
T
T
T
w
w
w1 X
w
X
X
1
1
w rT
w
w
w
−
Ψ? m
et
m
e tw = w
xt mt −
Ψ? m
et
m
e t w −→ 0 ,
w
w mT
w
wT
w
T
T
t=1
t=1
t=1
2
2
(30)
?
e 2 , . . . in K. By definition of K and Ψ , for all t > 1,
for some m
e 1, m
?
Ψ m
et
m
et
?
∈ (r, m) : C(r) 6 C ? (m) ⊆ H .
et
m
et =
Ψ m
m
et
Thus, the convex combination of the Ψ? m
et
m
e t belongs to co[H] and the convergence (28) is achieved. Under additional regularity assumptions (e.g., continuity of vex[C]
and cav[C ? ]), the stronger convergence (26) holds as can be seen by adapting the arguments
used in the second part of Section C.1.
However, the limitations of the approach of Bernstein and Shimkin (2015) are twofold.
First, as already underline in Section 5.2.4, the sets K or equivalently K need to be known
to the strategy; thus the game is not fully unknown. Second, there is no control on where
the m
e t or m
e t lie, and therefore, there is no reasonable hope to refine the convergence (28)
to a convergence to a set smaller than co[H] and defined in terms of mT as in our approach.
27
Mannor, Perchet, Stoltz
6.2. Approachability under sample path constraints
We generalize here the setting of regret minimization in known finite games under sample
path constraints, as introduced by Mannor et al. (2009) and further studied by Bernstein
and Shimkin (2015). The straightforward enough generalization is twofold: we deal with
approachability rather than just with regret; we consider unknown games.
Description of the problem in the case of unknown games. A vector in Kproj ⊂ Rd
now not only represents some payoff but also some cost. The aim of the player here is
to control the average payoff vector (to have it converge to the smallest expansion of a
given closed convex target set P) while abiding by some cost constraints (ensuring that the
average cost vector converges to a prescribed closed convex set Γ).
Formally, two matrices G and C, of respective sizes g × d and γ × d, associate with
a vector ma ∈ Kproj ⊂ Rd a payoff vector Gma ∈ Rg and a cost vector Cma ∈ Rγ . For
instance, when the decision-maker chooses a mixed action x ∈ ∆(A) and the vector of vector
payoffs is m ∈ K, she gets an instantaneous payoff G(x m) and suffers an instantaneous
cost C(x m). The admissible costs are represented by a closed convex set Γ ⊆ Rγ , while
some closed convex payoff set P ⊆ Rg is to be approached.
The question is in particular what the decision-maker should aim for: the target is
unknown. Following the general aim (3) and generalizing the aims of Mannor et al. (2009)
and Bernstein and Shimkin (2015), we assume that she wants the following convergences
to take place, uniformly over all strategies of the opponent: as T → ∞,
dp GrT , Pϕ(mT ) −→ 0
and
dp CrT , Γ −→ 0 ,
(31)
for some target function ϕ to be defined (being as small as possible). That is, she wants to
control her average payoff GrT as well as she can while ensuring that asymptotically, her
average cost CrT lies in the set Γ of admissible costs.
To make the problem meaningful and as in the original references, we assume that the
cost constraint is feasible.
Assumption 2 For all m ∈ K, there exists x ∈ ∆(A) such that G(x
m) ∈ Γ.
What the general result of Theorem 11 states. We consider mostly the following
response function x?Γ : for all m ∈ K,
n
o
x?Γ (m) ∈ arg min dp G(x m), P : x ∈ ∆(A) s.t. G(x m) ∈ Γ ,
which provides the instantaneous-best and cost-abiding response. The defining minimum
is indeed achieved by continuity as both P and Γ are closed sets. Since in addition P and
Γ are convex, the defining equation of x?Γ is a convex optimization problem under a convex
constraint and can be solved efficiently.
Of course, more general (preferably also cost-abiding) response functions can be considered. By a cost-abiding response function Ψ, we mean any response function such that
∀ m ∈ K,
C Ψ(m) m ∈ Γ .
This property is indeed satisfied by x?Γ .
28
Approachability in Unknown Games
We adapt the definition (11) of the target function based on some response function Ψ
to only consider payoffs: for all m ∈ K,
X
X
φΨ (m) = sup dp G
λi Ψ(mi ) mi , P : N > 1 and
λi mi = m .
i6N
i6N
?
A discussion below will explain why such goals, e.g., (31) with ϕ = φxΓ , are more ambitious
than the aims targeted in the original references, which essentially consisted of shooting for
(31) with ϕ = cav[φ? ] only and in restricted cases (uni-dimensional ones, g = 1), where for
all m ∈ K,
φ? (m) = dp G x?Γ (m) m , P .
(32)
Corollary 15 For all cost-abiding response functions Ψ, the strategy of Figure 2 ensures
that for all T > 1 and for all strategies of the opponent,
dp GrT , PφΨ (mT ) 6 LG 10 T −1/4 ln A + 3B2,max T −1/2
and
dp CrT , Γ 6 LC 10 T −1/4 ln A + 3B2,max T −1/2 ,
where LG , respectively, LC , is a norm on G, respectively, C, seen as a linear function from
Rd equipped with the `2 –norm to Rg , respectively, Rγ , equipped with the `p –norm.
In particular, the aim (31) is achieved.
Proof We apply Theorem 11 and use its notation. By (14) and by definition of LC ,
w
w
wCrT − CcT w 6 LC 10 T −1/4 ln A + 3B2,max T −1/2 .
p
Because Ψ was assumed to be cost-abiding and in view of the form (16) of cT , we have
CcT ∈ Γ and we thus have proved
dp CrT , Γ 6 LC 10 T −1/4 ln A + 3B2,max T −1/2 .
A similar argument, based on the fact that GcT ∈ PφΨ (mT ) by definition of φΨ , yields the
stated bound for dp GrT , PφΨ (mT ) .
What the extension of earlier results, e.g., Theorem 13, yields. As indicated several times already, Mannor et al. (2009) and Bernstein and Shimkin (2015) only considered
the case of regret minimization, i.e., a special case of approachability when G is a linear
form (g = 1) and P is an interval of the form [G∞ , +∞) where G∞ is a bound on the values
taken by G. We will discuss this special case below.
The strategies considered by Mannor et al. (2009) were not efficient (they relied on being
able to project on complicated sets or resorted to calibrated auxiliary strategies), unlike the
one studied by Bernstein and Shimkin (2015). We will thus focus on the latter. The (not
necessarily convex) target set considered therein is
H = (r, m) : Cr ∈ Γ and dp (Gr, P) 6 φ? (m) ,
29
Mannor, Perchet, Stoltz
where φ? was defined in (32). Because P is convex and G is linear, the function r ∈ Rd 7→
dp (Gr, P) is convex; see, e.g., Boyd and Vandenberghe (2004, Example 3.16). The convex
hull of H thus equals
co[H] = (r, m) : Cr ∈ Γ and dp (Gr, P) 6 cav[φ? ](m) .
To be able to compare the merits of the strategy by Bernstein and Shimkin (2015) to
Corollary 15, we first extend it to the case of unknown games, based on Theorem 13. To
that end we consider the same lifting as in (29) and apply similarly Theorem 13 to get (30)
as well, for the cost-abiding response function x?Γ . Using that in this case, by definition
of x?Γ ,
?
xΓ m
et
m
et
?
et
m
et =
Ψ m
∈ H,
m
et
the convergence (30) rewrites
w
T
w
1 X x?Γ m
et
w rT
−
w
m
et
w mT
T
{z
| t=1
∈ co[H]
m
et
w
w
w
w −→ 0
w
} 2
and entails the convergence of (rT , mT ) to co[H]. In particular, CrT → Γ. Under an
additional regularity assumption, e.g., the continuity of cav[φ? ], we also get (by adapting
the arguments used in the second part of Section C.1) the stronger convergence
n
o
lim sup dp (GrT , P) − cav[φ? ](mT ) 6 0 ,
T →∞
that is,
lim sup dp GrT , Pcav[φ? ](mT ) 6 0 .
T →∞
Summarizing, the convergence (31) is guaranteed with ϕ = cav[φ? ]; an inspection of the
arguments above shows that cav[φ? ] being actually uniformly continuous, the desired uniformity over the strategies of the opponent is achieved.
The same limitations to this approach as mentioned at the end of the previous section
arise as far as the concepts of unknown game and unknown target are concerned. First,
the set K needs to be known to the strategy and the game is not fully unknown. Second,
there is no control on where the m
e t lie, and therefore, there is no reasonable hope to refine
the convergence (31) with ϕ = cav[φ? ] into a convergence with a smaller target function ϕ.
?
In contrast, Corollary 15 provided such a refinement with ϕ = φxΓ , which by convexity of
r ∈ Rd 7→ dp (Gr, P) is smaller and possibly strictly smaller than cav[φ? ] (adapt Lemma 8
to prove the strict inequality).
A note on known games. However, Mannor et al. (2009, Section 5) exhibit a class of
cases when cav[φ? ] is the optimal target function: in known games, with scalar payoffs and
scalar constraints, and with set of constraints of the form Γ = (−∞, c0 ]. This amounts to
minimizing some constrained regret.
We thus briefly indicate what known games are in this context, as defined by Mannor
et al. (2009) and Bernstein and Shimkin (2015). Some linear scalar payoff function u :
30
Approachability in Unknown Games
∆(A) × ∆(B) → [0, G∞ ] and some linear vector-valued cost function v : ∆(A) × ∆(B) → Rγ
are given. (With no loss of generality we can assume that the payoff function takes values
in a bounded nonnegative interval.) The set K of our general formulation corresponds to
the vectors, as y describes ∆(B),
u( · , y)
v( · , y)
∈ R × Rγ .
The matrices G and C extract respectively the first component and all but the first component. Regret is considered, that is, the payoff set P to be be approached given the
constraints is [G∞ , +∞). The expansions are Pα = [G∞ − α, +∞). The distance of some
r ∈ R to some Pα equals (G∞ − α − r)+ .
In this context, convergences of the form (31) thus read
T
1X
v(xt , bt ) −→ Γ
T
and
lim inf
T →∞
t=1
where
uT
vT
T
1X
u(xt , bt ) − G∞ − ϕ (uT , v T )
> 0,
T
t=1
(33)
T
1 X u( · , bt )
,
=
v( · , bt )
T
t=1
and thus correspond to some constrained regret-minimization problems. Indeed, denoting
yT =
T
1X
δbt
T
t=1
the empirical frequency of actions bt ∈ B taken by the opponent, and recalling that u is
bounded by G∞ , we have, for instance, when ϕ = φ? ,
φ? (uT , v T ) = G∞ − u?Γ (y T )
where
u?Γ (y) = max u(x, y) : x ∈ ∆(A) s.t. v(x, y) ∈ Γ . (34)
The convergence (33) finally reads when ϕ = φ? :
T
1X
v(xt , bt ) −→ Γ
T
t=1
and
lim inf
T →∞
T
1X
?
u(xt , bt ) − uΓ (y T ) > 0 .
T
(35)
t=1
Just as we showed (in Section 3.1) that in general the target function φ? is not achievable,
Mannor et al. (2009, Section 3) showed that the constrained regret with respect to u?Γ (y T )
defined in (35) cannot be minimized.
The proposed relaxation was to consider its convexification vex[u?Γ ] instead in (35), which
corresponds to cav[φ? ] in (34). In this specific one-dimensional setting, the target function
?
φxΓ equals cav[φ? ]: our general theory provides no improvement. This is in line with the
optimality result for cav[φ? ] exhibited by Mannor et al. (2009, Section 5) in this case.
31
Mannor, Perchet, Stoltz
6.3. Approachability of an approachable set at a minimal cost
This is the dual problem of the previous problem: have the vector-valued payoffs approach
an approachable convex set while suffering some costs and trying to control the overall cost.
In this case, the set P is fixed and the α–expansions are in terms of the set of constraints
Γ. Actually, this is a problem symmetric to the previous one, when the roles of G and P
are exchanged with C and Γ.
Acknowledgments
Vianney Perchet acknowledges funding from the ANR, under grants ANR-10-BLAN-0112
and ANR-13-JS01-0004-01. Shie Mannor was partially supported by the ISF under contract
890015. Gilles Stoltz would like to thank Investissements d’Avenir (ANR-11-IDEX-0003 /
Labex Ecodec / ANR-11-LABX-0047) for financial support.
An extended abstract of this article appeared in the Proceedings of the 27th Annual
Conference on Learning Theory (COLT’2014), JMLR Workshop and Conference Proceedings, Volume 35, pages 339–355, 2014.
32
Approachability in Unknown Games
References
J. Abernethy, P.L. Bartlett, and E. Hazan. Blackwell approachability and no-regret learning
are equivalent. In Proceedings of COLT, pages 27–46, 2011.
P. Auer, N. Cesa-Bianchi, Y. Freund, and R.E. Schapire. The nonstochastic multiarmed
bandit problem. SIAM Journal on Computing, 32(1):48–77, 2002.
Y. Azar, U. Feige, M. Feldman, and M. Tennenholtz. Sequential decision making with
vector outcomes. In Proceedings of ITCS, 2014.
A. Bernstein and N. Shimkin. Response-based approachability with applications to generalized no-regret problems. Journal of Machine Learning Research, 16(Apr):747–773,
2015.
A. Bernstein, S. Mannor, and N. Shimkin. Opportunistic strategies for generalized no-regret
problems. In Proceedings of COLT, pages 158–171, 2013.
D. Blackwell. An analog of the minimax theorem for vector payoffs. Pacific Journal of
Mathematics, 6:1–8, 1956.
S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, Cambridge, UK, 2004.
N. Cesa-Bianchi and G. Lugosi. Potential-based algorithms in on-line prediction and game
theory. Machine Learning, 3(51):239–261, 2003.
N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University
Press, 2006.
N. Cesa-Bianchi, Y. Mansour, and G. Stoltz. Improved second-order bounds for prediction
with expert advice. Machine Learning, 66(2/3):321–352, 2007.
S. de Rooij, T. van Erven, P.D. Grünwald, and W. Koolen. Follow the leader if you can,
hedge if you must. Journal of Machine Learning Research, 15(Apr):1281–1316, 2014.
E. Even-Dar, R. Kleinberg, S. Mannor, and Y. Mansour. Online learning for global cost
functions. In Proceedings of COLT, 2009.
D. Foster and R. Vohra. Asymptotic calibration. Biometrika, 85:379–390, 1998.
D.P. Helmbold and M.Warmuth. Tracking the best expert. Machine Learning, 32(2):151–
178, 1998.
J.-B. Hiriart-Urruty and C. Lemaréchal. Fundamentals of Convex Analysis. Springer-Verlag,
2001.
T.-F. Hou. Approachability in a two-person game. The Annals of Mathematical Statistics,
42:735–744, 1971.
S. Mannor and G. Stoltz. A geometric proof of calibration. Mathematics of Operations
Research, 35:721–727, 2010.
33
Mannor, Perchet, Stoltz
S. Mannor and J. N. Tsitsiklis. Approachability in repeated games: Computational aspects
and a Stackelberg variant. Games and Economic Behavior, 66(1):315–325, 2009.
S. Mannor, J.N. Tsitsiklis, and J.Y. Yu. Online learning with sample path constraints.
Journal of Machine Learning Research, 10:569–590, 2009.
S. Mannor, V. Perchet, and G. Stoltz. Set-valued approachability and online learning with
partial monitoring. Journal of Machine Learning Research, 15(Oct):3247–3295, 2014.
J.-F. Mertens, S. Sorin, and S. Zamir. Repeated games. CORE Discussion papers 9420,
9421, 9422, Louvain-la-Neuve, Belgium, 1994.
V. Perchet. Approachability, regret and calibration: Implications and equivalences. Journal
of Dynamics and Games, 1(2):181–254, 2014.
X. Spinat. A necessary and sufficient condition for approachability. Mathematics of Operations Research, 27:31–44, 2002.
34
Approachability in Unknown Games
Appendix A. Link with regret minimization in unknown games
The problem of regret minimization can be encompassed as an instance of approachability.
We recall here why the knowledge of the payoff structure is not crucial for this very specific
problem. (This, of course, is not the case at all for general approachability problems.)
Indeed, with the notation of Section 2.2, the aim of regret minimization, in a known
finite game with payoff function s : A × B, is for the decision-maker to ensure that
lim sup
T →∞
T
T
1X
1X
s(at , bt ) − max
s(a, bt ) > 0
a∈A T
T
t=1
a.s.
t=1
This can be guaranteed by approaching (R− )A with the vector payoff function r : A × B →
RA defined by
r(a, b) = s(a0 , b) − s(a, b) a0 ∈A .
(36)
The necessary and sufficient condition for approachability of the closed convex set (R− )A
is satisfied for r. The condition (5) rewrites in our case
D
E
E D
∀ y ∈ ∆(B),
ret −e
ct , r(xt+1 , y)−e
ct = ret + , r(xt+1 , y)+ ret − = ret + , r(xt+1 , y) 6 0 ,
where ( · )+ and ( · )+ denote respectively the vectors formed by taking the nonnegative
and non-positive parts of the original components of the vector of interest. Now, using the
specific form of r, we see that
!
D
E
X
X
ret + , r(xt+1 , y) =
ret,a0 + s(a0 , y) −
ret,a0 + s(xt+1 , y) .
a0 ∈A
a0 ∈A
Either all components of ret are non-positive, i.e., ret is already in (R− )A , or we can choose
the mixed distribution xt+1 defined by
ret,a +
.
∀ a ∈ A,
xt+1,a = P
(37)
et,a0 +
a0 ∈A r
In the latter case, we then get
∀ y ∈ ∆(B),
ret − e
ct , r(xt+1 , y) − e
ct =
and (5) is in particular satisfied.
D
ret
E
,
r(x
,
y)
= 0,
t+1
+
The knowledge of s (or r) is not crucial here. Comments have to be made on the
specific choice of xt+1 : it is independent of the payoff structure s (or r), it only depends on
the past payoff vectors s( · , bτ ), where τ 6 t.
In particular, the strategy above to minimize the regret can be generalized in a straightforward way to the case of games with full monitoring but whose payoff structure is unknown. In these games, at each round, the opponent chooses a payoff vector
gt = (gt,a )a∈A ,
35
Mannor, Perchet, Stoltz
the decision-maker chooses an action at ∈ A and observes the entire vector gt , while wanting
to ensure that the regret vanishes,
T
T
1X
1X
lim sup
gt,at − max
gt,a > 0
a∈A T
T →∞ T
t=1
a.s.
(38)
t=1
It suffices to replace all occurrences of s( · , bt ) above by gt . In particular, the payoff function
r defined in (36) is to be replaced by the vectors of vector payoffs mt ∈ RA×A whose
components equal
∀ a ∈ A,
mt,a = gt,a0 − gt,a a0 ∈A .
A note on the bandit monitoring: the case of unknown games. In the case of an
unknown game (i.e., when the payoff structure is unknown and when only bandit monitoring
is available), the generic trick presented around (2) should be adapted, as indicated by
Footnote 1. Indeed, the only feedback available at the end of each round is gt,at and not
mt,at . The estimation to be performed is rather on the vectors gt than on the mt : for all
a ∈ A,
gt,a
gbt,a =
I
,
xt,a {a=at }
with the same constraints xt,a > γt for all t > 1, from which we define
m
b t,a = gbt,a0 − gbt,a a0 ∈A .
Substituting the estimates gbt in the strategy defined around (37) in lieu of the vectors
s( · , bt ) ensures that the regret vanishes.
36
Approachability in Unknown Games
Appendix B. Calculations associated with Examples 1 and 2
B.1. Proof of Lemma 3
Proof for Example 1 Assume by contradiction that the convergence (3) can be achieved
and consider any strategy of the decision maker to do so, which we denote by σ. It suffices
to consider the almost sure convergence (3), the stronger uniformity requirements stated
after it will not be invoked. All statements in the sequel hold almost surely, and quantities
like Tε and Tε0 should be thought of as random variables.
Imagine in a first time that opponent chooses, at every stage t > 1, the vectors mt = m† .
We have ϕ? (m† ) = 4, the smallest of the supremum norms of m†1 and m†2 . The aim is then
that the average payoffs rT converge to C4 . But this can be guaranteed only if the averages
of the chosen mixed actions xt converge to (1, 0). That is, given ε > 0, there exists some,
possibly large, integer Tε such that
3
rTε −
6 ε.
(39)
4
∞
Now, consider a second scenario. During the first Tε stages, the opponent chooses the
vectors mt = m† . By construction, as the strategy σ is fixed, (39) is ensured. Now, in
the next Tε stages, for Tε + 1 6 t 6 2Tε , assume that the opponent chooses the vectors
mt = m] , and denote by γε ∈ [0, 1] the average of the first components xt,1 ∈ [0, 1] of the
mixed actions xt selected by σ in this second set of stages. We have m2Tε = m(1/2) , where
1 † 1 ]
2.5
3.5
(1/2)
(1/2)
(1/2)
.
and m2
=
m
= m + m
has components
m1
=
2.5
3.5
2
2
Therefore the target set is Cϕ? (m2Tε ) = C2.5 . However, by definition of γε , we have
r2Tε
2T
1
1
1 Xε
1
1
1 5 − γε
5
4
]
= r Tε +
= rTε +
xt m = rTε +
γε
+ (1 − γε )
,
0
3
3γε
2
2
2
2
2
2
t=Tε +1
and therefore, because of (39),
r2Tε −
4 − γε /2
2 + 3γε /2
6
∞
ε
.
2
This entails that
ε
4 − γε /2
d∞ r2Tε , Cϕ? (m2Tε ) > d∞
, C2.5 − > 1 − ε/2 .
2 + 3γε /2
2
This construction can be repeated again after stage 2Tε , by choosing mt = m† till a
stage Tε0 is reached when
ε
3
r2Tε +Tε0 −
6 ;
4
2
∞
such a stage exists by the assumption that the convergence (3) is achieved by the strategy
σ. One can then similarly see that
d∞ r2Tε +2Tε0 , Cϕ? (m2T +2T 0 ) > 1 − ε/4 .
ε
37
ε
Mannor, Perchet, Stoltz
By repeating this over again and again, one proves that
lim sup d∞ rT , Cϕ? (mT ) > 1 ,
T →∞
which contradicts the assumption that σ ensures the convergence (3). The claim follows.
Proof for Example 2 (sketch) The same construction as for the previous example holds,
by switching between a first regime when mt = (−1, 1) ∈ R2 is chosen and at the end of
which the average payoff should be close to null, rTε 6 ε. Then, another regime of the
same length starts with mt = (1, 1) ∈ R2 and no matter what the decision-maker does, she
will get an average payoff of 1 in this regime. In total, at the end of the second regime,
r2Tε > 1/2 − ε while the target set is given by
Cϕ? (m2Tε ) = Cϕ? ((0,1)) = C0 = {0} ,
as
m2Tε = (0, 1) ∈ R2 .
This can be repeated over and over again.
B.2. Proof of Lemma 4
Proof for Example 1 We have cav[ϕ? ] ≡ 4. To prove this fact, we first compute ϕ? . For
ν ∈ [0, 1], the components of m(ν) = νm† + (1 − ν)m] equal
4−ν
4
3
(ν)
=
+ (1 − ν)
m1 = ν
3+ν
3
4
5 − 5ν
5
0
(ν)
.
=
+ (1 − ν)
and
m2 = ν
5ν
0
5
Therefore,
4−ν
5 − 5ν
ϕ? m(ν) = min max{4 − ν, 3 + ν}, max{5 − 5ν, 5ν} =
5ν
3 + ν
if
if
if
if
ν
ν
ν
ν
∈ [0, 1/4],
∈ [1/4, 1/2],
(40)
∈ [1/2, 3/4],
∈ [3/4, 1].
We note that ϕ? m† = ϕ? m] = 4 and that ϕ? 6 4, so that cav[ϕ? ] is identically equal
to 4 on the set K defined as the convex hull of m† and m] .
Smaller target functions ϕ such that the convergence (3) holds can be considered. This
proves in particular that (3) can also be guaranteed for the larger cav[ϕ? ]. Indeed,
ϕ1 m(ν) = max{4 − ν, 3 + ν}
is smaller than cav[ϕ? ] ≡ 4 (and even strictly smaller when ν 6= 0 and ν 6= 1); see Figure 4. In
addition, the convergence (3) can hold for it. Indeed, if the decision-maker plays xt = (1, 0)
at each round, i.e., always picks the first component of mt , then her average payoff equals
rT = mT,1 , where mT = m(ν T ) for some ν T ∈ [0, 1]. By definition of ϕ1 , the distance of
38
3 + ν
if ν ∈ [3/4, 1].
This last set of equalities can be seen to hold true by contemplating Figure 2.
Approachability in Unknown Games
m†
5
αx
4
ϕ?
m]
C
2.5
m]
m†
C4
C2.5
Figure 4:
Graphs of the functions ϕ? (bold solid line) and ϕ1 (dotted line).
Figure 2: Graphical representation of m† and m] , and of different expansions Cα (left); graphs
(ν)
of the functions
ϕ? (bold
solid
of α0 and
α1 is(dotted
lines),
αx , for we even have in
m1 to (−∞,
0]2 in
theline),
supremum
norm
precisely
ϕ1 and
m(ν)of .some
Therefore,
x ∈ [0,this
1] (thin
casesolid line).
d∞ rT , Cϕ1 (mT ) = 0 ,
which
proves by
in particular
that
theϕ?convergence
for ϕexample
1.
Proof [of Lemma
4] Assume
contradiction
that
is achievable(3)in holds
the above
and
consider any strategy of the decision maker to do so, which we denote by σ. Imagine in a first time
that Nature chooses,
at every
t > 1, the
vectors
mt = m† , which
amounts
to playing
νt = seemingly
1.
Proof
for stage
Example
2 The
computations
are more
involved
in this
simpler
?
?
?
2
Given that the average
of
the
ν
equals
ν
=
1
and
that
its
image
by
ϕ
equals
ϕ
(1)
=
4,
the
aim
t
T
example. As before, we start by computing ϕ . We refer to vectors m ∈ [−1, 1] chosen
is then to converge
C4 . opponent
But this can
guaranteed
if the
the chosen
xt by
converges
bytothe
asonly
m =be(v,
w) and to
theaverage
mixed of
actions
picked
the decision-maker by
to 1. That is, given
ε
>
0,
there
exists
some,
possibly
large,
integer
T
such
that
x = (x1 , 1 − x1 ), where x1 ∈ [0, 1]. The absoluteε value of a convex combination of v and w
is to be minimized. This is achieved
with
3
rTε −
6 ε.
(21)
4
(1,
0)
if
0
<
v
6
w
or
0
>
v > w,
∞
(0,
1)
if
0
<
w
<
v
or
0
>
w > v,
?
Tε stages, Nature chooses
x (v,
w) =the first
Now, consider a second scenario.
During
the vectors mt = m† .
|w|
|v|
, next Tε stages,
if for
v wT6ε +
0, 1 6 t 6 2Tε ,
Now, for the
By construction, as σ is fixed, (21) is ensured.
|v|] + |w| |v| + |w|
assume that Nature chooses the vectors mt = m , and denote by γ ∈ [0, 1] the average of the
which
to
xt ∈ [0, 1] selected
by σ.leads
The average
of the played vectorsis m2Tε corresponds to ν = 1/2, whose
min |v|, |w| if v w > 0,
image by ϕ? equals ϕ? (1/2) = 5/2. Therefore
thew)target
ϕ? (v,
= set is C2.5 . However, by definition of γ,
0
if v w 6 0.
we have
for all v, w ∈ [−1, 1]2 ,
2T
The concavification
of ϕ? thus admits
thefollowing
expression:
1
1 Xε
1
1
4
5
]
r2Tε = rTε +
xt m = rTε +
γ
+ (1 − γ)
3
2
2
2
2
|v − w| 0
t=Tε +1
cav[ϕ? ]
(v, w) = 1
−
.
2
1
1
5−γ
rT +
,
We replace a lengthy and=tedious
of3γthis expression
by the graphical illustrations
2 ε proof
2
provided by Figure 5.
Now, we consider the target
23 function ϕ1/2 defined as
|v + w|
ϕ1/2 (v, w) =
.
2
(41)
We denote mT = v T , wT . By playing xt = (1/2, 1/2) at each round, the decision-maker
ensures that
v T + wT
v T + wT
rT =
while
ϕ1/2 mT =
.
2
2
Thus, for this example again (in which the `p –norm can be chosen freely), we have
dp rT , Cϕ1/2 (mT ) = 0 .
39
1 + |v + w| /3
(1 + v)/2
?
(1 + w)/2
ϕx (v, w) =
(1 − v)/2
(1 − w)/2
if
if
if
if
if
|2w − w| 6 1 and |2v − w| 6 1,
2w − v > 1 and v + w > 0,
2v − w > 1 and v + w > 0,
2w − v 6 −1 and
v + w 6Perchet,
0,
Mannor,
Stoltz
2v − w 6 −1 and v + w > 0,
?
which shows that ϕx 6 cav[ϕ? ]. Admittedly, a picture would help: we provide one as Figure 4.
1
1
v
v
1−y
2
0
0
1+|x+y|
3
1−x
2
w
w
1+x
2
0
1
1+y
2
1
1
1−
0
11
|v−w|
2
1−
1
|v−w|
2
0
0
ϕ?
ϕ?
cav[ϕ? ]
0
cav[ϕ? ]
0
?
? (left)
? ] (right).
Figure
of ϕ
and of its concavification cav[ϕ? ] (right).
3: Representations
ϕ?of(left)
its
concavification
cav[ϕ
Figure Figure
4: Representations
in 2D –of3D
ϕx ?and3:ofRepresentations
α1/2
.
Figure 5: Representations of ϕ (left), of the alternative target function ϕ1/2 (center), and
of the concavification cav[ϕ? ] (right).
?
x can be improved. Indeed,
We conclude our
discussion
of this4]example
by construction
showing
thatLemma
even
Proof [of
4] previous
The sameexample
construction
the previous
Proof
[of Lemma
The same
as
forϕthe
holds,asbyforswitching
be- example holds, by switching be? , that is, somehow by having the unwas constructedtween
by choosing
the
response
function
Ψ
=
x
tween
regimes
mend
(−1,
1) isthe
chosen
andpayoff
at the should
end of which the average payoff should
regimes when mt = (−1, 1)
is chosen
andwhen
at the
which
average
t = of
? in mind. This response function, as the main algorithm below shows,
chievable target function
be close
to null,
rT the
6 ε.
Then,
another
regime
oft =
the same
length
be close ϕto null,
rT 6 ε. Then, another
regime
same
length
with
1) and
no starts with mt = (1, 1) and no
? ],m
The
convergence
(3)ofholds
for
ϕof
and
also
forstarts
cav[ϕ
since(1,the
latter
is larger than ϕ1/2
1/2anMore
s a short-term reply
to awhat
localthe
average
of vectors
payoffs.
reasonable
quantities
matter
whatwill
theget
decision-maker
does, of
she
getregime.
an average
payoff
of 1 in this regime. In total,
matter
decision-maker
does,vector
she
average
payoff
1 will
in this
In total,
?
(strictly
at some
points).
Indeed,
αϕ
follows
from
? , equals
hould be targeted;rTquite
surprisingly
we
that
relaxing
the
short-terms
1/2
rTbelow
>is,1/2
−
ε while
the
image
of? (0,
m6
= (0,
1)] by
ϕ? , equals
ϕ?the
(0, 1)fact
= 0. This
> 1/2
− ε whilelarger
theillustrate
target,
that
the
image
of
m
= the
(0, that
1)inequality
byis,expectations
ϕthe
1)cav[ϕ
0. This
2T
2T target,
lso results in better
long-term
payoffs.
For
instance,
if
the
decision
maker
knows
in
advance
that
that
can be repeated over and over again.
can be repeated over and over again.
thus,
he next vectors will be mt = (0, 1), she should not worry about getting ϕ? (0, 1) = 0, and
|v + w| + |v − w| = max |v|, |w| 6 2 .
?
laying xt = (1, 0); she could well be satisfied with ϕ (0, 1) = 1/2, and thus play xt = (1/2, 1/2).
Proof
[of
the
second
consider
w) = |v + w|/2, which is an achievthe second
part
of Lemma
6] We
α1/2of
(v,Lemma
w) = |v6]
+ We
w|/2,
which αis1/2
an(v,
achievturns out that theProof
same[of
argument
can be
performed
at each
(v,consider
w).? part
That
the inequality
α1/2
6
can
be strict
seen,
e.g.,
at(1/2,
(0,
0).
Again,
an illustration
target
function;
it (1/2,
indeed
suffices
to
play
x
=
1/2)
at each
able target
function;
it indeed able
suffices
to cav[ϕ
play
xt ] =
1/2)
atiseach
round.
The
inequality
t
round. The inequality
? ] follows from the fact that, |v + w| + |v − w| = max |v|, |w| 6 2. That this
? ] follows from the
α
6
cav[ϕ
α
6
cav[ϕ
fact
that,
|v
+
w|
+
|v
−
w|
=
max
|v|,
|w|
6
2.
That
this
5.
1/2partisofprovided
roof [of the second
Lemma 8] by
TheFigure
target1/2
function
ϕ = α1/2 already considered above is
? ]. (Thiswe
?
is seen, e.g.,
at (0,α0).
a conclusion,
inequality
can be strict is seen, inequality
e.g., at (0,can
0). be
As strict
a conclusion,
we have
≺ cav[ϕ
is have α1/2 ≺ cav[ϕ? ]. (This is
x
1/2 As
uch that α1/2 ≺ ϕ as can be seen on a picture.
illustrated by Figures 3 and 4.) illustrated by Figures 3 and 4.)
what we could
a “signvalues
compensation.”
The absolute values of
The B.3.
proof below
illustrates
whatThe
weproof
call a illustrates
“sign compensation.”
Thecall
absolute
of
Proof
of Lemma
8couldbelow
the convex
in (22)
can be (much)
smallerofthan the convex combinations of
the convex combinations considered
in (22)combinations
can be (much)considered
smaller than
the convex
combinations
? ]; it can indeed be
? ]; it in
the absolute
values
their
elements,
as
expression
cav[ϕcompensathe absolute
values
of their
elements,
considered
the expression
of considered
cav[ϕ
indeed
be a of“sign
Proof
for
Example
2asThe
proofofinbelow
illustrates
what
wecanthe
could
call
2
seenfor
in our
example
(given
seen in our
example
form
of
the
α ) that
tion.”
We(given
have the
here,
allsets
(v,Cw)
∈
R ,the form of the sets Cα ) that
( N
)
( N
)
N
N
X
X
X
X
( ?N
)
?
?
?
N
sup
λi 1xand
(vi , wi ) λi (v
(vii,, w
wii)) =: (v,
N w)
> 1 .X
and
λi (vi , wi ) = (v, w) .
cav[ϕ ](v, w) = sup λi xcav[ϕ
(vi X
, w](v,
(v=
>
i ) w)
i, w
i) : N
cav[ϕ? ] (v, w)i=1= 27
sup
λi x? (vi , wi=1
i)
(vi , wi=1
i ) : N > 1 and
i=1
λi=1
i (vi , wi ) = (v, w) ,
i=1
Proof [of
second to
part
Lemma
7] According
(8) Cand
given the form of the sets Cα , the
Proof [of the second part of Lemma
7] the
According
(8)of
and
given the
form of thetosets
α , the
?
?
target function ϕx is defined astarget function ϕx is defined as
)
( N
N
)
( N
)
( N
X
X
N
N
?
X
X
X
X
?
?
? (vi , wi ) :
λ
xi )? :(vN
, w ) = (v, w) .
ϕxsup(v, w) λi=x?sup
i, w
sup
(vi , wi ) λi (v
(vii,, w
wiiN
N1w)
>and
1 . and
ϕx (v, w) =
(vϕi ,xw(v,
(v=
>λi )i1xand
)) =:>(v,
(22) λλi (v
i (vii , wii) = (v, w) . (22)
i ) w)
i , iw
i=1
i=1
i=1
i=1
i=1
i=1
By a tedious case study consisting of identifying the worst convex decompositions
one then gets the explicit expression
ϕx
?
26
1 + |v + w| /3
(1 + v)/2
(1 + w)/2
(v, w) =
(1
− v)/2
(1 − w)/2
if
if
if
if
if
40
26
|2w − v| 6 1 and |2v − w| 6 1,
2w − v > 1 and v + w > 0,
2v − w > 1 and v + w > 0,
2w − v 6 −1 and v + w 6 0,
2v − w 6 −1 and v + w 6 0,
P
λi (vi , wi ),
(42)
Approachability in Unknown Games
to be compared to the closed-form expression obtained earlier for cav[ϕ? ], namely
|v − w|
cav[ϕ? ] (v, w) = 1 −
.
2
Admittedly, a picture would help: we provide one as Figure 6. We see (on the picture or
1
1
v
w
1+v
2
0
1−w
2
1−v
2
0
1
1+|v+w|
3
1−
1
|v−w|
2
0
1+w
2
ϕ?
cav[ϕ? ]
0
?
?
? ϕx of
Figure 1: Representations
in
and
α1/2 .and and
Figure
ϕ?of(left)
of its3D
concavification
cav[ϕ
Figure 6: Representations
in 2D3:– Representations
3D2Dof– 3D
ϕx of(left
and
center)
representation
of ] (right).
cav[ϕ? ].
Proof [of Lemma
4] The same construction as for the previous example holds, by switching bex? 6 cav[ϕ? ], and even that ϕx? ≺ cav[ϕ? ], by considering the
by direct calculations)
that
ϕ
tween regimes when mt = (−1, 1) is chosen and at the end of which the average payoff should
respective values be
1/3close
and to
1 null,
at r(0, 6
0).ε. Then, another regime of the same length starts with m = (1, 1) and no
t
T
matter what the decision-maker does, she will get an average payoff of 1 in this regime. In total,
x? is,
rT > 1/2
the target,
that
the
image
m2T
= (0,will
1) by
ϕ? , equals
ϕ? (0, 1) = 0. This
Proof for Example
1−
Weε while
will prove
that ϕ
=ϕ
thatofthe
result
follow
from the
1 so
? ] already
be repeated
over
and over
again. B.2.
inequality ϕ1 ≺can
cav[ϕ
proved
in Section
Indeed, as can be seen in the computations leading to the closed-form expression (40)
of ϕ? , we have Proof [of the second (
part of Lemma 6] We consider α1/2 (v, w) = |v + w|/2, which is an achiev
(1, 0) if ν ∈ [0, 1/4] ∪ [3/4, 1],
able target
x? function;
m(ν)
= it indeed suffices to play xt = (1/2, 1/2) at each
round. The inequality
?
1) the
if νfact
∈ [1/4,
α1/2 6 cav[ϕ ] follows(0,from
that, 3/4].
|v + w| + |v − w| = max |v|, |w| 6 2. That this
inequality can be strict is seen, e.g., at (0, 0). As a conclusion, we have α1/2 ≺ cav[ϕ? ]. (This is
Therefore,
(4.)(ν)
illustrated by Figures 3 and
m1
if ν ∈ [0, 1/4] ∪ [3/4, 1],
?
(ν)
(ν)
x m
m =
(ν)
m2
if ν ∈ [1/4, 3/4].
The proof below illustrates what we could call a “sign compensation.” The absolute values of
But for ν ∈ [1/4,
wecombinations
have the component-wise
inequality
the3/4],
convex
considered in (22)
can be (much) smaller than the convex combinations of
the absolute values
of
their
elements,
as
considered
in the expression of cav[ϕ? ]; it can indeed be
5 − 5ν
4−ν
(ν)
(ν)
m2 = (given the form
6 of the sets=Cαm) 1that,
seen in our example
5ν
3+ν
( N
)
N
(ν)
X
X
?
(ν)
(ν)
which entails that
for ?all
∈ [0,
m 6 m . Substicav[ϕ
](v,ν w)
= 1],
supagain component-wise,
λi x? (vi , wi ) (vxi , wm
λi (vi , wi ) = (v, w) .
i ) : N > 1 and 1
tuting in (10) and using that the supremum
distance
to
the
negative
orthant is i=1
increasing
i=1
with respect to component-wise inequalities, i.e.,
Proof [of the second part of Lemma 7] According
to (8) and given the form of the sets Cα , the
2
2
?
x
(v,
w)
∈
R
−
7
→
d
(v,
w),
(−∞,
0]
=
max
(v)+ , (w)+ ,
target function ϕ is∞defined as
)
( N
N
X
X
? 41
x?
λi x (vi , wi ) (vi , wi ) : N > 1 and
λi (vi , wi ) = (v, w) . (22)
ϕ (v, w) = sup
i=1
1
i=1
Mannor, Perchet, Stoltz
we get that
x?
ϕ
m
(ν)
(
= sup d∞
(
6 sup d∞
N
X
i=1
N
X
?
(νi )
λi x m
(ν )
(νi )
m
λi m1 i , (−∞, 0]2
i=1
2
, (−∞, 0]
!
!
: N > 1 and
: N > 1 and
(ν)
= d∞ m1 , (−∞, 0]2 = ϕ1 m(ν) .
?
N
X
i=1
N
X
i=1
λi νi = ν
)
λi νi = ν
)
The converse inequality ϕx > ϕ1 follows from the decomposition of any ν ∈ [0, 1] as the
convex combination of 1, with weight λ1 = ν, and 0, with weight λ2 = 1 − ν. In particular,
?
ϕx m(ν) > d∞ ν x? m(1)
m(1) + (1 − ν) x? m(0)
m(0) , (−∞, 0]2
(ν)
= d∞ m1 , (−∞, 0]2 = ϕ1 m(ν) ,
as both x? m(0) = x? m(1) = (1, 0) as indicated above.
42
Approachability in Unknown Games
Appendix C. Other technical proofs
C.1. Proof of two facts related to Definition 5
The comments after Definition 5 mentioned two facts, that we now prove. First, that
condition (8) is less restrictive in general than (7). Second, that for continuous target
functions ϕ, the two definitions (7) and (8) coincide, i.e., that that (8) entails (7).
Condition (8) is less restrictive in general than (7). Non-continuous target functions
need to be considered to that end. We consider a toy case when A is reduced to one element
(so that the decision-maker has no decision to make and has to play this action), and the
opponent player chooses elements in R, i.e., d = 1, and more precisely, K = [0, 1]. The
target set equals C = {0} and the target function ϕ is defined as
(
0 if m ∈ [0, 1),
ϕ(m) =
1 if m = 1.
Since d = A = 1, we can identify mt and rt . We consider the sequence mt = 1 − 1/t. We
have
(mT , rT ) −→ (1, 1) ∈ Gϕ ,
since Gϕ = (m, 0) : m ∈ [0, 1) ∪ {1} × [−1, 1] .
On the contrary, for all T , we have Cϕ(mT ) = {0} and therefore,
dp rT , Cϕ(mT ) = rT
does not converge to 0 as T → ∞ (this sequence converges to 1 actually).
Proof that (8) entails (7) under a continuity assumption. We consider a continuous
function ϕ : K → [0, +∞). To show that (8) entails (7), it suffices to show that there exists
a function f : (0, +∞) → (0, +∞) with f (ε) → 0 as ε → 0, such that for all (m, r) ∈ K ×Rd ,
dp r, Cϕ(m) 6 f dp (m, r), Gϕ ;
(43)
the required uniformities with respect to the strategies of the opponent will be carried over.
To that end, the continuity of ϕ will be exploited through the following two properties.
First, Gϕ is closed. Second, since K is bounded, ϕ is actually uniformly continuous: we
denote by ω : (0, +∞) → (0, +∞) its modulus of continuity in the `p –norm, which is a
non-decreasing function that satisfies ω(ε) → 0 as ε → 0.
We denote by (mG , rG ) the projection in `p –norm of (m, r) ∈ K × Rd onto the closed set
Gϕ . By definition, rG ∈ Cϕ(mG ) and
w
w
(44)
dp (m, r), Gϕ = w(m, r) − (mG , rG )wp > km − mG kp .
We also define an element rG0 ∈ Cϕ(m) as follows. If rG ∈ Cϕ(m) then we let rG0 = rG .
Otherwise, ϕ(mG ) − ϕ(m) > 0 and as rG ∈ Cϕ(mG ) , there exists an element r0 ∈ C such
that krG − r0 kp 6 ϕ(mG ). (We recall that C is a closed set and its expansions are closed
expansions.) We denote by d the vector
d=
1
(rG − r0 ) ;
ϕ(mG )
43
Mannor, Perchet, Stoltz
by construction, kdkp 6 1. We introduce a new point rG0 and provide a rewriting of rG :
rG0 = r0 + ϕ(m) d ∈ Cϕ(m)
and
rG = r0 + ϕ(mG ) d .
These two equalities yield that
krG − rG0 kp = ϕ(mG ) − ϕ(m) kdkp 6 ϕ(mG ) − ϕ(m) .
Summarizing, we have in all cases (whether rG belongs to Cϕ(m) or not)
w
krG − rG0 wp 6 ϕ(mG ) − ϕ(m) .
(45)
Since rG0 ∈ Cϕ(m) , we get, by a triangle inequality,
w
w
w
dp r, Cϕ(m) 6 kr − rG0 wp = w(m, r) − (m, rG0 )wp
w
w
w
w
w
w
6 w(m, r) − (mG , rG )wp + w(mG , rG ) − (m, rG )wp + w(m, rG ) − (m, rG0 )wp
6 2 dp (m, r), Gϕ + krG − rG0 kp ,
where the last inequality follows from (44). By (45) and the uniform continuity of ϕ, the
last term in the right-hand side of the display above can be bounded as
krG − rG0 kp 6 ϕ(mG ) − ϕ(m) 6 ω kmG − mk 6 ω dp (m, r), Gϕ ,
where for the last inequality, we used again (44) and the fact that ω is non-decreasing.
Putting all pieces together, we proved (43) with f : x 7→ 2x + ω(x).
C.2. A lemma used in the proof of Theorem 11
Lemma 16 Consider two positive numbers γ1 , γ2 and form the positive sequence (un ) defined by u1 = γ2 and
p
un+1 = un + 2γ1 (n + 1) un + γ2 (n + 1)2
for all n > 1. Then
∀ n > 1,
un 6 max 2γ12 , γ2 n3 .
Proof We proceed by induction and note that the relation is satisfied by construction for
n = 1. Assuming now
that it holds for some n > 1, we show that it is also true for n + 1.
Denoting C = max 2γ12 , γ2 , we get
p
√ p
un+1 = un + 2γ1 (n + 1) un + γ2 (n + 1)2 6 C n3 + 2γ1 C (n + 1) n3 + γ2 (n + 1)2 .
It suffices to show that the latter upper bound is smaller than C (n + 1)3 , which follows
from
√ p
√
2γ1 C (n + 1) n3 + γ2 (n + 1)2 6 2γ1 2C + γ2 n2 + 2γ2 n + γ2 6 3C n2 + 3C n + C ;
indeed, the first inequality comes from bounding (n+1)n3 by 2n4 and expanding
the (n+1)2
√
term, while the second inequality holds because C > γ2 and 2C > 2γ1 2C by definition
of C.
44
Approachability in Unknown Games
Appendix D. Some thoughts on the optimality of target functions
We first define a notion of optimality, based on the classical theory of mathematical orderings, with ≺ (see Definition 6) being seen as a strict partial order, with associated
non-strict partial order denoted by 4 (corresponding to the standard pointwise inequality
6 for functions).
D.1. On the existence of admissible target functions
Definition 17 A target function ϕ is admissible if it is achievable and if there exists no
other achievable target function ϕ0 such that ϕ0 ≺ ϕ.
There might exist several, even an infinite number of, admissible target functions, as
we will show below for Example 1. But there exists always at least one such admissible
function, as we show below in a non-constructive way. We unfortunately were unable to
exhibit general, concrete and closed-form admissible target functions.
Lemma 18 In any unknown game there exists at least one admissible mapping.
Proof The proof is based on an application of Zorn’s lemma. We prove below that the set
T of all achievable target functions ϕ : K → [0, +∞), which is partially ordered for 4, has
the property that every totally ordered subset TΘ = {ϕθ , θ ∈ Θ} has a lower bound in T .
In that case, Zorn’s lemma ensures that the set T contains at least one minimal element:
an element ϕ such that no other element ϕ ∈ T satisfies ϕ ≺ ϕ.
Given a totally ordered subset TΘ , we can define the target function
ϕΘ : m ∈ K 7−→ inf ϕθ (m) ;
θ∈Θ
ϕΘ is of course smaller than any element of TΘ . The point is to show that ϕΘ ∈ T , i.e.,
that ϕΘ is still achievable.
A property that we will use repeatedly below is that if two target functions are such
that ϕ 4 φ, then Gϕ ⊆ Gφ .
Now, by definition, the fact that the ϕθ are achievable means that the compact sets
Gϕθ are each approachable for the game with payoffs (x, m) ∈ ∆(A) × K 7→ (m, x m); in
particular, they are non empty. The compact set
\
GϕΘ =
Gϕθ
θ∈Θ
cannot be empty. Indeed, if it were, fixing any θ0 ∈ Θ, we would have that the subsets
Gϕθ0 \ Gϕθ cover the compact topological space Gϕθ0 . As these subsets are open sets in the
topological space Gϕθ0 , only finitely many of them would be needed for the covering, call
them Gϕθ0 \Gϕj , with j = 1, . . . , N . Since TΘ is totally ordered, one of the sets Gϕj is minimal
for inclusion ⊆, and therefore, one of the sets Gϕθ0 \ Gϕj is maximal for the inclusion ⊆,
say, the one corresponding to j = 1. Therefore, we would have Gϕθ0 \ Gϕ1 = Gϕθ0 . As TΘ is
totally ordered, we would either have ϕ1 4 ϕθ0 and Gϕ1 ⊆ Gϕθ0 , or ϕθ0 4 ϕ1 and Gϕθ0 ⊆ Gϕ1 .
This would lead to Gϕ1 = ∅ in the former case, and Gϕθ0 = ∅ in the latter case: in both
cases, to a contradiction.
45
Mannor, Perchet, Stoltz
In addition, we now prove that for all ε > 0, there exists θε ∈ Θ such that Gϕθε is
included in the open ε–expansion of GϕΘ , which we denote by GϕΘ ,ε . Indeed, denote by Hϕθ
the compact sets Hϕθ = Gϕθ \ GϕΘ ,ε . We have that
\
HϕΘ =
Hϕθ = GϕΘ \ GϕΘ ,ε = ∅ .
θ∈Θ
Therefore, by the same argument as above, we see that there must exist some θε such that
Hϕθε = ∅, which is exactly what we wanted to prove.
So, summarizing, we proved that GϕΘ is non empty and that each of its ε–open expansion is approachable (as it contains an approachable set). As in the proof of Lemma 10,
this means that GϕΘ is a 0–approachable set, thus an approachable set, or put differently,
that ϕΘ is achievable.
D.2. Illustration on Examples 1 and 2
Which response function Ψ should we choose in practice? And are target functions ϕΨ
always admissible? A convenient and natural choice in practice is Ψ = x? , but Example 2
?
shows that unfortunately, ϕx is not always admissible. Example 1 shows that many different target functions ϕΨ may be admissible. It is thus difficult to issue any general theory
on how to choose Ψ and even on the optimality of the class of target functions ϕΨ .
?
?
Example 2: unfortunately, ϕx is not admissible. Indeed, we have ϕ1/2 ≺ ϕx , as
can be seen by carefully comparing the closed-form expressions (41) and (42). On the other
hand, ϕ1/2 is achievable: it suffices to play xt = (1/2, 1/2) at each round. Actually, ϕ1/2 is
of the form ϕΨ for the constant response function Ψ ≡ (1/2, 1/2).
Example 1: all the target functions associated with the Ψx ≡ (x, 1 − x) are
admissible. We illustrate the general existence result of Lemma 18 by showing that in
Example 1 the target functions ϕx = ϕΨx associated with the constant response functions
Ψx ≡ (x, 1 − x) are admissible, for all x ∈ [0, 1]. This corresponds to the case when the
decision-maker chooses the mixed action (x, 1 − x) at all rounds. In particular, the proof of
?
Lemma 8 indicates that ϕ1 = ϕx , the latter is thus admissible unlike in Example 2.
Closed-form expressions of these target functions ϕx will be needed. For ν ∈ [0, 1], when
the decision-maker plays (x, 1 − x) while the vector of vector payoffs is m(ν) , she gets an
average payoff which we denote by r (x, 1 − x), m(ν) and which equals
r (x, 1 − x), m(ν) = x ν m†1 + (1 − ν)m]1 + (1 − x) ν m†2 + (1 − ν)m]2
3
4
0
5
=x ν
+ (1 − ν)
+ (1 − x) ν
+ (1 − ν)
4
3
5
0
x(4 − ν) + 5(1 − x)(1 − ν)
5 − x − ν(5 − 4x)
=
=
.
x(3 + ν) + 5(1 − x)ν
3x + ν(5 − 4x)
The underlying response function Ψx being constant,
no convex decomposition needs to be
(ν)
considered in the defining supremum for ϕx m
and the latter equals
ϕx m(ν) = d∞ r (x, 1 − x), m(ν) , (−∞, 0]2 = max 5 − x − ν(5 − 4x), 3x + ν(5 − 4x) .
46
Approachability in Unknown Games
Since ν 7→ 5 − x − ν(5 − 4x) is decreasing, and ν 7→ 3x + ν(5 − 4x) is increasing, and both
functions take the same value 5/2 − x/2 at ν = 1/2, we get
(
5 − x − ν(5 − 4x) if ν ∈ [0, 1/2],
(ν)
ϕx m
=
3x + ν(5 − 4x)
if ν ∈ [1/2, 1].
Our proof follows the methodology used to prove Lemma 3. We fix any strategy of the
decision-maker achieving a target function ϕ 6 ϕx , for some fixed x ∈ [0, 1], and we show
that necessarily, ϕ = ϕx . We provide a detailed proof of the equality only for m(ν) where ν
lies in the interval (3/4, 1]; but this proof can be adapted in a straightforward manner to
prove the equality as well the intervals [0, 1/4], [1/4, 1/2] and [1/2, 3/4]. As in the proof
of Lemma 3, it suffices to consider the almost sure statement of convergence as in (3); the
uniformity with respect to strategies of the opponent is not needed. All statements below
hold almost surely and the times T and T 0 should be thought of as random variables.
Our argument for ν ∈ [3/4, 1] is based on three sequences of mixed actions. For the
first one, assume that the opponent chooses m† (corresponding to ν = 1) during T stages,
where T can be made arbitrarily large. We denote by (vT , 1 − vT ) the average of the mixed
actions (xt , 1 − xt ) played by the decision-maker during these rounds. The average payoff
vector received equals
3vT
,
5 − vT
whose distance to the negative orthant is 5 − vT . Since ϕx (m† ) = 5 − x and the strategy
achieves ϕ(m† ), where by assumption ϕ(m† ) 6 ϕx (m† ), it holds that lim sup 5 − vT 6 5 − x
as T → ∞. For the sake of compactness, we will denote this fact by 5 − vT . 5 − x. This
entails that lim inf vT > x as T → ∞, a fact that we denote by vT & x.
During the next T stages, we assume that the opponent chooses m] (which corresponds
to ν = 0) and denote by (wT , 1 − wT ) the average of the mixed actions (xt , 1 − xt ) played
by the decision-maker during these rounds. The average payoff vectors received between
rounds T + 1 to 2T , on the one hand, and during rounds 1 to 2T , on the other hand, are
therefore respectively equal to
1 3vT + 5 − wT
5 − wT
and
,
3wT
2 5 − vT + 3wT
so that the distance of the latter to the negative orthant is given by
1
max 3vT + 5 − wT , 5 − vT + 3wT ,
2
which we know is asymptotically smaller than ϕ m(1/2) by achievability of ϕ, where by
assumption
ϕ m(1/2) 6 ϕx m(1/2) = 5/2 + x .
We thus obtained the following system of equations:
vT & x,
3vT − wT . 2x,
−vT + 3wT . 2x.
47
Mannor, Perchet, Stoltz
The sum of the last two inequalities is vT + wT . 2x. Together with the first inequality
vT & x, it leads to wT . x. Substituting in the second inequality, we get 3vT − wT ≈ 2x,
where the ≈ symbol denotes a convergence: 3vT −wT → 2x as T → ∞. Summing the proved
limits −9vT + 3wT ≈ −6x and 8vT & 8x yields −vT + 3wT & 2x, thus −vT + 3wT ≈ 2x.
From the latter limit and 3vT − wT ≈ 2x, we finally get vT ≈ x and wT ≈ x.
Consider now some ν > 3/4. We show that ϕ(ν) > ϕx (ν). To that end, assume that
after the T stages of m† the opponent switches instead to m(3/4) = (3/4) m† + (1/4) m]
during
1−ν
T0 =
T
ν − 3/4
rounds. Note that in this case, the average values of the coefficients for m† and m] used in
the first T + T 0 rounds are proportional to
1+
ν
3 1−ν
=
4 ν − 3/4
4ν − 3
and
0+
1 1−ν
1−ν
=
4 ν − 3/4
4ν − 3
(46)
that is, mT +T 0 = m(ν) = ν m† + (1 − ν) m] was played. We perform first some auxiliary
calculations: by multiplying the equalities in (46) by T , we see that the total number T + T 0
of rounds equals T + T 0 = T /(4ν − 3). In particular, we have
T /(T + T 0 ) = 4ν − 3
and
T 0 /(T + T 0 ) = 1 − (4ν − 3) = 4 − 4ν .
Finally, denoting by uT 0 the average mixed action played by the decision-maker in rounds
T + 1 to T + T 0 , we have that the average vector payoffs during rounds T + 1 to T + T 0 and
during rounds 1 to T + T 0 are respectively equal to
5 − uT 0 − (3/4)(5 − 4uT 0 )
2uT 0 + 5/4
r (uT 0 , 1 − uT 0 ), 3/4 =
=
15/4
3uT 0 + (3/4)(5 − 4uT 0 )
and
2uT 0 + 5/4
.
15/4
(47)
The overall average payoff is given by the distance of this
vector in the supremum norm to
the negative orthant and must be smaller than ϕ m(ν) in the limit, by achievability of ϕ.
However, the said distance of (47) to the negative orthant is bound to be larger than the
second component of (47), which equals
T
T + T0
3vT
5 − vT
+
T0
T + T0
(4ν − 3)(5 − vT ) + (4 − 4ν)
2uT 0 + 5/4
15/4
= (4ν−3)
3vT
5 − vT
+(4−4ν)
15
15
−→ (4ν − 3)(5 − x) + (4 − 4ν) = 5ν + 3x − 4νx = ϕx m(ν) ,
4
4
as T → ∞, where we substituted
the above
limit vT → x.
We thus proved ϕ m(ν) > ϕx m(ν) , as claimed.
48
| 10math.ST
|
Journal of Machine Learning Research (X) (manuscript under review)
Submitted 1/18; Published XX/XX
Expected Policy Gradients for Reinforcement Learning
Kamil Ciosek
Shimon Whiteson
[email protected]
[email protected]
arXiv:1801.03326v1 [stat.ML] 10 Jan 2018
Department of Computer Science, University of Oxford
Wolfson Building, Parks Road, Oxford OX1 3QD
Editor: David Blei and Bernhard Schölkopf
Abstract
We propose expected policy gradients (EPG), which unify stochastic policy gradients (SPG)
and deterministic policy gradients (DPG) for reinforcement learning. Inspired by expected
sarsa, EPG integrates (or sums) across actions when estimating the gradient, instead of
relying only on the action in the sampled trajectory. For continuous action spaces, we first
derive a practical result for Gaussian policies and quadric critics and then extend it to
an analytical method for the universal case, covering a broad class of actors and critics,
including Gaussian, exponential families, and reparameterised policies with bounded support.
For Gaussian policies, we show that it is optimal to explore using covariance proportional
to eH , where H is the scaled Hessian of the critic with respect to the actions. EPG also
provides a general framework for reasoning about policy gradient methods, which we use to
establish a new general policy gradient theorem, of which the stochastic and deterministic
policy gradient theorems are special cases. Furthermore, we prove that EPG reduces the
variance of the gradient estimates without requiring deterministic policies and with little
computational overhead. Finally, we show that EPG outperforms existing approaches on
six challenging domains involving the simulated control of physical systems.
Keywords: policy gradients, exploration, bounded actions, reinforcement learning, Markov
decision process (MDP)
1. Introduction
In reinforcement learning, an agent aims to learn an optimal behaviour policy from trajectories
sampled from the environment. In settings where it is feasible to explicitly represent the
policy, policy gradient methods (Sutton et al., 2000; Peters and Schaal, 2006, 2008b; Silver
et al., 2014), which optimise policies by gradient ascent, have enjoyed great success, especially
with large or continuous action spaces. The archetypal algorithm optimises an actor, i.e., a
policy, by following a policy gradient that is estimated using a critic, i.e., a value function.
The policy can be stochastic or deterministic, yielding stochastic policy gradients (SPG)
(Sutton et al., 2000) or deterministic policy gradients (DPG) (Silver et al., 2014). The theory
underpinning these methods is quite fragmented, as each approach has a separate policy
gradient theorem guaranteeing the policy gradient is unbiased under certain conditions.
Furthermore, both approaches have significant shortcomings. For SPG, variance in the
gradient estimates means that many trajectories are usually needed for learning. Since
gathering trajectories is typically expensive, there is a great need for more sample efficient
methods.
c X Kamil Ciosek and Shimon Whiteson.
License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided
at http://jmlr.org/papers/v/ciosekwhiteson17.html.
Ciosek and Whiteson
DPG’s use of deterministic policies mitigates the problem of variance in the gradient
but raises other difficulties. The theoretical support for DPG is limited since it assumes
a critic that approximates ∇a Q when in practice it approximates Q instead. In addition,
DPG learns off-policy 1 , which is undesirable when we want learning to take the cost of
exploration into account. More importantly, learning off-policy necessitates designing a
suitable exploration policy, which is difficult in practice. In fact, efficient exploration in DPG
is an open problem and most applications simply use independent Gaussian noise or the
Ornstein-Uhlenbeck heuristic (Uhlenbeck and Ornstein, 1930; Lillicrap et al., 2015).
This article, which extends our previous work (Ciosek and Whiteson, 2018), proposes a
new approach called expected policy gradients (EPG) that unifies policy gradients in a way
that yields both theoretical and practical insights. Inspired by expected sarsa (Sutton and
Barto, 1998; van Seijen et al., 2009), the main idea is to integrate across the action selected
by the stochastic policy when estimating the gradient, instead of relying only on the action
selected during the sampled trajectory.
The contributions of this paper are threefold. First, EPG enables two general theoretical
contributions (Section 3.1): 1) a new general policy gradient theorem, of which the stochastic
and deterministic policy gradient theorems are special cases, and 2) a proof that (Section 3.2)
EPG reduces the variance of the gradient estimates without requiring deterministic policies
and, for the Gaussian case, with no computational overhead over SPG. Second, we define
practical policy gradient methods. For the Gaussian case (Section 4), the EPG solution is not
only analytically tractable but also leads to a principled exploration strategy (Section 4.2)
for continuous problems, with an exploration covariance that is proportional to eH , where H
is the scaled Hessian of the critic with respect to the actions. We present empirical results
(Section 6) confirming that this new approach to exploration substantially outperforms
DPG with Ornstein-Uhlenbeck exploration in six challenging MuJoCo domains. Third, we
provide a way of deriving tractable EPG methods for the general case of policies coming
from a certain exponential family (Section 5) and for critics that can be reparameterised as
polynomials, thus yielding analytic EPG solutions that are tractable for a broad class of
problems and essentially making EPG a universal method. Finally, in Section 7, we relate
EPG to other RL approaches.
2. Background
A Markov decision process (Puterman, 2014) is a tuple (S, A, R, p, p0 , γ) where S is a set
of states, A is a set of actions (in practice either A = Rd or A is finite), R(s, a) is a
reward function, p(s0 | a, s) is a transition kernel, p0 is an initial state distribution, and
γ ∈ [0, 1) is a discount factor. A policy π(a | s) is a distribution over actions given a state.
We denote trajectories as τ π = (s0 , a0 , r0 , s1 , a1 , r1 , . . . ), where s0 ∼ p0 , at ∼ π(· | st−1 )
and rt is a sample
reward. A policy π induces a Markov process with transition kernel
R
0
pπ (s | s) = a dπ(a | s)p(s0 | a, s) where we use the symbol dπ(a | s) to denote Lebesgue
integration against the measure π(a | s) where s is fixed. We assume the induced Markov
process is ergodic with a single
P invariant measure defined for the whole state space. The
value function is V π = Eτ [ i γi ri ] where actions are sampled from π. The Q-function is
1. We show in this article that, in certain settings, off-policy DPG is equivalent to EPG, our on-policy
method.
2
Expected Policy Gradients
Qπ (a | s) = ER [r | s, a] + γEpπ (s0 |s) [V π (s0 ) | s] and the advantage function Ris Aπ (a | s) =
Qπ (a | s) − V π (s). An optimal policy maximises the total return J = s dp0 (s)V π (s).
Since we consider only on-policy learning with just one current policy, we drop the π
super/subscript where it is redundant.
If π is parameterised by θ, then stochastic policy gradients (SPG) (Sutton et al., 2000;
Peters and Schaal, 2006, 2008b) perform gradient ascent on ∇J, the gradient of J with
respect to θ (gradients without a subscript are always with respect to θ). For stochastic
policies, we have:
R
R
∇J = s dρ(s) a dπ(a | s)∇ log π(a | s)(Q(a, s) + b(s)),
(1)
where ρ is the discounted-ergodic occupancy measure, defined in the Appendix, and b(s) is
a
R baseline, which can be any function that depends on the state but not the action, since
a dπ(a | s)∇ log π(a | s)b(s) = 0. Typically, because of ergodicity and Lemma 18 (see
Appendix), we can approximate (1) from samples from a trajectory τ of length T :
ˆ = PT γ t ∇ log π(at | st )(Q̂(at , st ) + b(st )),
∇J
t=0
(2)
where Q̂(at , st ) is a critic, discussed below. If the policy is deterministic (we denote it π(s)),
we can use deterministic policy gradients (Silver et al., 2014) instead:
R
∇J = s dρ(s)∇π(s) [∇a Q(a, s)]a=π(s) .
(3)
This update is then approximated using samples:
h
i
ˆ = PT γ t ∇π(s) ∇a Q̂(a, st )
∇J
t=0
a=π(st )
.
(4)
Since the policy is deterministic, the problem of exploration is addressed using an external
source of noise, typically modelled using a zero-mean Ornstein-Uhlenbeck (OU) process
(Uhlenbeck and Ornstein, 1930; Lillicrap et al., 2015) parameterised by ψ and σ:
ni ← −ni−1 ψ + N (0, σI)
a ∼ π(s) + ni .
(5)
In (2) and (4), Q̂ is a critic that approximates Q and can be learned by sarsa (Rummery
and Niranjan, 1994; Sutton, 1996):
Q̂(at , st ) ←Q̂(at , st ) + α rt+1 + γ Q̂(st+1 , at+1 ) − Q̂(at , st ) .
(6)
Alternatively, we can use expected sarsa (Sutton and Barto, 1998; van Seijen et al., 2009),
which marginalises out at+1 , the distribution over which is specified by the known policy, to
reduce the variance in the update:
R
Q̂(st , at ) ← Q̂(at , st ) + α rt+1 + γ a dπ(a | s)Q̂(st+1 , a) − Q̂(at , st ) .
(7)
We could also use advantage learning (Baird et al., 1995) or LSTDQ (Lagoudakis and Parr,
2003). If the critic’s function approximator is compatible, then the actor, i.e., π, converges
(Sutton et al., 2000).
3
Ciosek and Whiteson
Instead of learning Q̂, we can set b(s) = −V (s) so that Q(a, s) + b(s) = A(s, a) and then
use the TD error δ(r, s0 , s) = r + γV (s0 ) − V (s) as an estimate of A(s, a) (Bhatnagar et al.,
2008):
ˆ = PT γ t ∇ log π(at | st )(r + γ V̂ (s0 ) − V̂ (s)),
∇J
t=0
(8)
where V̂ (s) is an approximate value function learned using any policy evaluation algorithm.
(8) works because E [δ(r, s0 , s) | a, s] = A(s, a), i.e., the TD error is an unbiased estimate
of the advantage function. The benefit of this approach is that it is sometimes easier to
approximate V than Q and that the return in the TD error is unprojected, i.e., it is not
distorted by function approximation. However, the TD error is noisy, introducing variance
in the gradient.
To cope with this variance, we can reduce the learning rate when the variance of the
gradient would otherwise explode, using, e.g., Adam (Kingma and Ba, 2014), natural
policy gradients (Kakade, 2002; Amari, 1998; Peters and Schaal, 2008a) or Newton’s method
(Furmston and Barber, 2012). However, this results in slow learning when the variance is
high. See Section 7 for further discussion on variance reduction techniques.
3. Expected Policy Gradients
In this section, we propose expected policy gradients (EPG). First, we introduce IπQ (s) to
denote the inner integral in (1):
Z
Z
∇J = dρ(s) dπ(a | s)∇ log π(a | s)(Q(a, s) + b(s))
s
|a
{z
}
IπQ (s)
Z
=
Z
dπ(a | s)∇ log π(a | s)Q(a, s)
dρ(s)
Zs
=
a
dρ(s)IπQ (s).
(9)
s
This suggests a new way to write the approximate gradient2 , using Lemma 18 (see Appendix):
ˆ =
∇J
T
X
t=0
γ t IπQ̂ (st ),
| {z }
where
IπQ̂ (s) =
Z
dπ(a | s)∇ log π(a | s)Q̂(a, s).
(10)
a
gt
This approach makes explicit that one step in estimating the gradient is to evaluate an
integral included in the term IπQ̂ (s). The main insight behind EPG is that, given a state,
IπQ̂ (s) is expressed fully in terms of known quantities. Hence we can manipulate it analytically
to obtain a formula or we can just compute the integral using numerical quadrature if an
analytical solution is impossible (in Section 5.1 we show that this is rare). For a discrete
action space, IπQ̂ (st ) becomes a sum over actions.
2. The idea behind EPG was also independently and concurrently developed as Mean Actor Critic (Asadi
et al., 2017), though only for discrete actions and without a supporting theoretical analysis.
4
Expected Policy Gradients
SPG as given in (2) performs this quadrature using a simple one-sample Monte Carlo
method as follows, using the action at ∼ π(· | st ).
IπQ̂ (s)
Z
dπ(a | s)∇ log π(a | s)Q̂(a, s) ≈ ∇ log π(at | st )(Q̂(at , st ) + b(st ))
=
a
Moreover, SPG assumes that the action at used in the above estimation is the same action
that is executed in the environment. However, relying on such a method is unnecessary.
In fact, the actions used to interact with the environment need not be used at all in the
evaluation of IˆπQ (s) since a is a bound variable in the definition of IπQ (s). The motivation is
thus similar to that of expected sarsa but applied to the actor’s gradient estimate instead of
the critic’s update rule. EPG, shown in Algorithm 1, uses (10) to form a policy gradient
algorithm that repeatedly estimates IˆπQ (s) with an integration subroutine.
Algorithm 1 Expected policy gradients
1: s ← s0 , t ← 0
2: initialise optimiser, initialise policy π parameterised by θ
3: while not converged do
4:
gt ← γ t do-integral(Q̂, s, πθ )
. gt is the estimated policy gradient as per (10)
5:
θ ← θ + optimiser.update(gt )
6:
a ∼ π(·, s)
7:
s0 , r ← simulator.perform-action(a)
8:
Q̂.update(s, a, r, s0 )
9:
t←t+1
10:
s ← s0
11: end while
One of the motivations of DPG was precisely that the simple one-sample Monte-Carlo
quadrature implicitly used by SPG often yields high variance gradient estimates, even with
a good baseline. To see why, consider Figure 1 (left). A simple Monte Carlo method
evaluates the integral by sampling one or more times from π(a | s) (blue) and evaluating
∇µ log π(a | s)Q(a, s) (red) as a function of a. A baseline can decrease the variance by
adding a multiple of ∇µ log π(a | s) to the red curve, but the problem remains that the
red curve has high values where the blue curve is almost zero. Consequently, substantial
variance persists, whatever the baseline, even with a simple linear Q-function, as shown in
Figure 1 (right). DPG addressed this problem for deterministic policies but EPG extends
it to stochastic ones. We show in Section 5 that an analytical EPG solution, and thus the
corresponding reduction in the variance, is possible for a wide array of critics. We also
discuss the rare case where numerical quadrature is necessary in Section 5.4.
We now provide our most general results, which apply to EPG in any setting.
3.1 General Policy Gradient Theorem
We begin by stating our most general result, showing that EPG can be seen as a generalisation
of both SPG and DPG. To do this, we first state a new general policy gradient theorem.
5
Ciosek and Whiteson
15
SPG update
10
5
−1.0
−0.5
0
0.0
variance of MC
policy PDF
0.5
1.0
action
−0.70
−0.60
−0.50
−0.40
−0.30
baseline
Figure 1: At left, π(a | s) for a Gaussian policy with mean µ = θ = 0 at a given state and
constant σ 2 (blue) and ∇θ log π(a | s)Q(a, s) for Q = 12 + 12 a (red). At right, the variance
of a simple single-sample Monte Carlo estimator as a function of the baseline. In a simple
multi-sample Monte Carlo method, the variance would go down as the number of samples.
Theorem 1 (General Policy Gradient Theorem) If π(·, s) is a normalised Lebesgue
measure for all s, then
Z
Z
∇J = dρ(s) ∇V (s) − dπ(a|s)∇Q(a, s) .
s
a
|
{z
}
IG (s)
Proof We begin by expanding the following expression.
R
R
R
R
R
0
0
s dρ(s) a dπ(a|s)∇Q(a, s) = Rs dρ(s) Ra dπ(a|s)∇(R(a, s) + γR s0 dp(s | s, a)V (s ))
= s dρ(s) a dπ(a|s)(∇R(a, s) +γ s0 dp(s0 | s, a)∇V (s0 ))
| {z }
0
R
R
= γ s dρ(s) s0 dpπ (s0 | s)∇V (s0 )
R
R
= s dρ(s)∇V (s) − s dp0 (s)∇V (s)
|
{z
}
∇J
R
= s dρ(s)∇V (s) − ∇J.
The first equality follows by expanding the definition of Q and the penultimate one follows
from Lemma 19 in the Appendix. Then the theorem follows by rearranging terms.
The crucial benefit of Theorem 1 is that it works for all policies, both stochastic and
deterministic, unifying previously separate derivations for the two settings. To show this,
in the following two corollaries, we use Theorem 1 to recover the stochastic policy gradient
theorem (Sutton et al., 2000) and the deterministic policy gradient theorem (Silver et al.,
2014), in each case by introducing additional assumptions to obtain a formula for IG (s)
expressible in terms of known quantities.
Corollary 2 (Stochastic Policy Gradient Theorem) If π(· | s) is differentiable, then
R
R
R
∇J = s dρ(s)IG (s) = s dρ(s) a dπ(a | s)∇ log π(a | s)Q(a, s).
6
Expected Policy Gradients
Proof We obtain the following by expanding ∇V .
R
R
R
∇V = ∇ a dπ(a|s)Q(a, s) = a da(∇π(a|s))Q(a, s) + a dπ(a|s)(∇Q(a, s))
R
We obtain IG (s) = a dπ(a | s)∇ log π(a | s)Q(a, s) = IπQ (s) by plugging this into the definition of IG (s). We obtain ∇J by invoking Theorem 1 and plugging in the above expression
for IG (s).
We now recover the DPG update introduced in (3).
Corollary 3 (Deterministic Policy Gradient Theorem) If π(· | s) is a Dirac-delta
measure (i.e., a deterministic policy) and Q(·, s) is differentiable, then
R
R
∇J = s dρ(s)IG (s) = s dρ(s)∇π(s) [∇a Q(a, s)]a=π(s) .
We overload
the notation of π slightly. We denote by π(s) the action taken at state s, i.e.
R
π(s) = a adπ(a | s), where π(· | s) is the corresponding Dirac-delta measure.
Proof We begin by expanding the term for ∇V (s), which will be useful later on.
∇V (s) = ∇Q(π(s), a) = [∇Q(a, s)]a=π(s) + [∇a Q(a, s)]a=π(s) ∇π(s)
(11)
The above results from applying the multivariate chain rule—observe that both π(s) and
Q(a, s) depend on the policy parameters θ; hence, the dependency appears twice in Q(π(s), s).
We proceed to obtain an expression for IG (s).
R
IG (s) = ∇V (s) − a dπ(a|s)∇Q(a, s)
= ∇V (s) − [∇Q(a, s)]a=π(s)
= ∇π(s) [∇a Q(a, s)]a=π(s) .
Here, the second equality follows by observing that the policy is a Dirac-delta and the third
one follows from using (11). We can then obtain ∇J by invoking Theorem 1 and plugging
in the above expression for IG (s).
These corollaries show that the choice between deterministic and stochastic policy
gradients is fundamentally a choice of quadrature method. Hence, the empirical success of
DPG relative to SPG (Silver et al., 2014; Lillicrap et al., 2015) can be understood in a new
light. In particular, it can be attributed, not to a fundamental limitation of stochastic policies
(indeed, stochastic policies are sometimes preferred), but instead to superior quadrature.
DPG integrates over Dirac-delta measures, which is known to be easy, while SPG typically
relies on simple Monte Carlo integration. Thanks to EPG, a deterministic approach is no
longer required to obtain a method with low variance.
3.2 Variance Analysis
We now prove that for any policy, the EPG estimator of (10) has lower variance than the
SPG estimator of (2).
7
Ciosek and Whiteson
Lemma 4 If for all s ∈ S, the random variable ∇ log π(a | s)Q̂(a, s) where a ∼ π(·|s) has
nonzero variance, then
hP
i
hP
i
∞
∞
t ∇ log π(a | s )(Q̂(a , s ) + b(s )) > V
t I Q̂ (s ) .
Vτ
γ
γ
π
t
t
t
t
t
τ
t
t=0
t=0
Proof Both random variables have the same mean so we need only show that:
2
2
P∞ t
P∞ t Q̂
Eτ
> Eτ
.
t=0 γ ∇ log π(at | st )(Q̂(at , st ) + b(st ))
t=0 γ Iπ (st )
We start by applying Lemma 21 to the lefthand side and setting X = X1 (st ) = γ t ∇ log π(at |
st )(Q̂(at , st ) + b(st )) where at ∼ π(at |st ). This shows that
2
P∞ t
Eτ
t=0 γ ∇ log π(at | st )(Q̂(at , st ) + b(st ))
is the total return of the MRP (p, p0 , u1 , γ 2 ), where
2
u1 = VX1 (x|s) [x] + EX1 (x|s) [x] + 2γEX1 (x|s) [x] Ep(s0 |s) V (s0 ) .
Likewise, applying Lemma 21 again to the righthandside, instantiating X as a deterministic
2
P∞ t Q̂
random variable X2 (st ) = IπQ̂ (st ), we have that Eτ
is the total return
t=0 γ Iπ (st )
of the MRP (p, p0 , u2 , γ 2 ), where
2
u2 = EX2 (x|s) [x] + 2γEX2 (x|s) [x] Ep(s0 |s) V (s0 ) .
Note that EX1 (x|s) [x] = EX2 (x|s) [x] and therefore u1 ≥ u2 . Furthermore, by assumption of
the lemma, the inequality is strict. The lemma then follows by applying Observation 22.
For convenience, Lemma 4 also assumes infinite length trajectories. However, this is not
a practical limitation since all policy gradient methods implicitly assume trajectories are
long enough to be modelled as infinite. Furthermore, a finite trajectory variant also holds,
though the proof is messier.
Lemma 4’s assumption is reasonable since the only way a random variable ∇ log π(a |
s)Q̂(a, s) could have zero variance is if it were the same for all actions in the policy’s support
(except for sets of measure zero), in which case optimising the policy would be unnecessary.
Since we know that both the estimators of (2) and (10) are unbiased, the estimator with
lower variance has lower MSE. Moreover, we observe that Lemma 4 holds for the case where
the computation of IπQ̂ is exact. Section 5 shows that this is often possible.
4. Expected policy gradients for Gaussian Policies
EPG is particularly useful when we make the common assumption of a Gaussian policy: we
can then perform the integration analytically under reasonable conditions. We show below
8
Expected Policy Gradients
Algorithm 2 Gaussian policy gradients
1: s ← s0 , t ← 0
2: initialise optimiser
3: while not converged do
4:
gt ← γ t do-integral-Gauss(Q̂, s, πθ )
5:
θ ← θ + optimiser.update(gt )
. policy parameters θ are updated using gradient
1/2
1/2
6:
Σs ← get-covariance(Q̂, s, πθ )
. Σs computed from scratch
7:
a ∼ π(· | s)
. π(· | s) = N (µs , Σs )
8:
s0 , r ← simulator.perform-action(a)
9:
Q̂.update(s, a, r, s0 )
10:
t←t+1
11:
s ← s0
12: end while
Algorithm 3 Gaussian integrals
function do-integral-Gauss(Q̂, s, πθ )
Q
2:
Iπ(s),µ
← (∇µs )∇a Q̂(a = µs , s)
s
1:
3:
4:
. Use Lemma 5
Q
Iπ(s),µ
s
return
end function
5:
function get-covariance(Q̂, s, πθ )
7:
H ← compute-Hessian(Q̂(µs , s))
8:
return σ0 ecH
9: end function
6:
. Use Lemma 6
(see Corollary 7) that the update to the policy mean computed by EPG is equivalent to
the DPG update. Moreover, we derive a simple formula for the covariance (see Lemma 6).
Algorithms 2 and 3 show the resulting special case of EPG, which we call Gaussian policy
gradients (GPG).
Surprisingly, GPG is on-policy but nonetheless fully equivalent to DPG, an off-policy
method, with a particular form of exploration. Hence, GPG, by specifying the policy’s
covariance, can be seen as a derivation of an exploration strategy for DPG. In this way,
GPG addresses an important open question. As we show in Section 6, this leads to improved
performance in practice.
The computational cost of GPG is small: while it must store a Hessian matrix H(a, s) =
∇2a Q̂(a, s), its size is only d × d, where A = Rd , which is typically small, e.g., d = 6 for
HalfCheetah-v1, one of the MuJoCo tasks we use for our experiments in Section 6. This
Hessian is the same size as the policy’s covariance matrix, which any policy gradient must
store anyway, and should not be confused with the Hessian with respect to the parameters of
the neural network, as used with Newton’s or natural gradient methods (Peters and Schaal,
2008a; Furmston et al., 2016), which can easily have thousands of entries. Hence, GPG
obtains EPG’s variance reduction essentially for free.
9
Ciosek and Whiteson
4.1 Analytical Quadrature for Gaussian Policies
We now derive a lemma supporting GPG.
Lemma 5 (Gaussian Policy Gradients) If the policy is Gaussian, i.e. π(·|s) ∼ N (µs , Σs )
1/2
with µs and Σs parameterised by θ, and the critic is of the form Q(a, s) = a> A(s)a +
a> B(s) + const, then
>
Q
Q
Q
Iπ (s) = Iπ(s),µs I
,
1/2
π(s),Σs
where the mean and covariance components are given by:
Q
Iπ(s),µ
= (∇µs )B(s),
s
IQ
1/2
π(s),Σs
1/2
= (∇Σ1/2
s )Σs A(s).
(12)
Proof For ease of presentation, we prove the lemma for a one-dimensional action space,
1/2
where µs , a ∈ R and Σs = σs is the standard deviation (we drop the suffix s in µs and
σs in the subsequent formulae). First, note that the constant term in the critic does not
influence the value of IπQ (s) since it depends only on the state and not on the action and
can be treated as a baseline. Observe that
Q
Iπ(s),µ
= (∇µ)Eπ [∇µ log π(a|s)Q(a, s)]
h
i
h
i
= (∇µ) Eπ ∇µ log π(a|s)a> B(s) Eπ ∇µ log π(a|s)a> A(s)a .
We consider the linear term and the quadric term separately. For the linear term we have:
a−µ
Eπ [∇µ log π(a|s)aB(s)] = Eπ
B(s)a
σ2
1
= 2 Eπ B(s)a2 − B(s)aµ =
σ
1
= 2 B(s)Eπ a2 − B(s)µEπ [a]
σ
1
= 2 B(s) σ 2 + µ2 − B(s)µµ = B(s).
σ
For the quadric term we have:
1
Eπ ∇µ log π(a|s)a2 A(s) = 2 Eπ a2 A(s)(a − µ) =
σ
A(s)
= 2 Eπ a3 − a2 µ =
σ
A(s) 3
= 2 µ + 3µσ 2 − (µ2 + σ 2 )µ =
σ
= 2A(s)µ
Summing the two terms yields:
Q
Iπ(s),µ
= ∇µ(2A(s)µ + B(s)).
10
Expected Policy Gradients
We now calculate the integrals for the standard deviation, again beginning with the
linear term:
(a − µ)2
1
Eπ [∇σ log π(a|s)aB(s)] = Eπ B(s)a
− B(s)a
σ3
σ
1 3
B(s)
2
2
= σ Eπ σ2 a − 2a µ + µ a − a
1
µ3 + 3µσ 2 − 2(µ2 + σ 2 )µ + µ3 − µ
= B(s)
σ
σ2
=
B(s)
σ (µ
− µ) = 0.
For the quadric term we have:
Eπ ∇σ log π(a|s)A(s)a2
2
2 (a − µ)
21
= Eπ A(s)a
− A(s)a
σ3
σ
1 4
A(s)
3
2
2
= σ Eπ σ2 (a − 2a µ + µ a ) − a2
1 4
= A(s)
(µ + 6µ2 σ 2 + 3σ 4 − 2µ(µ3 + 3µσ 2 ) + µ2 (µ2 + σ 2 )) − (µ2 + σ 2 )
σ
σ2
= 2A(s)σ.
Summing the two terms yields:
Q
Iπ(s),σ
= ∇σ(2A(s)σ).
The multivariate case (i.e., with a multi-dimensional action space) can be obtained using
the method developed in Section 5 later in the paper by observing that the multivariate
normal distribution is in the parametric family given by (15) with the sufficient statistic
vector T containing the vector a and the vectorised matrix aa> , both of which are polynomial
in a, and hence Lemma 8 is applicable.
While Lemma 5 requires the critic to be quadric in the actions, this assumption is not very
restrictive since the coefficients B(s) and A(s) can be arbitrary continuous functions of the
state, e.g., a neural network.
4.2 Exploration using the Hessian
Equation (12) suggests that we can include the covariance in the actor network and learn it
along with the mean, using the update rule:
1/2
Σ1/2
← Σ1/2
s
s + αΣs H(s).
(13)
However, another option is to compute the covariance from scratch at each iteration by
analytically computing the result of applying (13) infinitely many times, as in the following
lemma.
Lemma 6 (Exploration Limit) The iterative procedure defined by (13) applied n times
1/2
using the diminishing learning rate α = 1/n converges to Σs ∝ eH(s) as n → ∞.
11
Ciosek and Whiteson
eigenvalue increases
sharp maximum,
very little exploration
sharp minimum,
lots of exploration
moderate exploration
Figure 2: The parabolas show different possible curvatures of the critic Q̂(·, s). We set
exploration to be the strongest for sharp mimima, on the left side of the figure. The
exploration strength then decreases as we move towards the right. There is almost no
exploration to the far right, where we have a sharp maximum.
1/2
1/2
1/2
1/2
Proof Consider the sequence (Σs )1 = σ0 I, (Σs )n = (Σs )n−1 + n1 (Σs )n−1 H(s). We
diagonalise the Hessian as H(s) = U ΛU > for some orthonormal matrix U and obtain the
following expression for the n-th element of the sequence:
(Σ1/2
s )n+1 = (I +
1
1
H(s))n σ0 = U (I + Λ)n U > σ0 .
n
n
Since we have limn→∞ (1 − n1 λ)n = eλ for each eigenvalue of the Hessian, we obtain the
identity:
1
lim U (I + Λ)n U > σ0 = σ0 eH(s) .
n→∞
n
The practical implication of Lemma 6 is that, in a policy gradient method, it is justified3
to use Gaussian exploration with covariance proportional to ecH for some reward scaling
constant c. Thus, by exploring with (scaled) covariance ecH , we obtain a principled alternative
to the Ornstein-Uhlenbeck heuristic of (5). Our results below show that it also performs
much better in practice.
Lemma 6 has an intuitive interpretation. If H(s) has a large positive eigenvalue λ, then
Q̂(s, ·) has a sharp minimum along the corresponding eigenvector, and the corresponding
eigenvalue of Σ1/2 is eλ , i.e., also large. This is easiest to see with a one-dimensional action
space, where the Hessian and its only eigenvalue are just the same scalar. The exploration
mechanism in the one-dimensional case is illustrated in Figure 2. The idea is simple: the
larger the eigenvalue the worse the minimum we are in and the more exploration we need to
leave it. On the other hand, if λ is negative, then Q̂(s, ·) has a maximum and so eλ is small,
since exploration is not needed.
In the multi-dimensional case, the critic can have saddle points, as shown in Figure 3.
For the case shown in the figure, we explore little along the blue eigenvector (since the
intersection of Q(·, s) with the blue plane shows a maximum) and much more along the red
3. Lemma 6 relies crucially on the use of step sizes diminishing in the length of the trajectory, rather than
finite step sizes. Therefore, the step sequence serves as a useful intermediate stage between simply taking
one PG step of (13) and using finite step sizes, which would mean that the covariance would converge
either to zero or diverge to infinity.
12
Expected Policy Gradients
Q(·, s)
u1
u2
Figure 3: In multi-dimensional action spaces, the critic Q̂(·, s) can have saddle points. In
this case, we define exploration along each eigenvector separately.
eigenvector (since the intersection of Q(·, s) with the red plane shows a minimum, which we
want to escape). In essence, we apply the one-dimensional reasoning shown in Figure 2 to
each plane separately, where the planes are spanned by the corresponding eigenvector and
the z-axis. This way, we can escape saddle points and minima.4
4.3 Action Clipping
We now describe how GPG works in environments where the action space has bounded
support5 . This setting occurs frequently in practice, since real-world systems often have
physical constraints such as a bound on how fast a robot arm can accelerate. The typical
solution to this problem is simply to start with a policy πb with unbounded support and
then, when an action is to be taken, clip it to the desired range as follows
a ∼ π(a | s)
equivalent to a = max(min(b, 1), 0) with b ∼ πb (b | s).
(14)
The justification for this process is that we can simply treat the clipping operation
max(min(b, 1), 0) as part of the environment specification. Formally, this means that
we transform the original MDP M defined as M = (S, A, R, p, p0, γ) with A = [0, 1]d into
another MDP M 0 = (S, A0 , R, p0 , p0, γ), where A0 = Rd and p0 is defined as
p0 (s0 |b, s) = p(s0 | max(min(b, 1), 0), s).
Since M 0 has an unbounded action space, we can use the RL machinery for unbounded
actions to solve it. Since any MDP is guaranteed to have an optimal deterministic policy,
? : S → A. Now, π ? can be transformed into a policy
we call this deterministic solution πD
D
?
for M of the form max(min(πD (s), 1), 0). In practice, the MDP M 0 is never constructed
explicitly—the described process in equivalent to using an RL algorithm meant for A = Rd
and then, when the action is generated, simply clipping it (Algorithm 4).
4. Of course the optimisation is still local and there is no guarantee of finding a global optimum—we can
merely increase our chances.
5. We assume without loss of generality that the support interval is [0, 1].
13
Ciosek and Whiteson
Algorithm 4 Policy gradients with clipped actions.
1: s ← s0 , t ← 0
2: initialise optimiser, initialise policy π parameterised by θ
3: while not converged do
4:
gt ← γ t do-integral(Q̂b , s, πθ )
5:
θ ← θ + optimiser.update(gt )
6:
b ∼ π(·, s)
? (s), 1), 0).
7:
a = c(b)
. Clipping function c(b) = max(min(πD
0
8:
s , r ← simulator.perform-action(a)
9:
Q̂b .update(s, b, r, s0 )
. Update using the pre-clipping action b.
10:
t←t+1
11:
s ← s0
12: end while
g(b)
{b : g 0 (b) 6= 0}
π
bL
b1
b2
b
Figure 4: Vanishing gradients when using hard clipping. The agent cannot determine
whether b is too small or too large from b1 and b2 alone. It is necessary to sample from the
interval {b : g 0 (b) 6= 0} in order to obtain a meaningful policy update but this is unlikely for
the current policy (shown as the red curve).
However, while such an algorithm does not introduce new bias in the sense that reward
obtained in M and M 0 will be the same, it can lead to problems with slow convergence in
the policy gradient settings. To see why, consider Figure 4.
With hard clipping, the agent cannot distinguish between b1 and b2 since squashing
reduces them both to the same value, i.e., g(b1 ) = g(b2 ). Hence, the corresponding Q values
are identical and, based on trajectories using b1 and b2 , there is no way of knowing how
the mean of the policy should be adjusted. In order to get a useful gradient, a b? has to
be chosen which falls into the interval (−∞, bL ]. Since the b’s are samples from a Gaussian
with infinite support, it will eventually happen and a nonzero gradient will be obtained.
However, if this interval falls into a distant part of the tail of πb , convergence will be slow.
However, this problem is mitigated with GPG. To see why, consider Figure 5. Once the
policy shifts into the flat area, the critic becomes constant. A constant critic has a zero
Hessian, generating a boost to exploration by increasing the standard deviation of the policy,
making it much more likely that a point b < bL is sampled and a useful gradient is obtained.
14
Expected Policy Gradients
g(b)
{b : g 0 (b) 6= 0}
bL
b
Figure 5: GPG avoids the vanishing gradient problem. Once a policy (denoted in red)
enters the flat area entering the flat area b > bL , exploration immediately increases (the new
distribution is in blue).
Another way of mitigating the hard clipping problem is to use a differentiable squashing
function, which we describe in Section 5.
4.4 Quadric Critics and their Approximations
Gaussian policy gradients require a quadric critic given the state. This assumption, which is
different from assuming a quadric dependency on the state, is typically sufficient for two
reasons. First, discrete-time linear quadratic regulators (LQR) with time-varying feedback,
a class of problems widely studied in classical control theory, are known to have a Q-function
that is quadric in the action vector given the state (Crassidis and Junkins, 2011, Equation
8.81)6 . Second, it is often assumed (Li and Todorov, 2004) that a quadric critic (or a quadric
approximation to a general critic) is enough to capture enough local structure to preform
a policy optimisation step, in much the same way as Newton’s method for deterministic
unconstrained optimisation, which locally approximates a function with a quadric, can be
used to optimise a non-quadric function across several iterations. In Corollary 7 below, we
describe such an approximation method applied to GPG where we approximate Q with a
quadric function in the neighbourhood of the policy mean.
Corollary 7 (Approximate Gaussian Policy Gradients with an Arbitrary Critic)
1/2
1/2
If the policy is Gaussian, i.e. π(·|s) ∼ N (µs , Σs ) with µs and Σs parameterised by θ as
in Lemma 5 and any critic Q(a, s) doubly differentiable with respect to actions for each state,
1/2
1/2
Q
then Iπ(s),µ
≈ (∇µs )∇a Q(a = µs , s) and I Q
)Σs H(µs , s), where H(µs , s)
1/2 ≈ (∇Σs
s
π(s),Σs
is the Hessian of Q with respect to a, evaluated at µs for a fixed s.
Proof We begin by approximating the critic (for a given s) using the first two terms of the
Taylor expansion of Q in µs .
Q(a, s) ≈ Q(µs , s) + (a − µs )> [∇a Q(a, s)]a=µs + 12 (a − µs )> H(µs , s)(a − µs )
= 12 a> H(µs , s)a + a> [∇a Q(a, s)]a=µs − H(µs , s)µs + const.
6. Indeed, the Hessian discussed in Section 4.2 can be considered a type of reward model.
15
Ciosek and Whiteson
Because of the series truncation, the function on the righthand side is quadric and we can
then use Lemma 5:
Q
Iπ(s),µ
= ∇µs (2 12 H(µs , s)µs + [∇a Q(a, s)]a=µs − H(µs , s)µs )
s
= ∇µs [∇a Q(a, s)]a=µs
I
1/2
Q
1/2
π(s),Σs
1/2
= ∇Σ1/2 (2 12 H(µs , s)Σs ) = ∇Σ1/2 H(µs , s)Σs .
s
s
To actually obtain the Hessian, we could use automatic differentiation to compute it
analytically. Sometimes this may not be possible—for example when ReLU units are used,
the Hessian is always zero. In these cases, we can approximate the Hessian by generating a
number of random action-values around µs , computing the Q values, and (locally) fitting a
quadric, akin to sigma-point methods in control (Roth et al., 2016).
5. Universal Expected Policy Gradients
Having covered the most common case of continuous Gaussian policies, we now extend
the analysis to other policy classes. We provide two cases of such results in the following
sections: exponential family policies with multivariate polynomial critics (of arbitrary order)
and arbitrary policies (possessing a mean) with linear critics. Our main claim is that an
analytic solution to the EPG integral is possible for almost any system; hence we describe
EPG as a universal method.7
5.1 Exponential Family Policies and Polynomial Critics
We now describe a general technique to obtain analytic EPG updates for the case when
the policy belongs to a certain exponential family and the critic is an arbitrary polynomial.
This result is significant since polynomials can approximate any continuous function on a
bounded interval with arbitrary accuracy (Weierstrass, 1885; Stone, 1948). Since our result
holds for a nontrivial class of distributions in the exponential family, it implies that analytic
solutions for EPG can almost always be obtained in practice and hence that the Monte
Carlo sampling to estimate the inner integral that is typical in SPG is rarely necessary.
Lemma 8 (EPG for Exponential Families with Polynomial Sufficient Statistics)
Consider the class of policies parameterised by θ where:
>
π(a | s) = eηθ T (a)−Uηθ +W (a) ,
(15)
where each entry in the vector T (a) is a (possibly multivariate) polynomial in the entries of
the vector a. Moreover, assume that the critic Q̂(a) is (a possibly multivariate) polynomial
7. Of course no method can be truly universal for a completely arbitrary problem. Our claim is that EPG is
universal for the class of systems arising from lemmas in this section. However, this class is so broad
that we feel the term ‘universal‘ is justified. This is similar to the claim that neural networks based on
sigmoid nonlinearities are universal, even though then can only approximate continuous functions, as
opposed to completely arbitrary ones.
16
Expected Policy Gradients
in the entries of a. Then, the policy gradient update is a closed form expression in terms of
the uncentered moments of π(· | s):
>
IπQ (s) = (∇ηθ> )(CT>Q mπ ) − (∇Uηθ )(CQ
mπ ),
(16)
where CQ is the vector containing the coefficients of the polynomial Q, CT Q is the vector
containing the coefficients of the polynomial T (a)Q(a) (i.e., a multiplication of T and Q)
and mπ is a vector of uncentered moments of π (in the order matching the polynomials).
Proof We first rewrite the inner integral as an expectation.
Z
Q
Iπ (s) =
dπ(a | s)∇ log π(a | s)Q(a)
A
= Ea∼π [∇ log π(a | s)Q(a)]
h
i
= Ea∼π (∇(ηθ> T (a) − Uηθ + W (a)))Q(a)
i
h
= Ea∼π (∇ηθ> )T (a)Q(a) − (∇Uηθ )Q(a)
= (∇ηθ> )Ea∼π [T (a)Q(a)] − (∇Uηθ )Ea∼π [Q(a)] .
Since T (a) and Q(a) are polynomials, and the multiplication of polynomials is still polynomial,
both expectations are expectations of polynomials.
To compute the second expectation, we exploit the fact that, since Q is a polynomial, it
is a sum of monomial terms:
D
d
D
d
X
Y
X
Y
p (j)
p (j)
Ea∼π [Q(a)] = Ea∼π
ci
aj i =
ci Ea∼π
aj i
i=1
j=1
i=1
j=1
|
{z
cross-moment of π
}
hQ
i
pi (j)
d
On the right, the terms Ea∼π
a
, are the uncentered (pi (1), pi (2), . . . , pi (d))-crossj=1 j
moments of π. If we arrange the coefficients ci into the vector CQ and the cross-moments
into the vector mπ , we obtain the right term in (16). We can apply the same reasoning to
the product of T and Q to obtain the left term.
The cross-moments themselves can be obtained from the moment generating function
(MGF) of π. Indeed, for a distribution of the form of (15), the MGF of T (a) is guaranteed
to exist and has a closed form (Bickel and Doksum, 2006). Hence, the computation of the
moments reduces to the computation of derivatives. See details in Appendix A.2.
Note that the assumption that T and Q are polynomial is with respect to the action
a. The dependence on the state only appears in ηθ and Uηθ and can be arbitrary, e.g., a
multi-layered neural network.
Of course, while polynomials are universal approximators, they may not be the most
efficient or stable ones. The importance of Lemma 8 is currently mainly conceptual—analytic
EPG is possible for a universal class of approximators (polynomials) which shows that EPG
17
Ciosek and Whiteson
is analytically tractable in principle for any continuous Q-function.8 It is an open research
question whether more suitable universal approximators admitting analytic EPG solutions
can be identified.
5.2 Reparameterised Exponential Families and Reparameterised Critics
In Lemma 8, we assumed that the function T (a) (called the sufficient statistic of the
exponential family) is polynomial. We now relax this assumption. Our approach is to start
with a policy πb which does have a polynomial sufficient statistic and then introduce a
suitable reparameterisation function g : Rd → A. The policy is then defined as:
>
equivalent to a = g(b) with b ∼ πb (b | s) = eηθ T (b)−Uηθ +V (b) ,
a ∼ π(a | s)
where b is the random variable representing the action before the squashing. Assuming that
g −1 exists and the Jacobian ∇g is non-singular almost everywhere, the PDF9 of the policy
π can be written as:
π(a | s) = πb (g −1 (a) | s)
1
1
= πb (b | s)
.
−1
det ∇g(g (a))
det ∇g(b)
(17)
The following lemma develops an EPG method for such policies.
Lemma 9 Consider an invertible and differentiable function g. Define a policy π as in (17).
Assume that the Jacobian of g is nonsingular except on a set of πb -measure zero. Consider
a critic Q. Denote as Qb a reparameterised critic such that for all a, Qb (g −1 (a)) = Q(a).
Then the policy gradient update is given by the formula IπQ (s) = IπQbb (s).
Proof
IπQ (s)
Z
dπ(a | s)∇ log π(a | s)Q(a)
=
ZA
dπ(g(b) | s)∇ log π(g(b) | s)Q(g(b)) det ∇g(b)
=
ZR
d
dπb (b | s)∇ log π(g(b) | s)Qb (b)
=
d
ZR
=
Rd
dπb (b | s)(∇ log πb (b | s) − ∇ log det ∇g(b))Qb (b) = IπQbb (s).
|
{z
}
0
In the second equality, we perform the variable substitution a = g(b). In the third equality
we use (17) and the fact that Qb (g −1 (a)) = Q(a). In the fourth equality we again use (17)
and the fact that log det ∇g(b) = 0 since g is not parameterised by θ.
8. The universality of polynomials holds only for bounded intervals (Weierstrass, 1885), while the support
of the policy may be unbounded. We do not address the unbounded approximation case here other than
by saying that, in practice, the critic is learned from samples and is thus typically only accurate on a
bounded interval anyway.
9. We abuse notation slightly by using π(a | s) for both the probability distribution and its PDF.
18
Expected Policy Gradients
We are now ready to state our universality result. The idea is to obtain a reparameterised
version of EPG (and Lemma 8) by reparameterising the critic and the policy using the same
transformation g. We do so in the following corollary, which is the most general constructive
result in this article.
Corollary 10 (EPG for Exponential Families with Reparameterisation) Consider
the class of policies, parameterised by θ, defined as in (15). Consider reparameterisation
function g and define Tb , Vb and Qb as Tb (g −1 (a)) = T (a), Wb (g −1 (a)) = W (a) and
Qb (g −1 (a)) = Q(a) for every a. Assume the following:
1. g is invertible;
2. The Jacobian of g exists and is nonsingluar except on a set of πb -measure zero, where
πb is the reparameterised policy as in (17); and
3. Tb and Qb are polynomial as in Lemma 8.
Then a closed-form policy gradient update can be obtained as follows:
>
IπQ (s) = (∇ηθ> )(CT>b Qb mπb ) − (∇Uηθ )(CQ
mπb ).
b
(18)
Proof Apply Lemmas 9 and then 8.
Lemma 9 also has a practical application in case we want to deal with bounded action
spaces. As we discussed in Section 4.3, hard clipping can cause the problem of vanishing
gradients and the default solution should be to use GPG. In case we can’t use GPG, for
instance when the dimensionality of the action space is so large that computing the covariance
of the policy is too costly, we can alleviate the vanishing gradients problem by using a
strictly monotonic squashing function g. One implication of Lemma 9 is that, if we set πb
to be Gaussian, we can invoke Lemma 5 to obtain exact analytic updates for useful policy
classes such as Log-Normal and Logit-Normal (obtained by setting g to the sigmoid and the
exponential function respectively), as long as we choose our critic Q to be quadric in g −1 (a),
i.e., Qb is quadric in b. The reparameterised version of EPG is the same as Algorithm 4
except it uses a squashing function g instead of the clipping function c.
5.3 Aribtrary Policies and Linear Critics
Next, we consider the case where the stochastic policy is almost completely arbitrary, i.e., it
only has to possess a mean and need not even be in the already general exponential family
of policies used in Lemma 8 and Corollary 10, but the critic is constrained to be linear in
the actions. We have the following lemma, which is a slight modification of an observation
made in connection with the Q-Prop algorithm (Gu et al., 2016a, Eq. 7).
Lemma 11 (EPG for Arbitrary Stochastic Policies and Linear Critics) Consider an
arbitrary (nondegenerate) probability distribution π(· | s) which has a mean. Assume that
the critic Q̂(a) is of the form A>
s a for some coefficient vector As . Then the
R policy gradient
Q
>
update is given by Iπ (s) = As ∇µπ(·|s) where µπ(·|s) denotes the integral a adπ(a | s) (the
mean).
19
Ciosek and Whiteson
Proof
IπQ (s)
Z
∇π(a | s)Q(a | s)da
=
Za
∇π(a | s)A>
s ada
a
Z
π(a | s)ada
= A>
∇
s
a
|
{z
}
=
µπ(·|s)
= A>
s ∇(µπ(·|s) ).
Since DPG already provides the same result for Dirac-delta policies (see Corollary 3), we
conclude that using linear critics means we can have an analytic solution for any reasonable
policy class.
To see why the above lemma is useful, first consider systems that arise as a discretisation
of continuous time systems with a fine-enough time scale. If we assume that the true Q
is smooth in the actions and that the magnitude of the allowed action goes to zero as
the time step decreases, then a linear critic is sufficient as an approximation of Q because
we can approximate any smooth function with a linear function in any sufficiently small
neighbourhood of a given point and then choose the time step to be small enough so an
action does not leave that neighbourhood. We can then use Lemma 11 to perform policy
gradients with any policy.10
5.4 If All Else Fails: EPG with Numerical Quadrature
If, despite the broad framework shown in this article, an analytical solution is impossible,
we can still perform integration numerically. EPG can still be beneficial in these cases: if
the action space is low dimensional, numerical quadrature is cheap; if it is high dimensional,
it is still often worthwhile to balance the expense of simulating the system with the cost of
quadrature. Actually, even in the extreme case of expensive quadrature but cheap simulation,
the limited resources available for quadrature could still be better spent on EPG with smart
quadrature than SPG with simple Monte Carlo.
The crucial insight behind numerical EPG is that the integral given as
Z
Q̂
Iπ = dπ(a | s)∇ log π(a | s)Q̂(a, s)
a
only depends on two fully known quantities: the current policy π and the current approximate
critic Q̂. Therefore, we can use any standard numerical integration method to compute it.
The actions at which the integrand is evaluated do not have to be sampled—one can also
use a method such as the Gauss-Legendre quadrature where the abscissae are designed.
10. Of course the update derived in Lemma 11 only provides a direction in which to change the policy mean
(which means that exploration has to be performed using some other mechanism). This is because a
linear critic does not contain enough information to determine exploration.
20
Expected Policy Gradients
6. Experiments
While EPG has many potential uses, we focus on empirically evaluating one particular
application: exploration driven by the Hessian exponential (as introduced in Algorithm 2 and
Lemma 6), replacing the standard Ornstein-Uhlenbeck (OU) exploration in continuous action
domains. To this end, we apply EPG to five domains modelled with the MuJoCo physics
simulator (Todorov et al., 2012): HalfCheetah-v1, InvertedPendulum-v1, Reacher2d-v1,
Walker2d-v1, and InvertedDoublePendulum-v1 and compare its performance to DPG and
SPG. The experiments described here extend our previous conference work (Ciosek and
Whiteson, 2018) in two ways: we added the InvertedDoublePendulum-v1 domain and used
it for a detailed comparison with the PPO algorithm (Schulman et al., 2017).
In practice, EPG differs from deep DPG (Lillicrap et al., 2015; Silver et al., 2014) only
in the exploration strategy, though their theoretical underpinnings are also different. The
hyperparameters for DPG and those of EPG that are not related to exploration were taken
from an existing benchmark (Islam et al., 2017; Brockman et al., 2016). The exploration
hyperparameters for EPG were σ0 = 0.2 and c = 1.0 where the exploration covariance is
σ0 ecH . These values were obtained using a grid search from the set {0.2, 0.5, 1} for σ0 and
{0.5, 1.0, 2.0} for c over the HalfCheetah-v1 domain. Since c is just a constant scaling the
rewards, it is reasonable to set it to 1.0 whenever reward scaling is already used. Hence,
our exploration strategy has just one hyperparameter σ0 as opposed specifying a pair of
parameters (standard deviation and mean reversion constant) for OU. We used the same
learning parameters for the other domains. For SPG11 , we used OU exploration and a
constant diagonal covariance of 0.2 in the actor update (this approximately corresponds to
the average variance of the OU process over time). The other parameters for SPG are the
same as for the rest of the algorithm. For the learning curves, we obtained 90% confidence
intervals and show results of independent evaluation runs that used actions generated by
the policy mean without any exploration noise.
The Hessian in GPG is obtained using a sigma-point method, as follows. At each step,
the agent samples 100 action values from Q̂(·, s) and a quadric is fit to them in the L2 norm.
Since this is a least-squares problem, it can be accomplished by solving a linear system. The
Hessian computation could be greatly sped up by using an approximate method, or even
skipped completely if we used a quadric critic. However, we did not optimise this part of the
algorithm since the core message of GPG is that a Hessian is useful, not how to compute it
efficiently.
The results in Figure 6 show that EPG’s exploration strategy yields much better performance than DPG with OU. Furthermore, SPG does poorly, solving only the easiest domain
(InvertedPendulum-v1) reasonably quickly, achieving slow progress on HalfCheetah-v1, and
failing entirely on the other domains. This is not surprising since DPG was introduced
precisely to solve the problem of high variance SPG estimates on this type of task. In
InvertedPendulum-v1, SPG initially learns quickly, outperforming the other methods. This
is because noisy gradient updates provide a crude, indirect form of exploration that happens
to suit this problem. Clearly, this is inadequate for more complex domains: even for this
simple domain it leads to subpar performance late in learning.
11. We tried learning the covariance for SPG but the covariance estimate was unstable; no regularisation
hyperparameters we tested matched SPG’s performance with OU even on the simplest domain.
21
Ciosek and Whiteson
(a) HalfCheetah-v1
(b) InvertedPendulum-v1
1000
EPG (5 runs)
DPG (40 runs)
SPG (40 runs)
0
0
5000
EPG (40 runs)
DPG (40 runs)
SPG (40 runs)
40
120
200
280
20
40
(c) Reacher2d-v1
80
(d) Walker2d-v1
EPG (40 runs)
DPG (40 runs)
SPG (10 runs)
0
-15
2000
EPG (5 runs)
DPG (5 runs)
SPG (10 runs)
-5
60
50
150
250
350
450
250
500
750
1000
1250
10000
(e) InvertedDoublePendulum-v1
0
EPG (12 runs)
DPG (36 runs)
SPG (22 runs)
20
40
60
80
Figure 6: Learning curves (mean and 90% interval). Returns for Reacher2d-v1 are clipped
at -14. The number of independent training runs is in parentheses. Horizontal axis is scaled
in thousands of steps.
In addition, EPG typically learns more consistently than DPG with OU. In three tasks,
the empirical standard deviation across runs of EPG (σ̂EPG ) was substantially lower than
that of DPG (σ̂DPG ) at the end of learning, as shown in Table 1. For the other two domains,
the confidence intervals around the empirical standard deviations for DPG and EPG were
too wide to draw conclusions.
Surprisingly, for InvertedPendulum-v1, DPG’s learning curve declines late in learning.
The reason can be seen in the individual runs shown in Figure 7: both DPG and SPG suffer
from severe unlearning. This unlearning cannot be explained by exploration noise since the
22
Expected Policy Gradients
Domain
HalfCheetah-v1
σ̂DPG
1336.39
[1107.85, 1614.51]
[875.54, 1275.94]
InvertedPendulum-v1
291.26
0.00
[241.45, 351.88]
n/a
Reacher2d-v1
1.22
0.13
Walker2d-1
InvertedDoublePendulum-v1
σ̂EPG
1056.15
[0.63, 2.31]
[0.07, 0.26]
543.54
762.35
[450.58, 656.65]
[631.98, 921.00]
921.73
226.07
[756.01, 1125.63]
[157.20, 326.00]
10
30
50
70
90
1000
0
1000
0
0
1000
Table 1: Estimated standard deviation (mean and 90% interval) across runs after learning.
10
30
50
70
90
10
30
50
70
90
Figure 7: Three runs for EPG (left), DPG (middle) and SPG (right) for the
InvertedPendulum-v1 domain, demonstrating that EPG shows much less unlearning.
evaluation runs just use the mean action, without exploring. Instead, OU exploration in
DPG may be too coarse, causing the optimiser to exit good optima, while SPG unlearns due
to noise in the gradients. The noise also helps speed initial learning, as described above, but
this does not transfer to other domains. EPG avoids this problem by automatically reducing
the noise when it finds a good optimum, i.e., a Hessian with large negative eigenvalues as
described is Section 4.2.
The fact that EPG is stable in this way raises the question whether the instability of an
algorithm (i.e., an inverted or oscillating learning curve) is caused primarily by inefficient
exploration or by excessivly large differences between subsequent policies. To address it,
we compare our results with proximal policy pptimisation (PPO) (Schulman et al., 2017), a
policy gradient algorithm designed specifically to include a term penalising the difference
between successive policies. Comparing our EPG result for InvertedDoublePendulum-v1
in Figure 6e with PPO (Schulman et al., 2017, Figure 3, first row, third plot from left,
blue PPO curve), it is clear that EPG is more stable. This suggests that efficient adaptive
exploration of the type used by EPG is important for stability, even in this relatively simple
domain.
7. Related Work
In this section, we discuss the relationship between EPG and several other methods.
23
Ciosek and Whiteson
7.1 Sampling Methods for SPG
EPG has some similarities with VINE sampling (Schulman et al., 2015), which uses an
(intrinsically noisy) Monte Carlo quadrature with many samples. However, there are
important differences. First, VINE relies entirely on reward rollouts and does not use an
explicit critic. This means that VINE has to perform many independent rollouts of Q(·, s)
for each s, requiring a simulator with reset. A second, related difference is that VINE uses
the same actions in the estimation of IπQ̂ that it executes in the environment. While this is
necessary with purely Monte Carlo rollouts, Section 5.4 shows that there is no such need in
general if we have an explicit critic. Ultimately, the main weakness of VINE is that it is
a purely Monte Carlo method. However, the example in Figure 1 (Section 3) shows that
even with a computationally expensive many-sample Monte Carlo method, the problem of
variance in the gradient estimator remains, regardless of the baseline.
EPG is also related to variance minimisation techniques that interpolate between two
estimators (Gu et al., 2016a). However, EPG uses a quadric (not linear) critic, which is
crucial for exploration. Furthermore, it completely eliminates variance in the inner integral,
as opposed to just reducing it.
A more direct way of coping with variance in policy gradients is to simply reduce the
learning rate when the variance of the gradient would otherwise explode, using, e.g., Adam
(Kingma and Ba, 2014), natural policy gradients (Kakade, 2002; Amari, 1998; Peters and
Schaal, 2008a), trust region policy optimisation (Schulman et al., 2015), proximal policy
optimisation (Schulman et al., 2017), the adaptive step size method (Pirotta et al., 2013) or
Newton’s method (Furmston and Barber, 2012; Furmston et al., 2016; Parisi et al., 2016).
However, this results in slow learning when the variance is high.
7.2 Sarsa and Q-Learning
It has been known since the introduction of policy gradient methods (Sutton et al., 2000)
that they represent a kind of slow-motion policy improvement as opposed to a greedy
improvement performed by methods such as (expected) sarsa. The two main reasons for the
slow-motion improvement are that a greedy maximisation operator may not be available
(e.g., for continuous or large discrete action spaces) and that a greedy step may be too
large because the critic only approximates the value function. The argument for a sarsa-like
method is that it may converge faster and does not need an additional optimisation for the
actor. Recently, approaches combining the features of both methods have been investigated.
One-step Newton’s method for Q-functions that are quadric in the actions has been used to
produce a sarsa-like algorithm for continuous domains (Gu et al., 2016b), previously only
tractable with policy gradient methods. For discrete action spaces, softmax Q-learning, a
family of methods with a hybrid loss combining sarsa and Q-learning, has recently been
linked to policy gradients via an entropy term(O’Donoghue et al., 2016). In this paper,
GPG with Hessian-based exploration (Section 4.2) can be seen as another kind of hybrid.
Specifically, it changes the mean of the policy slowly, similar to a vanilla policy gradient
method, and computes the covariance greedily, similar to sarsa.
24
Expected Policy Gradients
7.3 DPG
The update for the policy mean obtained in Corollary 7 is the same as the DPG update,
linking the two methods:
IπQ (s) = [∇a Q(a, s)]a=µs ∇µs .
We now formalise the equivalences between EPG and DPG. First, any EPG method with
a linear critic (or an arbitrary critic approximated by the first term in the Taylor expansion)
is equivalent to DPG with actions from a given state s drawn from an exploration policy of
the form:
a ∼ π(s) + n(a|s),
where
Ea∼n [a | s] = 0.
Here, the PDF of the zero-mean exploration noise n(·|s) must not depend on the policy
parameters. This fact follows directly from Lemma 11, which says that, in essence, a linear
critic only gives information on how to shift the mean of the policy and no information
about other moments. Second, on-policy GPG with a quadric critic (or an arbitrary critic
approximated by the first two terms in the Taylor expansion) is equivalent to DPG with a
Gaussian exploration policy where the covariance is computed as in Section 4.2. This follows
from Corollary 7. Third, and most generally, for any critic at all (not necessarily quadric),
DPG is a kind of EPG for a particular choice of quadrature (using a Dirac measure). This
follows from Theorem 1.
Surprisingly, this means that DPG, normally considered to be off-policy, can also be seen
as on-policy when exploring with Gaussian noise defined as above for the quadric critic or
any noise for the linear critic. Furthermore, the compatible critic for DPG (Silver et al.,
2014) is indeed linear in the actions. Hence, this relationship holds whenever DPG uses a
compatible critic.12 Furthermore, Lemma 5 lends new legitimacy to the common practice of
replacing the critic required by the DPG theory, which approximates ∇a Q, with one that
approximates Q itself, as done in SPG and EPG.
7.4 Entropy-Based Methods
On-policy SPG sometimes includes an entropy term (Peters et al., 2010) in the gradient
in order to aid exploration by making the policy more stochastic. The gradient of the
differential entropy H(s) of the policy at state s is defined as follows.13
R
−∇H(s) = ∇ a dπ(a|s) log π(a|s)
R
R
= a da∇π(a|s) log π(a|s) + a dπ(a|s)∇ log π(a|s)
R
R
1
= a da∇π(a|s) log π(a|s) + a dπ(a|s) π(a|s)
∇π(a|s)
R
R
= a da∇π(a|s) log π(a|s) + ∇ a dπ(a|s)
| {z }
1
R
R
= a da∇π(a|s) log π(a|s) = a dπ(a|s)∇ log π(a|s) log π(a|s).
12. The notion of compatibility of a critic is different for stochastic and deterministic policy gradients.
13. For discrete action spaces, the same derivation with integrals replaced by sums holds for the entropy.
25
Ciosek and Whiteson
Typically, we add the entropy update to the policy gradient update with a weight α:
E
IG
(s) = IG (s) + α∇H(s)
R
= a dπ(a|s)∇ log π(a|s)(Q(a, s) − α log π(a|s)).
(19)
This equation makes clear that performing entropy regularisation is equivalent to using a
different critic with Q-values shifted by α log π(a|s). This holds for both EPG and SPG,
including SPG with discrete actions where the integral over actions is replaced with a
sum. This follows because adding entropy regularisation to the objective of optimising the
total discounted reward in an RL setting corresponds to shifting the reward function by a
term proportional to log π(a|s) (Neu et al., 2017; Nachum et al., 2017). Indeed, the path
consistency learning algorithm (Nachum et al., 2017) contains a formula similar to (19),
though we obtained ours independently.
Next, we derive a further specialisation of (19) for the case where the parameters θ are
shared between the actor and the critic. We start with the policy gradient identity given
by (9) and replace the true critic Q with the approximate critic Q̂. Since this holds for any
stochastic policy, we choose one of the form:
1 Q̂(a,s)
π(a|s) =
e
,
Z(s)
Z
where
Z(s) =
eQ̂(a,s) da.
(20)
a
For the continuous case, we assume that the integral in (20) converges for each state. Here,
we assume that the approximate critic is parameterised by θ. Because of the form of (20),
the policy is parameterised by θ as well. Now, for the policy class given by (20), we can
simplify the gradient update even further, obtaining:
E
IG
(s) =
R
a dπ(a|s)∇ log π(a|s)(Q̂(a, s)
=
R
a π(a|s)∇ log π(a|s)(Q̂(s, a)
− α log π(a|s))
− α log eQ̂(a,s) −α log Z(s))
| {z }
Q̂(a,s)
= (1 − α)
R
a π(a|s)∇ log π(a|s)Q̂(s, a)
= −(1 − α)∇H(s).
In the above derivation, we could drop the term log Z(s) since it does not depend on a, as
with a baseline. This shows that, in the case of sharing parameters between the critic and
the policy as above, methods such as A3C (Mnih et al., 2016), which have both an entropy
loss and a policy gradient loss, are redundant since entropy regularisation does nothing
except scale the learning rate.14 Alternatively, for this shared parameterisation, a policy
gradient method simply subtracts entropy from the policy. In practice, this means that
a policy gradient method with this kind of parameter sharing is quite similar to learning
the critic alone and simply acting according to the argmax of the Q values rather than
representing the policy explicitly, producing a method similar to sarsa.
14. In this argument, we ignore the effects of sampling on exploration.
26
Expected Policy Gradients
7.5 Off-Policy Actor-Critic
Off-policy learning with policy gradients typically follows the framework of off-policy actorcritic (Degris et al., 2012). Denote the behaviour policy as b(a | s) and the corresponding
discounted-ergodic measure as ρb . The method uses the following reweighting approximation:
Z
Z
∇J = dρ(s) dπ(a | s)∇ log π(a | s)(Q(a, s))
Za
Zs
≈ dρb (s) dπ(a | s)∇ log π(a | s)(Q(a, s)).
(21)
s
a
The approximation is necessary since, as the samples are generated using the policy b, it is
not known how to approximate the integral with ρ from samples, while it is easy to do so for
an integral with ρb . A natural off-policy version of EPG emerges from this approximation
(see Algorithm 5), which simply replaces the inner integral with IπQ :
Z
Z
Z
dρb (s) dπ(a | s)∇ log π(a | s)(Q(a, s)) = dρb (s)IπQ (s).
(22)
s
a
s
Here, we use an analytic solution to IπQ (s) as before. The importance sampling term π(a|s)
b(a|s)
does not appear because, as the integral is computed analytically, there is no sampling in
IπQ (s), much less sampling with an importance correction. Of course, the algorithm also
requires an off-policy critic for which an importance sampling correction is typically necessary.
Indeed, (22) makes clear that off-policy actor-critic differs from SPG in two places: the use
of ρb as in (21) and the use of an importance-sampled Monte Carlo estimator, rather than
regular Monte Carlo, for the inner integral.
Algorithm 5 Off-policy expected policy gradients with reweighting approximation
1: s ← s0 , t ← 0
2: initialise optimiser, initialise policy π parameterised by θ
3: while not converged do
4:
gt ← γ t do-integral(Q̂, s, πθ )
. gt is the estimated policy gradient as per (10)
5:
θ ← θ + optimiser.update(gt )
6:
a ∼ b(·, s)
7:
s0 , r ← simulator.perform-action(a)
8:
Q̂.update(s, a, r, s0 , π, b)
. Off-policy critic algorithm
9:
t←t+1
10:
s ← s0
11: end while
7.6 Value Gradient Methods
Value gradient methods (Fairbank, 2014; Fairbank and Alonso, 2012; Heess et al., 2015)
assume the same parameterisation of the policy as policy gradients, i.e., π is parameterised
by θ, and maximise J by recursively computing the gradient of the value function. In our
notation, the policy gradient has the following connection with the value gradient of the
27
Ciosek and Whiteson
initial state:
Z
∇J =
dp0 (s0 )∇V (s0 ).
(23)
s0
Value gradient methods use a recursive equation that computes ∇V (s) using ∇V (s0 ) where
s0 is the successor state. In practice, this means that a trajectory is truncated and the
computation goes backward from the last state all the way to s0 , where (23) is applied, so
that the resulting estimate of ∇J can be used to update the policy. The recursive formulae
for ∇V (s) are based on the differentiated Bellman equation:
Z
Z
0
0
∇V = ∇ dπ(a|s) R(s, a) + γ
p(s |a, s)V (s ) .
(24)
s0
a
Different value gradient methods differ in the form of the recursive update for the value
gradient obtained from (24). For example, stochastic value gradients (SVG) introduce a
reparameterisation both of π and p(s0 |a, s):
s0 ∼ p(·|a, s)
⇔
s0 = f (a, s, ξ) with ξ ∼ B1 ,
a ∼ π(·|s)
⇔
a = h(s, η) with η ∼ B2 .
Here, we denote the base noise distributions as B1 and B2 , while f and h are deterministic
functions. The function f can be thought of as an MDP transition model. SVG rewrites
(24) using the reparameterisation as follows:
∇V = ∇
=
R
η
R
η
dB2 (η) R(s, h(s, η)) + γ
R
dB2 (η) ∇R(s, h(s, η)) + γ
s0
R
ξ
dB1 (ξ)V (f (h(s, η), s, ξ)) =
dB1 (ξ)∇V (f (h(s, η), s, ξ)) .
|
{z
}
(25)
s0
Here, the quantities ∇R(s, h(s, η)) and ∇V (f (h(s, η), s, ξ)) can be computed by the chain
rule from the known reward model R, transition model f . SVG learns the approximate
ˆ η̂ from samples and using a sample-based approximation to (25) to obtain the
model fˆ, R̂, ξ,
value gradient recursion.
By contrast, we now derive a related but simpler value gradient method that does not
require a model or a reparameterised policy15 , starting with (24).
R
R
∇V (s) = ∇ a dπ(a|s) R(s, a) + γ s0 p(s0 |a, s)V (s0 )
R
R
R
= a da∇π(a|s)R(s, a) + γ∇π(a|s) s0 p(s0 |a, s)V (s0 ) + π(a|s)∇ s0 p(s0 |a, s)V (s0 )
R
= a,s0 π(a|s)p(s0 |a, s) (∇ log π(a|s)R(s, a) + ∇V (s0 ) + ∇ log π(a|s)V (s0 )) .
(26)
Now (26) can be approximated from samples:
ˆ (s) ≈ ∇ log π(a|s)R(s, a) + ∇V
ˆ (s0 ) + ∇ log π(a|s)V̂ (s0 ) .
∇V
15. SVG(∞) and SVG(1) require a model an a policy reparameterisation while SVG(0) requires only a
policy reparameterisation. However, SVG(0) is inefficient since it does not directly use the reward in the
computation of the value gradient.
28
Expected Policy Gradients
Policy Class
Normal, a ∈ Rd
Logit-Normal; a ∈ [0, 1]d
Log-Normal; a ∈ [0, ∞]d
any policy
Squashing
none
a = expit(b)
a = eb
none
Q̂
Analytic Update
) Q
a> As a + a> Bs
Iπ(s),µs = (∇µs )Bs ,
b> As b + b> Bs
1/2
1/2
IQ
)Σs As
1/2 = (∇Σs
b> As b + b> B s
π(s),Σs
Bs> a
IπQ (s) = Bs> ∇µπ(·|s)
Table 2: A summary of the most useful analytic results for expected policy gradients. For
bounded action spaces, we assume that the bounding interval is [0, 1] or [0, ∞].
Here, the pair (a, s0 ) corresponds to the action taken at s and the successor state. This
method requires learning a critic, while SVG requires a model.
An additional connection between value gradient methods and policy gradients
R is that,
since the quantity IG (s) in Theorem 1 can be written as IG (s) = ∇V (s) − γ s0 dpπ (s0 |
s)∇V (s0 ), we can think of this theorem as showing how to obtain a policy gradient from a
value gradient without backwards iteration.
8. Conclusions
This paper proposed a new framework for reasoning about policy gradient methods called
expected policy gradients (EPG) that integrates across the action selected by the stochastic
policy, thus reducing variance compared to existing stochastic policy gradient methods. We
proved a new general policy gradient theorem subsuming the stochastic and deterministic
policy gradient theorems, which covers any reasonable class of policies. We showed that
analytical results for the policy update exist and, in the most common cases, lead to
a practical algorithm (the analytic updates are summarised in Table 2). We also gave
universality results which state that, under certain broad conditions, the quadrature required
by EPG can be performed analytically. For Gaussian policies, we also developed a novel
approach to exploration that infers the exploration covariance from the Hessian of the critic.
The analysis of EPG yielded new insights about DPG and delineated the links between
the two methods. We also discussed the connections between EPG and other common RL
techniques, notably sarsa, Q-learning and entropy regularisation. Finally, we evaluated the
GPG algorithm in six practical domains, showing that it outperforms existing techniques.
Acknowledgments
This project has received funding from the European Research Council (ERC) under the
European Union’s Horizon 2020 research and innovation programme (grant agreement
number 637713). The experiments were made possible by a generous equipment grant from
NVIDIA.
29
Ciosek and Whiteson
Appendix A.
A.1 Proofs and Detailed Definitions
First, we prove two lemmas concerning the discounted-ergodic measure ρ(s) which have
been implicitly realised for some time but as far as we could find, never proved explicitly.
Definition 12 (Time-dependent occupancy)
p(s | t = 0) = p0 (s)
Z
p(s0 | t = i + 1) = p(s0 | s)p(s | t = i)
for
i≥0
s
Definition 13 (Truncated trajectory) Define the trajectory truncated after N steps as
τ N = (s0 , a0 , r0 , s1 , a1 , r1 , . . . , sN ).
Observation 14 (Expectation
wrt. truncated trajectory) Since τN = (s0 , s1 , s2 , . . . , sN )
QN −1
is associated with the density i=0 p(si+1 | si )p0 (s0 ), we have that
hP
i
N
i f (s ) =
EτN
γ
i
i=0
Q
P
R
N −1
N
i f (s ) ds ds . . . ds =
= s0 ,s1 ,...,sN
p(s
|
s
)
p
(s
)
γ
i+1
i
0
0
i
0 1
N
i=0
i=0
P R
QN −1
i
= N
i=0 s0 ,s1 ,...,sN p0 (s0 )
i=0 p(si+1 | si ) γ f (si )ds0 ds1 . . . dsN =
P R
i
= N
i=0 s p(s | t = i)γ f (s)ds
for any function f .
Definition 15 (Expectation with respect to infinite trajectory) For any bounded function f , we have
"∞
#
"N
#
X
X
i
i
Eτ
γ f (si ) , lim EτN
γ f (si ) .
N →∞
i=0
i=0
Here, the sum on the left-hand side is part of the symbol being defined.
Observation 16 (Property of expectation with respect to infinite trajectory)
hP
i
P∞ i
N
i f (s ) =
Eτ
γ
f
(s
)
=
lim
E
γ
i
i
N →∞ τN
i=0
i=0
PN R
= limN →∞ i=0 s p(s | t = i)γ i f (s)ds =
∞ Z
X
=
dp(s | t = i)γ i f (s)
i=0
s
for any bounded function f .
Definition 17 (Discounted-ergodic occupancy measure ρ)
ρ(s) =
∞
X
γ i p(s | t = i)
i=0
30
Expected Policy Gradients
The measure ρ is not normalised in general. Intuitively, it can be thought of as ‘marginalising
out’ the time in the system dynamics.
Lemma 18 (Discounted-ergodic property) For any bounded function f :
"
Z
ρ(s)f (s) = Eτ
s
∞
X
#
γ i f (si ) .
i=0
Proof
Eτ
"∞
X
#
i
γ f (si ) =
i=0
∞
X
γ
i
Z
p(s | t = i)f (s)ds =
s
i=0
Z "X
∞
s
#
i
γ p(s | t = i) f (s)ds
| i=0
{z
ρ(s)
}
Here, the first equality follows from Observation 16.
This property is useful since the expression on the left can be easily manipulated while the
expression on the right can be estimated from samples using Monte Carlo.
Lemma 19 (Generalised eigenfunction property) For any bounded function f :
Z
Z
γ
0
Z
dp(s | s)f (s ) =
dρ(s)
s0
s
0
s
Z
dρ(s)f (s) −
dp0 (s)f (s)
s
Proof
γ
R
s dρ(s)
R
s0
R
P
i
0
0
0
dp(s0 | s)f (s0 ) = γ ∞
i=0 γ s,s0 p(s | t = i)p(s | s)f (s )dsds =
P∞ i+1 R
0
0
= i=0 γ
0 dp(s | t = i + 1)f (s )
P∞ i R s 0
= i=1 γ s0 dp(s | t = i)f (s0 )
R
P∞ i R
0
0
=
dp
(s)f
(s)
0
i=0 γ s0 dp(s | t = i)f (s ) −
s
R
R
= s dρ(s)f (s) − s dp0 (s)f (s)
Here, the first equality follows form definition 17, the second one from definition 12. The
last equality follows again from definition 17.
Definition 20 (Markov Reward Process) A Markov Reward Process is a tuple (p, p0 , R, γ),
where p(s0 |s) is a transition kernel, p0 is the distribution over initial states, R(·|s) is a reward
distribution conditioned on the state and γ is the discount constant.
An MRP can be thought of as an MDP
R with a fixed policy and dynamics given by
marginalising out the actions pπ (s0 | s) = a dπ(a | s)p(s0 | a, s). Since this paper considers
the case of one policy, we abuse notation slightly by using the same symbol τ to denote trajectories including actions, i.e. (s0 , a0 , r0 , s1 , a1 , r1 , . . . ) and without them (s0 , r0 , s1 , r1 , . . . ).
31
Ciosek and Whiteson
Lemma 21 (Second Moment Bellman Equation) Consider a Markov Reward Process
(p, p0 , X, γ) where p(s0 | s) is a Markov process and X(· | s) is some probability density
function16 . Denote the value function of the MRP as V . Denote the second moment function
S as
!2
∞
X
S(s) = Eτ
γ t xt
s0 = s xt ∼ X(· | st ).
t=0
Then S is the value function of the MRP: (p, p0 , u, γ 2 ), where u(s) is a deterministic random
variable given by
2
u(s) = VX(x|s) [x] + EX(x|s) [x] + 2γEX(x|s) [x] Ep(s0 |s) V (s0 ) .
Proof
i
tx 2 s = s
γ
t
0
t=1
h
i
P
P∞ t 2
∞
tx +
= Eτ x20 + 2x0
γ
γ
x
s
=
s
t
t
0
t=1
t=1
h P
i
2
P∞ t
∞
tx 2 s = s
= Eτ x0 s0 = s + Eτ 2x0
γ
x
s
=
s
+
+
E
γ
t
0
τ
t
0
t=1
t=1
{z
}
|
|
{z
}
S(s) = Eτ
h
x0 +
P∞
u(s)
γ 2 Ep(s0 |s) [S(s0 )]
This is exactly the Bellman equation of the MRP (p, p0 , u, γ 2 ). The theorem follows since
the Bellman equation uniquely determines the value function.
Observation 22 (Dominated Value Functions) Consider two Markov Reward Processes
(p, p0 , X1 , γ) and (p, p0 , X2 , γ), where p(s0 | s) is a Markov process (common to both
MRPs) and X1 (s), X2 (s) are some deterministic random variables meeting the condition X1 (s) ≤ X2 (s) for every s. Then the value functions V1 and V2 of the respective MRPs
satisfy V1 (s) ≤ V2 (s) for every s. Moreover, if we have that X1 (s) < X2 (s) for all states,
then the inequality between value functions is strict.
Proof Follows trivially by expanding the value function as a series and comparing series
elementwise.
A.2 Computation of Moments for an Exponential Family
Consider the moment generating function of T (a), which we denote as MT , for the exponential
family of the form given in Equation 15.
MT (v) = eUv+ηθ −Uηθ
16. Note that while X occupies a place in the definition of the MRP usually called ‘reward distribution’, we
are using the symbol X, not R since we shall apply the lemma to Xes which are constructions distinct
from the reward of the MDP we are solving.
32
Expected Policy Gradients
It is well-known that MT is finite in a neighbourhood of the origin (Bickel and Doksum,
2006), and hence the cross moments can be obtained as:
K
Y
∂
p(j)
MT (v)
Ea∼π
T (a)j = p(1)
p(2)
∂ v1 , ∂ v2 , . . . , ∂ p(K) vK
v=0
j=1
Here, we denoted as K the size of the sufficient statistic (i.e. the length of the vector T (a)).
However, we seek the cross-moments of a, not T (a). If T (a) contains a subset of indices
which correspond to the vector a, then we can simply use the corresponding indices in the
above equation. On the other hand, if this is not the case, we can introduce an extended
0> 0
distribution π 0 (a | s) = eη θ T (a)−Uηθ +W (a) ., where T 0 is a vector concatenation of T and
a. We can then use the MGF of T 0 (a), restricted to a suitable set of indices, to get the
moments.
References
Shun-Ichi Amari. Natural gradient works efficiently in learning. Neural computation, 10(2):
251–276, 1998.
K. Asadi, C. Allen, M. Roderick, A.-r. Mohamed, G. Konidaris, and M. Littman. Mean
Actor Critic. ArXiv e-prints, September 2017.
Leemon Baird et al. Residual algorithms: Reinforcement learning with function approximation. In Proceedings of the twelfth international conference on machine learning, pages
30–37, 1995.
Shalabh Bhatnagar, Mohammad Ghavamzadeh, Mark Lee, and Richard S Sutton. Incremental natural actor-critic algorithms. In Advances in neural information processing systems,
pages 105–112, 2008.
Peter Bickel and Kjell Doksum. Mathematical Statistics, Basic Ideas and Selected Topics,
Vol. 1, (2nd Edition). Prentice Hall, 2 edition, 2006. ISBN 0132306379.
Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie
Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016.
Kamil Ciosek and Shimon Whiteson. Expected policy gradients. In AAAI 2018: Proceedings
of the Thirty-Second AAAI Conference on Artificial Intelligence, February 2018.
John L Crassidis and John L Junkins. Optimal estimation of dynamic systems. CRC press,
2011.
Thomas Degris, Martha White, and Richard S Sutton. Off-policy actor-critic. arXiv preprint
arXiv:1205.4839, 2012.
Michael Fairbank. Value-gradient learning. PhD thesis, City University London, 2014.
Michael Fairbank and Eduardo Alonso. Value-gradient learning. In Neural Networks
(IJCNN), The 2012 International Joint Conference on, pages 1–8. IEEE, 2012.
33
Ciosek and Whiteson
Thomas Furmston and David Barber. A unifying perspective of parametric policy search
methods for markov decision processes. In Advances in neural information processing
systems, pages 2717–2725, 2012.
Thomas Furmston, Guy Lever, and David Barber. Approximate newton methods for policy
search in markov decision processes. Journal of Machine Learning Research, 17:1–51,
2016.
Shixiang Gu, Timothy Lillicrap, Zoubin Ghahramani, Richard E Turner, and Sergey
Levine. Q-prop: Sample-efficient policy gradient with an off-policy critic. arXiv preprint
arXiv:1611.02247, 2016a.
Shixiang Gu, Timothy Lillicrap, Ilya Sutskever, and Sergey Levine. Continuous deep qlearning with model-based acceleration. In International Conference on Machine Learning,
pages 2829–2838, 2016b.
Nicolas Heess, Gregory Wayne, David Silver, Tim Lillicrap, Tom Erez, and Yuval Tassa.
Learning continuous control policies by stochastic value gradients. In Advances in Neural
Information Processing Systems, pages 2944–2952, 2015.
Riashat Islam, Peter Henderson, Maziar Gomrokchi, and Doina Precup. Reproducibility of
benchmarked deep reinforcement learning tasks for continuous control. arXiv preprint
arXiv:1708.04133, 2017.
Sham M Kakade. A natural policy gradient. In Advances in neural information processing
systems, pages 1531–1538, 2002.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv
preprint arXiv:1412.6980, 2014.
Michail G Lagoudakis and Ronald Parr. Least-squares policy iteration. Journal of machine
learning research, 4(Dec):1107–1149, 2003.
Weiwei Li and Emanuel Todorov. Iterative linear quadratic regulator design for nonlinear
biological movement systems. In ICINCO (1), pages 222–229, 2004.
Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval
Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement
learning. arXiv preprint arXiv:1509.02971, 2015.
Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap,
Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep
reinforcement learning. In International Conference on Machine Learning, pages 1928–
1937, 2016.
Ofir Nachum, Mohammad Norouzi, Kelvin Xu, and Dale Schuurmans. Bridging the gap
between value and policy based reinforcement learning. arXiv preprint arXiv:1702.08892,
2017.
34
Expected Policy Gradients
Gergely Neu, Anders Jonsson, and Vicenç Gómez. A unified view of entropy-regularized
markov decision processes. arXiv preprint arXiv:1705.07798, 2017.
Brendan O’Donoghue, Remi Munos, Koray Kavukcuoglu, and Volodymyr Mnih. Combining
policy gradient and q-learning. 2016.
Simone Parisi, Matteo Pirotta, and Marcello Restelli. Multi-objective reinforcement learning
through continuous pareto manifold approximation. Journal of Artificial Intelligence
Research, 57:187–227, 2016.
Jan Peters and Stefan Schaal. Policy gradient methods for robotics. In Intelligent Robots
and Systems, 2006 IEEE/RSJ International Conference on, pages 2219–2225. IEEE, 2006.
Jan Peters and Stefan Schaal. Natural actor-critic. Neurocomputing, 71(7):1180–1190, 2008a.
Jan Peters and Stefan Schaal. Reinforcement learning of motor skills with policy gradients.
Neural networks, 21(4):682–697, 2008b.
Jan Peters, Katharina Mülling, and Yasemin Altun. Relative entropy policy search. In
AAAI, pages 1607–1612. Atlanta, 2010.
Matteo Pirotta, Marcello Restelli, and Luca Bascetta. Adaptive step-size for policy gradient
methods. In Advances in Neural Information Processing Systems, pages 1394–1402, 2013.
Martin L Puterman. Markov decision processes: discrete stochastic dynamic programming.
John Wiley & Sons, 2014.
Michael Roth, Gustaf Hendeby, and Fredrik Gustafsson. Nonlinear kalman filters explained:
A tutorial on moment computations and sigma point methods. Journal of Advances in
Information Fusion, 11(1):47–70, 2016.
Gavin A Rummery and Mahesan Niranjan. On-line Q-learning using connectionist systems.
University of Cambridge, Department of Engineering, 1994.
John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust
region policy optimization. In Proceedings of the 32nd International Conference on
Machine Learning (ICML-15), pages 1889–1897, 2015.
John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal
policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin
Riedmiller. Deterministic policy gradient algorithms. In ICML, 2014.
Marshall H Stone. The generalized weierstrass approximation theorem. Mathematics
Magazine, 21(5):237–254, 1948.
Richard S Sutton. Generalization in reinforcement learning: Successful examples using
sparse coarse coding. Advances in neural information processing systems, pages 1038–1044,
1996.
35
Ciosek and Whiteson
Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction, volume 1.
MIT press Cambridge, 1998.
Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy
gradient methods for reinforcement learning with function approximation. In Advances in
neural information processing systems, pages 1057–1063, 2000.
Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for modelbased control. In Intelligent Robots and Systems (IROS), 2012 IEEE/RSJ International
Conference on, pages 5026–5033. IEEE, 2012.
George E Uhlenbeck and Leonard S Ornstein. On the theory of the brownian motion.
Physical review, 36(5):823, 1930.
Harm van Seijen, Hado van Hasselt, Shimon Whiteson, and Marco Wiering. A theoretical and empirical analysis of expected sarsa. In ADPRL 2009: Proceedings of the
IEEE Symposium on Adaptive Dynamic Programming and Reinforcement Learning, pages
177–184, March 2009. URL http://www.cs.ox.ac.uk/people/shimon.whiteson/pubs/
vanseijenadprl09.pdf.
Karl Weierstrass. Über die analytische Darstellbarkeit sogenannter willkürlicher Functionen
einer reellen Veränderlichen. Sitzungsberichte der Königlich Preußischen Akademie der
Wissenschaften zu Berlin, 2:633–639, 1885.
36
| 2cs.AI
|